Initial `llama` Go module

.gitattributes

@@ -1,3 +1,10 @@
 llm/ext_server/* linguist-vendored
+llama/*.cpp linguist-vendored
+llama/*.h linguist-vendored
+llama/*.c linguist-vendored
+llama/*.cu linguist-vendored
+llama/*.cuh linguist-vendored
+llama/ggml/*.cu linguist-vendored
+llama/ggml/*.cuh linguist-vendored
 * text=auto
 *.go text eol=lf

llama/.gitignore

@@ -0,0 +1,4 @@
+*.bin
+*.gguf
+*.lib
+*.exp

llama/README.md

@@ -0,0 +1,73 @@
+# `llama`
+
+This package integrates llama.cpp a Go package.
+
+- [x] CPU
+- [x] avx, avx2
+- [ ] avx512
+- [x] macOS Metal
+- [x] Windows CUDA
+- [x] Windows ROCm
+- [ ] Linux CUDA
+- [ ] Linux ROCm
+
+Extra build steps are required for CUDA and ROCm on Windows since `nvcc` and `hipcc` both require using msvc as the host compiler. For these small dlls are created:
+
+- `ggml-cuda.dll`
+- `ggml-hipblas.dll`
+
+> Note: it's important that memory is allocated and freed by the same compiler (e.g. msvc or mingw code). Issues from this should be rare, but there are some places where pointers are returned by the CUDA or HIP runtimes and freed elsewhere, causing a a crash. In a future change the same runtime should be used in both cases to avoid crashes.
+
+## Building
+
+```
+go build .
+```
+
+### AVX
+
+```shell
+go build -tags avx .
+```
+
+### AVX2
+
+```shell
+# go doesn't recognize `-mfma` as a valid compiler flag
+# see https://github.com/golang/go/issues/17895
+go env -w "CGO_CFLAGS_ALLOW=-mfma"
+go env -w "CGO_CXXFLAGS_ALLOW=-mfma"
+go build -tags=avx2 .
+```
+
+### CUDA
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive) then build ggml-cuda:
+
+```shell
+# builds ggml-cuda.dll
+./cuda.sh
+```
+
+Then build this package with the `cuda` tag:
+
+```shell
+go build -tags=cuda .
+```
+
+### ROCm
+
+```shell
+# builds ggml-hipblas.dll
+./hipblas.sh
+```
+
+```shell
+go build -tags=rocm .
+```
+
+## Syncing with llama.cpp
+
+```
+./sync.sh ../../llama.cpp
+```

llama/cmd/main.go

@@ -1,11 +0,0 @@
-package main
-
-import (
-	"fmt"
-
-	"github.com/ollama/ollama/llama"
-)
-
-func main() {
-	fmt.Println(llama.SystemInfo())
-}

llama/cuda.sh

@@ -0,0 +1,25 @@
+nvcc \
+    -t 12 \
+    --generate-code=arch=compute_50,code=[compute_50,sm_50] \
+    --generate-code=arch=compute_52,code=[compute_52,sm_52] \
+    --generate-code=arch=compute_61,code=[compute_61,sm_61] \
+    --generate-code=arch=compute_70,code=[compute_70,sm_70] \
+    --generate-code=arch=compute_75,code=[compute_75,sm_75] \
+    --generate-code=arch=compute_80,code=[compute_80,sm_80] \
+    -DGGML_CUDA_DMMV_X=32 \
+    -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 \
+    -DGGML_MULTIPLATFORM \
+    -DGGML_CUDA_MMV_Y=1 \
+    -DGGML_USE_CUDA=1 \
+    -DGGML_SHARED=1 \
+    -DGGML_BUILD=1 \
+    -DGGML_USE_LLAMAFILE \
+    --forward-unknown-to-host-compiler \
+    -use_fast_math \
+    -link \
+    -shared \
+    -I. \
+    -lcuda -lcublas -lcudart -lcublasLt \
+    -O3 \
+    -o ggml-cuda.dll \
+    ggml-cuda.cu ggml-cuda/*.cu ggml.c ggml-backend.c ggml-alloc.c ggml-quants.c

llama/ggml-alloc.c

@@ -1,985 +1,985 @@
-#include "ggml-alloc.h"
-#include "ggml-backend-impl.h"
-#include "ggml.h"
-#include "ggml-impl.h"
-#include <assert.h>
-#include <limits.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-#define MAX_FREE_BLOCKS 256
-
-//#define GGML_ALLOCATOR_DEBUG
-
-//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
-#define AT_PRINTF(...)
-
-
-static bool ggml_is_view(const struct ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
-static bool ggml_op_can_inplace(enum ggml_op op) {
-    switch (op) {
-        case GGML_OP_SCALE:
-        case GGML_OP_DIAG_MASK_ZERO:
-        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_ADD:
-        case GGML_OP_ADD1:
-        case GGML_OP_SUB:
-        case GGML_OP_MUL:
-        case GGML_OP_DIV:
-        case GGML_OP_SQR:
-        case GGML_OP_SQRT:
-        case GGML_OP_LOG:
-        case GGML_OP_UNARY:
-        case GGML_OP_ROPE:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_SOFT_MAX:
-            return true;
-
-        default:
-            return false;
-    }
-}
-
-static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
-    assert(alignment && !(alignment & (alignment - 1))); // power of 2
-    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
-    return offset + align;
-}
-
-// tallocr
-
-struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) {
-    void * base = ggml_backend_buffer_get_base(buffer);
-    size_t align = ggml_backend_buffer_get_alignment(buffer);
-
-    assert(align && !(align & (align - 1))); // power of 2
-
-    struct ggml_tallocr talloc = (struct ggml_tallocr) {
-        /*.buffer    = */ buffer,
-        /*.base      = */ base,
-        /*.alignment = */ align,
-        /*.offset    = */ aligned_offset(base, 0, align),
-    };
-    return talloc;
-}
-
-void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) {
-    size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
-    size = GGML_PAD(size, talloc->alignment);
-
-    if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
-        fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
-                __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
-        GGML_ASSERT(!"not enough space in the buffer");
-        return;
-    }
-
-    void * addr = (char *)ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
-    talloc->offset += size;
-
-    assert(((uintptr_t)addr % talloc->alignment) == 0);
-
-    ggml_backend_tensor_alloc(talloc->buffer, tensor, addr);
-}
-
-// dynamic tensor allocator
-
-struct free_block {
-    size_t offset;
-    size_t size;
-};
-
-struct ggml_dyn_tallocr {
-    size_t alignment;
-    int n_free_blocks;
-    struct free_block free_blocks[MAX_FREE_BLOCKS];
-    size_t max_size;
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    struct {
-        const struct ggml_tensor * tensor;
-        size_t offset;
-    } allocated_tensors[1024];
-#endif
-};
-
-#ifdef GGML_ALLOCATOR_DEBUG
-static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
-    for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i].tensor == NULL) {
-            alloc->allocated_tensors[i].tensor = tensor;
-            alloc->allocated_tensors[i].offset = offset;
-            return;
-        }
-    }
-    GGML_ASSERT(!"out of allocated_tensors");
-}
-static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
-    for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i].offset == offset) {
-            alloc->allocated_tensors[i].tensor = NULL;
-            return;
-        }
-    }
-    fprintf(stderr, "tried to free tensor %s not found\n", tensor->name);
-    GGML_ASSERT(!"tensor not found");
-}
-#endif
-
-static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
-    size = aligned_offset(NULL, size, alloc->alignment);
-
-    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
-
-    size_t max_avail = 0;
-
-    // find the best fitting free block besides the last block
-    int best_fit_block = -1;
-    size_t best_fit_size = SIZE_MAX;
-    for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
-        max_avail = MAX(max_avail, block->size);
-        if (block->size >= size && block->size <= best_fit_size) {
-            best_fit_block = i;
-            best_fit_size = block->size;
-        }
-    }
-
-    if (best_fit_block == -1) {
-        // the last block is our last resort
-        struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
-        max_avail = MAX(max_avail, block->size);
-        if (block->size >= size) {
-            best_fit_block = alloc->n_free_blocks - 1;
-        } else {
-            // this should never happen
-            fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
-                    __func__, size, max_avail);
-            GGML_ASSERT(!"not enough space in the buffer");
-            GGML_UNREACHABLE();
-        }
-    }
-
-    struct free_block * block = &alloc->free_blocks[best_fit_block];
-    size_t offset = block->offset;
-    block->offset = offset + size;
-    block->size -= size;
-    if (block->size == 0) {
-        // remove block if empty
-        alloc->n_free_blocks--;
-        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
-            alloc->free_blocks[j] = alloc->free_blocks[j+1];
-        }
-    }
-
-    AT_PRINTF("block %d, offset %zu\n", best_fit_block, offset);
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    add_allocated_tensor(alloc, offset, tensor);
-    size_t cur_max = offset + size;
-    if (cur_max > alloc->max_size) {
-        // sort allocated_tensors by offset
-        for (int i = 0; i < 1024; i++) {
-            for (int j = i + 1; j < 1024; j++) {
-                if (alloc->allocated_tensors[i].offset > alloc->allocated_tensors[j].offset) {
-                    const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
-                    size_t tmp_offset = alloc->allocated_tensors[i].offset;
-                    alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
-                    alloc->allocated_tensors[i].offset = alloc->allocated_tensors[j].offset;
-                    alloc->allocated_tensors[j].tensor = tmp_tensor;
-                    alloc->allocated_tensors[j].offset = tmp_offset;
-                }
-            }
-        }
-        fprintf(stderr, "max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
-        for (int i = 0; i < 1024; i++) {
-            if (alloc->allocated_tensors[i].tensor) {
-                fprintf(stderr, "%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
-                    alloc->allocated_tensors[i].offset,
-                    alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
-                    ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
-            }
-        }
-        fprintf(stderr, "\n");
-    }
-#endif
-
-    alloc->max_size = MAX(alloc->max_size, offset + size);
-
-    return offset;
-
-    GGML_UNUSED(tensor);
-}
-
-// this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, size_t size, const struct ggml_tensor * tensor) {
-    size = aligned_offset(NULL, size, alloc->alignment);
-
-    AT_PRINTF("%s: freeing %s at %zu (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, offset, size, alloc->n_free_blocks);
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    remove_allocated_tensor(alloc, offset, tensor);
-#endif
-
-    // see if we can merge with an existing block
-    for (int i = 0; i < alloc->n_free_blocks; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
-        // check if ptr is at the end of the block
-        if (block->offset + block->size == offset) {
-            block->size += size;
-            // check if we can merge with the next block
-            if (i < alloc->n_free_blocks - 1 && block->offset + block->size == alloc->free_blocks[i+1].offset) {
-                block->size += alloc->free_blocks[i+1].size;
-                alloc->n_free_blocks--;
-                for (int j = i+1; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
-                }
-            }
-            return;
-        }
-        // check if ptr is at the beginning of the block
-        if (offset + size == block->offset) {
-            block->offset = offset;
-            block->size += size;
-            // check if we can merge with the previous block
-            if (i > 0 && alloc->free_blocks[i-1].offset + alloc->free_blocks[i-1].size == block->offset) {
-                alloc->free_blocks[i-1].size += block->size;
-                alloc->n_free_blocks--;
-                for (int j = i; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
-                }
-            }
-            return;
-        }
-    }
-    // otherwise, add a new block
-    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
-    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
-    int insert_pos = 0;
-    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].offset < offset) {
-        insert_pos++;
-    }
-    // shift all blocks from insert_pos onward to make room for the new block
-    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
-        alloc->free_blocks[i] = alloc->free_blocks[i-1];
-    }
-    // insert the new block
-    alloc->free_blocks[insert_pos].offset = offset;
-    alloc->free_blocks[insert_pos].size = size;
-    alloc->n_free_blocks++;
-
-    GGML_UNUSED(tensor);
-}
-
-static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
-    alloc->n_free_blocks = 1;
-    alloc->free_blocks[0].offset = 0;
-    alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
-    alloc->max_size = 0;
-}
-
-static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
-    struct ggml_dyn_tallocr * alloc = (struct ggml_dyn_tallocr *)malloc(sizeof(struct ggml_dyn_tallocr));
-
-    *alloc = (struct ggml_dyn_tallocr) {
-        /*.alignment     = */ alignment,
-        /*.n_free_blocks = */ 0,
-        /*.free_blocks   = */ {{0}},
-        /*.max_size      = */ 0,
-#ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ {{0}},
-#endif
-    };
-
-    ggml_dyn_tallocr_reset(alloc);
-
-    return alloc;
-}
-
-static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
-    free(alloc);
-}
-
-static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
-    return alloc->max_size;
-}
-
-
-/////////////////////////////////////
-
-// graph allocator
-
-struct hash_node {
-    int n_children;
-    int n_views;
-    int buffer_id;
-    size_t offset; // offset within the buffer
-    bool allocated;
-};
-
-struct tensor_alloc {
-    size_t offset;
-    size_t size_max; // 0 = pre-allocated, unused, or view
-};
-
-struct leaf_alloc {
-    int buffer_id;
-    struct tensor_alloc leaf;
-};
-
-struct node_alloc {
-    int buffer_id;
-    struct tensor_alloc dst;
-    struct tensor_alloc src[GGML_MAX_SRC];
-};
-
-struct ggml_gallocr {
-    ggml_backend_buffer_type_t * bufts; // [n_buffers]
-    ggml_backend_buffer_t * buffers; // [n_buffers]
-    struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
-    int n_buffers;
-
-    struct ggml_hash_set hash_set;
-    struct hash_node * hash_values; // [hash_set.size]
-
-    struct node_alloc * node_allocs; // [n_nodes]
-    int n_nodes;
-
-    struct leaf_alloc * leaf_allocs; // [n_leafs]
-    int n_leafs;
-};
-
-ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
-    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(sizeof(struct ggml_gallocr), 1);
-    GGML_ASSERT(galloc != NULL);
-
-    galloc->bufts = calloc(sizeof(ggml_backend_buffer_type_t) * n_bufs, 1);
-    GGML_ASSERT(galloc->bufts != NULL);
-
-    galloc->buffers = calloc(sizeof(ggml_backend_buffer_t) * n_bufs, 1);
-    GGML_ASSERT(galloc->buffers != NULL);
-
-    galloc->buf_tallocs = calloc(sizeof(struct ggml_dyn_tallocr *) * n_bufs, 1);
-    GGML_ASSERT(galloc->buf_tallocs != NULL);
-
-    for (int i = 0; i < n_bufs; i++) {
-        galloc->bufts[i] = bufts[i];
-        galloc->buffers[i] = NULL;
-        size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
-        galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
-    }
-    galloc->n_buffers = n_bufs;
-
-    return galloc;
-}
-
-ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft) {
-    return ggml_gallocr_new_n(&buft, 1);
-}
-
-void ggml_gallocr_free(ggml_gallocr_t galloc) {
-    if (galloc == NULL) {
-        return;
-    }
-
-    for (int i = 0; i < galloc->n_buffers; i++) {
-        if (galloc->buffers != NULL) {
-            ggml_backend_buffer_free(galloc->buffers[i]);
-        }
-        if (galloc->buf_tallocs != NULL) {
-            ggml_dyn_tallocr_free(galloc->buf_tallocs[i]);
-        }
-    }
-
-    free(galloc->hash_set.keys);
-    free(galloc->hash_values);
-    free(galloc->bufts);
-    free(galloc->buffers);
-    free(galloc->buf_tallocs);
-    free(galloc->node_allocs);
-    free(galloc->leaf_allocs);
-    free(galloc);
-}
-
-typedef struct ggml_gallocr * ggml_gallocr_t;
-
-static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    size_t i = ggml_hash_find_or_insert(galloc->hash_set, t);
-    return &galloc->hash_values[i];
-}
-
-static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    return ggml_gallocr_hash_get(galloc, t)->allocated;
-}
-
-static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    hn->buffer_id = buffer_id;
-    hn->offset = offset;
-    hn->allocated = true;
-}
-
-static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
-}
-
-static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-
-    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
-        hn->allocated = true;
-        assert(hn->offset == 0);
-
-        // try to reuse a parent's buffer (inplace)
-        if (ggml_op_can_inplace(node->op)) {
-            for (int i = 0; i < GGML_MAX_SRC; i++) {
-                struct ggml_tensor * parent = node->src[i];
-                if (parent == NULL) {
-                    continue;
-                }
-
-                // if the node's data is external, then we cannot re-use it
-                if (!ggml_gallocr_is_own(galloc, parent)) {
-                    AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
-                    continue;
-                }
-
-                // outputs cannot be reused
-                if (parent->flags & GGML_TENSOR_FLAG_OUTPUT || (parent->view_src != NULL && parent->view_src->flags & GGML_TENSOR_FLAG_OUTPUT)) {
-                    AT_PRINTF("not reusing parent %s for %s as it is an output\n", parent->name, node->name);
-                    continue;
-                }
-
-                if (!ggml_are_same_layout(node, parent)) {
-                    AT_PRINTF("not reusing parent %s for %s as layouts are different\n", parent->name, node->name);
-                    continue;
-                }
-
-                struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
-                if (p_hn->n_children == 1 && p_hn->n_views == 0) {
-                    if (ggml_is_view(parent)) {
-                        struct ggml_tensor * view_src = parent->view_src;
-                        struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
-                        if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
-                            AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
-                            assert(view_src_hn->offset == p_hn->offset);
-                            hn->buffer_id = p_hn->buffer_id;
-                            hn->offset = p_hn->offset;
-                            p_hn->allocated = false; // avoid freeing the parent
-                            view_src_hn->allocated = false;
-                            return;
-                        }
-                    } else {
-                        AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
-                        hn->buffer_id = p_hn->buffer_id;
-                        hn->offset = p_hn->offset;
-                        p_hn->allocated = false; // avoid freeing the parent
-                        return;
-                    }
-                }
-            }
-        }
-        // allocate tensor from the buffer
-        struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
-        ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
-        size_t size = ggml_backend_buft_get_alloc_size(buft, node);
-        size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
-        hn->buffer_id = buffer_id;
-        hn->offset = offset;
-        return;
-    }
-}
-
-static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
-    // graph outputs are never freed
-    if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
-        AT_PRINTF("not freeing output %s\n", node->name);
-        return;
-    }
-
-    struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
-    ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    size_t offset = hn->offset;
-    size_t size = ggml_backend_buft_get_alloc_size(buft, node);
-    ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
-    hn->allocated = false;
-}
-
-static int get_node_buffer_id(const int * node_buffer_ids, int i) {
-    return node_buffer_ids ? node_buffer_ids[i] : 0;
-}
-
-static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
-    // clear hash tables
-    memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
-    memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
-
-    // allocate leafs
-    // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
-    }
-
-    // count number of children and views
-    // allocate other graph inputs and leafs first to avoid overwriting them
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-
-        // TODO: better way to add external dependencies
-        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
-        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
-        // itself is never used and should not be considered a dependency
-        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
-            struct ggml_tensor * view_src = node->view_src;
-            ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
-        }
-
-        if (node->flags & GGML_TENSOR_FLAG_INPUT) {
-            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
-        }
-
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
-            }
-
-            ggml_gallocr_hash_get(galloc, src)->n_children += 1;
-
-            // allocate explicit inputs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
-                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
-            }
-        }
-    }
-
-    // allocate tensors
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        int buffer_id = get_node_buffer_id(node_buffer_ids, i);
-
-        // allocate parents (only leafs need to be allocated at this point)
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * parent = node->src[j];
-            if (parent == NULL) {
-                continue;
-            }
-            ggml_gallocr_allocate_node(galloc, parent, buffer_id);
-        }
-
-        // allocate node
-        ggml_gallocr_allocate_node(galloc, node, buffer_id);
-
-        AT_PRINTF("exec: %s (%s) <= ", ggml_op_desc(node), node->name);
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * parent = node->src[j];
-            if (parent == NULL) {
-                continue;
-            }
-            AT_PRINTF("%s", parent->name);
-            if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
-                AT_PRINTF(", ");
-            }
-        }
-        AT_PRINTF("\n");
-
-        // update parents
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * parent = node->src[j];
-            if (parent == NULL) {
-                continue;
-            }
-            struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
-            p_hn->n_children -= 1;
-
-            AT_PRINTF("parent %s: %d children, %d views, allocated: %d\n",
-                parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);
-
-            if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                if (ggml_is_view(parent)) {
-                    struct ggml_tensor * view_src = parent->view_src;
-                    struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
-                    view_src_hn->n_views -= 1;
-                    AT_PRINTF("view_src %s: %d children, %d views\n",
-                        view_src->name, view_src_hn->n_children, view_src_hn->n_views);
-                    if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src_hn->allocated) {
-                        ggml_gallocr_free_node(galloc, view_src, buffer_id);
-                    }
-                }
-                else if (p_hn->allocated) {
-                    ggml_gallocr_free_node(galloc, parent, buffer_id);
-                }
-            }
-            AT_PRINTF("\n");
-        }
-    }
-}
-
-bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
-    size_t hash_size = graph->visited_hash_table.size;
-
-    // initialize hash table
-    if (galloc->hash_set.size < hash_size) {
-        free(galloc->hash_set.keys);
-        free(galloc->hash_values);
-        galloc->hash_set.size = hash_size;
-        galloc->hash_set.keys = calloc(sizeof(struct ggml_tensor *), hash_size);
-        galloc->hash_values   = calloc(sizeof(struct hash_node), hash_size);
-        GGML_ASSERT(galloc->hash_set.keys != NULL);
-        GGML_ASSERT(galloc->hash_values != NULL);
-    } else {
-        // reset hash table
-        memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * galloc->hash_set.size);
-        memset(galloc->hash_values,   0, sizeof(struct hash_node) * galloc->hash_set.size);
-    }
-
-    // reset allocators
-    for (int i = 0; i < galloc->n_buffers; i++) {
-        ggml_dyn_tallocr_reset(galloc->buf_tallocs[i]);
-    }
-
-    // allocate in hash table
-    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);
-
-    // set the node_allocs from the hash table
-    if (galloc->n_nodes < graph->n_nodes) {
-        free(galloc->node_allocs);
-        galloc->node_allocs = calloc(sizeof(struct node_alloc), graph->n_nodes);
-        GGML_ASSERT(galloc->node_allocs != NULL);
-    }
-    galloc->n_nodes = graph->n_nodes;
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        struct node_alloc * node_alloc = &galloc->node_allocs[i];
-        node_alloc->buffer_id = get_node_buffer_id(node_buffer_ids, i);
-        if (node->view_src || node->data) {
-            node_alloc->dst.offset = SIZE_MAX;
-            node_alloc->dst.size_max = 0;
-        } else {
-            struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-            node_alloc->dst.offset   = hn->offset;
-            node_alloc->dst.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
-        }
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (!src || src->view_src || src->data) {
-                node_alloc->src[j].offset = SIZE_MAX;
-                node_alloc->src[j].size_max = 0;
-            } else {
-                struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
-                node_alloc->src[j].offset   = hn->offset;
-                node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
-            }
-        }
-    }
-    if (galloc->n_leafs < graph->n_leafs) {
-        free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
-        GGML_ASSERT(galloc->leaf_allocs != NULL);
-    }
-    galloc->n_leafs = graph->n_leafs;
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
-        if (leaf->view_src || leaf->data) {
-            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
-            galloc->leaf_allocs[i].leaf.size_max = 0;
-        } else {
-            galloc->leaf_allocs[i].leaf.offset = hn->offset;
-            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
-        }
-    }
-
-    // reallocate buffers if needed
-    for (int i = 0; i < galloc->n_buffers; i++) {
-        size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
-        size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
-
-        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
-        if (new_size > cur_size || galloc->buffers[i] == NULL) {
-#ifndef NDEBUG
-            fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
-#endif
-            ggml_backend_buffer_free(galloc->buffers[i]);
-            galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
-            if (galloc->buffers[i] == NULL) {
-                fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
-                return false;
-            }
-        }
-    }
-
-    return true;
-}
-
-bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-    return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
-}
-
-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, int buffer_id, struct tensor_alloc * tensor_alloc) {
-    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
-
-    if (tensor->view_src != NULL) {
-        if (tensor->buffer == NULL) {
-            assert(tensor_alloc->offset == SIZE_MAX);
-            if (tensor->view_src->buffer == NULL) {
-                // this tensor was allocated without ggml-backend
-                return;
-            }
-            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
-        }
-    } else {
-        if (tensor->data == NULL) {
-            assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
-            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
-            void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
-        } else {
-            if (tensor->buffer == NULL) {
-                // this tensor was allocated without ggml-backend
-                return;
-            }
-        }
-    }
-}
-
-static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct node_alloc * nalloc, struct tensor_alloc * talloc) {
-    ggml_backend_buffer_type_t buft = galloc->bufts[nalloc->buffer_id];
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(buft, node);
-    return talloc->size_max >= node_size;
-}
-
-static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
-    if (galloc->n_nodes != graph->n_nodes) {
-#ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of nodes\n", __func__);
-#endif
-        return true;
-    }
-
-    if (galloc->n_leafs != graph->n_leafs) {
-#ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
-#endif
-        return true;
-    }
-
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        struct node_alloc * node_alloc = &galloc->node_allocs[i];
-
-        if (!ggml_gallocr_node_needs_realloc(galloc, node, node_alloc, &node_alloc->dst)) {
-#ifndef NDEBUG
-            fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
-#endif
-            return true;
-        }
-
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
-            }
-            if (!ggml_gallocr_node_needs_realloc(galloc, src, node_alloc, &node_alloc->src[j])) {
-#ifndef NDEBUG
-                fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
-#endif
-                return true;
-            }
-        }
-    }
-
-    return false;
-}
-
-bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
-    if (ggml_gallocr_needs_realloc(galloc, graph)) {
-        if (galloc->n_buffers == 1) {
-#ifndef NDEBUG
-            fprintf(stderr, "%s: reallocating buffers automatically\n", __func__);
-#endif
-            if (!ggml_gallocr_reserve(galloc, graph)) {
-                return false;
-            }
-        } else {
-#ifndef NDEBUG
-            fprintf(stderr, "%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
-#endif
-            return false;
-        }
-    }
-
-    // reset buffers
-    for (int i = 0; i < galloc->n_buffers; i++) {
-        if (galloc->buffers[i] != NULL) {
-            ggml_backend_buffer_reset(galloc->buffers[i]);
-        }
-    }
-
-    // allocate the graph tensors from the previous assignments
-    // leafs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, leaf_alloc->buffer_id, &leaf_alloc->leaf);
-    }
-    // nodes
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        struct node_alloc * node_alloc = &galloc->node_allocs[i];
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
-            }
-            ggml_gallocr_init_tensor(galloc, src, node_alloc->buffer_id, &node_alloc->src[j]);
-        }
-        ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
-    }
-
-    return true;
-}
-
-size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
-    GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
-
-    if (galloc->buffers[buffer_id] == NULL) {
-        return 0;
-    }
-    return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
-}
-
-// utils
-
-static bool alloc_tensor_range(struct ggml_context * ctx,
-        struct ggml_tensor * first, struct ggml_tensor * last,
-        ggml_backend_buffer_type_t buft, size_t size,
-        ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
-    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
-    if (buffer == NULL) {
-#ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
-#endif
-        for (size_t i = 0; i < *n_buffers; i++) {
-            ggml_backend_buffer_free(*buffers[i]);
-        }
-        free(*buffers);
-        return false;
-    }
-
-    struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);
-
-    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
-        if (t->data == NULL) {
-            if (t->view_src == NULL) {
-                ggml_tallocr_alloc(&tallocr, t);
-            } else if (t->buffer == NULL) {
-                ggml_backend_view_init(buffer, t);
-            }
-        } else {
-            if (t->view_src != NULL && t->buffer == NULL) {
-                // view of a pre-allocated tensor
-                ggml_backend_view_init(buffer, t);
-            }
-        }
-    }
-
-    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
-    (*buffers)[(*n_buffers)++] = buffer;
-
-    return true;
-}
-
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
-    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
-
-    size_t alignment = ggml_backend_buft_get_alignment(buft);
-    size_t max_size = ggml_backend_buft_get_max_size(buft);
-
-    ggml_backend_buffer_t * buffers = NULL;
-    size_t n_buffers = 0;
-
-    size_t cur_buf_size = 0;
-    struct ggml_tensor * first = ggml_get_first_tensor(ctx);
-    for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size_t this_size = 0;
-        if (t->data == NULL && t->view_src == NULL) {
-            this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
-        }
-
-        if (this_size > max_size) {
-            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
-                    __func__, t->name,
-                    ggml_backend_buft_name(buft),
-                    this_size, max_size);
-            for (size_t i = 0; i < n_buffers; i++) {
-                ggml_backend_buffer_free(buffers[i]);
-            }
-            free(buffers);
-            return NULL;
-        }
-
-        if ((cur_buf_size + this_size) > max_size) {
-            // allocate tensors in the current buffer
-            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
-                return NULL;
-            }
-            first = t;
-            cur_buf_size = this_size;
-        } else {
-            cur_buf_size += this_size;
-        }
-    }
-
-    // allocate remaining tensors
-    if (cur_buf_size > 0) {
-        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
-            return NULL;
-        }
-    }
-
-    if (n_buffers == 0) {
-#ifndef NDEBUG
-        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
-#endif
-        return NULL;
-    }
-
-    ggml_backend_buffer_t buffer;
-    if (n_buffers == 1) {
-        buffer = buffers[0];
-    } else {
-        buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
-    }
-    free(buffers);
-    return buffer;
-}
-
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
-    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
-}
+#include "ggml-alloc.h"
+#include "ggml-backend-impl.h"
+#include "ggml.h"
+#include "ggml-impl.h"
+#include <assert.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define MAX_FREE_BLOCKS 256
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
+#define AT_PRINTF(...)
+
+
+static bool ggml_is_view(const struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SOFT_MAX:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+// tallocr
+
+struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) {
+    void * base = ggml_backend_buffer_get_base(buffer);
+    size_t align = ggml_backend_buffer_get_alignment(buffer);
+
+    assert(align && !(align & (align - 1))); // power of 2
+
+    struct ggml_tallocr talloc = (struct ggml_tallocr) {
+        /*.buffer    = */ buffer,
+        /*.base      = */ base,
+        /*.alignment = */ align,
+        /*.offset    = */ aligned_offset(base, 0, align),
+    };
+    return talloc;
+}
+
+void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
+    size = GGML_PAD(size, talloc->alignment);
+
+    if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
+        fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
+                __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
+        GGML_ASSERT(!"not enough space in the buffer");
+        return;
+    }
+
+    void * addr = (char *)ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
+    talloc->offset += size;
+
+    assert(((uintptr_t)addr % talloc->alignment) == 0);
+
+    ggml_backend_tensor_alloc(talloc->buffer, tensor, addr);
+}
+
+// dynamic tensor allocator
+
+struct free_block {
+    size_t offset;
+    size_t size;
+};
+
+struct ggml_dyn_tallocr {
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    size_t max_size;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct {
+        const struct ggml_tensor * tensor;
+        size_t offset;
+    } allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i].tensor == NULL) {
+            alloc->allocated_tensors[i].tensor = tensor;
+            alloc->allocated_tensors[i].offset = offset;
+            return;
+        }
+    }
+    GGML_ASSERT(!"out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i].offset == offset) {
+            alloc->allocated_tensors[i].tensor = NULL;
+            return;
+        }
+    }
+    fprintf(stderr, "tried to free tensor %s not found\n", tensor->name);
+    GGML_ASSERT(!"tensor not found");
+}
+#endif
+
+static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block besides the last block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    if (best_fit_block == -1) {
+        // the last block is our last resort
+        struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size) {
+            best_fit_block = alloc->n_free_blocks - 1;
+        } else {
+            // this should never happen
+            fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
+                    __func__, size, max_avail);
+            GGML_ASSERT(!"not enough space in the buffer");
+            GGML_UNREACHABLE();
+        }
+    }
+
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    size_t offset = block->offset;
+    block->offset = offset + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    AT_PRINTF("block %d, offset %zu\n", best_fit_block, offset);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, offset, tensor);
+    size_t cur_max = offset + size;
+    if (cur_max > alloc->max_size) {
+        // sort allocated_tensors by offset
+        for (int i = 0; i < 1024; i++) {
+            for (int j = i + 1; j < 1024; j++) {
+                if (alloc->allocated_tensors[i].offset > alloc->allocated_tensors[j].offset) {
+                    const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
+                    size_t tmp_offset = alloc->allocated_tensors[i].offset;
+                    alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
+                    alloc->allocated_tensors[i].offset = alloc->allocated_tensors[j].offset;
+                    alloc->allocated_tensors[j].tensor = tmp_tensor;
+                    alloc->allocated_tensors[j].offset = tmp_offset;
+                }
+            }
+        }
+        fprintf(stderr, "max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i].tensor) {
+                fprintf(stderr, "%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
+                    alloc->allocated_tensors[i].offset,
+                    alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
+                    ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
+            }
+        }
+        fprintf(stderr, "\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, offset + size);
+
+    return offset;
+
+    GGML_UNUSED(tensor);
+}
+
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, size_t size, const struct ggml_tensor * tensor) {
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: freeing %s at %zu (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, offset, size, alloc->n_free_blocks);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, offset, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if (block->offset + block->size == offset) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && block->offset + block->size == alloc->free_blocks[i+1].offset) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if (offset + size == block->offset) {
+            block->offset = offset;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && alloc->free_blocks[i-1].offset + alloc->free_blocks[i-1].size == block->offset) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].offset < offset) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].offset = offset;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+
+    GGML_UNUSED(tensor);
+}
+
+static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
+    alloc->n_free_blocks = 1;
+    alloc->free_blocks[0].offset = 0;
+    alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
+    alloc->max_size = 0;
+}
+
+static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
+    struct ggml_dyn_tallocr * alloc = (struct ggml_dyn_tallocr *)malloc(sizeof(struct ggml_dyn_tallocr));
+
+    *alloc = (struct ggml_dyn_tallocr) {
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.max_size      = */ 0,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ {{0}},
+#endif
+    };
+
+    ggml_dyn_tallocr_reset(alloc);
+
+    return alloc;
+}
+
+static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
+    free(alloc);
+}
+
+static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
+    return alloc->max_size;
+}
+
+
+/////////////////////////////////////
+
+// graph allocator
+
+struct hash_node {
+    int n_children;
+    int n_views;
+    int buffer_id;
+    size_t offset; // offset within the buffer
+    bool allocated;
+};
+
+struct tensor_alloc {
+    size_t offset;
+    size_t size_max; // 0 = pre-allocated, unused, or view
+};
+
+struct leaf_alloc {
+    int buffer_id;
+    struct tensor_alloc leaf;
+};
+
+struct node_alloc {
+    int buffer_id;
+    struct tensor_alloc dst;
+    struct tensor_alloc src[GGML_MAX_SRC];
+};
+
+struct ggml_gallocr {
+    ggml_backend_buffer_type_t * bufts; // [n_buffers]
+    ggml_backend_buffer_t * buffers; // [n_buffers]
+    struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
+    int n_buffers;
+
+    struct ggml_hash_set hash_set;
+    struct hash_node * hash_values; // [hash_set.size]
+
+    struct node_alloc * node_allocs; // [n_nodes]
+    int n_nodes;
+
+    struct leaf_alloc * leaf_allocs; // [n_leafs]
+    int n_leafs;
+};
+
+ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
+    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
+    GGML_ASSERT(galloc != NULL);
+
+    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
+    GGML_ASSERT(galloc->bufts != NULL);
+
+    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
+    GGML_ASSERT(galloc->buffers != NULL);
+
+    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
+    GGML_ASSERT(galloc->buf_tallocs != NULL);
+
+    for (int i = 0; i < n_bufs; i++) {
+        galloc->bufts[i] = bufts[i];
+        galloc->buffers[i] = NULL;
+        size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
+        galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
+    }
+    galloc->n_buffers = n_bufs;
+
+    return galloc;
+}
+
+ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft) {
+    return ggml_gallocr_new_n(&buft, 1);
+}
+
+void ggml_gallocr_free(ggml_gallocr_t galloc) {
+    if (galloc == NULL) {
+        return;
+    }
+
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers != NULL) {
+            ggml_backend_buffer_free(galloc->buffers[i]);
+        }
+        if (galloc->buf_tallocs != NULL) {
+            ggml_dyn_tallocr_free(galloc->buf_tallocs[i]);
+        }
+    }
+
+    free(galloc->hash_set.keys);
+    free(galloc->hash_values);
+    free(galloc->bufts);
+    free(galloc->buffers);
+    free(galloc->buf_tallocs);
+    free(galloc->node_allocs);
+    free(galloc->leaf_allocs);
+    free(galloc);
+}
+
+typedef struct ggml_gallocr * ggml_gallocr_t;
+
+static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    size_t i = ggml_hash_find_or_insert(galloc->hash_set, t);
+    return &galloc->hash_values[i];
+}
+
+static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    return ggml_gallocr_hash_get(galloc, t)->allocated;
+}
+
+static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    hn->buffer_id = buffer_id;
+    hn->offset = offset;
+    hn->allocated = true;
+}
+
+static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
+}
+
+static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+
+    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
+        hn->allocated = true;
+        assert(hn->offset == 0);
+
+        // try to reuse a parent's buffer (inplace)
+        if (ggml_op_can_inplace(node->op)) {
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                struct ggml_tensor * parent = node->src[i];
+                if (parent == NULL) {
+                    continue;
+                }
+
+                // if the node's data is external, then we cannot re-use it
+                if (!ggml_gallocr_is_own(galloc, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                    continue;
+                }
+
+                // outputs cannot be reused
+                if (parent->flags & GGML_TENSOR_FLAG_OUTPUT || (parent->view_src != NULL && parent->view_src->flags & GGML_TENSOR_FLAG_OUTPUT)) {
+                    AT_PRINTF("not reusing parent %s for %s as it is an output\n", parent->name, node->name);
+                    continue;
+                }
+
+                if (!ggml_are_same_layout(node, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as layouts are different\n", parent->name, node->name);
+                    continue;
+                }
+
+                struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+                if (p_hn->n_children == 1 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = parent->view_src;
+                        struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
+                        if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                            AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                            assert(view_src_hn->offset == p_hn->offset);
+                            hn->buffer_id = p_hn->buffer_id;
+                            hn->offset = p_hn->offset;
+                            p_hn->allocated = false; // avoid freeing the parent
+                            view_src_hn->allocated = false;
+                            return;
+                        }
+                    } else {
+                        AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                        hn->buffer_id = p_hn->buffer_id;
+                        hn->offset = p_hn->offset;
+                        p_hn->allocated = false; // avoid freeing the parent
+                        return;
+                    }
+                }
+            }
+        }
+        // allocate tensor from the buffer
+        struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+        ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+        size_t size = ggml_backend_buft_get_alloc_size(buft, node);
+        size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
+        hn->buffer_id = buffer_id;
+        hn->offset = offset;
+        return;
+    }
+}
+
+static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    // graph outputs are never freed
+    if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
+        AT_PRINTF("not freeing output %s\n", node->name);
+        return;
+    }
+
+    struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+    ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    size_t offset = hn->offset;
+    size_t size = ggml_backend_buft_get_alloc_size(buft, node);
+    ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
+    hn->allocated = false;
+}
+
+static int get_node_buffer_id(const int * node_buffer_ids, int i) {
+    return node_buffer_ids ? node_buffer_ids[i] : 0;
+}
+
+static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    // clear hash tables
+    memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
+    memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
+
+    // allocate leafs
+    // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
+    }
+
+    // count number of children and views
+    // allocate other graph inputs and leafs first to avoid overwriting them
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
+            struct ggml_tensor * view_src = node->view_src;
+            ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
+        }
+
+        if (node->flags & GGML_TENSOR_FLAG_INPUT) {
+            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
+        }
+
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+
+            ggml_gallocr_hash_get(galloc, src)->n_children += 1;
+
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
+                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int buffer_id = get_node_buffer_id(node_buffer_ids, i);
+
+        // allocate parents (only leafs need to be allocated at this point)
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            ggml_gallocr_allocate_node(galloc, parent, buffer_id);
+        }
+
+        // allocate node
+        ggml_gallocr_allocate_node(galloc, node, buffer_id);
+
+        AT_PRINTF("exec: %s (%s) <= ", ggml_op_desc(node), node->name);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            AT_PRINTF("%s", parent->name);
+            if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                AT_PRINTF(", ");
+            }
+        }
+        AT_PRINTF("\n");
+
+        // update parents
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+            p_hn->n_children -= 1;
+
+            AT_PRINTF("parent %s: %d children, %d views, allocated: %d\n",
+                parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);
+
+            if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                if (ggml_is_view(parent)) {
+                    struct ggml_tensor * view_src = parent->view_src;
+                    struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
+                    view_src_hn->n_views -= 1;
+                    AT_PRINTF("view_src %s: %d children, %d views\n",
+                        view_src->name, view_src_hn->n_children, view_src_hn->n_views);
+                    if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src_hn->allocated) {
+                        ggml_gallocr_free_node(galloc, view_src, buffer_id);
+                    }
+                }
+                else if (p_hn->allocated) {
+                    ggml_gallocr_free_node(galloc, parent, buffer_id);
+                }
+            }
+            AT_PRINTF("\n");
+        }
+    }
+}
+
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    size_t hash_size = graph->visited_hash_table.size;
+
+    // initialize hash table
+    if (galloc->hash_set.size < hash_size) {
+        free(galloc->hash_set.keys);
+        free(galloc->hash_values);
+        galloc->hash_set.size = hash_size;
+        galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
+        galloc->hash_values   = calloc(hash_size, sizeof(struct hash_node));
+        GGML_ASSERT(galloc->hash_set.keys != NULL);
+        GGML_ASSERT(galloc->hash_values != NULL);
+    } else {
+        // reset hash table
+        memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * galloc->hash_set.size);
+        memset(galloc->hash_values,   0, sizeof(struct hash_node) * galloc->hash_set.size);
+    }
+
+    // reset allocators
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        ggml_dyn_tallocr_reset(galloc->buf_tallocs[i]);
+    }
+
+    // allocate in hash table
+    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);
+
+    // set the node_allocs from the hash table
+    if (galloc->n_nodes < graph->n_nodes) {
+        free(galloc->node_allocs);
+        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
+        GGML_ASSERT(galloc->node_allocs != NULL);
+    }
+    galloc->n_nodes = graph->n_nodes;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        node_alloc->buffer_id = get_node_buffer_id(node_buffer_ids, i);
+        if (node->view_src || node->data) {
+            node_alloc->dst.offset = SIZE_MAX;
+            node_alloc->dst.size_max = 0;
+        } else {
+            struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+            node_alloc->dst.offset   = hn->offset;
+            node_alloc->dst.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (!src || src->view_src || src->data) {
+                node_alloc->src[j].offset = SIZE_MAX;
+                node_alloc->src[j].size_max = 0;
+            } else {
+                struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
+                node_alloc->src[j].offset   = hn->offset;
+                node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
+            }
+        }
+    }
+    if (galloc->n_leafs < graph->n_leafs) {
+        free(galloc->leaf_allocs);
+        galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
+        GGML_ASSERT(galloc->leaf_allocs != NULL);
+    }
+    galloc->n_leafs = graph->n_leafs;
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
+        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
+        if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.size_max = 0;
+        } else {
+            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        }
+    }
+
+    // reallocate buffers if needed
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
+        size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
+
+        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
+        if (new_size > cur_size || galloc->buffers[i] == NULL) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+#endif
+            ggml_backend_buffer_free(galloc->buffers[i]);
+            galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
+            if (galloc->buffers[i] == NULL) {
+                fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                return false;
+            }
+        }
+    }
+
+    return true;
+}
+
+bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
+    return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
+}
+
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, int buffer_id, struct tensor_alloc * tensor_alloc) {
+    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+
+    if (tensor->view_src != NULL) {
+        if (tensor->buffer == NULL) {
+            assert(tensor_alloc->offset == SIZE_MAX);
+            if (tensor->view_src->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
+        }
+    } else {
+        if (tensor->data == NULL) {
+            assert(tensor_alloc->offset != SIZE_MAX);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
+            void * addr = (char *)base + tensor_alloc->offset;
+            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
+        } else {
+            if (tensor->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+        }
+    }
+}
+
+static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct node_alloc * nalloc, struct tensor_alloc * talloc) {
+    ggml_backend_buffer_type_t buft = galloc->bufts[nalloc->buffer_id];
+    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(buft, node);
+    return talloc->size_max >= node_size;
+}
+
+static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
+    if (galloc->n_nodes != graph->n_nodes) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: graph has different number of nodes\n", __func__);
+#endif
+        return true;
+    }
+
+    if (galloc->n_leafs != graph->n_leafs) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
+#endif
+        return true;
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+
+        if (!ggml_gallocr_node_needs_realloc(galloc, node, node_alloc, &node_alloc->dst)) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
+#endif
+            return true;
+        }
+
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            if (!ggml_gallocr_node_needs_realloc(galloc, src, node_alloc, &node_alloc->src[j])) {
+#ifndef NDEBUG
+                fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
+#endif
+                return true;
+            }
+        }
+    }
+
+    return false;
+}
+
+bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
+    if (ggml_gallocr_needs_realloc(galloc, graph)) {
+        if (galloc->n_buffers == 1) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: reallocating buffers automatically\n", __func__);
+#endif
+            if (!ggml_gallocr_reserve(galloc, graph)) {
+                return false;
+            }
+        } else {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
+#endif
+            return false;
+        }
+    }
+
+    // reset buffers
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers[i] != NULL) {
+            ggml_backend_buffer_reset(galloc->buffers[i]);
+        }
+    }
+
+    // allocate the graph tensors from the previous assignments
+    // leafs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
+        ggml_gallocr_init_tensor(galloc, leaf, leaf_alloc->buffer_id, &leaf_alloc->leaf);
+    }
+    // nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            ggml_gallocr_init_tensor(galloc, src, node_alloc->buffer_id, &node_alloc->src[j]);
+        }
+        ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
+    }
+
+    return true;
+}
+
+size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
+
+    if (galloc->buffers[buffer_id] == NULL) {
+        return 0;
+    }
+    return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
+}
+
+// utils
+
+static bool alloc_tensor_range(struct ggml_context * ctx,
+        struct ggml_tensor * first, struct ggml_tensor * last,
+        ggml_backend_buffer_type_t buft, size_t size,
+        ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
+    if (buffer == NULL) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
+#endif
+        for (size_t i = 0; i < *n_buffers; i++) {
+            ggml_backend_buffer_free(*buffers[i]);
+        }
+        free(*buffers);
+        return false;
+    }
+
+    struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);
+
+    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                ggml_tallocr_alloc(&tallocr, t);
+            } else if (t->buffer == NULL) {
+                ggml_backend_view_init(buffer, t);
+            }
+        } else {
+            if (t->view_src != NULL && t->buffer == NULL) {
+                // view of a pre-allocated tensor
+                ggml_backend_view_init(buffer, t);
+            }
+        }
+    }
+
+    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
+    (*buffers)[(*n_buffers)++] = buffer;
+
+    return true;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    size_t max_size = ggml_backend_buft_get_max_size(buft);
+
+    ggml_backend_buffer_t * buffers = NULL;
+    size_t n_buffers = 0;
+
+    size_t cur_buf_size = 0;
+    struct ggml_tensor * first = ggml_get_first_tensor(ctx);
+    for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        size_t this_size = 0;
+        if (t->data == NULL && t->view_src == NULL) {
+            this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+
+        if (this_size > max_size) {
+            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
+            // allocate tensors in the current buffer
+            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+                return NULL;
+            }
+            first = t;
+            cur_buf_size = this_size;
+        } else {
+            cur_buf_size += this_size;
+        }
+    }
+
+    // allocate remaining tensors
+    if (cur_buf_size > 0) {
+        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+            return NULL;
+        }
+    }
+
+    if (n_buffers == 0) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer;
+    if (n_buffers == 1) {
+        buffer = buffers[0];
+    } else {
+        buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
+    }
+    free(buffers);
+    return buffer;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
+    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
+}

llama/ggml-alloc.h

@@ -1,76 +1,76 @@
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-typedef struct ggml_backend * ggml_backend_t;
-
-// Tensor allocator
-struct ggml_tallocr {
-    ggml_backend_buffer_t buffer;
-    void * base;
-    size_t alignment;
-    size_t offset;
-};
-
-GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
-GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
-
-// Graph allocator
-/*
-  Example usage:
-    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
-
-    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
-    ggml_gallocr_reserve(galloc, build_graph(max_batch));
-
-    // allocate the graph
-    struct ggml_cgraph * graph = build_graph(batch);
-    ggml_gallocr_alloc_graph(galloc, graph);
-
-    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
-
-    // evaluate the graph
-    ggml_backend_graph_compute(backend, graph);
-*/
-
-// special tensor flags for use with the graph allocator:
-//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
-//   ggml_set_output(): output tensors are never freed and never overwritten
-
-typedef struct ggml_gallocr * ggml_gallocr_t;
-
-GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
-GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
-GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
-
-// pre-allocate buffers from a measure graph - does not allocate or modify the graph
-// call with a worst-case graph to avoid buffer reallocations
-// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
-// returns false if the buffer allocation failed
-GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
-GGML_API bool ggml_gallocr_reserve_n(
-    ggml_gallocr_t galloc,
-    struct ggml_cgraph * graph,
-    const int * node_buffer_ids,
-    const int * leaf_buffer_ids);
-
-// automatic reallocation if the topology changes when using a single buffer
-// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
-GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
-
-GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
-
-// Utils
-// Create a buffer and allocate all the tensors in a ggml_context
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
-
-#ifdef  __cplusplus
-}
-#endif
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct ggml_backend * ggml_backend_t;
+
+// Tensor allocator
+struct ggml_tallocr {
+    ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};
+
+GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
+
+// Graph allocator
+/*
+  Example usage:
+    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
+
+    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
+    ggml_gallocr_reserve(galloc, build_graph(max_batch));
+
+    // allocate the graph
+    struct ggml_cgraph * graph = build_graph(batch);
+    ggml_gallocr_alloc_graph(galloc, graph);
+
+    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
+
+    // evaluate the graph
+    ggml_backend_graph_compute(backend, graph);
+*/
+
+// special tensor flags for use with the graph allocator:
+//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
+//   ggml_set_output(): output tensors are never freed and never overwritten
+
+typedef struct ggml_gallocr * ggml_gallocr_t;
+
+GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
+GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
+GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
+
+// pre-allocate buffers from a measure graph - does not allocate or modify the graph
+// call with a worst-case graph to avoid buffer reallocations
+// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
+// returns false if the buffer allocation failed
+GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API bool ggml_gallocr_reserve_n(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);
+
+// automatic reallocation if the topology changes when using a single buffer
+// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
+GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+
+GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
+
+// Utils
+// Create a buffer and allocate all the tensors in a ggml_context
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
+
+#ifdef  __cplusplus
+}
+#endif

llama/ggml-backend-impl.h

@@ -1,141 +1,141 @@
-#pragma once
-
-// ggml-backend internal header
-
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
-    struct ggml_backend_buffer_type_i {
-        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
-        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
-        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
-        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
-        // check if tensor data is in host memory
-        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
-        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
-    };
-
-    struct ggml_backend_buffer_type {
-        struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
-    };
-
-    // buffer
-    typedef void * ggml_backend_buffer_context_t;
-
-    struct ggml_backend_buffer_i {
-        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
-        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
-        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
-        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-    };
-
-    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i  iface;
-        ggml_backend_buffer_type_t    buft;
-        ggml_backend_buffer_context_t context;
-        size_t size;
-        enum ggml_backend_buffer_usage usage;
-    };
-
-    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
-            struct ggml_backend_buffer_i           iface,
-                   ggml_backend_buffer_context_t   context,
-                   size_t                          size);
-
-    // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // buffer that contains a collection of buffers
-    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-
-    //
-    // Backend
-    //
-
-    typedef void * ggml_backend_context_t;
-
-    struct ggml_backend_i {
-        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
-
-        void (*GGML_CALL free)(ggml_backend_t backend);
-
-        // buffer allocation
-        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
-
-        // (optional) asynchronous tensor data access
-        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-        // (optional) complete all pending operations
-        void (*GGML_CALL synchronize)(ggml_backend_t backend);
-
-        // compute graph with a plan (not used currently)
-        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
-        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-        // compute graph with a plan
-        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-        // check if the backend supports an operation
-        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // (optional) event synchronization
-        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
-        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
-        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
-    };
-
-    struct ggml_backend {
-        ggml_guid_t guid;
-
-        struct ggml_backend_i iface;
-        ggml_backend_context_t context;
-    };
-
-    struct ggml_backend_event {
-        ggml_backend_t backend;
-        void * context;
-    };
-
-    //
-    // Backend registry
-    //
-
-    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
-
-    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
-
-#ifdef  __cplusplus
-}
-#endif
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    // buffer type
+    typedef void * ggml_backend_buffer_type_context_t;
+
+    struct ggml_backend_buffer_type_i {
+        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
+        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
+        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        // check if tensor data is in host memory
+        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
+        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
+    typedef void * ggml_backend_buffer_context_t;
+
+    struct ggml_backend_buffer_i {
+        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
+        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
+        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
+        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
+        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+    };
+
+    struct ggml_backend_buffer {
+        struct ggml_backend_buffer_i  iface;
+        ggml_backend_buffer_type_t    buft;
+        ggml_backend_buffer_context_t context;
+        size_t size;
+        enum ggml_backend_buffer_usage usage;
+    };
+
+    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
+                   ggml_backend_buffer_type_t      buft,
+            struct ggml_backend_buffer_i           iface,
+                   ggml_backend_buffer_context_t   context,
+                   size_t                          size);
+
+    // do not use directly, use ggml_backend_tensor_copy instead
+    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // buffer that contains a collection of buffers
+    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+
+    //
+    // Backend
+    //
+
+    typedef void * ggml_backend_context_t;
+
+    struct ggml_backend_i {
+        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
+
+        void (*GGML_CALL free)(ggml_backend_t backend);
+
+        // buffer allocation
+        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
+
+        // (optional) asynchronous tensor data access
+        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        // (optional) complete all pending operations
+        void (*GGML_CALL synchronize)(ggml_backend_t backend);
+
+        // compute graph with a plan (not used currently)
+        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
+        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph with a plan
+        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // compute graph without a plan (async)
+        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+        // check if the backend supports an operation
+        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
+        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
+        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
+    };
+
+    struct ggml_backend {
+        ggml_guid_t guid;
+
+        struct ggml_backend_i iface;
+        ggml_backend_context_t context;
+    };
+
+    struct ggml_backend_event {
+        ggml_backend_t backend;
+        void * context;
+    };
+
+    //
+    // Backend registry
+    //
+
+    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
+
+    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
+#ifdef  __cplusplus
+}
+#endif

llama/ggml-backend.c

@@ -8,6 +8,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <malloc.h>
 
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
@@ -56,7 +57,6 @@ bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
 }
 
 // backend buffer
-
 GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
                ggml_backend_buffer_type_t      buft,
         struct ggml_backend_buffer_i           iface,
@@ -79,6 +79,10 @@ const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
     return buffer->iface.get_name(buffer);
 }
 
+#define ggml_assert_aligned(ptr) \
+    GGML_ASSERT(((uintptr_t) (ptr))%GGML_MEM_ALIGN == 0)
+
+
 void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
     if (buffer == NULL) {
         return;
@@ -87,7 +91,12 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
     if (buffer->iface.free_buffer != NULL) {
         buffer->iface.free_buffer(buffer);
     }
+    
+    // TODO: this needs to be freed in cuda and hipblas backends because
+    // the cuda backend implementation compiled with msvc
+#if !defined(GGML_USE_CUDA) && !defined(GGML_USE_HIPBLAS)
     free(buffer);
+#endif
 }
 
 size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
@@ -822,7 +831,11 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_CPY:
-            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
         default:
@@ -1178,9 +1191,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 static char * fmt_size(size_t size) {
     static char buffer[128];
     if (size >= 1024*1024) {
-        sprintf(buffer, "%zuM", size/1024/1024);
+        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
     } else {
-        sprintf(buffer, "%zuK", size/1024);
+        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
     }
     return buffer;
 }
@@ -1721,23 +1734,23 @@ ggml_backend_sched_t ggml_backend_sched_new(
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
 
-    struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);
+    struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
 
     // initialize hash table
     sched->hash_set          = ggml_hash_set_new(graph_size);
-    sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
-    sched->tensor_copies     = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
+    sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
+    sched->tensor_copies     = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
 
     const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), nodes_size);
-    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), nodes_size);
+    sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
 
     sched->n_backends = n_backends;
 
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
     const int initial_splits_capacity = 16;
-    sched->splits = calloc(sizeof(sched->splits[0]), initial_splits_capacity);
+    sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
     sched->splits_capacity = initial_splits_capacity;
 
     for (int b = 0; b < n_backends; b++) {
@@ -1780,12 +1793,14 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
 
 void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
     // reset state for the next run
-    size_t hash_size = sched->hash_set.size;
-    memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
-    memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
-    memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
+    if (!sched->is_reset) {
+        size_t hash_size = sched->hash_set.size;
+        memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
+        memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
+        memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
 
-    sched->is_reset = true;
+        sched->is_reset = true;
+    }
     sched->is_alloc = false;
 }
 
@@ -1889,7 +1904,6 @@ void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * t
 
     tensor->buffer = buffer;
     tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
-    tensor->backend = tensor->view_src->backend;
     ggml_backend_buffer_init_tensor(buffer, tensor);
 }
 
@@ -1968,10 +1982,10 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
     struct ggml_hash_set hash_set = {
         /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(sizeof(hash_set.keys[0]), graph->visited_hash_table.size) // NOLINT
+        /* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
     };
-    struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]), hash_set.size); // NOLINT
-    bool * node_init = calloc(sizeof(node_init[0]), hash_set.size);
+    struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
+    bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
 
     struct ggml_init_params params = {
         /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),

llama/ggml-backend.h

@@ -1,233 +1,233 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-    typedef struct ggml_backend_event * ggml_backend_event_t;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
-    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
-
-    // buffer
-    enum ggml_backend_buffer_usage {
-        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
-        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
-    };
-
-    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
-
-    //
-    // Backend
-    //
-
-    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
-    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
-    GGML_API void         ggml_backend_free(ggml_backend_t backend);
-
-    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
-    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
-
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
-    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
-
-    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // asynchronous copy
-    // the copy is performed after all the currently queued operations in backend_src
-    // backend_dst will wait for the copy to complete before performing other operations
-    // automatic fallback to sync copy if async is not supported
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // events
-    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
-    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
-
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
-
-    // Create a backend buffer from an existing pointer
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
-
-    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    //
-    // Backend registry
-    //
-
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
-
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
-
-    //
-    // Backend scheduler
-    //
-
-    // The backend scheduler allows for multiple backends to be used together
-    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
-    // The backends are selected based on:
-    // - the backend that supports the operation
-    // - the location of the pre-allocated tensors (e.g. the weights)
-    /*
-      Example usage:
-
-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
-        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
-
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
-
-        // initialize buffers from a max size graph (optional)
-        reserve_graph = build_graph(sched, max_batch_size);
-
-        // manually assign nodes to a backend (optional, should not be needed in most cases)
-        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
-
-        ggml_backend_sched_reserve(sched, reserve_graph);
-
-        // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
-
-        // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
-    }
-    */
-
-    struct ggml_backend_sched;
-    typedef struct ggml_backend_sched * ggml_backend_sched_t;
-
-    // when ask == true, the scheduler wants to know if the user wants to observe this node
-    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
-    //
-    // when ask == false, the scheduler is passing the node tensor to the user for observation
-    // if the user returns false, the scheduler will cancel the graph compute
-    //
-    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
-
-    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
-    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
-
-    // Initialize backend buffers from a measure graph
-    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
-
-    // Get the number of splits of the last graph
-    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
-    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
-
-    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
-
-    // Allocate and compute graph on the backend scheduler
-    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
-
-    // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
-
-    // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
-
-    //
-    // Utils
-    //
-
-    struct ggml_backend_graph_copy {
-        ggml_backend_buffer_t buffer;
-        struct ggml_context * ctx_allocated;
-        struct ggml_context * ctx_unallocated;
-        struct ggml_cgraph * graph;
-    };
-
-    // Copy a graph to a different backend
-    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
-    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
-
-    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
-
-    // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
-
-    // Tensor initialization
-    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-
-
-#ifdef  __cplusplus
-}
-#endif
+#pragma once
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend_event * ggml_backend_event_t;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
+    //
+    // Backend buffer
+    //
+
+    // buffer type
+    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
+
+    // buffer
+    enum ggml_backend_buffer_usage {
+        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
+        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
+    };
+
+    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+
+    //
+    // Backend
+    //
+
+    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
+    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
+    GGML_API void         ggml_backend_free(ggml_backend_t backend);
+
+    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
+    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
+
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // asynchronous copy
+    // the copy is performed after all the currently queued operations in backend_src
+    // backend_dst will wait for the copy to complete before performing other operations
+    // automatic fallback to sync copy if async is not supported
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // events
+    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
+    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    // Create a backend buffer from an existing pointer
+    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+
+    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
+    //
+    // Backend registry
+    //
+
+    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
+
+    GGML_API size_t                     ggml_backend_reg_get_count(void);
+    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
+    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
+    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
+    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
+
+    //
+    // Backend scheduler
+    //
+
+    // The backend scheduler allows for multiple backends to be used together
+    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
+    // The backends are selected based on:
+    // - the backend that supports the operation
+    // - the location of the pre-allocated tensors (e.g. the weights)
+    /*
+      Example usage:
+
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
+        // preferrably to run on the same backend as the buffer
+        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+
+        // initialize buffers from a max size graph (optional)
+        reserve_graph = build_graph(sched, max_batch_size);
+
+        // manually assign nodes to a backend (optional, should not be needed in most cases)
+        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
+
+        ggml_backend_sched_reserve(sched, reserve_graph);
+
+        // compute
+        graph = build_graph(sched);
+        ggml_backend_sched_graph_compute(sched, graph);
+
+        // if there are graph inputs:
+        ggml_backend_sched_reset(sched);
+        ggml_backend_sched_alloc_graph(sched, graph);
+        ggml_backend_tensor_set(input_tensor, ...);
+        ggml_backend_sched_graph_compute(sched, graph);
+    }
+    */
+
+    struct ggml_backend_sched;
+    typedef struct ggml_backend_sched * ggml_backend_sched_t;
+
+    // when ask == true, the scheduler wants to know if the user wants to observe this node
+    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
+    //
+    // when ask == false, the scheduler is passing the node tensor to the user for observation
+    // if the user returns false, the scheduler will cancel the graph compute
+    //
+    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
+
+    // Initialize a backend scheduler
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+
+    // Initialize backend buffers from a measure graph
+    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+
+    // Get the number of splits of the last graph
+    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
+
+    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+
+    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+
+    // Allocate and compute graph on the backend scheduler
+    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
+
+    // Reset all assignments and allocators - must be called before changing the node backends
+    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
+
+    // Set a callback to be called for each resulting node during graph compute
+    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
+
+    //
+    // Utils
+    //
+
+    struct ggml_backend_graph_copy {
+        ggml_backend_buffer_t buffer;
+        struct ggml_context * ctx_allocated;
+        struct ggml_context * ctx_unallocated;
+        struct ggml_cgraph * graph;
+    };
+
+    // Copy a graph to a different backend
+    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
+    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
+
+    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+
+    // Compare the output of two backends
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+
+    // Tensor initialization
+    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
+    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+
+#ifdef  __cplusplus
+}
+#endif

llama/ggml-common.h

@@ -1,1853 +1,1853 @@
-#ifndef GGML_COMMON_DECL
-
-#if defined(GGML_COMMON_DECL_C)
-#include <stdint.h>
-
-typedef uint16_t ggml_half;
-typedef uint32_t ggml_half2;
-
-#define GGML_COMMON_AGGR
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_METAL)
-#include <metal_stdlib>
-
-typedef half  ggml_half;
-typedef half2 ggml_half2;
-
-#define GGML_COMMON_AGGR
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_CUDA)
-#include <cuda_fp16.h>
-#include <cstdint>
-
-typedef half  ggml_half;
-typedef half2 ggml_half2;
-
-#define GGML_COMMON_AGGR data
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_HIP)
-#include <hip/hip_fp16.h>
-#include <cstdint>
-
-typedef half  ggml_half;
-typedef half2 ggml_half2;
-
-#define GGML_COMMON_AGGR data
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_SYCL)
-#include <sycl/half_type.hpp>
-#include <cstdint>
-
-typedef sycl::half  ggml_half;
-typedef sycl::half2 ggml_half2;
-
-#define GGML_COMMON_AGGR data
-
-#define GGML_COMMON_DECL
-#endif
-
-#if defined(GGML_COMMON_DECL)
-
-#ifndef __cplusplus
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-#endif // __cplusplus
-
-// QK = number of values after dequantization
-// QK_K = super-block size
-
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif // GGML_QKK_64
-
-#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
-// QR = QK / number of values before dequantization
-// QI = number of 32 bit integers before dequantization
-
-#define QI4_0 (QK4_0 / (4 * QR4_0))
-#define QR4_0 2
-
-#define QI4_1 (QK4_1 / (4 * QR4_1))
-#define QR4_1 2
-
-#define QI5_0 (QK5_0 / (4 * QR5_0))
-#define QR5_0 2
-
-#define QI5_1 (QK5_1 / (4 * QR5_1))
-#define QR5_1 2
-
-#define QI8_0 (QK8_0 / (4 * QR8_0))
-#define QR8_0 1
-
-#define QI8_1 (QK8_1 / (4 * QR8_1))
-#define QR8_1 1
-
-#define QI2_K (QK_K / (4*QR2_K))
-#define QR2_K 4
-
-#define QI3_K (QK_K / (4*QR3_K))
-#define QR3_K 4
-
-#define QI4_K (QK_K / (4*QR4_K))
-#define QR4_K 2
-
-#define QI5_K (QK_K / (4*QR5_K))
-#define QR5_K 2
-
-#define QI6_K (QK_K / (4*QR6_K))
-#define QR6_K 2
-
-#define QI2_XXS (QK_K / (4*QR2_XXS))
-#define QR2_XXS 8
-
-#define QI2_XS (QK_K / (4*QR2_XS))
-#define QR2_XS 8
-
-#define QI2_S (QK_K / (4*QR2_S))
-#define QR2_S 8
-
-#define QI3_XXS (QK_K / (4*QR3_XXS))
-#define QR3_XXS 8
-
-#define QI3_XS (QK_K / (4*QR3_XS))
-#define QR3_XS 8
-
-#define QI1_S (QK_K / (4*QR1_S))
-#define QR1_S 8
-
-#define QI4_NL (QK4_NL / (4*QR4_NL))
-#define QR4_NL 2
-
-#if QK_K == 64
-#define QI4_XS QI4_NL
-#define QR4_XS QR4_NL
-#else
-#define QI4_XS (QK_K / (4*QR4_XS))
-#define QR4_XS 8
-#endif
-
-#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
-
-#define QK4_0 32
-typedef struct {
-    ggml_half d;           // delta
-    uint8_t qs[QK4_0 / 2]; // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-typedef struct {
-    union {
-        struct {
-            ggml_half d; // delta
-            ggml_half m; // min
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t qs[QK4_1 / 2]; // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-typedef struct {
-    ggml_half d;           // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-typedef struct {
-    union {
-        struct {
-            ggml_half d; // delta
-            ggml_half m; // min
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2]; // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-typedef struct {
-    ggml_half d;       // delta
-    int8_t  qs[QK8_0]; // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-typedef struct {
-    union {
-        struct {
-            ggml_half d; // delta
-            ggml_half s; // d * sum(qs[i])
-        } GGML_COMMON_AGGR;
-        ggml_half2 ds;
-    };
-    int8_t qs[QK8_1]; // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
-
-//
-// Super-block quantization structures
-//
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elements each
-// Effectively 2.625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    union {
-        struct {
-            ggml_half d;    // super-block scale for quantized scales
-            ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[12];    // scales, quantized with 6 bits
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
-
-// 4-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d[2];     // super-block scales/mins
-    uint8_t scales[2];  // 4-bit block scales/mins
-    uint8_t qs[QK_K/2]; // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    union {
-        struct {
-            ggml_half d;    // super-block scale for quantized scales
-            ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];           // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-// 5-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d;             // super-block scale
-    int8_t  scales[QK_K/16]; // 8-bit block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    union {
-        struct {
-            ggml_half d;    // super-block scale for quantized scales
-            ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];           // quants, high bit
-    uint8_t qs[QK_K/2];           // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    ggml_half d;             // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-// (Almost) "true" 2-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 2.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_half d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-// 2.3125 bpw quants
-typedef struct {
-    ggml_half d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
-
-// 2.5625 bpw quants
-typedef struct {
-    ggml_half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t scales[QK_K/32];
-} block_iq2_s;
-static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
-
-// (Almost) "true" 3-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 3.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_half d;
-    uint8_t qs[3*QK_K/8];
-} block_iq3_xxs;
-static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
-
-// 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    ggml_half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
-typedef struct {
-    ggml_half d;
-    uint8_t  qs[QK_K/8];
-    uint16_t qh[QK_K/32];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-// 1.75 bpw
-typedef struct {
-    uint8_t  qs[QK_K/8];      // grid index, low 8 bits
-    uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
-    uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
-} block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif
-
-// Used by IQ1_M quants
-typedef union {
-    ggml_half f16;
-    uint16_t  u16;
-} iq1m_scale_t;
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    ggml_half d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
-typedef struct {
-    ggml_half d;
-    uint16_t scales_h;
-    uint8_t  scales_l[QK_K/64];
-    uint8_t  qs[QK_K/2];
-} block_iq4_xs;
-static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
-
-#endif // GGML_COMMON_DECL
-#endif // GGML_COMMON_DECL
-
-////////////////////////////////////////////////////////////////////////////////
-
-#ifndef GGML_COMMON_IMPL
-
-#if defined(GGML_COMMON_IMPL_C)
-#include <stdint.h>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_METAL)
-#include <metal_stdlib>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
-#include <cstdint>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_SYCL)
-
-#include <cstdint>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#endif
-
-#if defined(GGML_COMMON_IMPL)
-
-GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8)
-    1, 2, 4, 8, 16, 32, 64, 128
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
-      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
-    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
-    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
-     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
-    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
-     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
-     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
-    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
-GGML_TABLE_END()
-
-//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
-    0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
-    0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
-    0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
-    0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
-    0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
-    0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
-    0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
-    0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
-    0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
-    0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
-    0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
-    0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
-    0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
-    0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
-    0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
-    0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
-    0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
-    0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
-    0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
-    0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
-    0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
-    0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
-    0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
-    0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
-    0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
-    0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
-    0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
-    0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
-    0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
-    0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
-    0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
-    0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
-GGML_TABLE_END()
-//#endif
-
-
-GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
-    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
-    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
-    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
-    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
-    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
-    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
-    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
-    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
-    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
-    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
-    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
-    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
-    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
-    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
-    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
-    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
-    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
-    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
-    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
-    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
-    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
-    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
-    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
-    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
-    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
-    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
-    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
-    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
-    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
-    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
-    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
-    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
-    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
-    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
-    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
-    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
-    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
-    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
-    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
-    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
-    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
-    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
-    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
-    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
-    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
-    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
-    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
-    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
-    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
-    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
-    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
-    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
-    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
-    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
-    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
-    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
-    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
-    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
-    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
-    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
-    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
-    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
-    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
-    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
-    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
-    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
-    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
-    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
-    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
-    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
-    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
-    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
-    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
-    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
-    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
-    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
-    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
-    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
-    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
-    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
-    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
-    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
-    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
-    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
-    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
-    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
-    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
-    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
-    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
-    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
-    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
-    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
-    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
-    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
-    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
-    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
-    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
-    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
-    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
-    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
-    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
-    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
-    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
-    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
-    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
-    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
-    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
-    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
-    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
-    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
-    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
-    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
-    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
-    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
-    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
-    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
-    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
-    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
-    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
-    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
-    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
-    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
-    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
-    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
-    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
-    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
-    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
-    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
-    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
-    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
-    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
-    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
-    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
-    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
-    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
-    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
-    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
-    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
-    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
-    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
-    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
-    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
-    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
-    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
-    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
-    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
-    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
-    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
-    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
-    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
-    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
-    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
-    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
-    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
-    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
-    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
-    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
-    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
-    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
-    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
-    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
-    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
-    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
-    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
-    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
-    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
-    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
-    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
-    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
-    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
-    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
-    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
-    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
-    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
-    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
-    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
-    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
-    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
-    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
-    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
-    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
-    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
-    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
-    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
-    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
-    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
-    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
-    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
-    0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
-    0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
-    0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
-    0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
-    0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
-    0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
-    0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
-    0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
-    0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
-    0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
-    0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
-    0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
-    0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
-    0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
-    0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
-    0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
-    0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
-    0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
-    0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
-    0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
-    0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
-    0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
-    0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
-    0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
-    0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
-    0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
-    0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
-    0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
-    0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
-    0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
-    0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
-    0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
-    0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
-    0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
-    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
-    0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
-    0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
-    0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
-    0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
-    0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
-    0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
-    0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
-    0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
-    0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
-    0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
-    0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
-    0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
-    0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
-    0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
-    0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
-    0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
-    0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
-    0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
-    0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
-    0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
-    0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
-    0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
-    0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
-    0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
-    0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
-    0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
-    0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
-    0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
-    0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
-    0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
-    0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
-    0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
-    0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
-    0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
-    0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
-    0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
-    0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
-    0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
-    0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
-    0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
-    0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
-    0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
-    0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
-    0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
-    0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
-    0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
-    0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
-    0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
-    0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
-    0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
-    0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
-    0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
-    0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
-    0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
-    0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
-    0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
-    0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
-    0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
-    0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
-    0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
-    0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
-    0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
-    0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
-    0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
-    0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
-    0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
-    0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
-    0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
-    0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
-    0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
-    0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
-    0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
-    0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
-    0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
-    0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
-    0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
-    0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
-    0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
-    0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
-    0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
-    0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
-    0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
-    0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
-    0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
-    0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
-    0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
-    0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
-    0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
-    0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
-    0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
-    0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
-    0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
-    0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
-    0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
-    0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
-    0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
-    0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
-    0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
-    0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
-    0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
-    0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
-    0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
-    0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
-    0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
-    0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
-    0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
-    0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
-    0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
-    0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
-    0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
-    0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
-    0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
-    0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
-    0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
-    0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
-    0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
-    0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
-    0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
-    0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
-    0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
-    0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
-    0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
-    0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
-    0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
-    0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
-    0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
-    0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
-    0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
-    0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
-    0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
-    0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
-    0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
-    0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
-    0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
-    0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
-    0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
-    0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
-    0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
-    0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
-    0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
-    0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
-    0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
-    0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
-    0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
-    0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
-    0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
-    0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
-    0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
-    0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
-    0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
-    0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
-    0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
-    0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
-    0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
-    0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
-    0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
-    0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
-    0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
-    0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
-    0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
-    0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
-    0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
-    0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
-    0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
-    0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
-    0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
-    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
-    0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
-    0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
-    0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
-    0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
-    0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
-    0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
-    0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
-    0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
-    0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
-    0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
-    0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
-    0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
-    0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
-    0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
-    0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
-    0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
-    0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
-    0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
-    0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
-    0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
-    0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
-    0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
-    0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
-    0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
-    0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
-    0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
-    0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
-    0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
-    0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
-    0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
-    0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
-    0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
-    0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
-    0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
-    0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
-    0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
-    0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
-    0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
-    0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
-    0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
-    0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
-    0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
-    0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
-    0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
-    0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
-    0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
-    0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
-    0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256)
-    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
-    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
-    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
-    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
-    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
-    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
-    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
-    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
-    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
-    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
-    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
-    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
-    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
-    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
-    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
-    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
-    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
-    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
-    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
-    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
-    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
-    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
-    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
-    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
-    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
-    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
-    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
-    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
-    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
-    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
-    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
-    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
-    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
-    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
-    0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
-    0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
-    0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
-    0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
-    0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
-    0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
-    0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
-    0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
-    0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
-    0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
-    0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
-    0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
-    0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
-    0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
-    0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
-    0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
-    0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
-    0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
-    0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
-    0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
-    0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
-    0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
-    0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
-    0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
-    0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
-    0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
-    0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
-    0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
-    0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
-    0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
-    0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
-    0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
-    0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
-    0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
-    0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
-    0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
-    0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
-    0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
-    0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
-    0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
-    0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
-    0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
-    0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
-    0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
-    0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
-    0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
-    0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
-    0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
-    0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
-    0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
-    0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
-    0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
-    0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
-    0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
-    0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
-    0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
-    0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
-    0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
-    0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
-    0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
-    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
-    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
-GGML_TABLE_END()
-
-#define NGRID_IQ1S 2048
-#define IQ1S_DELTA 0.125f
-#define IQ1M_DELTA 0.125f
-#if defined(GGML_COMMON_IMPL_C)
-GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
-    0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
-    0xffffffffffff0101, 0xffffffffff00ff00, 0xffffffffff000000, 0xffffffffff01ffff,
-    0xffffffffff01ff01, 0xffffffffff0101ff, 0xffffffffff010101, 0xffffffff00ff0000,
-    0xffffffff0000ff00, 0xffffffff000000ff, 0xffffffff00000001, 0xffffffff00010000,
-    0xffffffff01ffffff, 0xffffffff01ffff01, 0xffffffff01ff01ff, 0xffffffff01ff0101,
-    0xffffffff01000000, 0xffffffff0101ffff, 0xffffffff0101ff01, 0xffffffff010101ff,
-    0xffffffff01010101, 0xffffff00ffff00ff, 0xffffff00ffff0000, 0xffffff00ff00ff00,
-    0xffffff00ff0000ff, 0xffffff00ff000001, 0xffffff00ff000100, 0xffffff00ff000101,
-    0xffffff00ff010000, 0xffffff0000ffff00, 0xffffff0000ff0001, 0xffffff0000ff0100,
-    0xffffff000000ff01, 0xffffff0000000000, 0xffffff0000000101, 0xffffff000001ff00,
-    0xffffff00000100ff, 0xffffff0000010001, 0xffffff00000101ff, 0xffffff0001ff0000,
-    0xffffff000100ff00, 0xffffff00010000ff, 0xffffff0001000001, 0xffffff0001010000,
-    0xffffff01ffffffff, 0xffffff01ffffff01, 0xffffff01ffff01ff, 0xffffff01ffff0101,
-    0xffffff01ff000000, 0xffffff01ff01ffff, 0xffffff01ff01ff01, 0xffffff01ff0101ff,
-    0xffffff01ff010101, 0xffffff0100ff0000, 0xffffff010000ff00, 0xffffff0100000100,
-    0xffffff01000100ff, 0xffffff0100010100, 0xffffff0101ffffff, 0xffffff0101ffff01,
-    0xffffff0101ff01ff, 0xffffff0101ff0101, 0xffffff010100ff00, 0xffffff0101000000,
-    0xffffff0101000100, 0xffffff010101ffff, 0xffffff010101ff01, 0xffffff01010101ff,
-    0xffffff0101010101, 0xffff00ffff00ff00, 0xffff00ffff0000ff, 0xffff00ffff000001,
-    0xffff00ffff010000, 0xffff00ff00ffff00, 0xffff00ff00ff0100, 0xffff00ff00000000,
-    0xffff00ff00000101, 0xffff00ff000100ff, 0xffff00ff00010000, 0xffff00ff0100ff00,
-    0xffff00ff01000100, 0xffff00ff01010000, 0xffff0000ffffff00, 0xffff0000ffff00ff,
-    0xffff0000ffff0000, 0xffff0000ffff0001, 0xffff0000ff000000, 0xffff0000ff0001ff,
-    0xffff0000ff000101, 0xffff0000ff010100, 0xffff000000ffffff, 0xffff000000ff0000,
-    0xffff000000ff0101, 0xffff00000000ffff, 0xffff00000000ff00, 0xffff0000000000ff,
-    0xffff000000000000, 0xffff000000000001, 0xffff000000000100, 0xffff00000001ffff,
-    0xffff00000001ff01, 0xffff000000010000, 0xffff0000000101ff, 0xffff000000010101,
-    0xffff000001ffff00, 0xffff00000100ff00, 0xffff000001000000, 0xffff0000010001ff,
-    0xffff000001000101, 0xffff00000101ff00, 0xffff0000010100ff, 0xffff000001010000,
-    0xffff000001010001, 0xffff000001010100, 0xffff0001ff0000ff, 0xffff0001ff000100,
-    0xffff000100ffff00, 0xffff000100ff00ff, 0xffff00010000ffff, 0xffff00010000ff01,
-    0xffff000100000000, 0xffff0001000001ff, 0xffff00010001ffff, 0xffff00010001ff00,
-    0xffff000100010001, 0xffff000100010100, 0xffff000101ff0000, 0xffff00010100ff00,
-    0xffff0001010000ff, 0xffff000101000100, 0xffff01ffffffffff, 0xffff01ffffffff01,
-    0xffff01ffffff01ff, 0xffff01ffffff0101, 0xffff01ffff000000, 0xffff01ffff01ffff,
-    0xffff01ffff01ff01, 0xffff01ffff0101ff, 0xffff01ffff010101, 0xffff01ff00ff0000,
-    0xffff01ff0000ff00, 0xffff01ff00000001, 0xffff01ff00010000, 0xffff01ff01ffffff,
-    0xffff01ff01ffff01, 0xffff01ff01ff01ff, 0xffff01ff01ff0101, 0xffff01ff01000000,
-    0xffff01ff0101ffff, 0xffff01ff0101ff01, 0xffff01ff010101ff, 0xffff01ff01010101,
-    0xffff0100ffff0000, 0xffff0100ff00ff00, 0xffff0100ff0000ff, 0xffff0100ff000100,
-    0xffff0100ff0100ff, 0xffff0100ff010000, 0xffff010000ffff00, 0xffff01000000ffff,
-    0xffff01000000ff00, 0xffff010000000000, 0xffff01000001ff00, 0xffff0100000100ff,
-    0xffff010000010100, 0xffff01000100ff00, 0xffff0100010000ff, 0xffff010001000001,
-    0xffff010001000100, 0xffff010001010000, 0xffff0101ffffffff, 0xffff0101ffffff01,
-    0xffff0101ffff01ff, 0xffff0101ffff0101, 0xffff0101ff000000, 0xffff0101ff01ffff,
-    0xffff0101ff01ff01, 0xffff0101ff0101ff, 0xffff0101ff010101, 0xffff010100ff0000,
-    0xffff01010000ff00, 0xffff010100000100, 0xffff01010001ff00, 0xffff010100010000,
-    0xffff010101ffffff, 0xffff010101ffff01, 0xffff010101ff0000, 0xffff010101ff01ff,
-    0xffff010101ff0101, 0xffff010101000000, 0xffff01010101ffff, 0xffff01010101ff01,
-    0xffff0101010101ff, 0xffff010101010101, 0xff00ffffff00ffff, 0xff00ffffff00ff00,
-    0xff00ffffff0000ff, 0xff00ffffff000100, 0xff00ffffff0100ff, 0xff00ffffff010000,
-    0xff00ffff00ffff00, 0xff00ffff00ff00ff, 0xff00ffff0000ffff, 0xff00ffff00000000,
-    0xff00ffff000001ff, 0xff00ffff0001ff00, 0xff00ffff000100ff, 0xff00ffff00010000,
-    0xff00ffff00010100, 0xff00ffff0100ff00, 0xff00ffff010000ff, 0xff00ffff01000001,
-    0xff00ffff0101ff00, 0xff00ffff01010000, 0xff00ff00ffffff00, 0xff00ff00ffff00ff,
-    0xff00ff00ffff0001, 0xff00ff00ffff0100, 0xff00ff00ff00ffff, 0xff00ff00ff00ff01,
-    0xff00ff00ff000000, 0xff00ff00ff0001ff, 0xff00ff00ff01ff00, 0xff00ff00ff0100ff,
-    0xff00ff00ff010100, 0xff00ff0000ff0000, 0xff00ff0000ff0101, 0xff00ff000000ffff,
-    0xff00ff000000ff00, 0xff00ff000000ff01, 0xff00ff00000000ff, 0xff00ff0000000000,
-    0xff00ff0000000001, 0xff00ff0000000100, 0xff00ff000001ffff, 0xff00ff0000010000,
-    0xff00ff0001ff00ff, 0xff00ff000100ff01, 0xff00ff0001000000, 0xff00ff000101ff00,
-    0xff00ff00010100ff, 0xff00ff01ff00ff00, 0xff00ff01ff0000ff, 0xff00ff01ff000001,
-    0xff00ff01ff010000, 0xff00ff0100ffffff, 0xff00ff0100ff0001, 0xff00ff0100ff0100,
-    0xff00ff010000ff01, 0xff00ff0100000000, 0xff00ff01000001ff, 0xff00ff0100000101,
-    0xff00ff01000100ff, 0xff00ff0100010001, 0xff00ff0101ff0000, 0xff00ff010100ff00,
-    0xff00ff01010000ff, 0xff00ff0101000001, 0xff00ff0101010000, 0xff0000ffffffff00,
-    0xff0000ffffff0001, 0xff0000ffffff0100, 0xff0000ffff0000ff, 0xff0000ffff000000,
-    0xff0000ffff0001ff, 0xff0000ffff000100, 0xff0000ffff01ff00, 0xff0000ffff010001,
-    0xff0000ff00ffff00, 0xff0000ff00ff0000, 0xff0000ff00ff0001, 0xff0000ff00ff01ff,
-    0xff0000ff00ff0101, 0xff0000ff0000ff00, 0xff0000ff000000ff, 0xff0000ff00000000,
-    0xff0000ff00000001, 0xff0000ff00000100, 0xff0000ff0001ff01, 0xff0000ff00010000,
-    0xff0000ff000101ff, 0xff0000ff01ff00ff, 0xff0000ff01ff0100, 0xff0000ff0100ffff,
-    0xff0000ff010000ff, 0xff0000ff01000000, 0xff0000ff010001ff, 0xff0000ff01000100,
-    0xff0000ff01000101, 0xff0000ff0101ff00, 0xff0000ff010100ff, 0xff0000ff01010000,
-    0xff0000ff01010100, 0xff000000ffffff01, 0xff000000ffff0000, 0xff000000ffff0101,
-    0xff000000ff00ff00, 0xff000000ff0000ff, 0xff000000ff000000, 0xff000000ff000001,
-    0xff000000ff000100, 0xff000000ff01ffff, 0xff000000ff01ff01, 0xff000000ff010000,
-    0xff000000ff0101ff, 0xff000000ff010101, 0xff00000000ffff00, 0xff00000000ff00ff,
-    0xff00000000ff0000, 0xff00000000ff0001, 0xff0000000000ff00, 0xff0000000000ff01,
-    0xff000000000000ff, 0xff00000000000000, 0xff00000000000001, 0xff00000000000100,
-    0xff00000000000101, 0xff0000000001ff00, 0xff000000000100ff, 0xff00000000010000,
-    0xff00000000010001, 0xff00000000010100, 0xff00000001ffffff, 0xff00000001ffff01,
-    0xff00000001ff00ff, 0xff00000001ff0000, 0xff00000001ff01ff, 0xff00000001ff0101,
-    0xff0000000100ffff, 0xff0000000100ff00, 0xff000000010000ff, 0xff00000001000000,
-    0xff00000001000001, 0xff00000001000100, 0xff00000001000101, 0xff0000000101ffff,
-    0xff0000000101ff01, 0xff00000001010000, 0xff000001ffffff00, 0xff000001ffff00ff,
-    0xff000001ffff0000, 0xff000001ffff0001, 0xff000001ff000000, 0xff000001ff000001,
-    0xff000001ff0001ff, 0xff000001ff000101, 0xff000001ff01ff00, 0xff000001ff010001,
-    0xff00000100ffffff, 0xff00000100ffff01, 0xff00000100ff00ff, 0xff00000100ff0000,
-    0xff00000100ff01ff, 0xff00000100ff0101, 0xff0000010000ff00, 0xff00000100000000,
-    0xff00000100000001, 0xff000001000001ff, 0xff00000100000100, 0xff0000010001ff00,
-    0xff000001000100ff, 0xff00000100010000, 0xff000001000101ff, 0xff00000100010100,
-    0xff00000100010101, 0xff00000101ff0001, 0xff00000101ff0101, 0xff0000010100ff01,
-    0xff00000101000000, 0xff000001010100ff, 0xff00000101010100, 0xff0001ffff00ff00,
-    0xff0001ffff000001, 0xff0001ffff010000, 0xff0001ff00ffff00, 0xff0001ff00ff00ff,
-    0xff0001ff00ff0001, 0xff0001ff00ff0100, 0xff0001ff0000ffff, 0xff0001ff00000000,
-    0xff0001ff000001ff, 0xff0001ff00000101, 0xff0001ff0001ffff, 0xff0001ff0001ff00,
-    0xff0001ff000100ff, 0xff0001ff00010001, 0xff0001ff00010100, 0xff0001ff01ff0000,
-    0xff0001ff0100ff00, 0xff0001ff010000ff, 0xff0001ff01010000, 0xff000100ff00ffff,
-    0xff000100ff00ff01, 0xff000100ff000000, 0xff000100ff000101, 0xff000100ff01ff00,
-    0xff000100ff010000, 0xff00010000ffff01, 0xff00010000ff00ff, 0xff00010000ff0000,
-    0xff00010000ff01ff, 0xff0001000000ff00, 0xff000100000000ff, 0xff00010000000000,
-    0xff00010000000001, 0xff00010000000100, 0xff00010000000101, 0xff0001000001ffff,
-    0xff00010000010000, 0xff00010000010101, 0xff00010001ff0100, 0xff0001000100ff00,
-    0xff0001000100ff01, 0xff00010001000000, 0xff000100010001ff, 0xff0001000101ff00,
-    0xff00010001010001, 0xff00010001010100, 0xff000101ffff0100, 0xff000101ff000001,
-    0xff000101ff0100ff, 0xff000101ff010001, 0xff00010100ff00ff, 0xff00010100ff0001,
-    0xff00010100ff0100, 0xff0001010000ffff, 0xff0001010000ff01, 0xff00010100000000,
-    0xff000101000001ff, 0xff0001010001ff00, 0xff00010100010001, 0xff00010100010100,
-    0xff00010101ff0000, 0xff0001010100ff00, 0xff00010101000001, 0xff00010101000101,
-    0xff01ffffffffffff, 0xff01ffffffffff01, 0xff01ffffffff01ff, 0xff01ffffffff0101,
-    0xff01ffffff000000, 0xff01ffffff01ffff, 0xff01ffffff01ff01, 0xff01ffffff010000,
-    0xff01ffffff0101ff, 0xff01ffffff010101, 0xff01ffff00ff0000, 0xff01ffff0000ff00,
-    0xff01ffff00000100, 0xff01ffff0001ff00, 0xff01ffff00010000, 0xff01ffff01ffffff,
-    0xff01ffff01ffff01, 0xff01ffff01ff01ff, 0xff01ffff01ff0101, 0xff01ffff01000000,
-    0xff01ffff0101ffff, 0xff01ffff0101ff01, 0xff01ffff01010000, 0xff01ffff010101ff,
-    0xff01ffff01010101, 0xff01ff00ffff0000, 0xff01ff00ff00ff00, 0xff01ff00ff0000ff,
-    0xff01ff00ff000100, 0xff01ff00ff010000, 0xff01ff0000ffff01, 0xff01ff0000ff00ff,
-    0xff01ff0000ff0100, 0xff01ff0000000000, 0xff01ff00000001ff, 0xff01ff0000000101,
-    0xff01ff000001ff00, 0xff01ff00000100ff, 0xff01ff0000010000, 0xff01ff0000010001,
-    0xff01ff0001ff0000, 0xff01ff000100ffff, 0xff01ff0001000001, 0xff01ff0001000100,
-    0xff01ff0001010000, 0xff01ff01ffffff00, 0xff01ff01ffff01ff, 0xff01ff01ffff0101,
-    0xff01ff01ff00ff00, 0xff01ff01ff000000, 0xff01ff01ff01ffff, 0xff01ff01ff01ff01,
-    0xff01ff01ff0101ff, 0xff01ff01ff010101, 0xff01ff0100ff0000, 0xff01ff010000ff00,
-    0xff01ff0100000001, 0xff01ff0100000100, 0xff01ff0100010000, 0xff01ff0101ffff00,
-    0xff01ff0101ff01ff, 0xff01ff0101ff0101, 0xff01ff010100ff00, 0xff01ff0101000000,
-    0xff01ff010101ffff, 0xff01ff010101ff01, 0xff01ff01010101ff, 0xff01ff0101010101,
-    0xff0100ffffff0000, 0xff0100ffff0000ff, 0xff0100ffff000001, 0xff0100ffff000100,
-    0xff0100ffff010000, 0xff0100ff00ff00ff, 0xff0100ff00ff0000, 0xff0100ff00ff0001,
-    0xff0100ff00ff0100, 0xff0100ff0000ff01, 0xff0100ff00000000, 0xff0100ff000001ff,
-    0xff0100ff00000101, 0xff0100ff00010001, 0xff0100ff01ff0000, 0xff0100ff0100ff00,
-    0xff0100ff010000ff, 0xff0100ff01000100, 0xff0100ff0101ff00, 0xff0100ff01010000,
-    0xff010000ffff0100, 0xff010000ff000000, 0xff010000ff01ff00, 0xff010000ff010100,
-    0xff01000000ffffff, 0xff01000000ff0000, 0xff01000000ff01ff, 0xff0100000000ff00,
-    0xff010000000000ff, 0xff01000000000000, 0xff01000000000100, 0xff0100000001ff01,
-    0xff01000000010000, 0xff010000000101ff, 0xff01000001ff0100, 0xff0100000100ffff,
-    0xff010000010000ff, 0xff01000001000000, 0xff010000010001ff, 0xff01000001000101,
-    0xff0100000101ff00, 0xff010000010100ff, 0xff01000001010001, 0xff01000001010100,
-    0xff010001ffff0000, 0xff010001ff00ffff, 0xff010001ff00ff01, 0xff010001ff000100,
-    0xff010001ff010000, 0xff01000100ffff00, 0xff01000100ff0100, 0xff01000100000000,
-    0xff0100010001ffff, 0xff0100010001ff00, 0xff01000100010100, 0xff01000101ff00ff,
-    0xff01000101ff0001, 0xff0100010100ffff, 0xff01000101000101, 0xff0101ffffffffff,
-    0xff0101ffffffff01, 0xff0101ffffff01ff, 0xff0101ffffff0101, 0xff0101ffff000000,
-    0xff0101ffff01ffff, 0xff0101ffff01ff01, 0xff0101ffff0101ff, 0xff0101ffff010101,
-    0xff0101ff00ff0000, 0xff0101ff0000ff00, 0xff0101ff000000ff, 0xff0101ff00010000,
-    0xff0101ff01ffffff, 0xff0101ff01ffff01, 0xff0101ff01ff01ff, 0xff0101ff01ff0101,
-    0xff0101ff0101ffff, 0xff0101ff0101ff01, 0xff0101ff010101ff, 0xff0101ff01010101,
-    0xff010100ffff0100, 0xff010100ff00ff00, 0xff010100ff0000ff, 0xff010100ff000100,
-    0xff010100ff010000, 0xff01010000ff0001, 0xff01010000ff0100, 0xff0101000000ff01,
-    0xff01010000000000, 0xff0101000001ff00, 0xff010100000100ff, 0xff01010000010001,
-    0xff01010000010100, 0xff01010001ff0000, 0xff0101000100ffff, 0xff01010001000001,
-    0xff01010001000100, 0xff010100010100ff, 0xff01010001010000, 0xff010101ffffffff,
-    0xff010101ffffff01, 0xff010101ffff01ff, 0xff010101ffff0101, 0xff010101ff01ffff,
-    0xff010101ff01ff01, 0xff010101ff0101ff, 0xff010101ff010101, 0xff01010100ff0000,
-    0xff0101010000ff00, 0xff01010100000001, 0xff01010100000100, 0xff01010100010000,
-    0xff01010101ffffff, 0xff01010101ffff01, 0xff01010101ff01ff, 0xff01010101ff0101,
-    0xff01010101000000, 0xff0101010101ffff, 0xff0101010101ff01, 0xff010101010101ff,
-    0xff01010101010101, 0x00ffffffffff0000, 0x00ffffffff00ff00, 0x00ffffffff000001,
-    0x00ffffffff010000, 0x00ffffff00ff0100, 0x00ffffff0000ff01, 0x00ffffff00000000,
-    0x00ffffff000001ff, 0x00ffffff00000101, 0x00ffffff0001ff00, 0x00ffffff000100ff,
-    0x00ffffff00010001, 0x00ffffff010000ff, 0x00ffffff01000100, 0x00ffffff0101ff00,
-    0x00ffffff01010001, 0x00ffff00ffffffff, 0x00ffff00ffffff00, 0x00ffff00ffff00ff,
-    0x00ffff00ffff0001, 0x00ffff00ffff0100, 0x00ffff00ff00ff01, 0x00ffff00ff000000,
-    0x00ffff00ff000001, 0x00ffff00ff0001ff, 0x00ffff00ff000101, 0x00ffff00ff01ff00,
-    0x00ffff00ff010001, 0x00ffff00ff010100, 0x00ffff0000ff0000, 0x00ffff0000ff01ff,
-    0x00ffff0000ff0101, 0x00ffff000000ff00, 0x00ffff00000000ff, 0x00ffff0000000000,
-    0x00ffff0000000001, 0x00ffff0000000100, 0x00ffff0000000101, 0x00ffff0000010000,
-    0x00ffff00000101ff, 0x00ffff0000010101, 0x00ffff0001ffff00, 0x00ffff0001ff00ff,
-    0x00ffff0001ff0001, 0x00ffff000100ffff, 0x00ffff000100ff01, 0x00ffff0001000000,
-    0x00ffff000101ffff, 0x00ffff000101ff00, 0x00ffff000101ff01, 0x00ffff01ffff0000,
-    0x00ffff01ff00ff00, 0x00ffff01ff0000ff, 0x00ffff01ff000001, 0x00ffff01ff010000,
-    0x00ffff0100ffff00, 0x00ffff010000ff01, 0x00ffff0100000000, 0x00ffff0100000101,
-    0x00ffff01000100ff, 0x00ffff0100010100, 0x00ffff0101ff0100, 0x00ffff01010000ff,
-    0x00ffff0101010000, 0x00ff00ffffffff00, 0x00ff00ffff000000, 0x00ff00ffff000100,
-    0x00ff00ffff010100, 0x00ff00ff00ff0000, 0x00ff00ff00ff01ff, 0x00ff00ff00ff0101,
-    0x00ff00ff0000ff00, 0x00ff00ff000000ff, 0x00ff00ff00000000, 0x00ff00ff00000001,
-    0x00ff00ff0001ff00, 0x00ff00ff0001ff01, 0x00ff00ff00010000, 0x00ff00ff000101ff,
-    0x00ff00ff00010101, 0x00ff00ff01ffff00, 0x00ff00ff01ff0001, 0x00ff00ff01ff0100,
-    0x00ff00ff0100ffff, 0x00ff00ff0100ff01, 0x00ff00ff01000000, 0x00ff00ff0101ffff,
-    0x00ff00ff0101ff00, 0x00ff00ff01010100, 0x00ff0000ffffff00, 0x00ff0000ffffff01,
-    0x00ff0000ffff0000, 0x00ff0000ffff0101, 0x00ff0000ff00ff00, 0x00ff0000ff0000ff,
-    0x00ff0000ff000000, 0x00ff0000ff000001, 0x00ff0000ff000100, 0x00ff0000ff01ffff,
-    0x00ff0000ff010000, 0x00ff0000ff010101, 0x00ff000000ffff00, 0x00ff000000ff00ff,
-    0x00ff000000ff0000, 0x00ff000000ff0001, 0x00ff000000ff0100, 0x00ff00000000ffff,
-    0x00ff00000000ff00, 0x00ff0000000000ff, 0x00ff000000000000, 0x00ff000000000001,
-    0x00ff0000000001ff, 0x00ff000000000100, 0x00ff00000001ff00, 0x00ff0000000100ff,
-    0x00ff000000010000, 0x00ff000000010001, 0x00ff000000010100, 0x00ff000001ffff01,
-    0x00ff000001ff00ff, 0x00ff000001ff0000, 0x00ff000001ff01ff, 0x00ff00000100ff00,
-    0x00ff0000010000ff, 0x00ff000001000000, 0x00ff000001000001, 0x00ff000001000100,
-    0x00ff000001000101, 0x00ff000001010000, 0x00ff0000010101ff, 0x00ff000001010101,
-    0x00ff0001ffffff00, 0x00ff0001ffff0000, 0x00ff0001ffff0100, 0x00ff0001ff0000ff,
-    0x00ff0001ff000000, 0x00ff0001ff0001ff, 0x00ff0001ff000101, 0x00ff0001ff01ff00,
-    0x00ff0001ff0100ff, 0x00ff0001ff010100, 0x00ff000100ffffff, 0x00ff000100ffff01,
-    0x00ff000100ff0000, 0x00ff000100ff01ff, 0x00ff00010000ffff, 0x00ff00010000ff00,
-    0x00ff00010000ff01, 0x00ff000100000000, 0x00ff000100000001, 0x00ff000100000100,
-    0x00ff00010001ff01, 0x00ff000100010000, 0x00ff0001000101ff, 0x00ff000101ffff00,
-    0x00ff000101ff0000, 0x00ff000101ff0101, 0x00ff0001010000ff, 0x00ff000101000000,
-    0x00ff00010101ff00, 0x00ff0001010100ff, 0x00ff000101010001, 0x00ff01ffffff0000,
-    0x00ff01ffff00ff00, 0x00ff01ffff000000, 0x00ff01ffff000101, 0x00ff01ffff010000,
-    0x00ff01ff00ffff01, 0x00ff01ff00ff0100, 0x00ff01ff0000ffff, 0x00ff01ff00000000,
-    0x00ff01ff000001ff, 0x00ff01ff0001ff00, 0x00ff01ff000100ff, 0x00ff01ff00010001,
-    0x00ff01ff00010100, 0x00ff01ff01ff0000, 0x00ff01ff0100ff00, 0x00ff01ff010000ff,
-    0x00ff01ff01000001, 0x00ff01ff01000100, 0x00ff01ff01010000, 0x00ff0100ffffff00,
-    0x00ff0100ffff0000, 0x00ff0100ffff0001, 0x00ff0100ffff0101, 0x00ff0100ff00ffff,
-    0x00ff0100ff0000ff, 0x00ff0100ff000000, 0x00ff0100ff0001ff, 0x00ff0100ff01ff00,
-    0x00ff0100ff0100ff, 0x00ff0100ff010001, 0x00ff010000ffffff, 0x00ff010000ff0000,
-    0x00ff010000ff0101, 0x00ff01000000ff00, 0x00ff01000000ff01, 0x00ff0100000000ff,
-    0x00ff010000000000, 0x00ff010000000001, 0x00ff010000000100, 0x00ff01000001ffff,
-    0x00ff01000001ff01, 0x00ff010000010000, 0x00ff010000010001, 0x00ff010000010101,
-    0x00ff010001ff0001, 0x00ff010001ff0100, 0x00ff01000100ff01, 0x00ff010001000000,
-    0x00ff010001000001, 0x00ff0100010001ff, 0x00ff01000101ff00, 0x00ff0100010100ff,
-    0x00ff010001010001, 0x00ff010001010100, 0x00ff0101ff000001, 0x00ff010100ff00ff,
-    0x00ff010100ff0001, 0x00ff010100ff0100, 0x00ff010100000000, 0x00ff0101000001ff,
-    0x00ff010100000101, 0x00ff0101000100ff, 0x00ff010100010100, 0x00ff0101010000ff,
-    0x00ff010101010000, 0x0000ffffffffff00, 0x0000ffffffff00ff, 0x0000ffffffff0000,
-    0x0000ffffffff0001, 0x0000ffffffff0100, 0x0000ffffff00ff01, 0x0000ffffff000000,
-    0x0000ffffff000101, 0x0000ffffff01ff00, 0x0000ffffff0100ff, 0x0000ffffff010100,
-    0x0000ffff00ffffff, 0x0000ffff00ff0000, 0x0000ffff00ff01ff, 0x0000ffff0000ff00,
-    0x0000ffff000000ff, 0x0000ffff00000000, 0x0000ffff00000001, 0x0000ffff00000100,
-    0x0000ffff00010000, 0x0000ffff000101ff, 0x0000ffff01ff0001, 0x0000ffff01ff0100,
-    0x0000ffff01000000, 0x0000ffff010001ff, 0x0000ffff0101ffff, 0x0000ffff0101ff00,
-    0x0000ffff01010001, 0x0000ffff01010100, 0x0000ff00ffff0000, 0x0000ff00ffff01ff,
-    0x0000ff00ffff0100, 0x0000ff00ffff0101, 0x0000ff00ff00ff00, 0x0000ff00ff0000ff,
-    0x0000ff00ff000000, 0x0000ff00ff000001, 0x0000ff00ff0001ff, 0x0000ff00ff000100,
-    0x0000ff00ff01ffff, 0x0000ff00ff010000, 0x0000ff00ff010001, 0x0000ff00ff0101ff,
-    0x0000ff00ff010101, 0x0000ff0000ffff00, 0x0000ff0000ff00ff, 0x0000ff0000ff0000,
-    0x0000ff0000ff0001, 0x0000ff0000ff0100, 0x0000ff000000ffff, 0x0000ff000000ff00,
-    0x0000ff000000ff01, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
-    0x0000ff00000001ff, 0x0000ff0000000100, 0x0000ff0000000101, 0x0000ff000001ff00,
-    0x0000ff00000100ff, 0x0000ff0000010000, 0x0000ff0000010001, 0x0000ff0000010100,
-    0x0000ff0001ffff01, 0x0000ff0001ff0000, 0x0000ff000100ff00, 0x0000ff00010000ff,
-    0x0000ff0001000000, 0x0000ff0001000001, 0x0000ff0001000100, 0x0000ff000101ffff,
-    0x0000ff0001010000, 0x0000ff0001010101, 0x0000ff01ffffff00, 0x0000ff01ffff0001,
-    0x0000ff01ff00ff01, 0x0000ff01ff000000, 0x0000ff01ff000101, 0x0000ff01ff01ff00,
-    0x0000ff01ff0100ff, 0x0000ff0100ffff01, 0x0000ff0100ff0000, 0x0000ff0100ff0101,
-    0x0000ff010000ff00, 0x0000ff01000000ff, 0x0000ff0100000000, 0x0000ff0100000001,
-    0x0000ff0100000100, 0x0000ff010001ff01, 0x0000ff0100010000, 0x0000ff0101ff0000,
-    0x0000ff010100ffff, 0x0000ff010100ff01, 0x0000ff0101000000, 0x0000ff0101000100,
-    0x0000ff0101000101, 0x0000ff01010100ff, 0x000000ffffff00ff, 0x000000ffffff0000,
-    0x000000ffff00ff00, 0x000000ffff0000ff, 0x000000ffff000000, 0x000000ffff000001,
-    0x000000ffff0001ff, 0x000000ffff000100, 0x000000ffff01ff00, 0x000000ffff010000,
-    0x000000ffff0101ff, 0x000000ffff010101, 0x000000ff00ffff00, 0x000000ff00ff00ff,
-    0x000000ff00ff0000, 0x000000ff00ff0001, 0x000000ff00ff0100, 0x000000ff00ff0101,
-    0x000000ff0000ffff, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
-    0x000000ff00000001, 0x000000ff000001ff, 0x000000ff00000100, 0x000000ff00000101,
-    0x000000ff0001ff00, 0x000000ff0001ff01, 0x000000ff000100ff, 0x000000ff00010000,
-    0x000000ff00010001, 0x000000ff00010100, 0x000000ff01ffffff, 0x000000ff01ff01ff,
-    0x000000ff01ff0101, 0x000000ff0100ff00, 0x000000ff010000ff, 0x000000ff01000000,
-    0x000000ff01000001, 0x000000ff01000100, 0x000000ff0101ff00, 0x000000ff010100ff,
-    0x000000ff01010000, 0x000000ff01010101, 0x00000000ffffff00, 0x00000000ffffff01,
-    0x00000000ffff00ff, 0x00000000ffff0000, 0x00000000ffff0001, 0x00000000ffff0100,
-    0x00000000ff00ffff, 0x00000000ff00ff00, 0x00000000ff00ff01, 0x00000000ff0000ff,
-    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff000101,
-    0x00000000ff01ff00, 0x00000000ff0100ff, 0x00000000ff010000, 0x00000000ff010001,
-    0x00000000ff010100, 0x0000000000ffffff, 0x0000000000ffff00, 0x0000000000ffff01,
-    0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, 0x0000000000ff01ff,
-    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
-    0x00000000000000ff, 0x0000000000000000, 0x0000000000000001, 0x00000000000001ff,
-    0x0000000000000100, 0x0000000000000101, 0x000000000001ffff, 0x000000000001ff00,
-    0x00000000000100ff, 0x0000000000010000, 0x0000000000010001, 0x00000000000101ff,
-    0x0000000000010100, 0x0000000000010101, 0x0000000001ffff00, 0x0000000001ff00ff,
-    0x0000000001ff0000, 0x0000000001ff0100, 0x0000000001ff0101, 0x000000000100ffff,
-    0x000000000100ff00, 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001,
-    0x00000000010001ff, 0x0000000001000100, 0x000000000101ff00, 0x00000000010100ff,
-    0x0000000001010000, 0x0000000001010001, 0x0000000001010100, 0x00000001ffffffff,
-    0x00000001ffffff00, 0x00000001ffffff01, 0x00000001ffff00ff, 0x00000001ffff0001,
-    0x00000001ffff01ff, 0x00000001ffff0100, 0x00000001ff00ff00, 0x00000001ff0000ff,
-    0x00000001ff000000, 0x00000001ff0001ff, 0x00000001ff000100, 0x00000001ff01ffff,
-    0x00000001ff01ff00, 0x00000001ff01ff01, 0x00000001ff0100ff, 0x00000001ff010000,
-    0x00000001ff010001, 0x00000001ff0101ff, 0x00000001ff010100, 0x0000000100ffff00,
-    0x0000000100ff0000, 0x0000000100ff0001, 0x0000000100ff01ff, 0x0000000100ff0100,
-    0x0000000100ff0101, 0x000000010000ffff, 0x000000010000ff00, 0x000000010000ff01,
-    0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, 0x00000001000001ff,
-    0x0000000100000100, 0x0000000100000101, 0x000000010001ff00, 0x00000001000100ff,
-    0x0000000100010000, 0x0000000100010100, 0x0000000101ffff01, 0x0000000101ff0000,
-    0x0000000101ff0001, 0x0000000101ff01ff, 0x0000000101ff0100, 0x0000000101ff0101,
-    0x000000010100ff00, 0x0000000101000000, 0x0000000101000101, 0x000000010101ff01,
-    0x0000000101010000, 0x0000000101010001, 0x00000001010101ff, 0x0000000101010100,
-    0x000001ffffff00ff, 0x000001ffffff0000, 0x000001ffffff0001, 0x000001ffffff0100,
-    0x000001ffff00ffff, 0x000001ffff000000, 0x000001ffff0001ff, 0x000001ffff01ff00,
-    0x000001ffff010101, 0x000001ff00ff0000, 0x000001ff00ff01ff, 0x000001ff00ff0101,
-    0x000001ff0000ff00, 0x000001ff000000ff, 0x000001ff00000000, 0x000001ff00000001,
-    0x000001ff000001ff, 0x000001ff00000100, 0x000001ff0001ffff, 0x000001ff0001ff01,
-    0x000001ff000100ff, 0x000001ff00010000, 0x000001ff01ffff01, 0x000001ff01ff0100,
-    0x000001ff0100ffff, 0x000001ff0100ff01, 0x000001ff01000000, 0x000001ff010001ff,
-    0x000001ff0101ff00, 0x000001ff01010100, 0x00000100ffffff00, 0x00000100ffffff01,
-    0x00000100ffff0000, 0x00000100ffff0101, 0x00000100ff00ff00, 0x00000100ff0000ff,
-    0x00000100ff000000, 0x00000100ff000001, 0x00000100ff000100, 0x00000100ff010000,
-    0x0000010000ffff00, 0x0000010000ff00ff, 0x0000010000ff0000, 0x0000010000ff0001,
-    0x0000010000ff0100, 0x000001000000ffff, 0x000001000000ff00, 0x000001000000ff01,
-    0x00000100000000ff, 0x0000010000000000, 0x0000010000000001, 0x00000100000001ff,
-    0x0000010000000100, 0x0000010000000101, 0x000001000001ff00, 0x00000100000100ff,
-    0x0000010000010000, 0x0000010000010001, 0x0000010000010100, 0x0000010001ffff00,
-    0x0000010001ff0000, 0x0000010001ff0100, 0x000001000100ff00, 0x00000100010000ff,
-    0x0000010001000000, 0x0000010001000001, 0x00000100010001ff, 0x0000010001000100,
-    0x0000010001010000, 0x00000101ffff00ff, 0x00000101ffff01ff, 0x00000101ff000000,
-    0x00000101ff000101, 0x00000101ff01ffff, 0x00000101ff010000, 0x00000101ff010001,
-    0x00000101ff010100, 0x0000010100ff0000, 0x0000010100ff01ff, 0x0000010100ff0100,
-    0x000001010000ff00, 0x0000010100000000, 0x0000010100000001, 0x00000101000001ff,
-    0x0000010100000100, 0x000001010001ff01, 0x0000010100010000, 0x00000101000101ff,
-    0x0000010100010101, 0x0000010101ffff00, 0x0000010101ff0101, 0x000001010100ff01,
-    0x0000010101000000, 0x0000010101000001, 0x00000101010001ff, 0x0000010101000101,
-    0x000001010101ff00, 0x0001ffffffff0000, 0x0001ffffff0000ff, 0x0001ffffff000001,
-    0x0001ffffff000100, 0x0001ffffff010000, 0x0001ffff00ff00ff, 0x0001ffff0000ffff,
-    0x0001ffff00000000, 0x0001ffff00000001, 0x0001ffff000001ff, 0x0001ffff00000101,
-    0x0001ffff0001ff00, 0x0001ffff000100ff, 0x0001ffff00010001, 0x0001ffff00010100,
-    0x0001ffff01ffff00, 0x0001ffff01000001, 0x0001ffff01010000, 0x0001ff00ffffff00,
-    0x0001ff00ffff00ff, 0x0001ff00ffff0001, 0x0001ff00ffff0100, 0x0001ff00ff00ff01,
-    0x0001ff00ff000000, 0x0001ff00ff01ff00, 0x0001ff00ff01ff01, 0x0001ff00ff010001,
-    0x0001ff00ff010100, 0x0001ff0000ff0000, 0x0001ff0000ff0100, 0x0001ff000000ff00,
-    0x0001ff0000000000, 0x0001ff0000000001, 0x0001ff0000000100, 0x0001ff0000010000,
-    0x0001ff0000010001, 0x0001ff0000010101, 0x0001ff0001ff00ff, 0x0001ff0001ff0101,
-    0x0001ff000100ff01, 0x0001ff0001000000, 0x0001ff000101ff00, 0x0001ff0001010001,
-    0x0001ff0001010100, 0x0001ff01ff00ff00, 0x0001ff01ff000001, 0x0001ff01ff000100,
-    0x0001ff0100ffffff, 0x0001ff0100ffff00, 0x0001ff0100ff0001, 0x0001ff0100000000,
-    0x0001ff0100000001, 0x0001ff01000001ff, 0x0001ff010001ffff, 0x0001ff0101ff0000,
-    0x0001ff010100ff00, 0x0001ff0101000001, 0x0001ff0101010000, 0x000100ffff00ff00,
-    0x000100ffff00ff01, 0x000100ffff000000, 0x000100ffff000001, 0x000100ffff000101,
-    0x000100ffff01ff00, 0x000100ffff010001, 0x000100ffff010100, 0x000100ff00ffffff,
-    0x000100ff00ffff01, 0x000100ff00ff0000, 0x000100ff00ff01ff, 0x000100ff00ff0101,
-    0x000100ff0000ff00, 0x000100ff000000ff, 0x000100ff00000000, 0x000100ff00000001,
-    0x000100ff00000100, 0x000100ff00000101, 0x000100ff0001ffff, 0x000100ff0001ff01,
-    0x000100ff00010000, 0x000100ff01ff00ff, 0x000100ff01ff0000, 0x000100ff01ff0100,
-    0x000100ff0100ffff, 0x000100ff0100ff01, 0x000100ff010000ff, 0x000100ff01000000,
-    0x000100ff01000001, 0x000100ff010001ff, 0x000100ff01000101, 0x000100ff0101ff00,
-    0x000100ff010100ff, 0x000100ff01010100, 0x00010000ffff0000, 0x00010000ffff01ff,
-    0x00010000ffff0101, 0x00010000ff00ff00, 0x00010000ff000000, 0x00010000ff000001,
-    0x00010000ff000100, 0x0001000000ff00ff, 0x0001000000ff0000, 0x0001000000ff0001,
-    0x0001000000ff0100, 0x000100000000ffff, 0x000100000000ff00, 0x00010000000000ff,
-    0x0001000000000000, 0x0001000000000001, 0x0001000000000100, 0x000100000001ff00,
-    0x00010000000100ff, 0x0001000000010000, 0x0001000000010001, 0x0001000000010100,
-    0x0001000001ff0001, 0x0001000001ff0100, 0x0001000001ff0101, 0x000100000100ff00,
-    0x0001000001000000, 0x0001000001000001, 0x0001000001000100, 0x0001000001000101,
-    0x000100000101ff01, 0x0001000001010000, 0x0001000001010001, 0x00010000010101ff,
-    0x00010001ffffff01, 0x00010001ffff0100, 0x00010001ff000000, 0x00010001ff01ffff,
-    0x00010001ff010001, 0x00010001ff0101ff, 0x00010001ff010100, 0x0001000100ffffff,
-    0x0001000100ff0000, 0x0001000100ff01ff, 0x0001000100ff0101, 0x000100010000ff00,
-    0x00010001000000ff, 0x0001000100000000, 0x0001000100000001, 0x00010001000001ff,
-    0x0001000100000101, 0x000100010001ffff, 0x0001000100010000, 0x00010001000101ff,
-    0x0001000101ffffff, 0x0001000101ffff01, 0x0001000101ff0000, 0x0001000101ff0101,
-    0x00010001010000ff, 0x0001000101000001, 0x00010001010001ff, 0x0001000101000100,
-    0x000100010101ffff, 0x00010001010100ff, 0x0001000101010001, 0x0001000101010101,
-    0x000101ffff000001, 0x000101ffff000100, 0x000101ffff010000, 0x000101ff00ffff00,
-    0x000101ff0000ff01, 0x000101ff00000000, 0x000101ff00000101, 0x000101ff0001ff00,
-    0x000101ff00010100, 0x000101ff01ff0000, 0x000101ff0100ff00, 0x000101ff010001ff,
-    0x000101ff01010001, 0x00010100ffffff00, 0x00010100ffff00ff, 0x00010100ff00ffff,
-    0x00010100ff000000, 0x00010100ff01ff00, 0x00010100ff0100ff, 0x00010100ff010001,
-    0x00010100ff010100, 0x0001010000ffffff, 0x0001010000ffff00, 0x0001010000ff0000,
-    0x0001010000ff0001, 0x0001010000ff01ff, 0x000101000000ff00, 0x00010100000000ff,
-    0x0001010000000000, 0x0001010000000001, 0x0001010000000100, 0x000101000001ffff,
-    0x0001010000010000, 0x0001010000010101, 0x0001010001ffff01, 0x0001010001ff00ff,
-    0x0001010001ff0101, 0x0001010001000000, 0x000101000101ff00, 0x00010100010100ff,
-    0x0001010001010000, 0x0001010001010100, 0x00010101ff00ff00, 0x00010101ff000001,
-    0x00010101ff0001ff, 0x0001010100ffff00, 0x0001010100ff00ff, 0x0001010100ff0100,
-    0x000101010000ffff, 0x0001010100000000, 0x00010101000001ff, 0x0001010100000101,
-    0x00010101000100ff, 0x0001010100010000, 0x0001010100010100, 0x0001010101ff0001,
-    0x00010101010000ff, 0x00010101010001ff, 0x0001010101000101, 0x0001010101010001,
-    0x01ffffffffffffff, 0x01ffffffffffff01, 0x01ffffffffff01ff, 0x01ffffffffff0101,
-    0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, 0x01ffffffff010101,
-    0x01ffffff00ff0000, 0x01ffffff0000ffff, 0x01ffffff0000ff00, 0x01ffffff000000ff,
-    0x01ffffff00000001, 0x01ffffff00000100, 0x01ffffff00010000, 0x01ffffff01ffffff,
-    0x01ffffff01ffff01, 0x01ffffff01ff01ff, 0x01ffffff01ff0101, 0x01ffffff01000000,
-    0x01ffffff0101ffff, 0x01ffffff0101ff01, 0x01ffffff010101ff, 0x01ffffff01010101,
-    0x01ffff00ffff0000, 0x01ffff00ff00ff00, 0x01ffff00ff0000ff, 0x01ffff00ff000001,
-    0x01ffff00ff000100, 0x01ffff00ff010000, 0x01ffff0000ffff00, 0x01ffff0000ff00ff,
-    0x01ffff0000ff0100, 0x01ffff000000ffff, 0x01ffff000000ff01, 0x01ffff0000000000,
-    0x01ffff0000000001, 0x01ffff00000001ff, 0x01ffff0000000100, 0x01ffff00000100ff,
-    0x01ffff0000010001, 0x01ffff0000010100, 0x01ffff0001ff0000, 0x01ffff0001ff0100,
-    0x01ffff00010000ff, 0x01ffff0001000001, 0x01ffff0001000100, 0x01ffff0001010000,
-    0x01ffff01ffffffff, 0x01ffff01ffffff01, 0x01ffff01ffff01ff, 0x01ffff01ffff0101,
-    0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff01ff01, 0x01ffff01ff0101ff,
-    0x01ffff01ff010101, 0x01ffff010000ff00, 0x01ffff01000000ff, 0x01ffff0100000100,
-    0x01ffff0100010000, 0x01ffff0101ffffff, 0x01ffff0101ffff01, 0x01ffff0101ff01ff,
-    0x01ffff0101ff0101, 0x01ffff0101000000, 0x01ffff010101ffff, 0x01ffff010101ff01,
-    0x01ffff01010101ff, 0x01ffff0101010101, 0x01ff00ffff0000ff, 0x01ff00ffff000100,
-    0x01ff00ff00ffff00, 0x01ff00ff00ff00ff, 0x01ff00ff0000ff00, 0x01ff00ff00000000,
-    0x01ff00ff00000101, 0x01ff00ff0001ff00, 0x01ff00ff000100ff, 0x01ff00ff00010100,
-    0x01ff00ff010000ff, 0x01ff00ff01000100, 0x01ff0000ffffff00, 0x01ff0000ffff0100,
-    0x01ff0000ff00ff01, 0x01ff0000ff000000, 0x01ff0000ff000101, 0x01ff0000ff010001,
-    0x01ff0000ff010100, 0x01ff000000ffffff, 0x01ff000000ffff00, 0x01ff000000ff0000,
-    0x01ff000000ff01ff, 0x01ff00000000ff00, 0x01ff0000000000ff, 0x01ff000000000000,
-    0x01ff000000000001, 0x01ff000000000100, 0x01ff000000000101, 0x01ff000000010000,
-    0x01ff000000010001, 0x01ff0000000101ff, 0x01ff000000010101, 0x01ff000001ffff00,
-    0x01ff000001ff00ff, 0x01ff000001ff0001, 0x01ff000001ff0100, 0x01ff00000100ffff,
-    0x01ff00000100ff01, 0x01ff000001000000, 0x01ff0000010001ff, 0x01ff000001010001,
-    0x01ff0001ff00ff00, 0x01ff0001ff000001, 0x01ff0001ff000100, 0x01ff0001ff010000,
-    0x01ff000100ffff00, 0x01ff000100ff00ff, 0x01ff000100ff0100, 0x01ff000100ff0101,
-    0x01ff00010000ffff, 0x01ff000100000000, 0x01ff000100000100, 0x01ff000100000101,
-    0x01ff00010001ff00, 0x01ff000100010001, 0x01ff000100010101, 0x01ff000101ff0000,
-    0x01ff00010100ff00, 0x01ff000101000101, 0x01ff0001010100ff, 0x01ff01ffffffffff,
-    0x01ff01ffffffff01, 0x01ff01ffffff01ff, 0x01ff01ffffff0101, 0x01ff01ffff000000,
-    0x01ff01ffff01ffff, 0x01ff01ffff01ff01, 0x01ff01ffff0101ff, 0x01ff01ffff010101,
-    0x01ff01ff00ffff00, 0x01ff01ff00ff0000, 0x01ff01ff0000ff00, 0x01ff01ff000000ff,
-    0x01ff01ff00000100, 0x01ff01ff00010000, 0x01ff01ff00010100, 0x01ff01ff01ffffff,
-    0x01ff01ff01ffff01, 0x01ff01ff01ff01ff, 0x01ff01ff01ff0101, 0x01ff01ff01000000,
-    0x01ff01ff0101ffff, 0x01ff01ff0101ff01, 0x01ff01ff010101ff, 0x01ff01ff01010101,
-    0x01ff0100ffff0000, 0x01ff0100ffff0001, 0x01ff0100ff00ff00, 0x01ff0100ff0000ff,
-    0x01ff0100ff000001, 0x01ff0100ff010000, 0x01ff010000ffff00, 0x01ff010000ff00ff,
-    0x01ff010000ff0001, 0x01ff010000ff0100, 0x01ff01000000ffff, 0x01ff01000000ff01,
-    0x01ff010000000000, 0x01ff010000000101, 0x01ff01000001ff00, 0x01ff0100000100ff,
-    0x01ff010001ff0000, 0x01ff010001000001, 0x01ff010001000100, 0x01ff010001010000,
-    0x01ff0101ffffffff, 0x01ff0101ffffff01, 0x01ff0101ffff01ff, 0x01ff0101ffff0101,
-    0x01ff0101ff000000, 0x01ff0101ff01ffff, 0x01ff0101ff01ff01, 0x01ff0101ff0101ff,
-    0x01ff0101ff010101, 0x01ff010100ff0000, 0x01ff01010000ff00, 0x01ff0101000000ff,
-    0x01ff010100000001, 0x01ff010101ffffff, 0x01ff010101ffff01, 0x01ff010101ff01ff,
-    0x01ff010101ff0101, 0x01ff010101000000, 0x01ff01010101ffff, 0x01ff01010101ff01,
-    0x01ff0101010101ff, 0x01ff010101010101, 0x0100ffffffff0000, 0x0100ffffff00ff00,
-    0x0100ffffff000001, 0x0100ffffff0001ff, 0x0100ffffff000100, 0x0100ffffff010000,
-    0x0100ffff00ffff00, 0x0100ffff00ff0001, 0x0100ffff00ff0100, 0x0100ffff00000000,
-    0x0100ffff000001ff, 0x0100ffff00000101, 0x0100ffff00010100, 0x0100ffff00010101,
-    0x0100ffff01ff0000, 0x0100ffff0100ff00, 0x0100ffff010000ff, 0x0100ffff01000001,
-    0x0100ffff01000100, 0x0100ffff01010000, 0x0100ff00ffffff00, 0x0100ff00ffff00ff,
-    0x0100ff00ffff0001, 0x0100ff00ffff0100, 0x0100ff00ff00ffff, 0x0100ff00ff000000,
-    0x0100ff00ff0001ff, 0x0100ff00ff000101, 0x0100ff00ff01ff00, 0x0100ff00ff0100ff,
-    0x0100ff00ff010001, 0x0100ff00ff010100, 0x0100ff0000ffffff, 0x0100ff0000ff0000,
-    0x0100ff000000ffff, 0x0100ff000000ff00, 0x0100ff00000000ff, 0x0100ff0000000000,
-    0x0100ff0000000001, 0x0100ff0000000100, 0x0100ff000001ff01, 0x0100ff0000010000,
-    0x0100ff0001ff00ff, 0x0100ff0001ff0001, 0x0100ff000100ff01, 0x0100ff0001000000,
-    0x0100ff00010001ff, 0x0100ff000101ff00, 0x0100ff00010100ff, 0x0100ff0001010001,
-    0x0100ff0001010100, 0x0100ff01ffff0000, 0x0100ff01ff00ff00, 0x0100ff01ff0000ff,
-    0x0100ff01ff000100, 0x0100ff01ff010000, 0x0100ff0100ff00ff, 0x0100ff0100ff0001,
-    0x0100ff0100ff0100, 0x0100ff010000ffff, 0x0100ff010000ff01, 0x0100ff0100000000,
-    0x0100ff01000001ff, 0x0100ff0100010001, 0x0100ff0100010100, 0x0100ff0101ff0000,
-    0x0100ff01010000ff, 0x0100ff0101000001, 0x0100ff0101010100, 0x010000ffffffff00,
-    0x010000ffffff00ff, 0x010000ffffff0001, 0x010000ffff00ffff, 0x010000ffff000000,
-    0x010000ffff0001ff, 0x010000ffff010001, 0x010000ff00ffffff, 0x010000ff00ff0101,
-    0x010000ff0000ff00, 0x010000ff000000ff, 0x010000ff00000000, 0x010000ff00000001,
-    0x010000ff000001ff, 0x010000ff00000100, 0x010000ff0001ffff, 0x010000ff0001ff00,
-    0x010000ff0001ff01, 0x010000ff00010000, 0x010000ff01ff00ff, 0x010000ff01ff0001,
-    0x010000ff0100ff01, 0x010000ff010000ff, 0x010000ff01000000, 0x010000ff010001ff,
-    0x010000ff0101ff00, 0x010000ff01010100, 0x01000000ffffffff, 0x01000000ffff0000,
-    0x01000000ffff01ff, 0x01000000ffff0101, 0x01000000ff00ffff, 0x01000000ff00ff00,
-    0x01000000ff0000ff, 0x01000000ff000000, 0x01000000ff000001, 0x01000000ff000100,
-    0x01000000ff01ff00, 0x01000000ff010000, 0x01000000ff010100, 0x01000000ff010101,
-    0x0100000000ffff00, 0x0100000000ff00ff, 0x0100000000ff0000, 0x0100000000ff0001,
-    0x0100000000ff0100, 0x010000000000ffff, 0x010000000000ff00, 0x010000000000ff01,
-    0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, 0x01000000000001ff,
-    0x0100000000000100, 0x0100000000000101, 0x010000000001ff00, 0x01000000000100ff,
-    0x0100000000010000, 0x0100000000010001, 0x0100000000010100, 0x0100000001ffff00,
-    0x0100000001ff0000, 0x0100000001ff01ff, 0x010000000100ff00, 0x010000000100ff01,
-    0x01000000010000ff, 0x0100000001000000, 0x0100000001000001, 0x0100000001000100,
-    0x0100000001000101, 0x010000000101ffff, 0x010000000101ff01, 0x0100000001010000,
-    0x01000000010101ff, 0x0100000001010101, 0x01000001ffffff00, 0x01000001ffff00ff,
-    0x01000001ff00ffff, 0x01000001ff000000, 0x01000001ff000100, 0x01000001ff01ffff,
-    0x01000001ff010001, 0x01000001ff010100, 0x0100000100ff0000, 0x0100000100ff01ff,
-    0x0100000100ff0100, 0x010000010000ff00, 0x010000010000ff01, 0x0100000100000000,
-    0x0100000100000001, 0x0100000100000100, 0x0100000100010000, 0x01000001000101ff,
-    0x0100000101ffff01, 0x0100000101ff00ff, 0x0100000101ff0100, 0x0100000101ff0101,
-    0x010000010100ff01, 0x01000001010000ff, 0x0100000101000000, 0x01000001010100ff,
-    0x0100000101010001, 0x0100000101010100, 0x010001ffffff0000, 0x010001ffff000001,
-    0x010001ffff000100, 0x010001ffff010000, 0x010001ff00ffff00, 0x010001ff00ff0001,
-    0x010001ff0000ffff, 0x010001ff0000ff01, 0x010001ff00000000, 0x010001ff00000001,
-    0x010001ff00000101, 0x010001ff000100ff, 0x010001ff00010000, 0x010001ff01ff0000,
-    0x010001ff0100ff00, 0x010001ff01000001, 0x010001ff01000100, 0x010001ff01010000,
-    0x01000100ffff00ff, 0x01000100ffff0001, 0x01000100ffff0100, 0x01000100ff00ffff,
-    0x01000100ff00ff01, 0x01000100ff000000, 0x01000100ff0001ff, 0x01000100ff000101,
-    0x01000100ff01ffff, 0x01000100ff01ff00, 0x01000100ff0100ff, 0x01000100ff010001,
-    0x0100010000ffffff, 0x0100010000ffff01, 0x0100010000ff0000, 0x0100010000ff01ff,
-    0x0100010000ff0101, 0x010001000000ff00, 0x01000100000000ff, 0x0100010000000000,
-    0x0100010000000001, 0x0100010000000100, 0x010001000001ff01, 0x0100010000010000,
-    0x0100010000010001, 0x0100010000010101, 0x0100010001ffff00, 0x0100010001ff00ff,
-    0x010001000100ffff, 0x010001000100ff01, 0x0100010001000000, 0x0100010001000101,
-    0x010001000101ff00, 0x0100010001010001, 0x01000101ffff0000, 0x01000101ff000000,
-    0x01000101ff010000, 0x0100010100ff00ff, 0x0100010100ff0001, 0x0100010100ff0100,
-    0x010001010000ffff, 0x0100010100000000, 0x01000101000001ff, 0x010001010001ff00,
-    0x0100010101ff0000, 0x010001010100ff00, 0x01000101010000ff, 0x0100010101000000,
-    0x0100010101000001, 0x0101ffffffffffff, 0x0101ffffffffff01, 0x0101ffffffff01ff,
-    0x0101ffffffff0101, 0x0101ffffff000000, 0x0101ffffff01ffff, 0x0101ffffff01ff01,
-    0x0101ffffff0101ff, 0x0101ffffff010101, 0x0101ffff00ff0000, 0x0101ffff0000ff00,
-    0x0101ffff000000ff, 0x0101ffff00000001, 0x0101ffff00000100, 0x0101ffff01ffffff,
-    0x0101ffff01ffff01, 0x0101ffff01ff01ff, 0x0101ffff01ff0101, 0x0101ffff01000000,
-    0x0101ffff0101ffff, 0x0101ffff0101ff01, 0x0101ffff010101ff, 0x0101ffff01010101,
-    0x0101ff00ffff0000, 0x0101ff00ffff0100, 0x0101ff00ff00ff00, 0x0101ff00ff0000ff,
-    0x0101ff00ff000001, 0x0101ff00ff000100, 0x0101ff00ff000101, 0x0101ff0000ff0001,
-    0x0101ff0000ff0100, 0x0101ff000000ff00, 0x0101ff0000000000, 0x0101ff00000001ff,
-    0x0101ff0000000101, 0x0101ff000001ff00, 0x0101ff00000100ff, 0x0101ff0001ff0000,
-    0x0101ff000100ffff, 0x0101ff000100ff01, 0x0101ff0001000001, 0x0101ff0001000100,
-    0x0101ff01ffffff01, 0x0101ff01ffff01ff, 0x0101ff01ffff0101, 0x0101ff01ff00ffff,
-    0x0101ff01ff000100, 0x0101ff01ff01ff01, 0x0101ff01ff0101ff, 0x0101ff01ff010101,
-    0x0101ff0100ff0000, 0x0101ff010000ff00, 0x0101ff0100000001, 0x0101ff0100000100,
-    0x0101ff0100010000, 0x0101ff0101ffffff, 0x0101ff0101ffff01, 0x0101ff0101ff01ff,
-    0x0101ff0101ff0101, 0x0101ff0101000000, 0x0101ff010101ffff, 0x0101ff010101ff01,
-    0x0101ff01010101ff, 0x0101ff0101010101, 0x010100ffff000100, 0x010100ffff010000,
-    0x010100ff00ffff00, 0x010100ff00ff00ff, 0x010100ff0000ffff, 0x010100ff000000ff,
-    0x010100ff00000000, 0x010100ff000001ff, 0x010100ff00000101, 0x010100ff0001ff00,
-    0x010100ff00010000, 0x010100ff00010001, 0x010100ff000101ff, 0x010100ff00010100,
-    0x010100ff01ff0000, 0x01010000ffff0001, 0x01010000ffff0100, 0x01010000ff00ffff,
-    0x01010000ff00ff01, 0x01010000ff000000, 0x01010000ff0001ff, 0x01010000ff010001,
-    0x01010000ff010100, 0x0101000000ffff01, 0x0101000000ff0000, 0x010100000000ff00,
-    0x01010000000000ff, 0x0101000000000000, 0x0101000000000001, 0x0101000000000100,
-    0x0101000000010000, 0x0101000000010101, 0x0101000001ffff00, 0x0101000001ff00ff,
-    0x0101000001ff0000, 0x0101000001ff0001, 0x0101000001ff0100, 0x010100000100ff01,
-    0x0101000001000000, 0x01010000010001ff, 0x01010001ffff0000, 0x01010001ff00ff00,
-    0x01010001ff000001, 0x01010001ff000101, 0x01010001ff01ff00, 0x01010001ff010000,
-    0x0101000100ff00ff, 0x0101000100ff0001, 0x0101000100ff0101, 0x010100010000ff01,
-    0x0101000100000000, 0x0101000100000001, 0x01010001000001ff, 0x010100010001ffff,
-    0x010100010001ff01, 0x0101000101ff0001, 0x010100010100ffff, 0x0101000101000000,
-    0x0101000101000001, 0x0101000101000100, 0x010100010101ff00, 0x01010001010100ff,
-    0x0101000101010001, 0x010101ffffffffff, 0x010101ffffffff01, 0x010101ffffff01ff,
-    0x010101ffffff0101, 0x010101ffff01ffff, 0x010101ffff01ff01, 0x010101ffff0101ff,
-    0x010101ffff010101, 0x010101ff0000ff00, 0x010101ff000000ff, 0x010101ff00000001,
-    0x010101ff00000100, 0x010101ff01ffffff, 0x010101ff01ffff01, 0x010101ff01ff01ff,
-    0x010101ff01ff0101, 0x010101ff01000000, 0x010101ff0101ffff, 0x010101ff0101ff01,
-    0x010101ff010101ff, 0x010101ff01010101, 0x01010100ffff0000, 0x01010100ff0000ff,
-    0x01010100ff000100, 0x01010100ff01ff00, 0x01010100ff010000, 0x0101010000ffff00,
-    0x010101000000ffff, 0x0101010000000000, 0x0101010000000101, 0x010101000001ff00,
-    0x0101010000010001, 0x0101010000010100, 0x010101000100ffff, 0x0101010001000001,
-    0x01010101ffffffff, 0x01010101ffffff01, 0x01010101ffff01ff, 0x01010101ffff0101,
-    0x01010101ff01ffff, 0x01010101ff01ff01, 0x01010101ff0101ff, 0x01010101ff010101,
-    0x010101010000ff00, 0x01010101000000ff, 0x0101010100000001, 0x0101010101ffffff,
-    0x0101010101ffff01, 0x0101010101ff01ff, 0x0101010101ff0101, 0x0101010101000000,
-    0x010101010101ffff, 0x010101010101ff01, 0x01010101010101ff, 0x0101010101010101,
-GGML_TABLE_END()
-#else
-GGML_TABLE_BEGIN(uint32_t, iq1s_grid_gpu, NGRID_IQ1S)
-    0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
-    0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
-    0x02000000, 0x02000002, 0x02000200, 0x02000202, 0x02010101, 0x02020000, 0x02020002, 0x02020200,
-    0x02020202, 0x00000110, 0x00000111, 0x00010011, 0x00010110, 0x00010112, 0x00010211, 0x00010212,
-    0x00020111, 0x01000011, 0x01000112, 0x01000211, 0x01010012, 0x01010111, 0x01010212, 0x01020011,
-    0x01020110, 0x01020112, 0x01020210, 0x02000111, 0x02010011, 0x02010110, 0x02010112, 0x02020111,
-    0x00000020, 0x00000022, 0x00000220, 0x00000222, 0x00010121, 0x00020020, 0x00020022, 0x00020220,
-    0x00020222, 0x01000121, 0x01010021, 0x01010221, 0x01020120, 0x01020221, 0x02000020, 0x02000022,
-    0x02000220, 0x02000222, 0x02010021, 0x02010121, 0x02010221, 0x02020020, 0x02020022, 0x02020220,
-    0x02020222, 0x00011001, 0x00011100, 0x00011102, 0x00021101, 0x01001001, 0x01001201, 0x01011101,
-    0x01011202, 0x01021100, 0x01021101, 0x02011001, 0x02011201, 0x02021101, 0x00001011, 0x00001110,
-    0x00001111, 0x00001112, 0x00011111, 0x00011210, 0x00011212, 0x00021211, 0x01001010, 0x01001111,
-    0x01001212, 0x01011010, 0x01011011, 0x01011110, 0x01011111, 0x01011112, 0x01011211, 0x01021010,
-    0x01021012, 0x01021111, 0x01021210, 0x01021212, 0x02001011, 0x02011011, 0x02011111, 0x02011210,
-    0x02011212, 0x02021011, 0x02021110, 0x02021111, 0x02021112, 0x02021211, 0x00011120, 0x00011221,
-    0x01001021, 0x01001120, 0x01011020, 0x01011022, 0x01011121, 0x01011220, 0x01021020, 0x01021021,
-    0x01021122, 0x01021221, 0x02001121, 0x02011021, 0x02011120, 0x02011221, 0x00002000, 0x00002002,
-    0x00002200, 0x00002202, 0x00012101, 0x00022000, 0x00022002, 0x00022200, 0x00022202, 0x01002101,
-    0x01012001, 0x01012102, 0x01022101, 0x02002000, 0x02002002, 0x02002200, 0x02002202, 0x02012101,
-    0x02022000, 0x02022002, 0x02022200, 0x02022202, 0x00002111, 0x00012011, 0x00012110, 0x00012211,
-    0x00022110, 0x00022111, 0x01002011, 0x01012010, 0x01012011, 0x01012111, 0x01022011, 0x01022110,
-    0x01022211, 0x02012011, 0x02012110, 0x02012112, 0x02012211, 0x02022111, 0x00002020, 0x00002022,
-    0x00002220, 0x00002222, 0x00012121, 0x00022020, 0x00022022, 0x00022220, 0x00022222, 0x01002121,
-    0x01012021, 0x01012221, 0x01022021, 0x01022121, 0x02002020, 0x02002022, 0x02002121, 0x02002220,
-    0x02002222, 0x02012121, 0x02022020, 0x02022022, 0x02022220, 0x02022222, 0x00110000, 0x00110001,
-    0x00110100, 0x00110201, 0x00120100, 0x00120101, 0x01100001, 0x01100100, 0x01110000, 0x01110101,
-    0x01110200, 0x01120001, 0x01120100, 0x01120101, 0x01120201, 0x02110001, 0x02110100, 0x02110102,
-    0x02120001, 0x02120101, 0x00100011, 0x00100110, 0x00100112, 0x00100211, 0x00110010, 0x00110012,
-    0x00110111, 0x00110210, 0x00120011, 0x00120110, 0x00120211, 0x01100111, 0x01100212, 0x01110010,
-    0x01110011, 0x01110012, 0x01110110, 0x01110111, 0x01110112, 0x01110211, 0x01120010, 0x01120111,
-    0x02100110, 0x02110012, 0x02110111, 0x02120011, 0x02120110, 0x00110021, 0x00110120, 0x00110122,
-    0x00120121, 0x01100020, 0x01100122, 0x01100221, 0x01110022, 0x01110121, 0x01110220, 0x01110222,
-    0x01120120, 0x01120122, 0x02100121, 0x02110021, 0x02110120, 0x02110122, 0x02120121, 0x00101001,
-    0x00101102, 0x00101201, 0x00111100, 0x00111101, 0x00111200, 0x00111201, 0x00121001, 0x00121102,
-    0x01101001, 0x01101101, 0x01101102, 0x01101200, 0x01101202, 0x01111001, 0x01111100, 0x01111101,
-    0x01111102, 0x01111201, 0x01121002, 0x01121101, 0x01121200, 0x02101100, 0x02101201, 0x02111000,
-    0x02111100, 0x02111101, 0x02111200, 0x02111201, 0x02111202, 0x02121001, 0x02121100, 0x02121101,
-    0x02121201, 0x00101012, 0x00101111, 0x00101212, 0x00111011, 0x00111110, 0x00111111, 0x00111112,
-    0x00111211, 0x00121010, 0x00121012, 0x00121111, 0x00121210, 0x00121212, 0x01101011, 0x01101110,
-    0x01101111, 0x01101112, 0x01111011, 0x01111012, 0x01111110, 0x01111111, 0x01111112, 0x01111211,
-    0x01111212, 0x01121011, 0x01121110, 0x01121111, 0x01121112, 0x01121211, 0x02101010, 0x02101012,
-    0x02101110, 0x02101111, 0x02101210, 0x02101212, 0x02111010, 0x02111011, 0x02111110, 0x02111111,
-    0x02111112, 0x02111211, 0x02111212, 0x02121010, 0x02121012, 0x02121111, 0x00101021, 0x00101120,
-    0x00101121, 0x00101122, 0x00111121, 0x00111122, 0x00111220, 0x00111222, 0x00121021, 0x00121122,
-    0x01101020, 0x01101022, 0x01101120, 0x01101121, 0x01101220, 0x01101222, 0x01111021, 0x01111121,
-    0x01111122, 0x01111220, 0x01111221, 0x01121021, 0x01121120, 0x01121121, 0x01121220, 0x01121221,
-    0x01121222, 0x02101122, 0x02101222, 0x02111022, 0x02111121, 0x02121120, 0x02121221, 0x00112001,
-    0x00112102, 0x00122101, 0x01102001, 0x01102100, 0x01102102, 0x01102201, 0x01112000, 0x01112101,
-    0x01112200, 0x01112202, 0x01122000, 0x01122001, 0x01122100, 0x01122102, 0x01122201, 0x02102101,
-    0x02112001, 0x02112100, 0x02122101, 0x00112010, 0x00112012, 0x00112111, 0x00112212, 0x00122011,
-    0x00122111, 0x01102012, 0x01102110, 0x01102111, 0x01102210, 0x01112011, 0x01112110, 0x01112111,
-    0x01112112, 0x01112211, 0x01112212, 0x01122010, 0x01122111, 0x01122212, 0x02102211, 0x02112011,
-    0x02112012, 0x02112111, 0x02112210, 0x02122011, 0x02122112, 0x02122211, 0x00102221, 0x00112122,
-    0x00122120, 0x00122122, 0x01102120, 0x01102122, 0x01102221, 0x01112020, 0x01112022, 0x01112121,
-    0x01112220, 0x01122021, 0x01122122, 0x01122221, 0x02102121, 0x02112021, 0x02112122, 0x02112222,
-    0x00200000, 0x00200002, 0x00200200, 0x00200202, 0x00210101, 0x00220000, 0x00220002, 0x00220101,
-    0x00220200, 0x00220202, 0x01200101, 0x01210001, 0x01210201, 0x01220001, 0x01220101, 0x02200000,
-    0x02200002, 0x02200200, 0x02200202, 0x02210101, 0x02220000, 0x02220002, 0x02220101, 0x02220200,
-    0x02220202, 0x00200111, 0x00210011, 0x00210110, 0x00210211, 0x00220111, 0x01200012, 0x01200110,
-    0x01200211, 0x01210111, 0x01210210, 0x01210212, 0x01220011, 0x01220110, 0x01220111, 0x01220112,
-    0x02200111, 0x02210010, 0x02210112, 0x02210211, 0x02220111, 0x00200021, 0x00200220, 0x00200222,
-    0x00210021, 0x00210121, 0x00220020, 0x00220022, 0x00220220, 0x00220222, 0x01200121, 0x01210021,
-    0x01210122, 0x01210221, 0x01220121, 0x02200021, 0x02200220, 0x02200222, 0x02210021, 0x02210121,
-    0x02220020, 0x02220022, 0x02220220, 0x02220222, 0x00201101, 0x00211100, 0x00211102, 0x00211201,
-    0x00221101, 0x01201100, 0x01201101, 0x01201102, 0x01201201, 0x01211002, 0x01211101, 0x01211200,
-    0x01211202, 0x01221102, 0x02201101, 0x02211001, 0x02211100, 0x02211201, 0x02221001, 0x02221101,
-    0x00201211, 0x00211111, 0x00221011, 0x00221211, 0x01201010, 0x01201111, 0x01201210, 0x01211011,
-    0x01211110, 0x01211111, 0x01211211, 0x01221012, 0x01221111, 0x01221210, 0x02201211, 0x02211010,
-    0x02211110, 0x02211111, 0x02211210, 0x02211212, 0x02221011, 0x02221110, 0x02221112, 0x02221211,
-    0x00201121, 0x00211020, 0x00211022, 0x00211221, 0x00221121, 0x01201021, 0x01201221, 0x01211121,
-    0x01221020, 0x01221021, 0x01221221, 0x02201120, 0x02201122, 0x02211020, 0x02211222, 0x00202000,
-    0x00202002, 0x00202200, 0x00202202, 0x00212101, 0x00222000, 0x00222002, 0x00222200, 0x00222202,
-    0x01202101, 0x01212001, 0x01212100, 0x01222101, 0x02202000, 0x02202002, 0x02202200, 0x02202202,
-    0x02222000, 0x02222002, 0x02222200, 0x02222202, 0x00202211, 0x00212011, 0x00212110, 0x00212211,
-    0x00222111, 0x01202112, 0x01202211, 0x01212012, 0x01212111, 0x01222011, 0x01222110, 0x01222112,
-    0x01222211, 0x02202111, 0x02212010, 0x02212112, 0x02212211, 0x02222110, 0x02222111, 0x00202020,
-    0x00202022, 0x00202220, 0x00202222, 0x00222020, 0x00222022, 0x00222220, 0x00222222, 0x01202121,
-    0x01212021, 0x01212122, 0x01212221, 0x01222121, 0x02202020, 0x02202022, 0x02202220, 0x02202222,
-    0x02212121, 0x02222020, 0x02222022, 0x02222220, 0x02222222, 0x10000101, 0x10010001, 0x10010102,
-    0x10020101, 0x11000201, 0x11010002, 0x11010101, 0x11010200, 0x11010202, 0x11020001, 0x11020100,
-    0x11020102, 0x12010100, 0x12010201, 0x12020001, 0x12020102, 0x10000010, 0x10000011, 0x10000110,
-    0x10000112, 0x10000211, 0x10010012, 0x10010111, 0x10010112, 0x10010210, 0x10010212, 0x10020011,
-    0x10020112, 0x10020211, 0x11000111, 0x11000210, 0x11000212, 0x11010011, 0x11010110, 0x11010111,
-    0x11010112, 0x11010211, 0x11010212, 0x11020111, 0x11020210, 0x11020212, 0x12000011, 0x12000110,
-    0x12000112, 0x12010010, 0x12010012, 0x12010111, 0x12020010, 0x12020011, 0x12020012, 0x10000121,
-    0x10010021, 0x10010120, 0x10010122, 0x10020121, 0x11000021, 0x11010022, 0x11010121, 0x11010222,
-    0x11020120, 0x11020221, 0x12000221, 0x12010120, 0x12020121, 0x10001001, 0x10011101, 0x10011201,
-    0x10021201, 0x11001101, 0x11001200, 0x11001202, 0x11011001, 0x11011100, 0x11011101, 0x11011102,
-    0x11021001, 0x11021002, 0x11021101, 0x11021200, 0x11021202, 0x12001001, 0x12001102, 0x12001201,
-    0x12011000, 0x12011002, 0x12011101, 0x12021000, 0x12021001, 0x12021201, 0x10001011, 0x10001012,
-    0x10001111, 0x10001212, 0x10011011, 0x10011110, 0x10011111, 0x10011112, 0x10011211, 0x10021010,
-    0x10021111, 0x10021212, 0x11001011, 0x11001110, 0x11001111, 0x11001112, 0x11001211, 0x11011010,
-    0x11011011, 0x11011110, 0x11011111, 0x11011112, 0x11011210, 0x11011211, 0x11021011, 0x11021110,
-    0x11021111, 0x11021112, 0x11021211, 0x12001012, 0x12001110, 0x12001111, 0x12001210, 0x12011011,
-    0x12011110, 0x12011111, 0x12011112, 0x12011211, 0x12011212, 0x12021111, 0x12021210, 0x12021212,
-    0x10001021, 0x10001121, 0x10001221, 0x10011120, 0x10011121, 0x10011220, 0x10011222, 0x10021021,
-    0x10021120, 0x10021221, 0x11001020, 0x11001022, 0x11001121, 0x11001220, 0x11011020, 0x11011021,
-    0x11011022, 0x11011121, 0x11011122, 0x11011221, 0x11021022, 0x11021121, 0x11021220, 0x12001021,
-    0x12001121, 0x12001222, 0x12011120, 0x12011121, 0x12021021, 0x12021120, 0x12021122, 0x10002101,
-    0x10012001, 0x10012101, 0x10012202, 0x10022101, 0x11002002, 0x11002201, 0x11012000, 0x11012101,
-    0x11012200, 0x11022001, 0x11022100, 0x11022102, 0x11022201, 0x12002101, 0x12012001, 0x12012100,
-    0x12012102, 0x12012201, 0x12022101, 0x10002011, 0x10002111, 0x10002112, 0x10002212, 0x10012010,
-    0x10012110, 0x10012111, 0x10012210, 0x10022011, 0x10022110, 0x10022112, 0x11002010, 0x11002111,
-    0x11002212, 0x11012011, 0x11012012, 0x11012110, 0x11012111, 0x11012112, 0x11012211, 0x11022010,
-    0x11022012, 0x11022111, 0x11022112, 0x11022212, 0x12002112, 0x12002211, 0x12012012, 0x12012111,
-    0x12012112, 0x12012210, 0x12022011, 0x12022110, 0x12022112, 0x12022211, 0x10012122, 0x11002120,
-    0x11002122, 0x11002221, 0x11012121, 0x11012220, 0x11012222, 0x11022120, 0x11022221, 0x12012120,
-    0x12022121, 0x10100001, 0x10100100, 0x10100101, 0x10100102, 0x10100201, 0x10110002, 0x10110101,
-    0x10110202, 0x10120001, 0x10120100, 0x10120201, 0x11100000, 0x11100101, 0x11100200, 0x11110001,
-    0x11110100, 0x11110101, 0x11110102, 0x11110201, 0x11120101, 0x11120200, 0x12100102, 0x12100201,
-    0x12110101, 0x12110200, 0x12120000, 0x12120001, 0x12120102, 0x12120201, 0x10100111, 0x10100210,
-    0x10100211, 0x10100212, 0x10110011, 0x10110110, 0x10110111, 0x10110112, 0x10110210, 0x10110211,
-    0x10120010, 0x10120111, 0x10120112, 0x10120210, 0x10120212, 0x11100011, 0x11100110, 0x11100111,
-    0x11100112, 0x11100211, 0x11110010, 0x11110011, 0x11110012, 0x11110110, 0x11110111, 0x11110112,
-    0x11110210, 0x11110211, 0x11110212, 0x11120011, 0x11120110, 0x11120111, 0x11120112, 0x11120211,
-    0x12100012, 0x12100111, 0x12110011, 0x12110110, 0x12110111, 0x12110112, 0x12110211, 0x12120010,
-    0x12120111, 0x12120212, 0x10100021, 0x10100122, 0x10110022, 0x10110121, 0x10110222, 0x10120021,
-    0x10120120, 0x11100022, 0x11100121, 0x11100222, 0x11110021, 0x11110120, 0x11110121, 0x11110122,
-    0x11110221, 0x11120022, 0x11120121, 0x12100121, 0x12110020, 0x12110022, 0x12110121, 0x12110221,
-    0x12110222, 0x12120120, 0x10101100, 0x10101101, 0x10111001, 0x10111100, 0x10111101, 0x10111102,
-    0x10111200, 0x10111201, 0x10121001, 0x10121101, 0x10121200, 0x10121202, 0x11101001, 0x11101100,
-    0x11101101, 0x11101102, 0x11101201, 0x11101202, 0x11111000, 0x11111001, 0x11111100, 0x11111101,
-    0x11111102, 0x11111200, 0x11111201, 0x11111202, 0x11121001, 0x11121002, 0x11121100, 0x11121101,
-    0x11121102, 0x11121201, 0x12101000, 0x12101200, 0x12101202, 0x12111001, 0x12111100, 0x12111101,
-    0x12111102, 0x12111201, 0x12121001, 0x12121100, 0x12121101, 0x12121202, 0x10101011, 0x10101012,
-    0x10101110, 0x10101111, 0x10101112, 0x10101211, 0x10111010, 0x10111011, 0x10111012, 0x10111110,
-    0x10111111, 0x10111112, 0x10111211, 0x10111212, 0x10121011, 0x10121110, 0x10121111, 0x10121112,
-    0x10121211, 0x11101010, 0x11101011, 0x11101012, 0x11101110, 0x11101111, 0x11101112, 0x11101210,
-    0x11101211, 0x11111010, 0x11111011, 0x11111012, 0x11111110, 0x11111111, 0x11111112, 0x11111210,
-    0x11111211, 0x11111212, 0x11121010, 0x11121011, 0x11121110, 0x11121111, 0x11121112, 0x11121210,
-    0x11121211, 0x11121212, 0x12101011, 0x12101110, 0x12101111, 0x12101211, 0x12101212, 0x12111010,
-    0x12111011, 0x12111110, 0x12111111, 0x12111112, 0x12111210, 0x12111211, 0x12121011, 0x12121110,
-    0x12121111, 0x12121112, 0x12121211, 0x10101020, 0x10101021, 0x10101022, 0x10101120, 0x10101122,
-    0x10101220, 0x10101221, 0x10111021, 0x10111120, 0x10111121, 0x10111220, 0x10111221, 0x10121020,
-    0x10121021, 0x10121022, 0x10121120, 0x10121121, 0x10121122, 0x10121220, 0x10121221, 0x11101021,
-    0x11101121, 0x11101122, 0x11101220, 0x11101221, 0x11101222, 0x11111020, 0x11111021, 0x11111022,
-    0x11111120, 0x11111121, 0x11111122, 0x11111220, 0x11111221, 0x11111222, 0x11121021, 0x11121120,
-    0x11121121, 0x11121221, 0x12101022, 0x12101121, 0x12101122, 0x12101220, 0x12101221, 0x12101222,
-    0x12111021, 0x12111121, 0x12111222, 0x12121022, 0x12121121, 0x12121122, 0x12121220, 0x12121221,
-    0x10102100, 0x10102101, 0x10102102, 0x10102201, 0x10112000, 0x10112101, 0x10112200, 0x10122001,
-    0x10122202, 0x11102101, 0x11102200, 0x11102202, 0x11112001, 0x11112100, 0x11112101, 0x11112102,
-    0x11112200, 0x11112201, 0x11122000, 0x11122002, 0x11122100, 0x11122101, 0x12102002, 0x12102201,
-    0x12112000, 0x12112002, 0x12112101, 0x12112200, 0x12122001, 0x12122201, 0x10102011, 0x10102012,
-    0x10102111, 0x10102212, 0x10112011, 0x10112110, 0x10112111, 0x10112112, 0x10112211, 0x10122111,
-    0x11102011, 0x11102110, 0x11102111, 0x11102112, 0x11102211, 0x11112010, 0x11112011, 0x11112012,
-    0x11112110, 0x11112111, 0x11112112, 0x11112210, 0x11112211, 0x11112212, 0x11122011, 0x11122110,
-    0x11122111, 0x11122112, 0x11122211, 0x12102011, 0x12102111, 0x12102211, 0x12112011, 0x12112110,
-    0x12112111, 0x12112112, 0x12112210, 0x12112211, 0x12122111, 0x10102120, 0x10102220, 0x10112121,
-    0x10112222, 0x10122020, 0x10122121, 0x10122122, 0x10122221, 0x11102121, 0x11102220, 0x11102221,
-    0x11112021, 0x11112121, 0x11112122, 0x11112220, 0x11112221, 0x11122022, 0x11122121, 0x11122220,
-    0x11122222, 0x12102021, 0x12102222, 0x12112022, 0x12112121, 0x12112122, 0x12112220, 0x12112222,
-    0x12122021, 0x10200101, 0x10210100, 0x10210102, 0x10210201, 0x10220101, 0x11200100, 0x11210000,
-    0x11210101, 0x11210102, 0x11210200, 0x11210202, 0x11220001, 0x11220100, 0x11220102, 0x11220201,
-    0x12200001, 0x12210102, 0x12220101, 0x10200011, 0x10200110, 0x10200112, 0x10200211, 0x10210012,
-    0x10210111, 0x10220011, 0x10220012, 0x10220112, 0x10220211, 0x11200111, 0x11200211, 0x11210011,
-    0x11210111, 0x11210112, 0x11210211, 0x11220111, 0x11220112, 0x11220212, 0x12200110, 0x12200212,
-    0x12210012, 0x12210111, 0x12220011, 0x12220112, 0x12220211, 0x10210021, 0x10210122, 0x10210221,
-    0x11200020, 0x11200021, 0x11200122, 0x11210121, 0x11210122, 0x11210220, 0x11220020, 0x12200121,
-    0x12210021, 0x12210122, 0x12220121, 0x10211001, 0x10211002, 0x10211101, 0x10211102, 0x10211202,
-    0x10221001, 0x10221102, 0x10221201, 0x11201000, 0x11201002, 0x11201101, 0x11201200, 0x11201202,
-    0x11211001, 0x11211100, 0x11211101, 0x11211102, 0x11211201, 0x11211202, 0x11221000, 0x11221002,
-    0x11221101, 0x12201100, 0x12201101, 0x12201201, 0x12211000, 0x12211002, 0x12211100, 0x12211101,
-    0x12211102, 0x12211200, 0x12211202, 0x12221001, 0x12221100, 0x12221201, 0x10201111, 0x10201210,
-    0x10201212, 0x10211011, 0x10211111, 0x10211112, 0x10211211, 0x11201110, 0x11201111, 0x11201112,
-    0x11201211, 0x11211010, 0x11211011, 0x11211110, 0x11211111, 0x11211112, 0x11211211, 0x11221011,
-    0x11221110, 0x11221111, 0x11221112, 0x11221211, 0x12201112, 0x12201211, 0x12201212, 0x12211011,
-    0x12211111, 0x12211112, 0x12211211, 0x12211212, 0x12221012, 0x12221111, 0x12221112, 0x12221210,
-    0x10201022, 0x10201221, 0x10211121, 0x10221020, 0x10221122, 0x10221220, 0x10221221, 0x11201020,
-    0x11201121, 0x11201220, 0x11201222, 0x11211021, 0x11211120, 0x11211121, 0x11211122, 0x11211220,
-    0x11211222, 0x11221020, 0x11221121, 0x11221220, 0x12201020, 0x12201022, 0x12201121, 0x12201222,
-    0x12211120, 0x12211122, 0x12211220, 0x12211221, 0x12221020, 0x12221120, 0x12221122, 0x12221222,
-    0x10212102, 0x10212201, 0x10222101, 0x11202001, 0x11212002, 0x11212101, 0x11212202, 0x11222001,
-    0x11222201, 0x12202101, 0x12212001, 0x12212200, 0x12222102, 0x10202011, 0x10202110, 0x10212010,
-    0x10212111, 0x10222011, 0x10222110, 0x10222112, 0x10222211, 0x11202010, 0x11202011, 0x11202111,
-    0x11202112, 0x11202210, 0x11212011, 0x11212110, 0x11212111, 0x11212112, 0x11212211, 0x11222010,
-    0x11222111, 0x11222212, 0x12202012, 0x12202110, 0x12202212, 0x12212111, 0x12222011, 0x12222110,
-    0x12222111, 0x12222211, 0x10212021, 0x10212122, 0x10212220, 0x11202021, 0x11202120, 0x11202221,
-    0x11212020, 0x11212121, 0x11212220, 0x11212222, 0x11222120, 0x11222121, 0x11222221, 0x12202122,
-    0x12212120, 0x12212220, 0x12212222, 0x12222122, 0x20000000, 0x20000002, 0x20000200, 0x20000202,
-    0x20020000, 0x20020002, 0x20020200, 0x20020202, 0x21000101, 0x21010000, 0x21010001, 0x21010100,
-    0x21010102, 0x21010201, 0x21020101, 0x22000000, 0x22000002, 0x22000200, 0x22000202, 0x22010101,
-    0x22020000, 0x22020002, 0x22020200, 0x22020202, 0x20000111, 0x20010011, 0x20010110, 0x20010112,
-    0x20010211, 0x20020111, 0x21000011, 0x21000110, 0x21000211, 0x21010010, 0x21010012, 0x21010111,
-    0x21010112, 0x21010210, 0x21010211, 0x21020110, 0x21020112, 0x21020211, 0x22000111, 0x22000211,
-    0x22010110, 0x22010112, 0x22010211, 0x22020111, 0x20000020, 0x20000022, 0x20000220, 0x20000222,
-    0x20010121, 0x20020020, 0x20020022, 0x20020220, 0x20020222, 0x21010021, 0x21010120, 0x21010221,
-    0x21020121, 0x22000020, 0x22000022, 0x22000220, 0x22000222, 0x22010121, 0x22020020, 0x22020022,
-    0x22020220, 0x22020222, 0x20011100, 0x20011201, 0x21001001, 0x21001100, 0x21011001, 0x21011101,
-    0x21011202, 0x21021001, 0x21021100, 0x21021201, 0x22011100, 0x22011201, 0x20001011, 0x20001211,
-    0x20011012, 0x20011111, 0x20011212, 0x20021112, 0x20021211, 0x21001010, 0x21001011, 0x21001111,
-    0x21001210, 0x21011011, 0x21011110, 0x21011111, 0x21011112, 0x21011211, 0x21011212, 0x21021111,
-    0x21021112, 0x21021210, 0x21021212, 0x22001011, 0x22001110, 0x22001112, 0x22001211, 0x22011010,
-    0x22011012, 0x22011111, 0x22011210, 0x22021112, 0x20011021, 0x20011122, 0x20011221, 0x20021121,
-    0x21001021, 0x21001120, 0x21001221, 0x21001222, 0x21011020, 0x21011121, 0x21011221, 0x21011222,
-    0x21021021, 0x21021122, 0x21021222, 0x22001121, 0x22011021, 0x22011222, 0x22021120, 0x20002000,
-    0x20002002, 0x20002200, 0x20002202, 0x20012101, 0x20022000, 0x20022002, 0x20022200, 0x20022202,
-    0x21002001, 0x21002101, 0x21012001, 0x21012100, 0x21012201, 0x21022101, 0x21022201, 0x22002000,
-    0x22002002, 0x22002200, 0x22002202, 0x22012101, 0x22022000, 0x22022002, 0x22022200, 0x22022202,
-    0x20002111, 0x20002112, 0x20012011, 0x20012110, 0x20012112, 0x20022111, 0x21002011, 0x21002110,
-    0x21002112, 0x21002211, 0x21012010, 0x21012012, 0x21012111, 0x21012212, 0x21022011, 0x21022110,
-    0x22002111, 0x22012112, 0x22012211, 0x22022111, 0x20002020, 0x20002022, 0x20002220, 0x20002222,
-    0x20012121, 0x20022020, 0x20022022, 0x20022220, 0x20022222, 0x21002121, 0x21012021, 0x21012120,
-    0x21012122, 0x22002020, 0x22002022, 0x22002220, 0x22002222, 0x22012121, 0x22022020, 0x22022022,
-    0x22022220, 0x22022222, 0x20100101, 0x20110001, 0x20110102, 0x20110200, 0x20110201, 0x20120101,
-    0x21100001, 0x21100102, 0x21100201, 0x21110101, 0x21110200, 0x21110202, 0x21120201, 0x21120202,
-    0x22100101, 0x22110001, 0x22110100, 0x22110102, 0x22110201, 0x22120101, 0x20100011, 0x20100110,
-    0x20100112, 0x20100211, 0x20110010, 0x20110111, 0x20110210, 0x20110212, 0x20120011, 0x20120110,
-    0x20120112, 0x20120211, 0x21100010, 0x21100111, 0x21110010, 0x21110011, 0x21110110, 0x21110111,
-    0x21110112, 0x21110211, 0x21120012, 0x21120111, 0x22100110, 0x22100112, 0x22110012, 0x22110111,
-    0x22110210, 0x22120011, 0x22120110, 0x22120112, 0x22120211, 0x20100121, 0x20110021, 0x20110120,
-    0x20110221, 0x20120121, 0x21100120, 0x21100122, 0x21100221, 0x21110020, 0x21110022, 0x21110121,
-    0x21110220, 0x21120122, 0x21120221, 0x22100121, 0x22110120, 0x22110122, 0x22120221, 0x20101001,
-    0x20101100, 0x20101102, 0x20111000, 0x20111101, 0x20111200, 0x20121102, 0x21101000, 0x21101202,
-    0x21111001, 0x21111100, 0x21111101, 0x21111102, 0x21111200, 0x21111201, 0x21121000, 0x21121001,
-    0x21121002, 0x21121101, 0x22101100, 0x22101102, 0x22111002, 0x22111100, 0x22111101, 0x22111200,
-    0x22121001, 0x22121201, 0x20101010, 0x20101111, 0x20101210, 0x20101212, 0x20111010, 0x20111011,
-    0x20111110, 0x20111111, 0x20111112, 0x20111211, 0x20121011, 0x20121111, 0x20121211, 0x20121212,
-    0x21101011, 0x21101110, 0x21101111, 0x21101112, 0x21101211, 0x21111010, 0x21111011, 0x21111012,
-    0x21111110, 0x21111111, 0x21111112, 0x21111210, 0x21111211, 0x21111212, 0x21121011, 0x21121110,
-    0x21121111, 0x21121112, 0x21121211, 0x22101011, 0x22101111, 0x22101210, 0x22111011, 0x22111012,
-    0x22111110, 0x22111111, 0x22111112, 0x22111211, 0x22111212, 0x22121010, 0x22121012, 0x22121111,
-    0x22121210, 0x22121212, 0x20101021, 0x20101120, 0x20111020, 0x20111121, 0x20111221, 0x20121020,
-    0x20121122, 0x20121221, 0x21101121, 0x21101220, 0x21101221, 0x21111021, 0x21111022, 0x21111121,
-    0x21111122, 0x21111221, 0x21121121, 0x21121220, 0x22101022, 0x22101120, 0x22101221, 0x22101222,
-    0x22111022, 0x22111120, 0x22111121, 0x22121120, 0x22121122, 0x22121221, 0x20102101, 0x20112102,
-    0x20112201, 0x20122101, 0x21102001, 0x21102102, 0x21112000, 0x21112002, 0x21112101, 0x21112102,
-    0x21112202, 0x21122100, 0x21122101, 0x22102101, 0x22112001, 0x22112102, 0x22112201, 0x22122101,
-    0x20102110, 0x20102112, 0x20102211, 0x20112010, 0x20112012, 0x20112111, 0x20112210, 0x20112212,
-    0x20122010, 0x20122011, 0x20122110, 0x20122112, 0x21102010, 0x21102012, 0x21102111, 0x21102210,
-    0x21102212, 0x21112011, 0x21112110, 0x21112111, 0x21112112, 0x21112211, 0x21122012, 0x21122111,
-    0x21122112, 0x21122212, 0x22102011, 0x22102110, 0x22112010, 0x22112012, 0x22112111, 0x22112212,
-    0x22122011, 0x22122112, 0x20102121, 0x20112121, 0x20122121, 0x21102120, 0x21102122, 0x21102221,
-    0x21112020, 0x21112121, 0x21112220, 0x21122021, 0x22102121, 0x22112021, 0x22112120, 0x22112121,
-    0x22112122, 0x20200000, 0x20200002, 0x20200200, 0x20200202, 0x20210101, 0x20220000, 0x20220002,
-    0x20220200, 0x20220202, 0x21200101, 0x21210001, 0x21210100, 0x21210102, 0x21210201, 0x22200000,
-    0x22200002, 0x22200200, 0x22200202, 0x22210101, 0x22220000, 0x22220002, 0x22220200, 0x22220202,
-    0x20200111, 0x20200211, 0x20210011, 0x20210110, 0x20210112, 0x20210211, 0x20210212, 0x21200112,
-    0x21200211, 0x21210011, 0x21210111, 0x21210210, 0x21210212, 0x21220011, 0x21220110, 0x22200111,
-    0x22210010, 0x22210012, 0x22210112, 0x22210211, 0x20200022, 0x20200220, 0x20200222, 0x20210020,
-    0x20210221, 0x20220022, 0x20220220, 0x20220222, 0x21200121, 0x21210021, 0x21210122, 0x21210221,
-    0x21220121, 0x22200020, 0x22200022, 0x22200220, 0x22200222, 0x22210121, 0x22220020, 0x22220022,
-    0x22220220, 0x22220222, 0x20211201, 0x20221101, 0x21201001, 0x21201100, 0x21211000, 0x21211100,
-    0x21211101, 0x21211200, 0x21211202, 0x21221001, 0x21221101, 0x21221102, 0x21221200, 0x21221201,
-    0x22201101, 0x20201112, 0x20201211, 0x20211010, 0x20211012, 0x20211111, 0x20211210, 0x20221112,
-    0x20221211, 0x21201012, 0x21201111, 0x21211011, 0x21211110, 0x21211111, 0x21211112, 0x21211211,
-    0x21221111, 0x21221212, 0x22201011, 0x22201110, 0x22201111, 0x22201112, 0x22201211, 0x22211012,
-    0x22211111, 0x22211210, 0x20201121, 0x20211021, 0x20211122, 0x20211222, 0x20221021, 0x20221121,
-    0x21201120, 0x21201122, 0x21201222, 0x21211022, 0x21211121, 0x21211122, 0x21211220, 0x21221020,
-    0x21221022, 0x22201122, 0x22211020, 0x22211121, 0x22211122, 0x22211221, 0x22221021, 0x22221120,
-    0x22221122, 0x20202000, 0x20202002, 0x20202200, 0x20202202, 0x20222000, 0x20222002, 0x20222200,
-    0x20222202, 0x21212001, 0x21212100, 0x21212102, 0x21212201, 0x22202000, 0x22202002, 0x22202200,
-    0x22202202, 0x22212101, 0x22222000, 0x22222002, 0x22222200, 0x22222202, 0x20202111, 0x20212110,
-    0x20212211, 0x20222011, 0x20222111, 0x21202011, 0x21212010, 0x21212111, 0x21212212, 0x21222011,
-    0x21222112, 0x21222211, 0x22212010, 0x22212112, 0x20202020, 0x20202022, 0x20202220, 0x20202222,
-    0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
-    0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
-GGML_TABLE_END()
-#endif
-
-#endif // GGML_COMMON_IMPL
-#endif // GGML_COMMON_IMPL
+#ifndef GGML_COMMON_DECL
+
+#if defined(GGML_COMMON_DECL_C)
+#include <stdint.h>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+#define GGML_COMMON_AGGR
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_METAL)
+#include <metal_stdlib>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CUDA)
+#include <cuda_fp16.h>
+#include <cstdint>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_HIP)
+#include <hip/hip_fp16.h>
+#include <cstdint>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_SYCL)
+#include <sycl/half_type.hpp>
+#include <cstdint>
+
+typedef sycl::half  ggml_half;
+typedef sycl::half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#endif
+
+#if defined(GGML_COMMON_DECL)
+
+#ifndef __cplusplus
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+#endif // __cplusplus
+
+// QK = number of values after dequantization
+// QK_K = super-block size
+
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif // GGML_QKK_64
+
+#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
+// QR = QK / number of values before dequantization
+// QI = number of 32 bit integers before dequantization
+
+#define QI4_0 (QK4_0 / (4 * QR4_0))
+#define QR4_0 2
+
+#define QI4_1 (QK4_1 / (4 * QR4_1))
+#define QR4_1 2
+
+#define QI5_0 (QK5_0 / (4 * QR5_0))
+#define QR5_0 2
+
+#define QI5_1 (QK5_1 / (4 * QR5_1))
+#define QR5_1 2
+
+#define QI8_0 (QK8_0 / (4 * QR8_0))
+#define QR8_0 1
+
+#define QI8_1 (QK8_1 / (4 * QR8_1))
+#define QR8_1 1
+
+#define QI2_K (QK_K / (4*QR2_K))
+#define QR2_K 4
+
+#define QI3_K (QK_K / (4*QR3_K))
+#define QR3_K 4
+
+#define QI4_K (QK_K / (4*QR4_K))
+#define QR4_K 2
+
+#define QI5_K (QK_K / (4*QR5_K))
+#define QR5_K 2
+
+#define QI6_K (QK_K / (4*QR6_K))
+#define QR6_K 2
+
+#define QI2_XXS (QK_K / (4*QR2_XXS))
+#define QR2_XXS 8
+
+#define QI2_XS (QK_K / (4*QR2_XS))
+#define QR2_XS 8
+
+#define QI2_S (QK_K / (4*QR2_S))
+#define QR2_S 8
+
+#define QI3_XXS (QK_K / (4*QR3_XXS))
+#define QR3_XXS 8
+
+#define QI3_XS (QK_K / (4*QR3_XS))
+#define QR3_XS 8
+
+#define QI1_S (QK_K / (4*QR1_S))
+#define QR1_S 8
+
+#define QI4_NL (QK4_NL / (4*QR4_NL))
+#define QR4_NL 2
+
+#if QK_K == 64
+#define QI4_XS QI4_NL
+#define QR4_XS QR4_NL
+#else
+#define QI4_XS (QK_K / (4*QR4_XS))
+#define QR4_XS 8
+#endif
+
+#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
+
+#define QK4_0 32
+typedef struct {
+    ggml_half d;           // delta
+    uint8_t qs[QK4_0 / 2]; // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half m; // min
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t qs[QK4_1 / 2]; // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_half d;           // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half m; // min
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_half d;       // delta
+    int8_t  qs[QK8_0]; // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half s; // d * sum(qs[i])
+        } GGML_COMMON_AGGR;
+        ggml_half2 ds;
+    };
+    int8_t qs[QK8_1]; // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8]; // quants - high bit
+    uint8_t qs[QK_K/4];    // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_half d;           // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8]; // quants - high bit
+    uint8_t qs[QK_K/4];    // quants - low 2 bits
+    uint8_t scales[12];    // scales, quantized with 6 bits
+    ggml_half d;           // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_half d[2];     // super-block scales/mins
+    uint8_t scales[2];  // 4-bit block scales/mins
+    uint8_t qs[QK_K/2]; // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];           // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_half d;             // super-block scale
+    int8_t  scales[QK_K/16]; // 8-bit block scales
+    uint8_t qh[QK_K/8];      // quants, high bit
+    uint8_t qs[QK_K/2];      // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];           // quants, high bit
+    uint8_t qs[QK_K/2];           // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_half d;             // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+// (Almost) "true" 2-bit quantization.
+// Due to the need to use blocks as per ggml design, it ends up using
+// 2.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_half d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
+
+// 2.3125 bpw quants
+typedef struct {
+    ggml_half d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
+
+// 2.5625 bpw quants
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t scales[QK_K/32];
+} block_iq2_s;
+static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
+
+// (Almost) "true" 3-bit quantization.
+// Due to the need to use blocks as per ggml design, it ends up using
+// 3.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_half d;
+    uint8_t qs[3*QK_K/8];
+} block_iq3_xxs;
+static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
+
+// 3.4375 bpw
+#if QK_K == 64
+#define IQ3S_N_SCALE 2
+#else
+#define IQ3S_N_SCALE QK_K/64
+#endif
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t signs[QK_K/8];
+    uint8_t scales[IQ3S_N_SCALE];
+} block_iq3_s;
+static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
+
+typedef struct {
+    ggml_half d;
+    uint8_t  qs[QK_K/8];
+    uint16_t qh[QK_K/32];
+} block_iq1_s;
+static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
+
+// 1.75 bpw
+typedef struct {
+    uint8_t  qs[QK_K/8];      // grid index, low 8 bits
+    uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
+#if QK_K == 64
+    ggml_half d;
+#endif
+    uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
+} block_iq1_m;
+#if QK_K == 64
+static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
+#else
+static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
+#endif
+
+// Used by IQ1_M quants
+typedef union {
+    ggml_half f16;
+    uint16_t  u16;
+} iq1m_scale_t;
+
+// Non-linear quants
+#define QK4_NL 32
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK4_NL/2];
+} block_iq4_nl;
+static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
+
+#if QK_K == 64
+#define block_iq4_xs block_iq4_nl
+#else
+typedef struct {
+    ggml_half d;
+    uint16_t scales_h;
+    uint8_t  scales_l[QK_K/64];
+    uint8_t  qs[QK_K/2];
+} block_iq4_xs;
+static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
+#endif
+
+#endif // GGML_COMMON_DECL
+#endif // GGML_COMMON_DECL
+
+////////////////////////////////////////////////////////////////////////////////
+
+#ifndef GGML_COMMON_IMPL
+
+#if defined(GGML_COMMON_IMPL_C)
+#include <stdint.h>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_METAL)
+#include <metal_stdlib>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_SYCL)
+
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#endif
+
+#if defined(GGML_COMMON_IMPL)
+
+GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8)
+    1, 2, 4, 8, 16, 32, 64, 128
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
+      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
+    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
+    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
+     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
+    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
+     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
+     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
+    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
+GGML_TABLE_END()
+
+//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
+    0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
+    0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
+    0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
+    0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
+    0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
+    0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
+    0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
+    0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
+    0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
+    0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
+    0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
+    0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
+    0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
+    0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
+    0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
+    0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
+    0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
+    0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
+    0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
+    0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
+    0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
+    0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
+    0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
+    0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
+    0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
+    0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
+    0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
+    0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
+    0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
+    0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
+    0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
+    0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
+GGML_TABLE_END()
+//#endif
+
+
+GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
+    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
+    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
+    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
+    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
+    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
+    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
+    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
+    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
+    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
+    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
+    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
+    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
+    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
+    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
+    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
+    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
+    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
+    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
+    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
+    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
+    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
+    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
+    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
+    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
+    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
+    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
+    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
+    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
+    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
+    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
+    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
+    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
+    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
+    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
+    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
+    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
+    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
+    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
+    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
+    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
+    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
+    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
+    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
+    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
+    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
+    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
+    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
+    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
+    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
+    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
+    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
+    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
+    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
+    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
+    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
+    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
+    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
+    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512)
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
+    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
+    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
+    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
+    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
+    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
+    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
+    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
+    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
+    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
+    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
+    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
+    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
+    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
+    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
+    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
+    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
+    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
+    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
+    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
+    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
+    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
+    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
+    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
+    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
+    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
+    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
+    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
+    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
+    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
+    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
+    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
+    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
+    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
+    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
+    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
+    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
+    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
+    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
+    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
+    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
+    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
+    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
+    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
+    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
+    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
+    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
+    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
+    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
+    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
+    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
+    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
+    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
+    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
+    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
+    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
+    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
+    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
+    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
+    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
+    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
+    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
+    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
+    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
+    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
+    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
+    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
+    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
+    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
+    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
+    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
+    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
+    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
+    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
+    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
+    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
+    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
+    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
+    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
+    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
+    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
+    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
+    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
+    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
+    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
+    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
+    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
+    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
+    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
+    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
+    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
+    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
+    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
+    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
+    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
+    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
+    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
+    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
+    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
+    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
+    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
+    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
+    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
+    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
+    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
+    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
+    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
+    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
+    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
+    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
+    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
+    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
+    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
+    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
+    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024)
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
+    0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
+    0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
+    0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
+    0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
+    0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
+    0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
+    0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
+    0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
+    0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
+    0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
+    0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
+    0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
+    0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
+    0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
+    0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
+    0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
+    0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
+    0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
+    0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
+    0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
+    0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
+    0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
+    0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
+    0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
+    0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
+    0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
+    0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
+    0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
+    0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
+    0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
+    0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
+    0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
+    0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
+    0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
+    0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
+    0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
+    0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
+    0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
+    0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
+    0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
+    0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
+    0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
+    0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
+    0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
+    0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
+    0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
+    0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
+    0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
+    0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
+    0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
+    0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
+    0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
+    0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
+    0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
+    0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
+    0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
+    0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
+    0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
+    0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
+    0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
+    0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
+    0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
+    0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
+    0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
+    0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
+    0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
+    0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
+    0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
+    0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
+    0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
+    0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
+    0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
+    0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
+    0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
+    0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
+    0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
+    0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
+    0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
+    0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
+    0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
+    0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
+    0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
+    0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
+    0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
+    0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
+    0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
+    0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
+    0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
+    0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
+    0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
+    0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
+    0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
+    0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
+    0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
+    0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
+    0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
+    0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
+    0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
+    0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
+    0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
+    0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
+    0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
+    0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
+    0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
+    0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
+    0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
+    0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
+    0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
+    0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
+    0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
+    0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
+    0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
+    0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
+    0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
+    0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
+    0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
+    0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
+    0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
+    0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
+    0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
+    0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
+    0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
+    0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
+    0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
+    0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
+    0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
+    0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
+    0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
+    0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
+    0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
+    0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
+    0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
+    0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
+    0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
+    0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
+    0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
+    0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
+    0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
+    0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
+    0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
+    0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
+    0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
+    0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
+    0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
+    0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
+    0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
+    0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
+    0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
+    0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
+    0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
+    0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
+    0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
+    0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
+    0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
+    0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
+    0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
+    0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
+    0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
+    0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
+    0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
+    0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
+    0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
+    0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
+    0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
+    0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
+    0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
+    0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
+    0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
+    0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
+    0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
+    0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
+    0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
+    0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
+    0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
+    0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
+    0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
+    0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
+    0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
+    0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
+    0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
+    0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
+    0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
+    0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
+    0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
+    0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
+    0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
+    0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
+    0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
+    0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
+    0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
+    0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
+    0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
+    0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
+    0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
+    0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
+    0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
+    0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
+    0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
+    0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
+    0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
+    0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
+    0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
+    0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
+    0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
+    0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
+    0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
+    0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
+    0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
+    0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
+    0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
+    0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
+    0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
+    0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
+    0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
+    0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
+    0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
+    0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
+    0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
+    0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
+    0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
+    0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
+    0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
+    0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
+    0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
+    0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
+    0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
+    0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
+    0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
+    0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
+    0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
+    0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
+    0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
+    0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
+    0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
+    0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
+    0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
+    0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
+    0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
+    0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
+    0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
+    0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
+    0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256)
+    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
+    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
+    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
+    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
+    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
+    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
+    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
+    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
+    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
+    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
+    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
+    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
+    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
+    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
+    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
+    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
+    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
+    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
+    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
+    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
+    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
+    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
+    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
+    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
+    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
+    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
+    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
+    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
+    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
+    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
+    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
+    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
+    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
+    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
+    0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
+    0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
+    0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
+    0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
+    0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
+    0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
+    0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
+    0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
+    0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
+    0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
+    0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
+    0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
+    0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
+    0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
+    0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
+    0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
+    0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
+    0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
+    0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
+    0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
+    0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
+    0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
+    0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
+    0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
+    0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
+    0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
+    0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
+    0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
+    0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
+    0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
+    0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
+    0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
+    0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
+    0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
+    0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
+    0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
+    0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
+    0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
+    0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
+    0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
+    0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
+    0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
+    0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
+    0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
+    0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
+    0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
+    0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
+    0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
+    0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
+    0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
+    0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
+    0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
+    0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
+    0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
+    0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
+    0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
+    0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
+    0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
+    0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
+    0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
+    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
+    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
+GGML_TABLE_END()
+
+#define NGRID_IQ1S 2048
+#define IQ1S_DELTA 0.125f
+#define IQ1M_DELTA 0.125f
+#if defined(GGML_COMMON_IMPL_C)
+GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
+    0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
+    0xffffffffffff0101, 0xffffffffff00ff00, 0xffffffffff000000, 0xffffffffff01ffff,
+    0xffffffffff01ff01, 0xffffffffff0101ff, 0xffffffffff010101, 0xffffffff00ff0000,
+    0xffffffff0000ff00, 0xffffffff000000ff, 0xffffffff00000001, 0xffffffff00010000,
+    0xffffffff01ffffff, 0xffffffff01ffff01, 0xffffffff01ff01ff, 0xffffffff01ff0101,
+    0xffffffff01000000, 0xffffffff0101ffff, 0xffffffff0101ff01, 0xffffffff010101ff,
+    0xffffffff01010101, 0xffffff00ffff00ff, 0xffffff00ffff0000, 0xffffff00ff00ff00,
+    0xffffff00ff0000ff, 0xffffff00ff000001, 0xffffff00ff000100, 0xffffff00ff000101,
+    0xffffff00ff010000, 0xffffff0000ffff00, 0xffffff0000ff0001, 0xffffff0000ff0100,
+    0xffffff000000ff01, 0xffffff0000000000, 0xffffff0000000101, 0xffffff000001ff00,
+    0xffffff00000100ff, 0xffffff0000010001, 0xffffff00000101ff, 0xffffff0001ff0000,
+    0xffffff000100ff00, 0xffffff00010000ff, 0xffffff0001000001, 0xffffff0001010000,
+    0xffffff01ffffffff, 0xffffff01ffffff01, 0xffffff01ffff01ff, 0xffffff01ffff0101,
+    0xffffff01ff000000, 0xffffff01ff01ffff, 0xffffff01ff01ff01, 0xffffff01ff0101ff,
+    0xffffff01ff010101, 0xffffff0100ff0000, 0xffffff010000ff00, 0xffffff0100000100,
+    0xffffff01000100ff, 0xffffff0100010100, 0xffffff0101ffffff, 0xffffff0101ffff01,
+    0xffffff0101ff01ff, 0xffffff0101ff0101, 0xffffff010100ff00, 0xffffff0101000000,
+    0xffffff0101000100, 0xffffff010101ffff, 0xffffff010101ff01, 0xffffff01010101ff,
+    0xffffff0101010101, 0xffff00ffff00ff00, 0xffff00ffff0000ff, 0xffff00ffff000001,
+    0xffff00ffff010000, 0xffff00ff00ffff00, 0xffff00ff00ff0100, 0xffff00ff00000000,
+    0xffff00ff00000101, 0xffff00ff000100ff, 0xffff00ff00010000, 0xffff00ff0100ff00,
+    0xffff00ff01000100, 0xffff00ff01010000, 0xffff0000ffffff00, 0xffff0000ffff00ff,
+    0xffff0000ffff0000, 0xffff0000ffff0001, 0xffff0000ff000000, 0xffff0000ff0001ff,
+    0xffff0000ff000101, 0xffff0000ff010100, 0xffff000000ffffff, 0xffff000000ff0000,
+    0xffff000000ff0101, 0xffff00000000ffff, 0xffff00000000ff00, 0xffff0000000000ff,
+    0xffff000000000000, 0xffff000000000001, 0xffff000000000100, 0xffff00000001ffff,
+    0xffff00000001ff01, 0xffff000000010000, 0xffff0000000101ff, 0xffff000000010101,
+    0xffff000001ffff00, 0xffff00000100ff00, 0xffff000001000000, 0xffff0000010001ff,
+    0xffff000001000101, 0xffff00000101ff00, 0xffff0000010100ff, 0xffff000001010000,
+    0xffff000001010001, 0xffff000001010100, 0xffff0001ff0000ff, 0xffff0001ff000100,
+    0xffff000100ffff00, 0xffff000100ff00ff, 0xffff00010000ffff, 0xffff00010000ff01,
+    0xffff000100000000, 0xffff0001000001ff, 0xffff00010001ffff, 0xffff00010001ff00,
+    0xffff000100010001, 0xffff000100010100, 0xffff000101ff0000, 0xffff00010100ff00,
+    0xffff0001010000ff, 0xffff000101000100, 0xffff01ffffffffff, 0xffff01ffffffff01,
+    0xffff01ffffff01ff, 0xffff01ffffff0101, 0xffff01ffff000000, 0xffff01ffff01ffff,
+    0xffff01ffff01ff01, 0xffff01ffff0101ff, 0xffff01ffff010101, 0xffff01ff00ff0000,
+    0xffff01ff0000ff00, 0xffff01ff00000001, 0xffff01ff00010000, 0xffff01ff01ffffff,
+    0xffff01ff01ffff01, 0xffff01ff01ff01ff, 0xffff01ff01ff0101, 0xffff01ff01000000,
+    0xffff01ff0101ffff, 0xffff01ff0101ff01, 0xffff01ff010101ff, 0xffff01ff01010101,
+    0xffff0100ffff0000, 0xffff0100ff00ff00, 0xffff0100ff0000ff, 0xffff0100ff000100,
+    0xffff0100ff0100ff, 0xffff0100ff010000, 0xffff010000ffff00, 0xffff01000000ffff,
+    0xffff01000000ff00, 0xffff010000000000, 0xffff01000001ff00, 0xffff0100000100ff,
+    0xffff010000010100, 0xffff01000100ff00, 0xffff0100010000ff, 0xffff010001000001,
+    0xffff010001000100, 0xffff010001010000, 0xffff0101ffffffff, 0xffff0101ffffff01,
+    0xffff0101ffff01ff, 0xffff0101ffff0101, 0xffff0101ff000000, 0xffff0101ff01ffff,
+    0xffff0101ff01ff01, 0xffff0101ff0101ff, 0xffff0101ff010101, 0xffff010100ff0000,
+    0xffff01010000ff00, 0xffff010100000100, 0xffff01010001ff00, 0xffff010100010000,
+    0xffff010101ffffff, 0xffff010101ffff01, 0xffff010101ff0000, 0xffff010101ff01ff,
+    0xffff010101ff0101, 0xffff010101000000, 0xffff01010101ffff, 0xffff01010101ff01,
+    0xffff0101010101ff, 0xffff010101010101, 0xff00ffffff00ffff, 0xff00ffffff00ff00,
+    0xff00ffffff0000ff, 0xff00ffffff000100, 0xff00ffffff0100ff, 0xff00ffffff010000,
+    0xff00ffff00ffff00, 0xff00ffff00ff00ff, 0xff00ffff0000ffff, 0xff00ffff00000000,
+    0xff00ffff000001ff, 0xff00ffff0001ff00, 0xff00ffff000100ff, 0xff00ffff00010000,
+    0xff00ffff00010100, 0xff00ffff0100ff00, 0xff00ffff010000ff, 0xff00ffff01000001,
+    0xff00ffff0101ff00, 0xff00ffff01010000, 0xff00ff00ffffff00, 0xff00ff00ffff00ff,
+    0xff00ff00ffff0001, 0xff00ff00ffff0100, 0xff00ff00ff00ffff, 0xff00ff00ff00ff01,
+    0xff00ff00ff000000, 0xff00ff00ff0001ff, 0xff00ff00ff01ff00, 0xff00ff00ff0100ff,
+    0xff00ff00ff010100, 0xff00ff0000ff0000, 0xff00ff0000ff0101, 0xff00ff000000ffff,
+    0xff00ff000000ff00, 0xff00ff000000ff01, 0xff00ff00000000ff, 0xff00ff0000000000,
+    0xff00ff0000000001, 0xff00ff0000000100, 0xff00ff000001ffff, 0xff00ff0000010000,
+    0xff00ff0001ff00ff, 0xff00ff000100ff01, 0xff00ff0001000000, 0xff00ff000101ff00,
+    0xff00ff00010100ff, 0xff00ff01ff00ff00, 0xff00ff01ff0000ff, 0xff00ff01ff000001,
+    0xff00ff01ff010000, 0xff00ff0100ffffff, 0xff00ff0100ff0001, 0xff00ff0100ff0100,
+    0xff00ff010000ff01, 0xff00ff0100000000, 0xff00ff01000001ff, 0xff00ff0100000101,
+    0xff00ff01000100ff, 0xff00ff0100010001, 0xff00ff0101ff0000, 0xff00ff010100ff00,
+    0xff00ff01010000ff, 0xff00ff0101000001, 0xff00ff0101010000, 0xff0000ffffffff00,
+    0xff0000ffffff0001, 0xff0000ffffff0100, 0xff0000ffff0000ff, 0xff0000ffff000000,
+    0xff0000ffff0001ff, 0xff0000ffff000100, 0xff0000ffff01ff00, 0xff0000ffff010001,
+    0xff0000ff00ffff00, 0xff0000ff00ff0000, 0xff0000ff00ff0001, 0xff0000ff00ff01ff,
+    0xff0000ff00ff0101, 0xff0000ff0000ff00, 0xff0000ff000000ff, 0xff0000ff00000000,
+    0xff0000ff00000001, 0xff0000ff00000100, 0xff0000ff0001ff01, 0xff0000ff00010000,
+    0xff0000ff000101ff, 0xff0000ff01ff00ff, 0xff0000ff01ff0100, 0xff0000ff0100ffff,
+    0xff0000ff010000ff, 0xff0000ff01000000, 0xff0000ff010001ff, 0xff0000ff01000100,
+    0xff0000ff01000101, 0xff0000ff0101ff00, 0xff0000ff010100ff, 0xff0000ff01010000,
+    0xff0000ff01010100, 0xff000000ffffff01, 0xff000000ffff0000, 0xff000000ffff0101,
+    0xff000000ff00ff00, 0xff000000ff0000ff, 0xff000000ff000000, 0xff000000ff000001,
+    0xff000000ff000100, 0xff000000ff01ffff, 0xff000000ff01ff01, 0xff000000ff010000,
+    0xff000000ff0101ff, 0xff000000ff010101, 0xff00000000ffff00, 0xff00000000ff00ff,
+    0xff00000000ff0000, 0xff00000000ff0001, 0xff0000000000ff00, 0xff0000000000ff01,
+    0xff000000000000ff, 0xff00000000000000, 0xff00000000000001, 0xff00000000000100,
+    0xff00000000000101, 0xff0000000001ff00, 0xff000000000100ff, 0xff00000000010000,
+    0xff00000000010001, 0xff00000000010100, 0xff00000001ffffff, 0xff00000001ffff01,
+    0xff00000001ff00ff, 0xff00000001ff0000, 0xff00000001ff01ff, 0xff00000001ff0101,
+    0xff0000000100ffff, 0xff0000000100ff00, 0xff000000010000ff, 0xff00000001000000,
+    0xff00000001000001, 0xff00000001000100, 0xff00000001000101, 0xff0000000101ffff,
+    0xff0000000101ff01, 0xff00000001010000, 0xff000001ffffff00, 0xff000001ffff00ff,
+    0xff000001ffff0000, 0xff000001ffff0001, 0xff000001ff000000, 0xff000001ff000001,
+    0xff000001ff0001ff, 0xff000001ff000101, 0xff000001ff01ff00, 0xff000001ff010001,
+    0xff00000100ffffff, 0xff00000100ffff01, 0xff00000100ff00ff, 0xff00000100ff0000,
+    0xff00000100ff01ff, 0xff00000100ff0101, 0xff0000010000ff00, 0xff00000100000000,
+    0xff00000100000001, 0xff000001000001ff, 0xff00000100000100, 0xff0000010001ff00,
+    0xff000001000100ff, 0xff00000100010000, 0xff000001000101ff, 0xff00000100010100,
+    0xff00000100010101, 0xff00000101ff0001, 0xff00000101ff0101, 0xff0000010100ff01,
+    0xff00000101000000, 0xff000001010100ff, 0xff00000101010100, 0xff0001ffff00ff00,
+    0xff0001ffff000001, 0xff0001ffff010000, 0xff0001ff00ffff00, 0xff0001ff00ff00ff,
+    0xff0001ff00ff0001, 0xff0001ff00ff0100, 0xff0001ff0000ffff, 0xff0001ff00000000,
+    0xff0001ff000001ff, 0xff0001ff00000101, 0xff0001ff0001ffff, 0xff0001ff0001ff00,
+    0xff0001ff000100ff, 0xff0001ff00010001, 0xff0001ff00010100, 0xff0001ff01ff0000,
+    0xff0001ff0100ff00, 0xff0001ff010000ff, 0xff0001ff01010000, 0xff000100ff00ffff,
+    0xff000100ff00ff01, 0xff000100ff000000, 0xff000100ff000101, 0xff000100ff01ff00,
+    0xff000100ff010000, 0xff00010000ffff01, 0xff00010000ff00ff, 0xff00010000ff0000,
+    0xff00010000ff01ff, 0xff0001000000ff00, 0xff000100000000ff, 0xff00010000000000,
+    0xff00010000000001, 0xff00010000000100, 0xff00010000000101, 0xff0001000001ffff,
+    0xff00010000010000, 0xff00010000010101, 0xff00010001ff0100, 0xff0001000100ff00,
+    0xff0001000100ff01, 0xff00010001000000, 0xff000100010001ff, 0xff0001000101ff00,
+    0xff00010001010001, 0xff00010001010100, 0xff000101ffff0100, 0xff000101ff000001,
+    0xff000101ff0100ff, 0xff000101ff010001, 0xff00010100ff00ff, 0xff00010100ff0001,
+    0xff00010100ff0100, 0xff0001010000ffff, 0xff0001010000ff01, 0xff00010100000000,
+    0xff000101000001ff, 0xff0001010001ff00, 0xff00010100010001, 0xff00010100010100,
+    0xff00010101ff0000, 0xff0001010100ff00, 0xff00010101000001, 0xff00010101000101,
+    0xff01ffffffffffff, 0xff01ffffffffff01, 0xff01ffffffff01ff, 0xff01ffffffff0101,
+    0xff01ffffff000000, 0xff01ffffff01ffff, 0xff01ffffff01ff01, 0xff01ffffff010000,
+    0xff01ffffff0101ff, 0xff01ffffff010101, 0xff01ffff00ff0000, 0xff01ffff0000ff00,
+    0xff01ffff00000100, 0xff01ffff0001ff00, 0xff01ffff00010000, 0xff01ffff01ffffff,
+    0xff01ffff01ffff01, 0xff01ffff01ff01ff, 0xff01ffff01ff0101, 0xff01ffff01000000,
+    0xff01ffff0101ffff, 0xff01ffff0101ff01, 0xff01ffff01010000, 0xff01ffff010101ff,
+    0xff01ffff01010101, 0xff01ff00ffff0000, 0xff01ff00ff00ff00, 0xff01ff00ff0000ff,
+    0xff01ff00ff000100, 0xff01ff00ff010000, 0xff01ff0000ffff01, 0xff01ff0000ff00ff,
+    0xff01ff0000ff0100, 0xff01ff0000000000, 0xff01ff00000001ff, 0xff01ff0000000101,
+    0xff01ff000001ff00, 0xff01ff00000100ff, 0xff01ff0000010000, 0xff01ff0000010001,
+    0xff01ff0001ff0000, 0xff01ff000100ffff, 0xff01ff0001000001, 0xff01ff0001000100,
+    0xff01ff0001010000, 0xff01ff01ffffff00, 0xff01ff01ffff01ff, 0xff01ff01ffff0101,
+    0xff01ff01ff00ff00, 0xff01ff01ff000000, 0xff01ff01ff01ffff, 0xff01ff01ff01ff01,
+    0xff01ff01ff0101ff, 0xff01ff01ff010101, 0xff01ff0100ff0000, 0xff01ff010000ff00,
+    0xff01ff0100000001, 0xff01ff0100000100, 0xff01ff0100010000, 0xff01ff0101ffff00,
+    0xff01ff0101ff01ff, 0xff01ff0101ff0101, 0xff01ff010100ff00, 0xff01ff0101000000,
+    0xff01ff010101ffff, 0xff01ff010101ff01, 0xff01ff01010101ff, 0xff01ff0101010101,
+    0xff0100ffffff0000, 0xff0100ffff0000ff, 0xff0100ffff000001, 0xff0100ffff000100,
+    0xff0100ffff010000, 0xff0100ff00ff00ff, 0xff0100ff00ff0000, 0xff0100ff00ff0001,
+    0xff0100ff00ff0100, 0xff0100ff0000ff01, 0xff0100ff00000000, 0xff0100ff000001ff,
+    0xff0100ff00000101, 0xff0100ff00010001, 0xff0100ff01ff0000, 0xff0100ff0100ff00,
+    0xff0100ff010000ff, 0xff0100ff01000100, 0xff0100ff0101ff00, 0xff0100ff01010000,
+    0xff010000ffff0100, 0xff010000ff000000, 0xff010000ff01ff00, 0xff010000ff010100,
+    0xff01000000ffffff, 0xff01000000ff0000, 0xff01000000ff01ff, 0xff0100000000ff00,
+    0xff010000000000ff, 0xff01000000000000, 0xff01000000000100, 0xff0100000001ff01,
+    0xff01000000010000, 0xff010000000101ff, 0xff01000001ff0100, 0xff0100000100ffff,
+    0xff010000010000ff, 0xff01000001000000, 0xff010000010001ff, 0xff01000001000101,
+    0xff0100000101ff00, 0xff010000010100ff, 0xff01000001010001, 0xff01000001010100,
+    0xff010001ffff0000, 0xff010001ff00ffff, 0xff010001ff00ff01, 0xff010001ff000100,
+    0xff010001ff010000, 0xff01000100ffff00, 0xff01000100ff0100, 0xff01000100000000,
+    0xff0100010001ffff, 0xff0100010001ff00, 0xff01000100010100, 0xff01000101ff00ff,
+    0xff01000101ff0001, 0xff0100010100ffff, 0xff01000101000101, 0xff0101ffffffffff,
+    0xff0101ffffffff01, 0xff0101ffffff01ff, 0xff0101ffffff0101, 0xff0101ffff000000,
+    0xff0101ffff01ffff, 0xff0101ffff01ff01, 0xff0101ffff0101ff, 0xff0101ffff010101,
+    0xff0101ff00ff0000, 0xff0101ff0000ff00, 0xff0101ff000000ff, 0xff0101ff00010000,
+    0xff0101ff01ffffff, 0xff0101ff01ffff01, 0xff0101ff01ff01ff, 0xff0101ff01ff0101,
+    0xff0101ff0101ffff, 0xff0101ff0101ff01, 0xff0101ff010101ff, 0xff0101ff01010101,
+    0xff010100ffff0100, 0xff010100ff00ff00, 0xff010100ff0000ff, 0xff010100ff000100,
+    0xff010100ff010000, 0xff01010000ff0001, 0xff01010000ff0100, 0xff0101000000ff01,
+    0xff01010000000000, 0xff0101000001ff00, 0xff010100000100ff, 0xff01010000010001,
+    0xff01010000010100, 0xff01010001ff0000, 0xff0101000100ffff, 0xff01010001000001,
+    0xff01010001000100, 0xff010100010100ff, 0xff01010001010000, 0xff010101ffffffff,
+    0xff010101ffffff01, 0xff010101ffff01ff, 0xff010101ffff0101, 0xff010101ff01ffff,
+    0xff010101ff01ff01, 0xff010101ff0101ff, 0xff010101ff010101, 0xff01010100ff0000,
+    0xff0101010000ff00, 0xff01010100000001, 0xff01010100000100, 0xff01010100010000,
+    0xff01010101ffffff, 0xff01010101ffff01, 0xff01010101ff01ff, 0xff01010101ff0101,
+    0xff01010101000000, 0xff0101010101ffff, 0xff0101010101ff01, 0xff010101010101ff,
+    0xff01010101010101, 0x00ffffffffff0000, 0x00ffffffff00ff00, 0x00ffffffff000001,
+    0x00ffffffff010000, 0x00ffffff00ff0100, 0x00ffffff0000ff01, 0x00ffffff00000000,
+    0x00ffffff000001ff, 0x00ffffff00000101, 0x00ffffff0001ff00, 0x00ffffff000100ff,
+    0x00ffffff00010001, 0x00ffffff010000ff, 0x00ffffff01000100, 0x00ffffff0101ff00,
+    0x00ffffff01010001, 0x00ffff00ffffffff, 0x00ffff00ffffff00, 0x00ffff00ffff00ff,
+    0x00ffff00ffff0001, 0x00ffff00ffff0100, 0x00ffff00ff00ff01, 0x00ffff00ff000000,
+    0x00ffff00ff000001, 0x00ffff00ff0001ff, 0x00ffff00ff000101, 0x00ffff00ff01ff00,
+    0x00ffff00ff010001, 0x00ffff00ff010100, 0x00ffff0000ff0000, 0x00ffff0000ff01ff,
+    0x00ffff0000ff0101, 0x00ffff000000ff00, 0x00ffff00000000ff, 0x00ffff0000000000,
+    0x00ffff0000000001, 0x00ffff0000000100, 0x00ffff0000000101, 0x00ffff0000010000,
+    0x00ffff00000101ff, 0x00ffff0000010101, 0x00ffff0001ffff00, 0x00ffff0001ff00ff,
+    0x00ffff0001ff0001, 0x00ffff000100ffff, 0x00ffff000100ff01, 0x00ffff0001000000,
+    0x00ffff000101ffff, 0x00ffff000101ff00, 0x00ffff000101ff01, 0x00ffff01ffff0000,
+    0x00ffff01ff00ff00, 0x00ffff01ff0000ff, 0x00ffff01ff000001, 0x00ffff01ff010000,
+    0x00ffff0100ffff00, 0x00ffff010000ff01, 0x00ffff0100000000, 0x00ffff0100000101,
+    0x00ffff01000100ff, 0x00ffff0100010100, 0x00ffff0101ff0100, 0x00ffff01010000ff,
+    0x00ffff0101010000, 0x00ff00ffffffff00, 0x00ff00ffff000000, 0x00ff00ffff000100,
+    0x00ff00ffff010100, 0x00ff00ff00ff0000, 0x00ff00ff00ff01ff, 0x00ff00ff00ff0101,
+    0x00ff00ff0000ff00, 0x00ff00ff000000ff, 0x00ff00ff00000000, 0x00ff00ff00000001,
+    0x00ff00ff0001ff00, 0x00ff00ff0001ff01, 0x00ff00ff00010000, 0x00ff00ff000101ff,
+    0x00ff00ff00010101, 0x00ff00ff01ffff00, 0x00ff00ff01ff0001, 0x00ff00ff01ff0100,
+    0x00ff00ff0100ffff, 0x00ff00ff0100ff01, 0x00ff00ff01000000, 0x00ff00ff0101ffff,
+    0x00ff00ff0101ff00, 0x00ff00ff01010100, 0x00ff0000ffffff00, 0x00ff0000ffffff01,
+    0x00ff0000ffff0000, 0x00ff0000ffff0101, 0x00ff0000ff00ff00, 0x00ff0000ff0000ff,
+    0x00ff0000ff000000, 0x00ff0000ff000001, 0x00ff0000ff000100, 0x00ff0000ff01ffff,
+    0x00ff0000ff010000, 0x00ff0000ff010101, 0x00ff000000ffff00, 0x00ff000000ff00ff,
+    0x00ff000000ff0000, 0x00ff000000ff0001, 0x00ff000000ff0100, 0x00ff00000000ffff,
+    0x00ff00000000ff00, 0x00ff0000000000ff, 0x00ff000000000000, 0x00ff000000000001,
+    0x00ff0000000001ff, 0x00ff000000000100, 0x00ff00000001ff00, 0x00ff0000000100ff,
+    0x00ff000000010000, 0x00ff000000010001, 0x00ff000000010100, 0x00ff000001ffff01,
+    0x00ff000001ff00ff, 0x00ff000001ff0000, 0x00ff000001ff01ff, 0x00ff00000100ff00,
+    0x00ff0000010000ff, 0x00ff000001000000, 0x00ff000001000001, 0x00ff000001000100,
+    0x00ff000001000101, 0x00ff000001010000, 0x00ff0000010101ff, 0x00ff000001010101,
+    0x00ff0001ffffff00, 0x00ff0001ffff0000, 0x00ff0001ffff0100, 0x00ff0001ff0000ff,
+    0x00ff0001ff000000, 0x00ff0001ff0001ff, 0x00ff0001ff000101, 0x00ff0001ff01ff00,
+    0x00ff0001ff0100ff, 0x00ff0001ff010100, 0x00ff000100ffffff, 0x00ff000100ffff01,
+    0x00ff000100ff0000, 0x00ff000100ff01ff, 0x00ff00010000ffff, 0x00ff00010000ff00,
+    0x00ff00010000ff01, 0x00ff000100000000, 0x00ff000100000001, 0x00ff000100000100,
+    0x00ff00010001ff01, 0x00ff000100010000, 0x00ff0001000101ff, 0x00ff000101ffff00,
+    0x00ff000101ff0000, 0x00ff000101ff0101, 0x00ff0001010000ff, 0x00ff000101000000,
+    0x00ff00010101ff00, 0x00ff0001010100ff, 0x00ff000101010001, 0x00ff01ffffff0000,
+    0x00ff01ffff00ff00, 0x00ff01ffff000000, 0x00ff01ffff000101, 0x00ff01ffff010000,
+    0x00ff01ff00ffff01, 0x00ff01ff00ff0100, 0x00ff01ff0000ffff, 0x00ff01ff00000000,
+    0x00ff01ff000001ff, 0x00ff01ff0001ff00, 0x00ff01ff000100ff, 0x00ff01ff00010001,
+    0x00ff01ff00010100, 0x00ff01ff01ff0000, 0x00ff01ff0100ff00, 0x00ff01ff010000ff,
+    0x00ff01ff01000001, 0x00ff01ff01000100, 0x00ff01ff01010000, 0x00ff0100ffffff00,
+    0x00ff0100ffff0000, 0x00ff0100ffff0001, 0x00ff0100ffff0101, 0x00ff0100ff00ffff,
+    0x00ff0100ff0000ff, 0x00ff0100ff000000, 0x00ff0100ff0001ff, 0x00ff0100ff01ff00,
+    0x00ff0100ff0100ff, 0x00ff0100ff010001, 0x00ff010000ffffff, 0x00ff010000ff0000,
+    0x00ff010000ff0101, 0x00ff01000000ff00, 0x00ff01000000ff01, 0x00ff0100000000ff,
+    0x00ff010000000000, 0x00ff010000000001, 0x00ff010000000100, 0x00ff01000001ffff,
+    0x00ff01000001ff01, 0x00ff010000010000, 0x00ff010000010001, 0x00ff010000010101,
+    0x00ff010001ff0001, 0x00ff010001ff0100, 0x00ff01000100ff01, 0x00ff010001000000,
+    0x00ff010001000001, 0x00ff0100010001ff, 0x00ff01000101ff00, 0x00ff0100010100ff,
+    0x00ff010001010001, 0x00ff010001010100, 0x00ff0101ff000001, 0x00ff010100ff00ff,
+    0x00ff010100ff0001, 0x00ff010100ff0100, 0x00ff010100000000, 0x00ff0101000001ff,
+    0x00ff010100000101, 0x00ff0101000100ff, 0x00ff010100010100, 0x00ff0101010000ff,
+    0x00ff010101010000, 0x0000ffffffffff00, 0x0000ffffffff00ff, 0x0000ffffffff0000,
+    0x0000ffffffff0001, 0x0000ffffffff0100, 0x0000ffffff00ff01, 0x0000ffffff000000,
+    0x0000ffffff000101, 0x0000ffffff01ff00, 0x0000ffffff0100ff, 0x0000ffffff010100,
+    0x0000ffff00ffffff, 0x0000ffff00ff0000, 0x0000ffff00ff01ff, 0x0000ffff0000ff00,
+    0x0000ffff000000ff, 0x0000ffff00000000, 0x0000ffff00000001, 0x0000ffff00000100,
+    0x0000ffff00010000, 0x0000ffff000101ff, 0x0000ffff01ff0001, 0x0000ffff01ff0100,
+    0x0000ffff01000000, 0x0000ffff010001ff, 0x0000ffff0101ffff, 0x0000ffff0101ff00,
+    0x0000ffff01010001, 0x0000ffff01010100, 0x0000ff00ffff0000, 0x0000ff00ffff01ff,
+    0x0000ff00ffff0100, 0x0000ff00ffff0101, 0x0000ff00ff00ff00, 0x0000ff00ff0000ff,
+    0x0000ff00ff000000, 0x0000ff00ff000001, 0x0000ff00ff0001ff, 0x0000ff00ff000100,
+    0x0000ff00ff01ffff, 0x0000ff00ff010000, 0x0000ff00ff010001, 0x0000ff00ff0101ff,
+    0x0000ff00ff010101, 0x0000ff0000ffff00, 0x0000ff0000ff00ff, 0x0000ff0000ff0000,
+    0x0000ff0000ff0001, 0x0000ff0000ff0100, 0x0000ff000000ffff, 0x0000ff000000ff00,
+    0x0000ff000000ff01, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
+    0x0000ff00000001ff, 0x0000ff0000000100, 0x0000ff0000000101, 0x0000ff000001ff00,
+    0x0000ff00000100ff, 0x0000ff0000010000, 0x0000ff0000010001, 0x0000ff0000010100,
+    0x0000ff0001ffff01, 0x0000ff0001ff0000, 0x0000ff000100ff00, 0x0000ff00010000ff,
+    0x0000ff0001000000, 0x0000ff0001000001, 0x0000ff0001000100, 0x0000ff000101ffff,
+    0x0000ff0001010000, 0x0000ff0001010101, 0x0000ff01ffffff00, 0x0000ff01ffff0001,
+    0x0000ff01ff00ff01, 0x0000ff01ff000000, 0x0000ff01ff000101, 0x0000ff01ff01ff00,
+    0x0000ff01ff0100ff, 0x0000ff0100ffff01, 0x0000ff0100ff0000, 0x0000ff0100ff0101,
+    0x0000ff010000ff00, 0x0000ff01000000ff, 0x0000ff0100000000, 0x0000ff0100000001,
+    0x0000ff0100000100, 0x0000ff010001ff01, 0x0000ff0100010000, 0x0000ff0101ff0000,
+    0x0000ff010100ffff, 0x0000ff010100ff01, 0x0000ff0101000000, 0x0000ff0101000100,
+    0x0000ff0101000101, 0x0000ff01010100ff, 0x000000ffffff00ff, 0x000000ffffff0000,
+    0x000000ffff00ff00, 0x000000ffff0000ff, 0x000000ffff000000, 0x000000ffff000001,
+    0x000000ffff0001ff, 0x000000ffff000100, 0x000000ffff01ff00, 0x000000ffff010000,
+    0x000000ffff0101ff, 0x000000ffff010101, 0x000000ff00ffff00, 0x000000ff00ff00ff,
+    0x000000ff00ff0000, 0x000000ff00ff0001, 0x000000ff00ff0100, 0x000000ff00ff0101,
+    0x000000ff0000ffff, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
+    0x000000ff00000001, 0x000000ff000001ff, 0x000000ff00000100, 0x000000ff00000101,
+    0x000000ff0001ff00, 0x000000ff0001ff01, 0x000000ff000100ff, 0x000000ff00010000,
+    0x000000ff00010001, 0x000000ff00010100, 0x000000ff01ffffff, 0x000000ff01ff01ff,
+    0x000000ff01ff0101, 0x000000ff0100ff00, 0x000000ff010000ff, 0x000000ff01000000,
+    0x000000ff01000001, 0x000000ff01000100, 0x000000ff0101ff00, 0x000000ff010100ff,
+    0x000000ff01010000, 0x000000ff01010101, 0x00000000ffffff00, 0x00000000ffffff01,
+    0x00000000ffff00ff, 0x00000000ffff0000, 0x00000000ffff0001, 0x00000000ffff0100,
+    0x00000000ff00ffff, 0x00000000ff00ff00, 0x00000000ff00ff01, 0x00000000ff0000ff,
+    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff000101,
+    0x00000000ff01ff00, 0x00000000ff0100ff, 0x00000000ff010000, 0x00000000ff010001,
+    0x00000000ff010100, 0x0000000000ffffff, 0x0000000000ffff00, 0x0000000000ffff01,
+    0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, 0x0000000000ff01ff,
+    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
+    0x00000000000000ff, 0x0000000000000000, 0x0000000000000001, 0x00000000000001ff,
+    0x0000000000000100, 0x0000000000000101, 0x000000000001ffff, 0x000000000001ff00,
+    0x00000000000100ff, 0x0000000000010000, 0x0000000000010001, 0x00000000000101ff,
+    0x0000000000010100, 0x0000000000010101, 0x0000000001ffff00, 0x0000000001ff00ff,
+    0x0000000001ff0000, 0x0000000001ff0100, 0x0000000001ff0101, 0x000000000100ffff,
+    0x000000000100ff00, 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001,
+    0x00000000010001ff, 0x0000000001000100, 0x000000000101ff00, 0x00000000010100ff,
+    0x0000000001010000, 0x0000000001010001, 0x0000000001010100, 0x00000001ffffffff,
+    0x00000001ffffff00, 0x00000001ffffff01, 0x00000001ffff00ff, 0x00000001ffff0001,
+    0x00000001ffff01ff, 0x00000001ffff0100, 0x00000001ff00ff00, 0x00000001ff0000ff,
+    0x00000001ff000000, 0x00000001ff0001ff, 0x00000001ff000100, 0x00000001ff01ffff,
+    0x00000001ff01ff00, 0x00000001ff01ff01, 0x00000001ff0100ff, 0x00000001ff010000,
+    0x00000001ff010001, 0x00000001ff0101ff, 0x00000001ff010100, 0x0000000100ffff00,
+    0x0000000100ff0000, 0x0000000100ff0001, 0x0000000100ff01ff, 0x0000000100ff0100,
+    0x0000000100ff0101, 0x000000010000ffff, 0x000000010000ff00, 0x000000010000ff01,
+    0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, 0x00000001000001ff,
+    0x0000000100000100, 0x0000000100000101, 0x000000010001ff00, 0x00000001000100ff,
+    0x0000000100010000, 0x0000000100010100, 0x0000000101ffff01, 0x0000000101ff0000,
+    0x0000000101ff0001, 0x0000000101ff01ff, 0x0000000101ff0100, 0x0000000101ff0101,
+    0x000000010100ff00, 0x0000000101000000, 0x0000000101000101, 0x000000010101ff01,
+    0x0000000101010000, 0x0000000101010001, 0x00000001010101ff, 0x0000000101010100,
+    0x000001ffffff00ff, 0x000001ffffff0000, 0x000001ffffff0001, 0x000001ffffff0100,
+    0x000001ffff00ffff, 0x000001ffff000000, 0x000001ffff0001ff, 0x000001ffff01ff00,
+    0x000001ffff010101, 0x000001ff00ff0000, 0x000001ff00ff01ff, 0x000001ff00ff0101,
+    0x000001ff0000ff00, 0x000001ff000000ff, 0x000001ff00000000, 0x000001ff00000001,
+    0x000001ff000001ff, 0x000001ff00000100, 0x000001ff0001ffff, 0x000001ff0001ff01,
+    0x000001ff000100ff, 0x000001ff00010000, 0x000001ff01ffff01, 0x000001ff01ff0100,
+    0x000001ff0100ffff, 0x000001ff0100ff01, 0x000001ff01000000, 0x000001ff010001ff,
+    0x000001ff0101ff00, 0x000001ff01010100, 0x00000100ffffff00, 0x00000100ffffff01,
+    0x00000100ffff0000, 0x00000100ffff0101, 0x00000100ff00ff00, 0x00000100ff0000ff,
+    0x00000100ff000000, 0x00000100ff000001, 0x00000100ff000100, 0x00000100ff010000,
+    0x0000010000ffff00, 0x0000010000ff00ff, 0x0000010000ff0000, 0x0000010000ff0001,
+    0x0000010000ff0100, 0x000001000000ffff, 0x000001000000ff00, 0x000001000000ff01,
+    0x00000100000000ff, 0x0000010000000000, 0x0000010000000001, 0x00000100000001ff,
+    0x0000010000000100, 0x0000010000000101, 0x000001000001ff00, 0x00000100000100ff,
+    0x0000010000010000, 0x0000010000010001, 0x0000010000010100, 0x0000010001ffff00,
+    0x0000010001ff0000, 0x0000010001ff0100, 0x000001000100ff00, 0x00000100010000ff,
+    0x0000010001000000, 0x0000010001000001, 0x00000100010001ff, 0x0000010001000100,
+    0x0000010001010000, 0x00000101ffff00ff, 0x00000101ffff01ff, 0x00000101ff000000,
+    0x00000101ff000101, 0x00000101ff01ffff, 0x00000101ff010000, 0x00000101ff010001,
+    0x00000101ff010100, 0x0000010100ff0000, 0x0000010100ff01ff, 0x0000010100ff0100,
+    0x000001010000ff00, 0x0000010100000000, 0x0000010100000001, 0x00000101000001ff,
+    0x0000010100000100, 0x000001010001ff01, 0x0000010100010000, 0x00000101000101ff,
+    0x0000010100010101, 0x0000010101ffff00, 0x0000010101ff0101, 0x000001010100ff01,
+    0x0000010101000000, 0x0000010101000001, 0x00000101010001ff, 0x0000010101000101,
+    0x000001010101ff00, 0x0001ffffffff0000, 0x0001ffffff0000ff, 0x0001ffffff000001,
+    0x0001ffffff000100, 0x0001ffffff010000, 0x0001ffff00ff00ff, 0x0001ffff0000ffff,
+    0x0001ffff00000000, 0x0001ffff00000001, 0x0001ffff000001ff, 0x0001ffff00000101,
+    0x0001ffff0001ff00, 0x0001ffff000100ff, 0x0001ffff00010001, 0x0001ffff00010100,
+    0x0001ffff01ffff00, 0x0001ffff01000001, 0x0001ffff01010000, 0x0001ff00ffffff00,
+    0x0001ff00ffff00ff, 0x0001ff00ffff0001, 0x0001ff00ffff0100, 0x0001ff00ff00ff01,
+    0x0001ff00ff000000, 0x0001ff00ff01ff00, 0x0001ff00ff01ff01, 0x0001ff00ff010001,
+    0x0001ff00ff010100, 0x0001ff0000ff0000, 0x0001ff0000ff0100, 0x0001ff000000ff00,
+    0x0001ff0000000000, 0x0001ff0000000001, 0x0001ff0000000100, 0x0001ff0000010000,
+    0x0001ff0000010001, 0x0001ff0000010101, 0x0001ff0001ff00ff, 0x0001ff0001ff0101,
+    0x0001ff000100ff01, 0x0001ff0001000000, 0x0001ff000101ff00, 0x0001ff0001010001,
+    0x0001ff0001010100, 0x0001ff01ff00ff00, 0x0001ff01ff000001, 0x0001ff01ff000100,
+    0x0001ff0100ffffff, 0x0001ff0100ffff00, 0x0001ff0100ff0001, 0x0001ff0100000000,
+    0x0001ff0100000001, 0x0001ff01000001ff, 0x0001ff010001ffff, 0x0001ff0101ff0000,
+    0x0001ff010100ff00, 0x0001ff0101000001, 0x0001ff0101010000, 0x000100ffff00ff00,
+    0x000100ffff00ff01, 0x000100ffff000000, 0x000100ffff000001, 0x000100ffff000101,
+    0x000100ffff01ff00, 0x000100ffff010001, 0x000100ffff010100, 0x000100ff00ffffff,
+    0x000100ff00ffff01, 0x000100ff00ff0000, 0x000100ff00ff01ff, 0x000100ff00ff0101,
+    0x000100ff0000ff00, 0x000100ff000000ff, 0x000100ff00000000, 0x000100ff00000001,
+    0x000100ff00000100, 0x000100ff00000101, 0x000100ff0001ffff, 0x000100ff0001ff01,
+    0x000100ff00010000, 0x000100ff01ff00ff, 0x000100ff01ff0000, 0x000100ff01ff0100,
+    0x000100ff0100ffff, 0x000100ff0100ff01, 0x000100ff010000ff, 0x000100ff01000000,
+    0x000100ff01000001, 0x000100ff010001ff, 0x000100ff01000101, 0x000100ff0101ff00,
+    0x000100ff010100ff, 0x000100ff01010100, 0x00010000ffff0000, 0x00010000ffff01ff,
+    0x00010000ffff0101, 0x00010000ff00ff00, 0x00010000ff000000, 0x00010000ff000001,
+    0x00010000ff000100, 0x0001000000ff00ff, 0x0001000000ff0000, 0x0001000000ff0001,
+    0x0001000000ff0100, 0x000100000000ffff, 0x000100000000ff00, 0x00010000000000ff,
+    0x0001000000000000, 0x0001000000000001, 0x0001000000000100, 0x000100000001ff00,
+    0x00010000000100ff, 0x0001000000010000, 0x0001000000010001, 0x0001000000010100,
+    0x0001000001ff0001, 0x0001000001ff0100, 0x0001000001ff0101, 0x000100000100ff00,
+    0x0001000001000000, 0x0001000001000001, 0x0001000001000100, 0x0001000001000101,
+    0x000100000101ff01, 0x0001000001010000, 0x0001000001010001, 0x00010000010101ff,
+    0x00010001ffffff01, 0x00010001ffff0100, 0x00010001ff000000, 0x00010001ff01ffff,
+    0x00010001ff010001, 0x00010001ff0101ff, 0x00010001ff010100, 0x0001000100ffffff,
+    0x0001000100ff0000, 0x0001000100ff01ff, 0x0001000100ff0101, 0x000100010000ff00,
+    0x00010001000000ff, 0x0001000100000000, 0x0001000100000001, 0x00010001000001ff,
+    0x0001000100000101, 0x000100010001ffff, 0x0001000100010000, 0x00010001000101ff,
+    0x0001000101ffffff, 0x0001000101ffff01, 0x0001000101ff0000, 0x0001000101ff0101,
+    0x00010001010000ff, 0x0001000101000001, 0x00010001010001ff, 0x0001000101000100,
+    0x000100010101ffff, 0x00010001010100ff, 0x0001000101010001, 0x0001000101010101,
+    0x000101ffff000001, 0x000101ffff000100, 0x000101ffff010000, 0x000101ff00ffff00,
+    0x000101ff0000ff01, 0x000101ff00000000, 0x000101ff00000101, 0x000101ff0001ff00,
+    0x000101ff00010100, 0x000101ff01ff0000, 0x000101ff0100ff00, 0x000101ff010001ff,
+    0x000101ff01010001, 0x00010100ffffff00, 0x00010100ffff00ff, 0x00010100ff00ffff,
+    0x00010100ff000000, 0x00010100ff01ff00, 0x00010100ff0100ff, 0x00010100ff010001,
+    0x00010100ff010100, 0x0001010000ffffff, 0x0001010000ffff00, 0x0001010000ff0000,
+    0x0001010000ff0001, 0x0001010000ff01ff, 0x000101000000ff00, 0x00010100000000ff,
+    0x0001010000000000, 0x0001010000000001, 0x0001010000000100, 0x000101000001ffff,
+    0x0001010000010000, 0x0001010000010101, 0x0001010001ffff01, 0x0001010001ff00ff,
+    0x0001010001ff0101, 0x0001010001000000, 0x000101000101ff00, 0x00010100010100ff,
+    0x0001010001010000, 0x0001010001010100, 0x00010101ff00ff00, 0x00010101ff000001,
+    0x00010101ff0001ff, 0x0001010100ffff00, 0x0001010100ff00ff, 0x0001010100ff0100,
+    0x000101010000ffff, 0x0001010100000000, 0x00010101000001ff, 0x0001010100000101,
+    0x00010101000100ff, 0x0001010100010000, 0x0001010100010100, 0x0001010101ff0001,
+    0x00010101010000ff, 0x00010101010001ff, 0x0001010101000101, 0x0001010101010001,
+    0x01ffffffffffffff, 0x01ffffffffffff01, 0x01ffffffffff01ff, 0x01ffffffffff0101,
+    0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, 0x01ffffffff010101,
+    0x01ffffff00ff0000, 0x01ffffff0000ffff, 0x01ffffff0000ff00, 0x01ffffff000000ff,
+    0x01ffffff00000001, 0x01ffffff00000100, 0x01ffffff00010000, 0x01ffffff01ffffff,
+    0x01ffffff01ffff01, 0x01ffffff01ff01ff, 0x01ffffff01ff0101, 0x01ffffff01000000,
+    0x01ffffff0101ffff, 0x01ffffff0101ff01, 0x01ffffff010101ff, 0x01ffffff01010101,
+    0x01ffff00ffff0000, 0x01ffff00ff00ff00, 0x01ffff00ff0000ff, 0x01ffff00ff000001,
+    0x01ffff00ff000100, 0x01ffff00ff010000, 0x01ffff0000ffff00, 0x01ffff0000ff00ff,
+    0x01ffff0000ff0100, 0x01ffff000000ffff, 0x01ffff000000ff01, 0x01ffff0000000000,
+    0x01ffff0000000001, 0x01ffff00000001ff, 0x01ffff0000000100, 0x01ffff00000100ff,
+    0x01ffff0000010001, 0x01ffff0000010100, 0x01ffff0001ff0000, 0x01ffff0001ff0100,
+    0x01ffff00010000ff, 0x01ffff0001000001, 0x01ffff0001000100, 0x01ffff0001010000,
+    0x01ffff01ffffffff, 0x01ffff01ffffff01, 0x01ffff01ffff01ff, 0x01ffff01ffff0101,
+    0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff01ff01, 0x01ffff01ff0101ff,
+    0x01ffff01ff010101, 0x01ffff010000ff00, 0x01ffff01000000ff, 0x01ffff0100000100,
+    0x01ffff0100010000, 0x01ffff0101ffffff, 0x01ffff0101ffff01, 0x01ffff0101ff01ff,
+    0x01ffff0101ff0101, 0x01ffff0101000000, 0x01ffff010101ffff, 0x01ffff010101ff01,
+    0x01ffff01010101ff, 0x01ffff0101010101, 0x01ff00ffff0000ff, 0x01ff00ffff000100,
+    0x01ff00ff00ffff00, 0x01ff00ff00ff00ff, 0x01ff00ff0000ff00, 0x01ff00ff00000000,
+    0x01ff00ff00000101, 0x01ff00ff0001ff00, 0x01ff00ff000100ff, 0x01ff00ff00010100,
+    0x01ff00ff010000ff, 0x01ff00ff01000100, 0x01ff0000ffffff00, 0x01ff0000ffff0100,
+    0x01ff0000ff00ff01, 0x01ff0000ff000000, 0x01ff0000ff000101, 0x01ff0000ff010001,
+    0x01ff0000ff010100, 0x01ff000000ffffff, 0x01ff000000ffff00, 0x01ff000000ff0000,
+    0x01ff000000ff01ff, 0x01ff00000000ff00, 0x01ff0000000000ff, 0x01ff000000000000,
+    0x01ff000000000001, 0x01ff000000000100, 0x01ff000000000101, 0x01ff000000010000,
+    0x01ff000000010001, 0x01ff0000000101ff, 0x01ff000000010101, 0x01ff000001ffff00,
+    0x01ff000001ff00ff, 0x01ff000001ff0001, 0x01ff000001ff0100, 0x01ff00000100ffff,
+    0x01ff00000100ff01, 0x01ff000001000000, 0x01ff0000010001ff, 0x01ff000001010001,
+    0x01ff0001ff00ff00, 0x01ff0001ff000001, 0x01ff0001ff000100, 0x01ff0001ff010000,
+    0x01ff000100ffff00, 0x01ff000100ff00ff, 0x01ff000100ff0100, 0x01ff000100ff0101,
+    0x01ff00010000ffff, 0x01ff000100000000, 0x01ff000100000100, 0x01ff000100000101,
+    0x01ff00010001ff00, 0x01ff000100010001, 0x01ff000100010101, 0x01ff000101ff0000,
+    0x01ff00010100ff00, 0x01ff000101000101, 0x01ff0001010100ff, 0x01ff01ffffffffff,
+    0x01ff01ffffffff01, 0x01ff01ffffff01ff, 0x01ff01ffffff0101, 0x01ff01ffff000000,
+    0x01ff01ffff01ffff, 0x01ff01ffff01ff01, 0x01ff01ffff0101ff, 0x01ff01ffff010101,
+    0x01ff01ff00ffff00, 0x01ff01ff00ff0000, 0x01ff01ff0000ff00, 0x01ff01ff000000ff,
+    0x01ff01ff00000100, 0x01ff01ff00010000, 0x01ff01ff00010100, 0x01ff01ff01ffffff,
+    0x01ff01ff01ffff01, 0x01ff01ff01ff01ff, 0x01ff01ff01ff0101, 0x01ff01ff01000000,
+    0x01ff01ff0101ffff, 0x01ff01ff0101ff01, 0x01ff01ff010101ff, 0x01ff01ff01010101,
+    0x01ff0100ffff0000, 0x01ff0100ffff0001, 0x01ff0100ff00ff00, 0x01ff0100ff0000ff,
+    0x01ff0100ff000001, 0x01ff0100ff010000, 0x01ff010000ffff00, 0x01ff010000ff00ff,
+    0x01ff010000ff0001, 0x01ff010000ff0100, 0x01ff01000000ffff, 0x01ff01000000ff01,
+    0x01ff010000000000, 0x01ff010000000101, 0x01ff01000001ff00, 0x01ff0100000100ff,
+    0x01ff010001ff0000, 0x01ff010001000001, 0x01ff010001000100, 0x01ff010001010000,
+    0x01ff0101ffffffff, 0x01ff0101ffffff01, 0x01ff0101ffff01ff, 0x01ff0101ffff0101,
+    0x01ff0101ff000000, 0x01ff0101ff01ffff, 0x01ff0101ff01ff01, 0x01ff0101ff0101ff,
+    0x01ff0101ff010101, 0x01ff010100ff0000, 0x01ff01010000ff00, 0x01ff0101000000ff,
+    0x01ff010100000001, 0x01ff010101ffffff, 0x01ff010101ffff01, 0x01ff010101ff01ff,
+    0x01ff010101ff0101, 0x01ff010101000000, 0x01ff01010101ffff, 0x01ff01010101ff01,
+    0x01ff0101010101ff, 0x01ff010101010101, 0x0100ffffffff0000, 0x0100ffffff00ff00,
+    0x0100ffffff000001, 0x0100ffffff0001ff, 0x0100ffffff000100, 0x0100ffffff010000,
+    0x0100ffff00ffff00, 0x0100ffff00ff0001, 0x0100ffff00ff0100, 0x0100ffff00000000,
+    0x0100ffff000001ff, 0x0100ffff00000101, 0x0100ffff00010100, 0x0100ffff00010101,
+    0x0100ffff01ff0000, 0x0100ffff0100ff00, 0x0100ffff010000ff, 0x0100ffff01000001,
+    0x0100ffff01000100, 0x0100ffff01010000, 0x0100ff00ffffff00, 0x0100ff00ffff00ff,
+    0x0100ff00ffff0001, 0x0100ff00ffff0100, 0x0100ff00ff00ffff, 0x0100ff00ff000000,
+    0x0100ff00ff0001ff, 0x0100ff00ff000101, 0x0100ff00ff01ff00, 0x0100ff00ff0100ff,
+    0x0100ff00ff010001, 0x0100ff00ff010100, 0x0100ff0000ffffff, 0x0100ff0000ff0000,
+    0x0100ff000000ffff, 0x0100ff000000ff00, 0x0100ff00000000ff, 0x0100ff0000000000,
+    0x0100ff0000000001, 0x0100ff0000000100, 0x0100ff000001ff01, 0x0100ff0000010000,
+    0x0100ff0001ff00ff, 0x0100ff0001ff0001, 0x0100ff000100ff01, 0x0100ff0001000000,
+    0x0100ff00010001ff, 0x0100ff000101ff00, 0x0100ff00010100ff, 0x0100ff0001010001,
+    0x0100ff0001010100, 0x0100ff01ffff0000, 0x0100ff01ff00ff00, 0x0100ff01ff0000ff,
+    0x0100ff01ff000100, 0x0100ff01ff010000, 0x0100ff0100ff00ff, 0x0100ff0100ff0001,
+    0x0100ff0100ff0100, 0x0100ff010000ffff, 0x0100ff010000ff01, 0x0100ff0100000000,
+    0x0100ff01000001ff, 0x0100ff0100010001, 0x0100ff0100010100, 0x0100ff0101ff0000,
+    0x0100ff01010000ff, 0x0100ff0101000001, 0x0100ff0101010100, 0x010000ffffffff00,
+    0x010000ffffff00ff, 0x010000ffffff0001, 0x010000ffff00ffff, 0x010000ffff000000,
+    0x010000ffff0001ff, 0x010000ffff010001, 0x010000ff00ffffff, 0x010000ff00ff0101,
+    0x010000ff0000ff00, 0x010000ff000000ff, 0x010000ff00000000, 0x010000ff00000001,
+    0x010000ff000001ff, 0x010000ff00000100, 0x010000ff0001ffff, 0x010000ff0001ff00,
+    0x010000ff0001ff01, 0x010000ff00010000, 0x010000ff01ff00ff, 0x010000ff01ff0001,
+    0x010000ff0100ff01, 0x010000ff010000ff, 0x010000ff01000000, 0x010000ff010001ff,
+    0x010000ff0101ff00, 0x010000ff01010100, 0x01000000ffffffff, 0x01000000ffff0000,
+    0x01000000ffff01ff, 0x01000000ffff0101, 0x01000000ff00ffff, 0x01000000ff00ff00,
+    0x01000000ff0000ff, 0x01000000ff000000, 0x01000000ff000001, 0x01000000ff000100,
+    0x01000000ff01ff00, 0x01000000ff010000, 0x01000000ff010100, 0x01000000ff010101,
+    0x0100000000ffff00, 0x0100000000ff00ff, 0x0100000000ff0000, 0x0100000000ff0001,
+    0x0100000000ff0100, 0x010000000000ffff, 0x010000000000ff00, 0x010000000000ff01,
+    0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, 0x01000000000001ff,
+    0x0100000000000100, 0x0100000000000101, 0x010000000001ff00, 0x01000000000100ff,
+    0x0100000000010000, 0x0100000000010001, 0x0100000000010100, 0x0100000001ffff00,
+    0x0100000001ff0000, 0x0100000001ff01ff, 0x010000000100ff00, 0x010000000100ff01,
+    0x01000000010000ff, 0x0100000001000000, 0x0100000001000001, 0x0100000001000100,
+    0x0100000001000101, 0x010000000101ffff, 0x010000000101ff01, 0x0100000001010000,
+    0x01000000010101ff, 0x0100000001010101, 0x01000001ffffff00, 0x01000001ffff00ff,
+    0x01000001ff00ffff, 0x01000001ff000000, 0x01000001ff000100, 0x01000001ff01ffff,
+    0x01000001ff010001, 0x01000001ff010100, 0x0100000100ff0000, 0x0100000100ff01ff,
+    0x0100000100ff0100, 0x010000010000ff00, 0x010000010000ff01, 0x0100000100000000,
+    0x0100000100000001, 0x0100000100000100, 0x0100000100010000, 0x01000001000101ff,
+    0x0100000101ffff01, 0x0100000101ff00ff, 0x0100000101ff0100, 0x0100000101ff0101,
+    0x010000010100ff01, 0x01000001010000ff, 0x0100000101000000, 0x01000001010100ff,
+    0x0100000101010001, 0x0100000101010100, 0x010001ffffff0000, 0x010001ffff000001,
+    0x010001ffff000100, 0x010001ffff010000, 0x010001ff00ffff00, 0x010001ff00ff0001,
+    0x010001ff0000ffff, 0x010001ff0000ff01, 0x010001ff00000000, 0x010001ff00000001,
+    0x010001ff00000101, 0x010001ff000100ff, 0x010001ff00010000, 0x010001ff01ff0000,
+    0x010001ff0100ff00, 0x010001ff01000001, 0x010001ff01000100, 0x010001ff01010000,
+    0x01000100ffff00ff, 0x01000100ffff0001, 0x01000100ffff0100, 0x01000100ff00ffff,
+    0x01000100ff00ff01, 0x01000100ff000000, 0x01000100ff0001ff, 0x01000100ff000101,
+    0x01000100ff01ffff, 0x01000100ff01ff00, 0x01000100ff0100ff, 0x01000100ff010001,
+    0x0100010000ffffff, 0x0100010000ffff01, 0x0100010000ff0000, 0x0100010000ff01ff,
+    0x0100010000ff0101, 0x010001000000ff00, 0x01000100000000ff, 0x0100010000000000,
+    0x0100010000000001, 0x0100010000000100, 0x010001000001ff01, 0x0100010000010000,
+    0x0100010000010001, 0x0100010000010101, 0x0100010001ffff00, 0x0100010001ff00ff,
+    0x010001000100ffff, 0x010001000100ff01, 0x0100010001000000, 0x0100010001000101,
+    0x010001000101ff00, 0x0100010001010001, 0x01000101ffff0000, 0x01000101ff000000,
+    0x01000101ff010000, 0x0100010100ff00ff, 0x0100010100ff0001, 0x0100010100ff0100,
+    0x010001010000ffff, 0x0100010100000000, 0x01000101000001ff, 0x010001010001ff00,
+    0x0100010101ff0000, 0x010001010100ff00, 0x01000101010000ff, 0x0100010101000000,
+    0x0100010101000001, 0x0101ffffffffffff, 0x0101ffffffffff01, 0x0101ffffffff01ff,
+    0x0101ffffffff0101, 0x0101ffffff000000, 0x0101ffffff01ffff, 0x0101ffffff01ff01,
+    0x0101ffffff0101ff, 0x0101ffffff010101, 0x0101ffff00ff0000, 0x0101ffff0000ff00,
+    0x0101ffff000000ff, 0x0101ffff00000001, 0x0101ffff00000100, 0x0101ffff01ffffff,
+    0x0101ffff01ffff01, 0x0101ffff01ff01ff, 0x0101ffff01ff0101, 0x0101ffff01000000,
+    0x0101ffff0101ffff, 0x0101ffff0101ff01, 0x0101ffff010101ff, 0x0101ffff01010101,
+    0x0101ff00ffff0000, 0x0101ff00ffff0100, 0x0101ff00ff00ff00, 0x0101ff00ff0000ff,
+    0x0101ff00ff000001, 0x0101ff00ff000100, 0x0101ff00ff000101, 0x0101ff0000ff0001,
+    0x0101ff0000ff0100, 0x0101ff000000ff00, 0x0101ff0000000000, 0x0101ff00000001ff,
+    0x0101ff0000000101, 0x0101ff000001ff00, 0x0101ff00000100ff, 0x0101ff0001ff0000,
+    0x0101ff000100ffff, 0x0101ff000100ff01, 0x0101ff0001000001, 0x0101ff0001000100,
+    0x0101ff01ffffff01, 0x0101ff01ffff01ff, 0x0101ff01ffff0101, 0x0101ff01ff00ffff,
+    0x0101ff01ff000100, 0x0101ff01ff01ff01, 0x0101ff01ff0101ff, 0x0101ff01ff010101,
+    0x0101ff0100ff0000, 0x0101ff010000ff00, 0x0101ff0100000001, 0x0101ff0100000100,
+    0x0101ff0100010000, 0x0101ff0101ffffff, 0x0101ff0101ffff01, 0x0101ff0101ff01ff,
+    0x0101ff0101ff0101, 0x0101ff0101000000, 0x0101ff010101ffff, 0x0101ff010101ff01,
+    0x0101ff01010101ff, 0x0101ff0101010101, 0x010100ffff000100, 0x010100ffff010000,
+    0x010100ff00ffff00, 0x010100ff00ff00ff, 0x010100ff0000ffff, 0x010100ff000000ff,
+    0x010100ff00000000, 0x010100ff000001ff, 0x010100ff00000101, 0x010100ff0001ff00,
+    0x010100ff00010000, 0x010100ff00010001, 0x010100ff000101ff, 0x010100ff00010100,
+    0x010100ff01ff0000, 0x01010000ffff0001, 0x01010000ffff0100, 0x01010000ff00ffff,
+    0x01010000ff00ff01, 0x01010000ff000000, 0x01010000ff0001ff, 0x01010000ff010001,
+    0x01010000ff010100, 0x0101000000ffff01, 0x0101000000ff0000, 0x010100000000ff00,
+    0x01010000000000ff, 0x0101000000000000, 0x0101000000000001, 0x0101000000000100,
+    0x0101000000010000, 0x0101000000010101, 0x0101000001ffff00, 0x0101000001ff00ff,
+    0x0101000001ff0000, 0x0101000001ff0001, 0x0101000001ff0100, 0x010100000100ff01,
+    0x0101000001000000, 0x01010000010001ff, 0x01010001ffff0000, 0x01010001ff00ff00,
+    0x01010001ff000001, 0x01010001ff000101, 0x01010001ff01ff00, 0x01010001ff010000,
+    0x0101000100ff00ff, 0x0101000100ff0001, 0x0101000100ff0101, 0x010100010000ff01,
+    0x0101000100000000, 0x0101000100000001, 0x01010001000001ff, 0x010100010001ffff,
+    0x010100010001ff01, 0x0101000101ff0001, 0x010100010100ffff, 0x0101000101000000,
+    0x0101000101000001, 0x0101000101000100, 0x010100010101ff00, 0x01010001010100ff,
+    0x0101000101010001, 0x010101ffffffffff, 0x010101ffffffff01, 0x010101ffffff01ff,
+    0x010101ffffff0101, 0x010101ffff01ffff, 0x010101ffff01ff01, 0x010101ffff0101ff,
+    0x010101ffff010101, 0x010101ff0000ff00, 0x010101ff000000ff, 0x010101ff00000001,
+    0x010101ff00000100, 0x010101ff01ffffff, 0x010101ff01ffff01, 0x010101ff01ff01ff,
+    0x010101ff01ff0101, 0x010101ff01000000, 0x010101ff0101ffff, 0x010101ff0101ff01,
+    0x010101ff010101ff, 0x010101ff01010101, 0x01010100ffff0000, 0x01010100ff0000ff,
+    0x01010100ff000100, 0x01010100ff01ff00, 0x01010100ff010000, 0x0101010000ffff00,
+    0x010101000000ffff, 0x0101010000000000, 0x0101010000000101, 0x010101000001ff00,
+    0x0101010000010001, 0x0101010000010100, 0x010101000100ffff, 0x0101010001000001,
+    0x01010101ffffffff, 0x01010101ffffff01, 0x01010101ffff01ff, 0x01010101ffff0101,
+    0x01010101ff01ffff, 0x01010101ff01ff01, 0x01010101ff0101ff, 0x01010101ff010101,
+    0x010101010000ff00, 0x01010101000000ff, 0x0101010100000001, 0x0101010101ffffff,
+    0x0101010101ffff01, 0x0101010101ff01ff, 0x0101010101ff0101, 0x0101010101000000,
+    0x010101010101ffff, 0x010101010101ff01, 0x01010101010101ff, 0x0101010101010101,
+GGML_TABLE_END()
+#else
+GGML_TABLE_BEGIN(uint32_t, iq1s_grid_gpu, NGRID_IQ1S)
+    0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
+    0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
+    0x02000000, 0x02000002, 0x02000200, 0x02000202, 0x02010101, 0x02020000, 0x02020002, 0x02020200,
+    0x02020202, 0x00000110, 0x00000111, 0x00010011, 0x00010110, 0x00010112, 0x00010211, 0x00010212,
+    0x00020111, 0x01000011, 0x01000112, 0x01000211, 0x01010012, 0x01010111, 0x01010212, 0x01020011,
+    0x01020110, 0x01020112, 0x01020210, 0x02000111, 0x02010011, 0x02010110, 0x02010112, 0x02020111,
+    0x00000020, 0x00000022, 0x00000220, 0x00000222, 0x00010121, 0x00020020, 0x00020022, 0x00020220,
+    0x00020222, 0x01000121, 0x01010021, 0x01010221, 0x01020120, 0x01020221, 0x02000020, 0x02000022,
+    0x02000220, 0x02000222, 0x02010021, 0x02010121, 0x02010221, 0x02020020, 0x02020022, 0x02020220,
+    0x02020222, 0x00011001, 0x00011100, 0x00011102, 0x00021101, 0x01001001, 0x01001201, 0x01011101,
+    0x01011202, 0x01021100, 0x01021101, 0x02011001, 0x02011201, 0x02021101, 0x00001011, 0x00001110,
+    0x00001111, 0x00001112, 0x00011111, 0x00011210, 0x00011212, 0x00021211, 0x01001010, 0x01001111,
+    0x01001212, 0x01011010, 0x01011011, 0x01011110, 0x01011111, 0x01011112, 0x01011211, 0x01021010,
+    0x01021012, 0x01021111, 0x01021210, 0x01021212, 0x02001011, 0x02011011, 0x02011111, 0x02011210,
+    0x02011212, 0x02021011, 0x02021110, 0x02021111, 0x02021112, 0x02021211, 0x00011120, 0x00011221,
+    0x01001021, 0x01001120, 0x01011020, 0x01011022, 0x01011121, 0x01011220, 0x01021020, 0x01021021,
+    0x01021122, 0x01021221, 0x02001121, 0x02011021, 0x02011120, 0x02011221, 0x00002000, 0x00002002,
+    0x00002200, 0x00002202, 0x00012101, 0x00022000, 0x00022002, 0x00022200, 0x00022202, 0x01002101,
+    0x01012001, 0x01012102, 0x01022101, 0x02002000, 0x02002002, 0x02002200, 0x02002202, 0x02012101,
+    0x02022000, 0x02022002, 0x02022200, 0x02022202, 0x00002111, 0x00012011, 0x00012110, 0x00012211,
+    0x00022110, 0x00022111, 0x01002011, 0x01012010, 0x01012011, 0x01012111, 0x01022011, 0x01022110,
+    0x01022211, 0x02012011, 0x02012110, 0x02012112, 0x02012211, 0x02022111, 0x00002020, 0x00002022,
+    0x00002220, 0x00002222, 0x00012121, 0x00022020, 0x00022022, 0x00022220, 0x00022222, 0x01002121,
+    0x01012021, 0x01012221, 0x01022021, 0x01022121, 0x02002020, 0x02002022, 0x02002121, 0x02002220,
+    0x02002222, 0x02012121, 0x02022020, 0x02022022, 0x02022220, 0x02022222, 0x00110000, 0x00110001,
+    0x00110100, 0x00110201, 0x00120100, 0x00120101, 0x01100001, 0x01100100, 0x01110000, 0x01110101,
+    0x01110200, 0x01120001, 0x01120100, 0x01120101, 0x01120201, 0x02110001, 0x02110100, 0x02110102,
+    0x02120001, 0x02120101, 0x00100011, 0x00100110, 0x00100112, 0x00100211, 0x00110010, 0x00110012,
+    0x00110111, 0x00110210, 0x00120011, 0x00120110, 0x00120211, 0x01100111, 0x01100212, 0x01110010,
+    0x01110011, 0x01110012, 0x01110110, 0x01110111, 0x01110112, 0x01110211, 0x01120010, 0x01120111,
+    0x02100110, 0x02110012, 0x02110111, 0x02120011, 0x02120110, 0x00110021, 0x00110120, 0x00110122,
+    0x00120121, 0x01100020, 0x01100122, 0x01100221, 0x01110022, 0x01110121, 0x01110220, 0x01110222,
+    0x01120120, 0x01120122, 0x02100121, 0x02110021, 0x02110120, 0x02110122, 0x02120121, 0x00101001,
+    0x00101102, 0x00101201, 0x00111100, 0x00111101, 0x00111200, 0x00111201, 0x00121001, 0x00121102,
+    0x01101001, 0x01101101, 0x01101102, 0x01101200, 0x01101202, 0x01111001, 0x01111100, 0x01111101,
+    0x01111102, 0x01111201, 0x01121002, 0x01121101, 0x01121200, 0x02101100, 0x02101201, 0x02111000,
+    0x02111100, 0x02111101, 0x02111200, 0x02111201, 0x02111202, 0x02121001, 0x02121100, 0x02121101,
+    0x02121201, 0x00101012, 0x00101111, 0x00101212, 0x00111011, 0x00111110, 0x00111111, 0x00111112,
+    0x00111211, 0x00121010, 0x00121012, 0x00121111, 0x00121210, 0x00121212, 0x01101011, 0x01101110,
+    0x01101111, 0x01101112, 0x01111011, 0x01111012, 0x01111110, 0x01111111, 0x01111112, 0x01111211,
+    0x01111212, 0x01121011, 0x01121110, 0x01121111, 0x01121112, 0x01121211, 0x02101010, 0x02101012,
+    0x02101110, 0x02101111, 0x02101210, 0x02101212, 0x02111010, 0x02111011, 0x02111110, 0x02111111,
+    0x02111112, 0x02111211, 0x02111212, 0x02121010, 0x02121012, 0x02121111, 0x00101021, 0x00101120,
+    0x00101121, 0x00101122, 0x00111121, 0x00111122, 0x00111220, 0x00111222, 0x00121021, 0x00121122,
+    0x01101020, 0x01101022, 0x01101120, 0x01101121, 0x01101220, 0x01101222, 0x01111021, 0x01111121,
+    0x01111122, 0x01111220, 0x01111221, 0x01121021, 0x01121120, 0x01121121, 0x01121220, 0x01121221,
+    0x01121222, 0x02101122, 0x02101222, 0x02111022, 0x02111121, 0x02121120, 0x02121221, 0x00112001,
+    0x00112102, 0x00122101, 0x01102001, 0x01102100, 0x01102102, 0x01102201, 0x01112000, 0x01112101,
+    0x01112200, 0x01112202, 0x01122000, 0x01122001, 0x01122100, 0x01122102, 0x01122201, 0x02102101,
+    0x02112001, 0x02112100, 0x02122101, 0x00112010, 0x00112012, 0x00112111, 0x00112212, 0x00122011,
+    0x00122111, 0x01102012, 0x01102110, 0x01102111, 0x01102210, 0x01112011, 0x01112110, 0x01112111,
+    0x01112112, 0x01112211, 0x01112212, 0x01122010, 0x01122111, 0x01122212, 0x02102211, 0x02112011,
+    0x02112012, 0x02112111, 0x02112210, 0x02122011, 0x02122112, 0x02122211, 0x00102221, 0x00112122,
+    0x00122120, 0x00122122, 0x01102120, 0x01102122, 0x01102221, 0x01112020, 0x01112022, 0x01112121,
+    0x01112220, 0x01122021, 0x01122122, 0x01122221, 0x02102121, 0x02112021, 0x02112122, 0x02112222,
+    0x00200000, 0x00200002, 0x00200200, 0x00200202, 0x00210101, 0x00220000, 0x00220002, 0x00220101,
+    0x00220200, 0x00220202, 0x01200101, 0x01210001, 0x01210201, 0x01220001, 0x01220101, 0x02200000,
+    0x02200002, 0x02200200, 0x02200202, 0x02210101, 0x02220000, 0x02220002, 0x02220101, 0x02220200,
+    0x02220202, 0x00200111, 0x00210011, 0x00210110, 0x00210211, 0x00220111, 0x01200012, 0x01200110,
+    0x01200211, 0x01210111, 0x01210210, 0x01210212, 0x01220011, 0x01220110, 0x01220111, 0x01220112,
+    0x02200111, 0x02210010, 0x02210112, 0x02210211, 0x02220111, 0x00200021, 0x00200220, 0x00200222,
+    0x00210021, 0x00210121, 0x00220020, 0x00220022, 0x00220220, 0x00220222, 0x01200121, 0x01210021,
+    0x01210122, 0x01210221, 0x01220121, 0x02200021, 0x02200220, 0x02200222, 0x02210021, 0x02210121,
+    0x02220020, 0x02220022, 0x02220220, 0x02220222, 0x00201101, 0x00211100, 0x00211102, 0x00211201,
+    0x00221101, 0x01201100, 0x01201101, 0x01201102, 0x01201201, 0x01211002, 0x01211101, 0x01211200,
+    0x01211202, 0x01221102, 0x02201101, 0x02211001, 0x02211100, 0x02211201, 0x02221001, 0x02221101,
+    0x00201211, 0x00211111, 0x00221011, 0x00221211, 0x01201010, 0x01201111, 0x01201210, 0x01211011,
+    0x01211110, 0x01211111, 0x01211211, 0x01221012, 0x01221111, 0x01221210, 0x02201211, 0x02211010,
+    0x02211110, 0x02211111, 0x02211210, 0x02211212, 0x02221011, 0x02221110, 0x02221112, 0x02221211,
+    0x00201121, 0x00211020, 0x00211022, 0x00211221, 0x00221121, 0x01201021, 0x01201221, 0x01211121,
+    0x01221020, 0x01221021, 0x01221221, 0x02201120, 0x02201122, 0x02211020, 0x02211222, 0x00202000,
+    0x00202002, 0x00202200, 0x00202202, 0x00212101, 0x00222000, 0x00222002, 0x00222200, 0x00222202,
+    0x01202101, 0x01212001, 0x01212100, 0x01222101, 0x02202000, 0x02202002, 0x02202200, 0x02202202,
+    0x02222000, 0x02222002, 0x02222200, 0x02222202, 0x00202211, 0x00212011, 0x00212110, 0x00212211,
+    0x00222111, 0x01202112, 0x01202211, 0x01212012, 0x01212111, 0x01222011, 0x01222110, 0x01222112,
+    0x01222211, 0x02202111, 0x02212010, 0x02212112, 0x02212211, 0x02222110, 0x02222111, 0x00202020,
+    0x00202022, 0x00202220, 0x00202222, 0x00222020, 0x00222022, 0x00222220, 0x00222222, 0x01202121,
+    0x01212021, 0x01212122, 0x01212221, 0x01222121, 0x02202020, 0x02202022, 0x02202220, 0x02202222,
+    0x02212121, 0x02222020, 0x02222022, 0x02222220, 0x02222222, 0x10000101, 0x10010001, 0x10010102,
+    0x10020101, 0x11000201, 0x11010002, 0x11010101, 0x11010200, 0x11010202, 0x11020001, 0x11020100,
+    0x11020102, 0x12010100, 0x12010201, 0x12020001, 0x12020102, 0x10000010, 0x10000011, 0x10000110,
+    0x10000112, 0x10000211, 0x10010012, 0x10010111, 0x10010112, 0x10010210, 0x10010212, 0x10020011,
+    0x10020112, 0x10020211, 0x11000111, 0x11000210, 0x11000212, 0x11010011, 0x11010110, 0x11010111,
+    0x11010112, 0x11010211, 0x11010212, 0x11020111, 0x11020210, 0x11020212, 0x12000011, 0x12000110,
+    0x12000112, 0x12010010, 0x12010012, 0x12010111, 0x12020010, 0x12020011, 0x12020012, 0x10000121,
+    0x10010021, 0x10010120, 0x10010122, 0x10020121, 0x11000021, 0x11010022, 0x11010121, 0x11010222,
+    0x11020120, 0x11020221, 0x12000221, 0x12010120, 0x12020121, 0x10001001, 0x10011101, 0x10011201,
+    0x10021201, 0x11001101, 0x11001200, 0x11001202, 0x11011001, 0x11011100, 0x11011101, 0x11011102,
+    0x11021001, 0x11021002, 0x11021101, 0x11021200, 0x11021202, 0x12001001, 0x12001102, 0x12001201,
+    0x12011000, 0x12011002, 0x12011101, 0x12021000, 0x12021001, 0x12021201, 0x10001011, 0x10001012,
+    0x10001111, 0x10001212, 0x10011011, 0x10011110, 0x10011111, 0x10011112, 0x10011211, 0x10021010,
+    0x10021111, 0x10021212, 0x11001011, 0x11001110, 0x11001111, 0x11001112, 0x11001211, 0x11011010,
+    0x11011011, 0x11011110, 0x11011111, 0x11011112, 0x11011210, 0x11011211, 0x11021011, 0x11021110,
+    0x11021111, 0x11021112, 0x11021211, 0x12001012, 0x12001110, 0x12001111, 0x12001210, 0x12011011,
+    0x12011110, 0x12011111, 0x12011112, 0x12011211, 0x12011212, 0x12021111, 0x12021210, 0x12021212,
+    0x10001021, 0x10001121, 0x10001221, 0x10011120, 0x10011121, 0x10011220, 0x10011222, 0x10021021,
+    0x10021120, 0x10021221, 0x11001020, 0x11001022, 0x11001121, 0x11001220, 0x11011020, 0x11011021,
+    0x11011022, 0x11011121, 0x11011122, 0x11011221, 0x11021022, 0x11021121, 0x11021220, 0x12001021,
+    0x12001121, 0x12001222, 0x12011120, 0x12011121, 0x12021021, 0x12021120, 0x12021122, 0x10002101,
+    0x10012001, 0x10012101, 0x10012202, 0x10022101, 0x11002002, 0x11002201, 0x11012000, 0x11012101,
+    0x11012200, 0x11022001, 0x11022100, 0x11022102, 0x11022201, 0x12002101, 0x12012001, 0x12012100,
+    0x12012102, 0x12012201, 0x12022101, 0x10002011, 0x10002111, 0x10002112, 0x10002212, 0x10012010,
+    0x10012110, 0x10012111, 0x10012210, 0x10022011, 0x10022110, 0x10022112, 0x11002010, 0x11002111,
+    0x11002212, 0x11012011, 0x11012012, 0x11012110, 0x11012111, 0x11012112, 0x11012211, 0x11022010,
+    0x11022012, 0x11022111, 0x11022112, 0x11022212, 0x12002112, 0x12002211, 0x12012012, 0x12012111,
+    0x12012112, 0x12012210, 0x12022011, 0x12022110, 0x12022112, 0x12022211, 0x10012122, 0x11002120,
+    0x11002122, 0x11002221, 0x11012121, 0x11012220, 0x11012222, 0x11022120, 0x11022221, 0x12012120,
+    0x12022121, 0x10100001, 0x10100100, 0x10100101, 0x10100102, 0x10100201, 0x10110002, 0x10110101,
+    0x10110202, 0x10120001, 0x10120100, 0x10120201, 0x11100000, 0x11100101, 0x11100200, 0x11110001,
+    0x11110100, 0x11110101, 0x11110102, 0x11110201, 0x11120101, 0x11120200, 0x12100102, 0x12100201,
+    0x12110101, 0x12110200, 0x12120000, 0x12120001, 0x12120102, 0x12120201, 0x10100111, 0x10100210,
+    0x10100211, 0x10100212, 0x10110011, 0x10110110, 0x10110111, 0x10110112, 0x10110210, 0x10110211,
+    0x10120010, 0x10120111, 0x10120112, 0x10120210, 0x10120212, 0x11100011, 0x11100110, 0x11100111,
+    0x11100112, 0x11100211, 0x11110010, 0x11110011, 0x11110012, 0x11110110, 0x11110111, 0x11110112,
+    0x11110210, 0x11110211, 0x11110212, 0x11120011, 0x11120110, 0x11120111, 0x11120112, 0x11120211,
+    0x12100012, 0x12100111, 0x12110011, 0x12110110, 0x12110111, 0x12110112, 0x12110211, 0x12120010,
+    0x12120111, 0x12120212, 0x10100021, 0x10100122, 0x10110022, 0x10110121, 0x10110222, 0x10120021,
+    0x10120120, 0x11100022, 0x11100121, 0x11100222, 0x11110021, 0x11110120, 0x11110121, 0x11110122,
+    0x11110221, 0x11120022, 0x11120121, 0x12100121, 0x12110020, 0x12110022, 0x12110121, 0x12110221,
+    0x12110222, 0x12120120, 0x10101100, 0x10101101, 0x10111001, 0x10111100, 0x10111101, 0x10111102,
+    0x10111200, 0x10111201, 0x10121001, 0x10121101, 0x10121200, 0x10121202, 0x11101001, 0x11101100,
+    0x11101101, 0x11101102, 0x11101201, 0x11101202, 0x11111000, 0x11111001, 0x11111100, 0x11111101,
+    0x11111102, 0x11111200, 0x11111201, 0x11111202, 0x11121001, 0x11121002, 0x11121100, 0x11121101,
+    0x11121102, 0x11121201, 0x12101000, 0x12101200, 0x12101202, 0x12111001, 0x12111100, 0x12111101,
+    0x12111102, 0x12111201, 0x12121001, 0x12121100, 0x12121101, 0x12121202, 0x10101011, 0x10101012,
+    0x10101110, 0x10101111, 0x10101112, 0x10101211, 0x10111010, 0x10111011, 0x10111012, 0x10111110,
+    0x10111111, 0x10111112, 0x10111211, 0x10111212, 0x10121011, 0x10121110, 0x10121111, 0x10121112,
+    0x10121211, 0x11101010, 0x11101011, 0x11101012, 0x11101110, 0x11101111, 0x11101112, 0x11101210,
+    0x11101211, 0x11111010, 0x11111011, 0x11111012, 0x11111110, 0x11111111, 0x11111112, 0x11111210,
+    0x11111211, 0x11111212, 0x11121010, 0x11121011, 0x11121110, 0x11121111, 0x11121112, 0x11121210,
+    0x11121211, 0x11121212, 0x12101011, 0x12101110, 0x12101111, 0x12101211, 0x12101212, 0x12111010,
+    0x12111011, 0x12111110, 0x12111111, 0x12111112, 0x12111210, 0x12111211, 0x12121011, 0x12121110,
+    0x12121111, 0x12121112, 0x12121211, 0x10101020, 0x10101021, 0x10101022, 0x10101120, 0x10101122,
+    0x10101220, 0x10101221, 0x10111021, 0x10111120, 0x10111121, 0x10111220, 0x10111221, 0x10121020,
+    0x10121021, 0x10121022, 0x10121120, 0x10121121, 0x10121122, 0x10121220, 0x10121221, 0x11101021,
+    0x11101121, 0x11101122, 0x11101220, 0x11101221, 0x11101222, 0x11111020, 0x11111021, 0x11111022,
+    0x11111120, 0x11111121, 0x11111122, 0x11111220, 0x11111221, 0x11111222, 0x11121021, 0x11121120,
+    0x11121121, 0x11121221, 0x12101022, 0x12101121, 0x12101122, 0x12101220, 0x12101221, 0x12101222,
+    0x12111021, 0x12111121, 0x12111222, 0x12121022, 0x12121121, 0x12121122, 0x12121220, 0x12121221,
+    0x10102100, 0x10102101, 0x10102102, 0x10102201, 0x10112000, 0x10112101, 0x10112200, 0x10122001,
+    0x10122202, 0x11102101, 0x11102200, 0x11102202, 0x11112001, 0x11112100, 0x11112101, 0x11112102,
+    0x11112200, 0x11112201, 0x11122000, 0x11122002, 0x11122100, 0x11122101, 0x12102002, 0x12102201,
+    0x12112000, 0x12112002, 0x12112101, 0x12112200, 0x12122001, 0x12122201, 0x10102011, 0x10102012,
+    0x10102111, 0x10102212, 0x10112011, 0x10112110, 0x10112111, 0x10112112, 0x10112211, 0x10122111,
+    0x11102011, 0x11102110, 0x11102111, 0x11102112, 0x11102211, 0x11112010, 0x11112011, 0x11112012,
+    0x11112110, 0x11112111, 0x11112112, 0x11112210, 0x11112211, 0x11112212, 0x11122011, 0x11122110,
+    0x11122111, 0x11122112, 0x11122211, 0x12102011, 0x12102111, 0x12102211, 0x12112011, 0x12112110,
+    0x12112111, 0x12112112, 0x12112210, 0x12112211, 0x12122111, 0x10102120, 0x10102220, 0x10112121,
+    0x10112222, 0x10122020, 0x10122121, 0x10122122, 0x10122221, 0x11102121, 0x11102220, 0x11102221,
+    0x11112021, 0x11112121, 0x11112122, 0x11112220, 0x11112221, 0x11122022, 0x11122121, 0x11122220,
+    0x11122222, 0x12102021, 0x12102222, 0x12112022, 0x12112121, 0x12112122, 0x12112220, 0x12112222,
+    0x12122021, 0x10200101, 0x10210100, 0x10210102, 0x10210201, 0x10220101, 0x11200100, 0x11210000,
+    0x11210101, 0x11210102, 0x11210200, 0x11210202, 0x11220001, 0x11220100, 0x11220102, 0x11220201,
+    0x12200001, 0x12210102, 0x12220101, 0x10200011, 0x10200110, 0x10200112, 0x10200211, 0x10210012,
+    0x10210111, 0x10220011, 0x10220012, 0x10220112, 0x10220211, 0x11200111, 0x11200211, 0x11210011,
+    0x11210111, 0x11210112, 0x11210211, 0x11220111, 0x11220112, 0x11220212, 0x12200110, 0x12200212,
+    0x12210012, 0x12210111, 0x12220011, 0x12220112, 0x12220211, 0x10210021, 0x10210122, 0x10210221,
+    0x11200020, 0x11200021, 0x11200122, 0x11210121, 0x11210122, 0x11210220, 0x11220020, 0x12200121,
+    0x12210021, 0x12210122, 0x12220121, 0x10211001, 0x10211002, 0x10211101, 0x10211102, 0x10211202,
+    0x10221001, 0x10221102, 0x10221201, 0x11201000, 0x11201002, 0x11201101, 0x11201200, 0x11201202,
+    0x11211001, 0x11211100, 0x11211101, 0x11211102, 0x11211201, 0x11211202, 0x11221000, 0x11221002,
+    0x11221101, 0x12201100, 0x12201101, 0x12201201, 0x12211000, 0x12211002, 0x12211100, 0x12211101,
+    0x12211102, 0x12211200, 0x12211202, 0x12221001, 0x12221100, 0x12221201, 0x10201111, 0x10201210,
+    0x10201212, 0x10211011, 0x10211111, 0x10211112, 0x10211211, 0x11201110, 0x11201111, 0x11201112,
+    0x11201211, 0x11211010, 0x11211011, 0x11211110, 0x11211111, 0x11211112, 0x11211211, 0x11221011,
+    0x11221110, 0x11221111, 0x11221112, 0x11221211, 0x12201112, 0x12201211, 0x12201212, 0x12211011,
+    0x12211111, 0x12211112, 0x12211211, 0x12211212, 0x12221012, 0x12221111, 0x12221112, 0x12221210,
+    0x10201022, 0x10201221, 0x10211121, 0x10221020, 0x10221122, 0x10221220, 0x10221221, 0x11201020,
+    0x11201121, 0x11201220, 0x11201222, 0x11211021, 0x11211120, 0x11211121, 0x11211122, 0x11211220,
+    0x11211222, 0x11221020, 0x11221121, 0x11221220, 0x12201020, 0x12201022, 0x12201121, 0x12201222,
+    0x12211120, 0x12211122, 0x12211220, 0x12211221, 0x12221020, 0x12221120, 0x12221122, 0x12221222,
+    0x10212102, 0x10212201, 0x10222101, 0x11202001, 0x11212002, 0x11212101, 0x11212202, 0x11222001,
+    0x11222201, 0x12202101, 0x12212001, 0x12212200, 0x12222102, 0x10202011, 0x10202110, 0x10212010,
+    0x10212111, 0x10222011, 0x10222110, 0x10222112, 0x10222211, 0x11202010, 0x11202011, 0x11202111,
+    0x11202112, 0x11202210, 0x11212011, 0x11212110, 0x11212111, 0x11212112, 0x11212211, 0x11222010,
+    0x11222111, 0x11222212, 0x12202012, 0x12202110, 0x12202212, 0x12212111, 0x12222011, 0x12222110,
+    0x12222111, 0x12222211, 0x10212021, 0x10212122, 0x10212220, 0x11202021, 0x11202120, 0x11202221,
+    0x11212020, 0x11212121, 0x11212220, 0x11212222, 0x11222120, 0x11222121, 0x11222221, 0x12202122,
+    0x12212120, 0x12212220, 0x12212222, 0x12222122, 0x20000000, 0x20000002, 0x20000200, 0x20000202,
+    0x20020000, 0x20020002, 0x20020200, 0x20020202, 0x21000101, 0x21010000, 0x21010001, 0x21010100,
+    0x21010102, 0x21010201, 0x21020101, 0x22000000, 0x22000002, 0x22000200, 0x22000202, 0x22010101,
+    0x22020000, 0x22020002, 0x22020200, 0x22020202, 0x20000111, 0x20010011, 0x20010110, 0x20010112,
+    0x20010211, 0x20020111, 0x21000011, 0x21000110, 0x21000211, 0x21010010, 0x21010012, 0x21010111,
+    0x21010112, 0x21010210, 0x21010211, 0x21020110, 0x21020112, 0x21020211, 0x22000111, 0x22000211,
+    0x22010110, 0x22010112, 0x22010211, 0x22020111, 0x20000020, 0x20000022, 0x20000220, 0x20000222,
+    0x20010121, 0x20020020, 0x20020022, 0x20020220, 0x20020222, 0x21010021, 0x21010120, 0x21010221,
+    0x21020121, 0x22000020, 0x22000022, 0x22000220, 0x22000222, 0x22010121, 0x22020020, 0x22020022,
+    0x22020220, 0x22020222, 0x20011100, 0x20011201, 0x21001001, 0x21001100, 0x21011001, 0x21011101,
+    0x21011202, 0x21021001, 0x21021100, 0x21021201, 0x22011100, 0x22011201, 0x20001011, 0x20001211,
+    0x20011012, 0x20011111, 0x20011212, 0x20021112, 0x20021211, 0x21001010, 0x21001011, 0x21001111,
+    0x21001210, 0x21011011, 0x21011110, 0x21011111, 0x21011112, 0x21011211, 0x21011212, 0x21021111,
+    0x21021112, 0x21021210, 0x21021212, 0x22001011, 0x22001110, 0x22001112, 0x22001211, 0x22011010,
+    0x22011012, 0x22011111, 0x22011210, 0x22021112, 0x20011021, 0x20011122, 0x20011221, 0x20021121,
+    0x21001021, 0x21001120, 0x21001221, 0x21001222, 0x21011020, 0x21011121, 0x21011221, 0x21011222,
+    0x21021021, 0x21021122, 0x21021222, 0x22001121, 0x22011021, 0x22011222, 0x22021120, 0x20002000,
+    0x20002002, 0x20002200, 0x20002202, 0x20012101, 0x20022000, 0x20022002, 0x20022200, 0x20022202,
+    0x21002001, 0x21002101, 0x21012001, 0x21012100, 0x21012201, 0x21022101, 0x21022201, 0x22002000,
+    0x22002002, 0x22002200, 0x22002202, 0x22012101, 0x22022000, 0x22022002, 0x22022200, 0x22022202,
+    0x20002111, 0x20002112, 0x20012011, 0x20012110, 0x20012112, 0x20022111, 0x21002011, 0x21002110,
+    0x21002112, 0x21002211, 0x21012010, 0x21012012, 0x21012111, 0x21012212, 0x21022011, 0x21022110,
+    0x22002111, 0x22012112, 0x22012211, 0x22022111, 0x20002020, 0x20002022, 0x20002220, 0x20002222,
+    0x20012121, 0x20022020, 0x20022022, 0x20022220, 0x20022222, 0x21002121, 0x21012021, 0x21012120,
+    0x21012122, 0x22002020, 0x22002022, 0x22002220, 0x22002222, 0x22012121, 0x22022020, 0x22022022,
+    0x22022220, 0x22022222, 0x20100101, 0x20110001, 0x20110102, 0x20110200, 0x20110201, 0x20120101,
+    0x21100001, 0x21100102, 0x21100201, 0x21110101, 0x21110200, 0x21110202, 0x21120201, 0x21120202,
+    0x22100101, 0x22110001, 0x22110100, 0x22110102, 0x22110201, 0x22120101, 0x20100011, 0x20100110,
+    0x20100112, 0x20100211, 0x20110010, 0x20110111, 0x20110210, 0x20110212, 0x20120011, 0x20120110,
+    0x20120112, 0x20120211, 0x21100010, 0x21100111, 0x21110010, 0x21110011, 0x21110110, 0x21110111,
+    0x21110112, 0x21110211, 0x21120012, 0x21120111, 0x22100110, 0x22100112, 0x22110012, 0x22110111,
+    0x22110210, 0x22120011, 0x22120110, 0x22120112, 0x22120211, 0x20100121, 0x20110021, 0x20110120,
+    0x20110221, 0x20120121, 0x21100120, 0x21100122, 0x21100221, 0x21110020, 0x21110022, 0x21110121,
+    0x21110220, 0x21120122, 0x21120221, 0x22100121, 0x22110120, 0x22110122, 0x22120221, 0x20101001,
+    0x20101100, 0x20101102, 0x20111000, 0x20111101, 0x20111200, 0x20121102, 0x21101000, 0x21101202,
+    0x21111001, 0x21111100, 0x21111101, 0x21111102, 0x21111200, 0x21111201, 0x21121000, 0x21121001,
+    0x21121002, 0x21121101, 0x22101100, 0x22101102, 0x22111002, 0x22111100, 0x22111101, 0x22111200,
+    0x22121001, 0x22121201, 0x20101010, 0x20101111, 0x20101210, 0x20101212, 0x20111010, 0x20111011,
+    0x20111110, 0x20111111, 0x20111112, 0x20111211, 0x20121011, 0x20121111, 0x20121211, 0x20121212,
+    0x21101011, 0x21101110, 0x21101111, 0x21101112, 0x21101211, 0x21111010, 0x21111011, 0x21111012,
+    0x21111110, 0x21111111, 0x21111112, 0x21111210, 0x21111211, 0x21111212, 0x21121011, 0x21121110,
+    0x21121111, 0x21121112, 0x21121211, 0x22101011, 0x22101111, 0x22101210, 0x22111011, 0x22111012,
+    0x22111110, 0x22111111, 0x22111112, 0x22111211, 0x22111212, 0x22121010, 0x22121012, 0x22121111,
+    0x22121210, 0x22121212, 0x20101021, 0x20101120, 0x20111020, 0x20111121, 0x20111221, 0x20121020,
+    0x20121122, 0x20121221, 0x21101121, 0x21101220, 0x21101221, 0x21111021, 0x21111022, 0x21111121,
+    0x21111122, 0x21111221, 0x21121121, 0x21121220, 0x22101022, 0x22101120, 0x22101221, 0x22101222,
+    0x22111022, 0x22111120, 0x22111121, 0x22121120, 0x22121122, 0x22121221, 0x20102101, 0x20112102,
+    0x20112201, 0x20122101, 0x21102001, 0x21102102, 0x21112000, 0x21112002, 0x21112101, 0x21112102,
+    0x21112202, 0x21122100, 0x21122101, 0x22102101, 0x22112001, 0x22112102, 0x22112201, 0x22122101,
+    0x20102110, 0x20102112, 0x20102211, 0x20112010, 0x20112012, 0x20112111, 0x20112210, 0x20112212,
+    0x20122010, 0x20122011, 0x20122110, 0x20122112, 0x21102010, 0x21102012, 0x21102111, 0x21102210,
+    0x21102212, 0x21112011, 0x21112110, 0x21112111, 0x21112112, 0x21112211, 0x21122012, 0x21122111,
+    0x21122112, 0x21122212, 0x22102011, 0x22102110, 0x22112010, 0x22112012, 0x22112111, 0x22112212,
+    0x22122011, 0x22122112, 0x20102121, 0x20112121, 0x20122121, 0x21102120, 0x21102122, 0x21102221,
+    0x21112020, 0x21112121, 0x21112220, 0x21122021, 0x22102121, 0x22112021, 0x22112120, 0x22112121,
+    0x22112122, 0x20200000, 0x20200002, 0x20200200, 0x20200202, 0x20210101, 0x20220000, 0x20220002,
+    0x20220200, 0x20220202, 0x21200101, 0x21210001, 0x21210100, 0x21210102, 0x21210201, 0x22200000,
+    0x22200002, 0x22200200, 0x22200202, 0x22210101, 0x22220000, 0x22220002, 0x22220200, 0x22220202,
+    0x20200111, 0x20200211, 0x20210011, 0x20210110, 0x20210112, 0x20210211, 0x20210212, 0x21200112,
+    0x21200211, 0x21210011, 0x21210111, 0x21210210, 0x21210212, 0x21220011, 0x21220110, 0x22200111,
+    0x22210010, 0x22210012, 0x22210112, 0x22210211, 0x20200022, 0x20200220, 0x20200222, 0x20210020,
+    0x20210221, 0x20220022, 0x20220220, 0x20220222, 0x21200121, 0x21210021, 0x21210122, 0x21210221,
+    0x21220121, 0x22200020, 0x22200022, 0x22200220, 0x22200222, 0x22210121, 0x22220020, 0x22220022,
+    0x22220220, 0x22220222, 0x20211201, 0x20221101, 0x21201001, 0x21201100, 0x21211000, 0x21211100,
+    0x21211101, 0x21211200, 0x21211202, 0x21221001, 0x21221101, 0x21221102, 0x21221200, 0x21221201,
+    0x22201101, 0x20201112, 0x20201211, 0x20211010, 0x20211012, 0x20211111, 0x20211210, 0x20221112,
+    0x20221211, 0x21201012, 0x21201111, 0x21211011, 0x21211110, 0x21211111, 0x21211112, 0x21211211,
+    0x21221111, 0x21221212, 0x22201011, 0x22201110, 0x22201111, 0x22201112, 0x22201211, 0x22211012,
+    0x22211111, 0x22211210, 0x20201121, 0x20211021, 0x20211122, 0x20211222, 0x20221021, 0x20221121,
+    0x21201120, 0x21201122, 0x21201222, 0x21211022, 0x21211121, 0x21211122, 0x21211220, 0x21221020,
+    0x21221022, 0x22201122, 0x22211020, 0x22211121, 0x22211122, 0x22211221, 0x22221021, 0x22221120,
+    0x22221122, 0x20202000, 0x20202002, 0x20202200, 0x20202202, 0x20222000, 0x20222002, 0x20222200,
+    0x20222202, 0x21212001, 0x21212100, 0x21212102, 0x21212201, 0x22202000, 0x22202002, 0x22202200,
+    0x22202202, 0x22212101, 0x22222000, 0x22222002, 0x22222200, 0x22222202, 0x20202111, 0x20212110,
+    0x20212211, 0x20222011, 0x20222111, 0x21202011, 0x21212010, 0x21212111, 0x21212212, 0x21222011,
+    0x21222112, 0x21222211, 0x22212010, 0x22212112, 0x20202020, 0x20202022, 0x20202220, 0x20202222,
+    0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
+    0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
+GGML_TABLE_END()
+#endif
+
+#endif // GGML_COMMON_IMPL
+#endif // GGML_COMMON_IMPL

llama/ggml-cuda.cu

@@ -4,7 +4,6 @@
 
 #include "ggml-cuda/common.cuh"
 #include "ggml-cuda/acc.cuh"
-#include "ggml-cuda/alibi.cuh"
 #include "ggml-cuda/arange.cuh"
 #include "ggml-cuda/argsort.cuh"
 #include "ggml-cuda/binbcast.cuh"
@@ -14,6 +13,7 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/diagmask.cuh"
 #include "ggml-cuda/dmmv.cuh"
+#include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmq.cuh"
@@ -112,7 +112,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
         CUdevice device;
         CU_CHECK(cuDeviceGet(&device, id));
         CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@@ -140,6 +140,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpb = prop.sharedMemPerBlock;
+        info.devices[id].nsm  = prop.multiProcessorCount;
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -257,7 +258,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 };
 
 // pool with virtual memory
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
 struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -354,7 +355,7 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
 #endif // !defined(GGML_USE_HIPBLAS)
 
 std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
@@ -714,6 +715,9 @@ static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
 GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
     delete ctx;
+    
+    // HACK: this needs to be freed in msvc
+    free(buffer);
 }
 
 GGML_CALL static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -1645,7 +1649,7 @@ static void ggml_cuda_op_mul_mat(
     }
 }
 
-static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
+static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
     GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
@@ -1668,7 +1672,7 @@ static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const gg
     ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
 }
 
-static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
+static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
@@ -2203,6 +2207,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_RELU:
                     ggml_cuda_op_relu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_SIGMOID:
+                    ggml_cuda_op_sigmoid(ctx, dst);
+                    break;
                 case GGML_UNARY_OP_HARDSIGMOID:
                     ggml_cuda_op_hardsigmoid(ctx, dst);
                     break;
@@ -2275,9 +2282,6 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ROPE:
             ggml_cuda_op_rope(ctx, dst);
             break;
-        case GGML_OP_ALIBI:
-            ggml_cuda_op_alibi(ctx, dst);
-            break;
         case GGML_OP_IM2COL:
             ggml_cuda_op_im2col(ctx, dst);
             break;
@@ -2290,6 +2294,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ARGSORT:
             ggml_cuda_op_argsort(ctx, dst);
             break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            ggml_cuda_flash_attn_ext(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2405,44 +2412,318 @@ GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
     GGML_UNUSED(backend);
 }
 
+static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+    graph_node_properties->node_address = node->data;
+    graph_node_properties->node_op = node->op;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        graph_node_properties->ne[i] = node->ne[i];
+        graph_node_properties->nb[i] = node->nb[i];
+    }
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
+    }
+}
+
+static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+    if (node->data != graph_node_properties->node_address &&
+          node->op != GGML_OP_CPY &&
+          node->op != GGML_OP_VIEW) {
+        return false;
+    }
+
+    if (node->op != graph_node_properties->node_op) {
+        return false;
+    }
+
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (node->ne[i] != graph_node_properties->ne[i]) {
+            return false;
+        }
+        if (node->nb[i] != graph_node_properties->nb[i]) {
+            return false;
+        }
+    }
+
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (node->src[i] &&
+            node->src[i]->data != graph_node_properties->src_address[i] &&
+            node->op != GGML_OP_CPY &&
+            node->op != GGML_OP_VIEW
+        ) {
+            return false;
+        }
+    }
+    return true;
+}
+
 GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     ggml_cuda_set_device(cuda_ctx->device);
 
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_tensor * node = cgraph->nodes[i];
+#ifdef USE_CUDA_GRAPH
+    static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
 
-        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
-            continue;
+    // Objects required for CUDA Graph
+    if (cuda_ctx->cuda_graph == nullptr) {
+        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
+    }
+
+    bool use_cuda_graph = true;
+    bool cuda_graph_update_required = false;
+    // pointer to CUDA cpy kernel, which is required to identify
+    // kernel parameters which need updated in the graph for each token
+    void * ggml_cuda_cpy_fn_ptr = nullptr;
+
+    if (cuda_ctx->cuda_graph->graph == nullptr) {
+        if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
+            cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
+#ifndef NDEBUG
+            fprintf(stderr, "%s: disabling CUDA graphs due to GPU architecture\n", __func__);
+#endif
+        }
+    }
+
+    // Disable CUDA graphs in presence of env var, old GPU, use-case which is changing too rapidly,
+    // or previous graph capture failure.
+    // Also disable for multi-gpu for now. TO DO investigate
+    if (disable_cuda_graphs_due_to_env
+        || cuda_ctx->cuda_graph->disable_due_to_gpu_arch
+        || cuda_ctx->cuda_graph->disable_due_to_too_many_updates
+        || cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture) {
+        use_cuda_graph = false;
+    }
+
+    if (use_cuda_graph) {
+        if (cuda_ctx->cuda_graph->instance == nullptr) {
+            cuda_graph_update_required = true;
         }
 
+        // Check if the graph size has changed
+        if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
+            cuda_graph_update_required = true;
+            cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+        }
+
+        // Loop over nodes in GGML graph to determine if CUDA graph update is required
+        // and store properties to allow this comparison for the next token
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            bool has_matching_properties = true;
+            if (!cuda_graph_update_required) {
+                has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
+            }
+            if (!has_matching_properties) {
+                cuda_graph_update_required = true;
+            }
+            set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
+        }
+
+        // Loop over nodes in GGML graph to obtain info needed for CUDA graph
+        cuda_ctx->cuda_graph->updated_kernel_arg.clear();
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            ggml_tensor * node = cgraph->nodes[i];
+
+            if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
+                use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
 #ifndef NDEBUG
-        assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            if (node->src[j] != nullptr) {
-                assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
+                fprintf(stderr, "%s: disabling CUDA graphs due to split buffer\n", __func__);
+#endif
+            }
+
+            if (node->op == GGML_OP_MUL_MAT_ID) {
+                use_cuda_graph = false; // This node type is not supported by CUDA graph capture
+#ifndef NDEBUG
+                fprintf(stderr, "%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
+#endif
+            }
+
+            if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
+                // disable CUDA graphs for batch size > 1 for now.
+                // Changes in batch size or context size can cause changes to the grid size of some kernels.
+                use_cuda_graph = false;
+#ifndef NDEBUG
+                fprintf(stderr, "%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
+#endif
+            }
+
+            if (node->op == GGML_OP_CPY) {
+                // store the copy op parameter which changes with each token.
+                cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
+                if (ggml_cuda_cpy_fn_ptr == nullptr) {
+                    // store a pointer to the copy op CUDA kernel to identify it later
+                    ggml_cuda_cpy_fn_ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
+                }
+            }
+
+            if (!use_cuda_graph) {
+                break;
             }
         }
+
+        // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
+        if (use_cuda_graph && cuda_graph_update_required) {
+            cuda_ctx->cuda_graph->number_consecutive_updates++;
+        } else {
+            cuda_ctx->cuda_graph->number_consecutive_updates = 0;
+        }
+
+        if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
+            cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
+#ifndef NDEBUG
+            fprintf(stderr, "%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
+#endif
+        }
+    }
+
+    if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
+        CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
+    }
+
+#else
+    bool use_cuda_graph = false;
+    bool cuda_graph_update_required = false;
+#endif // USE_CUDA_GRAPH
+
+    bool graph_evaluated_or_captured = false;
+
+    while (!graph_evaluated_or_captured) {
+        // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
+        // With the use of CUDA graphs, the execution will be performed by the graph launch.
+        if (!use_cuda_graph || cuda_graph_update_required) {
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+
+                if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+                    continue;
+                }
+
+#ifndef NDEBUG
+                assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    if (node->src[j] != nullptr) {
+                        assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
+                    }
+                }
+#endif
+
+                bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
+                if (!ok) {
+                    fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+                }
+                GGML_ASSERT(ok);
+            }
+        }
+
+#ifdef USE_CUDA_GRAPH
+        if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
+            if (cuda_ctx->cuda_graph->graph != nullptr) {
+                CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
+                cuda_ctx->cuda_graph->graph = nullptr;
+            }
+            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
+
+#if 0
+            if (disable_cuda_graphs_due_to_failed_capture) {
+                use_cuda_graph = false;
+                cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
+#ifndef NDEBUG
+                fprintf(stderr, "%s: disabling CUDA graphs due to failed graph capture\n", __func__);
 #endif
+            } else {
+                graph_evaluated_or_captured = true; // CUDA graph has been captured
+            }
+#endif
+            graph_evaluated_or_captured = true; // CUDA graph has been captured
+        } else {
+            graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
+        }
+    }
+
+    if (use_cuda_graph) {
+        if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
+            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        }
+
+        // Perform update to graph (if required for this token), and change copy parameter (required for every token)
+
+        if (cuda_graph_update_required) {
+            // Extract nodes from graph
+            if (cuda_ctx->cuda_graph->num_nodes == 0) {
+                // First call with null argument gets number of nodes in graph
+                CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
+            }
+            // Subsequent call with non-null argument gets nodes
+            cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
+            cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
+            if (cuda_ctx->cuda_graph->num_nodes > 0) {
+                CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
+
+                // Loop over nodes, and extract kernel parameters from each node
+                for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
+                    cudaGraphNodeType node_type;
+                    CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
+                    if (node_type == cudaGraphNodeTypeKernel) {
+                        cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
+                        if (stat == cudaErrorInvalidDeviceFunction) {
+                            // Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
+                            // We don't need to update blas nodes, so clear error and move on.
+                            cudaGetLastError();
+                        } else {
+                            GGML_ASSERT(stat == cudaSuccess);
+                        }
+                    }
+                }
+            }
+        }
 
-        bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
-        if (!ok) {
-            fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+        // One of the arguments to the copy kernel is updated for each token, hence we need to
+        // replace that argument with the updated value in the CUDA graph
+        if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured
+            int k = 0;
+            for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
+                if (cuda_ctx->cuda_graph->params[i].func == ggml_cuda_cpy_fn_ptr) {
+                    char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
+                    cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
+                    CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
+                }
+            }
         }
-        GGML_ASSERT(ok);
+
+        // Update graph executable
+        cudaGraphExecUpdateResultInfo result_info;
+        cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
+        if (stat == cudaErrorGraphExecUpdateFailure) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: CUDA graph update failed\n", __func__);
+#endif
+            // The pre-existing graph exec cannot be updated due to violated constraints
+            // so instead clear error and re-instantiate
+            cudaGetLastError();
+            CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
+            cuda_ctx->cuda_graph->instance = nullptr;
+            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        } else {
+            GGML_ASSERT(stat == cudaSuccess);
+        }
+        // Launch graph
+        CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
+#else
+        graph_evaluated_or_captured = true;
+#endif // USE_CUDA_GRAPH
     }
 
     return GGML_STATUS_SUCCESS;
 }
 
 GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
@@ -2552,7 +2833,6 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_ROPE:
-        case GGML_OP_ALIBI:
         case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
@@ -2565,6 +2845,15 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
             return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+#else
+            if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
+                return true;
+            }
+            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         default:
             return false;
     }
@@ -2742,7 +3031,7 @@ GGML_CALL static ggml_backend_t ggml_backend_reg_cuda_init(const char * params,
     GGML_UNUSED(params);
 }
 
-extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();
+// extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();
 
 GGML_CALL int ggml_backend_cuda_reg_devices() {
     int device_count = ggml_backend_cuda_get_device_count();

llama/ggml-cuda.h

@@ -1,43 +1,45 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef GGML_USE_HIPBLAS
-#define GGML_CUDA_NAME "ROCm"
-#define GGML_CUBLAS_NAME "hipBLAS"
-#else
-#define GGML_CUDA_NAME "CUDA"
-#define GGML_CUBLAS_NAME "cuBLAS"
-#endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
-
-GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
-
-// device buffer
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-
-GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
-
-#ifdef  __cplusplus
-}
-#endif
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef GGML_USE_HIPBLAS
+#define GGML_CUDA_NAME "ROCm"
+#define GGML_CUBLAS_NAME "hipBLAS"
+#else
+#define GGML_CUDA_NAME "CUDA"
+#define GGML_CUBLAS_NAME "cuBLAS"
+#endif
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_CUDA_MAX_DEVICES       16
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
+
+GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
+
+// device buffer
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+
+GGML_API GGML_CALL int ggml_backend_cuda_reg_devices();
+
+GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+
+GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+
+#ifdef  __cplusplus
+}
+#endif

llama/ggml-cuda/acc.cu

@@ -0,0 +1,47 @@
+#include "acc.cuh"
+
+static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, int offset) {
+    const int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i >= ne) {
+        return;
+    }
+    int src1_idx = i - offset;
+    int oz = src1_idx / nb2;
+    int oy = (src1_idx - (oz * nb2)) / nb1;
+    int ox = src1_idx % nb1;
+    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
+        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
+    } else {
+        dst[i] = x[i];
+    }
+}
+
+static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
+    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
+}
+
+void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
+
+    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
+    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
+    int offset = dst->op_params[3] / 4; // offset in bytes
+
+    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
+}

llama/ggml-cuda/acc.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ACC_BLOCK_SIZE 256
+
+void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/alibi.cu

@@ -0,0 +1,63 @@
+#include "alibi.cuh"
+
+static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
+                                 const int n_heads_log2_floor, const float m0, const float m1) {
+    const int col = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int row = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i = row*ncols + col;
+
+    const int k = row/k_rows;
+
+    float m_k;
+    if (k < n_heads_log2_floor) {
+        m_k = powf(m0, k + 1);
+    } else {
+        m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
+    }
+
+    dst[i] = col * m_k + x[i];
+}
+
+static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
+                           const int k_rows, const int n_heads_log2_floor, const float m0,
+                           const float m1, cudaStream_t stream) {
+    const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1);
+    const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE);
+    const dim3 block_nums(num_blocks_x, nrows, 1);
+    alibi_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1);
+}
+
+void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t nrows = ggml_nrows(src0);
+
+    //const int n_past = ((int32_t *) dst->op_params)[0];
+    const int n_head = ((int32_t *) dst->op_params)[1];
+    float max_bias;
+    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
+
+    //GGML_ASSERT(ne01 + n_past == ne00);
+    GGML_ASSERT(n_head == ne02);
+
+    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
+
+    alibi_f32_cuda(src0_d, dst_d, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, stream);
+}

llama/ggml-cuda/alibi.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ALIBI_BLOCK_SIZE 32
+
+void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/arange.cu

@@ -0,0 +1,34 @@
+#include "arange.cuh"
+
+static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    dst[nidx] = start + step * nidx;
+}
+
+static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
+    arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start,  step);
+}
+
+void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    float start;
+    float stop;
+    float step;
+    memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
+    memcpy(&stop,  (float *)dst->op_params + 1, sizeof(float));
+    memcpy(&step,  (float *)dst->op_params + 2, sizeof(float));
+
+    int64_t steps = (int64_t)ceil((stop - start) / step);
+    GGML_ASSERT(ggml_nelements(dst) == steps);
+
+    arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
+}

llama/ggml-cuda/arange.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ARANGE_BLOCK_SIZE 256
+
+void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/argsort.cu

@@ -0,0 +1,103 @@
+#include "argsort.cuh"
+
+template<typename T>
+static inline __device__ void ggml_cuda_swap(T & a, T & b) {
+    T tmp = a;
+    a = b;
+    b = tmp;
+}
+
+template<ggml_sort_order order>
+static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
+    // bitonic sort
+    int col = threadIdx.x;
+    int row = blockIdx.y;
+
+    if (col >= ncols_pad) {
+        return;
+    }
+
+    const float * x_row = x + row * ncols;
+    extern __shared__ int dst_row[];
+
+    // initialize indices
+    dst_row[col] = col;
+
+    __syncthreads();
+
+    for (int k = 2; k <= ncols_pad; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= ncols ||
+                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
+                    }
+                } else {
+                    if (dst_row[ixj] >= ncols ||
+                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
+                    }
+                }
+            }
+            __syncthreads();
+        }
+    }
+
+    // copy the result to dst without the padding
+    if (col < ncols) {
+        dst[row * ncols + col] = dst_row[col];
+    }
+}
+
+static int next_power_of_2(int x) {
+    int n = 1;
+    while (n < x) {
+        n *= 2;
+    }
+    return n;
+}
+
+static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
+    // bitonic sort requires ncols to be power of 2
+    const int ncols_pad = next_power_of_2(ncols);
+
+    const dim3 block_dims(ncols_pad, 1, 1);
+    const dim3 block_nums(1, nrows, 1);
+    const size_t shared_mem = ncols_pad * sizeof(int);
+
+    GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+    } else if (order == GGML_SORT_ORDER_DESC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+    } else {
+        GGML_ASSERT(false);
+    }
+}
+
+void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
+
+    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
+}

llama/ggml-cuda/argsort.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/binbcast.cu

@@ -0,0 +1,280 @@
+#include "binbcast.cuh"
+
+static __device__ __forceinline__ float op_repeat(const float a, const float b) {
+    return b;
+    GGML_UNUSED(a);
+}
+
+static __device__ __forceinline__ float op_add(const float a, const float b) {
+    return a + b;
+}
+
+static __device__ __forceinline__ float op_mul(const float a, const float b) {
+    return a * b;
+}
+
+static __device__ __forceinline__ float op_div(const float a, const float b) {
+    return a / b;
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s00,*/ int s01, int s02, int s03,
+        /*int s10,*/ int s11, int s12, int s13) {
+    const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
+    const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
+    const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
+    const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
+
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
+        const int i10 = i0 % ne10;
+        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+    }
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s00,*/ int s01, int s02, int s03,
+        /*int s10,*/ int s11, int s12, int s13) {
+
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    const int i3 = i/(ne2*ne1*ne0);
+    const int i2 = (i/(ne1*ne0)) % ne2;
+    const int i1 = (i/ne0) % ne1;
+    const int i0 = i % ne0;
+
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    const int i10 = i0 % ne10;
+    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+}
+
+template<float (*bin_op)(const float, const float)>
+struct bin_bcast_cuda {
+    template<typename src0_t, typename src1_t, typename dst_t>
+    void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
+            const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
+            cudaStream_t stream) {
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        int nr0 = ne10/ne0;
+        int nr1 = ne11/ne1;
+        int nr2 = ne12/ne2;
+        int nr3 = ne13/ne3;
+
+        int nr[4] = { nr0, nr1, nr2, nr3 };
+
+        // collapse dimensions until first broadcast dimension
+        int64_t cne[] = {ne0, ne1, ne2, ne3};
+        int64_t cne0[] = {ne00, ne01, ne02, ne03};
+        int64_t cne1[] = {ne10, ne11, ne12, ne13};
+
+        size_t cnb[] = {nb0, nb1, nb2, nb3};
+        size_t cnb0[] = {nb00, nb01, nb02, nb03};
+        size_t cnb1[] = {nb10, nb11, nb12, nb13};
+
+        auto collapse = [](int64_t cne[]) {
+            cne[0] *= cne[1];
+            cne[1] = cne[2];
+            cne[2] = cne[3];
+            cne[3] = 1;
+        };
+
+        auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
+            cnb[1] *= cne[1];
+            cnb[2] *= cne[2];
+            cnb[3] *= cne[3];
+        };
+
+        if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+            for (int i = 0; i < 4; i++) {
+                if (nr[i] != 1) {
+                    break;
+                }
+                if (i > 0) {
+                    collapse_nb(cnb, cne);
+                    collapse_nb(cnb0, cne0);
+                    collapse_nb(cnb1, cne1);
+                    collapse(cne);
+                    collapse(cne0);
+                    collapse(cne1);
+                }
+            }
+        }
+
+        {
+            int64_t ne0 = cne[0];
+            int64_t ne1 = cne[1];
+            int64_t ne2 = cne[2];
+            int64_t ne3 = cne[3];
+
+            //int64_t ne00 = cne0[0]; GGML_UNUSED(ne00);
+            //int64_t ne01 = cne0[1]; GGML_UNUSED(ne01);
+            //int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
+            //int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
+
+            int64_t ne10 = cne1[0];
+            int64_t ne11 = cne1[1];
+            int64_t ne12 = cne1[2];
+            int64_t ne13 = cne1[3];
+
+            size_t nb0 = cnb[0];
+            size_t nb1 = cnb[1];
+            size_t nb2 = cnb[2];
+            size_t nb3 = cnb[3];
+
+            size_t nb00 = cnb0[0];
+            size_t nb01 = cnb0[1];
+            size_t nb02 = cnb0[2];
+            size_t nb03 = cnb0[3];
+
+            size_t nb10 = cnb1[0];
+            size_t nb11 = cnb1[1];
+            size_t nb12 = cnb1[2];
+            size_t nb13 = cnb1[3];
+
+            size_t s0 = nb0 / sizeof(dst_t);
+            size_t s1 = nb1 / sizeof(dst_t);
+            size_t s2 = nb2 / sizeof(dst_t);
+            size_t s3 = nb3 / sizeof(dst_t);
+
+            size_t s10 = nb10 / sizeof(src1_t);
+            size_t s11 = nb11 / sizeof(src1_t);
+            size_t s12 = nb12 / sizeof(src1_t);
+            size_t s13 = nb13 / sizeof(src1_t);
+
+            size_t s00 = nb00 / sizeof(src0_t);
+            size_t s01 = nb01 / sizeof(src0_t);
+            size_t s02 = nb02 / sizeof(src0_t);
+            size_t s03 = nb03 / sizeof(src0_t);
+
+            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
+
+            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
+
+            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
+
+            GGML_ASSERT(s0 == 1);
+            GGML_ASSERT(s00 == 1);
+            GGML_ASSERT(s10 == 1);
+
+            const int block_size = 128;
+
+            int64_t hne0 = std::max(ne0/2LL, 1LL);
+
+            dim3 block_dims;
+            block_dims.x = std::min<unsigned int>(hne0, block_size);
+            block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
+            block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
+
+            dim3 block_nums(
+                (hne0 + block_dims.x - 1) / block_dims.x,
+                (ne1 + block_dims.y - 1) / block_dims.y,
+                (ne2*ne3 + block_dims.z - 1) / block_dims.z
+            );
+
+            if (block_nums.z > 65535) {
+                // this is the maximum number of blocks in z dimension, fallback to 1D grid kernel
+                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
+                k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd,
+                    ne0, ne1, ne2, ne3,
+                    ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s00, */ s01, s02, s03,
+                    /* s10, */ s11, s12, s13);
+            } else {
+                k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd,
+                    ne0, ne1, ne2, ne3,
+                    ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s00, */ s01, s02, s03,
+                    /* s10, */ s11, s12, s13);
+            }
+        }
+    }
+};
+
+template<class op>
+static void ggml_cuda_op_bin_bcast(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) {
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
+    } else {
+        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
+            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ASSERT(false);
+    }
+}
+
+void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, dst->src[0], dst, nullptr, dst->src[0]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}

llama/ggml-cuda/binbcast.cuh

@@ -0,0 +1,6 @@
+#include "common.cuh"
+
+void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/clamp.cu

@@ -0,0 +1,34 @@
+#include "clamp.cuh"
+
+static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
+}
+
+static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
+    clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
+}
+
+
+void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float min;
+    float max;
+    memcpy(&min, dst->op_params, sizeof(float));
+    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
+
+    clamp_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream);
+}

llama/ggml-cuda/clamp.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CLAMP_BLOCK_SIZE 256
+
+void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/common.cuh

@@ -0,0 +1,683 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cuda.h"
+
+#include <memory>
+
+#if defined(GGML_USE_HIPBLAS)
+#define GGML_COMMON_DECL_HIP
+#define GGML_COMMON_IMPL_HIP
+#else
+#define GGML_COMMON_DECL_CUDA
+#define GGML_COMMON_IMPL_CUDA
+#endif
+#include "ggml-common.h"
+
+#include <cstdio>
+#include <array>
+#include <cassert>
+#include <cfloat>
+#include <string>
+#include <vector>
+
+#if defined(GGML_USE_HIPBLAS)
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
+#define cublasCreate hipblasCreate
+#define cublasDestroy hipblasDestroy
+#define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
+#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaHostRegister hipHostRegister
+#define cudaHostRegisterPortable hipHostRegisterPortable
+#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
+#define cudaHostUnregister hipHostUnregister
+#define cudaLaunchHostFunc hipLaunchHostFunc
+#ifdef GGML_HIP_UMA
+#define cudaMalloc hipMallocManaged
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
+#else
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#endif
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaMemsetAsync hipMemsetAsync
+#define cudaMemGetInfo hipMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamDestroy hipStreamDestroy
+#define cudaStreamFireAndForget hipStreamFireAndForget
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#define __trap abort
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
+#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
+#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
+#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
+#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
+#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
+#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
+#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+#else
+#include <cuda_runtime.h>
+#include <cuda.h>
+#include <cublas_v2.h>
+#include <cuda_fp16.h>
+
+#if CUDART_VERSION < 11020
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#endif // CUDART_VERSION < 11020
+
+#endif // defined(GGML_USE_HIPBLAS)
+
+#define STRINGIZE_IMPL(...) #__VA_ARGS__
+#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
+
+#define WARP_SIZE 32
+#define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
+#define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
+
+#define CC_PASCAL     600
+#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define CC_VOLTA      700
+#define CC_AMPERE     800
+#define CC_OFFSET_AMD 1000000
+#define CC_RDNA1      (CC_OFFSET_AMD + 1010)
+#define CC_RDNA2      (CC_OFFSET_AMD + 1030)
+#define CC_RDNA3      (CC_OFFSET_AMD + 1100)
+
+// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication
+// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant
+// for large computational tasks. the drawback is that this requires some extra amount of VRAM:
+// -  7B quantum model: +100-200 MB
+// - 13B quantum model: +200-400 MB
+//
+//#define GGML_CUDA_FORCE_MMQ
+
+// TODO: improve this to be correct for more hardware
+//       for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
+#if !defined(GGML_CUDA_FORCE_MMQ)
+#define CUDA_USE_TENSOR_CORES
+#endif
+
+#define MMVQ_MAX_BATCH_SIZE  8 // max batch size to use MMVQ kernels
+#define  MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available
+
+#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#define GGML_CUDA_MAX_STREAMS 8
+
+[[noreturn]]
+void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
+
+#define CUDA_CHECK_GEN(err, success, error_fn)                                      \
+     do {                                                                           \
+        auto err_ = (err);                                                          \
+        if (err_ != (success)) {                                                    \
+            ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_));    \
+        }                                                                           \
+    } while (0)
+
+#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
+
+#if CUDART_VERSION >= 12000
+    static const char * cublas_get_error_str(const cublasStatus_t err) {
+        return cublasGetStatusString(err);
+    }
+#else
+    static const char * cublas_get_error_str(const cublasStatus_t err) {
+        switch (err) {
+            case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
+            case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
+            case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
+            case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
+            case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
+            case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
+            case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
+            case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
+            case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
+            default: return "unknown error";
+        }
+    }
+#endif // CUDART_VERSION >= 12000
+
+#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
+
+#if !defined(GGML_USE_HIPBLAS)
+static const char * cu_get_error_str(CUresult err) {
+    const char * err_str;
+    cuGetErrorString(err, &err_str);
+    return err_str;
+}
+#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
+#endif
+
+#if CUDART_VERSION >= 11100
+#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
+#else
+#define GGML_CUDA_ASSUME(x)
+#endif // CUDART_VERSION >= 11100
+
+#ifdef GGML_CUDA_F16
+typedef half dfloat; // dequantize float
+typedef half2 dfloat2;
+#else
+typedef float dfloat; // dequantize float
+typedef float2 dfloat2;
+#endif //GGML_CUDA_F16
+
+#if defined(GGML_USE_HIPBLAS)
+#define __CUDA_ARCH__ 1300
+
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+    defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3
+#endif
+
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
+#define RDNA2
+#endif
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int &>(c);
+#else
+    int8x4_t c;
+    int16_t tmp;
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+        tmp = va[i] - vb[i];
+        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+        c[i] = tmp;
+    }
+    return reinterpret_cast<int &>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
+}
+
+static __device__ __forceinline__ int __vsub4(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
+static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
+#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
+    c = __builtin_amdgcn_sdot4(a, b, c, false);
+#elif defined(RDNA3)
+    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+#elif defined(__gfx1010__) || defined(__gfx900__)
+    int tmp1;
+    int tmp2;
+    asm("\n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        "
+        : "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
+        : "v"(a), "v"(b)
+    );
+#else
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+    c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
+#endif
+    return c;
+}
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+#endif // defined(GGML_USE_HIPBLAS)
+
+#define FP16_AVAILABLE (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+
+#define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+
+static bool fast_fp16_available(const int cc) {
+    return cc >= CC_PASCAL && cc != 610;
+}
+
+static bool fp16_mma_available(const int cc) {
+    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
+}
+
+[[noreturn]]
+static __device__ void no_device_code(
+    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
+           file_name, line, function_name, arch);
+    GGML_UNUSED(arch_list);
+#else
+    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
+           file_name, line, function_name, arch, arch_list);
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    __trap();
+
+    GGML_UNUSED(no_device_code); // suppress unused function warning
+}
+
+#ifdef __CUDA_ARCH__
+#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
+#else
+#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
+#endif // __CUDA_ARCH__
+
+static __device__ __forceinline__ float warp_reduce_sum(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    }
+    return x;
+}
+
+static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    }
+    return a;
+}
+
+static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+#if FP16_AVAILABLE
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
+        reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
+        reinterpret_cast<half&>(a.y) += __high2half(a_other);
+    }
+    return a;
+#else
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    }
+    return a;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+#else
+    NO_DEVICE_CODE;
+    return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ float warp_reduce_max(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    }
+    return x;
+}
+
+static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
+#if FP16_AVAILABLE
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+    return __float2half(fmaxf(__half2float(a), __half2float(b)));
+#else
+    return __hmax(a, b);
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+
+#else
+   NO_DEVICE_CODE;
+   GGML_UNUSED(b);
+   return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+
+#if CUDART_VERSION >= CUDART_HMAX
+    return __hmax2(a, b);
+#else
+    half2 ret;
+    reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
+    reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
+    return ret;
+#endif // CUDART_VERSION >= CUDART_HMAX
+
+#else
+    GGML_UNUSED(a);
+    GGML_UNUSED(b);
+    NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+}
+
+static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#pragma unroll
+   for (int mask = 16; mask > 0; mask >>= 1) {
+       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+   }
+   return x;
+#else
+   GGML_UNUSED(x);
+   NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+}
+
+#if CUDART_VERSION < CUDART_HMASK
+static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
+    const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
+    const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
+    return mask_low | mask_high;
+}
+#endif // CUDART_VERSION < 12000
+
+// TODO: move to ggml-common.h
+static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
+typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
+
+static __device__ __forceinline__ float get_alibi_slope(
+    const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return powf(base, exph);
+}
+
+//////////////////////
+
+struct ggml_cuda_device_info {
+    int device_count;
+
+    struct cuda_device_info {
+        int     cc;                 // compute capability
+        int     nsm;                // number of streaming multiprocessors
+        size_t  smpb;               // max. shared memory per block
+        bool    vmm;                // virtual memory support
+        size_t  vmm_granularity;    // granularity of virtual memory
+        size_t  total_vram;
+    };
+
+    cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {};
+
+    std::array<float, GGML_CUDA_MAX_DEVICES> default_tensor_split = {};
+};
+
+const ggml_cuda_device_info & ggml_cuda_info();
+
+void ggml_cuda_set_device(int device);
+int ggml_cuda_get_device();
+
+struct ggml_cuda_pool {
+    virtual ~ggml_cuda_pool() = default;
+
+    virtual void * alloc(size_t size, size_t * actual_size) = 0;
+    virtual void free(void * ptr, size_t size) = 0;
+};
+
+template<typename T>
+struct ggml_cuda_pool_alloc {
+    ggml_cuda_pool * pool = nullptr;
+    T * ptr = nullptr;
+    size_t actual_size = 0;
+
+    ggml_cuda_pool_alloc() = default;
+
+    explicit ggml_cuda_pool_alloc(ggml_cuda_pool & pool) : pool(&pool) {
+    }
+
+    ggml_cuda_pool_alloc(ggml_cuda_pool & pool, size_t size) : pool(&pool) {
+        alloc(size);
+    }
+
+    ~ggml_cuda_pool_alloc() {
+        if (ptr != nullptr) {
+            pool->free(ptr, actual_size);
+        }
+    }
+
+    // size is in number of elements
+    T * alloc(size_t size) {
+        GGML_ASSERT(pool != nullptr);
+        GGML_ASSERT(ptr == nullptr);
+        ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
+        return ptr;
+    }
+
+    T * alloc(ggml_cuda_pool & pool, size_t size) {
+        this->pool = &pool;
+        return alloc(size);
+    }
+
+    T * get() {
+        return ptr;
+    }
+
+    ggml_cuda_pool_alloc(const ggml_cuda_pool_alloc &) = delete;
+    ggml_cuda_pool_alloc(ggml_cuda_pool_alloc &&) = delete;
+    ggml_cuda_pool_alloc& operator=(const ggml_cuda_pool_alloc &) = delete;
+    ggml_cuda_pool_alloc& operator=(ggml_cuda_pool_alloc &&) = delete;
+};
+
+
+// backend interface
+
+struct ggml_tensor_extra_gpu {
+    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
+    cudaEvent_t events[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS]; // events for synchronizing multiple GPUs
+};
+
+
+#if (CUDART_VERSION >= 12000) && defined(GGML_CUDA_USE_GRAPHS)
+#define USE_CUDA_GRAPH
+#endif
+
+struct ggml_graph_node_properties {
+    void * node_address;
+    ggml_op node_op;
+    int64_t ne[GGML_MAX_DIMS];
+    size_t nb[GGML_MAX_DIMS];
+    void * src_address[GGML_MAX_SRC];
+};
+
+struct ggml_cuda_graph {
+#ifdef USE_CUDA_GRAPH
+    ~ggml_cuda_graph() {
+        if (instance != nullptr) {
+            CUDA_CHECK(cudaGraphExecDestroy(instance));
+        }
+        if (graph != nullptr) {
+            CUDA_CHECK(cudaGraphDestroy(graph));
+        }
+    }
+    cudaGraph_t graph = nullptr;
+    cudaGraphExec_t instance = nullptr;
+    size_t num_nodes = 0;
+    std::vector<cudaGraphNode_t> nodes;
+    std::vector<cudaKernelNodeParams> params;
+    bool disable_due_to_gpu_arch = false;
+    bool disable_due_to_too_many_updates = false;
+    bool disable_due_to_failed_graph_capture = false;
+    int number_consecutive_updates = 0;
+    std::vector<ggml_graph_node_properties> ggml_graph_properties;
+    std::vector<char **> updated_kernel_arg;
+#endif
+};
+
+struct ggml_backend_cuda_context {
+    int device;
+    std::string name;
+    cudaEvent_t copy_event = nullptr;
+
+    cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
+    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
+
+    std::unique_ptr<ggml_cuda_graph> cuda_graph;
+
+    explicit ggml_backend_cuda_context(int device) :
+        device(device),
+        name(GGML_CUDA_NAME + std::to_string(device)) {
+    }
+
+    ~ggml_backend_cuda_context() {
+        if (copy_event != nullptr) {
+            CUDA_CHECK(cudaEventDestroy(copy_event));
+        }
+        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
+            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
+                if (streams[i][j] != nullptr) {
+                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
+                }
+            }
+            if (cublas_handles[i] != nullptr) {
+                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+            }
+        }
+    }
+
+    cudaStream_t stream(int device, int stream) {
+        if (streams[device][stream] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUDA_CHECK(cudaStreamCreateWithFlags(&streams[device][stream], cudaStreamNonBlocking));
+        }
+        return streams[device][stream];
+    }
+
+    cudaStream_t stream() {
+        return stream(device, 0);
+    }
+
+    cublasHandle_t cublas_handle(int device) {
+        if (cublas_handles[device] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUBLAS_CHECK(cublasCreate(&cublas_handles[device]));
+            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH));
+        }
+        return cublas_handles[device];
+    }
+
+    cublasHandle_t cublas_handle() {
+        return cublas_handle(device);
+    }
+
+    // pool
+    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
+
+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
+
+    ggml_cuda_pool & pool(int device) {
+        if (pools[device] == nullptr) {
+            pools[device] = new_pool_for_device(device);
+        }
+        return *pools[device];
+    }
+
+    ggml_cuda_pool & pool() {
+        return pool(device);
+    }
+};

llama/ggml-cuda/concat.cu

@@ -0,0 +1,49 @@
+#include "concat.cuh"
+
+static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (blockIdx.z < ne02) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            blockIdx.z * ne0 * gridDim.y;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            (blockIdx.z - ne02) * ne0 *  gridDim.y;
+        dst[offset_dst] = y[offset_src];
+    }
+}
+
+static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2);
+    concat_f32<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+}
+
+void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+        concat_f32_cuda(src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4), dst_d + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], stream);
+    }
+}

llama/ggml-cuda/concat.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CONCAT_BLOCK_SIZE 256
+
+void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/convert.cu

@@ -0,0 +1,824 @@
+#include "convert.cuh"
+#include "dequantize.cuh"
+
+#define CUDA_Q8_0_NE_ALIGN 2048
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
+    const int64_t i = (int64_t)2*(blockDim.x*blockIdx.x + threadIdx.x);
+
+    if (i >= k) {
+        return;
+    }
+
+    const int64_t ib = i/qk; // block index
+    const int64_t iqs = (i%qk)/qr; // quant index
+    const int64_t iybs = i - i%qk; // y block start index
+    const int64_t y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(vx, ib, iqs, v);
+
+    y[iybs + iqs + 0]        = v.x;
+    y[iybs + iqs + y_offset] = v.y;
+}
+
+template <bool need_check>
+static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) {
+#if __CUDA_ARCH__ >= CC_PASCAL
+    constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
+
+    const int64_t   i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x;
+    const int * x0 = ((int *) vx) + blockIdx.x * nint;
+    half2 * y2 = (half2 *) (y + i0);
+
+    __shared__ int vals[nint];
+
+#pragma unroll
+    for (int ix0 = 0; ix0 < nint; ix0 += WARP_SIZE) {
+        if (need_check && i0*sizeof(block_q8_0)/QK8_0 + sizeof(int)*(ix0 + threadIdx.x) >= k*sizeof(block_q8_0)/QK8_0) {
+            break;
+        }
+
+        const int ix = ix0 + threadIdx.x;
+        vals[ix] = x0[ix];
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) {
+        if (need_check && i0 + iy + 2*threadIdx.x >= k) {
+            return;
+        }
+
+        const half * b0 = ((const half  *) vals) + (sizeof(block_q8_0)/sizeof(half)) * ((iy + 2*threadIdx.x)/QK8_0);
+        const half    d = *b0;
+        const char2  qs = ((const char2 *) (b0 + 1))[threadIdx.x % (QK8_0/2)];
+
+        y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
+    }
+#else
+    GGML_UNUSED(vx);
+    GGML_UNUSED(y);
+    GGML_UNUSED(k);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_PASCAL
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
+
+    const int64_t i = blockIdx.x;
+
+    // assume 32 threads
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_0 * x = (const block_q4_0 *)vx + ib;
+    const float d = __half2float(x->d);
+    const float dm = -8*d;
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l+ 0] = d * (q[l] & 0xF) + dm;
+        y[l+16] = d * (q[l] >>  4) + dm;
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
+
+    const int64_t i = blockIdx.x;
+
+    // assume 32 threads
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_1 * x = (const block_q4_1 *)vx + ib;
+    const float2 d = __half22float2(x->dm);
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l+ 0] = d.x * (q[l] & 0xF) + d.y;
+        y[l+16] = d.x * (q[l] >>  4) + d.y;
+    }
+}
+
+//================================== k-quants
+
+template<typename dst_t>
+static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_q2_K * x = (const block_q2_K *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t n   = tid/32;
+    const int64_t l   = tid - 32*n;
+    const int64_t is  = 8*n + l/16;
+
+    const uint8_t q = x[i].qs[32*n + l];
+    dst_t * y = yy + i*QK_K + 128*n;
+
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
+    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
+    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
+    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
+#else
+    const int64_t is = tid/16;  // 0 or 1
+    const int64_t il = tid%16;  // 0...15
+    const uint8_t q = x[i].qs[il] >> (2*is);
+    dst_t * y = yy + i*QK_K + 16*is + il;
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
+    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i = blockIdx.x;
+    const block_q3_K * x = (const block_q3_K *) vx;
+
+#if QK_K == 256
+    const int64_t r = threadIdx.x/4;
+    const int64_t tid = r/2;
+    const int64_t is0 = r%2;
+    const int64_t l0 = 16*is0 + 4*(threadIdx.x%4);
+    const int64_t n = tid / 4;
+    const int64_t j = tid - 4*n;
+
+    uint8_t m = 1 << (4*n + j);
+    int64_t is = 8*n + 2*j + is0;
+    int shift = 2*j;
+
+    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
+                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
+                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
+                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
+    float d_all = x[i].d;
+    float dl = d_all * (us - 32);
+
+    dst_t * y = yy + i*QK_K + 128*n + 32*j;
+    const uint8_t * q = x[i].qs + 32*n;
+    const uint8_t * hm = x[i].hmask;
+
+    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+#else
+    const int64_t tid = threadIdx.x;
+    const int64_t is  = tid/16;  // 0 or 1
+    const int64_t il  = tid%16;  // 0...15
+    const int64_t im  = il/8;    // 0...1
+    const int64_t in  = il%8;    // 0...7
+
+    dst_t * y = yy + i*QK_K + 16*is + il;
+
+    const uint8_t q = x[i].qs[il] >> (2*is);
+    const uint8_t h = x[i].hmask[in] >> (2*is + im);
+    const float   d = (float)x[i].d;
+
+    if (is == 0) {
+        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
+        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
+    } else {
+        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
+        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
+    }
+#endif
+
+}
+
+#if QK_K == 256
+static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63; m = q[j + 4] & 63;
+    } else {
+        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+#endif
+
+template<typename dst_t>
+static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q4_K * x = (const block_q4_K *) vx;
+
+    const int64_t i = blockIdx.x;
+
+#if QK_K == 256
+    // assume 32 threads
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t is  = 2*il;
+    const int64_t n   = 4;
+
+    dst_t * y = yy + i*QK_K + 64*il + n*ir;
+
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
+
+    const uint8_t * q = x[i].qs + 32*il + n*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+    for (int l = 0; l < n; ++l) {
+        y[l + 0] = d1 * (q[l] & 0xF) - m1;
+        y[l +32] = d2 * (q[l] >>  4) - m2;
+    }
+#else
+    const int64_t tid = threadIdx.x;
+    const uint8_t * q = x[i].qs;
+    dst_t * y = yy + i*QK_K;
+    const float d = (float)x[i].dm[0];
+    const float m = (float)x[i].dm[1];
+    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
+    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
+#endif
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q5_K * x = (const block_q5_K *) vx;
+
+    const int64_t i = blockIdx.x;
+
+#if QK_K == 256
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/16;   // il is in 0...3
+    const int64_t ir  = tid%16;   // ir is in 0...15
+    const int64_t is  = 2*il;     // is is in 0...6
+
+    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
+
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
+
+    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
+    const uint8_t * qh = x[i].qh + 2*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+
+    uint8_t   hm  = 1 << (2*il);
+    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
+    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
+    hm <<= 1;
+    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
+    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
+#else
+    const int64_t tid = threadIdx.x;
+    const uint8_t q = x[i].qs[tid];
+    const int64_t im = tid/8;  // 0...3
+    const int64_t in = tid%8;  // 0...7
+    const int64_t is = tid/16; // 0 or 1
+    const uint8_t h = x[i].qh[in] >> im;
+    const float d = x[i].d;
+    dst_t * y = yy + i*QK_K + tid;
+    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
+    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
+#endif
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q6_K * x = (const block_q6_K *) vx;
+
+    const int64_t i = blockIdx.x;
+#if QK_K == 256
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t tid = threadIdx.x;
+    const int64_t ip  = tid/32;   // ip is 0 or 1
+    const int64_t il  = tid - 32*ip; // 0...32
+    const int64_t is  = 8*ip + il/16;
+
+    dst_t * y = yy + i*QK_K + 128*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t * ql = x[i].ql + 64*ip + il;
+    const uint8_t   qh = x[i].qh[32*ip + il];
+    const int8_t  * sc = x[i].scales + is;
+
+    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+#else
+
+    // assume 32 threads
+    const int64_t tid = threadIdx.x;
+    const int64_t ip  = tid/16;         // 0 or 1
+    const int64_t il  = tid - 16*ip;    // 0...15
+
+    dst_t * y = yy + i*QK_K + 16*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t   ql = x[i].ql[16*ip + il];
+    const uint8_t   qh = x[i].qh[il] >> (2*ip);
+    const int8_t  * sc = x[i].scales;
+
+    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+#endif
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq2_xs * x = (const block_iq2_xs *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq2_s * x = (const block_iq2_s *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t  * q3 = x[i].qs + 8*ib;
+    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq3_s * x = (const block_iq3_s *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * qs = x[i].qs + 8*ib;
+    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
+    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
+    const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
+    const uint8_t signs = x[i].signs[4*ib + il];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq1_s * x = (const block_iq1_s  *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq1_m * x = (const block_iq1_m  *) vx;
+
+    const int64_t tid = threadIdx.x;
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * sc = (const uint16_t *)x[i].scales;
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
+    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
+    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+#else
+    NO_DEVICE_CODE;
+#endif
+
+}
+
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = (float)x[ib].d;
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+
+}
+
+#if QK_K != 64
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i   = blockIdx.x;
+    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
+    const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+}
+#endif
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
+    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+}
+
+static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN;
+    if (k % CUDA_Q8_0_NE_ALIGN == 0) {
+        const bool need_check = false;
+        dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
+    } else {
+        const bool need_check = true;
+        dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
+    }
+}
+
+template<typename dst_t>
+static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
+#else
+    dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
+#else
+    dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb32 = k / 32;
+    const int nb = (k + 255) / 256;
+    dequantize_block_q4_0<<<nb, 32, 0, stream>>>(vx, y, nb32);
+}
+
+template<typename dst_t>
+static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb32 = k / 32;
+    const int nb = (k + 255) / 256;
+    dequantize_block_q4_1<<<nb, 32, 0, stream>>>(vx, y, nb32);
+}
+
+template<typename dst_t>
+static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
+#else
+    dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
+#else
+    dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+#if QK_K == 64
+    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
+#else
+    dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
+#endif
+}
+
+template <typename src_t, typename dst_t>
+static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
+    const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    const src_t * x = (src_t *) vx;
+
+    y[i] = x[i];
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+}
+
+to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= CC_PASCAL) {
+                return dequantize_block_q8_0_f16_cuda;
+            }
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_cuda;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_cuda;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_cuda;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_cuda;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_cuda;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_cuda;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_cuda;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_cuda;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_cuda;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_cuda;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_cuda;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_cuda;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half>;
+        default:
+            return nullptr;
+    }
+}

llama/ggml-cuda/convert.cuh

@@ -0,0 +1,13 @@
+#include "common.cuh"
+
+#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
+
+template<typename T>
+using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
+
+typedef to_t_cuda_t<float> to_fp32_cuda_t;
+typedef to_t_cuda_t<half> to_fp16_cuda_t;
+
+to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
+
+to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);

llama/ggml-cuda/cpy.cu

@@ -0,0 +1,490 @@
+#include "cpy.cuh"
+
+typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
+
+static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    float * dsti = (float *) cdsti;
+
+    *dsti = *xi;
+}
+
+static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    half * dsti = (half *) cdsti;
+
+    *dsti = __float2half(*xi);
+}
+
+static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
+    const half * xi = (const half *) cxi;
+    half * dsti = (half *) cdsti;
+
+    *dsti = *xi;
+}
+
+static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
+    const half * xi = (const half *) cxi;
+    float * dsti = (float *) cdsti;
+
+    *dsti = *xi;
+}
+
+template <cpy_kernel_t cpy_1>
+static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
+                                   const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                                   const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13) {
+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= ne) {
+        return;
+    }
+
+    // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
+    // then combine those indices with the corresponding byte offsets to get the total offsets
+    const int64_t i03 = i/(ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int64_t i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int64_t x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int64_t i13 = i/(ne10 * ne11 * ne12);
+    const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
+
+    cpy_1(cx + x_offset, cdst + dst_offset);
+}
+
+static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q8_0 * dsti = (block_q8_0 *) cdsti;
+
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = xi[j];
+        amax = fmaxf(amax, fabsf(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = xi[j]*id;
+
+        dsti->qs[j] = roundf(x0);
+    }
+}
+
+static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_0 * dsti = (block_q4_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = xi[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = xi[0       + j]*id;
+        const float x1 = xi[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_1 * dsti = (block_q4_1 *) cdsti;
+
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = xi[j];
+
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->dm.x = d;
+    dsti->dm.y = vmin;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (xi[0       + j] - vmin)*id;
+        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q5_0 * dsti = (block_q5_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK5_0; ++j) {
+        const float v = xi[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = xi[0       + j]*id;
+        const float x1 = xi[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
+
+        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+    memcpy(dsti->qh, &qh, sizeof(qh));
+}
+
+static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q5_1 * dsti = (block_q5_1 *) cdsti;
+
+    float min = xi[0];
+    float max = xi[0];
+
+    for (int j = 1; j < QK5_1; ++j) {
+        const float v = xi[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d  = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->dm.x = d;
+    dsti->dm.y = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (xi[0       + j] - min)*id;
+        const float x1 = (xi[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+    memcpy(dsti->qh, &qh, sizeof(qh));
+}
+
+
+static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_NL; ++j) {
+        const float v = xi[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    float d = vmax / kvalues_iq4nl[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = xi[0        + j]*id;
+        const float x1 = xi[QK4_NL/2 + j]*id;
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
+        dsti->qs[j] = xi0 | (xi1 << 4);
+        const float v0 = kvalues_iq4nl[xi0];
+        const float v1 = kvalues_iq4nl[xi1];
+        const float w0 = xi[0        + j]*xi[0        + j];
+        const float w1 = xi[QK4_NL/2 + j]*xi[QK4_NL/2 + j];
+        sumqx += w0*v0*xi[j] + w1*v1*xi[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+    }
+
+    dsti->d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
+template <cpy_kernel_t cpy_blck, int qk>
+static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
+                                 const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                                 const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                 const int nb12, const int nb13) {
+    const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i03 = i/(ne00 * ne01 * ne02);
+    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int i13 = i/(ne10 * ne11 * ne12);
+    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
+
+    cpy_blck(cx + x_offset, cdst + dst_offset);
+}
+
+static void ggml_cpy_f16_f32_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_f32_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_f16_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q8_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK8_0 == 0);
+    const int num_blocks = ne / QK8_0;
+    cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q4_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_0 == 0);
+    const int num_blocks = ne / QK4_0;
+    cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q4_1_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_1 == 0);
+    const int num_blocks = ne / QK4_1;
+    cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q5_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK5_0 == 0);
+    const int num_blocks = ne / QK5_0;
+    cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q5_1_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK5_1 == 0);
+    const int num_blocks = ne / QK5_1;
+    cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_iq4_nl_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_NL == 0);
+    const int num_blocks = ne / QK4_NL;
+    cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f16_f16_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+    const int64_t ne = ggml_nelements(src0);
+    GGML_ASSERT(ne == ggml_nelements(src1));
+
+    GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
+    GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+
+    //GGML_ASSERT(src0->ne[3] == 1);
+
+    const int64_t nb00 = src0->nb[0];
+    const int64_t nb01 = src0->nb[1];
+    const int64_t nb02 = src0->nb[2];
+    const int64_t nb03 = src0->nb[3];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+
+    //GGML_ASSERT(src1->ne[3] == 1);
+
+    const int64_t nb10 = src1->nb[0];
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2];
+    const int64_t nb13 = src1->nb[3];
+
+    cudaStream_t main_stream = ctx.stream();
+
+    char * src0_ddc = (char *) src0->data;
+    char * src1_ddc = (char *) src1->data;
+
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
+        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
+        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
+        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else {
+        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ASSERT(false);
+    }
+}
+
+void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    ggml_cuda_cpy(ctx, src0, dst);
+}
+
+void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+            return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
+            return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
+            return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
+            return (void*) cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
+            return (void*) cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
+            return (void*) cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
+            return (void*) cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
+            return (void*) cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+            return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+            return (void*) cpy_f32_f16<cpy_1_f16_f32>;
+    } else {
+        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ASSERT(false);
+    }
+}
+

llama/ggml-cuda/cpy.cuh

@@ -0,0 +1,9 @@
+#include "common.cuh"
+
+#define CUDA_CPY_BLOCK_SIZE 32
+
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+
+void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1);

llama/ggml-cuda/dequantize.cuh

@@ -0,0 +1,103 @@
+#include "common.cuh"
+
+static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
+
+#ifdef GGML_CUDA_F16
+    v = __hsub2(v, {8.0f, 8.0f});
+    v = __hmul2(v, {d, d});
+#else
+    v.x = (v.x - 8.0f) * d;
+    v.y = (v.y - 8.0f) * d;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
+
+#ifdef GGML_CUDA_F16
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+#else
+    v.x = (v.x * d) + m;
+    v.y = (v.y * d) + m;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
+
+#ifdef GGML_CUDA_F16
+    v = __hsub2(v, {16.0f, 16.0f});
+    v = __hmul2(v, {d, d});
+#else
+    v.x = (v.x - 16.0f) * d;
+    v.y = (v.y - 16.0f) * d;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
+
+#ifdef GGML_CUDA_F16
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+#else
+    v.x = (v.x * d) + m;
+    v.y = (v.y * d) + m;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    v.x = x[ib].qs[iqs + 0];
+    v.y = x[ib].qs[iqs + 1];
+
+#ifdef GGML_CUDA_F16
+    v = __hmul2(v, {d, d});
+#else
+    v.x *= d;
+    v.y *= d;
+#endif // GGML_CUDA_F16
+}

llama/ggml-cuda/diagmask.cu

@@ -0,0 +1,40 @@
+#include "diagmask.cuh"
+
+static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
+    const int col = blockDim.y*blockIdx.y + threadIdx.y;
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int i = row*ncols + col;
+    //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i];
+    //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
+    dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
+}
+
+static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) {
+    const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1);
+    const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE;
+    const dim3 block_nums(nrows_x, block_num_x, 1);
+    diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
+}
+
+void ggml_cuda_op_diag_mask_inf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int nrows0 = ggml_nrows(src0);
+
+    const int n_past = ((int32_t *) dst->op_params)[0];
+
+    diag_mask_inf_f32_cuda(src0_d, dst_d, ne00, nrows0, ne01, n_past, stream);
+}

llama/ggml-cuda/diagmask.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
+
+void ggml_cuda_op_diag_mask_inf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/dmmv.cu

@@ -0,0 +1,813 @@
+#include "dmmv.cuh"
+#include "dequantize.cuh"
+#include "convert.cuh"
+
+#ifndef K_QUANTS_PER_ITERATION
+#define K_QUANTS_PER_ITERATION 2
+#else
+static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
+#endif
+
+static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q2_K * x = (const block_q2_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+#if QK_K == 256
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
+    const int q_offset = 32*im + l0;
+    const int s_offset = 8*im;
+    const int y_offset = 128*im + l0;
+
+    uint32_t aux[4];
+    const uint8_t * d = (const uint8_t *)aux;
+    const uint8_t * m = (const uint8_t *)(aux + 2);
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y = yy + i * QK_K + y_offset;
+        const uint8_t * q = x[i].qs + q_offset;
+
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
+
+        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
+        aux[0] = a[0] & 0x0f0f0f0f;
+        aux[1] = a[1] & 0x0f0f0f0f;
+        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
+        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
+
+        float sum1 = 0, sum2 = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
+            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
+
+        }
+        tmp += dall * sum1 - dmin * sum2;
+
+    }
+#else
+    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
+    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
+    const int offset = tid * K_QUANTS_PER_ITERATION;
+
+    uint32_t uaux[2];
+    const uint8_t * d = (const uint8_t *)uaux;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+
+        const float   * y = yy + i * QK_K + offset;
+        const uint8_t * q = x[i].qs + offset;
+        const uint32_t * s = (const uint32_t *)x[i].scales;
+
+        uaux[0] = s[0] & 0x0f0f0f0f;
+        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
+
+        const float2 dall = __half22float2(x[i].dm);
+
+        float sum1 = 0, sum2 = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            const uint8_t ql = q[l];
+            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
+                  + y[l+16] * d[1] * ((ql >> 2) & 3)
+                  + y[l+32] * d[2] * ((ql >> 4) & 3)
+                  + y[l+48] * d[3] * ((ql >> 6) & 3);
+            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
+        }
+        tmp += dall.x * sum1 - dall.y * sum2;
+    }
+#endif
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q3_K * x = (const block_q3_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+#if QK_K == 256
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
+
+    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
+    const int step = 16/K_QUANTS_PER_ITERATION;
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0....15 or 0...7
+
+    const uint8_t m = 1 << (4*im);
+
+    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
+    const int q_offset =  32*im + l0;
+    const int y_offset = 128*im + l0;
+
+    uint16_t utmp[4];
+    const int8_t * s = (const int8_t *)utmp;
+
+    const uint16_t s_shift = 4*im;
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + y_offset;
+        const uint8_t * q = x[i].qs + q_offset;
+        const uint8_t * h = x[i].hmask + l0;
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
+        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
+
+        const float d = x[i].d;
+
+        float sum = 0;
+        for (int l = 0; l < n; ++l) {
+            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
+            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
+        }
+        tmp += d * sum;
+
+    }
+#else
+
+    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
+    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
+    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
+    const int in = offset/8;                                 // 0 or 1
+    const int im = offset%8;                                 // 0...7
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+
+        const float   * y = yy + i * QK_K + offset;
+        const uint8_t * q = x[i].qs + offset;
+        const uint8_t * s = x[i].scales;
+
+        const float dall = (float)x[i].d;
+
+        float sum = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            const uint8_t hl = x[i].hmask[im+l] >> in;
+            const uint8_t ql = q[l];
+            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
+                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
+                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
+                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
+        }
+        tmp += sum;
+    }
+#endif
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q4_K * x = (const block_q4_K *)vx + ib0;
+
+#if QK_K == 256
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
+
+    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
+
+    const int il  = tid/step;                            // 0...3
+    const int ir  = tid - step*il;                       // 0...7 or 0...3
+    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+#if K_QUANTS_PER_ITERATION == 2
+    uint32_t q32[4];
+    const uint8_t * q4 = (const uint8_t *)q32;
+#else
+    uint16_t q16[4];
+    const uint8_t * q4 = (const uint8_t *)q16;
+#endif
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y1 = yy + i*QK_K + y_offset;
+        const float   * y2 = y1 + 128;
+
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+#if K_QUANTS_PER_ITERATION == 2
+        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
+        const uint32_t * q2 = q1 + 16;
+
+        q32[0] = q1[0] & 0x0f0f0f0f;
+        q32[1] = q1[0] & 0xf0f0f0f0;
+        q32[2] = q2[0] & 0x0f0f0f0f;
+        q32[3] = q2[0] & 0xf0f0f0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 4; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
+            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#else
+        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
+        const uint16_t * q2 = q1 + 32;
+
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[0] & 0xf0f0;
+        q16[2] = q2[0] & 0x0f0f;
+        q16[3] = q2[0] & 0xf0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 2; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
+            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#endif
+
+    }
+#else
+    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
+    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
+
+    const int step = tid * K_QUANTS_PER_ITERATION;
+
+    uint16_t aux16[2];
+    const uint8_t * s = (const uint8_t *)aux16;
+
+    float tmp = 0;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+        const uint8_t * q = x[i].qs + step;
+        const float   * y = yy + i*QK_K + step;
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+        const float d = (float)x[i].dm[0];
+        const float m = (float)x[i].dm[1];
+        float sum = 0.f;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
+                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
+                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
+                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
+        }
+        tmp += sum;
+    }
+
+#endif
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
+
+    const int row = blockIdx.x;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q5_K * x = (const block_q5_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+#if QK_K == 256
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid = threadIdx.x/2;  // 0...15
+    const int ix  = threadIdx.x%2;
+
+    const int il  = tid/4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
+    const int n   = 2;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    const uint8_t hm1  = 1 << (2*im);
+    const uint8_t hm2  = hm1 << 4;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+    uint16_t q16[8];
+    const uint8_t * q4 = (const uint8_t *)q16;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2) {
+
+        const uint8_t * ql1 = x[i].qs + q_offset;
+        const uint8_t * qh  = x[i].qh + l0;
+        const float   * y1  = yy + i*QK_K + y_offset;
+        const float   * y2  = y1 + 128;
+
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+        float4 sum = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        const uint16_t * q1 = (const uint16_t *)ql1;
+        const uint16_t * q2 = q1 + 32;
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[8] & 0x0f0f;
+        q16[2] = (q1[0] >> 4) & 0x0f0f;
+        q16[3] = (q1[8] >> 4) & 0x0f0f;
+        q16[4] = q2[0] & 0x0f0f;
+        q16[5] = q2[8] & 0x0f0f;
+        q16[6] = (q2[0] >> 4) & 0x0f0f;
+        q16[7] = (q2[8] >> 4) & 0x0f0f;
+        for (int l = 0; l < n; ++l) {
+            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
+            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
+            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
+            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
+            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
+        }
+        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
+    }
+
+#else
+    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
+    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
+    const int step = tid * K_QUANTS_PER_ITERATION;
+    const int im = step/8;
+    const int in = step%8;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+        const uint8_t * q = x[i].qs + step;
+        const int8_t  * s = x[i].scales;
+        const float   * y = yy + i*QK_K + step;
+        const float     d = x[i].d;
+        float sum = 0.f;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            const uint8_t h = x[i].qh[in+j] >> im;
+            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
+                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
+                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
+                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
+        }
+        tmp += sum;
+    }
+#endif
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q6_K * x = (const block_q6_K *)vx + ib0;
+
+#if QK_K == 256
+
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+#if K_QUANTS_PER_ITERATION == 1
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
+    const int is = 0;
+#else
+    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
+    const int is = in / 4;
+#endif
+    const int ql_offset = 64*im + l0;
+    const int qh_offset = 32*im + l0;
+    const int s_offset  =  8*im + is;
+    const int y_offset = 128*im + l0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + y_offset;
+        const uint8_t * ql = x[i].ql + ql_offset;
+        const uint8_t * qh = x[i].qh + qh_offset;
+        const int8_t  * s  = x[i].scales + s_offset;
+
+        const float d = x[i].d;
+
+#if K_QUANTS_PER_ITERATION == 1
+        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
+        tmp += sum;
+#else
+        float sum = 0;
+        for (int l = 0; l < 4; ++l) {
+            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
+        }
+        tmp += sum;
+#endif
+
+    }
+
+#else
+
+    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...7
+    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0...3
+
+    const int step = tid * K_QUANTS_PER_ITERATION;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + step;
+        const uint8_t * ql = x[i].ql + step;
+        const uint8_t * qh = x[i].qh + step;
+        const int8_t  * s  = x[i].scales;
+
+        const float d = x[i+0].d;
+
+        float sum = 0;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
+                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
+                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
+                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
+        }
+        tmp += sum;
+
+    }
+
+#endif
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const half * x = (const half *) vx;
+
+    // automatic half -> float type cast if dfloat == float
+    v.x = x[ib + iqs + 0];
+    v.y = x[ib + iqs + 1];
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
+static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
+    // qk = quantized weights per x block
+    // qr = number of quantized weights per data value in x block
+    const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int tid = threadIdx.x;
+
+    const int iter_stride = 2*GGML_CUDA_DMMV_X;
+    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+// partial sum for each thread
+#ifdef GGML_CUDA_F16
+    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
+#else
+    float tmp = 0.0f;
+#endif // GGML_CUDA_F16
+
+    for (int i = 0; i < ncols; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int64_t ib = ((int64_t)row*ncols + col)/qk; // x block index
+        const int iqs = (col%qk)/qr; // x quant index
+        const int iybs = col - col%qk; // y block start index
+
+// processing >2 values per i iter is faster for fast GPUs
+#pragma unroll
+        for (int j = 0; j < vals_per_iter; j += 2) {
+            // process 2 vals per j iter
+
+            // dequantize
+            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
+            dfloat2 v;
+            dequantize_kernel(vx, ib, iqs + j/qr, v);
+
+            // matrix multiplication
+            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+#ifdef GGML_CUDA_F16
+            tmp += __hmul2(v, {
+                y[iybs + iqs + j/qr + 0],
+                y[iybs + iqs + j/qr + y_offset]
+            });
+#else
+            tmp += v.x * y[iybs + iqs + j/qr + 0];
+            tmp += v.y * y[iybs + iqs + j/qr + y_offset];
+#endif // GGML_CUDA_F16
+        }
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (tid == 0) {
+#ifdef GGML_CUDA_F16
+        dst[row] = tmp.x + tmp.y;
+#else
+        dst[row] = tmp;
+#endif // GGML_CUDA_F16
+    }
+}
+
+static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const dim3 block_dims(32, 1, 1);
+    dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<1, 1, convert_f16>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+void ggml_cuda_op_dequantize_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+    GGML_UNUSED(ctx);
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
+#ifdef GGML_CUDA_F16
+    ggml_cuda_pool_alloc<half> src1_dfloat_a(ctx.pool());
+    half * src1_dfloat = nullptr; // dfloat == half
+
+    bool src1_convert_f16 =
+        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
+        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
+        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
+
+    if (src1_convert_f16) {
+        src1_dfloat = src1_dfloat_a.alloc(ne00);
+        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        GGML_ASSERT(to_fp16_cuda != nullptr);
+        to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
+    }
+#else
+    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
+#endif // GGML_CUDA_F16
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_F16:
+            convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}

llama/ggml-cuda/dmmv.cuh

@@ -0,0 +1,18 @@
+#include "common.cuh"
+
+// dmmv = dequantize_mul_mat_vec
+
+// TODO: remove this?
+#ifndef GGML_CUDA_DMMV_X
+#define GGML_CUDA_DMMV_X 32
+#endif
+
+#ifndef GGML_CUDA_MMV_Y
+#define GGML_CUDA_MMV_Y 1
+#endif
+
+void ggml_cuda_op_dequantize_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);

llama/ggml-cuda/fattn-common.cuh

@@ -0,0 +1,162 @@
+#include "common.cuh"
+
+#include <cstdint>
+
+#define FATTN_KQ_STRIDE       256
+#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
+#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
+
+typedef void (* fattn_kernel_t)(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3);
+
+template<int D, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_combine_results(
+        const float  * __restrict__ VKQ_parts,
+        const float2 * __restrict__ VKQ_meta,
+        float * __restrict__ dst) {
+    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
+    dst       +=                 D * gridDim.y*blockIdx.x;
+
+    const int tid = threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float2 meta[parallel_blocks];
+    if (tid < 2*parallel_blocks) {
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+    }
+
+    __syncthreads();
+
+    float kqmax = meta[0].x;
+#pragma unroll
+    for (int l = 1; l < parallel_blocks; ++l) {
+        kqmax = max(kqmax, meta[l].x);
+    }
+
+    float VKQ_numerator   = 0.0f;
+    float VKQ_denominator = 0.0f;
+#pragma unroll
+    for (int l = 0; l < parallel_blocks; ++l) {
+        const float diff = meta[l].x - kqmax;
+        const float KQ_max_scale = expf(diff);
+        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
+        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
+
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_denominator += KQ_max_scale * meta[l].y;
+    }
+
+    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+}
+
+template <int D, int parallel_blocks>
+void launch_fattn(ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, int nwarps, int cols_per_block) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->type == GGML_TYPE_F32);
+    GGML_ASSERT(K->type == GGML_TYPE_F16);
+    GGML_ASSERT(V->type == GGML_TYPE_F16);
+    GGML_ASSERT(KQV->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
+    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
+                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
+
+    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t main_stream = ctx.stream();
+
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    const dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    fattn_kernel<<<blocks_num, block_dim, shmem, main_stream>>>(
+        (const char *) Q->data,
+        (const char *) K->data,
+        (const char *) V->data,
+        mask ? ((const char *) mask->data) : nullptr,
+        (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+        scale, max_bias, m0, m1, n_head_log2,
+        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+        mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+        Q->nb[1], Q->nb[2], Q->nb[3],
+        K->nb[1], K->nb[2], K->nb[3],
+        KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+    );
+    CUDA_CHECK(cudaGetLastError());
+
+    if ((parallel_blocks) == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}

llama/ggml-cuda/fattn-tile-f16.cu

@@ -0,0 +1,312 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F16 64
+
+template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ half KQ[ncols*FATTN_KQ_STRIDE_TILE_F16];
+    half2 * KQ2 = (half2 *) KQ;
+
+    __shared__ half2 KV_tmp[FATTN_KQ_STRIDE_TILE_F16][D/2 + 1]; // Pad D to avoid memory bank conflicts.
+
+    half kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -HALF_MAX_HALF;
+    }
+    half2 kqsum[ncols/nwarps] = {{0.0f, 0.0f}};
+
+    half2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ half2 Q_h2[ncols][D/2];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        half kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                KV_tmp[i_KQ][k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+            }
+        }
+
+        __syncthreads();
+
+        half2 sum2[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE][ncols/nwarps] = {{{0.0f, 0.0f}}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D/2; ++k_KQ) {
+            half2 K_k[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE];
+            half2 Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_h2[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE]*Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                half sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F16 + i_KQ] = sum;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]));
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F16/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const half2 diff = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] - __half2half2(kqmax[j0/nwarps]);
+                const half2 val = h2exp(diff);
+                kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + val;
+                KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] = val;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp[k][i] = V_h2[(k_VKQ_0 + k)*stride_KV2 + i];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += 2) {
+            half2  V_k[(D/2)/WARP_SIZE][2];
+            half2 KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE][0] = KV_tmp[k0 + 0][i];
+                V_k[i0/WARP_SIZE][1] = KV_tmp[k0 + 1][i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + k0/2];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][0]* __low2half2(KQ_k[j0/nwarps]);
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][1]*__high2half2(KQ_k[j0/nwarps]);
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val /= __half2half2(kqsum_j);
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}

llama/ggml-cuda/fattn-tile-f16.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/fattn-tile-f32.cu

@@ -0,0 +1,309 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f32.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F32 32
+
+template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ float KQ[ncols*FATTN_KQ_STRIDE_TILE_F32];
+
+    __shared__ float KV_tmp[FATTN_KQ_STRIDE_TILE_F32][D + 1]; // Pad D to avoid memory bank conflicts.
+    float2 * KV_tmp2 = (float2 *) KV_tmp;
+
+    float kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols/nwarps] = {0.0f};
+
+    float2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ float Q_f[ncols][D];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
+            float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x];
+            Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
+            Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
+                const half2 tmp = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
+                KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] =  __low2float(tmp);
+                KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
+            }
+        }
+
+        __syncthreads();
+
+        float sum[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE][ncols/nwarps] = {{0.0f}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D; ++k_KQ) {
+            float K_k[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE];
+            float Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_f[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE] * Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F32 + i_KQ] = sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+            float kqsum_add = 0.0f;
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F32; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const float diff = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] - kqmax[j0/nwarps];
+                const float val = expf(diff);
+                kqsum_add += val;
+                KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] = val;
+            }
+            kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
+                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F32; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp2[k*(D/2) + i].x =  __low2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+                KV_tmp2[k*(D/2) + i].y = __high2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < FATTN_KQ_STRIDE_TILE_F32; ++k) {
+            float2 V_k[(D/2)/WARP_SIZE];
+            float  KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE] = KV_tmp2[k*(D/2) + i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + k];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE].x += V_k[i0/WARP_SIZE].x*KQ_k[j0/nwarps];
+                    VKQ[j0/nwarps][i0/WARP_SIZE].y += V_k[i0/WARP_SIZE].y*KQ_k[j0/nwarps];
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        float kqsum_j = kqsum[j_VKQ_0/nwarps];
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val.x /= kqsum_j;
+                dst_val.y /= kqsum_j;
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}

llama/ggml-cuda/fattn-tile-f32.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/fattn-vec-f16.cu

@@ -0,0 +1,326 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ half KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -HALF_MAX_HALF;
+    }
+    half2 * KQ2 = (half2 *) KQ;
+
+    half kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -HALF_MAX_HALF;
+    }
+    half kqsum[ncols] = {0.0f};
+
+    __shared__ half kqmax_shared[ncols][WARP_SIZE];
+    __shared__ half kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    half2 VKQ[ncols] = {{0.0f, 0.0f}};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
+        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
+        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
+        half kqmax_new = kqmax[0];
+        half kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            half2 sum2[ncols] = {{0.0f, 0.0f}};
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum2[j] = warp_reduce_sum(sum2[j]);
+                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
+                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                if (ncols == 1) {
+                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
+                } else {
+                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
+                }
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < D; k0 += 2) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
+                break;
+            }
+
+            half2 V_k;
+            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
+            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid < ncols) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * KQV = dst;
+    ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    constexpr int cols_per_block  = 1;
+    constexpr int parallel_blocks = 4;
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 256: {
+            constexpr int      D = 256;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_vec_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] == 1) {
+        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block  = 2;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block  = 4;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block  = 8;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block  = 8;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}

llama/ggml-cuda/fattn-vec-f16.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/fattn-vec-f32.cu

@@ -0,0 +1,275 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f32.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -FLT_MAX/2.0f;
+    }
+
+    float kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols] = {0.0f};
+
+    __shared__ float kqmax_shared[ncols][WARP_SIZE];
+    __shared__ float kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    float2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            Q_h2[j][i0/WARP_SIZE]    = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE].x *= scale;
+            Q_h2[j][i0/WARP_SIZE].y *= scale;
+        }
+    }
+
+    float VKQ[ncols] = {0.0f};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            float sum[ncols] = {0.0f};
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum[j] +=  __low2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].x;
+                    sum[j] += __high2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].y;
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum[j] = warp_reduce_sum(sum[j]);
+                sum[j] += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum[j]);
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum[j];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const float val = expf(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= KQ_max_scale;
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < D; ++k) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
+                break;
+            }
+
+            const float V_ki = __half2float(V_h[(k_VKQ_0 + k)*stride_KV + tid]);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_ki*KQ[j*D + k];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        float dst_val = VKQ[j_VKQ];
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid < ncols) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    }
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_vec_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block  = 1;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block  = 2;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block  = 4;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block  = 8;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block  = 8;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}

llama/ggml-cuda/fattn-vec-f32.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/fattn.cu

@@ -0,0 +1,638 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+#include "fattn-tile-f32.cuh"
+#include "fattn-vec-f16.cuh"
+#include "fattn-vec-f32.cuh"
+#include "fattn.cuh"
+
+#include <cstdint>
+
+#if FP16_MMA_AVAILABLE
+#include <mma.h>
+#endif
+
+// D == head size, VKQ_stride == num VKQ rows calculated in parallel:
+template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_MMA_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.
+    const int ip  =        blockIdx.x % parallel_blocks;  // Index in group of blocks running for the same column in parallel.
+
+    static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE.");
+    static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16.");
+    constexpr int frag_m = ncols == 8 ? 32 : 16;
+    constexpr int frag_n = ncols == 8 ?  8 : 16;
+    static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0.");
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_a,    frag_m, frag_n, 16, half, nvcuda::wmma::row_major> frag_a_K;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_a,    frag_m, frag_n, 16, half, nvcuda::wmma::col_major> frag_a_V;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_b,    frag_m, frag_n, 16, half, nvcuda::wmma::col_major> frag_b;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::accumulator, frag_m, frag_n, 16, half>                          frag_c_VKQ;
+
+    constexpr int KQ_stride_tc  = nwarps*frag_m; // Number of KQ rows calculated in parallel.
+    constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy.
+    static_assert(VKQ_ratio <= nwarps, "VKQ_ratio must be <= nwarps.");
+
+    // Pad internal representation of KQ, KQV to reduce shared memory bank conflicts:
+    constexpr int D_padded = D + 8;
+    constexpr int kqs_padded = FATTN_KQ_STRIDE + 8;
+    constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float * Q_f   = (const float *) (Q + nb02* blockIdx.y              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.y / gqa_ratio));
+    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
+    const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
+
+    const int stride_Q  = nb01 / sizeof(float);
+    const int stride_KV = nb11 / sizeof(half);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+    const half2 slope2 = make_half2(slopef, slopef);
+
+    frag_b Q_b[D/16][ncols/frag_n];
+
+    // A single buffer for temporarily holding tiles of KQ and VKQ parts:
+    constexpr int mem_KQ = ncols*kqs_padded*kqar;
+    constexpr int mem_VKQ_parts = VKQ_ratio*ncols*D_padded;
+    __shared__ half KQ[mem_KQ >= mem_VKQ_parts ? mem_KQ : mem_VKQ_parts];
+    float * KQ_f = (float *) KQ;
+    half2 * KQ2 = (half2 *) KQ;
+
+    float    KQ_rowsum_f[ncols/nwarps] = {0.0f};
+    float       KQ_max_f[ncols/nwarps];
+    float KQ_max_scale_f[ncols/nwarps] = {0.0f};
+
+#pragma unroll
+    for (int j = 0; j < ncols/nwarps; ++j) {
+        KQ_max_f[j] = -FLT_MAX/2.0f;
+    }
+
+    half2    KQ_rowsum_h2[ncols/nwarps] = {{0.0f, 0.0f}};
+    half2       KQ_max_h2[ncols/nwarps];
+    half2 KQ_max_scale_h2[ncols/nwarps] = {{0.0f, 0.0f}};
+
+#pragma unroll
+    for (int j = 0; j < ncols/nwarps; ++j) {
+        KQ_max_h2[j] = make_half2(-HALF_MAX_HALF, -HALF_MAX_HALF);
+    }
+
+    __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice.
+    half2 * VKQ2 = (half2 *) VKQ;
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D/2 && i >= D/2) {
+                break;
+            }
+            VKQ2[j*(D_padded/2) + i] = make_half2(0.0f, 0.0f);
+        }
+    }
+
+    // Convert Q to half and apply scale, temporarily store in KQ:
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D && i >= D) {
+                break;
+            }
+            KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f;
+        }
+    }
+
+    __syncthreads();
+
+    // Load Q into tensor core fragments/registers since it will be used frequently:
+#pragma unroll
+    for (int i0 = 0; i0 < D; i0 += 16) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+            nvcuda::wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded);
+        }
+    }
+
+    __syncthreads();
+
+    // Iterate over ne11 == previous tokens:
+    for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) {
+        // Calculate tile of KQ:
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
+            frag_c_KQ KQ_c[ncols/frag_n];
+#pragma unroll
+            for (int j = 0; j < ncols/frag_n; ++j) {
+                nvcuda::wmma::fill_fragment(KQ_c[j], 0.0f);
+            }
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) {
+                frag_a_K K_a;
+                nvcuda::wmma::load_matrix_sync(K_a, K_h + (k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
+#pragma unroll
+                for (int j = 0; j < ncols/frag_n; ++j) {
+                    nvcuda::wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]);
+                }
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+                nvcuda::wmma::store_matrix_sync((KQ_acc_t *) KQ + j0*kqs_padded + i_KQ_0 + frag_m*threadIdx.y, KQ_c[j0/frag_n], kqs_padded, nvcuda::wmma::mem_col_major);
+            }
+        }
+
+        __syncthreads();
+
+        // Calculate softmax for each KQ column using the current max. value.
+        // The divisor is stored in KQ_rowsum and will be applied at the end.
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (std::is_same<KQ_acc_t, float>::value) {
+                float KQ_f_tmp[FATTN_KQ_STRIDE / WARP_SIZE];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ_f_tmp[k0/WARP_SIZE] = KQ_f[j*kqs_padded + k];
+                }
+
+                float KQ_max_new = KQ_max_f[j0/nwarps];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ_f_tmp[k0/WARP_SIZE] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
+                    KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/WARP_SIZE]);
+                }
+                KQ_max_new = warp_reduce_max(KQ_max_new);
+
+                const float diff = KQ_max_f[j0/nwarps] - KQ_max_new;
+                KQ_max_scale_f[j0/nwarps] = expf(diff);
+                if (diff <= SOFTMAX_FTZ_THRESHOLD) {
+                    KQ_max_scale_f[j0/nwarps] = 0.0f;
+                }
+                KQ_max_f[j0/nwarps] = KQ_max_new;
+
+                float KQ_rowsum_add = 0.0f;
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    const float diff = KQ_f_tmp[k0/WARP_SIZE] - KQ_max_f[j0/nwarps];
+                    KQ_f_tmp[k0/WARP_SIZE] = expf(diff);
+                    if (diff <= SOFTMAX_FTZ_THRESHOLD) {
+                        KQ_f_tmp[k0/WARP_SIZE] = 0.0f;
+                    }
+                    KQ_rowsum_add += KQ_f_tmp[k0/WARP_SIZE];
+                    KQ[j*(kqar*kqs_padded) + k] = KQ_f_tmp[k0/WARP_SIZE];
+                }
+                KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add);
+
+                // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
+                KQ_rowsum_f[j0/nwarps] = KQ_max_scale_f[j0/nwarps]*KQ_rowsum_f[j0/nwarps] + KQ_rowsum_add;
+            } else {
+                half2 KQ2_tmp[FATTN_KQ_STRIDE/(2*WARP_SIZE)];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ2_tmp[k0/WARP_SIZE] = KQ2[j*(kqs_padded/2) + k];
+                }
+
+                half2 KQ_max_new = KQ_max_h2[j0/nwarps];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ2_tmp[k0/WARP_SIZE] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
+                    KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/WARP_SIZE]);
+                }
+                KQ_max_new = __half2half2(warp_reduce_max(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
+                const half2 diff = KQ_max_h2[j0/nwarps] - KQ_max_new;
+                KQ_max_scale_h2[j0/nwarps] = h2exp(diff);
+                const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
+                *((uint32_t *) &KQ_max_scale_h2[j0/nwarps]) &= ftz_mask;
+                KQ_max_h2[j0/nwarps] = KQ_max_new;
+
+                half2 KQ_rowsum_add = make_half2(0.0f, 0.0f);
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    const half2 diff = KQ2_tmp[k0/WARP_SIZE] - KQ_max_h2[j0/nwarps];
+                    KQ2_tmp[k0/WARP_SIZE] = h2exp(diff);
+                    const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
+                    *((uint32_t *) &KQ2_tmp[k0/WARP_SIZE]) &= ftz_mask;
+                    KQ_rowsum_add += KQ2_tmp[k0/WARP_SIZE];
+                    KQ2[j*(kqs_padded/2) + k] = KQ2_tmp[k0/WARP_SIZE];
+                }
+                KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add);
+
+                // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
+                KQ_rowsum_h2[j0/nwarps] = KQ_max_scale_h2[j0/nwarps]*KQ_rowsum_h2[j0/nwarps] + KQ_rowsum_add;
+            }
+        }
+
+        __syncthreads();
+
+        frag_b KQ_b[FATTN_KQ_STRIDE/(VKQ_ratio*16)][ncols/frag_n];
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+#pragma unroll
+            for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) {
+                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
+                nvcuda::wmma::load_matrix_sync(
+                    KQ_b[k0/(VKQ_ratio*16)][j0/frag_n],
+                    KQ + j0*(kqar*kqs_padded) + k,
+                    kqar*kqs_padded);
+            }
+        }
+
+        frag_c_VKQ VKQ_c[D/VKQ_stride][ncols/frag_n];
+#pragma unroll
+        for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += VKQ_stride) {
+#pragma unroll
+            for (int j = 0; j < ncols/frag_n; ++j) {
+                nvcuda::wmma::fill_fragment(VKQ_c[i_VKQ_0/VKQ_stride][j], 0.0f);
+            }
+
+#pragma unroll
+            for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) {
+                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
+
+                frag_a_V v_a;
+                nvcuda::wmma::load_matrix_sync(v_a, V_h + (k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
+#pragma unroll
+                for (int j = 0; j < ncols/frag_n; ++j) {
+                    nvcuda::wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]);
+                }
+            }
+        }
+
+        __syncthreads();
+
+        const int offset_k = (threadIdx.y % VKQ_ratio) * (ncols*D_padded);
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += VKQ_stride) {
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+                nvcuda::wmma::store_matrix_sync(
+                    KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
+                    VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n],
+                    D_padded, nvcuda::wmma::mem_col_major);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            half2 VKQ_scale;
+            if (std::is_same<KQ_acc_t, float>::value) {
+                VKQ_scale = make_half2(KQ_max_scale_f[j0/nwarps], KQ_max_scale_f[j0/nwarps]);
+            } else {
+                VKQ_scale = KQ_max_scale_h2[j0/nwarps];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+                if (i0 + WARP_SIZE > D/2 && i >= D/2) {
+                    break;
+                }
+
+                half2 VKQ_add = make_half2(0.0f, 0.0f);
+#pragma unroll
+                for (int l = 0; l < VKQ_ratio; ++l) {
+                    VKQ_add += KQ2[l*(ncols*D_padded/2) + j*(D_padded/2) + i];
+                }
+                VKQ2[j*(D_padded/2) + i] = VKQ_scale*VKQ2[j*(D_padded/2) + i] + VKQ_add;
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j_VKQ = j0 + threadIdx.y;
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+
+        float KQ_rowsum_j;
+        if (std::is_same<KQ_acc_t, float>::value) {
+            KQ_rowsum_j = KQ_rowsum_f[j0/nwarps];
+        } else {
+            KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]);
+        }
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D && i >= D) {
+                break;
+            }
+            float dst_val = VKQ[j_VKQ*D_padded + i];
+            if (parallel_blocks == 1) {
+                dst_val /= KQ_rowsum_j;
+            }
+            dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val;
+        }
+
+        if (parallel_blocks == 1 || threadIdx.x != 0) {
+            continue;
+        }
+
+        float2 dst_meta_val;
+        if (std::is_same<KQ_acc_t, float>::value) {
+            dst_meta_val.x = KQ_max_f[j0/nwarps];
+        } else {
+            dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
+        }
+        dst_meta_val.y = KQ_rowsum_j;
+        dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val;
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_MMA_AVAILABLE
+}
+
+constexpr int get_max_power_of_2(int x) {
+    return x % 2 == 0 ? 2*get_max_power_of_2(x/2) : 1;
+}
+
+static_assert(get_max_power_of_2(1) == 1, "Test failed.");
+static_assert(get_max_power_of_2(2) == 2, "Test failed.");
+static_assert(get_max_power_of_2(4) == 4, "Test failed.");
+static_assert(get_max_power_of_2(6) == 2, "Test failed.");
+
+// Number of VKQ rows calculated in parallel:
+constexpr int get_VKQ_stride(int D, int nwarps, int frag_m) {
+    return (get_max_power_of_2(D/frag_m) < nwarps ? get_max_power_of_2(D/frag_m) : nwarps)*frag_m;
+}
+
+static_assert(get_VKQ_stride(128, 1, 32) ==  32, "Test failed.");
+static_assert(get_VKQ_stride(128, 2, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride(128, 4, 32) == 128, "Test failed.");
+static_assert(get_VKQ_stride( 64, 1, 32) ==  32, "Test failed.");
+static_assert(get_VKQ_stride( 64, 2, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride( 64, 4, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride( 80, 1, 16) ==  16, "Test failed.");
+static_assert(get_VKQ_stride( 80, 2, 16) ==  16, "Test failed.");
+static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
+
+template <int D, int cols_per_block, int nwarps, typename KQ_acc_t>
+void launch_fattn_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+
+    constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
+    const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
+    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
+
+    if (4*blocks_num_pb1 < 2*nsm) {
+        constexpr int parallel_blocks = 4;
+        fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        return;
+    }
+    if (2*blocks_num_pb1 < 2*nsm) {
+        constexpr int parallel_blocks = 2;
+        fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        return;
+    }
+    constexpr int parallel_blocks = 1;
+    fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+}
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    ggml_cuda_set_device(ctx.device);
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int32_t precision = KQV->op_params[2];
+
+    // On AMD the tile kernels perform poorly, use the vec kernel instead:
+    if (cc >= CC_OFFSET_AMD) {
+        if (precision == GGML_PREC_DEFAULT) {
+            ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fast_fp16_available(cc)) {
+        if (Q->ne[1] <= 8) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fp16_mma_available(cc)) {
+        if (Q->ne[1] <= 8) {
+            ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+        }
+        return;
+    }
+
+    if (precision != GGML_PREC_DEFAULT) {
+        if (Q->ne[1] == 1 && (Q->ne[0] == 64 || Q->ne[0] == 128)) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            return;
+        }
+
+        if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
+            constexpr int cols_per_block = 16;
+            constexpr int nwarps         =  4;
+            switch (Q->ne[0]) {
+                case 64:
+                    launch_fattn_f16< 64, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 80:
+                    launch_fattn_f16< 80, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 96:
+                    launch_fattn_f16< 96, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 112:
+                    launch_fattn_f16<112, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 128:
+                    launch_fattn_f16<128, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 256:
+                    launch_fattn_f16<256, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                default:
+                    GGML_ASSERT(false);
+                    break;
+            }
+        } else {
+            constexpr int cols_per_block = 32;
+            constexpr int nwarps         =  4;
+            switch (Q->ne[0]) {
+                case 64:
+                    launch_fattn_f16< 64, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 80:
+                    launch_fattn_f16< 80, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 96:
+                    launch_fattn_f16< 96, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 112:
+                    launch_fattn_f16<112, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                case 128:
+                    launch_fattn_f16<128, cols_per_block, nwarps, float>(ctx, dst);
+                    break;
+                // case 256:
+                //     launch_fattn_f16<256, cols_per_block, nwarps, float>(ctx, dst);
+                //     break;
+                default:
+                    GGML_ASSERT(false);
+                    break;
+            }
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
+        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 8 && Q->ne[0] % WARP_SIZE == 0) {
+        constexpr int cols_per_block = 8;
+        constexpr int nwarps         = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 96:
+                launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 128:
+                launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 256:
+                launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 16;
+        constexpr int nwarps         =  4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 80:
+                launch_fattn_f16< 80, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 96:
+                launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 112:
+                launch_fattn_f16<112, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 128:
+                launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            case 256:
+                launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int nwarps         =  4;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
+            break;
+        case 80:
+            launch_fattn_f16< 80, cols_per_block, nwarps, half>(ctx, dst);
+            break;
+        case 96:
+            launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
+            break;
+        case 112:
+            launch_fattn_f16<112, cols_per_block, nwarps, half>(ctx, dst);
+            break;
+        case 128:
+            launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
+            break;
+        case 256:
+            launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+    return;
+}

llama/ggml-cuda/fattn.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/getrows.cu

@@ -0,0 +1,178 @@
+#include "getrows.cuh"
+#include "dequantize.cuh"
+
+template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void k_get_rows(
+            const void * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
+
+    const int ib = i00/qk; // block index
+    const int iqs = (i00%qk)/qr; // quant index
+    const int iybs = i00 - i00%qk; // dst block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(src0_row, ib, iqs, v);
+
+    dst_row[iybs + iqs + 0]        = v.x;
+    dst_row[iybs + iqs + y_offset] = v.y;
+}
+
+template<typename src0_t, typename dst_t>
+static __global__ void k_get_rows_float(
+            const src0_t * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
+    dst_row[i00] = src0_row[i00];
+}
+
+template<int qk, int qr, dequantize_kernel_t dq>
+static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                            const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    GGML_UNUSED(dst);
+}
+
+template<typename src0_t>
+static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                                const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    GGML_UNUSED(dst);
+}
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
+
+    const int32_t * src1_i32 = (const int32_t *) src1_d;
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        default:
+            // TODO: k-quants
+            fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
+            GGML_ASSERT(false);
+            break;
+    }
+}

llama/ggml-cuda/getrows.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_GET_ROWS_BLOCK_SIZE 256
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/im2col.cu

@@ -0,0 +1,104 @@
+#include "im2col.cuh"
+
+template <typename T>
+static  __global__ void im2col_kernel(
+        const float * x, T * dst, int64_t batch_offset,
+        int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
+        int s0, int s1, int p0, int p1, int d0, int d1) {
+    const int64_t i = threadIdx.x + blockIdx.x * blockDim.x;
+    if (i >= pelements) {
+        return;
+    }
+
+    const int64_t  ksize = OW * (KH > 1 ? KW : 1);
+    const int64_t  kx = i / ksize;
+    const int64_t  kd = kx * ksize;
+    const int64_t  ky = (i - kd) / OW;
+    const int64_t  ix = i % OW;
+
+    const int64_t  oh = blockIdx.y;
+    const int64_t  batch = blockIdx.z / IC;
+    const int64_t  ic = blockIdx.z % IC;
+
+    const int64_t iiw = ix * s0 + kx * d0 - p0;
+    const int64_t iih = oh * s1 + ky * d1 - p1;
+
+    const int64_t offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = ic * offset_delta + batch * batch_offset;
+        dst[offset_dst] = x[offset_src + iih * IW + iiw];
+    }
+}
+
+template <typename T>
+static void im2col_cuda(const float * x, T* dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+    const int parallel_elements = OW * KW * KH;
+    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
+    dim3 block_nums(num_blocks, OH, batch * IC);
+    im2col_kernel<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
+}
+
+static void im2col_cuda_f16(const float * x, half * dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+
+    im2col_cuda<half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+}
+
+static void im2col_cuda_f32(const float * x, float * dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+
+    im2col_cuda<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+}
+
+void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const int64_t IC = src1->ne[is_2D ? 2 : 1];
+    const int64_t IH = is_2D ? src1->ne[1] : 1;
+    const int64_t IW =         src1->ne[0];
+
+    const int64_t KH = is_2D ? src0->ne[1] : 1;
+    const int64_t KW =         src0->ne[0];
+
+    const int64_t OH = is_2D ? dst->ne[2] : 1;
+    const int64_t OW =         dst->ne[1];
+
+    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t batch = src1->ne[3];
+    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+
+    if(dst->type == GGML_TYPE_F16) {
+        im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+    } else {
+        im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+    }
+}

llama/ggml-cuda/im2col.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_IM2COL_BLOCK_SIZE 256
+
+void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/mmq.cu

@@ -0,0 +1,2255 @@
+#include "mmq.cuh"
+#include "vecdotq.cuh"
+
+typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
+typedef void (*load_tiles_cuda_t)(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row);
+typedef float (*vec_dot_q_mul_mat_cuda_t)(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ms, const int & i, const int & j, const int & k);
+typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v);
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh);
+    GGML_UNUSED(x_sc);
+
+    __shared__ int  tile_x_qs[mmq_y * (WARP_SIZE)       + mmq_y];
+    __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI4_0) + mmq_y/QI4_0];
+
+    *x_ql = tile_x_qs;
+    *x_dm = (half2 *) tile_x_d;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI4_0;
+    const int kqsx = k % QI4_0;
+
+    const block_q4_0 * bx0 = (const block_q4_0 *) vx;
+
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
+        // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) {
+        int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+    const float * x_dmf = (const float *) x_dm;
+
+    int u[2*VDR_Q4_0_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE];
+    }
+
+    return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
+        (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0],
+         y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    __shared__ int   tile_x_qs[mmq_y * (WARP_SIZE) +     + mmq_y];
+    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_1) + mmq_y/QI4_1];
+
+    *x_ql = tile_x_qs;
+    *x_dm = tile_x_dm;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI4_1;
+    const int kqsx = k % QI4_1;
+
+    const block_q4_1 * bx0 = (const block_q4_1 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
+        int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+
+    int u[2*VDR_Q4_1_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE];
+    }
+
+    return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
+        (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1],
+         y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    __shared__ int  tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
+    __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI5_0) + mmq_y/QI5_0];
+
+    *x_ql = tile_x_ql;
+    *x_dm = (half2 *) tile_x_d;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI5_0;
+    const int kqsx = k % QI5_0;
+
+    const block_q5_0 * bx0 = (const block_q5_0 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        const int ql = get_int_from_uint8(bxi->qs, kqsx);
+        const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0));
+
+        int qs0 = (ql >>  0)   & 0x0F0F0F0F;
+        qs0    |= (qh <<  4)   & 0x00000010;  // 0 ->  4
+        qs0    |= (qh << 11)   & 0x00001000;  // 1 -> 12
+        qs0    |= (qh << 18)   & 0x00100000;  // 2 -> 20
+        qs0    |= (qh << 25)   & 0x10000000;  // 3 -> 28
+        qs0     = __vsubss4(qs0, 0x10101010); // subtract 16
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
+
+        int qs1 = (ql >>  4)   & 0x0F0F0F0F;
+        qs1    |= (qh >> 12)   & 0x00000010;  // 16 ->  4
+        qs1    |= (qh >>  5)   & 0x00001000;  // 17 -> 12
+        qs1    |= (qh <<  2)   & 0x00100000;  // 18 -> 20
+        qs1    |= (qh <<  9)   & 0x10000000;  // 19 -> 28
+        qs1     = __vsubss4(qs1, 0x10101010); // subtract 16
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
+    const int kbxd = k % blocks_per_tile_x_row;
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) {
+        int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q5_0_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+    const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0;
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    int u[2*VDR_Q5_0_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE];
+    }
+
+    return vec_dot_q8_0_q8_1_impl<QR5_0*VDR_Q5_0_Q8_1_MMQ>
+        (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    __shared__ int   tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
+    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_1) + mmq_y/QI5_1];
+
+    *x_ql = tile_x_ql;
+    *x_dm = tile_x_dm;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset < nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI5_1;
+    const int kqsx = k % QI5_1;
+
+    const block_q5_1 * bx0 = (const block_q5_1 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
+        const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1));
+
+        int qs0 = (ql >>  0) & 0x0F0F0F0F;
+        qs0    |= (qh <<  4) & 0x00000010; // 0 ->  4
+        qs0    |= (qh << 11) & 0x00001000; // 1 -> 12
+        qs0    |= (qh << 18) & 0x00100000; // 2 -> 20
+        qs0    |= (qh << 25) & 0x10000000; // 3 -> 28
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
+
+        int qs1 = (ql >>  4) & 0x0F0F0F0F;
+        qs1    |= (qh >> 12) & 0x00000010; // 16 ->  4
+        qs1    |= (qh >>  5) & 0x00001000; // 17 -> 12
+        qs1    |= (qh <<  2) & 0x00100000; // 18 -> 20
+        qs1    |= (qh <<  9) & 0x10000000; // 19 -> 28
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) {
+        int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q5_1_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+    const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1;
+
+    int u[2*VDR_Q5_1_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE];
+    }
+
+    return vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
+        (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    __shared__ int  tile_x_qs[mmq_y * (WARP_SIZE)       + mmq_y];
+    __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI8_0) + mmq_y/QI8_0];
+
+    *x_ql = tile_x_qs;
+    *x_dm = (half2 *) tile_x_d;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI8_0;
+    const int kqsx = k % QI8_0;
+    float * x_dmf = (float *) x_dm;
+
+    const block_q8_0 * bx0 = (const block_q8_0 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI8_0;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) {
+        int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q8_0_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
+
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMQ>
+        (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0],
+         y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q2_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh);
+
+    __shared__ int   tile_x_ql[mmq_y * (WARP_SIZE)       + mmq_y];
+    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI2_K) + mmq_y/QI2_K];
+    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/4)     + mmq_y/4];
+
+    *x_ql = tile_x_ql;
+    *x_dm = tile_x_dm;
+    *x_sc = tile_x_sc;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI2_K;
+    const int kqsx = k % QI2_K;
+
+    const block_q2_K * bx0 = (const block_q2_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI2_K;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) {
+        int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
+        int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4);
+
+        x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4));
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q2_K_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh);
+
+    const int kbx = k / QI2_K;
+    const int ky  = (k % QI2_K) * QR2_K;
+    const float * y_df = (const float *) y_ds;
+
+    int v[QR2_K*VDR_Q2_K_Q8_1_MMQ];
+
+    const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2);
+    const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2));
+
+#pragma unroll
+    for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) {
+        v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303;
+    }
+
+    const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4;
+
+    const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE;
+    return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q3_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+
+    __shared__ int   tile_x_ql[mmq_y * (WARP_SIZE)       + mmq_y];
+    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI3_K) + mmq_y/QI3_K];
+    __shared__ int   tile_x_qh[mmq_y * (WARP_SIZE/2)     + mmq_y/2];
+    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/4)     + mmq_y/4];
+
+    *x_ql = tile_x_ql;
+    *x_dm = tile_x_dm;
+    *x_qh = tile_x_qh;
+    *x_sc = tile_x_sc;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI3_K;
+    const int kqsx = k % QI3_K;
+
+    const block_q3_K * bx0 = (const block_q3_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI3_K;
+    const int kbxd = k % blocks_per_tile_x_row;
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) {
+        int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) {
+        int i = i0 + i_offset * 2 + k / (WARP_SIZE/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2);
+
+        // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
+        x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2));
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
+        int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4);
+
+        const int ksc = k % (QI3_K/4);
+
+        const int ksc_low = ksc % (QI3_K/8);
+        const int shift_low = 4 * (ksc / (QI3_K/8));
+        const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
+
+        const int ksc_high = QI3_K/8;
+        const int shift_high = 2 * ksc;
+        const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
+
+        const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
+
+        x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q3_K_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+
+    const int kbx  = k / QI3_K;
+    const int ky  = (k % QI3_K) * QR3_K;
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4;
+
+    int v[QR3_K*VDR_Q3_K_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) {
+        const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2);
+        const int shift = 2 * ((ky % 32) / 8);
+        const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303;
+
+        const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8);
+        const int vlh = (vh << 2) & 0x04040404;
+
+        v[l] = __vsubss4(vll, vlh);
+    }
+
+    const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE;
+    return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh);
+
+    __shared__ int   tile_x_ql[mmq_y * (WARP_SIZE)       + mmq_y];
+    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_K) + mmq_y/QI4_K];
+    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/8)     + mmq_y/8];
+
+    *x_ql = tile_x_ql;
+    *x_dm = tile_x_dm;
+    *x_sc = tile_x_sc;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI4_K; // == 0 if QK_K == 256
+    const int kqsx = k % QI4_K; // == k if QK_K == 256
+
+    const block_q4_K * bx0 = (const block_q4_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256
+    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) {
+        int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+#if QK_K == 256
+        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
+#else
+        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
+#endif
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8);
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = k % (WARP_SIZE/8);
+
+        // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
+        int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
+        scales8    |= (scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030; // upper 2 bits
+
+        x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh);
+
+    const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8);
+
+    const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE;
+    return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8,
+                                      x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh);
+
+    __shared__ int   tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
+    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_K) + mmq_y/QI5_K];
+    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/8)     + mmq_y/8];
+
+    *x_ql = tile_x_ql;
+    *x_dm = tile_x_dm;
+    *x_sc = tile_x_sc;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI5_K; // == 0 if QK_K == 256
+    const int kqsx = k % QI5_K; // == k if QK_K == 256
+
+    const block_q5_K * bx0 = (const block_q5_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx;
+        const int ky = QR5_K*kqsx;
+
+        const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4));
+        const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010;
+        const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010;
+
+        const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0;
+        const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4);
+
+        x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0;
+        x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256
+    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) {
+        int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+#if QK_K == 256
+        x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
+#endif
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8);
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = k % (WARP_SIZE/8);
+
+        // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
+        int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
+        scales8    |= (scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030; // upper 2 bits
+
+        x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh);
+
+    const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8);
+
+    const int index_x = i * (QR5_K*WARP_SIZE + 1) +  QR5_K*k;
+    const int index_y = j * WARP_SIZE             + (QR5_K*k) % WARP_SIZE;
+    return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8,
+                                      x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]);
+}
+
+template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q6_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
+    GGML_UNUSED(x_qh);
+
+    __shared__ int   tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
+    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI6_K) + mmq_y/QI6_K];
+    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/8)     + mmq_y/8];
+
+    *x_ql = tile_x_ql;
+    *x_dm = tile_x_dm;
+    *x_sc = tile_x_sc;
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
+    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
+    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
+    GGML_UNUSED(x_qh);
+
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI6_K; // == 0 if QK_K == 256
+    const int kqsx = k % QI6_K; // == k if QK_K == 256
+
+    const block_q6_K * bx0 = (const block_q6_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx;
+        const int ky = QR6_K*kqsx;
+
+        const int ql = get_int_from_uint8(bxi->ql, kqsx);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4));
+        const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030;
+        const int qh1 =  (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4))))       & 0x30303030;
+
+        const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0;
+        const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2);
+
+        x_ql[i * (2*WARP_SIZE + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
+        x_ql[i * (2*WARP_SIZE + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256
+    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) {
+        int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4;
+
+        x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8));
+    }
+}
+
+static __device__ __forceinline__ float vec_dot_q6_K_q8_1_mul_mat(
+    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
+    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
+    GGML_UNUSED(x_qh);
+
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]);
+
+    const int index_x = i * (QR6_K*WARP_SIZE + 1) +  QR6_K*k;
+    const int index_y = j * WARP_SIZE             + (QR6_K*k) % WARP_SIZE;
+    return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]);
+}
+
+#define  MMQ_X_Q4_0_RDNA2  64
+#define  MMQ_Y_Q4_0_RDNA2  128
+#define NWARPS_Q4_0_RDNA2  8
+#define  MMQ_X_Q4_0_RDNA1  64
+#define  MMQ_Y_Q4_0_RDNA1  64
+#define NWARPS_Q4_0_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q4_0_AMPERE 4
+#define  MMQ_Y_Q4_0_AMPERE 32
+#define NWARPS_Q4_0_AMPERE 4
+#else
+#define  MMQ_X_Q4_0_AMPERE 64
+#define  MMQ_Y_Q4_0_AMPERE 128
+#define NWARPS_Q4_0_AMPERE 4
+#endif
+#define  MMQ_X_Q4_0_PASCAL 64
+#define  MMQ_Y_Q4_0_PASCAL 64
+#define NWARPS_Q4_0_PASCAL 8
+
+template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
+              allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
+static __device__ __forceinline__ void mul_mat_q(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    const int blocks_per_row_x = ncols_x / qk;
+    const int blocks_per_col_y = nrows_y / QK8_1;
+    const int blocks_per_warp = WARP_SIZE / qi;
+
+    const int & ncols_dst = ncols_y;
+
+    const int row_dst_0 = blockIdx.x*mmq_y;
+    const int & row_x_0 = row_dst_0;
+
+    const int col_dst_0 = blockIdx.y*mmq_x;
+    const int & col_y_0 = col_dst_0;
+
+    int   * tile_x_ql = nullptr;
+    half2 * tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+    allocate_tiles(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc);
+
+    __shared__ int    tile_y_qs[mmq_x * WARP_SIZE];
+    __shared__ half2  tile_y_ds[mmq_x * WARP_SIZE/QI8_1];
+
+    float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}};
+
+    for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) {
+
+        load_tiles(x + row_x_0*blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc,
+                   threadIdx.y, nrows_x-row_x_0-1, threadIdx.x, blocks_per_row_x);
+
+#pragma unroll
+        for (int ir = 0; ir < qr; ++ir) {
+            const int kqs = ir*WARP_SIZE + threadIdx.x;
+            const int kbxd = kqs / QI8_1;
+
+#pragma unroll
+            for (int i = 0; i < mmq_x; i += nwarps) {
+                const int col_y_eff = min(col_y_0 + threadIdx.y + i, ncols_y-1); // to prevent out-of-bounds memory accesses
+
+                const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd];
+
+                const int index_y = (threadIdx.y + i) * WARP_SIZE + kqs % WARP_SIZE;
+                tile_y_qs[index_y] = get_int_from_int8_aligned(by0->qs, threadIdx.x % QI8_1);
+            }
+
+#pragma unroll
+            for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) {
+                const int ids = (ids0 + threadIdx.y * QI8_1 + threadIdx.x / (WARP_SIZE/QI8_1)) % mmq_x;
+                const int kby = threadIdx.x % (WARP_SIZE/QI8_1);
+                const int col_y_eff = min(col_y_0 + ids, ncols_y-1);
+
+                // if the sum is not needed it's faster to transform the scale to f32 ahead of time
+                const half2 * dsi_src = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + ir*(WARP_SIZE/QI8_1) + kby].ds;
+                half2       * dsi_dst = &tile_y_ds[ids * (WARP_SIZE/QI8_1) + kby];
+                if (need_sum) {
+                    *dsi_dst = *dsi_src;
+                } else {
+                    float * dfi_dst = (float *) dsi_dst;
+                    *dfi_dst = __low2float(*dsi_src);
+                }
+            }
+
+            __syncthreads();
+
+// #pragma unroll // unrolling this loop causes too much register pressure
+            for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) {
+#pragma unroll
+                for (int j = 0; j < mmq_x; j += nwarps) {
+#pragma unroll
+                    for (int i = 0; i < mmq_y; i += WARP_SIZE) {
+                        sum[i/WARP_SIZE][j/nwarps] += vec_dot(
+                            tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, tile_y_qs, tile_y_ds,
+                            threadIdx.x + i, threadIdx.y + j, k);
+                    }
+                }
+            }
+
+            __syncthreads();
+        }
+    }
+
+#pragma unroll
+    for (int j = 0; j < mmq_x; j += nwarps) {
+        const int col_dst = col_dst_0 + j + threadIdx.y;
+
+        if (col_dst >= ncols_dst) {
+            return;
+        }
+
+#pragma unroll
+        for (int i = 0; i < mmq_y; i += WARP_SIZE) {
+            const int row_dst = row_dst_0 + threadIdx.x + i;
+
+            if (row_dst >= nrows_dst) {
+                continue;
+            }
+
+            dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps];
+        }
+    }
+}
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_0_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    mul_mat_q4_0(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q4_0_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q4_0_RDNA2;
+    const int nwarps = NWARPS_Q4_0_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q4_0_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q4_0_RDNA1;
+    const int nwarps = NWARPS_Q4_0_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
+        load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q4_0_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q4_0_AMPERE;
+    const int nwarps = NWARPS_Q4_0_AMPERE;
+
+    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
+        load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q4_0_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q4_0_PASCAL;
+    const int nwarps = NWARPS_Q4_0_PASCAL;
+
+    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
+        load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q4_0_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q4_1_RDNA2  64
+#define  MMQ_Y_Q4_1_RDNA2  128
+#define NWARPS_Q4_1_RDNA2  8
+#define  MMQ_X_Q4_1_RDNA1  64
+#define  MMQ_Y_Q4_1_RDNA1  64
+#define NWARPS_Q4_1_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q4_1_AMPERE 4
+#define  MMQ_Y_Q4_1_AMPERE 32
+#define NWARPS_Q4_1_AMPERE 4
+#else
+#define  MMQ_X_Q4_1_AMPERE 64
+#define  MMQ_Y_Q4_1_AMPERE 128
+#define NWARPS_Q4_1_AMPERE 4
+#endif
+#define  MMQ_X_Q4_1_PASCAL 64
+#define  MMQ_Y_Q4_1_PASCAL 64
+#define NWARPS_Q4_1_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_PASCAL, 2)
+#endif // __CUDA_ARCH__ < CC_VOLTA
+    mul_mat_q4_1(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q4_1_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q4_1_RDNA2;
+    const int nwarps = NWARPS_Q4_1_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q4_1_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q4_1_RDNA1;
+    const int nwarps = NWARPS_Q4_1_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
+        load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q4_1_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q4_1_AMPERE;
+    const int nwarps = NWARPS_Q4_1_AMPERE;
+
+    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
+        load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q4_1_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q4_1_PASCAL;
+    const int nwarps = NWARPS_Q4_1_PASCAL;
+
+    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
+        load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q4_1_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q5_0_RDNA2  64
+#define  MMQ_Y_Q5_0_RDNA2  128
+#define NWARPS_Q5_0_RDNA2  8
+#define  MMQ_X_Q5_0_RDNA1  64
+#define  MMQ_Y_Q5_0_RDNA1  64
+#define NWARPS_Q5_0_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q5_0_AMPERE 4
+#define  MMQ_Y_Q5_0_AMPERE 32
+#define NWARPS_Q5_0_AMPERE 4
+#else
+#define  MMQ_X_Q5_0_AMPERE 128
+#define  MMQ_Y_Q5_0_AMPERE 64
+#define NWARPS_Q5_0_AMPERE 4
+#endif
+#define  MMQ_X_Q5_0_PASCAL 64
+#define  MMQ_Y_Q5_0_PASCAL 64
+#define NWARPS_Q5_0_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q5_0_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    mul_mat_q5_0(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q5_0_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q5_0_RDNA2;
+    const int nwarps = NWARPS_Q5_0_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q5_0_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q5_0_RDNA1;
+    const int nwarps = NWARPS_Q5_0_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
+        load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q5_0_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q5_0_AMPERE;
+    const int nwarps = NWARPS_Q5_0_AMPERE;
+
+    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
+        load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q5_0_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q5_0_PASCAL;
+    const int nwarps = NWARPS_Q5_0_PASCAL;
+
+    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
+        load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q5_0_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q5_1_RDNA2  64
+#define  MMQ_Y_Q5_1_RDNA2  128
+#define NWARPS_Q5_1_RDNA2  8
+#define  MMQ_X_Q5_1_RDNA1  64
+#define  MMQ_Y_Q5_1_RDNA1  64
+#define NWARPS_Q5_1_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q5_1_AMPERE 4
+#define  MMQ_Y_Q5_1_AMPERE 32
+#define NWARPS_Q5_1_AMPERE 4
+#else
+#define  MMQ_X_Q5_1_AMPERE 128
+#define  MMQ_Y_Q5_1_AMPERE 64
+#define NWARPS_Q5_1_AMPERE 4
+#endif
+#define  MMQ_X_Q5_1_PASCAL 64
+#define  MMQ_Y_Q5_1_PASCAL 64
+#define NWARPS_Q5_1_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q5_1_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+mul_mat_q5_1(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q5_1_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q5_1_RDNA2;
+    const int nwarps = NWARPS_Q5_1_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q5_1_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q5_1_RDNA1;
+    const int nwarps = NWARPS_Q5_1_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
+        load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q5_1_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q5_1_AMPERE;
+    const int nwarps = NWARPS_Q5_1_AMPERE;
+
+    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
+        load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q5_1_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q5_1_PASCAL;
+    const int nwarps = NWARPS_Q5_1_PASCAL;
+
+    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
+        load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q5_1_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q8_0_RDNA2  64
+#define  MMQ_Y_Q8_0_RDNA2  128
+#define NWARPS_Q8_0_RDNA2  8
+#define  MMQ_X_Q8_0_RDNA1  64
+#define  MMQ_Y_Q8_0_RDNA1  64
+#define NWARPS_Q8_0_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q8_0_AMPERE 4
+#define  MMQ_Y_Q8_0_AMPERE 32
+#define NWARPS_Q8_0_AMPERE 4
+#else
+#define  MMQ_X_Q8_0_AMPERE 128
+#define  MMQ_Y_Q8_0_AMPERE 64
+#define NWARPS_Q8_0_AMPERE 4
+#endif
+#define  MMQ_X_Q8_0_PASCAL 64
+#define  MMQ_Y_Q8_0_PASCAL 64
+#define NWARPS_Q8_0_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q8_0_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    mul_mat_q8_0(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q8_0_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q8_0_RDNA2;
+    const int nwarps = NWARPS_Q8_0_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q8_0_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q8_0_RDNA1;
+    const int nwarps = NWARPS_Q8_0_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
+        load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q8_0_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q8_0_AMPERE;
+    const int nwarps = NWARPS_Q8_0_AMPERE;
+
+    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
+        load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q8_0_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q8_0_PASCAL;
+    const int nwarps = NWARPS_Q8_0_PASCAL;
+
+    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
+        load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q8_0_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q2_K_RDNA2  64
+#define  MMQ_Y_Q2_K_RDNA2  128
+#define NWARPS_Q2_K_RDNA2  8
+#define  MMQ_X_Q2_K_RDNA1  128
+#define  MMQ_Y_Q2_K_RDNA1  32
+#define NWARPS_Q2_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q2_K_AMPERE 4
+#define  MMQ_Y_Q2_K_AMPERE 32
+#define NWARPS_Q2_K_AMPERE 4
+#else
+#define  MMQ_X_Q2_K_AMPERE 64
+#define  MMQ_Y_Q2_K_AMPERE 128
+#define NWARPS_Q2_K_AMPERE 4
+#endif
+#define  MMQ_X_Q2_K_PASCAL 64
+#define  MMQ_Y_Q2_K_PASCAL 64
+#define NWARPS_Q2_K_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q2_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+mul_mat_q2_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q2_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q2_K_RDNA2;
+    const int nwarps = NWARPS_Q2_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q2_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q2_K_RDNA1;
+    const int nwarps = NWARPS_Q2_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
+        load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q2_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q2_K_AMPERE;
+    const int nwarps = NWARPS_Q2_K_AMPERE;
+
+    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
+        load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q2_K_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q2_K_PASCAL;
+    const int nwarps = NWARPS_Q2_K_PASCAL;
+
+    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
+        load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q2_K_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q3_K_RDNA2  128
+#define  MMQ_Y_Q3_K_RDNA2  64
+#define NWARPS_Q3_K_RDNA2  8
+#define  MMQ_X_Q3_K_RDNA1  32
+#define  MMQ_Y_Q3_K_RDNA1  128
+#define NWARPS_Q3_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q3_K_AMPERE 4
+#define  MMQ_Y_Q3_K_AMPERE 32
+#define NWARPS_Q3_K_AMPERE 4
+#else
+#define  MMQ_X_Q3_K_AMPERE 128
+#define  MMQ_Y_Q3_K_AMPERE 128
+#define NWARPS_Q3_K_AMPERE 4
+#endif
+#define  MMQ_X_Q3_K_PASCAL 64
+#define  MMQ_Y_Q3_K_PASCAL 64
+#define NWARPS_Q3_K_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
+    __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_PASCAL, 2)
+#endif // __CUDA_ARCH__ < CC_VOLTA
+    mul_mat_q3_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q3_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q3_K_RDNA2;
+    const int nwarps = NWARPS_Q3_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q3_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q3_K_RDNA1;
+    const int nwarps = NWARPS_Q3_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
+        load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q3_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q3_K_AMPERE;
+    const int nwarps = NWARPS_Q3_K_AMPERE;
+
+    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
+        load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q3_K_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q3_K_PASCAL;
+    const int nwarps = NWARPS_Q3_K_PASCAL;
+
+    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
+        load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q3_K_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q4_K_RDNA2  64
+#define  MMQ_Y_Q4_K_RDNA2  128
+#define NWARPS_Q4_K_RDNA2  8
+#define  MMQ_X_Q4_K_RDNA1  32
+#define  MMQ_Y_Q4_K_RDNA1  64
+#define NWARPS_Q4_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q4_K_AMPERE 4
+#define  MMQ_Y_Q4_K_AMPERE 32
+#define NWARPS_Q4_K_AMPERE 4
+#else
+#define  MMQ_X_Q4_K_AMPERE 64
+#define  MMQ_Y_Q4_K_AMPERE 128
+#define NWARPS_Q4_K_AMPERE 4
+#endif
+#define  MMQ_X_Q4_K_PASCAL 64
+#define  MMQ_Y_Q4_K_PASCAL 64
+#define NWARPS_Q4_K_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_PASCAL, 2)
+#endif // __CUDA_ARCH__ < CC_VOLTA
+    mul_mat_q4_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q4_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q4_K_RDNA2;
+    const int nwarps = NWARPS_Q4_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q4_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q4_K_RDNA1;
+    const int nwarps = NWARPS_Q4_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
+        load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q4_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q4_K_AMPERE;
+    const int nwarps = NWARPS_Q4_K_AMPERE;
+
+    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
+        load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q4_K_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q4_K_PASCAL;
+    const int nwarps = NWARPS_Q4_K_PASCAL;
+
+    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
+        load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q4_K_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q5_K_RDNA2  64
+#define  MMQ_Y_Q5_K_RDNA2  128
+#define NWARPS_Q5_K_RDNA2  8
+#define  MMQ_X_Q5_K_RDNA1  32
+#define  MMQ_Y_Q5_K_RDNA1  64
+#define NWARPS_Q5_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q5_K_AMPERE 4
+#define  MMQ_Y_Q5_K_AMPERE 32
+#define NWARPS_Q5_K_AMPERE 4
+#else
+#define  MMQ_X_Q5_K_AMPERE 64
+#define  MMQ_Y_Q5_K_AMPERE 128
+#define NWARPS_Q5_K_AMPERE 4
+#endif
+#define  MMQ_X_Q5_K_PASCAL 64
+#define  MMQ_Y_Q5_K_PASCAL 64
+#define NWARPS_Q5_K_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q5_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+mul_mat_q5_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q5_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q5_K_RDNA2;
+    const int nwarps = NWARPS_Q5_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q5_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q5_K_RDNA1;
+    const int nwarps = NWARPS_Q5_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
+        load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q5_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q5_K_AMPERE;
+    const int nwarps = NWARPS_Q5_K_AMPERE;
+
+    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
+        load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q5_K_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q5_K_PASCAL;
+    const int nwarps = NWARPS_Q5_K_PASCAL;
+
+    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
+        load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q5_K_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+#define  MMQ_X_Q6_K_RDNA2  64
+#define  MMQ_Y_Q6_K_RDNA2  128
+#define NWARPS_Q6_K_RDNA2  8
+#define  MMQ_X_Q6_K_RDNA1  32
+#define  MMQ_Y_Q6_K_RDNA1  64
+#define NWARPS_Q6_K_RDNA1  8
+#if defined(CUDA_USE_TENSOR_CORES)
+#define  MMQ_X_Q6_K_AMPERE 4
+#define  MMQ_Y_Q6_K_AMPERE 32
+#define NWARPS_Q6_K_AMPERE 4
+#else
+#define  MMQ_X_Q6_K_AMPERE 64
+#define  MMQ_Y_Q6_K_AMPERE 64
+#define NWARPS_Q6_K_AMPERE 4
+#endif
+#define  MMQ_X_Q6_K_PASCAL 64
+#define  MMQ_Y_Q6_K_PASCAL 64
+#define NWARPS_Q6_K_PASCAL 8
+
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
+    __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_PASCAL, 2)
+#endif // __CUDA_ARCH__ < CC_VOLTA
+    mul_mat_q6_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q6_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q6_K_RDNA2;
+    const int nwarps = NWARPS_Q6_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q6_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q6_K_RDNA1;
+    const int nwarps = NWARPS_Q6_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
+        load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
+    const int mmq_x  =  MMQ_X_Q6_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q6_K_AMPERE;
+    const int nwarps = NWARPS_Q6_K_AMPERE;
+
+    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
+        load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int mmq_x  =  MMQ_X_Q6_K_PASCAL;
+    const int mmq_y  =  MMQ_Y_Q6_K_PASCAL;
+    const int nwarps = NWARPS_Q6_K_PASCAL;
+
+    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
+        load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+#else
+    GGML_UNUSED(vec_dot_q6_K_q8_1_mul_mat);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+}
+
+static void ggml_mul_mat_q4_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q4_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_0_RDNA2;
+        nwarps = NWARPS_Q4_0_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q4_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_0_RDNA1;
+        nwarps = NWARPS_Q4_0_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q4_0_AMPERE;
+        mmq_y  =  MMQ_Y_Q4_0_AMPERE;
+        nwarps = NWARPS_Q4_0_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q4_0_PASCAL;
+        mmq_y  =  MMQ_Y_Q4_0_PASCAL;
+        nwarps = NWARPS_Q4_0_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q4_0<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q4_0<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q4_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q4_1_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_1_RDNA2;
+        nwarps = NWARPS_Q4_1_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q4_1_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_1_RDNA1;
+        nwarps = NWARPS_Q4_1_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q4_1_AMPERE;
+        mmq_y  =  MMQ_Y_Q4_1_AMPERE;
+        nwarps = NWARPS_Q4_1_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q4_1_PASCAL;
+        mmq_y  =  MMQ_Y_Q4_1_PASCAL;
+        nwarps = NWARPS_Q4_1_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q4_1<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q4_1<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q5_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q5_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_0_RDNA2;
+        nwarps = NWARPS_Q5_0_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q5_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_0_RDNA1;
+        nwarps = NWARPS_Q5_0_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q5_0_AMPERE;
+        mmq_y  =  MMQ_Y_Q5_0_AMPERE;
+        nwarps = NWARPS_Q5_0_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q5_0_PASCAL;
+        mmq_y  =  MMQ_Y_Q5_0_PASCAL;
+        nwarps = NWARPS_Q5_0_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q5_0<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q5_0<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q5_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q5_1_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_1_RDNA2;
+        nwarps = NWARPS_Q5_1_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q5_1_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_1_RDNA1;
+        nwarps = NWARPS_Q5_1_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q5_1_AMPERE;
+        mmq_y  =  MMQ_Y_Q5_1_AMPERE;
+        nwarps = NWARPS_Q5_1_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q5_1_PASCAL;
+        mmq_y  =  MMQ_Y_Q5_1_PASCAL;
+        nwarps = NWARPS_Q5_1_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q5_1<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q5_1<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q8_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q8_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q8_0_RDNA2;
+        nwarps = NWARPS_Q8_0_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q8_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q8_0_RDNA1;
+        nwarps = NWARPS_Q8_0_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q8_0_AMPERE;
+        mmq_y  =  MMQ_Y_Q8_0_AMPERE;
+        nwarps = NWARPS_Q8_0_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q8_0_PASCAL;
+        mmq_y  =  MMQ_Y_Q8_0_PASCAL;
+        nwarps = NWARPS_Q8_0_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q8_0<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q8_0<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q2_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q2_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q2_K_RDNA2;
+        nwarps = NWARPS_Q2_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q2_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q2_K_RDNA1;
+        nwarps = NWARPS_Q2_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q2_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q2_K_AMPERE;
+        nwarps = NWARPS_Q2_K_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q2_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q2_K_PASCAL;
+        nwarps = NWARPS_Q2_K_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q2_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q2_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q3_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+#if QK_K == 256
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q3_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q3_K_RDNA2;
+        nwarps = NWARPS_Q3_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q3_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q3_K_RDNA1;
+        nwarps = NWARPS_Q3_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q3_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q3_K_AMPERE;
+        nwarps = NWARPS_Q3_K_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q3_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q3_K_PASCAL;
+        nwarps = NWARPS_Q3_K_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+#endif
+}
+
+static void ggml_mul_mat_q4_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q4_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_K_RDNA2;
+        nwarps = NWARPS_Q4_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q4_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_K_RDNA1;
+        nwarps = NWARPS_Q4_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q4_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q4_K_AMPERE;
+        nwarps = NWARPS_Q4_K_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q4_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q4_K_PASCAL;
+        nwarps = NWARPS_Q4_K_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q4_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q4_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q5_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q5_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_K_RDNA2;
+        nwarps = NWARPS_Q5_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q5_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_K_RDNA1;
+        nwarps = NWARPS_Q5_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q5_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q5_K_AMPERE;
+        nwarps = NWARPS_Q5_K_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q5_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q5_K_PASCAL;
+        nwarps = NWARPS_Q5_K_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q5_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q5_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+static void ggml_mul_mat_q6_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
+    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q6_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q6_K_RDNA2;
+        nwarps = NWARPS_Q6_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q6_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q6_K_RDNA1;
+        nwarps = NWARPS_Q6_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
+        mmq_x  =  MMQ_X_Q6_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q6_K_AMPERE;
+        nwarps = NWARPS_Q6_K_AMPERE;
+    } else if (compute_capability >= MIN_CC_DP4A) {
+        mmq_x  =  MMQ_X_Q6_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q6_K_PASCAL;
+        nwarps = NWARPS_Q6_K_PASCAL;
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const dim3 block_nums(block_num_x, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        mul_mat_q6_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    } else {
+        const bool need_check = true;
+        mul_mat_q6_K<need_check><<<block_nums, block_dims, 0, stream>>>
+            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+    }
+}
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t ne10 = src1->ne[0];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+
+    int id = ggml_cuda_get_device();
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_mul_mat_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            ggml_mul_mat_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            ggml_mul_mat_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            ggml_mul_mat_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_mul_mat_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            ggml_mul_mat_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            ggml_mul_mat_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            ggml_mul_mat_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            ggml_mul_mat_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            ggml_mul_mat_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
+}
+
+bool ggml_cuda_supports_mmq(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            return true;
+        default:
+            return false;
+    }
+}

llama/ggml-cuda/mmq.cuh

@@ -0,0 +1,9 @@
+#include "common.cuh"
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_supports_mmq(enum ggml_type type);

llama/ggml-cuda/mmvq.cu

@@ -0,0 +1,404 @@
+#include "mmvq.cuh"
+#include "vecdotq.cuh"
+
+typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
+
+template <int ncols_y, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+// tell the compiler to use as many registers as it wants, see nwarps definition below
+__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void mul_mat_vec_q(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+    constexpr int nwarps              = 1;
+    constexpr int rows_per_cuda_block = 1;
+#else
+    constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
+    constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+
+    const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    const     int row0 = rows_per_cuda_block*blockIdx.x;
+    const     int blocks_per_row_x = ncols_x / qk;
+    const     int blocks_per_col_y = nrows_y / QK8_1;
+    constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
+
+        // x block quant index when casting the quants to int
+        const int kqs = vdr * (tid % (qi/vdr));
+
+#pragma unroll
+        for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp[j][i] += vec_dot_q_cuda(
+                    &x[kbx + (row0 + i)*blocks_per_row_x], &y[j*blocks_per_col_y + kby], kqs);
+            }
+        }
+    }
+
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE];
+    if (threadIdx.y > 0) {
+#pragma unroll
+        for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+            }
+        }
+    }
+    __syncthreads();
+    if (threadIdx.y > 0) {
+        return;
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+        for (int i = 0; i < rows_per_cuda_block; ++i) {
+#pragma unroll
+            for (int l = 0; l < nwarps-1; ++l) {
+                tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+            }
+            tmp[j][i] = warp_reduce_sum(tmp[j][i]);
+        }
+
+        if (threadIdx.x < rows_per_cuda_block) {
+            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
+        }
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot>
+static void mul_mat_vec_q_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    GGML_ASSERT(ncols_x % qk == 0);
+    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
+
+    int id = ggml_cuda_get_device();
+
+    int64_t nwarps = 1;
+    int64_t rows_per_cuda_block = 1;
+
+    if (ggml_cuda_info().devices[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2
+        switch(ncols_y) {
+            case 1:
+                nwarps = 4;
+                rows_per_cuda_block = 1;
+                break;
+            case 2:
+            case 3:
+            case 4:
+                nwarps = 4;
+                rows_per_cuda_block = 2;
+                break;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                nwarps = 2;
+                rows_per_cuda_block = 2;
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+    }
+    const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
+    const dim3 block_nums(nblocks, 1, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    switch (ncols_y) {
+        case 1:
+            mul_mat_vec_q<1, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 2:
+            mul_mat_vec_q<2, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 3:
+            mul_mat_vec_q<3, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 4:
+            mul_mat_vec_q<4, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 5:
+            mul_mat_vec_q<5, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 6:
+            mul_mat_vec_q<6, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 7:
+            mul_mat_vec_q<7, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 8:
+            mul_mat_vec_q<8, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}
+
+static void mul_mat_vec_q4_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q4_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK4_1, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q8_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q2_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q3_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q4_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q6_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_xxs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI2_XXS, block_iq2_xxs, 1, vec_dot_iq2_xxs_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_xs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI2_XS, block_iq2_xs, 1, vec_dot_iq2_xs_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI2_S, block_iq2_s, 1, vec_dot_iq2_s_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq3_xxs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq1_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI1_S, block_iq1_s, 1, vec_dot_iq1_s_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq1_m_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI1_S, block_iq1_m, 1, vec_dot_iq1_m_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq4_nl_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK4_NL, QI4_NL, block_iq4_nl, VDR_Q4_0_Q8_1_MMVQ, vec_dot_iq4_nl_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq4_xs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI4_XS, block_iq4_xs, 1, vec_dot_iq4_xs_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq3_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<QK_K, QI3_XS, block_iq3_s, 1, vec_dot_iq3_s_q8_1>
+        (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+void ggml_cuda_op_mul_mat_vec_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    const int64_t ne10 = src1->ne[0];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    int id = ggml_cuda_get_device();
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}

llama/ggml-cuda/mmvq.cuh

@@ -0,0 +1,7 @@
+#include "common.cuh"
+
+void ggml_cuda_op_mul_mat_vec_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);

llama/ggml-cuda/norm.cu

@@ -0,0 +1,215 @@
+#include "norm.cuh"
+
+template <int block_size>
+static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    float2 mean_var = make_float2(0.f, 0.f);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row*ncols + col];
+        mean_var.x += xi;
+        mean_var.y += xi * xi;
+    }
+
+    // sum up partial sums
+    mean_var = warp_reduce_sum(mean_var);
+    if (block_size > WARP_SIZE) {
+        __shared__ float2 s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = mean_var;
+        }
+        __syncthreads();
+        mean_var = s_sum[lane_id];
+        mean_var = warp_reduce_sum(mean_var);
+    }
+
+    const float mean = mean_var.x / ncols;
+    const float var = mean_var.y / ncols - mean * mean;
+    const float inv_std = rsqrtf(var + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
+    }
+}
+
+template <int block_size>
+static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
+    // blockIdx.x: num_groups idx
+    // threadIdx.x: block_size idx
+    int start = blockIdx.x * group_size;
+    int end = start + group_size;
+
+    start += threadIdx.x;
+
+    if (end >= ne_elements) {
+        end = ne_elements;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += block_size) {
+        tmp += x[j];
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += block_size) {
+        float xi = x[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float variance = tmp / group_size;
+    float scale = rsqrtf(variance + eps);
+    for (int j = start; j < end; j += block_size) {
+        dst[j] *= scale;
+    }
+}
+
+template <int block_size>
+static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row*ncols + col];
+        tmp += xi * xi;
+    }
+
+    // sum up partial sums
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    const float mean = tmp / ncols;
+    const float scale = rsqrtf(mean + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = scale * x[row*ncols + col];
+    }
+}
+
+static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
+}
+
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
+    static const float eps = 1e-6f;
+    if (group_size < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    }
+}
+
+static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
+}
+
+void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
+}
+
+void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int num_groups = dst->op_params[0];
+    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    rms_norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
+}

llama/ggml-cuda/norm.cuh

@@ -0,0 +1,7 @@
+#include "common.cuh"
+
+void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/pad.cu

@@ -0,0 +1,49 @@
+#include "pad.cuh"
+
+static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
+    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
+    // blockIdx.y: idx of ne1
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        dst[offset_dst] = 0.0f;
+    }
+}
+
+static void pad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    pad_f32_cuda(src0_d, dst_d,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}

llama/ggml-cuda/pad.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_PAD_BLOCK_SIZE 256
+
+void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/pool2d.cu

@@ -0,0 +1,94 @@
+#include "pool2d.cuh"
+
+template <typename Ti, typename To>
+static  __global__ void pool2d_nchw_kernel(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const Ti* src, To* dst, const enum ggml_op_pool op) {
+    int idx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (idx >= parallel_elements) {
+        return;
+    }
+
+    const int I_HW = ih * iw;
+    const int O_HW = oh * ow;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / ow;
+    const int cur_ow = idx % O_HW % ow;
+    const Ti* i_ptr = src + nc * I_HW;
+    To* o_ptr = dst + nc * O_HW;
+    const int start_h = cur_oh * sh - ph;
+    const int bh = max(0, start_h);
+    const int eh = min(ih, start_h + kh);
+    const int start_w = cur_ow * sw - pw;
+    const int bw = max(0, start_w);
+    const int ew = min(iw, start_w + kw);
+    const To scale = 1. / (kh * kw);
+    To res = 0;
+
+    switch (op) {
+        case GGML_OP_POOL_AVG: res = 0; break;
+        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        default: assert(false);
+    }
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+#if __CUDA_ARCH__ >= 350
+            Ti cur = __ldg(i_ptr + i * iw + j);
+#else
+            Ti cur = i_ptr[i * iw + j];
+#endif
+            switch (op) {
+                case GGML_OP_POOL_AVG: res += cur * scale; break;
+                case GGML_OP_POOL_MAX: res = max(res, (To)cur); break;
+                default: assert(false);
+            }
+        }
+    }
+    o_ptr[cur_oh * ow + cur_ow] = res;
+}
+
+static void pool2d_nchw_kernel_f32_f32_cuda(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op,
+        cudaStream_t stream) {
+
+    const int num_blocks = (parallel_elements + CUDA_POOL2D_BLOCK_SIZE - 1) / CUDA_POOL2D_BLOCK_SIZE;
+    dim3 block_nums(num_blocks);
+    pool2d_nchw_kernel<<<block_nums, CUDA_POOL2D_BLOCK_SIZE, 0, stream>>>(ih, iw, oh, ow, kh, kw, sh, sw, ph, pw, parallel_elements, src, dst, op);
+}
+
+void ggml_cuda_op_pool2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
+
+    const int64_t IH = src0->ne[1];
+    const int64_t IW = src0->ne[0];
+
+    const int64_t N = dst->ne[3];
+    const int64_t OC = dst->ne[2];
+    const int64_t OH = dst->ne[1];
+    const int64_t OW = dst->ne[0];
+
+    const int parallel_elements = N * OC * OH * OW;
+
+    pool2d_nchw_kernel_f32_f32_cuda(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, parallel_elements, src0_d, dst_d, op, stream);
+}

llama/ggml-cuda/pool2d.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_POOL2D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pool2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/quantize.cu

@@ -0,0 +1,45 @@
+#include "quantize.cuh"
+
+static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx_padded) {
+    const int64_t ix = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (ix >= kx_padded) {
+        return;
+    }
+
+    const int64_t iy = (int64_t)blockDim.y*blockIdx.y + threadIdx.y;
+
+    const int64_t i_padded = (int64_t)iy*kx_padded + ix;
+
+    block_q8_1 * y = (block_q8_1 *) vy;
+
+    const int64_t ib = i_padded / QK8_1; // block index
+    const int64_t iqs = i_padded % QK8_1; // quant index
+
+    const float xi = ix < kx ? x[iy*kx + ix] : 0.0f;
+    float amax = fabsf(xi);
+    float sum = xi;
+
+    amax = warp_reduce_max(amax);
+    sum = warp_reduce_sum(sum);
+
+    const float d = amax / 127;
+    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
+
+    y[ib].qs[iqs] = q;
+
+    if (iqs > 0) {
+        return;
+    }
+
+    reinterpret_cast<half&>(y[ib].ds.x) = d;
+    reinterpret_cast<half&>(y[ib].ds.y) = sum;
+}
+
+void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream) {
+    const int64_t block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+    const dim3 num_blocks(block_num_x, ky, 1);
+    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);
+}
+

llama/ggml-cuda/quantize.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_QUANTIZE_BLOCK_SIZE 256
+
+void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream);

llama/ggml-cuda/rope.cu

@@ -0,0 +1,308 @@
+#include "rope.cuh"
+
+struct rope_corr_dims {
+    float v[4];
+};
+
+static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static __device__ void rope_yarn(
+    float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
+    float * cos_theta, float * sin_theta
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
+    }
+    *cos_theta = cosf(theta) * mscale;
+    *sin_theta = sinf(theta) * mscale;
+}
+
+// rope == RoPE == rotary positional embedding
+template<typename T, bool has_pos>
+static __global__ void rope(
+    const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims
+) {
+    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+    const int i = row*ncols + col;
+    const int i2 = row/p_delta_rows;
+
+    const int p = has_pos ? pos[i2] : 0;
+    const float theta_base = p*powf(freq_base, -float(col)/ncols);
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + 1];
+
+    dst[i + 0] = x0*cos_theta - x1*sin_theta;
+    dst[i + 1] = x0*sin_theta + x1*cos_theta;
+}
+
+template<typename T, bool has_pos>
+static __global__ void rope_neox(
+    const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims
+) {
+    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+    const int ib = col / n_dims;
+    const int ic = col % n_dims;
+
+    if (ib > 0) {
+        const int i = row*ncols + ib*n_dims + ic;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i  = row*ncols + ib*n_dims + ic/2;
+    const int i2 = row/p_delta_rows;
+
+    float cur_rot = inv_ndims * ic - ib;
+
+    const int p = has_pos ? pos[i2] : 0;
+    const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f);
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + n_dims/2];
+
+    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
+    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
+}
+
+static __global__ void rope_glm_f32(
+    const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
+    int n_ctx
+) {
+    const int col = blockDim.x*blockIdx.x + threadIdx.x;
+    const int half_n_dims = ncols/4;
+
+    if (col >= half_n_dims) {
+        return;
+    }
+
+    const int row = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i = row*ncols + col;
+    const int i2 = row/p_delta_rows;
+
+    const float col_theta_scale = powf(freq_base, -2.0f*col/ncols);
+     // FIXME: this is likely wrong
+    const int p = pos != nullptr ? pos[i2] : 0;
+
+    const float theta = min(p, n_ctx - 2)*freq_scale*col_theta_scale;
+    const float sin_theta = sinf(theta);
+    const float cos_theta = cosf(theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + half_n_dims];
+
+    dst[i + 0]           = x0*cos_theta - x1*sin_theta;
+    dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta;
+
+    const float block_theta = ((float)max(p - n_ctx - 2, 0))*col_theta_scale;
+    const float sin_block_theta = sinf(block_theta);
+    const float cos_block_theta = cosf(block_theta);
+
+    const float x2 = x[i + half_n_dims * 2];
+    const float x3 = x[i + half_n_dims * 3];
+
+    dst[i + half_n_dims * 2] = x2*cos_block_theta - x3*sin_block_theta;
+    dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta;
+}
+
+
+template<typename T>
+static void rope_cuda(
+    const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
+) {
+    GGML_ASSERT(ncols % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nrows, num_blocks_x, 1);
+    if (pos == nullptr) {
+        rope<T, false><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
+        );
+    } else {
+        rope<T, true><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
+        );
+    }
+}
+
+template<typename T>
+static void rope_neox_cuda(
+    const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
+) {
+    GGML_ASSERT(ncols % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nrows, num_blocks_x, 1);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float inv_ndims = -1.0f / n_dims;
+
+    if (pos == nullptr) {
+        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+            theta_scale, inv_ndims
+        );
+    } else {
+        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+            theta_scale, inv_ndims
+        );
+    }
+}
+
+static void rope_glm_f32_cuda(
+    const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, int n_ctx, cudaStream_t stream
+) {
+    GGML_ASSERT(ncols % 4 == 0);
+    const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1);
+    const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE;
+    const dim3 block_nums(num_blocks_x, nrows, 1);
+    rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx);
+}
+
+static void rope_cuda_f16(
+    const half * x, half * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
+
+    rope_cuda<half>(x, dst, ncols, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
+}
+
+static void rope_cuda_f32(
+    const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
+
+    rope_cuda<float>(x, dst, ncols, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
+}
+
+static void rope_neox_cuda_f16(
+    const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
+
+    rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
+}
+
+static void rope_neox_cuda_f32(
+    const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
+) {
+
+    rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
+}
+
+void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const int64_t nrows = ggml_nrows(src0);
+
+    //const int n_past      = ((int32_t *) dst->op_params)[0];
+    const int n_dims      = ((int32_t *) dst->op_params)[1];
+    const int mode        = ((int32_t *) dst->op_params)[2];
+    const int n_ctx       = ((int32_t *) dst->op_params)[3];
+    const int n_orig_ctx  = ((int32_t *) dst->op_params)[4];
+
+    // RoPE alteration for extended context
+    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+
+    const int32_t * pos = nullptr;
+    if ((mode & 1) == 0) {
+        GGML_ASSERT(src1->type == GGML_TYPE_I32);
+        GGML_ASSERT(src1->ne[0] == ne2);
+        pos = (const int32_t *) src1_d;
+    }
+
+    const bool is_neox = mode & 2;
+    const bool is_glm  = mode & 4;
+
+    rope_corr_dims corr_dims;
+    ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
+
+    // compute
+    if (is_glm) {
+        GGML_ASSERT(false);
+        rope_glm_f32_cuda(src0_d, dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, stream);
+    } else if (is_neox) {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_neox_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_neox_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, stream
+            );
+        } else {
+            GGML_ASSERT(false);
+        }
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, stream
+            );
+        } else {
+            GGML_ASSERT(false);
+        }
+    }
+}

llama/ggml-cuda/rope.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ROPE_BLOCK_SIZE 256
+
+void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/scale.cu

@@ -0,0 +1,31 @@
+#include "scale.cuh"
+
+static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = scale * x[i];
+}
+
+static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
+    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
+}
+
+void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float scale;
+    memcpy(&scale, dst->op_params, sizeof(float));
+
+    scale_f32_cuda(src0_d, dst_d, scale, ggml_nelements(src0), stream);
+}

llama/ggml-cuda/scale.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_SCALE_BLOCK_SIZE 256
+
+void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/softmax.cu

@@ -0,0 +1,205 @@
+#include "common.cuh"
+#include "softmax.cuh"
+
+template <typename T>
+static __device__ __forceinline__ float t2f32(T val) {
+    return (float) val;
+}
+
+template <>
+__device__ float __forceinline__ t2f32<half>(half val) {
+    return __half2float(val);
+}
+
+template <bool vals_smem, int ncols_template, int block_size_template, typename T>
+static __global__ void soft_max_f32(const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
+    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+
+    const int tid  = threadIdx.x;
+    const int rowx = blockIdx.x;
+    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
+
+    const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
+
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+
+    const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
+
+    extern __shared__ float data_soft_max_f32[];
+    float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
+    // shared memory buffer to cache values between iterations:
+    float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + (int64_t)rowx*ncols;
+
+    float max_val = -INFINITY;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const int64_t ix = (int64_t)rowx*ncols + col;
+        const int64_t iy = (int64_t)rowy*ncols + col;
+
+        const float val = x[ix]*scale + (mask ? slope*t2f32(mask[iy]) : 0.0f);
+
+        vals[col] = val;
+        max_val = max(max_val, val);
+    }
+
+    // find the max value in the block
+    max_val = warp_reduce_max(max_val);
+    if (block_size > WARP_SIZE) {
+        if (warp_id == 0) {
+            buf_iw[lane_id] = -INFINITY;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = max_val;
+        }
+        __syncthreads();
+
+        max_val = buf_iw[lane_id];
+        max_val = warp_reduce_max(max_val);
+    }
+
+    float tmp = 0.0f; // partial sum
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const float val = expf(vals[col] - max_val);
+        tmp += val;
+        vals[col] = val;
+    }
+
+    // find the sum of exps in the block
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __syncthreads();
+        if (warp_id == 0) {
+            buf_iw[lane_id] = 0.0f;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = tmp;
+        }
+        __syncthreads();
+
+        tmp = buf_iw[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    const float inv_sum = 1.0f / tmp;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            return;
+        }
+
+        const int64_t idst = (int64_t)rowx*ncols + col;
+        dst[idst] = vals[col] * inv_sum;
+    }
+}
+
+template<typename T>
+static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+    int nth = WARP_SIZE;
+    while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
+    const dim3 block_dims(nth,     1, 1);
+    const dim3 block_nums(nrows_x, 1, 1);
+    const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
+    static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
+
+    const uint32_t n_head      = nrows_x/nrows_y;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    if (shmem < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
+        switch (ncols_x) {
+            case 32:
+                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 64:
+                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 128:
+                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 256:
+                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 512:
+                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 1024:
+                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 2048:
+                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 4096:
+                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            default:
+                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+        }
+    } else {
+        const size_t shmem_low = WARP_SIZE*sizeof(float);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+    }
+}
+
+void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    const float * src0_d = (const float *)src0->data;
+    const void  * src1_d = src1 ? (const void *)src1->data : nullptr;
+
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
+
+    const int64_t ne00    = src0->ne[0];
+    const int64_t nrows_x = ggml_nrows(src0);
+    const int64_t nrows_y = src0->ne[1];
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
+
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
+
+    if (use_f16) {
+        const half * src1_dd = (const half *)src1_d;
+
+        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+    } else {
+        const float * src1_dd = (const float *)src1_d;
+
+        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+    }
+}

llama/ggml-cuda/softmax.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_SOFT_MAX_BLOCK_SIZE 1024
+
+void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/sumrows.cu

@@ -0,0 +1,40 @@
+#include "sumrows.cuh"
+
+static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col == 0) {
+        dst[row] = sum;
+    }
+}
+
+static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+}
+
+void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
+}

llama/ggml-cuda/sumrows.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/tsembd.cu

@@ -0,0 +1,47 @@
+#include "tsembd.cuh"
+
+static __global__ void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) {
+    // blockIDx.y: idx of timesteps->ne[0]
+    // blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE
+    int i = blockIdx.y;
+    int j = threadIdx.x + blockIdx.x * blockDim.x;
+    float * embed_data = (float *)((char *)dst +  i*nb1);
+
+    if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
+        embed_data[dim] = 0.f;
+    }
+
+    int half = dim / 2;
+    if (j >= half) {
+        return;
+    }
+
+    float timestep = timesteps[i];
+    float freq = (float)expf(-logf(max_period) * j / half);
+    float arg = timestep * freq;
+    embed_data[j] = cosf(arg);
+    embed_data[j + half] = sinf(arg);
+}
+
+static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1,
+                                        const int dim, const int max_period, cudaStream_t stream) {
+    int half_ceil = (dim + 1) / 2;
+    int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne00, 1);
+    timestep_embedding_f32<<<gridDim, CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE, 0, stream>>>(x, dst, nb1, dim, max_period);
+}
+
+void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int dim = dst->op_params[0];
+    const int max_period = dst->op_params[1];
+
+    timestep_embedding_f32_cuda(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
+}

llama/ggml-cuda/tsembd.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
+
+void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/unary.cu

@@ -0,0 +1,266 @@
+#include "unary.cuh"
+
+static __global__ void gelu_f32(const float * x, float * dst, const int k) {
+    const float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    float xi = x[i];
+    dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi)));
+}
+
+static __global__ void gelu_quick_f32(const float * x, float * dst, int k) {
+    const float GELU_QUICK_COEF = -1.702f;
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
+}
+
+static __global__ void silu_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] / (1.0f + expf(-x[i]));
+}
+
+static __global__ void tanh_f32(const float * x, float * dst, int k) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = tanhf(x[i]);
+}
+
+static __global__ void relu_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fmaxf(x[i], 0);
+}
+
+static __global__ void sigmoid_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = 1.0f / (1.0f + expf(-x[i]));
+}
+
+static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
+}
+
+static __global__ void hardswish_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
+}
+
+static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
+}
+
+static __global__ void sqr_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * x[i];
+}
+
+static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
+    gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
+    gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
+    silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
+    tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
+    relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void sigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SIGMOID_BLOCK_SIZE - 1) / CUDA_SIGMOID_BLOCK_SIZE;
+    sigmoid_f32<<<num_blocks, CUDA_SIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE;
+    hardsigmoid_f32<<<num_blocks, CUDA_HARDSIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE;
+    hardswish_f32<<<num_blocks, CUDA_HARDSWISH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
+    leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
+}
+
+static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
+    sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    silu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_quick_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    tanh_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    hardsigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
+    leaky_relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), negative_slope, stream);
+}
+
+void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}

llama/ggml-cuda/unary.cuh

@@ -0,0 +1,30 @@
+#include "common.cuh"
+
+#define CUDA_GELU_BLOCK_SIZE 256
+#define CUDA_SILU_BLOCK_SIZE 256
+#define CUDA_TANH_BLOCK_SIZE 256
+#define CUDA_RELU_BLOCK_SIZE 256
+#define CUDA_SIGMOID_BLOCK_SIZE 256
+#define CUDA_HARDSIGMOID_BLOCK_SIZE 256
+#define CUDA_HARDSWISH_BLOCK_SIZE 256
+#define CUDA_SQR_BLOCK_SIZE 256
+
+void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/upscale.cu

@@ -0,0 +1,51 @@
+#include "upscale.cuh"
+
+static __global__ void upscale_f32(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3) {
+    int index = threadIdx.x + blockIdx.x * blockDim.x;
+    if (index >= ne10 * ne11 * ne12 * ne13) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
+
+    int i00 = i10 / sf0;
+    int i01 = i11 / sf1;
+    int i02 = i12 / sf2;
+    int i03 = i13 / sf3;
+
+    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
+}
+
+static void upscale_f32_cuda(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        cudaStream_t stream) {
+    int dst_size = ne10 * ne11 * ne12 * ne13;
+    int num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+
+    upscale_f32<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
+}
+
+void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const float sf0 = (float)dst->ne[0]/src0->ne[0];
+    const float sf1 = (float)dst->ne[1]/src0->ne[1];
+    const float sf2 = (float)dst->ne[2]/src0->ne[2];
+    const float sf3 = (float)dst->ne[3]/src0->ne[3];
+
+    upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
+}

llama/ggml-cuda/upscale.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_UPSCALE_BLOCK_SIZE 256
+
+void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/vecdotq.cuh

@@ -0,0 +1,1280 @@
+#include "common.cuh"
+
+static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) {
+    const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
+
+    int x32 = 0;
+    x32 |= x16[0] <<  0;
+    x32 |= x16[1] << 16;
+
+    return x32;
+}
+
+static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) {
+    const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
+
+    int x32 = 0;
+    x32 |= x16[0] <<  0;
+    x32 |= x16[1] << 16;
+
+    return x32;
+}
+
+static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) {
+    return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
+}
+
+static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) {
+    return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
+}
+
+
+// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
+// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
+
+#define VDR_Q4_0_Q8_1_MMVQ 2
+#define VDR_Q4_0_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl(
+    const int * v, const int * u, const float & d4, const half2 & ds8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
+        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
+
+        // SIMD dot product of quantized values
+        sumi = __dp4a(vi0, u[2*i+0], sumi);
+        sumi = __dp4a(vi1, u[2*i+1], sumi);
+    }
+
+    const float2 ds8f = __half22float2(ds8);
+
+    // second part effectively subtracts 8 from each quant value
+    return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q4_1_Q8_1_MMVQ 2
+#define VDR_Q4_1_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl(
+    const int * v, const int * u, const half2 & dm4, const half2 & ds8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
+        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
+
+        // SIMD dot product of quantized values
+        sumi = __dp4a(vi0, u[2*i+0], sumi);
+        sumi = __dp4a(vi1, u[2*i+1], sumi);
+    }
+
+#ifdef GGML_CUDA_F16
+    const float2 tmp = __half22float2(__hmul2(dm4, ds8));
+    const float d4d8 = tmp.x;
+    const float m4s8 = tmp.y;
+#else
+    const float2 dm4f = __half22float2(dm4);
+    const float2 ds8f = __half22float2(ds8);
+    const float d4d8 = dm4f.x * ds8f.x;
+    const float m4s8 = dm4f.y * ds8f.y;
+#endif // GGML_CUDA_F16
+
+    // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
+    return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q5_0_Q8_1_MMVQ 2
+#define VDR_Q5_0_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl(
+    const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
+        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
+        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
+        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
+        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
+        sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
+
+        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
+        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
+        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
+        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
+        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
+        sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
+    }
+
+    const float2 ds8f = __half22float2(ds8);
+
+    // second part effectively subtracts 16 from each quant value
+    return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q5_1_Q8_1_MMVQ 2
+#define VDR_Q5_1_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl(
+    const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
+        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
+        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
+        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
+        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
+        sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
+
+        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
+        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
+        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
+        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
+        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
+        sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
+    }
+
+#ifdef GGML_CUDA_F16
+    const float2 tmp = __half22float2(__hmul2(dm5, ds8));
+    const float d5d8 = tmp.x;
+    const float m5s8 = tmp.y;
+#else
+    const float2 dm5f = __half22float2(dm5);
+    const float2 ds8f = __half22float2(ds8);
+    const float d5d8 = dm5f.x * ds8f.x;
+    const float m5s8 = dm5f.y * ds8f.y;
+#endif // GGML_CUDA_F16
+
+    // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
+    return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q8_0_Q8_1_MMVQ 2
+#define VDR_Q8_0_Q8_1_MMQ 8
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl(
+    const int * v, const int * u, const float & d8_0, const float & d8_1) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        // SIMD dot product of quantized values
+        sumi = __dp4a(v[i], u[i], sumi);
+    }
+
+    return d8_0*d8_1 * sumi;
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl(
+    const int * v, const int * u, const half2 & dm8, const half2 & ds8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        // SIMD dot product of quantized values
+        sumi = __dp4a(v[i], u[i], sumi);
+    }
+
+#ifdef GGML_CUDA_F16
+    const float2 tmp = __half22float2(__hmul2(dm8, ds8));
+    const float d8d8 = tmp.x;
+    const float m8s8 = tmp.y;
+#else
+    const float2 dm8f = __half22float2(dm8);
+    const float2 ds8f = __half22float2(ds8);
+    const float d8d8 = dm8f.x * ds8f.x;
+    const float m8s8 = dm8f.y * ds8f.y;
+#endif // GGML_CUDA_F16
+
+    // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
+    return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q2_K_Q8_1_MMVQ 1
+#define VDR_Q2_K_Q8_1_MMQ  2
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
+    const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
+    const half2 & dm2, const float * __restrict__ d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR2_K; ++i) {
+        const int sc = scales[2*i];
+
+        const int vi = (v >> (2*i)) & 0x03030303;
+
+        sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product
+
+        // fill int with 4x m
+        int m = sc >> 4;
+        m |= m <<  8;
+        m |= m << 16;
+        sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values
+    }
+
+    const float2 dm2f = __half22float2(dm2);
+
+    return dm2f.x*sumf_d - dm2f.y*sumf_m;
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+// contiguous u/y values
+static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
+    const half2 & dm2, const float & d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi_d = 0;
+    int sumi_m = 0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) {
+        int sumi_d_sc = 0;
+
+        const int sc = scales[i0 / (QI8_1/2)];
+
+        // fill int with 4x m
+        int m = sc >> 4;
+        m |= m <<  8;
+        m |= m << 16;
+
+#pragma unroll
+        for (int i = i0; i < i0 + QI8_1/2; ++i) {
+            sumi_d_sc = __dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product
+            sumi_m    = __dp4a(m,    u[i], sumi_m); // multiply sum of q8_1 values with m
+        }
+
+        sumi_d += sumi_d_sc * (sc & 0xF);
+    }
+
+    const float2 dm2f = __half22float2(dm2);
+
+    return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m);
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q3_K_Q8_1_MMVQ 1
+#define VDR_Q3_K_Q8_1_MMQ  2
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
+    const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales,
+    const int & scale_offset, const float & d3, const float * __restrict__ d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR3_K; ++i) {
+        const int isc = scale_offset + 2*i;
+
+        const int isc_low = isc % (QK_K/32);
+        const int sc_shift_low = 4 * (isc / (QK_K/32));
+        const int sc_low  = (scales[isc_low] >> sc_shift_low) & 0xF;
+
+        const int isc_high = isc % (QK_K/64);
+        const int sc_shift_high = 2 * (isc / (QK_K/64));
+        const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4;
+
+        const int sc = (sc_low | sc_high) - 32;
+
+        const int vil = (vl >> (2*i)) & 0x03030303;
+
+        const int vih = ((vh >> i) << 2) & 0x04040404;
+
+        const int vi = __vsubss4(vil, vih);
+
+        sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
+    }
+
+    return d3 * sumf;
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+// contiguous u/y values
+static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales,
+    const float & d3, const float & d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    int sumi = 0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) {
+        int sumi_sc = 0;
+
+        for (int i = i0; i < i0 + QI8_1/2; ++i) {
+            sumi_sc = __dp4a(v[i], u[i], sumi_sc); // SIMD dot product
+        }
+
+        sumi += sumi_sc * scales[i0 / (QI8_1/2)];
+    }
+
+    return d3*d8 * sumi;
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q4_K_Q8_1_MMVQ 2
+#define VDR_Q4_K_Q8_1_MMQ  8
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
+    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR4_K; ++i) {
+        const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F;
+        const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F;
+
+        const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product
+        const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u
+
+        sumf_d += d8[i] * (dot1 * sc[i]);
+        sumf_m += d8[i] * (dot2 * m[i]);  // multiply constant part of q4_K with sum of q8_1 values
+    }
+
+    const float2 dm4f = __half22float2(dm4);
+
+    return dm4f.x*sumf_d - dm4f.y*sumf_m;
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+// contiguous u/y values
+static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) {
+        int sumi_d = 0;
+
+#pragma unroll
+        for (int j = 0; j < QI8_1; ++j) {
+            sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product
+        }
+
+        const float2 ds8f = __half22float2(ds8[i]);
+
+        sumf_d += ds8f.x * (sc[i] * sumi_d);
+        sumf_m += ds8f.y *   m[i]; // sum of q8_1 block * q4_K min val
+    }
+
+    const float2 dm4f = __half22float2(dm4);
+
+    return dm4f.x*sumf_d - dm4f.y*sumf_m;
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q5_K_Q8_1_MMVQ 2
+#define VDR_Q5_K_Q8_1_MMQ  8
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
+    const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K; ++i) {
+        const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F;
+        const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F;
+
+        const int vh0i = ((vh[0] >> i) << 4) & 0x10101010;
+        const int vh1i = ((vh[1] >> i) << 4) & 0x10101010;
+
+        const int v0i = vl0i | vh0i;
+        const int v1i = vl1i | vh1i;
+
+        const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product
+        const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u
+
+        sumf_d += d8[i] * (dot1 * sc[i]);
+        sumf_m += d8[i] * (dot2 * m[i]);
+
+    }
+
+    const float2 dm5f = __half22float2(dm5);
+
+    return dm5f.x*sumf_d - dm5f.y*sumf_m;
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+// contiguous u/y values
+static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) {
+        int sumi_d = 0;
+
+#pragma unroll
+        for (int j = 0; j < QI8_1; ++j) {
+            sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product
+        }
+
+        const float2 ds8f = __half22float2(ds8[i]);
+
+        sumf_d += ds8f.x * (sc[i] * sumi_d);
+        sumf_m += ds8f.y *   m[i]; // sum of q8_1 block * q4_K min val
+    }
+
+    const float2 dm4f = __half22float2(dm4);
+
+    return dm4f.x*sumf_d - dm4f.y*sumf_m;
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+#define VDR_Q6_K_Q8_1_MMVQ 1
+#define VDR_Q6_K_Q8_1_MMQ  8
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
+    const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales,
+    const float & d, const float * __restrict__ d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR6_K; ++i) {
+        const int sc = scales[4*i];
+
+        const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
+
+        const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
+
+        const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32
+
+        sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
+    }
+
+    return d*sumf;
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+// contiguous u/y values
+static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc,
+    const float & d6, const float * __restrict__ d8) {
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    float sumf_d = 0.0f;
+
+#pragma unroll
+    for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) {
+        int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale
+
+#pragma unroll
+        for (int i = i0; i < i0 + 2; ++i) {
+            sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product
+            sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product
+
+            sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product
+            sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product
+        }
+
+        sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y);
+    }
+
+    return d6 * sumf_d;
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
+
+    int v[VDR_Q4_0_Q8_1_MMVQ];
+    int u[2*VDR_Q4_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
+        v[i]     = get_int_from_uint8(bq4_0->qs, iqs + i);
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
+    }
+
+    return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
+}
+
+
+static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
+
+    int v[VDR_Q4_1_Q8_1_MMVQ];
+    int u[2*VDR_Q4_1_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) {
+        v[i]    = get_int_from_uint8_aligned(bq4_1->qs, iqs + i);
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1);
+    }
+
+    return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
+}
+
+static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
+
+    int vl[VDR_Q5_0_Q8_1_MMVQ];
+    int vh[VDR_Q5_0_Q8_1_MMVQ];
+    int  u[2*VDR_Q5_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) {
+        vl[i]    = get_int_from_uint8(bq5_0->qs, iqs + i);
+        vh[i]    = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i));
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0);
+    }
+
+    return vec_dot_q5_0_q8_1_impl<VDR_Q5_0_Q8_1_MMVQ>(vl, vh, u, bq5_0->d, bq8_1->ds);
+}
+
+static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
+
+    int vl[VDR_Q5_1_Q8_1_MMVQ];
+    int vh[VDR_Q5_1_Q8_1_MMVQ];
+    int  u[2*VDR_Q5_1_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) {
+        vl[i]   = get_int_from_uint8_aligned(bq5_1->qs, iqs + i);
+        vh[i]   = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i));
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1);
+    }
+
+    return vec_dot_q5_1_q8_1_impl<VDR_Q5_1_Q8_1_MMVQ>(vl, vh, u, bq5_1->dm, bq8_1->ds);
+}
+
+static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
+
+    int v[VDR_Q8_0_Q8_1_MMVQ];
+    int u[VDR_Q8_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) {
+        v[i] = get_int_from_int8(bq8_0->qs, iqs + i);
+        u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+    }
+
+    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
+}
+
+static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q2_K * bq2_K = (const block_q2_K *) vbq;
+
+    const int bq8_offset = QR2_K * (iqs / QI8_1);
+    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
+
+    const uint8_t * scales = bq2_K->scales + scale_offset;
+
+    const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs);
+    int    u[QR2_K];
+    float d8[QR2_K];
+
+#pragma unroll
+    for (int i = 0; i < QR2_K; ++ i) {
+        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
+    }
+
+    return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
+}
+
+static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q3_K * bq3_K = (const block_q3_K *) vbq;
+
+    const int bq8_offset = QR3_K * (iqs / (QI3_K/2));
+    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
+
+    const float d = bq3_K->d;
+
+    const int vl = get_int_from_uint8(bq3_K->qs, iqs);
+
+    // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
+    const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset;
+
+    int    u[QR3_K];
+    float d8[QR3_K];
+
+#pragma unroll
+    for (int i = 0; i < QR3_K; ++i) {
+        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
+    }
+
+    return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
+}
+
+static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+#ifndef GGML_QKK_64
+    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
+
+    int    v[2];
+    int    u[2*QR4_K];
+    float d8[QR4_K];
+
+    // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
+    const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
+
+    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
+    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
+    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
+    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
+
+    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
+    v[0] = q4[0];
+    v[1] = q4[4];
+
+    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
+
+    for (int i = 0; i < QR4_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i] = __low2float(bq8i->ds);
+
+        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
+        u[2*i+0] = q8[0];
+        u[2*i+1] = q8[4];
+    }
+
+    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
+
+#else
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+    uint16_t aux16[2];
+    const uint8_t * s = (const uint8_t *)aux16;
+
+    const uint16_t * a = (const uint16_t *)bq4_K->scales;
+    aux16[0] = a[0] & 0x0f0f;
+    aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+    const float dall = bq4_K->dm[0];
+    const float dmin = bq4_K->dm[1];
+
+    const float d8_1 = __low2float(bq8_1[0].ds);
+    const float d8_2 = __low2float(bq8_1[1].ds);
+
+    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
+    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
+    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
+    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
+
+    const int * q4 = (const int *)bq4_K->qs + (iqs/2);
+    const int v1 = q4[0];
+    const int v2 = q4[4];
+
+    const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0));
+    const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
+    const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
+    const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0));
+
+    sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
+    sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
+
+    return dall * sumf_d - dmin * sumf_m;
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+#ifndef GGML_QKK_64
+    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
+
+    int   vl[2];
+    int   vh[2];
+    int    u[2*QR5_K];
+    float d8[QR5_K];
+
+    const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
+    const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
+    const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
+
+    vl[0] = ql[0];
+    vl[1] = ql[4];
+
+    vh[0] = qh[0] >> bq8_offset;
+    vh[1] = qh[4] >> bq8_offset;
+
+    const uint16_t * scales = (const uint16_t *)bq5_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i] = __low2float(bq8i->ds);
+
+        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
+        u[2*i+0] = q8[0];
+        u[2*i+1] = q8[4];
+    }
+
+    return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
+
+#else
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
+
+    const int8_t * s = bq5_K->scales;
+
+    const float d = bq5_K->d;
+
+    const float d8_1 = __low2half(bq8_1[0].ds);
+    const float d8_2 = __low2half(bq8_1[1].ds);
+
+    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
+    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
+    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
+    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
+
+    const int * ql = (const int *)bq5_K->qs + (iqs/2);
+    const int vl1 = ql[0];
+    const int vl2 = ql[4];
+
+    const int step = 4 * (iqs/2); // 0, 4, 8, 12
+    const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
+    const int in = step%8; // 0, 4, 0, 4
+    const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
+
+    const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
+    const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
+    const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
+    const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
+
+    const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1])
+                       + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]);
+
+    return d * sumf_d;
+
+#else
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_q6_K * bq6_K = (const block_q6_K *) vbq;
+
+    const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4);
+    const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8);
+    const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4));
+
+    const int vl = get_int_from_uint8(bq6_K->ql, iqs);
+    const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift;
+
+    const int8_t * scales = bq6_K->scales + scale_offset;
+
+    int    u[QR6_K];
+    float d8[QR6_K];
+
+#pragma unroll
+    for (int i = 0; i < QR6_K; ++i) {
+        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
+        d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds);
+    }
+
+    return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
+}
+
+static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if QK_K == 256
+    const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
+
+#if QR2_XXS == 8
+    const int ib32 = iqs;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    uint32_t aux32 = q2[2] | (q2[3] << 16);
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+        const uint8_t  signs = ksigns_iq2xs[aux32 & 127];
+        for (int j = 0; j < 8; ++j) {
+            sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+        }
+        q8 += 8;
+        aux32 >>= 7;
+    }
+    const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f;
+    return d * sumi;
+#else
+    // iqs is 0...15
+    const int ib32 = iqs/2;
+    const int il = iqs%2;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f;
+    const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127];
+    const int8_t * q8 = bq8_1[ib32].qs + 16*il;
+    int sumi1 = 0, sumi2 = 0;
+    for (int j = 0; j < 8; ++j) {
+        sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1);
+        sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1);
+    }
+    return d * (sumi1 + sumi2);
+#endif
+#else
+    NO_DEVICE_CODE;
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
+
+    const int ib32 = iqs;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    const uint8_t ls1 = bq2->scales[ib32] & 0xf;
+    const uint8_t ls2 = bq2->scales[ib32] >>  4;
+    int sumi1 = 0;
+    for (int l = 0; l < 2; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
+        const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
+        sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
+        sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
+        q8 += 8;
+    }
+    int sumi2 = 0;
+    for (int l = 2; l < 4; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
+        const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
+        sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
+        sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
+    return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
+#else
+    GGML_UNUSED(ksigns64);
+    NO_DEVICE_CODE;
+#endif
+#else
+    GGML_UNUSED(ksigns64);
+    NO_DEVICE_CODE;
+#endif
+}
+
+// TODO
+static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
+
+    const int ib32 = iqs;
+    const int8_t  * q8 = bq8_1[ib32].qs;
+    const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32;
+    const uint8_t ls1 = bq2->scales[ib32] & 0xf;
+    const uint8_t ls2 = bq2->scales[ib32] >>  4;
+    int sumi1 = 0;
+    for (int l = 0; l < 2; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
+        const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
+        const uint32_t signs1 = __vcmpeq4(((signs[l] >>  4) * 0x01010101) & 0x08040201, 0x08040201);
+        const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
+        const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
+        sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
+        sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
+        q8 += 8;
+    }
+    int sumi2 = 0;
+    for (int l = 2; l < 4; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
+        const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
+        const uint32_t signs1 = __vcmpeq4(((signs[l] >>  4) * 0x01010101) & 0x08040201, 0x08040201);
+        const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
+        const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
+        sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
+        sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
+    return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
+#else
+    GGML_UNUSED(ksigns64);
+    NO_DEVICE_CODE;
+#endif
+#else
+    GGML_UNUSED(ksigns64);
+    NO_DEVICE_CODE;
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
+
+    const int ib32 = iqs;
+    const uint8_t  * q3 = bq2->qs + 8*ib32;
+    const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    uint32_t aux32 = gas[0] | (gas[1] << 16);
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0];
+        const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1];
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127));
+        const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]);
+        sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
+        sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
+        q8 += 8;
+        aux32 >>= 7;
+    }
+    const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f;
+    return d * sumi;
+#else
+    NO_DEVICE_CODE;
+#endif
+#else
+    NO_DEVICE_CODE;
+#endif
+}
+
+// TODO: don't use lookup table for signs
+static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
+
+    const int ib32 = iqs;
+    const uint8_t  * qs = bq2->qs + 8*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256));
+        const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256));
+        uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
+        uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >>  4) * 0x01010101) & 0x08040201, 0x08040201);
+        const int grid_l = __vsub4(grid1[0] ^ signs0, signs0);
+        const int grid_h = __vsub4(grid2[0] ^ signs1, signs1);
+        sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
+        sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds);
+    return d * sumi;
+#else
+    NO_DEVICE_CODE;
+#endif
+#else
+    NO_DEVICE_CODE;
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if QK_K == 256
+    const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
+
+    const int ib32 = iqs;
+    int sumi = 0;
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const int * q8 = (const int *)bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
+        int grid0 = grid[0] & 0x0f0f0f0f;
+        int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
+        sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi));
+    }
+#else
+    const int8_t * q8 = bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
+        for (int j = 0; j < 4; ++j) {
+            sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
+        }
+        q8 += 8;
+    }
+#endif
+    const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA;
+    const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1);
+    const float d = d1q * __low2float (bq8_1[ib32].ds);
+    const float m = d1q * __high2float(bq8_1[ib32].ds);
+    return d * sumi + m * delta;
+#else
+    NO_DEVICE_CODE;
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if QK_K == 256
+    const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
+
+    const int ib32 = iqs;
+    int   sumi[2] = {0, 0};
+    float sumf[2] = {0.f, 0.f};
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const int * q8 = (const int *)bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8)));
+        int grid0 = grid[0] & 0x0f0f0f0f;
+        int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
+        sumi[l/2] = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi[l/2]));
+        const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
+        const int sumy = __dp4a(q8[2*l+1], 0x01010101, __dp4a(q8[2*l+0], 0x01010101, 0));
+        sumf[l/2] += delta*sumy;
+    }
+#else
+    const int8_t * q8 = bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
+        int sumy = 0;
+        for (int j = 0; j < 4; ++j) {
+            sumi[l/2] += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
+            sumy += q8[j] + q8[j+4];
+        }
+        const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
+        sumf[l/2] += delta*sumy;
+        q8 += 8;
+    }
+#endif
+    iq1m_scale_t scale;
+    const uint16_t * sc = (const uint16_t *)bq1->scales;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds);
+    return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
+#else
+    NO_DEVICE_CODE;
+#endif
+}
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values,
+        int & val1, int & val2) {
+
+    uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32;
+    aux32 = q4 & 0x0f0f0f0f;
+    uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8);
+    uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8);
+    val1 = v1 | (v2 << 16);
+    aux32 = (q4 >> 4) & 0x0f0f0f0f;
+    v1 = values[q8[0]] | (values[q8[1]] << 8);
+    v2 = values[q8[2]] | (values[q8[3]] << 8);
+    val2 = v1 | (v2 << 16);
+}
+#endif
+
+static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_iq4_nl * bq = (const block_iq4_nl *) vbq;
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs;
+    const int32_t  * q8 = (const int32_t  *)bq8_1->qs + iqs;
+
+    const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
+
+    int v1, v2;
+    int sumi1 = 0, sumi2 = 0;
+    for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
+        const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16);
+        get_int_from_table_16(aux, values, v1, v2);
+        sumi1 = __dp4a(v1, q8[l+0], sumi1);
+        sumi2 = __dp4a(v2, q8[l+4], sumi2);
+    }
+
+#else
+    const uint8_t * q4 = bq->qs + 4*iqs;
+    const int8_t  * q8 = bq8_1->qs + 4*iqs;
+
+    int sumi1 = 0, sumi2 = 0;
+    for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) {
+        sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf];
+        sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >>  4];
+    }
+#endif
+    const float d = (float)bq->d * __low2float(bq8_1->ds);
+    return d * (sumi1 + sumi2);
+}
+
+static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+#if QK_K == 256
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+
+    const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
+    const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
+
+    // iqs is 0...7
+    const int ib32 = iqs;
+    const int32_t  * q8 = (const int *)bq8_1[ib32].qs;
+    const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32;
+    const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
+    int v1, v2;
+    int sumi1 = 0, sumi2 = 0;
+    for (int j = 0; j < 4; ++j) {
+        get_int_from_table_16(q4[j], values, v1, v2);
+        sumi1 = __dp4a(v1, q8[j+0], sumi1);
+        sumi2 = __dp4a(v2, q8[j+4], sumi2);
+    }
+    return d * (sumi1 + sumi2);
+
+#else
+    NO_DEVICE_CODE;
+#endif
+#else
+    return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);
+#endif
+}

llama/ggml-impl.h

@@ -11,6 +11,89 @@
 #include <string.h> // memcpy
 #include <math.h>   // fabsf
 
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
+ * Subnormals shall be flushed to zero, and NANs will be quiet.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    if (!(u.i & 0x7f800000)) { /* subnormal */
+        h.bits = (u.i & 0x80000000) >> 16; /* flush to zero */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -37,15 +120,22 @@ extern "C" {
 #ifndef __F16C__
 #define __F16C__
 #endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __SSE3__
 #define __SSE3__
 #endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
 #endif
 
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
+#if defined(__ARM_NEON)
 
 // if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
 //
@@ -53,8 +143,262 @@ extern "C" {
 //
 #include <arm_neon.h>
 
+#ifdef _MSC_VER
+
+typedef uint16_t ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
 typedef __fp16 ggml_fp16_internal_t;
 
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddvq_s16(int16x8_t v) {
+    return
+        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
+        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
+        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
+        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
 #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
 #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
 
@@ -75,8 +419,6 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 
 #else
 
-typedef uint16_t ggml_fp16_internal_t;
-
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@@ -221,7 +563,7 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 
 #endif // __F16C__
 
-#endif // __ARM_NEON
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
 
 // precomputed f32 table for f16 (256 KB)
 // defined in ggml.c, initialized in ggml_init()

llama/ggml-metal.m → llama/ggml-metal-darwin_arm64.m

@@ -40,14 +40,17 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_CLAMP,
     GGML_METAL_KERNEL_TYPE_TANH,
     GGML_METAL_KERNEL_TYPE_RELU,
+    GGML_METAL_KERNEL_TYPE_SIGMOID,
     GGML_METAL_KERNEL_TYPE_GELU,
     GGML_METAL_KERNEL_TYPE_GELU_4,
     GGML_METAL_KERNEL_TYPE_GELU_QUICK,
     GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
     GGML_METAL_KERNEL_TYPE_SILU,
     GGML_METAL_KERNEL_TYPE_SILU_4,
-    GGML_METAL_KERNEL_TYPE_SOFT_MAX,
-    GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4,
     GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,
     GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,
@@ -167,7 +170,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F16,
-    GGML_METAL_KERNEL_TYPE_ALIBI_F32,
     GGML_METAL_KERNEL_TYPE_IM2COL_F16,
     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
@@ -177,6 +179,14 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
     GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
@@ -255,11 +265,20 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
 
 static void * ggml_metal_host_malloc(size_t n) {
     void * data = NULL;
+
+#if TARGET_OS_OSX
+    kern_return_t err = vm_allocate((vm_map_t) mach_task_self(), (void *) &data, n, VM_FLAGS_ANYWHERE);
+    if (err != KERN_SUCCESS) {
+        GGML_METAL_LOG_ERROR("%s: error: vm_allocate failed\n", __func__);
+        return NULL;
+    }
+#else
     const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
     if (result != 0) {
         GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
         return NULL;
     }
+#endif
 
     return data;
 }
@@ -443,7 +462,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         }
 
         /*
-            GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
+            GGML_METAL_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
                     (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
                     (int) kernel->pipeline.threadExecutionWidth); \
         */
@@ -459,172 +478,182 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 return NULL; \
             } \
         } else { \
-            GGML_METAL_LOG_WARN("%s: skipping %-32s (not supported)\n", __func__, "kernel_"#name); \
+            GGML_METAL_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \
         }
 
         // simd_sum and simd_max requires MTLGPUFamilyApple7
 
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                       add,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                   add_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                       mul,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                   mul_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                       div,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                   div_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                     scale,                  true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                   scale_4,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                     clamp,                  true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                      tanh,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                      relu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                      gelu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4,                    gelu_4,                 true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                gelu_quick,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,              gelu_quick_4,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                      silu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                    silu_4,                 true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX,                  soft_max,               ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,                soft_max_4,             ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,             diag_mask_inf,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,           diag_mask_inf_8,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,              get_rows_f32,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,              get_rows_f16,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,             get_rows_q4_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,             get_rows_q4_1,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,             get_rows_q5_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,             get_rows_q5_1,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,             get_rows_q8_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,             get_rows_q2_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,             get_rows_q3_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,             get_rows_q4_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,             get_rows_q5_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,             get_rows_q6_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,          get_rows_iq2_xxs,       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,           get_rows_iq2_xs,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,          get_rows_iq3_xxs,       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,            get_rows_iq3_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,            get_rows_iq2_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,            get_rows_iq1_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,            get_rows_iq1_m,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,           get_rows_iq4_nl,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,           get_rows_iq4_xs,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                  rms_norm,               ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                group_norm,             ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                      norm,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,            mul_mv_f32_f32,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,            mul_mv_f16_f16,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,            mul_mv_f16_f32,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,       mul_mv_f16_f32_1row,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,         mul_mv_f16_f32_l4,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,           mul_mv_q4_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,           mul_mv_q4_1_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,           mul_mv_q5_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,           mul_mv_q5_1_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,           mul_mv_q8_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,           mul_mv_q2_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,           mul_mv_q3_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,           mul_mv_q4_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,           mul_mv_q5_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,           mul_mv_q6_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,        mul_mv_iq2_xxs_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,         mul_mv_iq2_xs_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,        mul_mv_iq3_xxs_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,          mul_mv_iq3_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,          mul_mv_iq2_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,          mul_mv_iq1_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,          mul_mv_iq1_m_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,         mul_mv_iq4_nl_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,         mul_mv_iq4_xs_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,      ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,         mul_mv_id_f16_f16,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,         mul_mv_id_f16_f32,      ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,    mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,      mul_mv_id_f16_f32_l4,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,        mul_mv_id_q4_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,        mul_mv_id_q4_1_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,        mul_mv_id_q5_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,        mul_mv_id_q5_1_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,        mul_mv_id_q8_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,        mul_mv_id_q2_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,        mul_mv_id_q3_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,        mul_mv_id_q4_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,        mul_mv_id_q5_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,        mul_mv_id_q6_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,     mul_mv_id_iq2_xxs_f32,  ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,      mul_mv_id_iq2_xs_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,     mul_mv_id_iq3_xxs_f32,  ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,       mul_mv_id_iq3_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,       mul_mv_id_iq2_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,       mul_mv_id_iq1_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,       mul_mv_id_iq1_m_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,      mul_mv_id_iq4_nl_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,      mul_mv_id_iq4_xs_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,            mul_mm_f16_f32,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,           mul_mm_q4_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,           mul_mm_q4_1_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,           mul_mm_q5_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,           mul_mm_q5_1_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,           mul_mm_q8_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,           mul_mm_q2_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,           mul_mm_q3_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,           mul_mm_q4_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,           mul_mm_q5_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,           mul_mm_q6_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,        mul_mm_iq2_xxs_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,         mul_mm_iq2_xs_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,        mul_mm_iq3_xxs_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,          mul_mm_iq3_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,          mul_mm_iq2_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,          mul_mm_iq1_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,          mul_mm_iq1_m_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,         mul_mm_iq4_nl_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,         mul_mm_iq4_xs_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,         mul_mm_id_f16_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,        mul_mm_id_q4_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,        mul_mm_id_q4_1_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,        mul_mm_id_q5_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,        mul_mm_id_q5_1_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,        mul_mm_id_q8_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,        mul_mm_id_q2_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,        mul_mm_id_q3_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,        mul_mm_id_q4_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,        mul_mm_id_q5_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,        mul_mm_id_q6_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,     mul_mm_id_iq2_xxs_f32,  ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,      mul_mm_id_iq2_xs_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,     mul_mm_id_iq3_xxs_f32,  ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,       mul_mm_id_iq3_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,       mul_mm_id_iq2_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,       mul_mm_id_iq1_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,       mul_mm_id_iq1_m_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,      mul_mm_id_iq4_nl_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,      mul_mm_id_iq4_xs_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,               true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                  rope_f16,               true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                 alibi_f32,              true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                im2col_f16,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                im2col_f32,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,               upscale_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                   pad_f32,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,    timestep_embedding_f32, true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                arange_f32,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,       argsort_f32_i32_asc,    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,      argsort_f32_i32_desc,   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,            leaky_relu_f32,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,               cpy_f32_f16,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,               cpy_f32_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,              cpy_f32_q8_0,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,              cpy_f32_q4_0,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,              cpy_f32_q4_1,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,              cpy_f32_q5_0,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,              cpy_f32_q5_1,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,            cpy_f32_iq4_nl,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,               cpy_f16_f16,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,               cpy_f16_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                    concat,                 true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                       sqr,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                  sum_rows,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                           add,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                       add_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                           mul,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                       mul_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                           div,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                       div_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                         scale,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                       scale_4,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                         clamp,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                          tanh,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                          relu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID,                       sigmoid,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                          gelu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4,                        gelu_4,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                    gelu_quick,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,                  gelu_quick_4,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                          silu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                        silu_4,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,                  soft_max_f16,                   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,                soft_max_f16_4,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,                  soft_max_f32,                   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4,                soft_max_f32_4,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,                 diag_mask_inf,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,               diag_mask_inf_8,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,                  get_rows_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,                  get_rows_f16,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,                 get_rows_q4_0,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,                 get_rows_q4_1,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,                 get_rows_q5_0,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,                 get_rows_q5_1,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,                 get_rows_q8_0,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,                 get_rows_q2_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,                 get_rows_q3_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,                 get_rows_q4_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,                 get_rows_q5_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,                 get_rows_q6_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,              get_rows_iq2_xxs,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,               get_rows_iq2_xs,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,              get_rows_iq3_xxs,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,                get_rows_iq3_s,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,                get_rows_iq2_s,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,                get_rows_iq1_s,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,                get_rows_iq1_m,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,               get_rows_iq4_nl,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,               get_rows_iq4_xs,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,                  get_rows_i32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                      rms_norm,                       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                    group_norm,                     ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                          norm,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                mul_mv_f32_f32,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,                mul_mv_f16_f16,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,                mul_mv_f16_f32,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,           mul_mv_f16_f32_1row,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,             mul_mv_f16_f32_l4,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,               mul_mv_q4_0_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,               mul_mv_q4_1_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,               mul_mv_q5_0_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,               mul_mv_q5_1_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,               mul_mv_q8_0_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,               mul_mv_q2_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,               mul_mv_q3_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,               mul_mv_q4_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,               mul_mv_q5_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,               mul_mv_q6_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,            mul_mv_iq2_xxs_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,             mul_mv_iq2_xs_f32,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,            mul_mv_iq3_xxs_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,              mul_mv_iq3_s_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,              mul_mv_iq2_s_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,              mul_mv_iq1_s_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,              mul_mv_iq1_m_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,             mul_mv_iq4_nl_f32,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,             mul_mv_iq4_xs_f32,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,             mul_mv_id_f32_f32,              ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,             mul_mv_id_f16_f16,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,             mul_mv_id_f16_f32,              ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,        mul_mv_id_f16_f32_1row,         ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,          mul_mv_id_f16_f32_l4,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,            mul_mv_id_q4_0_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,            mul_mv_id_q4_1_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,            mul_mv_id_q5_0_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,            mul_mv_id_q5_1_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,            mul_mv_id_q8_0_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,            mul_mv_id_q2_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,            mul_mv_id_q3_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,            mul_mv_id_q4_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,            mul_mv_id_q5_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,            mul_mv_id_q6_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,         mul_mv_id_iq2_xxs_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,          mul_mv_id_iq2_xs_f32,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,         mul_mv_id_iq3_xxs_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,           mul_mv_id_iq3_s_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,           mul_mv_id_iq2_s_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,           mul_mv_id_iq1_s_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,           mul_mv_id_iq1_m_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,          mul_mv_id_iq4_nl_f32,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,          mul_mv_id_iq4_xs_f32,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,                mul_mm_f32_f32,                 ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,                mul_mm_f16_f32,                 ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,               mul_mm_q4_0_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,               mul_mm_q4_1_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,               mul_mm_q5_0_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,               mul_mm_q5_1_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,               mul_mm_q8_0_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,               mul_mm_q2_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,               mul_mm_q3_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,               mul_mm_q4_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,               mul_mm_q5_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,               mul_mm_q6_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,            mul_mm_iq2_xxs_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,             mul_mm_iq2_xs_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,            mul_mm_iq3_xxs_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,              mul_mm_iq3_s_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,              mul_mm_iq2_s_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,              mul_mm_iq1_s_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,              mul_mm_iq1_m_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,             mul_mm_iq4_nl_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,             mul_mm_iq4_xs_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,             mul_mm_id_f32_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,             mul_mm_id_f16_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,            mul_mm_id_q4_0_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,            mul_mm_id_q4_1_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,            mul_mm_id_q5_0_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,            mul_mm_id_q5_1_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,            mul_mm_id_q8_0_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,            mul_mm_id_q2_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,            mul_mm_id_q3_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,            mul_mm_id_q4_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,            mul_mm_id_q5_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,            mul_mm_id_q6_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,         mul_mm_id_iq2_xxs_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,          mul_mm_id_iq2_xs_f32,           ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,         mul_mm_id_iq3_xxs_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,           mul_mm_id_iq3_s_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,           mul_mm_id_iq2_s_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,           mul_mm_id_iq1_s_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,           mul_mm_id_iq1_m_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,          mul_mm_id_iq4_nl_f32,           ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,          mul_mm_id_iq4_xs_f32,           ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                      rope_f32,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                      rope_f16,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                    im2col_f16,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                    arange_f32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,          argsort_f32_i32_desc,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,                leaky_relu_f32,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,        flash_attn_ext_f16_h64,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,        flash_attn_ext_f16_h80,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,        flash_attn_ext_f16_h96,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,       flash_attn_ext_f16_h112,        ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,       flash_attn_ext_f16_h128,        ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,       flash_attn_ext_f16_h256,        ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,   flash_attn_ext_vec_f16_h128,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,   flash_attn_ext_vec_f16_h256,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,                   cpy_f32_f16,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                   cpy_f32_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,                  cpy_f32_q8_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,                  cpy_f32_q4_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,                  cpy_f32_q4_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,                  cpy_f32_q5_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,                  cpy_f32_q5_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,                cpy_f32_iq4_nl,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,                   cpy_f16_f16,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,                   cpy_f16_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                        concat,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                           sqr,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                      sum_rows,                       true);
     }
 
     [metal_library release];
@@ -703,6 +732,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
@@ -730,7 +760,6 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_GROUP_NORM:
             return ctx->support_simdgroup_reduction;
         case GGML_OP_NORM:
-        case GGML_OP_ALIBI:
         case GGML_OP_ROPE:
         case GGML_OP_IM2COL:
             return true;
@@ -744,6 +773,8 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_ARGSORT:
         case GGML_OP_LEAKY_RELU:
             return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+            return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             return ctx->support_simdgroup_reduction &&
@@ -782,7 +813,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_GET_ROWS:
             {
-                return op->ne[3] == 1;
+                return op->src[0]->type != GGML_TYPE_BF16 && op->ne[3] == 1;
             }
         default:
             return false;
@@ -1164,24 +1195,24 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     } break;
                 case GGML_OP_CLAMP:
-                {
-                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
+                    {
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
 
-                    float min;
-                    float max;
-                    memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
-                    memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
+                        float min;
+                        float max;
+                        memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
+                        memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
 
-                    [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&min length:sizeof(min) atIndex:2];
-                    [encoder setBytes:&max length:sizeof(max) atIndex:3];
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
+                        [encoder setBytes:&min length:sizeof(min) atIndex:2];
+                        [encoder setBytes:&max length:sizeof(max) atIndex:3];
 
-                    const int64_t n = ggml_nelements(dst);
+                        const int64_t n = ggml_nelements(dst);
 
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
+                        [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
                 case GGML_OP_UNARY:
                     switch (ggml_get_unary_op(gf->nodes[i])) {
                         // we are not taking into account the strides, so for now require contiguous tensors
@@ -1209,6 +1240,18 @@ static enum ggml_status ggml_metal_graph_compute(
 
                                 const int64_t n = ggml_nelements(dst);
 
+                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            } break;
+                        case GGML_UNARY_OP_SIGMOID:
+                            {
+                                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline;
+
+                                [encoder setComputePipelineState:pipeline];
+                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                const int64_t n = ggml_nelements(dst);
+
                                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                             } break;
                         case GGML_UNARY_OP_GELU:
@@ -1326,20 +1369,32 @@ static enum ggml_status ggml_metal_graph_compute(
                     } break;
                 case GGML_OP_SOFT_MAX:
                     {
+                        GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
+
                         int nth = 32; // SIMD width
 
                         id<MTLComputePipelineState> pipeline = nil;
 
+                        const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
+
                         if (ne00%4 == 0) {
-                            while (nth < ne00/4 && nth < 256) {
+                            while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) {
                                 nth *= 2;
                             }
-                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_4].pipeline;
+                            if (use_f16) {
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline;
+                            } else {
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline;
+                            }
                         } else {
-                            while (nth < ne00 && nth < 1024) {
+                            while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
                                 nth *= 2;
                             }
-                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline;
+                            if (use_f16) {
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline;
+                            } else {
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline;
+                            }
                         }
 
                         float scale;
@@ -1351,8 +1406,8 @@ static enum ggml_status ggml_metal_graph_compute(
                         const int64_t nrows_x = ggml_nrows(src0);
                         const int64_t nrows_y = src0->ne[1];
 
-                        const uint32_t n_head_kv   = nrows_x/nrows_y;
-                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+                        const uint32_t n_head      = nrows_x/nrows_y;
+                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
                         const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
                         const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
@@ -1364,20 +1419,15 @@ static enum ggml_status ggml_metal_graph_compute(
                         } else {
                             [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
                         }
-                        if (id_src2) {
-                            [encoder setBuffer:id_src2 offset:offs_src2   atIndex:2];
-                        } else {
-                            [encoder setBuffer:id_src0 offset:offs_src0   atIndex:2];
-                        }
-                        [encoder setBuffer:id_dst   offset:offs_dst          atIndex:3];
-                        [encoder setBytes:&ne00     length:sizeof(ne00)      atIndex:4];
-                        [encoder setBytes:&ne01     length:sizeof(ne01)      atIndex:5];
-                        [encoder setBytes:&ne02     length:sizeof(ne02)      atIndex:6];
-                        [encoder setBytes:&scale    length:sizeof(scale)     atIndex:7];
-                        [encoder setBytes:&max_bias length:sizeof(max_bias)  atIndex:8];
-                        [encoder setBytes:&m0       length:sizeof(m0)        atIndex:9];
-                        [encoder setBytes:&m1       length:sizeof(m1)        atIndex:10];
-                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:11];
+                        [encoder setBuffer:id_dst      offset:offs_dst            atIndex:2];
+                        [encoder setBytes:&ne00        length:sizeof(ne00)        atIndex:3];
+                        [encoder setBytes:&ne01        length:sizeof(ne01)        atIndex:4];
+                        [encoder setBytes:&ne02        length:sizeof(ne02)        atIndex:5];
+                        [encoder setBytes:&scale       length:sizeof(scale)       atIndex:6];
+                        [encoder setBytes:&max_bias    length:sizeof(max_bias)    atIndex:7];
+                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:8];
+                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:9];
+                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
                         [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                         [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -2182,49 +2232,6 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                     } break;
-                case GGML_OP_ALIBI:
-                    {
-                        GGML_ASSERT((src0t == GGML_TYPE_F32));
-
-                        const int nth = MIN(1024, ne00);
-
-                        //const int n_past = ((int32_t *) dst->op_params)[0];
-                        const int n_head = ((int32_t *) dst->op_params)[1];
-
-                        float max_bias;
-                        memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-                        const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-                        const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ALIBI_F32].pipeline;
-
-                        [encoder setComputePipelineState:pipeline];
-                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                        [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                        [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
-                        [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
-                        [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
-                        [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
-                        [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
-                        [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
-                        [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
-                        [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
-                        [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
-                        [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
-                        [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
-                        [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
-                        [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
-                        [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
-                        [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
-                        [encoder setBytes:&m0   length:sizeof(   float) atIndex:18];
-                        [encoder setBytes:&m1   length:sizeof(   float) atIndex:19];
-                        [encoder setBytes:&n_heads_log2_floor   length:sizeof(int) atIndex:20];
-
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                    } break;
                 case GGML_OP_ROPE:
                     {
                         GGML_ASSERT(ne10 == ne02);
@@ -2346,7 +2353,10 @@ static enum ggml_status ggml_metal_graph_compute(
                     {
                         GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
-                        const int sf = dst->op_params[0];
+                        const float sf0 = (float)ne0/src0->ne[0];
+                        const float sf1 = (float)ne1/src0->ne[1];
+                        const float sf2 = (float)ne2/src0->ne[2];
+                        const float sf3 = (float)ne3/src0->ne[3];
 
                         const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
 
@@ -2369,7 +2379,10 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
                         [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
                         [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
-                        [encoder setBytes:&sf   length:sizeof(sf)   atIndex:18];
+                        [encoder setBytes:&sf0  length:sizeof(sf0)  atIndex:18];
+                        [encoder setBytes:&sf1  length:sizeof(sf1)  atIndex:19];
+                        [encoder setBytes:&sf2  length:sizeof(sf2)  atIndex:20];
+                        [encoder setBytes:&sf3  length:sizeof(sf3)  atIndex:21];
 
                         const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
 
@@ -2503,6 +2516,180 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     } break;
+                case GGML_OP_FLASH_ATTN_EXT:
+                    {
+                        GGML_ASSERT(ne00 % 4  == 0);
+                        GGML_ASSERT(ne11 % 32 == 0);
+
+                        GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                        GGML_ASSERT(ggml_are_same_shape (src1, src2));
+
+                        struct ggml_tensor * src3 = gf->nodes[i]->src[3];
+
+                        size_t offs_src3 = 0;
+
+                        id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
+
+                        GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
+                        GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
+                                "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
+
+                        const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
+                        const uint64_t nb21 = src2 ? src2->nb[1] : 0;
+                        const uint64_t nb22 = src2 ? src2->nb[2] : 0;
+                        const uint64_t nb23 = src2 ? src2->nb[3] : 0;
+
+                        const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
+                      //const int64_t  ne31 = src3 ? src3->ne[1] : 0;
+                        const int64_t  ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
+                        const int64_t  ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33);
+
+                        const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30);
+                        const uint64_t nb31 = src3 ? src3->nb[1] : 0;
+                        const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
+                        const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
+
+                        const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
+
+                        float scale;
+                        float max_bias;
+
+                        memcpy(&scale,    ((int32_t *) dst->op_params) + 0, sizeof(scale));
+                        memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
+
+                        const uint32_t n_head      = src0->ne[2];
+                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+                        const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+                        id<MTLComputePipelineState> pipeline = nil;
+
+                        bool use_vec_kernel = false;
+
+                        if (ne01 >= 4 || (ne00%128 != 0)) {
+                            switch (ne00) {
+                                case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
+                                case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
+                                case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
+                                case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
+                                case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
+                                case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
+                                default:
+                                          {
+                                              GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                              GGML_METAL_LOG_ERROR("add template specialization for this size\n");
+                                              GGML_ASSERT(false && "add template specialization for this size");
+                                          }
+                            }
+                        } else {
+                            use_vec_kernel = true;
+
+                            switch (ne00) {
+                                case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
+                                case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
+                                default:
+                                          {
+                                              GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                              GGML_METAL_LOG_ERROR("add template specialization for this size\n");
+                                              GGML_ASSERT(false && "add template specialization for this size");
+                                          }
+                            }
+                        }
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0     offset:offs_src0           atIndex:0];
+                        [encoder setBuffer:id_src1     offset:offs_src1           atIndex:1];
+                        [encoder setBuffer:id_src2     offset:offs_src2           atIndex:2];
+                        if (id_src3) {
+                            [encoder setBuffer:id_src3     offset:offs_src3           atIndex:3];
+                        } else {
+                            [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
+                        }
+                        [encoder setBuffer:id_dst      offset:offs_dst            atIndex:4];
+                        [encoder setBytes:&ne01        length:sizeof( int64_t)    atIndex:5];
+                        [encoder setBytes:&ne02        length:sizeof( int64_t)    atIndex:6];
+                        [encoder setBytes:&ne03        length:sizeof( int64_t)    atIndex:7];
+                        [encoder setBytes:&nb01        length:sizeof(uint64_t)    atIndex:8];
+                        [encoder setBytes:&nb02        length:sizeof(uint64_t)    atIndex:9];
+                        [encoder setBytes:&nb03        length:sizeof(uint64_t)    atIndex:10];
+                        [encoder setBytes:&ne11        length:sizeof( int64_t)    atIndex:11];
+                        [encoder setBytes:&ne12        length:sizeof( int64_t)    atIndex:12];
+                        [encoder setBytes:&ne13        length:sizeof( int64_t)    atIndex:13];
+                        [encoder setBytes:&nb11        length:sizeof(uint64_t)    atIndex:14];
+                        [encoder setBytes:&nb12        length:sizeof(uint64_t)    atIndex:15];
+                        [encoder setBytes:&nb13        length:sizeof(uint64_t)    atIndex:16];
+                        [encoder setBytes:&nb21        length:sizeof(uint64_t)    atIndex:17];
+                        [encoder setBytes:&nb22        length:sizeof(uint64_t)    atIndex:18];
+                        [encoder setBytes:&nb23        length:sizeof(uint64_t)    atIndex:19];
+                        [encoder setBytes:&nb31        length:sizeof(uint64_t)    atIndex:20];
+                        [encoder setBytes:&ne1         length:sizeof( int64_t)    atIndex:21];
+                        [encoder setBytes:&ne2         length:sizeof( int64_t)    atIndex:22];
+                        [encoder setBytes:&scale       length:sizeof(   float)    atIndex:23];
+                        [encoder setBytes:&max_bias    length:sizeof(   float)    atIndex:24];
+                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:25];
+                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:26];
+                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27];
+
+                        if (!use_vec_kernel) {
+                            // half8x8 kernel
+                            const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
+                            const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
+
+                            GGML_ASSERT(nqptg <= 32);
+                            GGML_ASSERT(nqptg  % 8  == 0);
+                            GGML_ASSERT(ncpsg  % 32 == 0);
+
+                            int64_t nsgmax = 2;
+
+                            while (true) {
+                                const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2);
+                                if (smem > ctx->device.maxThreadgroupMemoryLength) {
+                                    break;
+                                }
+                                nsgmax *= 2;
+                            }
+                            nsgmax /= 2;
+
+                            // simdgroups per threadgroup (a.k.a. warps)
+                            const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
+
+                            const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2);
+
+                            //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
+                            GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
+
+                            [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                        } else {
+                            // half1x4 kernel
+                            const int64_t nqptg = 1;  // queries per threadgroup    !! sync with kernel template arguments !!
+                            const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
+
+                            GGML_ASSERT(nqptg <= 32);
+                            GGML_ASSERT(nqptg  % 1  == 0);
+                            GGML_ASSERT(ncpsg  % 32 == 0);
+
+                            // simdgroups per threadgroup (a.k.a. warps)
+                            const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32));
+
+                            int64_t nsg = 1;
+                            while (nsg <= nsgt) {
+                                nsg *= 2;
+                            }
+                            nsg /= 2;
+
+                            const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2);
+
+                            //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
+                            GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
+                            [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                        }
+                    } break;
                 case GGML_OP_DUP:
                 case GGML_OP_CPY:
                 case GGML_OP_CONT:
@@ -2590,6 +2777,11 @@ static enum ggml_status ggml_metal_graph_compute(
         MTLCommandBufferStatus status = [command_buffer status];
         if (status != MTLCommandBufferStatusCompleted) {
             GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            if (status == MTLCommandBufferStatusError) {
+                NSString * error_code = [command_buffer error].localizedDescription;
+                GGML_METAL_LOG_INFO("error: %s\n", [error_code UTF8String]);
+            }
+
             return GGML_STATUS_FAILED;
         }
     }
@@ -2646,7 +2838,11 @@ GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_
     ggml_backend_metal_free_device();
 
     if (ctx->owned) {
+#if TARGET_OS_OSX
+        vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)ctx->all_data, ctx->all_size);
+#else
         free(ctx->all_data);
+#endif
     }
 
     free(ctx);
@@ -2706,10 +2902,13 @@ GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backe
     UNUSED(buft);
 }
 
-static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
+static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
+#ifndef GGML_METAL_NDEBUG
 #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
     if (@available(macOS 10.12, iOS 16.0, *)) {
-        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
                 device.currentAllocatedSize / 1024.0 / 1024.0,
                 device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
@@ -2719,10 +2918,15 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
             GGML_METAL_LOG_INFO("\n");
         }
     } else {
-        GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
+                device.currentAllocatedSize / 1024.0 / 1024.0);
     }
+#endif
 #endif
     UNUSED(device);
+    UNUSED(size_aligned);
 }
 
 GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -2742,22 +2946,23 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff
     ctx->owned = true;
     ctx->n_buffers = 1;
 
-    ctx->buffers[0].data = ctx->all_data;
-    ctx->buffers[0].size = size;
-    ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
-                    length:size_aligned
-                    options:MTLResourceStorageModeShared
-                    deallocator:nil];
+    if (ctx->all_data != NULL) {
+        ctx->buffers[0].data = ctx->all_data;
+        ctx->buffers[0].size = size;
+        ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
+                        length:size_aligned
+                        options:MTLResourceStorageModeShared
+                        deallocator:nil];
+    }
 
-    if (ctx->buffers[0].metal == nil) {
+    if (ctx->all_data == NULL || ctx->buffers[0].metal == nil) {
         GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
         free(ctx);
         ggml_backend_metal_free_device();
         return NULL;
     }
 
-    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
-    ggml_backend_metal_log_allocated_size(device);
+    //ggml_backend_metal_log_allocated_size(device, size_aligned);
 
     return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
 }
@@ -2844,7 +3049,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
             return false;
         }
 
-        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+        ggml_backend_metal_log_allocated_size(device, size_aligned);
 
         ++ctx->n_buffers;
     } else {
@@ -2867,7 +3072,8 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
                 return false;
             }
 
-            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
+            ggml_backend_metal_log_allocated_size(device, size_step_aligned);
+
             if (i + size_step < size) {
                 GGML_METAL_LOG_INFO("\n");
             }
@@ -2876,8 +3082,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
         }
     }
 
-    ggml_backend_metal_log_allocated_size(device);
-
     return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
 }
 

llama/ggml-metal-embed.metal

@@ -1,8098 +0,0 @@
-#define GGML_COMMON_DECL_METAL
-#define GGML_COMMON_IMPL_METAL
-#ifndef GGML_COMMON_DECL
-
-#if defined(GGML_COMMON_DECL_C)
-#include <stdint.h>
-
-typedef uint16_t ggml_half;
-typedef uint32_t ggml_half2;
-
-#define GGML_COMMON_AGGR
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_METAL)
-#include <metal_stdlib>
-
-typedef half  ggml_half;
-typedef half2 ggml_half2;
-
-#define GGML_COMMON_AGGR
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_CUDA)
-#include <cuda_fp16.h>
-#include <cstdint>
-
-typedef half  ggml_half;
-typedef half2 ggml_half2;
-
-#define GGML_COMMON_AGGR data
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_HIP)
-#include <hip/hip_fp16.h>
-#include <cstdint>
-
-typedef half  ggml_half;
-typedef half2 ggml_half2;
-
-#define GGML_COMMON_AGGR data
-
-#define GGML_COMMON_DECL
-#elif defined(GGML_COMMON_DECL_SYCL)
-#include <sycl/half_type.hpp>
-#include <cstdint>
-
-typedef sycl::half  ggml_half;
-typedef sycl::half2 ggml_half2;
-
-#define GGML_COMMON_AGGR data
-
-#define GGML_COMMON_DECL
-#endif
-
-#if defined(GGML_COMMON_DECL)
-
-#ifndef __cplusplus
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-#endif // __cplusplus
-
-// QK = number of values after dequantization
-// QK_K = super-block size
-
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif // GGML_QKK_64
-
-#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
-// QR = QK / number of values before dequantization
-// QI = number of 32 bit integers before dequantization
-
-#define QI4_0 (QK4_0 / (4 * QR4_0))
-#define QR4_0 2
-
-#define QI4_1 (QK4_1 / (4 * QR4_1))
-#define QR4_1 2
-
-#define QI5_0 (QK5_0 / (4 * QR5_0))
-#define QR5_0 2
-
-#define QI5_1 (QK5_1 / (4 * QR5_1))
-#define QR5_1 2
-
-#define QI8_0 (QK8_0 / (4 * QR8_0))
-#define QR8_0 1
-
-#define QI8_1 (QK8_1 / (4 * QR8_1))
-#define QR8_1 1
-
-#define QI2_K (QK_K / (4*QR2_K))
-#define QR2_K 4
-
-#define QI3_K (QK_K / (4*QR3_K))
-#define QR3_K 4
-
-#define QI4_K (QK_K / (4*QR4_K))
-#define QR4_K 2
-
-#define QI5_K (QK_K / (4*QR5_K))
-#define QR5_K 2
-
-#define QI6_K (QK_K / (4*QR6_K))
-#define QR6_K 2
-
-#define QI2_XXS (QK_K / (4*QR2_XXS))
-#define QR2_XXS 8
-
-#define QI2_XS (QK_K / (4*QR2_XS))
-#define QR2_XS 8
-
-#define QI2_S (QK_K / (4*QR2_S))
-#define QR2_S 8
-
-#define QI3_XXS (QK_K / (4*QR3_XXS))
-#define QR3_XXS 8
-
-#define QI3_XS (QK_K / (4*QR3_XS))
-#define QR3_XS 8
-
-#define QI1_S (QK_K / (4*QR1_S))
-#define QR1_S 8
-
-#define QI4_NL (QK4_NL / (4*QR4_NL))
-#define QR4_NL 2
-
-#if QK_K == 64
-#define QI4_XS QI4_NL
-#define QR4_XS QR4_NL
-#else
-#define QI4_XS (QK_K / (4*QR4_XS))
-#define QR4_XS 8
-#endif
-
-#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
-
-#define QK4_0 32
-typedef struct {
-    ggml_half d;           // delta
-    uint8_t qs[QK4_0 / 2]; // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-typedef struct {
-    union {
-        struct {
-            ggml_half d; // delta
-            ggml_half m; // min
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t qs[QK4_1 / 2]; // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-typedef struct {
-    ggml_half d;           // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-typedef struct {
-    union {
-        struct {
-            ggml_half d; // delta
-            ggml_half m; // min
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2]; // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-typedef struct {
-    ggml_half d;       // delta
-    int8_t  qs[QK8_0]; // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-typedef struct {
-    union {
-        struct {
-            ggml_half d; // delta
-            ggml_half s; // d * sum(qs[i])
-        } GGML_COMMON_AGGR;
-        ggml_half2 ds;
-    };
-    int8_t qs[QK8_1]; // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
-
-//
-// Super-block quantization structures
-//
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elements each
-// Effectively 2.625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    union {
-        struct {
-            ggml_half d;    // super-block scale for quantized scales
-            ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[12];    // scales, quantized with 6 bits
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
-
-// 4-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d[2];     // super-block scales/mins
-    uint8_t scales[2];  // 4-bit block scales/mins
-    uint8_t qs[QK_K/2]; // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    union {
-        struct {
-            ggml_half d;    // super-block scale for quantized scales
-            ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];           // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-// 5-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d;             // super-block scale
-    int8_t  scales[QK_K/16]; // 8-bit block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    union {
-        struct {
-            ggml_half d;    // super-block scale for quantized scales
-            ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
-        ggml_half2 dm;
-    };
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];           // quants, high bit
-    uint8_t qs[QK_K/2];           // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    ggml_half d;             // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-// (Almost) "true" 2-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 2.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_half d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-// 2.3125 bpw quants
-typedef struct {
-    ggml_half d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
-
-// 2.5625 bpw quants
-typedef struct {
-    ggml_half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t scales[QK_K/32];
-} block_iq2_s;
-static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
-
-// (Almost) "true" 3-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 3.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_half d;
-    uint8_t qs[3*QK_K/8];
-} block_iq3_xxs;
-static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
-
-// 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    ggml_half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
-typedef struct {
-    ggml_half d;
-    uint8_t  qs[QK_K/8];
-    uint16_t qh[QK_K/32];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-// 1.75 bpw
-typedef struct {
-    uint8_t  qs[QK_K/8];      // grid index, low 8 bits
-    uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
-    uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
-} block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif
-
-// Used by IQ1_M quants
-typedef union {
-    ggml_half f16;
-    uint16_t  u16;
-} iq1m_scale_t;
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    ggml_half d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
-typedef struct {
-    ggml_half d;
-    uint16_t scales_h;
-    uint8_t  scales_l[QK_K/64];
-    uint8_t  qs[QK_K/2];
-} block_iq4_xs;
-static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
-
-#endif // GGML_COMMON_DECL
-#endif // GGML_COMMON_DECL
-
-////////////////////////////////////////////////////////////////////////////////
-
-#ifndef GGML_COMMON_IMPL
-
-#if defined(GGML_COMMON_IMPL_C)
-#include <stdint.h>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_METAL)
-#include <metal_stdlib>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
-#include <cstdint>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_SYCL)
-
-#include <cstdint>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#endif
-
-#if defined(GGML_COMMON_IMPL)
-
-GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8)
-    1, 2, 4, 8, 16, 32, 64, 128
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
-      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
-    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
-    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
-     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
-    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
-     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
-     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
-    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
-GGML_TABLE_END()
-
-//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
-    0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
-    0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
-    0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
-    0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
-    0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
-    0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
-    0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
-    0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
-    0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
-    0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
-    0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
-    0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
-    0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
-    0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
-    0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
-    0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
-    0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
-    0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
-    0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
-    0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
-    0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
-    0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
-    0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
-    0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
-    0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
-    0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
-    0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
-    0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
-    0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
-    0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
-    0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
-    0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
-GGML_TABLE_END()
-//#endif
-
-
-GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
-    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
-    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
-    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
-    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
-    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
-    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
-    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
-    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
-    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
-    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
-    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
-    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
-    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
-    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
-    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
-    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
-    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
-    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
-    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
-    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
-    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
-    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
-    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
-    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
-    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
-    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
-    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
-    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
-    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
-    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
-    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
-    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
-    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
-    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
-    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
-    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
-    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
-    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
-    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
-    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
-    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
-    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
-    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
-    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
-    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
-    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
-    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
-    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
-    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
-    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
-    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
-    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
-    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
-    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
-    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
-    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
-    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
-    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
-    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
-    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
-    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
-    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
-    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
-    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
-    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
-    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
-    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
-    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
-    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
-    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
-    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
-    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
-    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
-    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
-    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
-    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
-    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
-    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
-    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
-    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
-    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
-    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
-    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
-    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
-    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
-    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
-    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
-    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
-    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
-    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
-    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
-    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
-    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
-    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
-    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
-    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
-    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
-    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
-    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
-    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
-    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
-    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
-    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
-    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
-    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
-    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
-    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
-    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
-    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
-    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
-    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
-    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
-    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
-    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
-    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
-    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
-    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
-    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
-    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
-    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
-    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
-    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
-    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
-    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
-    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
-    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
-    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
-    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
-    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
-    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
-    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
-    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
-    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
-    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
-    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
-    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
-    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
-    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
-    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
-    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
-    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
-    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
-    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
-    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
-    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
-    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
-    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
-    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
-    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
-    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
-    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
-    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
-    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
-    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
-    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
-    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
-    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
-    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
-    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
-    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
-    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
-    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
-    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
-    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
-    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
-    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
-    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
-    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
-    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
-    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
-    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
-    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
-    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
-    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
-    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
-    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
-    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
-    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
-    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
-    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
-    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
-    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
-    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
-    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
-    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
-    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
-    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
-    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
-    0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
-    0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
-    0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
-    0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
-    0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
-    0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
-    0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
-    0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
-    0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
-    0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
-    0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
-    0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
-    0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
-    0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
-    0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
-    0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
-    0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
-    0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
-    0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
-    0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
-    0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
-    0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
-    0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
-    0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
-    0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
-    0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
-    0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
-    0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
-    0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
-    0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
-    0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
-    0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
-    0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
-    0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
-    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
-    0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
-    0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
-    0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
-    0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
-    0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
-    0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
-    0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
-    0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
-    0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
-    0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
-    0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
-    0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
-    0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
-    0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
-    0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
-    0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
-    0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
-    0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
-    0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
-    0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
-    0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
-    0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
-    0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
-    0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
-    0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
-    0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
-    0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
-    0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
-    0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
-    0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
-    0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
-    0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
-    0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
-    0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
-    0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
-    0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
-    0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
-    0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
-    0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
-    0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
-    0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
-    0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
-    0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
-    0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
-    0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
-    0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
-    0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
-    0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
-    0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
-    0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
-    0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
-    0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
-    0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
-    0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
-    0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
-    0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
-    0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
-    0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
-    0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
-    0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
-    0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
-    0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
-    0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
-    0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
-    0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
-    0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
-    0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
-    0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
-    0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
-    0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
-    0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
-    0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
-    0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
-    0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
-    0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
-    0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
-    0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
-    0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
-    0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
-    0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
-    0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
-    0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
-    0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
-    0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
-    0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
-    0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
-    0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
-    0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
-    0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
-    0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
-    0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
-    0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
-    0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
-    0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
-    0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
-    0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
-    0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
-    0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
-    0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
-    0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
-    0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
-    0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
-    0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
-    0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
-    0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
-    0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
-    0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
-    0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
-    0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
-    0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
-    0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
-    0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
-    0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
-    0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
-    0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
-    0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
-    0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
-    0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
-    0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
-    0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
-    0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
-    0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
-    0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
-    0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
-    0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
-    0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
-    0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
-    0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
-    0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
-    0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
-    0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
-    0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
-    0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
-    0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
-    0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
-    0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
-    0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
-    0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
-    0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
-    0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
-    0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
-    0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
-    0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
-    0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
-    0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
-    0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
-    0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
-    0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
-    0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
-    0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
-    0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
-    0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
-    0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
-    0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
-    0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
-    0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
-    0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
-    0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
-    0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
-    0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
-    0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
-    0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
-    0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
-    0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
-    0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
-    0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
-    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
-    0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
-    0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
-    0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
-    0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
-    0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
-    0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
-    0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
-    0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
-    0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
-    0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
-    0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
-    0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
-    0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
-    0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
-    0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
-    0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
-    0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
-    0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
-    0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
-    0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
-    0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
-    0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
-    0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
-    0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
-    0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
-    0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
-    0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
-    0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
-    0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
-    0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
-    0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
-    0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
-    0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
-    0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
-    0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
-    0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
-    0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
-    0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
-    0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
-    0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
-    0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
-    0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
-    0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
-    0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
-    0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
-    0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
-    0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
-    0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256)
-    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
-    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
-    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
-    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
-    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
-    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
-    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
-    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
-    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
-    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
-    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
-    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
-    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
-    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
-    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
-    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
-    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
-    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
-    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
-    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
-    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
-    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
-    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
-    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
-    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
-    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
-    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
-    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
-    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
-    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
-    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
-    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
-    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
-    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
-    0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
-    0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
-    0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
-    0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
-    0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
-    0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
-    0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
-    0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
-    0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
-    0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
-    0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
-    0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
-    0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
-    0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
-    0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
-    0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
-    0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
-    0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
-    0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
-    0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
-    0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
-    0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
-    0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
-    0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
-    0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
-    0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
-    0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
-    0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
-    0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
-    0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
-    0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
-    0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
-    0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
-    0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
-    0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
-    0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
-    0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
-    0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
-    0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
-    0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
-    0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
-    0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
-    0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
-    0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
-    0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
-    0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
-    0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
-    0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
-    0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
-    0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
-    0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
-    0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
-    0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
-    0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
-    0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
-    0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
-    0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
-    0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
-    0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
-    0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
-    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
-    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
-GGML_TABLE_END()
-
-#define NGRID_IQ1S 2048
-#define IQ1S_DELTA 0.125f
-#define IQ1M_DELTA 0.125f
-#if defined(GGML_COMMON_IMPL_C)
-GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
-    0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
-    0xffffffffffff0101, 0xffffffffff00ff00, 0xffffffffff000000, 0xffffffffff01ffff,
-    0xffffffffff01ff01, 0xffffffffff0101ff, 0xffffffffff010101, 0xffffffff00ff0000,
-    0xffffffff0000ff00, 0xffffffff000000ff, 0xffffffff00000001, 0xffffffff00010000,
-    0xffffffff01ffffff, 0xffffffff01ffff01, 0xffffffff01ff01ff, 0xffffffff01ff0101,
-    0xffffffff01000000, 0xffffffff0101ffff, 0xffffffff0101ff01, 0xffffffff010101ff,
-    0xffffffff01010101, 0xffffff00ffff00ff, 0xffffff00ffff0000, 0xffffff00ff00ff00,
-    0xffffff00ff0000ff, 0xffffff00ff000001, 0xffffff00ff000100, 0xffffff00ff000101,
-    0xffffff00ff010000, 0xffffff0000ffff00, 0xffffff0000ff0001, 0xffffff0000ff0100,
-    0xffffff000000ff01, 0xffffff0000000000, 0xffffff0000000101, 0xffffff000001ff00,
-    0xffffff00000100ff, 0xffffff0000010001, 0xffffff00000101ff, 0xffffff0001ff0000,
-    0xffffff000100ff00, 0xffffff00010000ff, 0xffffff0001000001, 0xffffff0001010000,
-    0xffffff01ffffffff, 0xffffff01ffffff01, 0xffffff01ffff01ff, 0xffffff01ffff0101,
-    0xffffff01ff000000, 0xffffff01ff01ffff, 0xffffff01ff01ff01, 0xffffff01ff0101ff,
-    0xffffff01ff010101, 0xffffff0100ff0000, 0xffffff010000ff00, 0xffffff0100000100,
-    0xffffff01000100ff, 0xffffff0100010100, 0xffffff0101ffffff, 0xffffff0101ffff01,
-    0xffffff0101ff01ff, 0xffffff0101ff0101, 0xffffff010100ff00, 0xffffff0101000000,
-    0xffffff0101000100, 0xffffff010101ffff, 0xffffff010101ff01, 0xffffff01010101ff,
-    0xffffff0101010101, 0xffff00ffff00ff00, 0xffff00ffff0000ff, 0xffff00ffff000001,
-    0xffff00ffff010000, 0xffff00ff00ffff00, 0xffff00ff00ff0100, 0xffff00ff00000000,
-    0xffff00ff00000101, 0xffff00ff000100ff, 0xffff00ff00010000, 0xffff00ff0100ff00,
-    0xffff00ff01000100, 0xffff00ff01010000, 0xffff0000ffffff00, 0xffff0000ffff00ff,
-    0xffff0000ffff0000, 0xffff0000ffff0001, 0xffff0000ff000000, 0xffff0000ff0001ff,
-    0xffff0000ff000101, 0xffff0000ff010100, 0xffff000000ffffff, 0xffff000000ff0000,
-    0xffff000000ff0101, 0xffff00000000ffff, 0xffff00000000ff00, 0xffff0000000000ff,
-    0xffff000000000000, 0xffff000000000001, 0xffff000000000100, 0xffff00000001ffff,
-    0xffff00000001ff01, 0xffff000000010000, 0xffff0000000101ff, 0xffff000000010101,
-    0xffff000001ffff00, 0xffff00000100ff00, 0xffff000001000000, 0xffff0000010001ff,
-    0xffff000001000101, 0xffff00000101ff00, 0xffff0000010100ff, 0xffff000001010000,
-    0xffff000001010001, 0xffff000001010100, 0xffff0001ff0000ff, 0xffff0001ff000100,
-    0xffff000100ffff00, 0xffff000100ff00ff, 0xffff00010000ffff, 0xffff00010000ff01,
-    0xffff000100000000, 0xffff0001000001ff, 0xffff00010001ffff, 0xffff00010001ff00,
-    0xffff000100010001, 0xffff000100010100, 0xffff000101ff0000, 0xffff00010100ff00,
-    0xffff0001010000ff, 0xffff000101000100, 0xffff01ffffffffff, 0xffff01ffffffff01,
-    0xffff01ffffff01ff, 0xffff01ffffff0101, 0xffff01ffff000000, 0xffff01ffff01ffff,
-    0xffff01ffff01ff01, 0xffff01ffff0101ff, 0xffff01ffff010101, 0xffff01ff00ff0000,
-    0xffff01ff0000ff00, 0xffff01ff00000001, 0xffff01ff00010000, 0xffff01ff01ffffff,
-    0xffff01ff01ffff01, 0xffff01ff01ff01ff, 0xffff01ff01ff0101, 0xffff01ff01000000,
-    0xffff01ff0101ffff, 0xffff01ff0101ff01, 0xffff01ff010101ff, 0xffff01ff01010101,
-    0xffff0100ffff0000, 0xffff0100ff00ff00, 0xffff0100ff0000ff, 0xffff0100ff000100,
-    0xffff0100ff0100ff, 0xffff0100ff010000, 0xffff010000ffff00, 0xffff01000000ffff,
-    0xffff01000000ff00, 0xffff010000000000, 0xffff01000001ff00, 0xffff0100000100ff,
-    0xffff010000010100, 0xffff01000100ff00, 0xffff0100010000ff, 0xffff010001000001,
-    0xffff010001000100, 0xffff010001010000, 0xffff0101ffffffff, 0xffff0101ffffff01,
-    0xffff0101ffff01ff, 0xffff0101ffff0101, 0xffff0101ff000000, 0xffff0101ff01ffff,
-    0xffff0101ff01ff01, 0xffff0101ff0101ff, 0xffff0101ff010101, 0xffff010100ff0000,
-    0xffff01010000ff00, 0xffff010100000100, 0xffff01010001ff00, 0xffff010100010000,
-    0xffff010101ffffff, 0xffff010101ffff01, 0xffff010101ff0000, 0xffff010101ff01ff,
-    0xffff010101ff0101, 0xffff010101000000, 0xffff01010101ffff, 0xffff01010101ff01,
-    0xffff0101010101ff, 0xffff010101010101, 0xff00ffffff00ffff, 0xff00ffffff00ff00,
-    0xff00ffffff0000ff, 0xff00ffffff000100, 0xff00ffffff0100ff, 0xff00ffffff010000,
-    0xff00ffff00ffff00, 0xff00ffff00ff00ff, 0xff00ffff0000ffff, 0xff00ffff00000000,
-    0xff00ffff000001ff, 0xff00ffff0001ff00, 0xff00ffff000100ff, 0xff00ffff00010000,
-    0xff00ffff00010100, 0xff00ffff0100ff00, 0xff00ffff010000ff, 0xff00ffff01000001,
-    0xff00ffff0101ff00, 0xff00ffff01010000, 0xff00ff00ffffff00, 0xff00ff00ffff00ff,
-    0xff00ff00ffff0001, 0xff00ff00ffff0100, 0xff00ff00ff00ffff, 0xff00ff00ff00ff01,
-    0xff00ff00ff000000, 0xff00ff00ff0001ff, 0xff00ff00ff01ff00, 0xff00ff00ff0100ff,
-    0xff00ff00ff010100, 0xff00ff0000ff0000, 0xff00ff0000ff0101, 0xff00ff000000ffff,
-    0xff00ff000000ff00, 0xff00ff000000ff01, 0xff00ff00000000ff, 0xff00ff0000000000,
-    0xff00ff0000000001, 0xff00ff0000000100, 0xff00ff000001ffff, 0xff00ff0000010000,
-    0xff00ff0001ff00ff, 0xff00ff000100ff01, 0xff00ff0001000000, 0xff00ff000101ff00,
-    0xff00ff00010100ff, 0xff00ff01ff00ff00, 0xff00ff01ff0000ff, 0xff00ff01ff000001,
-    0xff00ff01ff010000, 0xff00ff0100ffffff, 0xff00ff0100ff0001, 0xff00ff0100ff0100,
-    0xff00ff010000ff01, 0xff00ff0100000000, 0xff00ff01000001ff, 0xff00ff0100000101,
-    0xff00ff01000100ff, 0xff00ff0100010001, 0xff00ff0101ff0000, 0xff00ff010100ff00,
-    0xff00ff01010000ff, 0xff00ff0101000001, 0xff00ff0101010000, 0xff0000ffffffff00,
-    0xff0000ffffff0001, 0xff0000ffffff0100, 0xff0000ffff0000ff, 0xff0000ffff000000,
-    0xff0000ffff0001ff, 0xff0000ffff000100, 0xff0000ffff01ff00, 0xff0000ffff010001,
-    0xff0000ff00ffff00, 0xff0000ff00ff0000, 0xff0000ff00ff0001, 0xff0000ff00ff01ff,
-    0xff0000ff00ff0101, 0xff0000ff0000ff00, 0xff0000ff000000ff, 0xff0000ff00000000,
-    0xff0000ff00000001, 0xff0000ff00000100, 0xff0000ff0001ff01, 0xff0000ff00010000,
-    0xff0000ff000101ff, 0xff0000ff01ff00ff, 0xff0000ff01ff0100, 0xff0000ff0100ffff,
-    0xff0000ff010000ff, 0xff0000ff01000000, 0xff0000ff010001ff, 0xff0000ff01000100,
-    0xff0000ff01000101, 0xff0000ff0101ff00, 0xff0000ff010100ff, 0xff0000ff01010000,
-    0xff0000ff01010100, 0xff000000ffffff01, 0xff000000ffff0000, 0xff000000ffff0101,
-    0xff000000ff00ff00, 0xff000000ff0000ff, 0xff000000ff000000, 0xff000000ff000001,
-    0xff000000ff000100, 0xff000000ff01ffff, 0xff000000ff01ff01, 0xff000000ff010000,
-    0xff000000ff0101ff, 0xff000000ff010101, 0xff00000000ffff00, 0xff00000000ff00ff,
-    0xff00000000ff0000, 0xff00000000ff0001, 0xff0000000000ff00, 0xff0000000000ff01,
-    0xff000000000000ff, 0xff00000000000000, 0xff00000000000001, 0xff00000000000100,
-    0xff00000000000101, 0xff0000000001ff00, 0xff000000000100ff, 0xff00000000010000,
-    0xff00000000010001, 0xff00000000010100, 0xff00000001ffffff, 0xff00000001ffff01,
-    0xff00000001ff00ff, 0xff00000001ff0000, 0xff00000001ff01ff, 0xff00000001ff0101,
-    0xff0000000100ffff, 0xff0000000100ff00, 0xff000000010000ff, 0xff00000001000000,
-    0xff00000001000001, 0xff00000001000100, 0xff00000001000101, 0xff0000000101ffff,
-    0xff0000000101ff01, 0xff00000001010000, 0xff000001ffffff00, 0xff000001ffff00ff,
-    0xff000001ffff0000, 0xff000001ffff0001, 0xff000001ff000000, 0xff000001ff000001,
-    0xff000001ff0001ff, 0xff000001ff000101, 0xff000001ff01ff00, 0xff000001ff010001,
-    0xff00000100ffffff, 0xff00000100ffff01, 0xff00000100ff00ff, 0xff00000100ff0000,
-    0xff00000100ff01ff, 0xff00000100ff0101, 0xff0000010000ff00, 0xff00000100000000,
-    0xff00000100000001, 0xff000001000001ff, 0xff00000100000100, 0xff0000010001ff00,
-    0xff000001000100ff, 0xff00000100010000, 0xff000001000101ff, 0xff00000100010100,
-    0xff00000100010101, 0xff00000101ff0001, 0xff00000101ff0101, 0xff0000010100ff01,
-    0xff00000101000000, 0xff000001010100ff, 0xff00000101010100, 0xff0001ffff00ff00,
-    0xff0001ffff000001, 0xff0001ffff010000, 0xff0001ff00ffff00, 0xff0001ff00ff00ff,
-    0xff0001ff00ff0001, 0xff0001ff00ff0100, 0xff0001ff0000ffff, 0xff0001ff00000000,
-    0xff0001ff000001ff, 0xff0001ff00000101, 0xff0001ff0001ffff, 0xff0001ff0001ff00,
-    0xff0001ff000100ff, 0xff0001ff00010001, 0xff0001ff00010100, 0xff0001ff01ff0000,
-    0xff0001ff0100ff00, 0xff0001ff010000ff, 0xff0001ff01010000, 0xff000100ff00ffff,
-    0xff000100ff00ff01, 0xff000100ff000000, 0xff000100ff000101, 0xff000100ff01ff00,
-    0xff000100ff010000, 0xff00010000ffff01, 0xff00010000ff00ff, 0xff00010000ff0000,
-    0xff00010000ff01ff, 0xff0001000000ff00, 0xff000100000000ff, 0xff00010000000000,
-    0xff00010000000001, 0xff00010000000100, 0xff00010000000101, 0xff0001000001ffff,
-    0xff00010000010000, 0xff00010000010101, 0xff00010001ff0100, 0xff0001000100ff00,
-    0xff0001000100ff01, 0xff00010001000000, 0xff000100010001ff, 0xff0001000101ff00,
-    0xff00010001010001, 0xff00010001010100, 0xff000101ffff0100, 0xff000101ff000001,
-    0xff000101ff0100ff, 0xff000101ff010001, 0xff00010100ff00ff, 0xff00010100ff0001,
-    0xff00010100ff0100, 0xff0001010000ffff, 0xff0001010000ff01, 0xff00010100000000,
-    0xff000101000001ff, 0xff0001010001ff00, 0xff00010100010001, 0xff00010100010100,
-    0xff00010101ff0000, 0xff0001010100ff00, 0xff00010101000001, 0xff00010101000101,
-    0xff01ffffffffffff, 0xff01ffffffffff01, 0xff01ffffffff01ff, 0xff01ffffffff0101,
-    0xff01ffffff000000, 0xff01ffffff01ffff, 0xff01ffffff01ff01, 0xff01ffffff010000,
-    0xff01ffffff0101ff, 0xff01ffffff010101, 0xff01ffff00ff0000, 0xff01ffff0000ff00,
-    0xff01ffff00000100, 0xff01ffff0001ff00, 0xff01ffff00010000, 0xff01ffff01ffffff,
-    0xff01ffff01ffff01, 0xff01ffff01ff01ff, 0xff01ffff01ff0101, 0xff01ffff01000000,
-    0xff01ffff0101ffff, 0xff01ffff0101ff01, 0xff01ffff01010000, 0xff01ffff010101ff,
-    0xff01ffff01010101, 0xff01ff00ffff0000, 0xff01ff00ff00ff00, 0xff01ff00ff0000ff,
-    0xff01ff00ff000100, 0xff01ff00ff010000, 0xff01ff0000ffff01, 0xff01ff0000ff00ff,
-    0xff01ff0000ff0100, 0xff01ff0000000000, 0xff01ff00000001ff, 0xff01ff0000000101,
-    0xff01ff000001ff00, 0xff01ff00000100ff, 0xff01ff0000010000, 0xff01ff0000010001,
-    0xff01ff0001ff0000, 0xff01ff000100ffff, 0xff01ff0001000001, 0xff01ff0001000100,
-    0xff01ff0001010000, 0xff01ff01ffffff00, 0xff01ff01ffff01ff, 0xff01ff01ffff0101,
-    0xff01ff01ff00ff00, 0xff01ff01ff000000, 0xff01ff01ff01ffff, 0xff01ff01ff01ff01,
-    0xff01ff01ff0101ff, 0xff01ff01ff010101, 0xff01ff0100ff0000, 0xff01ff010000ff00,
-    0xff01ff0100000001, 0xff01ff0100000100, 0xff01ff0100010000, 0xff01ff0101ffff00,
-    0xff01ff0101ff01ff, 0xff01ff0101ff0101, 0xff01ff010100ff00, 0xff01ff0101000000,
-    0xff01ff010101ffff, 0xff01ff010101ff01, 0xff01ff01010101ff, 0xff01ff0101010101,
-    0xff0100ffffff0000, 0xff0100ffff0000ff, 0xff0100ffff000001, 0xff0100ffff000100,
-    0xff0100ffff010000, 0xff0100ff00ff00ff, 0xff0100ff00ff0000, 0xff0100ff00ff0001,
-    0xff0100ff00ff0100, 0xff0100ff0000ff01, 0xff0100ff00000000, 0xff0100ff000001ff,
-    0xff0100ff00000101, 0xff0100ff00010001, 0xff0100ff01ff0000, 0xff0100ff0100ff00,
-    0xff0100ff010000ff, 0xff0100ff01000100, 0xff0100ff0101ff00, 0xff0100ff01010000,
-    0xff010000ffff0100, 0xff010000ff000000, 0xff010000ff01ff00, 0xff010000ff010100,
-    0xff01000000ffffff, 0xff01000000ff0000, 0xff01000000ff01ff, 0xff0100000000ff00,
-    0xff010000000000ff, 0xff01000000000000, 0xff01000000000100, 0xff0100000001ff01,
-    0xff01000000010000, 0xff010000000101ff, 0xff01000001ff0100, 0xff0100000100ffff,
-    0xff010000010000ff, 0xff01000001000000, 0xff010000010001ff, 0xff01000001000101,
-    0xff0100000101ff00, 0xff010000010100ff, 0xff01000001010001, 0xff01000001010100,
-    0xff010001ffff0000, 0xff010001ff00ffff, 0xff010001ff00ff01, 0xff010001ff000100,
-    0xff010001ff010000, 0xff01000100ffff00, 0xff01000100ff0100, 0xff01000100000000,
-    0xff0100010001ffff, 0xff0100010001ff00, 0xff01000100010100, 0xff01000101ff00ff,
-    0xff01000101ff0001, 0xff0100010100ffff, 0xff01000101000101, 0xff0101ffffffffff,
-    0xff0101ffffffff01, 0xff0101ffffff01ff, 0xff0101ffffff0101, 0xff0101ffff000000,
-    0xff0101ffff01ffff, 0xff0101ffff01ff01, 0xff0101ffff0101ff, 0xff0101ffff010101,
-    0xff0101ff00ff0000, 0xff0101ff0000ff00, 0xff0101ff000000ff, 0xff0101ff00010000,
-    0xff0101ff01ffffff, 0xff0101ff01ffff01, 0xff0101ff01ff01ff, 0xff0101ff01ff0101,
-    0xff0101ff0101ffff, 0xff0101ff0101ff01, 0xff0101ff010101ff, 0xff0101ff01010101,
-    0xff010100ffff0100, 0xff010100ff00ff00, 0xff010100ff0000ff, 0xff010100ff000100,
-    0xff010100ff010000, 0xff01010000ff0001, 0xff01010000ff0100, 0xff0101000000ff01,
-    0xff01010000000000, 0xff0101000001ff00, 0xff010100000100ff, 0xff01010000010001,
-    0xff01010000010100, 0xff01010001ff0000, 0xff0101000100ffff, 0xff01010001000001,
-    0xff01010001000100, 0xff010100010100ff, 0xff01010001010000, 0xff010101ffffffff,
-    0xff010101ffffff01, 0xff010101ffff01ff, 0xff010101ffff0101, 0xff010101ff01ffff,
-    0xff010101ff01ff01, 0xff010101ff0101ff, 0xff010101ff010101, 0xff01010100ff0000,
-    0xff0101010000ff00, 0xff01010100000001, 0xff01010100000100, 0xff01010100010000,
-    0xff01010101ffffff, 0xff01010101ffff01, 0xff01010101ff01ff, 0xff01010101ff0101,
-    0xff01010101000000, 0xff0101010101ffff, 0xff0101010101ff01, 0xff010101010101ff,
-    0xff01010101010101, 0x00ffffffffff0000, 0x00ffffffff00ff00, 0x00ffffffff000001,
-    0x00ffffffff010000, 0x00ffffff00ff0100, 0x00ffffff0000ff01, 0x00ffffff00000000,
-    0x00ffffff000001ff, 0x00ffffff00000101, 0x00ffffff0001ff00, 0x00ffffff000100ff,
-    0x00ffffff00010001, 0x00ffffff010000ff, 0x00ffffff01000100, 0x00ffffff0101ff00,
-    0x00ffffff01010001, 0x00ffff00ffffffff, 0x00ffff00ffffff00, 0x00ffff00ffff00ff,
-    0x00ffff00ffff0001, 0x00ffff00ffff0100, 0x00ffff00ff00ff01, 0x00ffff00ff000000,
-    0x00ffff00ff000001, 0x00ffff00ff0001ff, 0x00ffff00ff000101, 0x00ffff00ff01ff00,
-    0x00ffff00ff010001, 0x00ffff00ff010100, 0x00ffff0000ff0000, 0x00ffff0000ff01ff,
-    0x00ffff0000ff0101, 0x00ffff000000ff00, 0x00ffff00000000ff, 0x00ffff0000000000,
-    0x00ffff0000000001, 0x00ffff0000000100, 0x00ffff0000000101, 0x00ffff0000010000,
-    0x00ffff00000101ff, 0x00ffff0000010101, 0x00ffff0001ffff00, 0x00ffff0001ff00ff,
-    0x00ffff0001ff0001, 0x00ffff000100ffff, 0x00ffff000100ff01, 0x00ffff0001000000,
-    0x00ffff000101ffff, 0x00ffff000101ff00, 0x00ffff000101ff01, 0x00ffff01ffff0000,
-    0x00ffff01ff00ff00, 0x00ffff01ff0000ff, 0x00ffff01ff000001, 0x00ffff01ff010000,
-    0x00ffff0100ffff00, 0x00ffff010000ff01, 0x00ffff0100000000, 0x00ffff0100000101,
-    0x00ffff01000100ff, 0x00ffff0100010100, 0x00ffff0101ff0100, 0x00ffff01010000ff,
-    0x00ffff0101010000, 0x00ff00ffffffff00, 0x00ff00ffff000000, 0x00ff00ffff000100,
-    0x00ff00ffff010100, 0x00ff00ff00ff0000, 0x00ff00ff00ff01ff, 0x00ff00ff00ff0101,
-    0x00ff00ff0000ff00, 0x00ff00ff000000ff, 0x00ff00ff00000000, 0x00ff00ff00000001,
-    0x00ff00ff0001ff00, 0x00ff00ff0001ff01, 0x00ff00ff00010000, 0x00ff00ff000101ff,
-    0x00ff00ff00010101, 0x00ff00ff01ffff00, 0x00ff00ff01ff0001, 0x00ff00ff01ff0100,
-    0x00ff00ff0100ffff, 0x00ff00ff0100ff01, 0x00ff00ff01000000, 0x00ff00ff0101ffff,
-    0x00ff00ff0101ff00, 0x00ff00ff01010100, 0x00ff0000ffffff00, 0x00ff0000ffffff01,
-    0x00ff0000ffff0000, 0x00ff0000ffff0101, 0x00ff0000ff00ff00, 0x00ff0000ff0000ff,
-    0x00ff0000ff000000, 0x00ff0000ff000001, 0x00ff0000ff000100, 0x00ff0000ff01ffff,
-    0x00ff0000ff010000, 0x00ff0000ff010101, 0x00ff000000ffff00, 0x00ff000000ff00ff,
-    0x00ff000000ff0000, 0x00ff000000ff0001, 0x00ff000000ff0100, 0x00ff00000000ffff,
-    0x00ff00000000ff00, 0x00ff0000000000ff, 0x00ff000000000000, 0x00ff000000000001,
-    0x00ff0000000001ff, 0x00ff000000000100, 0x00ff00000001ff00, 0x00ff0000000100ff,
-    0x00ff000000010000, 0x00ff000000010001, 0x00ff000000010100, 0x00ff000001ffff01,
-    0x00ff000001ff00ff, 0x00ff000001ff0000, 0x00ff000001ff01ff, 0x00ff00000100ff00,
-    0x00ff0000010000ff, 0x00ff000001000000, 0x00ff000001000001, 0x00ff000001000100,
-    0x00ff000001000101, 0x00ff000001010000, 0x00ff0000010101ff, 0x00ff000001010101,
-    0x00ff0001ffffff00, 0x00ff0001ffff0000, 0x00ff0001ffff0100, 0x00ff0001ff0000ff,
-    0x00ff0001ff000000, 0x00ff0001ff0001ff, 0x00ff0001ff000101, 0x00ff0001ff01ff00,
-    0x00ff0001ff0100ff, 0x00ff0001ff010100, 0x00ff000100ffffff, 0x00ff000100ffff01,
-    0x00ff000100ff0000, 0x00ff000100ff01ff, 0x00ff00010000ffff, 0x00ff00010000ff00,
-    0x00ff00010000ff01, 0x00ff000100000000, 0x00ff000100000001, 0x00ff000100000100,
-    0x00ff00010001ff01, 0x00ff000100010000, 0x00ff0001000101ff, 0x00ff000101ffff00,
-    0x00ff000101ff0000, 0x00ff000101ff0101, 0x00ff0001010000ff, 0x00ff000101000000,
-    0x00ff00010101ff00, 0x00ff0001010100ff, 0x00ff000101010001, 0x00ff01ffffff0000,
-    0x00ff01ffff00ff00, 0x00ff01ffff000000, 0x00ff01ffff000101, 0x00ff01ffff010000,
-    0x00ff01ff00ffff01, 0x00ff01ff00ff0100, 0x00ff01ff0000ffff, 0x00ff01ff00000000,
-    0x00ff01ff000001ff, 0x00ff01ff0001ff00, 0x00ff01ff000100ff, 0x00ff01ff00010001,
-    0x00ff01ff00010100, 0x00ff01ff01ff0000, 0x00ff01ff0100ff00, 0x00ff01ff010000ff,
-    0x00ff01ff01000001, 0x00ff01ff01000100, 0x00ff01ff01010000, 0x00ff0100ffffff00,
-    0x00ff0100ffff0000, 0x00ff0100ffff0001, 0x00ff0100ffff0101, 0x00ff0100ff00ffff,
-    0x00ff0100ff0000ff, 0x00ff0100ff000000, 0x00ff0100ff0001ff, 0x00ff0100ff01ff00,
-    0x00ff0100ff0100ff, 0x00ff0100ff010001, 0x00ff010000ffffff, 0x00ff010000ff0000,
-    0x00ff010000ff0101, 0x00ff01000000ff00, 0x00ff01000000ff01, 0x00ff0100000000ff,
-    0x00ff010000000000, 0x00ff010000000001, 0x00ff010000000100, 0x00ff01000001ffff,
-    0x00ff01000001ff01, 0x00ff010000010000, 0x00ff010000010001, 0x00ff010000010101,
-    0x00ff010001ff0001, 0x00ff010001ff0100, 0x00ff01000100ff01, 0x00ff010001000000,
-    0x00ff010001000001, 0x00ff0100010001ff, 0x00ff01000101ff00, 0x00ff0100010100ff,
-    0x00ff010001010001, 0x00ff010001010100, 0x00ff0101ff000001, 0x00ff010100ff00ff,
-    0x00ff010100ff0001, 0x00ff010100ff0100, 0x00ff010100000000, 0x00ff0101000001ff,
-    0x00ff010100000101, 0x00ff0101000100ff, 0x00ff010100010100, 0x00ff0101010000ff,
-    0x00ff010101010000, 0x0000ffffffffff00, 0x0000ffffffff00ff, 0x0000ffffffff0000,
-    0x0000ffffffff0001, 0x0000ffffffff0100, 0x0000ffffff00ff01, 0x0000ffffff000000,
-    0x0000ffffff000101, 0x0000ffffff01ff00, 0x0000ffffff0100ff, 0x0000ffffff010100,
-    0x0000ffff00ffffff, 0x0000ffff00ff0000, 0x0000ffff00ff01ff, 0x0000ffff0000ff00,
-    0x0000ffff000000ff, 0x0000ffff00000000, 0x0000ffff00000001, 0x0000ffff00000100,
-    0x0000ffff00010000, 0x0000ffff000101ff, 0x0000ffff01ff0001, 0x0000ffff01ff0100,
-    0x0000ffff01000000, 0x0000ffff010001ff, 0x0000ffff0101ffff, 0x0000ffff0101ff00,
-    0x0000ffff01010001, 0x0000ffff01010100, 0x0000ff00ffff0000, 0x0000ff00ffff01ff,
-    0x0000ff00ffff0100, 0x0000ff00ffff0101, 0x0000ff00ff00ff00, 0x0000ff00ff0000ff,
-    0x0000ff00ff000000, 0x0000ff00ff000001, 0x0000ff00ff0001ff, 0x0000ff00ff000100,
-    0x0000ff00ff01ffff, 0x0000ff00ff010000, 0x0000ff00ff010001, 0x0000ff00ff0101ff,
-    0x0000ff00ff010101, 0x0000ff0000ffff00, 0x0000ff0000ff00ff, 0x0000ff0000ff0000,
-    0x0000ff0000ff0001, 0x0000ff0000ff0100, 0x0000ff000000ffff, 0x0000ff000000ff00,
-    0x0000ff000000ff01, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
-    0x0000ff00000001ff, 0x0000ff0000000100, 0x0000ff0000000101, 0x0000ff000001ff00,
-    0x0000ff00000100ff, 0x0000ff0000010000, 0x0000ff0000010001, 0x0000ff0000010100,
-    0x0000ff0001ffff01, 0x0000ff0001ff0000, 0x0000ff000100ff00, 0x0000ff00010000ff,
-    0x0000ff0001000000, 0x0000ff0001000001, 0x0000ff0001000100, 0x0000ff000101ffff,
-    0x0000ff0001010000, 0x0000ff0001010101, 0x0000ff01ffffff00, 0x0000ff01ffff0001,
-    0x0000ff01ff00ff01, 0x0000ff01ff000000, 0x0000ff01ff000101, 0x0000ff01ff01ff00,
-    0x0000ff01ff0100ff, 0x0000ff0100ffff01, 0x0000ff0100ff0000, 0x0000ff0100ff0101,
-    0x0000ff010000ff00, 0x0000ff01000000ff, 0x0000ff0100000000, 0x0000ff0100000001,
-    0x0000ff0100000100, 0x0000ff010001ff01, 0x0000ff0100010000, 0x0000ff0101ff0000,
-    0x0000ff010100ffff, 0x0000ff010100ff01, 0x0000ff0101000000, 0x0000ff0101000100,
-    0x0000ff0101000101, 0x0000ff01010100ff, 0x000000ffffff00ff, 0x000000ffffff0000,
-    0x000000ffff00ff00, 0x000000ffff0000ff, 0x000000ffff000000, 0x000000ffff000001,
-    0x000000ffff0001ff, 0x000000ffff000100, 0x000000ffff01ff00, 0x000000ffff010000,
-    0x000000ffff0101ff, 0x000000ffff010101, 0x000000ff00ffff00, 0x000000ff00ff00ff,
-    0x000000ff00ff0000, 0x000000ff00ff0001, 0x000000ff00ff0100, 0x000000ff00ff0101,
-    0x000000ff0000ffff, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
-    0x000000ff00000001, 0x000000ff000001ff, 0x000000ff00000100, 0x000000ff00000101,
-    0x000000ff0001ff00, 0x000000ff0001ff01, 0x000000ff000100ff, 0x000000ff00010000,
-    0x000000ff00010001, 0x000000ff00010100, 0x000000ff01ffffff, 0x000000ff01ff01ff,
-    0x000000ff01ff0101, 0x000000ff0100ff00, 0x000000ff010000ff, 0x000000ff01000000,
-    0x000000ff01000001, 0x000000ff01000100, 0x000000ff0101ff00, 0x000000ff010100ff,
-    0x000000ff01010000, 0x000000ff01010101, 0x00000000ffffff00, 0x00000000ffffff01,
-    0x00000000ffff00ff, 0x00000000ffff0000, 0x00000000ffff0001, 0x00000000ffff0100,
-    0x00000000ff00ffff, 0x00000000ff00ff00, 0x00000000ff00ff01, 0x00000000ff0000ff,
-    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff000101,
-    0x00000000ff01ff00, 0x00000000ff0100ff, 0x00000000ff010000, 0x00000000ff010001,
-    0x00000000ff010100, 0x0000000000ffffff, 0x0000000000ffff00, 0x0000000000ffff01,
-    0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, 0x0000000000ff01ff,
-    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
-    0x00000000000000ff, 0x0000000000000000, 0x0000000000000001, 0x00000000000001ff,
-    0x0000000000000100, 0x0000000000000101, 0x000000000001ffff, 0x000000000001ff00,
-    0x00000000000100ff, 0x0000000000010000, 0x0000000000010001, 0x00000000000101ff,
-    0x0000000000010100, 0x0000000000010101, 0x0000000001ffff00, 0x0000000001ff00ff,
-    0x0000000001ff0000, 0x0000000001ff0100, 0x0000000001ff0101, 0x000000000100ffff,
-    0x000000000100ff00, 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001,
-    0x00000000010001ff, 0x0000000001000100, 0x000000000101ff00, 0x00000000010100ff,
-    0x0000000001010000, 0x0000000001010001, 0x0000000001010100, 0x00000001ffffffff,
-    0x00000001ffffff00, 0x00000001ffffff01, 0x00000001ffff00ff, 0x00000001ffff0001,
-    0x00000001ffff01ff, 0x00000001ffff0100, 0x00000001ff00ff00, 0x00000001ff0000ff,
-    0x00000001ff000000, 0x00000001ff0001ff, 0x00000001ff000100, 0x00000001ff01ffff,
-    0x00000001ff01ff00, 0x00000001ff01ff01, 0x00000001ff0100ff, 0x00000001ff010000,
-    0x00000001ff010001, 0x00000001ff0101ff, 0x00000001ff010100, 0x0000000100ffff00,
-    0x0000000100ff0000, 0x0000000100ff0001, 0x0000000100ff01ff, 0x0000000100ff0100,
-    0x0000000100ff0101, 0x000000010000ffff, 0x000000010000ff00, 0x000000010000ff01,
-    0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, 0x00000001000001ff,
-    0x0000000100000100, 0x0000000100000101, 0x000000010001ff00, 0x00000001000100ff,
-    0x0000000100010000, 0x0000000100010100, 0x0000000101ffff01, 0x0000000101ff0000,
-    0x0000000101ff0001, 0x0000000101ff01ff, 0x0000000101ff0100, 0x0000000101ff0101,
-    0x000000010100ff00, 0x0000000101000000, 0x0000000101000101, 0x000000010101ff01,
-    0x0000000101010000, 0x0000000101010001, 0x00000001010101ff, 0x0000000101010100,
-    0x000001ffffff00ff, 0x000001ffffff0000, 0x000001ffffff0001, 0x000001ffffff0100,
-    0x000001ffff00ffff, 0x000001ffff000000, 0x000001ffff0001ff, 0x000001ffff01ff00,
-    0x000001ffff010101, 0x000001ff00ff0000, 0x000001ff00ff01ff, 0x000001ff00ff0101,
-    0x000001ff0000ff00, 0x000001ff000000ff, 0x000001ff00000000, 0x000001ff00000001,
-    0x000001ff000001ff, 0x000001ff00000100, 0x000001ff0001ffff, 0x000001ff0001ff01,
-    0x000001ff000100ff, 0x000001ff00010000, 0x000001ff01ffff01, 0x000001ff01ff0100,
-    0x000001ff0100ffff, 0x000001ff0100ff01, 0x000001ff01000000, 0x000001ff010001ff,
-    0x000001ff0101ff00, 0x000001ff01010100, 0x00000100ffffff00, 0x00000100ffffff01,
-    0x00000100ffff0000, 0x00000100ffff0101, 0x00000100ff00ff00, 0x00000100ff0000ff,
-    0x00000100ff000000, 0x00000100ff000001, 0x00000100ff000100, 0x00000100ff010000,
-    0x0000010000ffff00, 0x0000010000ff00ff, 0x0000010000ff0000, 0x0000010000ff0001,
-    0x0000010000ff0100, 0x000001000000ffff, 0x000001000000ff00, 0x000001000000ff01,
-    0x00000100000000ff, 0x0000010000000000, 0x0000010000000001, 0x00000100000001ff,
-    0x0000010000000100, 0x0000010000000101, 0x000001000001ff00, 0x00000100000100ff,
-    0x0000010000010000, 0x0000010000010001, 0x0000010000010100, 0x0000010001ffff00,
-    0x0000010001ff0000, 0x0000010001ff0100, 0x000001000100ff00, 0x00000100010000ff,
-    0x0000010001000000, 0x0000010001000001, 0x00000100010001ff, 0x0000010001000100,
-    0x0000010001010000, 0x00000101ffff00ff, 0x00000101ffff01ff, 0x00000101ff000000,
-    0x00000101ff000101, 0x00000101ff01ffff, 0x00000101ff010000, 0x00000101ff010001,
-    0x00000101ff010100, 0x0000010100ff0000, 0x0000010100ff01ff, 0x0000010100ff0100,
-    0x000001010000ff00, 0x0000010100000000, 0x0000010100000001, 0x00000101000001ff,
-    0x0000010100000100, 0x000001010001ff01, 0x0000010100010000, 0x00000101000101ff,
-    0x0000010100010101, 0x0000010101ffff00, 0x0000010101ff0101, 0x000001010100ff01,
-    0x0000010101000000, 0x0000010101000001, 0x00000101010001ff, 0x0000010101000101,
-    0x000001010101ff00, 0x0001ffffffff0000, 0x0001ffffff0000ff, 0x0001ffffff000001,
-    0x0001ffffff000100, 0x0001ffffff010000, 0x0001ffff00ff00ff, 0x0001ffff0000ffff,
-    0x0001ffff00000000, 0x0001ffff00000001, 0x0001ffff000001ff, 0x0001ffff00000101,
-    0x0001ffff0001ff00, 0x0001ffff000100ff, 0x0001ffff00010001, 0x0001ffff00010100,
-    0x0001ffff01ffff00, 0x0001ffff01000001, 0x0001ffff01010000, 0x0001ff00ffffff00,
-    0x0001ff00ffff00ff, 0x0001ff00ffff0001, 0x0001ff00ffff0100, 0x0001ff00ff00ff01,
-    0x0001ff00ff000000, 0x0001ff00ff01ff00, 0x0001ff00ff01ff01, 0x0001ff00ff010001,
-    0x0001ff00ff010100, 0x0001ff0000ff0000, 0x0001ff0000ff0100, 0x0001ff000000ff00,
-    0x0001ff0000000000, 0x0001ff0000000001, 0x0001ff0000000100, 0x0001ff0000010000,
-    0x0001ff0000010001, 0x0001ff0000010101, 0x0001ff0001ff00ff, 0x0001ff0001ff0101,
-    0x0001ff000100ff01, 0x0001ff0001000000, 0x0001ff000101ff00, 0x0001ff0001010001,
-    0x0001ff0001010100, 0x0001ff01ff00ff00, 0x0001ff01ff000001, 0x0001ff01ff000100,
-    0x0001ff0100ffffff, 0x0001ff0100ffff00, 0x0001ff0100ff0001, 0x0001ff0100000000,
-    0x0001ff0100000001, 0x0001ff01000001ff, 0x0001ff010001ffff, 0x0001ff0101ff0000,
-    0x0001ff010100ff00, 0x0001ff0101000001, 0x0001ff0101010000, 0x000100ffff00ff00,
-    0x000100ffff00ff01, 0x000100ffff000000, 0x000100ffff000001, 0x000100ffff000101,
-    0x000100ffff01ff00, 0x000100ffff010001, 0x000100ffff010100, 0x000100ff00ffffff,
-    0x000100ff00ffff01, 0x000100ff00ff0000, 0x000100ff00ff01ff, 0x000100ff00ff0101,
-    0x000100ff0000ff00, 0x000100ff000000ff, 0x000100ff00000000, 0x000100ff00000001,
-    0x000100ff00000100, 0x000100ff00000101, 0x000100ff0001ffff, 0x000100ff0001ff01,
-    0x000100ff00010000, 0x000100ff01ff00ff, 0x000100ff01ff0000, 0x000100ff01ff0100,
-    0x000100ff0100ffff, 0x000100ff0100ff01, 0x000100ff010000ff, 0x000100ff01000000,
-    0x000100ff01000001, 0x000100ff010001ff, 0x000100ff01000101, 0x000100ff0101ff00,
-    0x000100ff010100ff, 0x000100ff01010100, 0x00010000ffff0000, 0x00010000ffff01ff,
-    0x00010000ffff0101, 0x00010000ff00ff00, 0x00010000ff000000, 0x00010000ff000001,
-    0x00010000ff000100, 0x0001000000ff00ff, 0x0001000000ff0000, 0x0001000000ff0001,
-    0x0001000000ff0100, 0x000100000000ffff, 0x000100000000ff00, 0x00010000000000ff,
-    0x0001000000000000, 0x0001000000000001, 0x0001000000000100, 0x000100000001ff00,
-    0x00010000000100ff, 0x0001000000010000, 0x0001000000010001, 0x0001000000010100,
-    0x0001000001ff0001, 0x0001000001ff0100, 0x0001000001ff0101, 0x000100000100ff00,
-    0x0001000001000000, 0x0001000001000001, 0x0001000001000100, 0x0001000001000101,
-    0x000100000101ff01, 0x0001000001010000, 0x0001000001010001, 0x00010000010101ff,
-    0x00010001ffffff01, 0x00010001ffff0100, 0x00010001ff000000, 0x00010001ff01ffff,
-    0x00010001ff010001, 0x00010001ff0101ff, 0x00010001ff010100, 0x0001000100ffffff,
-    0x0001000100ff0000, 0x0001000100ff01ff, 0x0001000100ff0101, 0x000100010000ff00,
-    0x00010001000000ff, 0x0001000100000000, 0x0001000100000001, 0x00010001000001ff,
-    0x0001000100000101, 0x000100010001ffff, 0x0001000100010000, 0x00010001000101ff,
-    0x0001000101ffffff, 0x0001000101ffff01, 0x0001000101ff0000, 0x0001000101ff0101,
-    0x00010001010000ff, 0x0001000101000001, 0x00010001010001ff, 0x0001000101000100,
-    0x000100010101ffff, 0x00010001010100ff, 0x0001000101010001, 0x0001000101010101,
-    0x000101ffff000001, 0x000101ffff000100, 0x000101ffff010000, 0x000101ff00ffff00,
-    0x000101ff0000ff01, 0x000101ff00000000, 0x000101ff00000101, 0x000101ff0001ff00,
-    0x000101ff00010100, 0x000101ff01ff0000, 0x000101ff0100ff00, 0x000101ff010001ff,
-    0x000101ff01010001, 0x00010100ffffff00, 0x00010100ffff00ff, 0x00010100ff00ffff,
-    0x00010100ff000000, 0x00010100ff01ff00, 0x00010100ff0100ff, 0x00010100ff010001,
-    0x00010100ff010100, 0x0001010000ffffff, 0x0001010000ffff00, 0x0001010000ff0000,
-    0x0001010000ff0001, 0x0001010000ff01ff, 0x000101000000ff00, 0x00010100000000ff,
-    0x0001010000000000, 0x0001010000000001, 0x0001010000000100, 0x000101000001ffff,
-    0x0001010000010000, 0x0001010000010101, 0x0001010001ffff01, 0x0001010001ff00ff,
-    0x0001010001ff0101, 0x0001010001000000, 0x000101000101ff00, 0x00010100010100ff,
-    0x0001010001010000, 0x0001010001010100, 0x00010101ff00ff00, 0x00010101ff000001,
-    0x00010101ff0001ff, 0x0001010100ffff00, 0x0001010100ff00ff, 0x0001010100ff0100,
-    0x000101010000ffff, 0x0001010100000000, 0x00010101000001ff, 0x0001010100000101,
-    0x00010101000100ff, 0x0001010100010000, 0x0001010100010100, 0x0001010101ff0001,
-    0x00010101010000ff, 0x00010101010001ff, 0x0001010101000101, 0x0001010101010001,
-    0x01ffffffffffffff, 0x01ffffffffffff01, 0x01ffffffffff01ff, 0x01ffffffffff0101,
-    0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, 0x01ffffffff010101,
-    0x01ffffff00ff0000, 0x01ffffff0000ffff, 0x01ffffff0000ff00, 0x01ffffff000000ff,
-    0x01ffffff00000001, 0x01ffffff00000100, 0x01ffffff00010000, 0x01ffffff01ffffff,
-    0x01ffffff01ffff01, 0x01ffffff01ff01ff, 0x01ffffff01ff0101, 0x01ffffff01000000,
-    0x01ffffff0101ffff, 0x01ffffff0101ff01, 0x01ffffff010101ff, 0x01ffffff01010101,
-    0x01ffff00ffff0000, 0x01ffff00ff00ff00, 0x01ffff00ff0000ff, 0x01ffff00ff000001,
-    0x01ffff00ff000100, 0x01ffff00ff010000, 0x01ffff0000ffff00, 0x01ffff0000ff00ff,
-    0x01ffff0000ff0100, 0x01ffff000000ffff, 0x01ffff000000ff01, 0x01ffff0000000000,
-    0x01ffff0000000001, 0x01ffff00000001ff, 0x01ffff0000000100, 0x01ffff00000100ff,
-    0x01ffff0000010001, 0x01ffff0000010100, 0x01ffff0001ff0000, 0x01ffff0001ff0100,
-    0x01ffff00010000ff, 0x01ffff0001000001, 0x01ffff0001000100, 0x01ffff0001010000,
-    0x01ffff01ffffffff, 0x01ffff01ffffff01, 0x01ffff01ffff01ff, 0x01ffff01ffff0101,
-    0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff01ff01, 0x01ffff01ff0101ff,
-    0x01ffff01ff010101, 0x01ffff010000ff00, 0x01ffff01000000ff, 0x01ffff0100000100,
-    0x01ffff0100010000, 0x01ffff0101ffffff, 0x01ffff0101ffff01, 0x01ffff0101ff01ff,
-    0x01ffff0101ff0101, 0x01ffff0101000000, 0x01ffff010101ffff, 0x01ffff010101ff01,
-    0x01ffff01010101ff, 0x01ffff0101010101, 0x01ff00ffff0000ff, 0x01ff00ffff000100,
-    0x01ff00ff00ffff00, 0x01ff00ff00ff00ff, 0x01ff00ff0000ff00, 0x01ff00ff00000000,
-    0x01ff00ff00000101, 0x01ff00ff0001ff00, 0x01ff00ff000100ff, 0x01ff00ff00010100,
-    0x01ff00ff010000ff, 0x01ff00ff01000100, 0x01ff0000ffffff00, 0x01ff0000ffff0100,
-    0x01ff0000ff00ff01, 0x01ff0000ff000000, 0x01ff0000ff000101, 0x01ff0000ff010001,
-    0x01ff0000ff010100, 0x01ff000000ffffff, 0x01ff000000ffff00, 0x01ff000000ff0000,
-    0x01ff000000ff01ff, 0x01ff00000000ff00, 0x01ff0000000000ff, 0x01ff000000000000,
-    0x01ff000000000001, 0x01ff000000000100, 0x01ff000000000101, 0x01ff000000010000,
-    0x01ff000000010001, 0x01ff0000000101ff, 0x01ff000000010101, 0x01ff000001ffff00,
-    0x01ff000001ff00ff, 0x01ff000001ff0001, 0x01ff000001ff0100, 0x01ff00000100ffff,
-    0x01ff00000100ff01, 0x01ff000001000000, 0x01ff0000010001ff, 0x01ff000001010001,
-    0x01ff0001ff00ff00, 0x01ff0001ff000001, 0x01ff0001ff000100, 0x01ff0001ff010000,
-    0x01ff000100ffff00, 0x01ff000100ff00ff, 0x01ff000100ff0100, 0x01ff000100ff0101,
-    0x01ff00010000ffff, 0x01ff000100000000, 0x01ff000100000100, 0x01ff000100000101,
-    0x01ff00010001ff00, 0x01ff000100010001, 0x01ff000100010101, 0x01ff000101ff0000,
-    0x01ff00010100ff00, 0x01ff000101000101, 0x01ff0001010100ff, 0x01ff01ffffffffff,
-    0x01ff01ffffffff01, 0x01ff01ffffff01ff, 0x01ff01ffffff0101, 0x01ff01ffff000000,
-    0x01ff01ffff01ffff, 0x01ff01ffff01ff01, 0x01ff01ffff0101ff, 0x01ff01ffff010101,
-    0x01ff01ff00ffff00, 0x01ff01ff00ff0000, 0x01ff01ff0000ff00, 0x01ff01ff000000ff,
-    0x01ff01ff00000100, 0x01ff01ff00010000, 0x01ff01ff00010100, 0x01ff01ff01ffffff,
-    0x01ff01ff01ffff01, 0x01ff01ff01ff01ff, 0x01ff01ff01ff0101, 0x01ff01ff01000000,
-    0x01ff01ff0101ffff, 0x01ff01ff0101ff01, 0x01ff01ff010101ff, 0x01ff01ff01010101,
-    0x01ff0100ffff0000, 0x01ff0100ffff0001, 0x01ff0100ff00ff00, 0x01ff0100ff0000ff,
-    0x01ff0100ff000001, 0x01ff0100ff010000, 0x01ff010000ffff00, 0x01ff010000ff00ff,
-    0x01ff010000ff0001, 0x01ff010000ff0100, 0x01ff01000000ffff, 0x01ff01000000ff01,
-    0x01ff010000000000, 0x01ff010000000101, 0x01ff01000001ff00, 0x01ff0100000100ff,
-    0x01ff010001ff0000, 0x01ff010001000001, 0x01ff010001000100, 0x01ff010001010000,
-    0x01ff0101ffffffff, 0x01ff0101ffffff01, 0x01ff0101ffff01ff, 0x01ff0101ffff0101,
-    0x01ff0101ff000000, 0x01ff0101ff01ffff, 0x01ff0101ff01ff01, 0x01ff0101ff0101ff,
-    0x01ff0101ff010101, 0x01ff010100ff0000, 0x01ff01010000ff00, 0x01ff0101000000ff,
-    0x01ff010100000001, 0x01ff010101ffffff, 0x01ff010101ffff01, 0x01ff010101ff01ff,
-    0x01ff010101ff0101, 0x01ff010101000000, 0x01ff01010101ffff, 0x01ff01010101ff01,
-    0x01ff0101010101ff, 0x01ff010101010101, 0x0100ffffffff0000, 0x0100ffffff00ff00,
-    0x0100ffffff000001, 0x0100ffffff0001ff, 0x0100ffffff000100, 0x0100ffffff010000,
-    0x0100ffff00ffff00, 0x0100ffff00ff0001, 0x0100ffff00ff0100, 0x0100ffff00000000,
-    0x0100ffff000001ff, 0x0100ffff00000101, 0x0100ffff00010100, 0x0100ffff00010101,
-    0x0100ffff01ff0000, 0x0100ffff0100ff00, 0x0100ffff010000ff, 0x0100ffff01000001,
-    0x0100ffff01000100, 0x0100ffff01010000, 0x0100ff00ffffff00, 0x0100ff00ffff00ff,
-    0x0100ff00ffff0001, 0x0100ff00ffff0100, 0x0100ff00ff00ffff, 0x0100ff00ff000000,
-    0x0100ff00ff0001ff, 0x0100ff00ff000101, 0x0100ff00ff01ff00, 0x0100ff00ff0100ff,
-    0x0100ff00ff010001, 0x0100ff00ff010100, 0x0100ff0000ffffff, 0x0100ff0000ff0000,
-    0x0100ff000000ffff, 0x0100ff000000ff00, 0x0100ff00000000ff, 0x0100ff0000000000,
-    0x0100ff0000000001, 0x0100ff0000000100, 0x0100ff000001ff01, 0x0100ff0000010000,
-    0x0100ff0001ff00ff, 0x0100ff0001ff0001, 0x0100ff000100ff01, 0x0100ff0001000000,
-    0x0100ff00010001ff, 0x0100ff000101ff00, 0x0100ff00010100ff, 0x0100ff0001010001,
-    0x0100ff0001010100, 0x0100ff01ffff0000, 0x0100ff01ff00ff00, 0x0100ff01ff0000ff,
-    0x0100ff01ff000100, 0x0100ff01ff010000, 0x0100ff0100ff00ff, 0x0100ff0100ff0001,
-    0x0100ff0100ff0100, 0x0100ff010000ffff, 0x0100ff010000ff01, 0x0100ff0100000000,
-    0x0100ff01000001ff, 0x0100ff0100010001, 0x0100ff0100010100, 0x0100ff0101ff0000,
-    0x0100ff01010000ff, 0x0100ff0101000001, 0x0100ff0101010100, 0x010000ffffffff00,
-    0x010000ffffff00ff, 0x010000ffffff0001, 0x010000ffff00ffff, 0x010000ffff000000,
-    0x010000ffff0001ff, 0x010000ffff010001, 0x010000ff00ffffff, 0x010000ff00ff0101,
-    0x010000ff0000ff00, 0x010000ff000000ff, 0x010000ff00000000, 0x010000ff00000001,
-    0x010000ff000001ff, 0x010000ff00000100, 0x010000ff0001ffff, 0x010000ff0001ff00,
-    0x010000ff0001ff01, 0x010000ff00010000, 0x010000ff01ff00ff, 0x010000ff01ff0001,
-    0x010000ff0100ff01, 0x010000ff010000ff, 0x010000ff01000000, 0x010000ff010001ff,
-    0x010000ff0101ff00, 0x010000ff01010100, 0x01000000ffffffff, 0x01000000ffff0000,
-    0x01000000ffff01ff, 0x01000000ffff0101, 0x01000000ff00ffff, 0x01000000ff00ff00,
-    0x01000000ff0000ff, 0x01000000ff000000, 0x01000000ff000001, 0x01000000ff000100,
-    0x01000000ff01ff00, 0x01000000ff010000, 0x01000000ff010100, 0x01000000ff010101,
-    0x0100000000ffff00, 0x0100000000ff00ff, 0x0100000000ff0000, 0x0100000000ff0001,
-    0x0100000000ff0100, 0x010000000000ffff, 0x010000000000ff00, 0x010000000000ff01,
-    0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, 0x01000000000001ff,
-    0x0100000000000100, 0x0100000000000101, 0x010000000001ff00, 0x01000000000100ff,
-    0x0100000000010000, 0x0100000000010001, 0x0100000000010100, 0x0100000001ffff00,
-    0x0100000001ff0000, 0x0100000001ff01ff, 0x010000000100ff00, 0x010000000100ff01,
-    0x01000000010000ff, 0x0100000001000000, 0x0100000001000001, 0x0100000001000100,
-    0x0100000001000101, 0x010000000101ffff, 0x010000000101ff01, 0x0100000001010000,
-    0x01000000010101ff, 0x0100000001010101, 0x01000001ffffff00, 0x01000001ffff00ff,
-    0x01000001ff00ffff, 0x01000001ff000000, 0x01000001ff000100, 0x01000001ff01ffff,
-    0x01000001ff010001, 0x01000001ff010100, 0x0100000100ff0000, 0x0100000100ff01ff,
-    0x0100000100ff0100, 0x010000010000ff00, 0x010000010000ff01, 0x0100000100000000,
-    0x0100000100000001, 0x0100000100000100, 0x0100000100010000, 0x01000001000101ff,
-    0x0100000101ffff01, 0x0100000101ff00ff, 0x0100000101ff0100, 0x0100000101ff0101,
-    0x010000010100ff01, 0x01000001010000ff, 0x0100000101000000, 0x01000001010100ff,
-    0x0100000101010001, 0x0100000101010100, 0x010001ffffff0000, 0x010001ffff000001,
-    0x010001ffff000100, 0x010001ffff010000, 0x010001ff00ffff00, 0x010001ff00ff0001,
-    0x010001ff0000ffff, 0x010001ff0000ff01, 0x010001ff00000000, 0x010001ff00000001,
-    0x010001ff00000101, 0x010001ff000100ff, 0x010001ff00010000, 0x010001ff01ff0000,
-    0x010001ff0100ff00, 0x010001ff01000001, 0x010001ff01000100, 0x010001ff01010000,
-    0x01000100ffff00ff, 0x01000100ffff0001, 0x01000100ffff0100, 0x01000100ff00ffff,
-    0x01000100ff00ff01, 0x01000100ff000000, 0x01000100ff0001ff, 0x01000100ff000101,
-    0x01000100ff01ffff, 0x01000100ff01ff00, 0x01000100ff0100ff, 0x01000100ff010001,
-    0x0100010000ffffff, 0x0100010000ffff01, 0x0100010000ff0000, 0x0100010000ff01ff,
-    0x0100010000ff0101, 0x010001000000ff00, 0x01000100000000ff, 0x0100010000000000,
-    0x0100010000000001, 0x0100010000000100, 0x010001000001ff01, 0x0100010000010000,
-    0x0100010000010001, 0x0100010000010101, 0x0100010001ffff00, 0x0100010001ff00ff,
-    0x010001000100ffff, 0x010001000100ff01, 0x0100010001000000, 0x0100010001000101,
-    0x010001000101ff00, 0x0100010001010001, 0x01000101ffff0000, 0x01000101ff000000,
-    0x01000101ff010000, 0x0100010100ff00ff, 0x0100010100ff0001, 0x0100010100ff0100,
-    0x010001010000ffff, 0x0100010100000000, 0x01000101000001ff, 0x010001010001ff00,
-    0x0100010101ff0000, 0x010001010100ff00, 0x01000101010000ff, 0x0100010101000000,
-    0x0100010101000001, 0x0101ffffffffffff, 0x0101ffffffffff01, 0x0101ffffffff01ff,
-    0x0101ffffffff0101, 0x0101ffffff000000, 0x0101ffffff01ffff, 0x0101ffffff01ff01,
-    0x0101ffffff0101ff, 0x0101ffffff010101, 0x0101ffff00ff0000, 0x0101ffff0000ff00,
-    0x0101ffff000000ff, 0x0101ffff00000001, 0x0101ffff00000100, 0x0101ffff01ffffff,
-    0x0101ffff01ffff01, 0x0101ffff01ff01ff, 0x0101ffff01ff0101, 0x0101ffff01000000,
-    0x0101ffff0101ffff, 0x0101ffff0101ff01, 0x0101ffff010101ff, 0x0101ffff01010101,
-    0x0101ff00ffff0000, 0x0101ff00ffff0100, 0x0101ff00ff00ff00, 0x0101ff00ff0000ff,
-    0x0101ff00ff000001, 0x0101ff00ff000100, 0x0101ff00ff000101, 0x0101ff0000ff0001,
-    0x0101ff0000ff0100, 0x0101ff000000ff00, 0x0101ff0000000000, 0x0101ff00000001ff,
-    0x0101ff0000000101, 0x0101ff000001ff00, 0x0101ff00000100ff, 0x0101ff0001ff0000,
-    0x0101ff000100ffff, 0x0101ff000100ff01, 0x0101ff0001000001, 0x0101ff0001000100,
-    0x0101ff01ffffff01, 0x0101ff01ffff01ff, 0x0101ff01ffff0101, 0x0101ff01ff00ffff,
-    0x0101ff01ff000100, 0x0101ff01ff01ff01, 0x0101ff01ff0101ff, 0x0101ff01ff010101,
-    0x0101ff0100ff0000, 0x0101ff010000ff00, 0x0101ff0100000001, 0x0101ff0100000100,
-    0x0101ff0100010000, 0x0101ff0101ffffff, 0x0101ff0101ffff01, 0x0101ff0101ff01ff,
-    0x0101ff0101ff0101, 0x0101ff0101000000, 0x0101ff010101ffff, 0x0101ff010101ff01,
-    0x0101ff01010101ff, 0x0101ff0101010101, 0x010100ffff000100, 0x010100ffff010000,
-    0x010100ff00ffff00, 0x010100ff00ff00ff, 0x010100ff0000ffff, 0x010100ff000000ff,
-    0x010100ff00000000, 0x010100ff000001ff, 0x010100ff00000101, 0x010100ff0001ff00,
-    0x010100ff00010000, 0x010100ff00010001, 0x010100ff000101ff, 0x010100ff00010100,
-    0x010100ff01ff0000, 0x01010000ffff0001, 0x01010000ffff0100, 0x01010000ff00ffff,
-    0x01010000ff00ff01, 0x01010000ff000000, 0x01010000ff0001ff, 0x01010000ff010001,
-    0x01010000ff010100, 0x0101000000ffff01, 0x0101000000ff0000, 0x010100000000ff00,
-    0x01010000000000ff, 0x0101000000000000, 0x0101000000000001, 0x0101000000000100,
-    0x0101000000010000, 0x0101000000010101, 0x0101000001ffff00, 0x0101000001ff00ff,
-    0x0101000001ff0000, 0x0101000001ff0001, 0x0101000001ff0100, 0x010100000100ff01,
-    0x0101000001000000, 0x01010000010001ff, 0x01010001ffff0000, 0x01010001ff00ff00,
-    0x01010001ff000001, 0x01010001ff000101, 0x01010001ff01ff00, 0x01010001ff010000,
-    0x0101000100ff00ff, 0x0101000100ff0001, 0x0101000100ff0101, 0x010100010000ff01,
-    0x0101000100000000, 0x0101000100000001, 0x01010001000001ff, 0x010100010001ffff,
-    0x010100010001ff01, 0x0101000101ff0001, 0x010100010100ffff, 0x0101000101000000,
-    0x0101000101000001, 0x0101000101000100, 0x010100010101ff00, 0x01010001010100ff,
-    0x0101000101010001, 0x010101ffffffffff, 0x010101ffffffff01, 0x010101ffffff01ff,
-    0x010101ffffff0101, 0x010101ffff01ffff, 0x010101ffff01ff01, 0x010101ffff0101ff,
-    0x010101ffff010101, 0x010101ff0000ff00, 0x010101ff000000ff, 0x010101ff00000001,
-    0x010101ff00000100, 0x010101ff01ffffff, 0x010101ff01ffff01, 0x010101ff01ff01ff,
-    0x010101ff01ff0101, 0x010101ff01000000, 0x010101ff0101ffff, 0x010101ff0101ff01,
-    0x010101ff010101ff, 0x010101ff01010101, 0x01010100ffff0000, 0x01010100ff0000ff,
-    0x01010100ff000100, 0x01010100ff01ff00, 0x01010100ff010000, 0x0101010000ffff00,
-    0x010101000000ffff, 0x0101010000000000, 0x0101010000000101, 0x010101000001ff00,
-    0x0101010000010001, 0x0101010000010100, 0x010101000100ffff, 0x0101010001000001,
-    0x01010101ffffffff, 0x01010101ffffff01, 0x01010101ffff01ff, 0x01010101ffff0101,
-    0x01010101ff01ffff, 0x01010101ff01ff01, 0x01010101ff0101ff, 0x01010101ff010101,
-    0x010101010000ff00, 0x01010101000000ff, 0x0101010100000001, 0x0101010101ffffff,
-    0x0101010101ffff01, 0x0101010101ff01ff, 0x0101010101ff0101, 0x0101010101000000,
-    0x010101010101ffff, 0x010101010101ff01, 0x01010101010101ff, 0x0101010101010101,
-GGML_TABLE_END()
-#else
-GGML_TABLE_BEGIN(uint32_t, iq1s_grid_gpu, NGRID_IQ1S)
-    0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
-    0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
-    0x02000000, 0x02000002, 0x02000200, 0x02000202, 0x02010101, 0x02020000, 0x02020002, 0x02020200,
-    0x02020202, 0x00000110, 0x00000111, 0x00010011, 0x00010110, 0x00010112, 0x00010211, 0x00010212,
-    0x00020111, 0x01000011, 0x01000112, 0x01000211, 0x01010012, 0x01010111, 0x01010212, 0x01020011,
-    0x01020110, 0x01020112, 0x01020210, 0x02000111, 0x02010011, 0x02010110, 0x02010112, 0x02020111,
-    0x00000020, 0x00000022, 0x00000220, 0x00000222, 0x00010121, 0x00020020, 0x00020022, 0x00020220,
-    0x00020222, 0x01000121, 0x01010021, 0x01010221, 0x01020120, 0x01020221, 0x02000020, 0x02000022,
-    0x02000220, 0x02000222, 0x02010021, 0x02010121, 0x02010221, 0x02020020, 0x02020022, 0x02020220,
-    0x02020222, 0x00011001, 0x00011100, 0x00011102, 0x00021101, 0x01001001, 0x01001201, 0x01011101,
-    0x01011202, 0x01021100, 0x01021101, 0x02011001, 0x02011201, 0x02021101, 0x00001011, 0x00001110,
-    0x00001111, 0x00001112, 0x00011111, 0x00011210, 0x00011212, 0x00021211, 0x01001010, 0x01001111,
-    0x01001212, 0x01011010, 0x01011011, 0x01011110, 0x01011111, 0x01011112, 0x01011211, 0x01021010,
-    0x01021012, 0x01021111, 0x01021210, 0x01021212, 0x02001011, 0x02011011, 0x02011111, 0x02011210,
-    0x02011212, 0x02021011, 0x02021110, 0x02021111, 0x02021112, 0x02021211, 0x00011120, 0x00011221,
-    0x01001021, 0x01001120, 0x01011020, 0x01011022, 0x01011121, 0x01011220, 0x01021020, 0x01021021,
-    0x01021122, 0x01021221, 0x02001121, 0x02011021, 0x02011120, 0x02011221, 0x00002000, 0x00002002,
-    0x00002200, 0x00002202, 0x00012101, 0x00022000, 0x00022002, 0x00022200, 0x00022202, 0x01002101,
-    0x01012001, 0x01012102, 0x01022101, 0x02002000, 0x02002002, 0x02002200, 0x02002202, 0x02012101,
-    0x02022000, 0x02022002, 0x02022200, 0x02022202, 0x00002111, 0x00012011, 0x00012110, 0x00012211,
-    0x00022110, 0x00022111, 0x01002011, 0x01012010, 0x01012011, 0x01012111, 0x01022011, 0x01022110,
-    0x01022211, 0x02012011, 0x02012110, 0x02012112, 0x02012211, 0x02022111, 0x00002020, 0x00002022,
-    0x00002220, 0x00002222, 0x00012121, 0x00022020, 0x00022022, 0x00022220, 0x00022222, 0x01002121,
-    0x01012021, 0x01012221, 0x01022021, 0x01022121, 0x02002020, 0x02002022, 0x02002121, 0x02002220,
-    0x02002222, 0x02012121, 0x02022020, 0x02022022, 0x02022220, 0x02022222, 0x00110000, 0x00110001,
-    0x00110100, 0x00110201, 0x00120100, 0x00120101, 0x01100001, 0x01100100, 0x01110000, 0x01110101,
-    0x01110200, 0x01120001, 0x01120100, 0x01120101, 0x01120201, 0x02110001, 0x02110100, 0x02110102,
-    0x02120001, 0x02120101, 0x00100011, 0x00100110, 0x00100112, 0x00100211, 0x00110010, 0x00110012,
-    0x00110111, 0x00110210, 0x00120011, 0x00120110, 0x00120211, 0x01100111, 0x01100212, 0x01110010,
-    0x01110011, 0x01110012, 0x01110110, 0x01110111, 0x01110112, 0x01110211, 0x01120010, 0x01120111,
-    0x02100110, 0x02110012, 0x02110111, 0x02120011, 0x02120110, 0x00110021, 0x00110120, 0x00110122,
-    0x00120121, 0x01100020, 0x01100122, 0x01100221, 0x01110022, 0x01110121, 0x01110220, 0x01110222,
-    0x01120120, 0x01120122, 0x02100121, 0x02110021, 0x02110120, 0x02110122, 0x02120121, 0x00101001,
-    0x00101102, 0x00101201, 0x00111100, 0x00111101, 0x00111200, 0x00111201, 0x00121001, 0x00121102,
-    0x01101001, 0x01101101, 0x01101102, 0x01101200, 0x01101202, 0x01111001, 0x01111100, 0x01111101,
-    0x01111102, 0x01111201, 0x01121002, 0x01121101, 0x01121200, 0x02101100, 0x02101201, 0x02111000,
-    0x02111100, 0x02111101, 0x02111200, 0x02111201, 0x02111202, 0x02121001, 0x02121100, 0x02121101,
-    0x02121201, 0x00101012, 0x00101111, 0x00101212, 0x00111011, 0x00111110, 0x00111111, 0x00111112,
-    0x00111211, 0x00121010, 0x00121012, 0x00121111, 0x00121210, 0x00121212, 0x01101011, 0x01101110,
-    0x01101111, 0x01101112, 0x01111011, 0x01111012, 0x01111110, 0x01111111, 0x01111112, 0x01111211,
-    0x01111212, 0x01121011, 0x01121110, 0x01121111, 0x01121112, 0x01121211, 0x02101010, 0x02101012,
-    0x02101110, 0x02101111, 0x02101210, 0x02101212, 0x02111010, 0x02111011, 0x02111110, 0x02111111,
-    0x02111112, 0x02111211, 0x02111212, 0x02121010, 0x02121012, 0x02121111, 0x00101021, 0x00101120,
-    0x00101121, 0x00101122, 0x00111121, 0x00111122, 0x00111220, 0x00111222, 0x00121021, 0x00121122,
-    0x01101020, 0x01101022, 0x01101120, 0x01101121, 0x01101220, 0x01101222, 0x01111021, 0x01111121,
-    0x01111122, 0x01111220, 0x01111221, 0x01121021, 0x01121120, 0x01121121, 0x01121220, 0x01121221,
-    0x01121222, 0x02101122, 0x02101222, 0x02111022, 0x02111121, 0x02121120, 0x02121221, 0x00112001,
-    0x00112102, 0x00122101, 0x01102001, 0x01102100, 0x01102102, 0x01102201, 0x01112000, 0x01112101,
-    0x01112200, 0x01112202, 0x01122000, 0x01122001, 0x01122100, 0x01122102, 0x01122201, 0x02102101,
-    0x02112001, 0x02112100, 0x02122101, 0x00112010, 0x00112012, 0x00112111, 0x00112212, 0x00122011,
-    0x00122111, 0x01102012, 0x01102110, 0x01102111, 0x01102210, 0x01112011, 0x01112110, 0x01112111,
-    0x01112112, 0x01112211, 0x01112212, 0x01122010, 0x01122111, 0x01122212, 0x02102211, 0x02112011,
-    0x02112012, 0x02112111, 0x02112210, 0x02122011, 0x02122112, 0x02122211, 0x00102221, 0x00112122,
-    0x00122120, 0x00122122, 0x01102120, 0x01102122, 0x01102221, 0x01112020, 0x01112022, 0x01112121,
-    0x01112220, 0x01122021, 0x01122122, 0x01122221, 0x02102121, 0x02112021, 0x02112122, 0x02112222,
-    0x00200000, 0x00200002, 0x00200200, 0x00200202, 0x00210101, 0x00220000, 0x00220002, 0x00220101,
-    0x00220200, 0x00220202, 0x01200101, 0x01210001, 0x01210201, 0x01220001, 0x01220101, 0x02200000,
-    0x02200002, 0x02200200, 0x02200202, 0x02210101, 0x02220000, 0x02220002, 0x02220101, 0x02220200,
-    0x02220202, 0x00200111, 0x00210011, 0x00210110, 0x00210211, 0x00220111, 0x01200012, 0x01200110,
-    0x01200211, 0x01210111, 0x01210210, 0x01210212, 0x01220011, 0x01220110, 0x01220111, 0x01220112,
-    0x02200111, 0x02210010, 0x02210112, 0x02210211, 0x02220111, 0x00200021, 0x00200220, 0x00200222,
-    0x00210021, 0x00210121, 0x00220020, 0x00220022, 0x00220220, 0x00220222, 0x01200121, 0x01210021,
-    0x01210122, 0x01210221, 0x01220121, 0x02200021, 0x02200220, 0x02200222, 0x02210021, 0x02210121,
-    0x02220020, 0x02220022, 0x02220220, 0x02220222, 0x00201101, 0x00211100, 0x00211102, 0x00211201,
-    0x00221101, 0x01201100, 0x01201101, 0x01201102, 0x01201201, 0x01211002, 0x01211101, 0x01211200,
-    0x01211202, 0x01221102, 0x02201101, 0x02211001, 0x02211100, 0x02211201, 0x02221001, 0x02221101,
-    0x00201211, 0x00211111, 0x00221011, 0x00221211, 0x01201010, 0x01201111, 0x01201210, 0x01211011,
-    0x01211110, 0x01211111, 0x01211211, 0x01221012, 0x01221111, 0x01221210, 0x02201211, 0x02211010,
-    0x02211110, 0x02211111, 0x02211210, 0x02211212, 0x02221011, 0x02221110, 0x02221112, 0x02221211,
-    0x00201121, 0x00211020, 0x00211022, 0x00211221, 0x00221121, 0x01201021, 0x01201221, 0x01211121,
-    0x01221020, 0x01221021, 0x01221221, 0x02201120, 0x02201122, 0x02211020, 0x02211222, 0x00202000,
-    0x00202002, 0x00202200, 0x00202202, 0x00212101, 0x00222000, 0x00222002, 0x00222200, 0x00222202,
-    0x01202101, 0x01212001, 0x01212100, 0x01222101, 0x02202000, 0x02202002, 0x02202200, 0x02202202,
-    0x02222000, 0x02222002, 0x02222200, 0x02222202, 0x00202211, 0x00212011, 0x00212110, 0x00212211,
-    0x00222111, 0x01202112, 0x01202211, 0x01212012, 0x01212111, 0x01222011, 0x01222110, 0x01222112,
-    0x01222211, 0x02202111, 0x02212010, 0x02212112, 0x02212211, 0x02222110, 0x02222111, 0x00202020,
-    0x00202022, 0x00202220, 0x00202222, 0x00222020, 0x00222022, 0x00222220, 0x00222222, 0x01202121,
-    0x01212021, 0x01212122, 0x01212221, 0x01222121, 0x02202020, 0x02202022, 0x02202220, 0x02202222,
-    0x02212121, 0x02222020, 0x02222022, 0x02222220, 0x02222222, 0x10000101, 0x10010001, 0x10010102,
-    0x10020101, 0x11000201, 0x11010002, 0x11010101, 0x11010200, 0x11010202, 0x11020001, 0x11020100,
-    0x11020102, 0x12010100, 0x12010201, 0x12020001, 0x12020102, 0x10000010, 0x10000011, 0x10000110,
-    0x10000112, 0x10000211, 0x10010012, 0x10010111, 0x10010112, 0x10010210, 0x10010212, 0x10020011,
-    0x10020112, 0x10020211, 0x11000111, 0x11000210, 0x11000212, 0x11010011, 0x11010110, 0x11010111,
-    0x11010112, 0x11010211, 0x11010212, 0x11020111, 0x11020210, 0x11020212, 0x12000011, 0x12000110,
-    0x12000112, 0x12010010, 0x12010012, 0x12010111, 0x12020010, 0x12020011, 0x12020012, 0x10000121,
-    0x10010021, 0x10010120, 0x10010122, 0x10020121, 0x11000021, 0x11010022, 0x11010121, 0x11010222,
-    0x11020120, 0x11020221, 0x12000221, 0x12010120, 0x12020121, 0x10001001, 0x10011101, 0x10011201,
-    0x10021201, 0x11001101, 0x11001200, 0x11001202, 0x11011001, 0x11011100, 0x11011101, 0x11011102,
-    0x11021001, 0x11021002, 0x11021101, 0x11021200, 0x11021202, 0x12001001, 0x12001102, 0x12001201,
-    0x12011000, 0x12011002, 0x12011101, 0x12021000, 0x12021001, 0x12021201, 0x10001011, 0x10001012,
-    0x10001111, 0x10001212, 0x10011011, 0x10011110, 0x10011111, 0x10011112, 0x10011211, 0x10021010,
-    0x10021111, 0x10021212, 0x11001011, 0x11001110, 0x11001111, 0x11001112, 0x11001211, 0x11011010,
-    0x11011011, 0x11011110, 0x11011111, 0x11011112, 0x11011210, 0x11011211, 0x11021011, 0x11021110,
-    0x11021111, 0x11021112, 0x11021211, 0x12001012, 0x12001110, 0x12001111, 0x12001210, 0x12011011,
-    0x12011110, 0x12011111, 0x12011112, 0x12011211, 0x12011212, 0x12021111, 0x12021210, 0x12021212,
-    0x10001021, 0x10001121, 0x10001221, 0x10011120, 0x10011121, 0x10011220, 0x10011222, 0x10021021,
-    0x10021120, 0x10021221, 0x11001020, 0x11001022, 0x11001121, 0x11001220, 0x11011020, 0x11011021,
-    0x11011022, 0x11011121, 0x11011122, 0x11011221, 0x11021022, 0x11021121, 0x11021220, 0x12001021,
-    0x12001121, 0x12001222, 0x12011120, 0x12011121, 0x12021021, 0x12021120, 0x12021122, 0x10002101,
-    0x10012001, 0x10012101, 0x10012202, 0x10022101, 0x11002002, 0x11002201, 0x11012000, 0x11012101,
-    0x11012200, 0x11022001, 0x11022100, 0x11022102, 0x11022201, 0x12002101, 0x12012001, 0x12012100,
-    0x12012102, 0x12012201, 0x12022101, 0x10002011, 0x10002111, 0x10002112, 0x10002212, 0x10012010,
-    0x10012110, 0x10012111, 0x10012210, 0x10022011, 0x10022110, 0x10022112, 0x11002010, 0x11002111,
-    0x11002212, 0x11012011, 0x11012012, 0x11012110, 0x11012111, 0x11012112, 0x11012211, 0x11022010,
-    0x11022012, 0x11022111, 0x11022112, 0x11022212, 0x12002112, 0x12002211, 0x12012012, 0x12012111,
-    0x12012112, 0x12012210, 0x12022011, 0x12022110, 0x12022112, 0x12022211, 0x10012122, 0x11002120,
-    0x11002122, 0x11002221, 0x11012121, 0x11012220, 0x11012222, 0x11022120, 0x11022221, 0x12012120,
-    0x12022121, 0x10100001, 0x10100100, 0x10100101, 0x10100102, 0x10100201, 0x10110002, 0x10110101,
-    0x10110202, 0x10120001, 0x10120100, 0x10120201, 0x11100000, 0x11100101, 0x11100200, 0x11110001,
-    0x11110100, 0x11110101, 0x11110102, 0x11110201, 0x11120101, 0x11120200, 0x12100102, 0x12100201,
-    0x12110101, 0x12110200, 0x12120000, 0x12120001, 0x12120102, 0x12120201, 0x10100111, 0x10100210,
-    0x10100211, 0x10100212, 0x10110011, 0x10110110, 0x10110111, 0x10110112, 0x10110210, 0x10110211,
-    0x10120010, 0x10120111, 0x10120112, 0x10120210, 0x10120212, 0x11100011, 0x11100110, 0x11100111,
-    0x11100112, 0x11100211, 0x11110010, 0x11110011, 0x11110012, 0x11110110, 0x11110111, 0x11110112,
-    0x11110210, 0x11110211, 0x11110212, 0x11120011, 0x11120110, 0x11120111, 0x11120112, 0x11120211,
-    0x12100012, 0x12100111, 0x12110011, 0x12110110, 0x12110111, 0x12110112, 0x12110211, 0x12120010,
-    0x12120111, 0x12120212, 0x10100021, 0x10100122, 0x10110022, 0x10110121, 0x10110222, 0x10120021,
-    0x10120120, 0x11100022, 0x11100121, 0x11100222, 0x11110021, 0x11110120, 0x11110121, 0x11110122,
-    0x11110221, 0x11120022, 0x11120121, 0x12100121, 0x12110020, 0x12110022, 0x12110121, 0x12110221,
-    0x12110222, 0x12120120, 0x10101100, 0x10101101, 0x10111001, 0x10111100, 0x10111101, 0x10111102,
-    0x10111200, 0x10111201, 0x10121001, 0x10121101, 0x10121200, 0x10121202, 0x11101001, 0x11101100,
-    0x11101101, 0x11101102, 0x11101201, 0x11101202, 0x11111000, 0x11111001, 0x11111100, 0x11111101,
-    0x11111102, 0x11111200, 0x11111201, 0x11111202, 0x11121001, 0x11121002, 0x11121100, 0x11121101,
-    0x11121102, 0x11121201, 0x12101000, 0x12101200, 0x12101202, 0x12111001, 0x12111100, 0x12111101,
-    0x12111102, 0x12111201, 0x12121001, 0x12121100, 0x12121101, 0x12121202, 0x10101011, 0x10101012,
-    0x10101110, 0x10101111, 0x10101112, 0x10101211, 0x10111010, 0x10111011, 0x10111012, 0x10111110,
-    0x10111111, 0x10111112, 0x10111211, 0x10111212, 0x10121011, 0x10121110, 0x10121111, 0x10121112,
-    0x10121211, 0x11101010, 0x11101011, 0x11101012, 0x11101110, 0x11101111, 0x11101112, 0x11101210,
-    0x11101211, 0x11111010, 0x11111011, 0x11111012, 0x11111110, 0x11111111, 0x11111112, 0x11111210,
-    0x11111211, 0x11111212, 0x11121010, 0x11121011, 0x11121110, 0x11121111, 0x11121112, 0x11121210,
-    0x11121211, 0x11121212, 0x12101011, 0x12101110, 0x12101111, 0x12101211, 0x12101212, 0x12111010,
-    0x12111011, 0x12111110, 0x12111111, 0x12111112, 0x12111210, 0x12111211, 0x12121011, 0x12121110,
-    0x12121111, 0x12121112, 0x12121211, 0x10101020, 0x10101021, 0x10101022, 0x10101120, 0x10101122,
-    0x10101220, 0x10101221, 0x10111021, 0x10111120, 0x10111121, 0x10111220, 0x10111221, 0x10121020,
-    0x10121021, 0x10121022, 0x10121120, 0x10121121, 0x10121122, 0x10121220, 0x10121221, 0x11101021,
-    0x11101121, 0x11101122, 0x11101220, 0x11101221, 0x11101222, 0x11111020, 0x11111021, 0x11111022,
-    0x11111120, 0x11111121, 0x11111122, 0x11111220, 0x11111221, 0x11111222, 0x11121021, 0x11121120,
-    0x11121121, 0x11121221, 0x12101022, 0x12101121, 0x12101122, 0x12101220, 0x12101221, 0x12101222,
-    0x12111021, 0x12111121, 0x12111222, 0x12121022, 0x12121121, 0x12121122, 0x12121220, 0x12121221,
-    0x10102100, 0x10102101, 0x10102102, 0x10102201, 0x10112000, 0x10112101, 0x10112200, 0x10122001,
-    0x10122202, 0x11102101, 0x11102200, 0x11102202, 0x11112001, 0x11112100, 0x11112101, 0x11112102,
-    0x11112200, 0x11112201, 0x11122000, 0x11122002, 0x11122100, 0x11122101, 0x12102002, 0x12102201,
-    0x12112000, 0x12112002, 0x12112101, 0x12112200, 0x12122001, 0x12122201, 0x10102011, 0x10102012,
-    0x10102111, 0x10102212, 0x10112011, 0x10112110, 0x10112111, 0x10112112, 0x10112211, 0x10122111,
-    0x11102011, 0x11102110, 0x11102111, 0x11102112, 0x11102211, 0x11112010, 0x11112011, 0x11112012,
-    0x11112110, 0x11112111, 0x11112112, 0x11112210, 0x11112211, 0x11112212, 0x11122011, 0x11122110,
-    0x11122111, 0x11122112, 0x11122211, 0x12102011, 0x12102111, 0x12102211, 0x12112011, 0x12112110,
-    0x12112111, 0x12112112, 0x12112210, 0x12112211, 0x12122111, 0x10102120, 0x10102220, 0x10112121,
-    0x10112222, 0x10122020, 0x10122121, 0x10122122, 0x10122221, 0x11102121, 0x11102220, 0x11102221,
-    0x11112021, 0x11112121, 0x11112122, 0x11112220, 0x11112221, 0x11122022, 0x11122121, 0x11122220,
-    0x11122222, 0x12102021, 0x12102222, 0x12112022, 0x12112121, 0x12112122, 0x12112220, 0x12112222,
-    0x12122021, 0x10200101, 0x10210100, 0x10210102, 0x10210201, 0x10220101, 0x11200100, 0x11210000,
-    0x11210101, 0x11210102, 0x11210200, 0x11210202, 0x11220001, 0x11220100, 0x11220102, 0x11220201,
-    0x12200001, 0x12210102, 0x12220101, 0x10200011, 0x10200110, 0x10200112, 0x10200211, 0x10210012,
-    0x10210111, 0x10220011, 0x10220012, 0x10220112, 0x10220211, 0x11200111, 0x11200211, 0x11210011,
-    0x11210111, 0x11210112, 0x11210211, 0x11220111, 0x11220112, 0x11220212, 0x12200110, 0x12200212,
-    0x12210012, 0x12210111, 0x12220011, 0x12220112, 0x12220211, 0x10210021, 0x10210122, 0x10210221,
-    0x11200020, 0x11200021, 0x11200122, 0x11210121, 0x11210122, 0x11210220, 0x11220020, 0x12200121,
-    0x12210021, 0x12210122, 0x12220121, 0x10211001, 0x10211002, 0x10211101, 0x10211102, 0x10211202,
-    0x10221001, 0x10221102, 0x10221201, 0x11201000, 0x11201002, 0x11201101, 0x11201200, 0x11201202,
-    0x11211001, 0x11211100, 0x11211101, 0x11211102, 0x11211201, 0x11211202, 0x11221000, 0x11221002,
-    0x11221101, 0x12201100, 0x12201101, 0x12201201, 0x12211000, 0x12211002, 0x12211100, 0x12211101,
-    0x12211102, 0x12211200, 0x12211202, 0x12221001, 0x12221100, 0x12221201, 0x10201111, 0x10201210,
-    0x10201212, 0x10211011, 0x10211111, 0x10211112, 0x10211211, 0x11201110, 0x11201111, 0x11201112,
-    0x11201211, 0x11211010, 0x11211011, 0x11211110, 0x11211111, 0x11211112, 0x11211211, 0x11221011,
-    0x11221110, 0x11221111, 0x11221112, 0x11221211, 0x12201112, 0x12201211, 0x12201212, 0x12211011,
-    0x12211111, 0x12211112, 0x12211211, 0x12211212, 0x12221012, 0x12221111, 0x12221112, 0x12221210,
-    0x10201022, 0x10201221, 0x10211121, 0x10221020, 0x10221122, 0x10221220, 0x10221221, 0x11201020,
-    0x11201121, 0x11201220, 0x11201222, 0x11211021, 0x11211120, 0x11211121, 0x11211122, 0x11211220,
-    0x11211222, 0x11221020, 0x11221121, 0x11221220, 0x12201020, 0x12201022, 0x12201121, 0x12201222,
-    0x12211120, 0x12211122, 0x12211220, 0x12211221, 0x12221020, 0x12221120, 0x12221122, 0x12221222,
-    0x10212102, 0x10212201, 0x10222101, 0x11202001, 0x11212002, 0x11212101, 0x11212202, 0x11222001,
-    0x11222201, 0x12202101, 0x12212001, 0x12212200, 0x12222102, 0x10202011, 0x10202110, 0x10212010,
-    0x10212111, 0x10222011, 0x10222110, 0x10222112, 0x10222211, 0x11202010, 0x11202011, 0x11202111,
-    0x11202112, 0x11202210, 0x11212011, 0x11212110, 0x11212111, 0x11212112, 0x11212211, 0x11222010,
-    0x11222111, 0x11222212, 0x12202012, 0x12202110, 0x12202212, 0x12212111, 0x12222011, 0x12222110,
-    0x12222111, 0x12222211, 0x10212021, 0x10212122, 0x10212220, 0x11202021, 0x11202120, 0x11202221,
-    0x11212020, 0x11212121, 0x11212220, 0x11212222, 0x11222120, 0x11222121, 0x11222221, 0x12202122,
-    0x12212120, 0x12212220, 0x12212222, 0x12222122, 0x20000000, 0x20000002, 0x20000200, 0x20000202,
-    0x20020000, 0x20020002, 0x20020200, 0x20020202, 0x21000101, 0x21010000, 0x21010001, 0x21010100,
-    0x21010102, 0x21010201, 0x21020101, 0x22000000, 0x22000002, 0x22000200, 0x22000202, 0x22010101,
-    0x22020000, 0x22020002, 0x22020200, 0x22020202, 0x20000111, 0x20010011, 0x20010110, 0x20010112,
-    0x20010211, 0x20020111, 0x21000011, 0x21000110, 0x21000211, 0x21010010, 0x21010012, 0x21010111,
-    0x21010112, 0x21010210, 0x21010211, 0x21020110, 0x21020112, 0x21020211, 0x22000111, 0x22000211,
-    0x22010110, 0x22010112, 0x22010211, 0x22020111, 0x20000020, 0x20000022, 0x20000220, 0x20000222,
-    0x20010121, 0x20020020, 0x20020022, 0x20020220, 0x20020222, 0x21010021, 0x21010120, 0x21010221,
-    0x21020121, 0x22000020, 0x22000022, 0x22000220, 0x22000222, 0x22010121, 0x22020020, 0x22020022,
-    0x22020220, 0x22020222, 0x20011100, 0x20011201, 0x21001001, 0x21001100, 0x21011001, 0x21011101,
-    0x21011202, 0x21021001, 0x21021100, 0x21021201, 0x22011100, 0x22011201, 0x20001011, 0x20001211,
-    0x20011012, 0x20011111, 0x20011212, 0x20021112, 0x20021211, 0x21001010, 0x21001011, 0x21001111,
-    0x21001210, 0x21011011, 0x21011110, 0x21011111, 0x21011112, 0x21011211, 0x21011212, 0x21021111,
-    0x21021112, 0x21021210, 0x21021212, 0x22001011, 0x22001110, 0x22001112, 0x22001211, 0x22011010,
-    0x22011012, 0x22011111, 0x22011210, 0x22021112, 0x20011021, 0x20011122, 0x20011221, 0x20021121,
-    0x21001021, 0x21001120, 0x21001221, 0x21001222, 0x21011020, 0x21011121, 0x21011221, 0x21011222,
-    0x21021021, 0x21021122, 0x21021222, 0x22001121, 0x22011021, 0x22011222, 0x22021120, 0x20002000,
-    0x20002002, 0x20002200, 0x20002202, 0x20012101, 0x20022000, 0x20022002, 0x20022200, 0x20022202,
-    0x21002001, 0x21002101, 0x21012001, 0x21012100, 0x21012201, 0x21022101, 0x21022201, 0x22002000,
-    0x22002002, 0x22002200, 0x22002202, 0x22012101, 0x22022000, 0x22022002, 0x22022200, 0x22022202,
-    0x20002111, 0x20002112, 0x20012011, 0x20012110, 0x20012112, 0x20022111, 0x21002011, 0x21002110,
-    0x21002112, 0x21002211, 0x21012010, 0x21012012, 0x21012111, 0x21012212, 0x21022011, 0x21022110,
-    0x22002111, 0x22012112, 0x22012211, 0x22022111, 0x20002020, 0x20002022, 0x20002220, 0x20002222,
-    0x20012121, 0x20022020, 0x20022022, 0x20022220, 0x20022222, 0x21002121, 0x21012021, 0x21012120,
-    0x21012122, 0x22002020, 0x22002022, 0x22002220, 0x22002222, 0x22012121, 0x22022020, 0x22022022,
-    0x22022220, 0x22022222, 0x20100101, 0x20110001, 0x20110102, 0x20110200, 0x20110201, 0x20120101,
-    0x21100001, 0x21100102, 0x21100201, 0x21110101, 0x21110200, 0x21110202, 0x21120201, 0x21120202,
-    0x22100101, 0x22110001, 0x22110100, 0x22110102, 0x22110201, 0x22120101, 0x20100011, 0x20100110,
-    0x20100112, 0x20100211, 0x20110010, 0x20110111, 0x20110210, 0x20110212, 0x20120011, 0x20120110,
-    0x20120112, 0x20120211, 0x21100010, 0x21100111, 0x21110010, 0x21110011, 0x21110110, 0x21110111,
-    0x21110112, 0x21110211, 0x21120012, 0x21120111, 0x22100110, 0x22100112, 0x22110012, 0x22110111,
-    0x22110210, 0x22120011, 0x22120110, 0x22120112, 0x22120211, 0x20100121, 0x20110021, 0x20110120,
-    0x20110221, 0x20120121, 0x21100120, 0x21100122, 0x21100221, 0x21110020, 0x21110022, 0x21110121,
-    0x21110220, 0x21120122, 0x21120221, 0x22100121, 0x22110120, 0x22110122, 0x22120221, 0x20101001,
-    0x20101100, 0x20101102, 0x20111000, 0x20111101, 0x20111200, 0x20121102, 0x21101000, 0x21101202,
-    0x21111001, 0x21111100, 0x21111101, 0x21111102, 0x21111200, 0x21111201, 0x21121000, 0x21121001,
-    0x21121002, 0x21121101, 0x22101100, 0x22101102, 0x22111002, 0x22111100, 0x22111101, 0x22111200,
-    0x22121001, 0x22121201, 0x20101010, 0x20101111, 0x20101210, 0x20101212, 0x20111010, 0x20111011,
-    0x20111110, 0x20111111, 0x20111112, 0x20111211, 0x20121011, 0x20121111, 0x20121211, 0x20121212,
-    0x21101011, 0x21101110, 0x21101111, 0x21101112, 0x21101211, 0x21111010, 0x21111011, 0x21111012,
-    0x21111110, 0x21111111, 0x21111112, 0x21111210, 0x21111211, 0x21111212, 0x21121011, 0x21121110,
-    0x21121111, 0x21121112, 0x21121211, 0x22101011, 0x22101111, 0x22101210, 0x22111011, 0x22111012,
-    0x22111110, 0x22111111, 0x22111112, 0x22111211, 0x22111212, 0x22121010, 0x22121012, 0x22121111,
-    0x22121210, 0x22121212, 0x20101021, 0x20101120, 0x20111020, 0x20111121, 0x20111221, 0x20121020,
-    0x20121122, 0x20121221, 0x21101121, 0x21101220, 0x21101221, 0x21111021, 0x21111022, 0x21111121,
-    0x21111122, 0x21111221, 0x21121121, 0x21121220, 0x22101022, 0x22101120, 0x22101221, 0x22101222,
-    0x22111022, 0x22111120, 0x22111121, 0x22121120, 0x22121122, 0x22121221, 0x20102101, 0x20112102,
-    0x20112201, 0x20122101, 0x21102001, 0x21102102, 0x21112000, 0x21112002, 0x21112101, 0x21112102,
-    0x21112202, 0x21122100, 0x21122101, 0x22102101, 0x22112001, 0x22112102, 0x22112201, 0x22122101,
-    0x20102110, 0x20102112, 0x20102211, 0x20112010, 0x20112012, 0x20112111, 0x20112210, 0x20112212,
-    0x20122010, 0x20122011, 0x20122110, 0x20122112, 0x21102010, 0x21102012, 0x21102111, 0x21102210,
-    0x21102212, 0x21112011, 0x21112110, 0x21112111, 0x21112112, 0x21112211, 0x21122012, 0x21122111,
-    0x21122112, 0x21122212, 0x22102011, 0x22102110, 0x22112010, 0x22112012, 0x22112111, 0x22112212,
-    0x22122011, 0x22122112, 0x20102121, 0x20112121, 0x20122121, 0x21102120, 0x21102122, 0x21102221,
-    0x21112020, 0x21112121, 0x21112220, 0x21122021, 0x22102121, 0x22112021, 0x22112120, 0x22112121,
-    0x22112122, 0x20200000, 0x20200002, 0x20200200, 0x20200202, 0x20210101, 0x20220000, 0x20220002,
-    0x20220200, 0x20220202, 0x21200101, 0x21210001, 0x21210100, 0x21210102, 0x21210201, 0x22200000,
-    0x22200002, 0x22200200, 0x22200202, 0x22210101, 0x22220000, 0x22220002, 0x22220200, 0x22220202,
-    0x20200111, 0x20200211, 0x20210011, 0x20210110, 0x20210112, 0x20210211, 0x20210212, 0x21200112,
-    0x21200211, 0x21210011, 0x21210111, 0x21210210, 0x21210212, 0x21220011, 0x21220110, 0x22200111,
-    0x22210010, 0x22210012, 0x22210112, 0x22210211, 0x20200022, 0x20200220, 0x20200222, 0x20210020,
-    0x20210221, 0x20220022, 0x20220220, 0x20220222, 0x21200121, 0x21210021, 0x21210122, 0x21210221,
-    0x21220121, 0x22200020, 0x22200022, 0x22200220, 0x22200222, 0x22210121, 0x22220020, 0x22220022,
-    0x22220220, 0x22220222, 0x20211201, 0x20221101, 0x21201001, 0x21201100, 0x21211000, 0x21211100,
-    0x21211101, 0x21211200, 0x21211202, 0x21221001, 0x21221101, 0x21221102, 0x21221200, 0x21221201,
-    0x22201101, 0x20201112, 0x20201211, 0x20211010, 0x20211012, 0x20211111, 0x20211210, 0x20221112,
-    0x20221211, 0x21201012, 0x21201111, 0x21211011, 0x21211110, 0x21211111, 0x21211112, 0x21211211,
-    0x21221111, 0x21221212, 0x22201011, 0x22201110, 0x22201111, 0x22201112, 0x22201211, 0x22211012,
-    0x22211111, 0x22211210, 0x20201121, 0x20211021, 0x20211122, 0x20211222, 0x20221021, 0x20221121,
-    0x21201120, 0x21201122, 0x21201222, 0x21211022, 0x21211121, 0x21211122, 0x21211220, 0x21221020,
-    0x21221022, 0x22201122, 0x22211020, 0x22211121, 0x22211122, 0x22211221, 0x22221021, 0x22221120,
-    0x22221122, 0x20202000, 0x20202002, 0x20202200, 0x20202202, 0x20222000, 0x20222002, 0x20222200,
-    0x20222202, 0x21212001, 0x21212100, 0x21212102, 0x21212201, 0x22202000, 0x22202002, 0x22202200,
-    0x22202202, 0x22212101, 0x22222000, 0x22222002, 0x22222200, 0x22222202, 0x20202111, 0x20212110,
-    0x20212211, 0x20222011, 0x20222111, 0x21202011, 0x21212010, 0x21212111, 0x21212212, 0x21222011,
-    0x21222112, 0x21222211, 0x22212010, 0x22212112, 0x20202020, 0x20202022, 0x20202220, 0x20202222,
-    0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
-    0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
-GGML_TABLE_END()
-#endif
-
-#endif // GGML_COMMON_IMPL
-#endif // GGML_COMMON_IMPL
-
-#include <metal_stdlib>
-
-using namespace metal;
-
-#define MAX(x, y) ((x) > (y) ? (x) : (y))
-#define MIN(x, y) ((x) < (y) ? (x) : (y))
-#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
-
-#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
-
-enum ggml_sort_order {
-    GGML_SORT_ORDER_ASC,
-    GGML_SORT_ORDER_DESC,
-};
-
-// general-purpose kernel for addition, multiplication and division of two tensors
-// pros: works for non-contiguous tensors, supports broadcast across all dims
-// cons: not very efficient
-kernel void kernel_add(
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        constant  int64_t & offs,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
-
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
-
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) + *((device float *)(src1_ptr + i10*nb10));
-    }
-}
-
-kernel void kernel_mul(
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
-
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
-
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) * *((device float *)(src1_ptr + i10*nb10));
-    }
-}
-
-kernel void kernel_div(
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
-
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
-
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) / *((device float *)(src1_ptr + i10*nb10));
-    }
-}
-
-// assumption: src1 is a row
-// broadcast src1 into src0
-kernel void kernel_add_row(
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
-        constant   uint64_t & nb [[buffer(28)]],
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] + src1[tpig % nb];
-}
-
-kernel void kernel_mul_row(
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
-        constant   uint64_t & nb  [[buffer(28)]],
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * src1[tpig % nb];
-}
-
-kernel void kernel_div_row(
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
-        constant   uint64_t & nb  [[buffer(28)]],
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] / src1[tpig % nb];
-}
-
-kernel void kernel_scale(
-        device const float * src0,
-        device       float * dst,
-        constant     float & scale,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * scale;
-}
-
-kernel void kernel_scale_4(
-        device const float4 * src0,
-        device       float4 * dst,
-        constant     float  & scale,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * scale;
-}
-
-kernel void kernel_clamp(
-        device const float * src0,
-        device       float * dst,
-        constant     float & min,
-        constant     float & max,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] < min ? min : (src0[tpig] > max ? max : src0[tpig]);
-}
-
-kernel void kernel_relu(
-        device const float * src0,
-        device       float * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = max(0.0f, src0[tpig]);
-}
-
-kernel void kernel_tanh(
-        device const float * src0,
-        device       float * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    device const float & x = src0[tpig];
-    dst[tpig] = precise::tanh(x);
-}
-
-constant float GELU_COEF_A     = 0.044715f;
-constant float GELU_QUICK_COEF = -1.702f;
-constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
-
-kernel void kernel_gelu(
-    device const float * src0,
-    device       float * dst,
-    uint tpig[[thread_position_in_grid]]) {
-    device const float & x = src0[tpig];
-
-    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
-}
-
-kernel void kernel_gelu_4(
-    device const float4 * src0,
-    device       float4 * dst,
-    uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-
-    // BEWARE !!!
-    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
-    // This was observed with Falcon 7B and 40B models
-    //
-    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
-}
-
-kernel void kernel_gelu_quick(
-    device const float * src0,
-    device       float * dst,
-    uint tpig[[thread_position_in_grid]]) {
-    device const float & x = src0[tpig];
-
-    dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
-}
-
-kernel void kernel_gelu_quick_4(
-    device const float4 * src0,
-    device       float4 * dst,
-    uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-
-    dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
-}
-
-kernel void kernel_silu(
-        device const float * src0,
-        device       float * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    device const float & x = src0[tpig];
-    dst[tpig] = x / (1.0f + exp(-x));
-}
-
-kernel void kernel_silu_4(
-        device const float4 * src0,
-        device       float4 * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-    dst[tpig] = x / (1.0f + exp(-x));
-}
-
-kernel void kernel_sqr(
-        device const float * src0,
-        device       float * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * src0[tpig];
-}
-
-kernel void kernel_sum_rows(
-        device const float * src0,
-        device       float * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tpig[[thread_position_in_grid]]) {
-    int64_t i3 = tpig.z;
-    int64_t i2 = tpig.y;
-    int64_t i1 = tpig.x;
-
-    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
-        return;
-    }
-
-    device const float * src_row = (device const float *) ((device const char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
-    device       float * dst_row = (device       float *) ((device       char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
-
-    float row_sum = 0;
-
-    for (int64_t i0 = 0; i0 < ne00; i0++) {
-        row_sum += src_row[i0];
-    }
-
-    dst_row[0] = row_sum;
-}
-
-kernel void kernel_soft_max(
-        device const float * src0,
-        device const float * src1,
-        device const float * src2,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint32_t & n_head_log2,
-        threadgroup  float * buf [[threadgroup(0)]],
-        uint  tgpig[[threadgroup_position_in_grid]],
-        uint  tpitg[[thread_position_in_threadgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint    ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = (tgpig) / (ne02*ne01);
-    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
-    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
-
-    device const float * psrc0 =         src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-    device const float * pmask = src1 != src0 ? src1                               + i01*ne00 : nullptr;
-    device const float * ppos  = src2 != src0 ? src2                                          : nullptr;
-    device       float * pdst  =         dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    float slope = 0.0f;
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int64_t h = i02;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slope = pow(base, exp);
-    }
-
-    // parallel max
-    float lmax = -INFINITY;
-
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
-    }
-
-    // find the max value in the block
-    float max_val = simd_max(lmax);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = -INFINITY;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = max_val;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        max_val = buf[tiisg];
-        max_val = simd_max(max_val);
-    }
-
-    // parallel sum
-    float lsum = 0.0f;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
-        lsum += exp_psrc0;
-        pdst[i00] = exp_psrc0;
-    }
-
-    // This barrier fixes a failing test
-    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
-    threadgroup_barrier(mem_flags::mem_none);
-
-    float sum = simd_sum(lsum);
-
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = sum;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        sum = buf[tiisg];
-        sum = simd_sum(sum);
-    }
-
-    const float inv_sum = 1.0f/sum;
-
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        pdst[i00] *= inv_sum;
-    }
-}
-
-kernel void kernel_soft_max_4(
-        device const float * src0,
-        device const float * src1,
-        device const float * src2,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint32_t & n_head_log2,
-        threadgroup  float * buf [[threadgroup(0)]],
-        uint  tgpig[[threadgroup_position_in_grid]],
-        uint  tpitg[[thread_position_in_threadgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint    ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = (tgpig) / (ne02*ne01);
-    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
-    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
-
-    device const float4 * psrc4 =                (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
-    device const float4 * pmask = src1 != src0 ? (device const float4 *)(src1 +                                      i01*ne00) : nullptr;
-    device const float4 * ppos  = src2 != src0 ? (device const float4 *)(src2)                                                 : nullptr;
-    device       float4 * pdst4 =                (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
-
-    float slope = 0.0f;
-
-    if (max_bias > 0.0f) {
-        const int64_t h = i02;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slope = pow(base, exp);
-    }
-
-    // parallel max
-    float4 lmax4 = -INFINITY;
-
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
-    }
-
-    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
-
-    float max_val = simd_max(lmax);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = -INFINITY;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = max_val;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        max_val = buf[tiisg];
-        max_val = simd_max(max_val);
-    }
-
-    // parallel sum
-    float4 lsum4 = 0.0f;
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
-        lsum4 += exp_psrc4;
-        pdst4[i00] = exp_psrc4;
-    }
-
-    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
-
-    // This barrier fixes a failing test
-    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
-    threadgroup_barrier(mem_flags::mem_none);
-
-    float sum = simd_sum(lsum);
-
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = sum;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        sum = buf[tiisg];
-        sum = simd_sum(sum);
-    }
-
-    const float inv_sum = 1.0f/sum;
-
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        pdst4[i00] *= inv_sum;
-    }
-}
-
-kernel void kernel_diag_mask_inf(
-        device const float * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant       int & n_past,
-        uint3 tpig[[thread_position_in_grid]]) {
-    const int64_t i02 = tpig[2];
-    const int64_t i01 = tpig[1];
-    const int64_t i00 = tpig[0];
-
-    if (i00 > n_past + i01) {
-        dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
-    } else {
-        dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
-    }
-}
-
-kernel void kernel_diag_mask_inf_8(
-        device const float4 * src0,
-        device       float4 * dst,
-        constant    int64_t & ne00,
-        constant    int64_t & ne01,
-        constant        int & n_past,
-        uint3 tpig[[thread_position_in_grid]]) {
-
-    const int64_t i = 2*tpig[0];
-
-    dst[i+0] = src0[i+0];
-    dst[i+1] = src0[i+1];
-    int64_t i4 = 4*i;
-    const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
-    const int64_t i01 = i4/(ne00);      i4 -= i01*ne00;
-    const int64_t i00 = i4;
-    for (int k = 3; k >= 0; --k) {
-        if (i00 + 4 + k <= n_past + i01) {
-            break;
-        }
-        dst[i+1][k] = -INFINITY;
-        if (i00 + k > n_past + i01) {
-            dst[i][k] = -INFINITY;
-        }
-    }
-}
-
-kernel void kernel_norm(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant     float & eps,
-        threadgroup float  * sum [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
-    // MEAN
-    // parallel sum
-    sum[tpitg] = 0.0f;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        sum[tpitg] += x[i00];
-    }
-    // reduce
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg/2; i > 0; i /= 2) {
-        if (tpitg < i) {
-            sum[tpitg] += sum[tpitg + i];
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-    const float mean  = sum[0] / ne00;
-
-    // recenter and VARIANCE
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    device float * y = dst + tgpig*ne00;
-    sum[tpitg] = 0.0f;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        y[i00] = x[i00] - mean;
-        sum[tpitg] += y[i00] * y[i00];
-    }
-
-    // reduce
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg/2; i > 0; i /= 2) {
-        if (tpitg < i) {
-            sum[tpitg] += sum[tpitg + i];
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-    const float variance = sum[0] / ne00;
-
-    const float scale = 1.0f/sqrt(variance + eps);
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        y[i00] = y[i00] * scale;
-    }
-}
-
-kernel void kernel_rms_norm(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant     float & eps,
-        threadgroup float  * buf [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
-
-    float4 sumf = 0;
-    float all_sum = 0;
-
-    // parallel sum
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        sumf += x[i00] * x[i00];
-    }
-    all_sum = sumf[0] + sumf[1] + sumf[2] + sumf[3];
-    all_sum = simd_sum(all_sum);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = all_sum;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        all_sum = buf[tiisg];
-        all_sum = simd_sum(all_sum);
-    }
-
-    const float mean  = all_sum/ne00;
-    const float scale = 1.0f/sqrt(mean + eps);
-
-    device float4 * y = (device float4 *) (dst + tgpig*ne00);
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        y[i00] = x[i00] * scale;
-    }
-}
-
-kernel void kernel_group_norm(
-        device const float * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int32_t & n_groups,
-        constant     float & eps,
-        threadgroup float  * buf [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    const int64_t ne = ne00*ne01*ne02;
-    const int64_t gs = ne00*ne01*((ne02 + n_groups - 1) / n_groups);
-
-    int start = tgpig * gs;
-    int end   = start + gs;
-
-    start += tpitg;
-
-    if (end >= ne) {
-        end = ne;
-    }
-
-    float tmp = 0.0f; // partial sum for thread in warp
-
-    for (int j = start; j < end; j += ntg) {
-        tmp += src0[j];
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    tmp = simd_sum(tmp);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = tmp;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        tmp = buf[tiisg];
-        tmp = simd_sum(tmp);
-    }
-
-    const float mean = tmp / gs;
-    tmp = 0.0f;
-
-    for (int j = start; j < end; j += ntg) {
-        float xi = src0[j] - mean;
-        dst[j] = xi;
-        tmp += xi * xi;
-    }
-
-    tmp = simd_sum(tmp);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = tmp;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        tmp = buf[tiisg];
-        tmp = simd_sum(tmp);
-    }
-
-    const float variance = tmp / gs;
-    const float scale = 1.0f/sqrt(variance + eps);
-    for (int j = start; j < end; j += ntg) {
-        dst[j] *= scale;
-    }
-}
-
-// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
-// il indicates where the q4 quants begin (0 or QK4_0/4)
-// we assume that the yl's have been multiplied with the appropriate scale factor
-// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
-inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
-    float d = qb_curr->d;
-
-    float2 acc = 0.f;
-
-    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
-
-    for (int i = 0; i < 8; i+=2) {
-        acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
-                + yl[i + 1] * (qs[i / 2] & 0x0F00);
-        acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
-                + yl[i + 9] * (qs[i / 2] & 0xF000);
-    }
-    return d * (sumy * -8.f + acc[0] + acc[1]);
-}
-
-// function for calculate inner product between half a q4_1 block and 16 floats (yl), sumy is SUM(yl[i])
-// il indicates where the q4 quants begin (0 or QK4_0/4)
-// we assume that the yl's have been multiplied with the appropriate scale factor
-// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
-inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
-    float d = qb_curr->d;
-    float m = qb_curr->m;
-
-    float2 acc = 0.f;
-
-    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
-
-    for (int i = 0; i < 8; i+=2) {
-        acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
-                + yl[i + 1] * (qs[i / 2] & 0x0F00);
-        acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
-                + yl[i + 9] * (qs[i / 2] & 0xF000);
-    }
-    return d * (acc[0] + acc[1]) + sumy * m;
-}
-
-// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
-// il indicates where the q5 quants begin (0 or QK5_0/4)
-// we assume that the yl's have been multiplied with the appropriate scale factor
-// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
-inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
-    float d = qb_curr->d;
-
-    float2 acc = 0.f;
-
-    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 3 + il/2);
-           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
-
-    for (int i = 0; i < 8; i+=2) {
-        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
-                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
-        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
-                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
-    }
-    return d * (sumy * -16.f + acc[0] + acc[1]);
-}
-
-// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
-// il indicates where the q5 quants begin (0 or QK5_1/4)
-// we assume that the yl's have been multiplied with the appropriate scale factor
-// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
-inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
-    float d = qb_curr->d;
-    float m = qb_curr->m;
-
-    float2 acc = 0.f;
-
-    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 4 + il/2);
-           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
-
-    for (int i = 0; i < 8; i+=2) {
-        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
-                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
-        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
-                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
-    }
-    return d * (acc[0] + acc[1]) + sumy * m;
-}
-
-// putting them in the kernel cause a significant performance penalty
-#define N_DST 4        // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2  // number of SIMD groups in a thread group
-//Note: This is a template, but strictly speaking it only applies to
-//      quantizations where the block size is 32. It also does not
-//      guard against the number of rows not being divisible by
-//      N_DST, so this is another explicit assumption of the implementation.
-template<typename block_q_type, int nr, int nsg, int nw>
-void mul_vec_q_n_f32_impl(
-        device const void  * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3 tgpig, uint tiisg, uint sgitg) {
-    const int nb = ne00/QK4_0;
-
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * nsg + sgitg) * nr;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q_type * x = (device const block_q_type *) src0 + offset0;
-    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[16]; // src1 vector cache
-    float sumf[nr] = {0.f};
-
-    const int ix = (tiisg/2);
-    const int il = (tiisg%2)*8;
-
-    device const float * yb = y + ix * QK4_0 + il;
-
-    // each thread in a SIMD group deals with half a block.
-    for (int ib = ix; ib < nb; ib += nw/2) {
-        float sumy = 0;
-        for (int i = 0; i < 8; i += 2) {
-            sumy += yb[i] + yb[i+1];
-            yl[i+0] = yb[i+ 0];
-            yl[i+1] = yb[i+ 1]/256.f;
-
-            sumy += yb[i+16] + yb[i+17];
-            yl[i+8] = yb[i+16]/16.f;
-            yl[i+9] = yb[i+17]/4096.f;
-        }
-
-        for (int row = 0; row < nr; row++) {
-            sumf[row] += block_q_n_dot_y(x+ib+row*nb, sumy, yl, il);
-        }
-
-        yb += QK4_0 * 16;
-    }
-
-    for (int row = 0; row < nr; ++row) {
-        const float tot = simd_sum(sumf[row]);
-        if (tiisg == 0 && first_row + row < ne01) {
-            dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
-        }
-    }
-}
-
-kernel void kernel_mul_mv_q4_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
-}
-
-kernel void kernel_mul_mv_q4_1_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
-}
-
-kernel void kernel_mul_mv_q5_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
-}
-
-kernel void kernel_mul_mv_q5_1_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
-}
-
-
-#define NB_Q8_0 8
-
-void kernel_mul_mv_q8_0_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-    const int nr  = N_DST;
-    const int nsg = N_SIMDGROUP;
-    const int nw  = N_SIMDWIDTH;
-
-    const int nb = ne00/QK8_0;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * nsg + sgitg) * nr;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[NB_Q8_0];
-    float sumf[nr]={0.f};
-
-    const int ix = tiisg/4;
-    const int il = tiisg%4;
-
-    device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
-
-    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
-    for (int ib = ix; ib < nb; ib += nw/4) {
-        for (int i = 0; i < NB_Q8_0; ++i) {
-            yl[i] = yb[i];
-        }
-
-        for (int row = 0; row < nr; row++) {
-            device const int8_t * qs = x[ib+row*nb].qs + NB_Q8_0*il;
-            float sumq = 0.f;
-            for (int iq = 0; iq < NB_Q8_0; ++iq) {
-                sumq += qs[iq] * yl[iq];
-            }
-            sumf[row] += sumq*x[ib+row*nb].d;
-        }
-
-        yb += NB_Q8_0 * nw;
-    }
-
-    for (int row = 0; row < nr; ++row) {
-        const float tot = simd_sum(sumf[row]);
-        if (tiisg == 0 && first_row + row < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_q8_0_f32")]]
-kernel void kernel_mul_mv_q8_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    kernel_mul_mv_q8_0_f32_impl(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
-}
-
-#define N_F32_F32 4
-
-void kernel_mul_mv_f32_f32_impl(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb00,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                   int64_t   ne10,
-                   int64_t   ne11,
-                   int64_t   ne12,
-                  uint64_t   nb10,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                     uint    r2,
-                     uint    r3,
-                     uint3   tgpig,
-                     uint    tiisg) {
-
-    const int64_t r0 = tgpig.x;
-    const int64_t rb = tgpig.y*N_F32_F32;
-    const int64_t im = tgpig.z;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
-
-    device const float * x = (device const float *) (src0 + offset0);
-
-    if (ne00 < 128) {
-        for (int row = 0; row < N_F32_F32; ++row) {
-            int r1 = rb + row;
-            if (r1 >= ne11) {
-                break;
-            }
-
-            device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
-
-            float sumf = 0;
-            for (int i = tiisg; i < ne00; i += 32) {
-                sumf += (float) x[i] * (float) y[i];
-            }
-
-            float all_sum = simd_sum(sumf);
-            if (tiisg == 0) {
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-            }
-        }
-    } else {
-        device const float4 * x4 = (device const float4 *)x;
-        for (int row = 0; row < N_F32_F32; ++row) {
-            int r1 = rb + row;
-            if (r1 >= ne11) {
-                break;
-            }
-
-            device const float  * y  = (device const float  *) (src1 + r1*nb11 + im*nb12);
-            device const float4 * y4 = (device const float4 *) y;
-
-            float sumf = 0;
-            for (int i = tiisg; i < ne00/4; i += 32) {
-                for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
-            }
-
-            float all_sum = simd_sum(sumf);
-            if (tiisg == 0) {
-                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-            }
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_f32_f32")]]
-kernel void kernel_mul_mv_f32_f32(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
-    kernel_mul_mv_f32_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
-}
-
-#define N_F16_F16 4
-
-kernel void kernel_mul_mv_f16_f16(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
-
-    const int64_t r0 = tgpig.x;
-    const int64_t rb = tgpig.y*N_F16_F16;
-    const int64_t im = tgpig.z;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
-
-    device const half * x = (device const half *) (src0 + offset0);
-
-    if (ne00 < 128) {
-        for (int row = 0; row < N_F16_F16; ++row) {
-            int r1 = rb + row;
-            if (r1 >= ne11) {
-                break;
-            }
-
-            device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12);
-
-            float sumf = 0;
-            for (int i = tiisg; i < ne00; i += 32) {
-                sumf += (half) x[i] * (half) y[i];
-            }
-
-            float all_sum = simd_sum(sumf);
-            if (tiisg == 0) {
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-            }
-        }
-    } else {
-        device const half4 * x4 = (device const half4 *)x;
-        for (int row = 0; row < N_F16_F16; ++row) {
-            int r1 = rb + row;
-            if (r1 >= ne11) {
-                break;
-            }
-
-            device const half  * y  = (device const half  *) (src1 + r1*nb11 + im*nb12);
-            device const half4 * y4 = (device const half4 *) y;
-
-            float sumf = 0;
-            for (int i = tiisg; i < ne00/4; i += 32) {
-                for (int k = 0; k < 4; ++k) sumf += (half) x4[i][k] * y4[i][k];
-            }
-
-            float all_sum = simd_sum(sumf);
-            if (tiisg == 0) {
-                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (half) x[i] * y[i];
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-            }
-        }
-    }
-}
-
-void kernel_mul_mv_f16_f32_1row_impl(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
-
-    device const half  * x = (device const half  *) (src0 + offset0);
-    device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
-
-    float sumf = 0;
-    if (ne00 < 128) {
-        for (int i = tiisg; i < ne00; i += 32) {
-            sumf += (float) x[i] * (float) y[i];
-        }
-        float all_sum = simd_sum(sumf);
-        if (tiisg == 0) {
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-        }
-    } else {
-        device const half4  * x4 = (device const half4  *) x;
-        device const float4 * y4 = (device const float4 *) y;
-        for (int i = tiisg; i < ne00/4; i += 32) {
-            for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k];
-        }
-        float all_sum = simd_sum(sumf);
-        if (tiisg == 0) {
-            for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_f16_f32_1row")]]
-kernel void kernel_mul_mv_f16_f32_1row(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
-    kernel_mul_mv_f16_f32_1row_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
-}
-
-#define N_F16_F32 4
-
-void kernel_mul_mv_f16_f32_impl(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb00,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                   int64_t   ne10,
-                   int64_t   ne11,
-                   int64_t   ne12,
-                  uint64_t   nb10,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-                   uint3     tgpig,
-                   uint      tiisg) {
-
-    const int64_t r0 = tgpig.x;
-    const int64_t rb = tgpig.y*N_F16_F32;
-    const int64_t im = tgpig.z;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
-
-    device const half * x = (device const half *) (src0 + offset0);
-
-    if (ne00 < 128) {
-        for (int row = 0; row < N_F16_F32; ++row) {
-            int r1 = rb + row;
-            if (r1 >= ne11) {
-                break;
-            }
-
-            device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
-
-            float sumf = 0;
-            for (int i = tiisg; i < ne00; i += 32) {
-                sumf += (float) x[i] * (float) y[i];
-            }
-
-            float all_sum = simd_sum(sumf);
-            if (tiisg == 0) {
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-            }
-        }
-    } else {
-        device const half4 * x4 = (device const half4 *)x;
-        for (int row = 0; row < N_F16_F32; ++row) {
-            int r1 = rb + row;
-            if (r1 >= ne11) {
-                break;
-            }
-
-            device const float  * y  = (device const float  *) (src1 + r1*nb11 + im*nb12);
-            device const float4 * y4 = (device const float4 *) y;
-
-            float sumf = 0;
-            for (int i = tiisg; i < ne00/4; i += 32) {
-                for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
-            }
-
-            float all_sum = simd_sum(sumf);
-            if (tiisg == 0) {
-                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-            }
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_f16_f32")]]
-kernel void kernel_mul_mv_f16_f32(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]]) {
-    kernel_mul_mv_f16_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
-}
-
-// Assumes row size (ne00) is a multiple of 4
-kernel void kernel_mul_mv_f16_f32_l4(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]]) {
-
-    const int nrows = ne11;
-    const int64_t r0 = tgpig.x;
-    const int64_t im = tgpig.z;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
-
-    device const half4 * x4 = (device const half4 *) (src0 + offset0);
-
-    for (int r1 = 0; r1 < nrows; ++r1) {
-        device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12);
-
-        float sumf = 0;
-        for (int i = tiisg; i < ne00/4; i += 32) {
-            for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
-        }
-
-        float all_sum = simd_sum(sumf);
-        if (tiisg == 0) {
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
-        }
-    }
-}
-
-kernel void kernel_alibi_f32(
-        device const float * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        constant     float & m0,
-        constant     float & m1,
-        constant       int & n_heads_log2_floor,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-  //const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
-
-    const int64_t k = i3*ne3 + i2;
-
-    float m_k;
-    if (k < n_heads_log2_floor) {
-        m_k = pow(m0, k + 1);
-    } else {
-        m_k = pow(m1, 2 * (k - n_heads_log2_floor) + 1);
-    }
-
-    device       char * dst_row = (device char *) dst + i3*nb3 + i2*nb2 + i1*nb1;
-    device const char * src_row = (device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        const  float   src_v = *(device float *)(src_row + i00*nb00);
-        device float * dst_v =  (device float *)(dst_row + i00*nb0);
-        *dst_v = i00 * m_k + src_v;
-    }
-}
-
-static float rope_yarn_ramp(const float low, const float high, const int i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
-// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
-static void rope_yarn(
-    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
-    thread float * cos_theta, thread float * sin_theta
-) {
-    // Get n-d rotational scaling corrected for extrapolation
-    float theta_interp = freq_scale * theta_extrap;
-    float theta = theta_interp;
-    if (ext_factor != 0.0f) {
-        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
-        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-
-        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
-    }
-    *cos_theta = cos(theta) * mscale;
-    *sin_theta = sin(theta) * mscale;
-}
-
-// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
-// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
-static float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
-    return n_dims * log(n_orig_ctx / (n_rot * 2 * M_PI_F)) / (2 * log(base));
-}
-
-static void rope_yarn_corr_dims(
-    int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
-) {
-    // start and end correction dims
-    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
-    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
-}
-
-typedef void (rope_t)(
-        device const    void * src0,
-        device const int32_t * src1,
-        device         float * dst,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant     int64_t & ne03,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant    uint64_t & nb03,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant     int64_t & ne2,
-        constant     int64_t & ne3,
-        constant    uint64_t & nb0,
-        constant    uint64_t & nb1,
-        constant    uint64_t & nb2,
-        constant    uint64_t & nb3,
-        constant         int & n_past,
-        constant         int & n_dims,
-        constant         int & mode,
-        constant         int & n_orig_ctx,
-        constant       float & freq_base,
-        constant       float & freq_scale,
-        constant       float & ext_factor,
-        constant       float & attn_factor,
-        constant       float & beta_fast,
-        constant       float & beta_slow,
-        uint  tiitg[[thread_index_in_threadgroup]],
-        uint3 tptg[[threads_per_threadgroup]],
-        uint3 tgpig[[threadgroup_position_in_grid]]);
-
-template<typename T>
-kernel void kernel_rope(
-        device const    void * src0,
-        device const int32_t * src1,
-        device         float * dst,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant     int64_t & ne03,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant    uint64_t & nb03,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant     int64_t & ne2,
-        constant     int64_t & ne3,
-        constant    uint64_t & nb0,
-        constant    uint64_t & nb1,
-        constant    uint64_t & nb2,
-        constant    uint64_t & nb3,
-        constant         int & n_past,
-        constant         int & n_dims,
-        constant         int & mode,
-        constant         int & n_orig_ctx,
-        constant       float & freq_base,
-        constant       float & freq_scale,
-        constant       float & ext_factor,
-        constant       float & attn_factor,
-        constant       float & beta_fast,
-        constant       float & beta_slow,
-        uint  tiitg[[thread_index_in_threadgroup]],
-        uint3 tptg[[threads_per_threadgroup]],
-        uint3 tgpig[[threadgroup_position_in_grid]]) {
-    const int64_t i3 = tgpig[2];
-    const int64_t i2 = tgpig[1];
-    const int64_t i1 = tgpig[0];
-
-    const bool is_neox = mode & 2;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
-
-    device const int32_t * pos = src1;
-
-    const int64_t p = pos[i2];
-
-    const float theta_0 = (float)p;
-    const float inv_ndims = -1.f/n_dims;
-
-    if (!is_neox) {
-        for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
-
-            const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
-            float cos_theta, sin_theta;
-            rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
-
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-            const T x0 = src[0];
-            const T x1 = src[1];
-
-            dst_data[0] = x0*cos_theta - x1*sin_theta;
-            dst_data[1] = x0*sin_theta + x1*cos_theta;
-        }
-    } else {
-        for (int64_t ic = 2*tiitg; ic < ne0; ic += 2*tptg.x) {
-            if (ic < n_dims) {
-                const int64_t ib = 0;
-
-                // simplified from `(ib * n_dims + ic) * inv_ndims`
-                const float cur_rot = inv_ndims*ic - ib;
-
-                const float theta = theta_0 * pow(freq_base, cur_rot);
-                float cos_theta, sin_theta;
-                rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
-
-                const int64_t i0 = ib*n_dims + ic/2;
-
-                device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                const float x0 = src[0];
-                const float x1 = src[n_dims/2];
-
-                dst_data[0]        = x0*cos_theta - x1*sin_theta;
-                dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
-            } else {
-                const int64_t i0 = ic;
-
-                device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                dst_data[0] = src[0];
-                dst_data[1] = src[1];
-            }
-        }
-    }
-}
-
-template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
-template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
-
-typedef void (im2col_t)(
-        device const float * x,
-        device        char * dst,
-        constant   int32_t & ofs0,
-        constant   int32_t & ofs1,
-        constant   int32_t & IW,
-        constant   int32_t & IH,
-        constant   int32_t & CHW,
-        constant   int32_t & s0,
-        constant   int32_t & s1,
-        constant   int32_t & p0,
-        constant   int32_t & p1,
-        constant   int32_t & d0,
-        constant   int32_t & d1,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3  tgpg[[threadgroups_per_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]);
-
-template <typename T>
-kernel void kernel_im2col(
-        device const float * x,
-        device        char * dst,
-        constant   int32_t & ofs0,
-        constant   int32_t & ofs1,
-        constant   int32_t & IW,
-        constant   int32_t & IH,
-        constant   int32_t & CHW,
-        constant   int32_t & s0,
-        constant   int32_t & s1,
-        constant   int32_t & p0,
-        constant   int32_t & p1,
-        constant   int32_t & d0,
-        constant   int32_t & d1,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3  tgpg[[threadgroups_per_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0;
-    const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1;
-
-    const int32_t offset_dst =
-        (tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
-        (tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]);
-
-    device T * pdst = (device T *) (dst);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        pdst[offset_dst] = 0.0f;
-    } else {
-        const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1;
-        pdst[offset_dst] = x[offset_src + iih * IW + iiw];
-    }
-}
-
-template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
-template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
-
-kernel void kernel_upscale_f32(
-    device  const char * src0,
-    device        char * dst,
-    constant   int64_t & ne00,
-    constant   int64_t & ne01,
-    constant   int64_t & ne02,
-    constant   int64_t & ne03,
-    constant  uint64_t & nb00,
-    constant  uint64_t & nb01,
-    constant  uint64_t & nb02,
-    constant  uint64_t & nb03,
-    constant   int64_t & ne0,
-    constant   int64_t & ne1,
-    constant   int64_t & ne2,
-    constant   int64_t & ne3,
-    constant  uint64_t & nb0,
-    constant  uint64_t & nb1,
-    constant  uint64_t & nb2,
-    constant  uint64_t & nb3,
-    constant   int32_t & sf,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
-
-    const int64_t i03 = i3;
-    const int64_t i02 = i2;
-    const int64_t i01 = i1/sf;
-
-    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
-    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        dst_ptr[i0] = src0_ptr[i0/sf];
-    }
-}
-
-kernel void kernel_pad_f32(
-    device  const char * src0,
-    device        char * dst,
-    constant   int64_t & ne00,
-    constant   int64_t & ne01,
-    constant   int64_t & ne02,
-    constant   int64_t & ne03,
-    constant  uint64_t & nb00,
-    constant  uint64_t & nb01,
-    constant  uint64_t & nb02,
-    constant  uint64_t & nb03,
-    constant   int64_t & ne0,
-    constant   int64_t & ne1,
-    constant   int64_t & ne2,
-    constant   int64_t & ne3,
-    constant  uint64_t & nb0,
-    constant  uint64_t & nb1,
-    constant  uint64_t & nb2,
-    constant  uint64_t & nb3,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
-
-    const int64_t i03 = i3;
-    const int64_t i02 = i2;
-    const int64_t i01 = i1;
-
-    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
-    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
-
-    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
-        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-            if (i0 < ne00) {
-                dst_ptr[i0] = src0_ptr[i0];
-            } else {
-                dst_ptr[i0] = 0.0f;
-            }
-        }
-
-        return;
-    }
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        dst_ptr[i0] = 0.0f;
-    }
-}
-
-kernel void kernel_arange_f32(
-    device        char * dst,
-    constant   int64_t & ne0,
-    constant   float   & start,
-    constant   float   & step,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    device float * dst_ptr = (device float *) dst;
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        dst_ptr[i0] = start + step * i0;
-    }
-}
-
-kernel void kernel_timestep_embedding_f32(
-    device  const char * src0,
-    device        char * dst,
-    constant  uint64_t & nb1,
-    constant  int      & dim,
-    constant  int      & max_period,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    int i = tgpig.x;
-    device float * embed_data = (device float *)(dst +  i*nb1);
-
-    int half_ = dim / 2;
-    for (int j = tpitg.x; j < half_; j += ntg.x) {
-        float timestep = ((device float *)src0)[i];
-        float freq = (float)exp(-log((float)max_period) * j / half_);
-        float arg = timestep * freq;
-        embed_data[j        ] = cos(arg);
-        embed_data[j + half_] = sin(arg);
-    }
-
-    if (dim % 2 != 0 && tpitg.x == 0) {
-        embed_data[dim] = 0.f;
-    }
-}
-
-// bitonic sort implementation following the CUDA kernels as reference
-typedef void (argsort_t)(
-        device const float  * x,
-        device     int32_t  * dst,
-        constant   int64_t  & ncols,
-        constant   int64_t  & ncols_pad,
-        threadgroup int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]);
-
-template<ggml_sort_order order>
-kernel void kernel_argsort_f32_i32(
-        device const float   * x,
-        device       int32_t * dst,
-        constant     int64_t & ncols,
-        constant     int64_t & ncols_pad,
-        threadgroup int32_t  * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]) {
-    // bitonic sort
-    int col = tpitg[0];
-    int row = tgpig[1];
-
-    if (col >= ncols_pad) return;
-
-    device const float   * x_row   = x + row * ncols;
-    threadgroup int32_t  * dst_row = shared_values;
-
-    // initialize indices
-    dst_row[col] = col;
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    for (int k = 2; k <= ncols_pad; k *= 2) {
-        for (int j = k / 2; j > 0; j /= 2) {
-            int ixj = col ^ j;
-            if (ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= ncols ||
-                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
-                    ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
-                    }
-                } else {
-                    if (dst_row[ixj] >= ncols ||
-                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
-                    ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
-                    }
-                }
-            }
-            threadgroup_barrier(mem_flags::mem_threadgroup);
-        }
-    }
-
-    // copy the result to dst without the padding
-    if (col < ncols) {
-        dst[row * ncols + col] = dst_row[col];
-    }
-}
-
-template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
-template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
-
-kernel void kernel_leaky_relu_f32(
-        device const float * src0,
-        device       float * dst,
-        constant     float & slope,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope;
-}
-
-kernel void kernel_cpy_f16_f16(
-        device  const half * src0,
-        device        half * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
-
-    device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-        dst_data[i00] = src[0];
-    }
-}
-
-kernel void kernel_cpy_f16_f32(
-        device  const half * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
-
-    device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-        dst_data[i00] = src[0];
-    }
-}
-
-kernel void kernel_cpy_f32_f16(
-        device const float * src0,
-        device        half * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
-
-    device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        dst_data[i00] = src[0];
-    }
-}
-
-kernel void kernel_cpy_f32_f32(
-        device const float * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
-
-    device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        dst_data[i00] = src[0];
-    }
-}
-
-kernel void kernel_cpy_f32_q8_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK8_0;
-
-    device block_q8_0 * dst_data = (device block_q8_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x*QK8_0; i00 < ne00; i00 += ntg.x*QK8_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        float amax = 0.0f; // absolute max
-
-        for (int j = 0; j < QK8_0; j++) {
-            const float v = src[j];
-            amax = MAX(amax, fabs(v));
-        }
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK8_0].d = d;
-
-        for (int j = 0; j < QK8_0; ++j) {
-            const float x0 = src[j]*id;
-
-            dst_data[i00/QK8_0].qs[j] = round(x0);
-        }
-    }
-}
-
-kernel void kernel_cpy_f32_q4_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_0;
-
-    device block_q4_0 * dst_data = (device block_q4_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x*QK4_0; i00 < ne00; i00 += ntg.x*QK4_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -8;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_0].d = d;
-
-        for (int j = 0; j < QK4_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK4_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
-
-            dst_data[i00/QK4_0].qs[j]  = xi0;
-            dst_data[i00/QK4_0].qs[j] |= xi1 << 4;
-        }
-    }
-}
-
-kernel void kernel_cpy_f32_q4_1(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_1;
-
-    device block_q4_1 * dst_data = (device block_q4_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x*QK4_1; i00 < ne00; i00 += ntg.x*QK4_1) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
-
-        for (int j = 0; j < QK4_1; j++) {
-            const float v = src[j];
-            if (min > v) min = v;
-            if (max < v) max = v;
-        }
-
-        const float d = (max - min) / ((1 << 4) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_1].d = d;
-        dst_data[i00/QK4_1].m = min;
-
-        for (int j = 0; j < QK4_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK4_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
-
-            dst_data[i00/QK4_1].qs[j]  = xi0;
-            dst_data[i00/QK4_1].qs[j] |= xi1 << 4;
-        }
-    }
-}
-
-kernel void kernel_cpy_f32_q5_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_0;
-
-    device block_q5_0 * dst_data = (device block_q5_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x*QK5_0; i00 < ne00; i00 += ntg.x*QK5_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK5_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -16;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK5_0].d = d;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK5_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
-            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
-
-            dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_0].qh[j] = qh8[j];
-        }
-    }
-}
-
-kernel void kernel_cpy_f32_q5_1(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_1;
-
-    device block_q5_1 * dst_data = (device block_q5_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x*QK5_1; i00 < ne00; i00 += ntg.x*QK5_1) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        float max = src[0];
-        float min = src[0];
-
-        for (int j = 1; j < QK5_1; j++) {
-            const float v = src[j];
-            min = v < min ? v : min;
-            max = v > max ? v : max;
-        }
-
-        const float d = (max - min) / 31;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK5_1].d = d;
-        dst_data[i00/QK5_1].m = min;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK5_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-            dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_1].qh[j] = qh8[j];
-        }
-    }
-}
-
-static inline int best_index_int8(int n, constant float * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
-    }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
-}
-
-constexpr constant static float kvalues_iq4nl_f[16] = {
-    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
-};
-
-kernel void kernel_cpy_f32_iq4_nl(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_NL;
-
-    device block_iq4_nl * dst_data = (device block_iq4_nl *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-    for (int64_t i00 = tpitg.x*QK4_NL; i00 < ne00; i00 += ntg.x*QK4_NL) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / kvalues_iq4nl_f[0];
-        const float id = d ? 1.0f/d : 0.0f;
-
-        float sumqx = 0, sumq2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            const float x0 = src[0        + j]*id;
-            const float x1 = src[QK4_NL/2 + j]*id;
-
-            const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
-            const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
-
-            dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4);
-
-            const float v0 = kvalues_iq4nl_f[xi0];
-            const float v1 = kvalues_iq4nl_f[xi1];
-            const float w0 = src[0        + j]*src[0        + j];
-            const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
-            sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
-            sumq2 += w0*v0*v0 + w1*v1*v1;
-
-        }
-
-        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
-
-    }
-}
-
-kernel void kernel_concat(
-    device  const char * src0,
-    device  const char * src1,
-    device        char * dst,
-    constant   int64_t & ne00,
-    constant   int64_t & ne01,
-    constant   int64_t & ne02,
-    constant   int64_t & ne03,
-    constant  uint64_t & nb00,
-    constant  uint64_t & nb01,
-    constant  uint64_t & nb02,
-    constant  uint64_t & nb03,
-    constant   int64_t & ne10,
-    constant   int64_t & ne11,
-    constant   int64_t & ne12,
-    constant   int64_t & ne13,
-    constant  uint64_t & nb10,
-    constant  uint64_t & nb11,
-    constant  uint64_t & nb12,
-    constant  uint64_t & nb13,
-    constant   int64_t & ne0,
-    constant   int64_t & ne1,
-    constant   int64_t & ne2,
-    constant   int64_t & ne3,
-    constant  uint64_t & nb0,
-    constant  uint64_t & nb1,
-    constant  uint64_t & nb2,
-    constant  uint64_t & nb3,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
-
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
-
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        if (i02 < ne02) {
-            ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0];
-            src0_ptr += ntg.x*nb00;
-        } else {
-            ((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0];
-            src1_ptr += ntg.x*nb10;
-        }
-        dst_ptr += ntg.x*nb0;
-    }
-}
-
-void kernel_mul_mv_q2_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q2_K * x = (device const block_q2_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int step = sizeof(block_q2_K) * nb;
-
-#if QK_K == 256
-    const int ix = tiisg/8;  // 0...3
-    const int it = tiisg%8;  // 0...7
-    const int iq = it/4;     // 0 or 1
-    const int ir = it%4;     // 0...3
-    const int is = (8*ir)/16;// 0 or 1
-
-    device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir;
-
-    for (int ib = ix; ib < nb; ib += 4) {
-
-        float4 sumy = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
-            yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
-            yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
-            yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
-        }
-
-        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales + 8*iq + is;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
-        device const half     * dh = &x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
-            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
-                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
-                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
-                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
-                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
-                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
-                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
-                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
-            }
-            float dall = dh[0];
-            float dmin = dh[1] * 1.f/16.f;
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
-                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
-                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
-                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
-                         dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
-
-            qs += step/2;
-            sc += step;
-            dh += step/2;
-        }
-
-        y4 += 4 * QK_K;
-    }
-#else
-    const int ix = tiisg/2;  // 0...15
-    const int it = tiisg%2;  // 0...1
-
-    device const float * y4 = y + ix * QK_K + 8 * it;
-
-    for (int ib = ix; ib < nb; ib += 16) {
-
-        float4 sumy = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
-            yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
-            yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
-            yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
-        }
-
-        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
-        device const half     * dh = &x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
-            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
-                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
-                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
-                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
-                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
-                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
-                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
-                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
-            }
-
-            float dall = dh[0];
-            float dmin = dh[1];
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
-                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
-                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
-                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
-                         dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
-
-            qs += step/2;
-            sc += step;
-            dh += step/2;
-        }
-
-        y4 += 16 * QK_K;
-    }
-#endif
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_q2_K_f32")]]
-kernel void kernel_mul_mv_q2_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
-}
-
-#if QK_K == 256
-void kernel_mul_mv_q3_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-
-    //const uint16_t kmask1 = 0x3030;
-    //const uint16_t kmask2 = 0x0f0f;
-
-    const int tid = tiisg/4;
-    const int ix  = tiisg%4;
-    const int ip  = tid/4;          // 0 or 1
-    const int il  = 2*((tid%4)/2);  // 0 or 2
-    const int ir  = tid%2;
-    const int n   = 8;
-    const int l0  = n*ir;
-
-    // One would think that the Metal compiler would figure out that ip and il can only have
-    // 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it
-    // with these two tales.
-    //
-    // Possible masks for the high bit
-    const ushort4 mm[4] = {{0x0001, 0x0100, 0x0002, 0x0200},  // ip = 0, il = 0
-                           {0x0004, 0x0400, 0x0008, 0x0800},  // ip = 0, il = 2
-                           {0x0010, 0x1000, 0x0020, 0x2000},  // ip = 1, il = 0
-                           {0x0040, 0x4000, 0x0080, 0x8000}}; // ip = 1, il = 2
-
-    // Possible masks for the low 2 bits
-    const int4 qm[2] = {{0x0003, 0x0300, 0x000c, 0x0c00}, {0x0030, 0x3000, 0x00c0, 0xc000}};
-
-    const ushort4 hm = mm[2*ip + il/2];
-
-    const int shift = 2*il;
-    const float    v1 = il == 0 ? 4.f : 64.f;
-    const float    v2 = 4.f * v1;
-
-    const uint16_t s_shift1 = 4*ip;
-    const uint16_t s_shift2 = s_shift1 + il;
-
-    const int q_offset = 32*ip + l0;
-    const int y_offset = 128*ip + 32*il + l0;
-
-    const int step = sizeof(block_q3_K) * nb / 2;
-
-    device const float * y1 = yy + ix*QK_K + y_offset;
-
-    uint32_t scales32, aux32;
-    thread uint16_t * scales16 = (thread uint16_t *)&scales32;
-    thread const int8_t * scales = (thread const int8_t *)&scales32;
-
-    float sumf1[2] = {0.f};
-    float sumf2[2] = {0.f};
-    for (int i = ix; i < nb; i += 4) {
-
-        for (int l = 0; l < 8; ++l) {
-            yl[l+ 0] = y1[l+ 0];
-            yl[l+ 8] = y1[l+16];
-            yl[l+16] = y1[l+32];
-            yl[l+24] = y1[l+48];
-        }
-
-        device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
-        device const uint16_t * h = (device const uint16_t *)(x[i].hmask + l0);
-        device const uint16_t * a = (device const uint16_t *)(x[i].scales);
-        device const half * dh = &x[i].d;
-
-        for (int row = 0; row < 2; ++row) {
-
-            const float d_all = (float)dh[0];
-
-            scales16[0] = a[4];
-            scales16[1] = a[5];
-            aux32 = ((scales32 >> s_shift2) << 4) & 0x30303030;
-            scales16[0] = a[il+0];
-            scales16[1] = a[il+1];
-            scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32;
-
-            float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0;
-            for (int l = 0; l < n; l += 2) {
-                const int32_t qs = q[l/2];
-                s1 += yl[l+0] * (qs & qm[il/2][0]);
-                s2 += yl[l+1] * (qs & qm[il/2][1]);
-                s3 += ((h[l/2] & hm[0]) ? 0.f : yl[l+0]) + ((h[l/2] & hm[1]) ? 0.f : yl[l+1]);
-                s4 += yl[l+16] * (qs & qm[il/2][2]);
-                s5 += yl[l+17] * (qs & qm[il/2][3]);
-                s6 += ((h[l/2] & hm[2]) ? 0.f : yl[l+16]) + ((h[l/2] & hm[3]) ? 0.f : yl[l+17]);
-            }
-            float d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
-            float d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
-            sumf1[row] += d1 * (scales[0] - 32);
-            sumf2[row] += d2 * (scales[2] - 32);
-
-            s1 = s2 = s3 = s4 = s5 = s6 = 0;
-            for (int l = 0; l < n; l += 2) {
-                const int32_t qs = q[l/2+8];
-                s1 += yl[l+8] * (qs & qm[il/2][0]);
-                s2 += yl[l+9] * (qs & qm[il/2][1]);
-                s3 += ((h[l/2+8] & hm[0]) ? 0.f : yl[l+8]) + ((h[l/2+8] & hm[1]) ? 0.f : yl[l+9]);
-                s4 += yl[l+24] * (qs & qm[il/2][2]);
-                s5 += yl[l+25] * (qs & qm[il/2][3]);
-                s6 += ((h[l/2+8] & hm[2]) ? 0.f : yl[l+24]) + ((h[l/2+8] & hm[3]) ? 0.f : yl[l+25]);
-            }
-            d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
-            d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
-            sumf1[row] += d1 * (scales[1] - 32);
-            sumf2[row] += d2 * (scales[3] - 32);
-
-            q  += step;
-            h  += step;
-            a  += step;
-            dh += step;
-
-        }
-
-        y1 += 4 * QK_K;
-
-    }
-
-    for (int row = 0; row < 2; ++row) {
-        const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
-        sumf1[row] = simd_sum(sumf);
-    }
-    if (tiisg == 0) {
-        for (int row = 0; row < 2; ++row) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = sumf1[row];
-        }
-    }
-}
-#else
-void kernel_mul_mv_q3_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
-
-    const int row = 2 * r0 + sgitg;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    const int ix = tiisg/4;
-    const int il = 4 * (tiisg%4);// 0, 4, 8, 12
-    const int iq = il/8;         // 0, 0, 1, 1
-    const int in = il%8;         // 0, 4, 0, 4
-
-    float2 sum = {0.f, 0.f};
-
-    for (int i = ix; i < nb; i += 8) {
-
-        const float d_all = (float)(x[i].d);
-
-        device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
-        device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
-        device const uint16_t * s = (device const uint16_t *)(x[i].scales);
-        device const float    * y = yy + i * QK_K + il;
-
-        const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
-        const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
-        const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
-        const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
-
-        for (int l = 0; l < 4; l += 2) {
-            const uint16_t hm = h[l/2] >> iq;
-            sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 :  4))
-                    + y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
-                    + y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
-                    + y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
-            sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
-                    + y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
-                    + y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
-                    + y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
-        }
-
-    }
-    const float sumf = sum[0] + sum[1] * 1.f/256.f;
-
-    const float tot = simd_sum(sumf);
-    if (tiisg == 0) {
-        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
-    }
-
-}
-#endif
-
-[[host_name("kernel_mul_mv_q3_K_f32")]]
-kernel void kernel_mul_mv_q3_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
-}
-
-#if QK_K == 256
-void kernel_mul_mv_q4_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int ix = tiisg/8;  // 0...3
-    const int it = tiisg%8;  // 0...7
-    const int iq = it/4;     // 0 or 1
-    const int ir = it%4;     // 0...3
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-    //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int first_row = r0 * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[16];
-    float yh[16];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int step = sizeof(block_q4_K) * nb / 2;
-
-    device const float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
-
-    uint16_t sc16[4];
-    thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
-
-    for (int ib = ix; ib < nb; ib += 4) {
-
-        float4 sumy = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+0] = y4[i+  0]; sumy[0] += yl[i+0];
-            yl[i+8] = y4[i+ 32]; sumy[1] += yl[i+8];
-            yh[i+0] = y4[i+128]; sumy[2] += yh[i+0];
-            yh[i+8] = y4[i+160]; sumy[3] += yh[i+8];
-        }
-
-        device const uint16_t * sc = (device const uint16_t *)x[ib].scales + iq;
-        device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
-        device const half     * dh = &x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            sc16[0] = sc[0] & kmask1;
-            sc16[1] = sc[2] & kmask1;
-            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
-            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
-
-            device const uint16_t * q2 = q1 + 32;
-
-            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
-            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+0] * (q1[i/2] & 0x000F);
-                acc1[1] += yl[i+1] * (q1[i/2] & 0x0F00);
-                acc1[2] += yl[i+8] * (q1[i/2] & 0x00F0);
-                acc1[3] += yl[i+9] * (q1[i/2] & 0xF000);
-                acc2[0] += yh[i+0] * (q2[i/2] & 0x000F);
-                acc2[1] += yh[i+1] * (q2[i/2] & 0x0F00);
-                acc2[2] += yh[i+8] * (q2[i/2] & 0x00F0);
-                acc2[3] += yh[i+9] * (q2[i/2] & 0xF000);
-            }
-
-            float dall = dh[0];
-            float dmin = dh[1];
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8[0] +
-                                 (acc1[2] + 1.f/256.f * acc1[3]) * sc8[1] * 1.f/16.f +
-                                 (acc2[0] + 1.f/256.f * acc2[1]) * sc8[4] +
-                                 (acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
-                         dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
-
-            q1 += step;
-            sc += step;
-            dh += step;
-        }
-
-        y4 += 4 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-#else
-void kernel_mul_mv_q4_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int ix = tiisg/4;  // 0...7
-    const int it = tiisg%4;  // 0...3
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-    const int first_row = r0 * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[8];
-    float yh[8];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int step = sizeof(block_q4_K) * nb / 2;
-
-    device const float * y4 = y + ix * QK_K + 8 * it;
-
-    uint16_t sc16[4];
-
-    for (int ib = ix; ib < nb; ib += 8) {
-
-        float2 sumy = {0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i] = y4[i+ 0]; sumy[0] += yl[i];
-            yh[i] = y4[i+32]; sumy[1] += yh[i];
-        }
-
-        device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
-        device const half     * dh = x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            sc16[0] = sc[0] & 0x000f;
-            sc16[1] = sc[0] & 0x0f00;
-            sc16[2] = sc[0] & 0x00f0;
-            sc16[3] = sc[0] & 0xf000;
-
-            float2 acc1 = {0.f, 0.f};
-            float2 acc2 = {0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
-                acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
-                acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
-                acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
-            }
-
-            float dall = dh[0];
-            float dmin = dh[1];
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
-                                 (acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
-                         dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
-
-            qs += step;
-            sc += step;
-            dh += step;
-        }
-
-        y4 += 8 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-#endif
-
-[[host_name("kernel_mul_mv_q4_K_f32")]]
-kernel void kernel_mul_mv_q4_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_q4_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
-}
-
-void kernel_mul_mv_q5_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q5_K * x = (device const block_q5_K *) src0 + first_row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float sumf[2]={0.f};
-
-    const int step = sizeof(block_q5_K) * nb;
-
-#if QK_K == 256
-#
-    float yl[16], yh[16];
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = tiisg/4;
-    const int ix  = tiisg%4;
-    const int iq  = tid/4;
-    const int ir  = tid%4;
-    const int n   = 8;
-
-    const int l0 = n*ir;
-    const int q_offset = 32*iq + l0;
-    const int y_offset = 64*iq + l0;
-
-    const uint8_t hm1 = 1u << (2*iq);
-    const uint8_t hm2 = hm1 << 1;
-    const uint8_t hm3 = hm1 << 4;
-    const uint8_t hm4 = hm2 << 4;
-
-    uint16_t sc16[4];
-    thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
-
-    device const float * y1 = yy + ix*QK_K + y_offset;
-
-    for (int i = ix; i < nb; i += 4) {
-
-        device const uint8_t * q1 = x[i].qs + q_offset;
-        device const uint8_t * qh = x[i].qh + l0;
-        device const half * dh = &x[i].d;
-        device const uint16_t * a = (device const uint16_t *)x[i].scales + iq;
-
-        device const float * y2 = y1 + 128;
-        float4 sumy = {0.f, 0.f, 0.f, 0.f};
-        for (int l = 0; l < 8; ++l) {
-            yl[l+0] = y1[l+ 0]; sumy[0] += yl[l+0];
-            yl[l+8] = y1[l+32]; sumy[1] += yl[l+8];
-            yh[l+0] = y2[l+ 0]; sumy[2] += yh[l+0];
-            yh[l+8] = y2[l+32]; sumy[3] += yh[l+8];
-        }
-
-        for (int row = 0; row < 2; ++row) {
-
-            device const uint8_t * q2 = q1 + 64;
-
-            sc16[0] = a[0] & kmask1;
-            sc16[1] = a[2] & kmask1;
-            sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
-            sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
-
-            float4 acc1 = {0.f};
-            float4 acc2 = {0.f};
-            for (int l = 0; l < n; ++l) {
-                uint8_t h = qh[l];
-                acc1[0] += yl[l+0] * (q1[l] & 0x0F);
-                acc1[1] += yl[l+8] * (q1[l] & 0xF0);
-                acc1[2] += yh[l+0] * (q2[l] & 0x0F);
-                acc1[3] += yh[l+8] * (q2[l] & 0xF0);
-                acc2[0] += h & hm1 ? yl[l+0] : 0.f;
-                acc2[1] += h & hm2 ? yl[l+8] : 0.f;
-                acc2[2] += h & hm3 ? yh[l+0] : 0.f;
-                acc2[3] += h & hm4 ? yh[l+8] : 0.f;
-            }
-            const float dall = dh[0];
-            const float dmin = dh[1];
-            sumf[row] += dall * (sc8[0] * (acc1[0] +  16.f*acc2[0]) +
-                                 sc8[1] * (acc1[1]/16.f + 16.f*acc2[1]) +
-                                 sc8[4] * (acc1[2] +  16.f*acc2[2]) +
-                                 sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
-                         dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
-
-            q1 += step;
-            qh += step;
-            dh += step/2;
-            a  += step/2;
-
-        }
-
-        y1 += 4 * QK_K;
-
-    }
-#else
-    float yl[8], yh[8];
-
-    const int il = 4 * (tiisg/8);  // 0, 4, 8, 12
-    const int ix = tiisg%8;
-    const int iq = il/8;         // 0, 0, 1, 1
-    const int in = il%8;         // 0, 4, 0, 4
-
-    device const float * y = yy + ix*QK_K + il;
-
-    for (int i = ix; i < nb; i += 8) {
-
-        for (int l = 0; l < 4; ++l) {
-            yl[l+0] = y[l+ 0];
-            yl[l+4] = y[l+16];
-            yh[l+0] = y[l+32];
-            yh[l+4] = y[l+48];
-        }
-
-        device const half * dh = &x[i].d;
-        device const uint8_t * q = x[i].qs + il;
-        device const uint8_t * h = x[i].qh + in;
-        device const int8_t  * s = x[i].scales;
-
-        for (int row = 0; row < 2; ++row) {
-
-            const float d = dh[0];
-
-            float2 acc = {0.f, 0.f};
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t hl = h[l] >> iq;
-                acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
-                        + yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
-                acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
-                        + yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
-            }
-            sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
-
-            q += step;
-            h += step;
-            s += step;
-            dh += step/2;
-
-        }
-
-        y += 8 * QK_K;
-    }
-#endif
-
-    for (int row = 0; row < 2; ++row) {
-        const float tot = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_q5_K_f32")]]
-kernel void kernel_mul_mv_q5_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_q5_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
-}
-
-void kernel_mul_mv_q6_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const uint8_t kmask1 = 0x03;
-    const uint8_t kmask2 = 0x0C;
-    const uint8_t kmask3 = 0x30;
-    const uint8_t kmask4 = 0xC0;
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int     im = tgpig.z;
-
-    const int row = 2 * r0 + sgitg;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q6_K * x = (device const block_q6_K *) src0 + row * nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float sumf = 0;
-
-#if QK_K == 256
-    const int tid  = tiisg/2;
-    const int ix   = tiisg%2;
-    const int ip   = tid/8;         // 0 or 1
-    const int il   = tid%8;
-    const int n    = 4;
-    const int l0   = n*il;
-    const int is   = 8*ip + l0/16;
-
-    const int y_offset = 128*ip + l0;
-    const int q_offset_l = 64*ip + l0;
-    const int q_offset_h = 32*ip + l0;
-
-    for (int i = ix; i < nb; i += 2) {
-
-        device const uint8_t * q1 = x[i].ql + q_offset_l;
-        device const uint8_t * q2 = q1 + 32;
-        device const uint8_t * qh = x[i].qh + q_offset_h;
-        device const int8_t  * sc = x[i].scales + is;
-
-        device const float * y = yy + i * QK_K + y_offset;
-
-        const float dall = x[i].d;
-
-        float4 sums = {0.f, 0.f, 0.f, 0.f};
-        for (int l = 0; l < n; ++l) {
-            sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
-            sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
-            sums[2] += y[l+64] * ((int8_t)((q1[l]  >> 4) | ((qh[l] & kmask3) << 0)) - 32);
-            sums[3] += y[l+96] * ((int8_t)((q2[l]  >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
-        }
-
-        sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
-
-    }
-
-#else
-    const int ix  = tiisg/4;
-    const int il  = 4*(tiisg%4);
-
-    for (int i = ix; i < nb; i += 8) {
-        device const float * y = yy + i * QK_K + il;
-        device const uint8_t * ql = x[i].ql + il;
-        device const uint8_t * qh = x[i].qh + il;
-        device const int8_t  * s  = x[i].scales;
-
-        const float d = x[i].d;
-
-        float4 sums = {0.f, 0.f, 0.f, 0.f};
-        for (int l = 0; l < 4; ++l) {
-            sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
-            sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
-            sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >>  4) | ((qh[l] & kmask3) >> 0)) - 32);
-            sums[3] += y[l+48] * ((int8_t)((ql[l+16] >>  4) | ((qh[l] & kmask4) >> 2)) - 32);
-        }
-        sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]);
-    }
-
-#endif
-
-    const float tot = simd_sum(sumf);
-    if (tiisg == 0) {
-        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
-    }
-}
-
-[[host_name("kernel_mul_mv_q6_K_f32")]]
-kernel void kernel_mul_mv_q6_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
-}
-
-// ======================= "True" 2-bit
-
-void kernel_mul_mv_iq2_xxs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_iq2_xxs * x = (device const block_iq2_xxs *) src0 + ib_row + offset0;
-    device const float         * y = (device const float         *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
-    {
-        int nval = 4;
-        int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i];
-        nval = 2;
-        pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    const int ix = tiisg;
-
-    device const float * y4 = y + 32 * ix;
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
-        for (int i = 0; i < 32; ++i) {
-            yl[i] = y4[i];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq2_xxs * xr = x + ibl;
-        device const uint16_t * q2 = xr->qs + 4 * ib;
-        device const half * dh = &xr->d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            const float db = dh[0];
-            device const uint8_t * aux8 = (device const uint8_t *)q2;
-            const uint32_t aux32 = q2[2] | (q2[3] << 16);
-            const float d = db * (0.5f + (aux32 >> 28));
-
-            float sum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + aux8[l]);
-                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-                }
-            }
-            sumf[row] += d * sum;
-
-            dh += nb*sizeof(block_iq2_xxs)/2;
-            q2 += nb*sizeof(block_iq2_xxs)/2;
-        }
-
-        y4 += 32 * 32;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq2_xxs_f32")]]
-kernel void kernel_mul_mv_iq2_xxs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-}
-
-void kernel_mul_mv_iq2_xs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_iq2_xs * x = (device const block_iq2_xs *) src0 + ib_row + offset0;
-    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 512);
-    {
-        int nval = 8;
-        int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i];
-        nval = 2;
-        pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    const int ix = tiisg;
-
-    device const float * y4 = y + 32 * ix;
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
-        for (int i = 0; i < 32; ++i) {
-            yl[i] = y4[i];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq2_xs * xr = x + ibl;
-        device const uint16_t * q2 = xr->qs + 4 * ib;
-        device const uint8_t  * sc = xr->scales + ib;
-        device const half * dh = &xr->d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            const float db = dh[0];
-            const uint8_t ls1 = sc[0] & 0xf;
-            const uint8_t ls2 = sc[0] >>  4;
-            const float d1 = db * (0.5f + ls1);
-            const float d2 = db * (0.5f + ls2);
-
-            float sum1 = 0, sum2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
-                const uint8_t signs = shared_signs[(q2[l] >> 9)];
-                for (int j = 0; j < 8; ++j) {
-                    sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-                }
-            }
-            for (int l = 2; l < 4; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
-                const uint8_t signs = shared_signs[(q2[l] >> 9)];
-                for (int j = 0; j < 8; ++j) {
-                    sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-                }
-            }
-            sumf[row] += d1 * sum1 + d2 * sum2;
-
-            dh += nb*sizeof(block_iq2_xs)/2;
-            q2 += nb*sizeof(block_iq2_xs)/2;
-            sc += nb*sizeof(block_iq2_xs);
-        }
-
-        y4 += 32 * 32;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq2_xs_f32")]]
-kernel void kernel_mul_mv_iq2_xs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-}
-
-void kernel_mul_mv_iq3_xxs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_iq3_xxs * x = (device const block_iq3_xxs *) src0 + ib_row + offset0;
-    device const float         * y = (device const float         *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
-    {
-        int nval = 4;
-        int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq3xxs_grid[pos + i];
-        nval = 2;
-        pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    const int ix = tiisg;
-
-    device const float * y4 = y + 32 * ix;
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
-        for (int i = 0; i < 32; ++i) {
-            yl[i] = y4[i];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq3_xxs * xr = x + ibl;
-        device const uint8_t  * q3 = xr->qs + 8 * ib;
-        device const uint16_t * gas = (device const uint16_t *)(xr->qs + QK_K/4) + 2 * ib;
-        device const half * dh = &xr->d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            const float db = dh[0];
-            const uint32_t aux32 = gas[0] | (gas[1] << 16);
-            const float d = db * (0.5f + (aux32 >> 28));
-
-            float2 sum = {0};
-            for (int l = 0; l < 4; ++l) {
-                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + q3[2*l+0]);
-                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + q3[2*l+1]);
-                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sum[0] += yl[8*l + j + 0] * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
-                    sum[1] += yl[8*l + j + 4] * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
-                }
-            }
-            sumf[row] += d * (sum[0] + sum[1]);
-
-            dh  += nb*sizeof(block_iq3_xxs)/2;
-            q3  += nb*sizeof(block_iq3_xxs);
-            gas += nb*sizeof(block_iq3_xxs)/2;
-        }
-
-        y4 += 32 * 32;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq3_xxs_f32")]]
-kernel void kernel_mul_mv_iq3_xxs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-}
-
-void kernel_mul_mv_iq3_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_iq3_s * x = (device const block_iq3_s *) src0 + ib_row + offset0;
-    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
-    {
-        int nval = 8;
-        int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq3s_grid[pos + i];
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    const int ix = tiisg;
-
-    device const float * y4 = y + 32 * ix;
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
-        for (int i = 0; i < 32; ++i) {
-            yl[i] = y4[i];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq3_s * xr = x + ibl;
-        device const uint8_t * qs = xr->qs + 8 * ib;
-        device const uint8_t * qh = xr->qh + ib;
-        device const uint8_t * sc = xr->scales + (ib/2);
-        device const uint8_t * signs = xr->signs + 4 * ib;
-        device const half * dh = &xr->d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            const float db = dh[0];
-            const float d = db * (1 + 2*((sc[0] >> 4*(ib%2)) & 0xf));
-
-            float2 sum = {0};
-            for (int l = 0; l < 4; ++l) {
-                const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? values + 256 : values;
-                const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? values + 256 : values;
-                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]);
-                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]);
-                for (int j = 0; j < 4; ++j) {
-                    sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
-                    sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
-                }
-            }
-            sumf[row] += d * (sum[0] + sum[1]);
-
-            dh  += nb*sizeof(block_iq3_s)/2;
-            qs  += nb*sizeof(block_iq3_s);
-            qh  += nb*sizeof(block_iq3_s);
-            sc  += nb*sizeof(block_iq3_s);
-            signs += nb*sizeof(block_iq3_s);
-        }
-
-        y4 += 32 * 32;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq3_s_f32")]]
-kernel void kernel_mul_mv_iq3_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-}
-
-void kernel_mul_mv_iq2_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_iq2_s * x = (device const block_iq2_s *) src0 + ib_row + offset0;
-    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-    //threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
-    //{
-    //    int nval = 32;
-    //    int pos  = (32*sgitg + tiisg)*nval;
-    //    for (int i = 0; i < nval; ++i) values[pos + i] = iq2s_grid[pos + i];
-    //    threadgroup_barrier(mem_flags::mem_threadgroup);
-    //}
-
-    const int ix = tiisg;
-
-    device const float * y4 = y + 32 * ix;
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
-        for (int i = 0; i < 32; ++i) {
-            yl[i] = y4[i];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq2_s * xr = x + ibl;
-        device const uint8_t * qs = xr->qs + 4 * ib;
-        device const uint8_t * qh = xr->qh + ib;
-        device const uint8_t * sc = xr->scales + ib;
-        device const uint8_t * signs = qs + QK_K/8;
-        device const half * dh = &xr->d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            const float db = dh[0];
-            const float d1 = db * (0.5f + (sc[0] & 0xf));
-            const float d2 = db * (0.5f + (sc[0] >>  4));
-
-            float2 sum = {0};
-            for (int l = 0; l < 2; ++l) {
-                //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
-                //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
-                constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
-                constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sum[0] += yl[8*l + j +  0] * grid1[j] * select(1, -1, signs[l+0] & kmask_iq2xs[j]);
-                    sum[1] += yl[8*l + j + 16] * grid2[j] * select(1, -1, signs[l+2] & kmask_iq2xs[j]);
-                }
-            }
-            sumf[row] += d1 * sum[0] + d2 * sum[1];
-
-            dh  += nb*sizeof(block_iq2_s)/2;
-            qs  += nb*sizeof(block_iq2_s);
-            qh  += nb*sizeof(block_iq2_s);
-            sc  += nb*sizeof(block_iq2_s);
-            signs += nb*sizeof(block_iq2_s);
-        }
-
-        y4 += 32 * 32;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq2_s_f32")]]
-kernel void kernel_mul_mv_iq2_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq2_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-}
-
-void kernel_mul_mv_iq1_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_value,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-    device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
-    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-    const int ix = tiisg;
-
-    device const float * y4 = y + 32 * ix;
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
-        float sumy = 0;
-        for (int i = 0; i < 32; ++i) {
-            yl[i] = y4[i];
-            sumy += yl[i];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq1_s * xr = x + ibl;
-        device const uint8_t  * qs = xr->qs + 4 * ib;
-        device const uint16_t * qh = xr->qh + ib;
-        device const half     * dh = &xr->d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
-            constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 5) & 0x700)));
-            constant uint8_t * grid3 = (constant uint8_t *)(iq1s_grid_gpu + (qs[2] | ((qh[0] << 2) & 0x700)));
-            constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[0] >> 1) & 0x700)));
-
-            float sum = 0;
-            for (int j = 0; j < 4; ++j) {
-                sum += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4)
-                     + yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4)
-                     + yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
-                     + yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
-            }
-            sumf[row] += (float)dh[0] * (sum + sumy * (qh[0] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA)) * (2*((qh[0] >> 12) & 7) + 1);
-
-            dh += nb*sizeof(block_iq1_s)/2;
-            qs += nb*sizeof(block_iq1_s);
-            qh += nb*sizeof(block_iq1_s)/2;
-        }
-
-        y4 += 32 * 32;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-
-void kernel_mul_mv_iq1_m_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_value,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-    device const block_iq1_m * x = (device const block_iq1_m *) src0 + ib_row + offset0;
-    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[32];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-    const int ix = tiisg;
-
-    device const float * y4 = y + 32 * ix;
-
-#if QK_K != 64
-    iq1m_scale_t scale;
-#endif
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
-        float4 sumy = {0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
-            yl[i+ 8] = y4[i+ 8]; sumy[1] += yl[i+ 8];
-            yl[i+16] = y4[i+16]; sumy[2] += yl[i+16];
-            yl[i+24] = y4[i+24]; sumy[3] += yl[i+24];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq1_m * xr = x + ibl;
-        device const uint8_t  * qs = xr->qs + 4 * ib;
-        device const uint8_t  * qh = xr->qh + 2 * ib;
-        device const uint16_t * sc = (device const uint16_t *)xr->scales;
-
-        for (int row = 0; row < N_DST; row++) {
-
-#if QK_K != 64
-            scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-#endif
-
-            constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
-            constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
-            constant uint8_t * grid3 = (constant uint8_t *)(iq1s_grid_gpu + (qs[2] | ((qh[1] << 8) & 0x700)));
-            constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[1] << 4) & 0x700)));
-
-            float2 sum = {0.f};
-            for (int j = 0; j < 4; ++j) {
-                sum[0] += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4)
-                        + yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4);
-                sum[1] += yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
-                        + yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
-            }
-            const float delta1 = sumy[0] * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[1] * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
-            const float delta2 = sumy[2] * (qh[1] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[3] * (qh[1] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
-#if QK_K == 64
-            const float d = (float) *((device const half *)(sc - 1));
-            sumf[row] += d * ((sum[0] + delta1) * (2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1) +
-                              (sum[1] + delta2) * (2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1));
-#else
-            sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
-                                             (sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
-#endif
-
-            sc += nb*sizeof(block_iq1_m)/2;
-            qs += nb*sizeof(block_iq1_m);
-            qh += nb*sizeof(block_iq1_m);
-        }
-
-        y4 += 32 * 32;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-
-void kernel_mul_mv_iq4_nl_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values_i8,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
-    const int nb = ne00/QK4_NL;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-    const int first_row = (r0 * 2 + sgitg) * 2;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-    device const block_iq4_nl * x = (device const block_iq4_nl *) src0 + ib_row + offset0;
-    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
-
-    const int ix = tiisg/2;  // 0...15
-    const int it = tiisg%2;  // 0 or 1
-
-    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    float4 yl[4];
-    float sumf[2]={0.f}, all_sum;
-
-    device const float * yb = y + ix * QK4_NL + it * 8;
-
-    uint32_t aux32[2];
-    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
-
-    float4 qf1, qf2;
-
-    for (int ib = ix; ib < nb; ib += 16) {
-
-        device const float4 * y4 = (device const float4 *)yb;
-        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
-
-        for (int row = 0; row < 2; ++row) {
-
-            device const block_iq4_nl & xb = x[row*nb + ib];
-            device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
-
-            float4 acc1 = {0.f}, acc2 = {0.f};
-
-            aux32[0] = q4[0] | (q4[1] << 16);
-            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
-            aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
-            acc1 += yl[0] * qf1;
-            acc2 += yl[1] * qf2;
-
-            aux32[0] = q4[2] | (q4[3] << 16);
-            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
-            aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
-            acc1 += yl[2] * qf1;
-            acc2 += yl[3] * qf2;
-
-            acc1 += acc2;
-
-            sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
-
-        }
-
-        yb += 16 * QK4_NL;
-    }
-
-    for (int row = 0; row < 2; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-
-#if QK_K != 64
-void kernel_mul_mv_iq4_xs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values_i8,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-    const int first_row = (r0 * 2 + sgitg) * 2;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-    device const block_iq4_xs * x = (device const block_iq4_xs *) src0 + ib_row + offset0;
-    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
-
-    const int ix = tiisg/16;  // 0 or 1
-    const int it = tiisg%16;  // 0...15
-    const int ib = it/2;
-    const int il = it%2;
-
-    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    float4 yl[4];
-    float sumf[2]={0.f}, all_sum;
-
-    device const float * yb = y + ix * QK_K + ib * 32 + il * 8;
-
-    uint32_t aux32[2];
-    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
-
-    float4 qf1, qf2;
-
-    for (int ibl = ix; ibl < nb; ibl += 2) {
-
-        device const float4 * y4 = (device const float4 *)yb;
-        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
-
-        for (int row = 0; row < 2; ++row) {
-
-            device const block_iq4_xs & xb = x[row*nb + ibl];
-            device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
-
-            float4 acc1 = {0.f}, acc2 = {0.f};
-
-            aux32[0] = q4[0] & 0x0f0f0f0f;
-            aux32[1] = (q4[0] >> 4) & 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
-            acc1 += yl[0] * qf1;
-            acc2 += yl[1] * qf2;
-
-            aux32[0] = q4[1] & 0x0f0f0f0f;
-            aux32[1] = (q4[1] >> 4) & 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
-            acc1 += yl[2] * qf1;
-            acc2 += yl[3] * qf2;
-
-            acc1 += acc2;
-
-            const int ls = (((xb.scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((xb.scales_h >> 2*ib) & 3) << 4)) - 32;
-            sumf[row] += (float)xb.d * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
-
-        }
-
-        yb += 2 * QK_K;
-    }
-
-    for (int row = 0; row < 2; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-#endif
-
-[[host_name("kernel_mul_mv_iq1_s_f32")]]
-kernel void kernel_mul_mv_iq1_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
-}
-
-[[host_name("kernel_mul_mv_iq1_m_f32")]]
-kernel void kernel_mul_mv_iq1_m_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq1_m_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
-}
-
-[[host_name("kernel_mul_mv_iq4_nl_f32")]]
-kernel void kernel_mul_mv_iq4_nl_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-}
-
-[[host_name("kernel_mul_mv_iq4_xs_f32")]]
-kernel void kernel_mul_mv_iq4_xs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-
-#if QK_K == 64
-    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-#else
-    kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-#endif
-}
-
-//============================= templates and their specializations =============================
-
-// NOTE: this is not dequantizing - we are simply fitting the template
-template <typename type4x4>
-void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
-    float4x4 temp = *(((device float4x4 *)src));
-    for (int i = 0; i < 16; i++){
-        reg[i/4][i%4] = temp[i/4][i%4];
-    }
-}
-
-template <typename type4x4>
-void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
-    half4x4 temp = *(((device half4x4 *)src));
-    for (int i = 0; i < 16; i++){
-        reg[i/4][i%4] = temp[i/4][i%4];
-    }
-}
-
-template <typename type4x4>
-void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
-    const float d1 = il ? (xb->d / 16.h) : xb->d;
-    const float d2 = d1 / 256.f;
-    const float md = -8.h * xb->d;
-    const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = mask0 << 8;
-
-    for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
-        reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
-    }
-}
-
-template <typename type4x4>
-void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
-    const float d1 = il ? (xb->d / 16.h) : xb->d;
-    const float d2 = d1 / 256.f;
-    const float  m = xb->m;
-    const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = mask0 << 8;
-
-    for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
-        reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
-    }
-}
-
-template <typename type4x4>
-void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
-    const float d = xb->d;
-    const float md = -16.h * xb->d;
-    const ushort mask = il ? 0x00F0 : 0x000F;
-
-    const uint32_t qh = *((device const uint32_t *)xb->qh);
-
-    const int x_mv = il ? 4 : 0;
-
-    const int gh_mv = il ? 12 : 0;
-    const int gh_bk = il ?  0 : 4;
-
-    for (int i = 0; i < 8; i++) {
-        // extract the 5-th bits for x0 and x1
-        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
-        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
-
-        // combine the 4-bits from qs with the 5th bit
-        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
-        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
-
-        reg[i/2][2*(i%2)+0] = d * x0 + md;
-        reg[i/2][2*(i%2)+1] = d * x1 + md;
-    }
-}
-
-template <typename type4x4>
-void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
-    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
-    const float d = xb->d;
-    const float m = xb->m;
-    const ushort mask = il ? 0x00F0 : 0x000F;
-
-    const uint32_t qh = *((device const uint32_t *)xb->qh);
-
-    const int x_mv = il ? 4 : 0;
-
-    const int gh_mv = il ? 12 : 0;
-    const int gh_bk = il ?  0 : 4;
-
-    for (int i = 0; i < 8; i++) {
-        // extract the 5-th bits for x0 and x1
-        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
-        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
-
-        // combine the 4-bits from qs with the 5th bit
-        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
-        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
-
-        reg[i/2][2*(i%2)+0] = d * x0 + m;
-        reg[i/2][2*(i%2)+1] = d * x1 + m;
-    }
-}
-
-template <typename type4x4>
-void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
-    device const int8_t * qs = ((device const int8_t *)xb->qs);
-    const half d = xb->d;
-
-    for (int i = 0; i < 16; i++) {
-        reg[i/4][i%4] = (qs[i + 16*il] * d);
-    }
-}
-
-template <typename type4x4>
-void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
-    const float d = xb->d;
-    const float min = xb->dmin;
-    device const uint8_t * q = (device const uint8_t *)xb->qs;
-    float dl, ml;
-    uint8_t sc = xb->scales[il];
-
-#if QK_K == 256
-    q = q + 32*(il/8) + 16*(il&1);
-    il = (il/2)%4;
-#endif
-    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
-    }
-}
-
-template <typename type4x4>
-void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
-    const half d_all = xb->d;
-    device const uint8_t * q = (device const uint8_t *)xb->qs;
-    device const uint8_t * h = (device const uint8_t *)xb->hmask;
-    device const int8_t * scales = (device const int8_t *)xb->scales;
-
-#if QK_K == 256
-    q = q + 32 * (il/8) + 16 * (il&1);
-    h = h + 16 * (il&1);
-    uint8_t m = 1 << (il/2);
-    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
-                                 ((il/4)>0 ? 12  : 3);
-    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
-    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
-    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
-                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
-    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
-    const float ml = 4.f * dl;
-
-    il = (il/2) & 3;
-    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    dl *= coef;
-
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
-    }
-#else
-    float    kcoef = il&1 ? 1.f/16.f : 1.f;
-    uint16_t kmask = il&1 ? 0xF0     : 0x0F;
-    float    dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
-    float    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uint8_t  mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    uint8_t  m = 1<<(il*2);
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
-    }
-#endif
-}
-
-static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
-    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
-                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
-}
-
-template <typename type4x4>
-void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
-    device const uchar * q = xb->qs;
-
-#if QK_K == 256
-    short is = (il/4) * 2;
-    q = q + (il/4) * 32 + 16 * (il&1);
-    il = il & 3;
-    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d   = il < 2 ? xb->d : xb->d / 16.h;
-    const float min = xb->dmin;
-    const float dl = d * sc[0];
-    const float ml = min * sc[1];
-#else
-    (void) get_scale_min_k4_just2;
-
-    q = q + 16 * (il&1);
-    device const uint8_t * s = xb->scales;
-    device const half2 * dh = (device const half2 *)xb->d;
-    const float2 d = (float2)dh[0];
-    const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
-    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1] * (s[1]>>4);
-#endif
-    const ushort mask = il<2 ? 0x0F : 0xF0;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
-    }
-}
-
-template <typename type4x4>
-void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
-    device const uint8_t * q  = xb->qs;
-    device const uint8_t * qh = xb->qh;
-
-#if QK_K == 256
-    short is = (il/4) * 2;
-    q  = q + 32 * (il/4) + 16 * (il&1);
-    qh = qh + 16 * (il&1);
-    uint8_t ul = 1 << (il/2);
-    il = il & 3;
-    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d = il < 2 ? xb->d : xb->d / 16.f;
-    const float min = xb->dmin;
-    const float dl = d * sc[0];
-    const float ml = min * sc[1];
-
-    const ushort mask  = il<2 ? 0x0F : 0xF0;
-    const float qh_val = il<2 ? 16.f : 256.f;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
-    }
-#else
-    q = q + 16 * (il&1);
-    device const int8_t * s = xb->scales;
-    const float dl = xb->d * s[il];
-    uint8_t m = 1<<(il*2);
-    const float  coef = il<2 ? 1.f  : 1.f/16.f;
-    const ushort mask = il<2 ? 0x0F : 0xF0;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
-    }
-#endif
-}
-
-template <typename type4x4>
-void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
-    const half d_all = xb->d;
-    device const uint8_t * ql = (device const uint8_t *)xb->ql;
-    device const uint8_t * qh = (device const uint8_t *)xb->qh;
-    device const int8_t * scales = (device const int8_t *)xb->scales;
-
-#if QK_K == 256
-    ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
-    qh = qh + 32*(il/8) + 16*(il&1);
-    float sc = scales[(il%2) + 2 * ((il/2))];
-    il = (il/2) & 3;
-#else
-    ql = ql + 16 * (il&1);
-    float sc = scales[il];
-#endif
-    const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
-    const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
-    const float       coef = il>1 ? 1.f/16.f          : 1.f;
-    const float ml = d_all * sc * 32.f;
-    const float dl = d_all * sc * coef;
-    for (int i = 0; i < 16; ++i) {
-        const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
-                            : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
-        reg[i/4][i%4] = dl * q - ml;
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
-    device const uint16_t * q2 = xb->qs + 4*ib32;
-    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
-    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
-    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
-    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
-    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
-    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
-    for (int i = 0; i < 8; ++i) {
-        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
-    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
-    for (int i = 0; i < 8; ++i) {
-        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    device const uint16_t * q2 = xb->qs + 4*ib32;
-    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
-    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
-    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
-    for (int i = 0; i < 8; ++i) {
-        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
-    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
-    for (int i = 0; i < 8; ++i) {
-        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    device const uint8_t * q3 = xb->qs + 8*ib32;
-    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
-    const uint32_t aux32 = gas[0] | (gas[1] << 16);
-    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
-    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
-    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
-    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
-    for (int i = 0; i < 4; ++i) {
-        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
-        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
-    }
-    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
-    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
-    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
-    for (int i = 0; i < 4; ++i) {
-        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
-        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    device const uint8_t * qs = xb->qs + 8*ib32;
-    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
-    const uint8_t qh = xb->qh[ib32] >> 4*il;
-    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
-    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
-    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
-    for (int i = 0; i < 4; ++i) {
-        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
-        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
-    }
-    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
-    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
-    for (int i = 0; i < 4; ++i) {
-        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
-        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
-    device const uint8_t * signs = qs + QK_K/8;
-    const uint8_t qh = xb->qh[ib32] >> 4*il;
-    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
-    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
-    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
-    for (int i = 0; i < 8; ++i) {
-        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
-        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const int ib32 = il/2;
-    il = il%2;
-    const float d = xb->d;
-    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
-    device const uint16_t * qh = xb->qh;
-    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
-    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
-    const uint16_t h = qh[ib32] >> 6*il;
-    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
-    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
-    for (int i = 0; i < 4; ++i) {
-        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
-        reg[1][i] = dl * (grid1[i] >>  4) + ml;
-        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
-        reg[3][i] = dl * (grid2[i] >>  4) + ml;
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const int ib32 = il/2;
-    il = il%2;
-    device const uint16_t * sc = (device const uint16_t *)xb->scales;
-#if QK_K == 64
-    const float d = xb->d;
-#else
-    iq1m_scale_t scale;
-    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-    const float d = scale.f16;
-#endif
-    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
-    device const uint8_t * qh = xb->qh + 2*ib32 + il;
-#if QK_K == 64
-    const float dl  = d * (2*((sc[ib32/2] >> (8*(ib32%2)+4*il)) & 0xf) + 1);
-#else
-    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
-#endif
-    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
-    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
-    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
-    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
-    for (int i = 0; i < 4; ++i) {
-        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
-        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
-        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
-        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
-    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
-    const float d = xb->d;
-    uint32_t aux32;
-    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
-    for (int i = 0; i < 4; ++i) {
-        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
-        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
-        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
-        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
-        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
-#if QK_K == 64
-    dequantize_iq4_nl(xb, il, reg);
-#else
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const int ib32 = il/2;
-    il = il%2;
-    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
-    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
-    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
-    const float d = (float)xb->d * (ls - 32);
-    uint32_t aux32;
-    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
-    for (int i = 0; i < 4; ++i) {
-        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
-        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
-        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
-        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
-        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
-    }
-#endif
-}
-
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
-kernel void kernel_get_rows(
-        device const  void * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    //const int64_t i = tgpig;
-    //const int64_t r = ((device int32_t *) src1)[i];
-
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
-
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
-
-    const int64_t i02 = i11;
-
-    for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) {
-        float4x4 temp;
-        dequantize_func(
-            ((device const block_q *) ((device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp);
-        *(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
-    }
-}
-
-kernel void kernel_get_rows_f32(
-        device const  void * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
-
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
-
-    const int64_t i02 = i11;
-
-    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
-        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((device float *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
-    }
-}
-
-kernel void kernel_get_rows_f16(
-        device const  void * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
-
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
-
-    const int64_t i02 = i11;
-
-    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
-        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((device half *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
-    }
-}
-
-kernel void kernel_get_rows_i32(
-        device const  void * src0,
-        device const  char * src1,
-        device     int32_t * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int64_t & ne10,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        uint3                tgpig[[threadgroup_position_in_grid]],
-        uint                 tiitg[[thread_index_in_threadgroup]],
-        uint3                tptg [[threads_per_threadgroup]]) {
-    const int64_t i10 = tgpig.x;
-    const int64_t i11 = tgpig.y;
-
-    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
-
-    const int64_t i02 = i11;
-
-    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
-        ((device int32_t *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((device int32_t *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
-    }
-}
-
-
-#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
-#define BLOCK_SIZE_K 32
-#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
-#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
-#define THREAD_PER_BLOCK 128
-#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
-#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
-#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
-#define SG_MAT_ROW 8
-
-// each block_q contains 16*nl weights
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-void kernel_mul_mm_impl(device const  uchar * src0,
-                        device const  uchar * src1,
-                        device        float * dst,
-                        constant    int64_t & ne00,
-                        constant    int64_t & ne02,
-                        constant   uint64_t & nb01,
-                        constant   uint64_t & nb02,
-                        constant    int64_t & ne12,
-                        constant   uint64_t & nb10,
-                        constant   uint64_t & nb11,
-                        constant   uint64_t & nb12,
-                        constant    int64_t & ne0,
-                        constant    int64_t & ne1,
-                        constant       uint & r2,
-                        constant       uint & r3,
-                        threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                        uint3                 tgpig[[threadgroup_position_in_grid]],
-                        uint                  tiitg[[thread_index_in_threadgroup]],
-                        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
-
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-    const uint im = tgpig.z;
-
-    // if this block is of 64x32 shape or smaller
-    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
-
-    // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
-
-    simdgroup_half8x8  ma[4];
-    simdgroup_float8x8 mb[2];
-    simdgroup_float8x8 c_res[8];
-    for (int i = 0; i < 8; i++){
-        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
-    }
-
-    short il = (tiitg % THREAD_PER_ROW);
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
-    ushort offset1 = il/nl;
-
-    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
-    device const float   * y = (device const float   *)(src1
-        + nb12 * im
-        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
-        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
-
-    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
-        // load data and store to threadgroup memory
-        half4x4 temp_a;
-        dequantize_func(x, il, temp_a);
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        #pragma unroll(16)
-        for (int i = 0; i < 16; i++) {
-            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
-            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
-        }
-
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
-
-        il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
-        y += BLOCK_SIZE_K;
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // load matrices from threadgroup memory and conduct outer products
-        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
-
-        #pragma unroll(4)
-        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
-            #pragma unroll(4)
-            for (int i = 0; i < 4; i++) {
-                simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
-            }
-            simdgroup_barrier(mem_flags::mem_none);
-            #pragma unroll(2)
-            for (int i = 0; i < 2; i++) {
-                simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
-            }
-
-            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
-
-            #pragma unroll(8)
-            for (int i = 0; i < 8; i++){
-                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
-            }
-        }
-    }
-
-    if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
-        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
-                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
-        for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
-        }
-    } else {
-        // block is smaller than 64x32, we should avoid writing data outside of the matrix
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-        for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
-        if (sgitg == 0) {
-            for (int i = 0; i < n_rows; i++) {
-                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                    *(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
-                }
-            }
-        }
-    }
-}
-
-// same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in rowids
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-void kernel_mul_mm_id_impl(
-        device const  uchar * src0,
-        device const  uchar * src1,
-        threadgroup ushort2 * rowids,
-        device        float * dst,
-        constant    int64_t & ne00,
-        constant    int64_t & ne02,
-        constant   uint64_t & nb01,
-        constant   uint64_t & nb02,
-        constant    int64_t & ne11,
-        constant    int64_t & ne12,
-        constant   uint64_t & nb10,
-        constant   uint64_t & nb11,
-        constant   uint64_t & nb12,
-        constant    int64_t & ne0,
-                    int64_t   ne1,
-                    int64_t   ne0ne1,
-        threadgroup   uchar * shared_memory,
-        uint3                 tgpig[[threadgroup_position_in_grid]],
-        uint                  tiitg[[thread_index_in_threadgroup]],
-        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
-
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-
-    if (r1 * BLOCK_SIZE_N >= ne1) return;
-
-    // if this block is of 64x32 shape or smaller
-    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
-
-    // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
-
-    simdgroup_half8x8  ma[4];
-    simdgroup_float8x8 mb[2];
-    simdgroup_float8x8 c_res[8];
-    for (int i = 0; i < 8; i++){
-        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
-    }
-    short il = (tiitg % THREAD_PER_ROW);
-
-    ushort offset1 = il/nl;
-
-    threadgroup const auto & id = rowids[r1 * BLOCK_SIZE_N + thread_col];
-
-    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01) + offset1;
-    device const float   * y = (device const float   *)(src1
-        + nb12 * id[1]
-        + nb11 * (id[0] % ne11)
-        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
-
-    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
-        // load data and store to threadgroup memory
-        half4x4 temp_a;
-        dequantize_func(x, il, temp_a);
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        for (int i = 0; i < 16; i++) {
-            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
-            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
-        }
-
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
-
-        il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
-        y += BLOCK_SIZE_K;
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // load matrices from threadgroup memory and conduct outer products
-        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
-
-        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
-            for (int i = 0; i < 4; i++) {
-                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
-            }
-            simdgroup_barrier(mem_flags::mem_none);
-            for (int i = 0; i < 2; i++) {
-                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
-            }
-
-            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
-
-            for (int i = 0; i < 8; i++){
-                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
-            }
-        }
-    }
-
-    {
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-        for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        device float * C = dst + (BLOCK_SIZE_M * r0);
-        if (sgitg == 0) {
-            for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                threadgroup const auto & jid = rowids[r1 * BLOCK_SIZE_N + j];
-                int joff =  jid[0] * ne0 + jid[1] * ne0ne1;
-                for (int i = 0; i < n_rows; i++) {
-                    *(C + i + joff) = *(temp_str + i + j * BLOCK_SIZE_M);
-                }
-            }
-        }
-    }
-}
-
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-kernel void kernel_mul_mm(device const  uchar * src0,
-                          device const  uchar * src1,
-                          device        float * dst,
-                          constant    int64_t & ne00,
-                          constant    int64_t & ne02,
-                          constant   uint64_t & nb01,
-                          constant   uint64_t & nb02,
-                          constant    int64_t & ne12,
-                          constant   uint64_t & nb10,
-                          constant   uint64_t & nb11,
-                          constant   uint64_t & nb12,
-                          constant    int64_t & ne0,
-                          constant    int64_t & ne1,
-                          constant       uint & r2,
-                          constant       uint & r3,
-                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                          uint3                 tgpig[[threadgroup_position_in_grid]],
-                          uint                  tiitg[[thread_index_in_threadgroup]],
-                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-    kernel_mul_mm_impl<block_q, nl, dequantize_func>(
-        src0,
-        src1,
-        dst,
-        ne00,
-        ne02,
-        nb01,
-        nb02,
-        ne12,
-        nb10,
-        nb11,
-        nb12,
-        ne0,
-        ne1,
-        r2,
-        r3,
-        shared_memory,
-        tgpig,
-        tiitg,
-        sgitg);
-}
-
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-kernel void kernel_mul_mm_id(
-        device const   uchar * src0s,
-        device const   uchar * src1,
-        device         float * dst,
-        device const   uchar * ids,
-        constant     int64_t & nei0,
-        constant     int64_t & nei1,
-        constant    uint64_t & nbi1,
-        constant     int64_t & ne00,
-        constant     int64_t & ne02,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant     int64_t & ne11,
-        constant     int64_t & ne12,
-        constant     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        threadgroup    uchar * shared_memory [[threadgroup(0)]],
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int32_t i02 = tgpig.z;
-    tgpig.z = 0;
-
-    device const uchar * src0 = src0s + i02*nb02;
-
-    // row indices
-    threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shared_memory + 8192);
-
-    // TODO: parallelize this loop
-    int64_t _ne1 = 0;
-    for (ushort ii1 = 0; ii1 < nei1; ii1++) {
-        for (ushort ii0 = 0; ii0 < nei0; ii0++) {
-            int32_t id = ((device int32_t *) (ids + ii1*nbi1))[ii0];
-            if (id == i02) {
-                //if (tiitg == 0) {
-                    rowids[_ne1] = ushort2(ii0, ii1);
-                //}
-                _ne1++;
-            }
-        }
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
-        src0,
-        src1,
-        rowids,
-        dst,
-        ne00,
-        ne02,
-        nb01,
-        nb02,
-        ne11,
-        ne12,
-        nb10,
-        nb11,
-        nb12,
-        ne0,
-        _ne1,
-        ne0*ne1,
-        shared_memory,
-        tgpig,
-        tiitg,
-        sgitg);
-}
-
-#if QK_K == 256
-#define QK_NL 16
-#else
-#define QK_NL 4
-#endif
-
-//
-// get rows
-//
-
-typedef void (get_rows_t)(
-        device const void * src0,
-        device const char * src1,
-        device      float * dst,
-        constant  int64_t & ne00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant  int64_t & ne10,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        uint3, uint, uint3);
-
-//template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
-//template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
-template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
-template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
-template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
-template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
-template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
-template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_t kernel_get_rows<block_iq3_s,   QK_NL, dequantize_iq3_s>;
-template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_t kernel_get_rows<block_iq2_s,   QK_NL, dequantize_iq2_s>;
-template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_t kernel_get_rows<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_t kernel_get_rows<block_iq1_m,   QK_NL, dequantize_iq1_m>;
-template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_t kernel_get_rows<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  2,     dequantize_iq4_xs>;
-#else
-template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
-
-//
-// matrix-matrix multiplication
-//
-
-typedef decltype(kernel_mul_mm<float4x4, 1, dequantize_f32>) mat_mm_t;
-
-template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mat_mm_t kernel_mul_mm<float4x4,      1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mat_mm_t kernel_mul_mm<half4x4,       1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_0,    2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_1,    2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_0,    2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_1,    2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q8_0,    2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q2_K,    QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q3_K,    QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_K,    QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_K,    QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q6_K,    QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq3_s,   QK_NL, dequantize_iq3_s>;
-template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq2_s,   QK_NL, dequantize_iq2_s>;
-template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_m,   QK_NL, dequantize_iq1_m>;
-template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_xs>;
-#else
-template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
-
-//
-// indirect matrix-matrix multiplication
-//
-
-typedef decltype(kernel_mul_mm_id<float4x4, 1, dequantize_f32>) mat_mm_id_t;
-
-template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<float4x4,      1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<half4x4,       1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0,    2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1,    2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0,    2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1,    2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0,    2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K,    QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K,    QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K,    QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K,    QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K,    QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_s,   QK_NL, dequantize_iq3_s>;
-template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_s,   QK_NL, dequantize_iq2_s>;
-template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_m,   QK_NL, dequantize_iq1_m>;
-template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  2,     dequantize_iq4_xs>;
-#else
-template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
-
-//
-// matrix-vector multiplication
-//
-
-typedef void (kernel_mul_mv_impl_t)(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb00,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                   int64_t   ne10,
-                   int64_t   ne11,
-                   int64_t   ne12,
-                  uint64_t   nb10,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-                   uint3     tgpig,
-                   uint      tiisg);
-
-typedef void (kernel_mul_mv2_impl_t)(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg);
-
-template<kernel_mul_mv_impl_t impl_fn>
-void mmv_fn(
-        device const    char * src0,
-        device const    char * src1,
-        device         float * dst,
-                     int64_t   ne00,
-                     int64_t   ne01,
-                     int64_t   ne02,
-                    uint64_t   nb00,
-                    uint64_t   nb01,
-                    uint64_t   nb02,
-                     int64_t   ne10,
-                     int64_t   ne11,
-                     int64_t   ne12,
-                     int64_t   ne13,
-                    uint64_t   nb10,
-                    uint64_t   nb11,
-                    uint64_t   nb12,
-                     int64_t   ne0,
-                     int64_t   ne1,
-                    uint64_t   nb1,
-                        uint   r2,
-                        uint   r3,
-        threadgroup int8_t   * shared_values,
-        uint3                  tgpig,
-        uint                   tiitg,
-        uint                   tiisg,
-        uint                   sgitg) {
-    impl_fn(src0,src1,dst,ne00,ne01,ne02,nb00,nb01,nb02,ne10,ne11,ne12,nb10,nb11,nb12,ne0,ne1,r2,r3,tgpig,tiisg);
-}
-
-template<kernel_mul_mv2_impl_t impl_fn>
-void mmv_fn(
-        device const    char * src0,
-        device const    char * src1,
-        device         float * dst,
-                     int64_t   ne00,
-                     int64_t   ne01,
-                     int64_t   ne02,
-                    uint64_t   nb00,
-                    uint64_t   nb01,
-                    uint64_t   nb02,
-                     int64_t   ne10,
-                     int64_t   ne11,
-                     int64_t   ne12,
-                     int64_t   ne13,
-                    uint64_t   nb10,
-                    uint64_t   nb11,
-                    uint64_t   nb12,
-                     int64_t   ne0,
-                     int64_t   ne1,
-                    uint64_t   nb1,
-                        uint   r2,
-                        uint   r3,
-        threadgroup int8_t   * shared_values,
-        uint3                  tgpig,
-        uint                   tiitg,
-        uint                   tiisg,
-        uint                   sgitg) {
-    impl_fn(src0,(const device float *)src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,shared_values,tgpig,tiisg,sgitg);
-}
-
-typedef decltype(mmv_fn<kernel_mul_mv_f32_f32_impl>) mul_mv_impl_fn_t;
-
-template<mul_mv_impl_fn_t impl_fn>
-kernel void kernel_mul_mv_id(
-        device const    char * src0s,
-        device const    char * src1,
-        device         float * dst,
-        device const    char * ids,
-        constant     int64_t & nei0,
-        constant     int64_t & nei1,
-        constant    uint64_t & nbi1,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant     int64_t & ne10,
-        constant     int64_t & ne11,
-        constant     int64_t & ne12,
-        constant     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        threadgroup int8_t   * shared_values [[threadgroup(0)]],
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   tiisg[[thread_index_in_simdgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
-    const int iid1 = tgpig.z/nei0;
-    const int idx = tgpig.z%nei0;
-
-    tgpig.z = 0;
-
-    const int32_t i02 = ((device const int32_t *) (ids + iid1*nbi1))[idx];
-
-    const int64_t i11 = idx % ne11;
-    const int64_t i12 = iid1;
-
-    const int64_t i1 = idx;
-    const int64_t i2 = i12;
-
-    device const char * src0_cur = src0s + i02*nb02;
-    device const char * src1_cur = src1 + i11*nb11 + i12*nb12;
-    device      float * dst_cur  = dst + i1*ne0 + i2*ne1*ne0;
-
-    impl_fn(
-        /* src0 */ src0_cur,
-        /* src1 */ src1_cur,
-        /* dst  */ dst_cur,
-        /* ne00 */ ne00,
-        /* ne01 */ ne01,
-        /* ne02 */ 1,//ne02,
-        /* nb00 */ nb00,
-        /* nb01 */ nb01,
-        /* nb02 */ nb02,
-        /* ne10 */ ne10,
-        /* ne11 */ 1,//ne11,
-        /* ne12 */ 1,//ne12,
-        /* ne13 */ 1,//ne13,
-        /* nb10 */ nb10,
-        /* nb11 */ nb11,
-        /* nb12 */ nb12,
-        /* ne0  */ ne0,
-        /* ne1  */ 1,//ne1,
-        /* nb1  */ nb1,
-        /* r2   */ 1,
-        /* r3   */ 1,
-        shared_values,
-        tgpig,
-        tiitg,
-        tiisg,
-        sgitg);
-}
-
-typedef decltype(kernel_mul_mv_id<mmv_fn<kernel_mul_mv_f32_f32_impl>>) kernel_mul_mv_id_t;
-
-template [[host_name("kernel_mul_mv_id_f32_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_f32_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_f16_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_f16_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_q8_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q8_0_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_q4_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
-template [[host_name("kernel_mul_mv_id_q4_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
-template [[host_name("kernel_mul_mv_id_q5_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
-template [[host_name("kernel_mul_mv_id_q5_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
-template [[host_name("kernel_mul_mv_id_q2_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q2_K_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_q3_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q3_K_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_q4_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q4_K_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_q5_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q5_K_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_q6_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q6_K_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq1_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq1_s_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq1_m_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq1_m_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq2_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xxs_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq2_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xs_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_xxs_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq3_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_s_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq2_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl>>;
-template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl>>;
-#if QK_K != 64
-template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl>>;
-#endif
-

llama/ggml-metal.h

@@ -1,66 +1,66 @@
-// An interface allowing to compute ggml_cgraph with Metal
-//
-// This is a fully functional interface that extends ggml with GPU support for Apple devices.
-// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
-//
-// How it works?
-//
-// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
-// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
-// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
-//
-// You only need to make sure that all memory buffers that you used during the graph creation
-// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
-// used during the graph evaluation to determine the arguments of the compute kernels.
-//
-// Synchronization between device and host memory (for example for input and output tensors)
-// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stddef.h>
-#include <stdbool.h>
-
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 64
-
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//
-// backend API
-// user-code should use only these functions
-//
-
-GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
-
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
-
-GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
-
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
-// helper to check if the device supports a specific family
-// ideally, the user code should be doing these checks
-// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
-
-// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
-
-#ifdef __cplusplus
-}
-#endif
-
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stddef.h>
+#include <stdbool.h>
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 64
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// backend API
+// user-code should use only these functions
+//
+
+GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
+GGML_API ggml_backend_t ggml_backend_metal_init(void);
+
+GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+
+GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+
+GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+
+// helper to check if the device supports a specific family
+// ideally, the user code should be doing these checks
+// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+
+// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
+GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+
+#ifdef __cplusplus
+}
+#endif
+

llama/ggml-quants.h

@@ -1,133 +1,133 @@
-#pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-// Dequantization
-void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-// Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-void iq2xs_init_impl(enum ggml_type type);
-void iq2xs_free_impl(enum ggml_type type);
-void iq3xs_init_impl(int grid_size);
-void iq3xs_free_impl(int grid_size);
-
-#ifdef __cplusplus
-}
-#endif
-
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dequantization
+void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+void iq2xs_init_impl(enum ggml_type type);
+void iq2xs_free_impl(enum ggml_type type);
+void iq3xs_init_impl(int grid_size);
+void iq3xs_free_impl(int grid_size);
+
+#ifdef __cplusplus
+}
+#endif
+

llama/ggml.h

@@ -326,14 +326,20 @@ extern "C" {
     // get ggml_status name string
     GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status);
 
+    // ieee 754-2008 half-precision float16
+    // todo: make this not an integral type
     typedef uint16_t ggml_fp16_t;
-
-    // convert FP16 <-> FP32
-    GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);
-    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
-
-    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n);
-    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n);
+    GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t);
+    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float);
+    GGML_API void        ggml_fp16_to_fp32_row(const ggml_fp16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_fp16_row(const float *, ggml_fp16_t *, int64_t);
+
+    // google brain half-precision bfloat16
+    typedef struct { uint16_t bits; } ggml_bf16_t;
+    GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
+    GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
+    GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
 
     struct ggml_object;
     struct ggml_context;
@@ -370,6 +376,7 @@ extern "C" {
         GGML_TYPE_I64     = 27,
         GGML_TYPE_F64     = 28,
         GGML_TYPE_IQ1_M   = 29,
+        GGML_TYPE_BF16    = 30,
         GGML_TYPE_COUNT,
     };
 
@@ -410,6 +417,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ2_S   = 21, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
+        GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
     };
 
     // available tensor operations:
@@ -460,7 +468,6 @@ extern "C" {
         GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
-        GGML_OP_ALIBI,
         GGML_OP_CLAMP,
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
@@ -475,6 +482,7 @@ extern "C" {
         GGML_OP_LEAKY_RELU,
 
         GGML_OP_FLASH_ATTN,
+        GGML_OP_FLASH_ATTN_EXT,
         GGML_OP_FLASH_FF,
         GGML_OP_FLASH_ATTN_BACK,
         GGML_OP_SSM_CONV,
@@ -511,6 +519,7 @@ extern "C" {
         GGML_UNARY_OP_TANH,
         GGML_UNARY_OP_ELU,
         GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_SIGMOID,
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
@@ -556,7 +565,8 @@ extern "C" {
     // n-dimensional tensor
     struct ggml_tensor {
         enum ggml_type         type;
-        enum ggml_backend_type backend;
+
+        GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
 
         struct ggml_backend_buffer * buffer;
 
@@ -757,11 +767,14 @@ extern "C" {
     GGML_API           bool ggml_is_3d        (const struct ggml_tensor * tensor);
     GGML_API           int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
-    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+    GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+    GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
 
+    GGML_API bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes);
+
     // main
 
     GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -1063,6 +1076,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sigmoid_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_gelu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1417,15 +1438,13 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // fused soft_max(a*scale + mask + pos[i]*(ALiBi slope))
+    // fused soft_max(a*scale + mask*(ALiBi slope))
     // mask is optional
-    // pos is required when max_bias > 0.0f
     // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * mask,
-            struct ggml_tensor  * pos,
             float                 scale,
             float                 max_bias);
 
@@ -1527,16 +1546,6 @@ extern "C" {
             float                 xpos_base,
             bool                  xpos_down);
 
-    // alibi position embedding
-    // in-place, returns view(a)
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_alibi(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   n_past,
-            int                   n_head,
-            float                 bias_max),
-        "use ggml_soft_max_ext instead (will be removed in Mar 2024)");
-
     // clamp
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_clamp(
@@ -1666,12 +1675,24 @@ extern "C" {
             float                 p1);
 
     // nearest interpolate
+    // multiplies ne0 and ne1 by scale factor
     // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_upscale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   scale_factor);
 
+    // nearest interpolate
+    // nearest interpolate to specified dimensions
+    // used in tortoise.cpp
+    GGML_API struct ggml_tensor * ggml_upscale_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   ne0,
+            int                   ne1,
+            int                   ne2,
+            int                   ne3);
+
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
             struct ggml_context * ctx,
@@ -1720,6 +1741,26 @@ extern "C" {
             struct ggml_tensor  * v,
             bool                  masked);
 
+#define GGML_KQ_MASK_PAD 32
+
+    // q:    [n_embd, n_batch,     n_head,    1]
+    // k:    [n_embd, n_kv,        n_head_kv, 1]
+    // v:    [n_embd, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,   n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd, n_head,      n_batch,   1] !! permuted !!
+    GGML_API struct ggml_tensor * ggml_flash_attn_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * q,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * mask,
+            float                 scale,
+            float                 max_bias);
+
+    GGML_API void ggml_flash_attn_ext_set_prec(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec);
+
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,

llama/hipblas.sh

@@ -0,0 +1,91 @@
+hipcc \
+    -parallel-jobs=12 \
+    -O3 \
+    --offload-arch=gfx1100 \
+    -DGGML_USE_CUDA \
+    -DGGML_BUILD=1 \
+    -DGGML_SHARED=1 \
+    -DGGML_CUDA_DMMV_X=32 \
+    -DGGML_CUDA_MMV_Y=1 \
+    -DGGML_SCHED_MAX_COPIES=4 \
+    -DGGML_USE_HIPBLAS \
+    -DGGML_USE_LLAMAFILE \
+    -DK_QUANTS_PER_ITERATION=2 \
+    -D_CRT_SECURE_NO_WARNINGS \
+    -Xclang --dependent-lib=msvcrt -Wl,/subsystem:console \
+    -Wno-expansion-to-defined \
+    -Wno-invalid-noreturn \
+    -Wno-ignored-attributes \
+    -Wno-pass-failed \
+    -Wno-deprecated-declarations \
+    -I. \
+    -lhipblas -lamdhip64 -lrocblas \
+    -shared \
+    -o ggml-hipblas.dll \
+    ggml-cuda.cu ggml-cuda/*.cu ggml.c ggml-backend.c ggml-alloc.c ggml-quants.c
+
+#   --offload-arch=gfx900 \
+#     --offload-arch=gfx906:xnack- \
+#     --offload-arch=gfx908:xnack- \
+#     --offload-arch=gfx90a:xnack+ \
+#     --offload-arch=gfx90a:xnack- \
+#     --offload-arch=gfx940 \
+#     --offload-arch=gfx941 \
+#     --offload-arch=gfx942 \
+#     --offload-arch=gfx1010 \
+#     --offload-arch=gfx1012 \
+#     --offload-arch=gfx1030 \
+#     --offload-arch=gfx1100 \
+#     --offload-arch=gfx1101 \
+#     --offload-arch=gfx1102 \
+
+# hipcc \
+#     -v \
+#     -parallel-jobs=12 \
+#     --offload-arch=gfx900 \
+#     --offload-arch=gfx940 \
+#     --offload-arch=gfx941 \
+#     --offload-arch=gfx942 \
+#     --offload-arch=gfx1010 \
+#     --offload-arch=gfx1012 \
+#     --offload-arch=gfx1030 \
+#     --offload-arch=gfx1100 \
+#     --offload-arch=gfx1101 \
+#     --offload-arch=gfx1102 \
+#     -DLLAMA_BUILD \
+#     -DLLAMA_SHARED \
+#     -DK_QUANTS_PER_ITERATION=2 \
+#     -DGGML_USE_CUDA=1 \
+#     -DGGML_USE_HIPBLAS \
+#     -DGGML_CUDA_DMMV_X=32 \
+#     -DGGML_CUDA_MMV_Y=1 \
+#     -DGGML_CUDA_FORCE_DMMV \
+#     -DNDEBUG \
+#     -D_DLL \
+#     -D_MT \
+#     -D_XOPEN_SOURCE=600 \
+#     -D_CRT_SECURE_NO_WARNINGS \
+#     -Xclang \
+#     --dependent-lib=msvcrt \
+#     -std=gnu++14 \
+#     -Wmissing-declarations \
+#     -Wmissing-noreturn \
+#     -Wall \
+#     -Wextra \
+#     -Wpedantic \
+#     -Wcast-qual \
+#     -Wno-unused-function \
+#     -Wunreachable-code-break \
+#     -Wunreachable-code-return \
+#     -Wmissing-prototypes \
+#     -Wextra-semi \
+#     -Wno-cast-qual \
+#     -I. \
+#     -lhipblas -lamdhip64 -lrocblas \
+#     -Wno-expansion-to-defined \
+#     -Wno-invalid-noreturn \
+#     -Wno-deprecated-declarations \
+#     -mllvm \
+#     -x hip \
+#     -shared -o ggml-cuda.dll \
+#     ggml-cuda.cu ggml-cuda/*.cu ggml.c ggml-backend.c ggml-alloc.c ggml-quants.c 

llama/llama.go

@@ -1,16 +1,203 @@
-// TODO: embed the metal library
-//
-//go:generate bash metal.sh
 package llama
 
-// #cgo CFLAGS: -I.
-// #cgo CXXFLAGS: -std=c++11 -DGGML_USE_METAL
-// #cgo darwin,arm64 LDFLAGS: -framework Foundation -framework Metal -framework MetalKit -framework Accelerate
+// #cgo darwin,arm64 CFLAGS: -std=c11 -DGGML_USE_METAL -DGGML_METAL_EMBED_LIBRARY -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
+// #cgo darwin,arm64 CXXFLAGS: -std=c++11 -DGGML_USE_METAL -DGGML_METAL_EMBED_LIBRARY -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
+// #cgo darwin,amd64 CXXFLAGS: -std=c++11
+// #cgo darwin,arm64 LDFLAGS: ggml-metal.o -framework Foundation -framework Metal -framework MetalKit -framework Accelerate
+// #cgo darwin,amd64 LDFLAGS: -framework Foundation -framework Accelerate
+// #cgo windows LDFLAGS: -lmsvcrt
+// #cgo avx CFLAGS: -mavx
+// #cgo avx CXXFLAGS: -mavx
+// #cgo avx2 CFLAGS: -mavx -mavx2 -mfma
+// #cgo avx2 CXXFLAGS: -mavx -mavx2 -mfma
+// #cgo avx2 LDFLAGS: -lm
+// #cgo cuda CFLAGS: -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_MULTIPLATFORM -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+// #cgo cuda CXXFLAGS: -std=c++11 -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_MULTIPLATFORM -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+// #cgo rocm CXXFLAGS: -std=c++11 -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_MULTIPLATFORM -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+// #cgo windows,cuda LDFLAGS: -L. -L"C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.3/lib/x64" -lggml-cuda -lcuda -lcudart -lcublas -lcublasLt
+// #cgo windows,rocm LDFLAGS: -L. -L"C:/Program Files/AMD/ROCm/5.7/lib" -lggml-hipblas -lhipblas -lamdhip64 -lrocblas
 // #include <stdlib.h>
 // #include "llama.h"
 import "C"
+import (
+	"fmt"
+	"runtime"
+	"strings"
+	"unsafe"
+)
+
+type Token int32
+type Pos int32
+type SeqId int32
 
 // SystemInfo is an unused example of calling llama.cpp functions using CGo
-func SystemInfo() string {
+func PrintSystemInfo() string {
 	return C.GoString(C.llama_print_system_info())
 }
+
+func BackendInit() {
+	C.llama_backend_init()
+}
+
+type ContextParams struct {
+	c C.struct_llama_context_params
+}
+
+func NewContextParams() ContextParams {
+	params := C.llama_context_default_params()
+	params.seed = C.uint(1234)
+	params.n_ctx = C.uint(2048)
+	params.n_threads = C.uint(runtime.NumCPU())
+	params.n_threads_batch = params.n_threads
+	return ContextParams{c: params}
+}
+
+type ModelParams struct {
+	c C.struct_llama_model_params
+}
+
+func NewModelParams() ModelParams {
+	params := C.llama_model_default_params()
+	params.n_gpu_layers = 999
+	return ModelParams{c: params}
+}
+
+type Context struct {
+	c *C.struct_llama_context
+}
+
+func (c *Context) Decode(batch Batch) error {
+	// Positive return values does not mean a fatal error, but rather a warning.
+	//   0 - success
+	//   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+	// < 0 - error
+	code := int(C.llama_decode(c.c, batch.c))
+
+	if code < 0 {
+		return fmt.Errorf("llama_decode failed with code %d", code)
+	}
+
+	if code > 0 {
+		return fmt.Errorf("could not find a KV slot for the batch - try reducing the size of the batch or increase the context. code: %d\n", code)
+	}
+
+	return nil
+}
+
+func (c *Context) getModel() *Model {
+	return &Model{c: C.llama_get_model(c.c)}
+}
+
+func (c *Context) SampleTokenGreedy(batch Batch) Token {
+	nv := c.getModel().NumVocab()
+
+	// TODO(jmorganca): split this up into different functions
+	candidates := (*C.struct_llama_token_data)(C.malloc(C.size_t(nv) * C.size_t(unsafe.Sizeof(C.struct_llama_token_data{}))))
+	defer C.free(unsafe.Pointer(candidates))
+
+	// get most recent logits
+	logits := C.llama_get_logits_ith(c.c, C.int(batch.NumTokens()-1))
+	for i := 0; i < int(nv); i++ {
+		ptr := (*C.struct_llama_token_data)(unsafe.Pointer(uintptr(unsafe.Pointer(candidates)) + uintptr(i)*unsafe.Sizeof(C.struct_llama_token_data{})))
+		ptr.id = C.int(i)
+		ptr.logit = unsafe.Slice(logits, nv)[i]
+		ptr.p = 0.0
+	}
+
+	return Token(C.llama_sample_token_greedy(c.c, &C.llama_token_data_array{
+		data:   candidates,
+		size:   C.size_t(nv),
+		sorted: C.bool(false),
+	}))
+}
+
+func LoadModelFromFile(modelPath string, params ModelParams) *Model {
+	return &Model{c: C.llama_load_model_from_file(C.CString(modelPath), params.c)}
+}
+
+func NewContextWithModel(model *Model, params ContextParams) *Context {
+	return &Context{c: C.llama_new_context_with_model(model.c, params.c)}
+}
+
+func (m *Model) NumVocab() int {
+	return int(C.llama_n_vocab(m.c))
+}
+
+func (m *Model) TokenIsEog(token Token) bool {
+	return bool(C.llama_token_is_eog(m.c, C.llama_token(token)))
+}
+
+type Batch struct {
+	c C.struct_llama_batch
+}
+
+func NewBatch(nTokens int, nSeqs int, nCtx int) Batch {
+	return Batch{c: C.llama_batch_init(C.int(nTokens), C.int(nSeqs), C.int(nCtx))}
+}
+
+func (b *Batch) NumTokens() int {
+	return int(b.c.n_tokens)
+}
+
+func (b *Batch) Add(token Token, pos Pos, seqIds []SeqId, logits bool) {
+	unsafe.Slice(b.c.token, 512)[b.c.n_tokens] = C.llama_token(token)
+	unsafe.Slice(b.c.pos, 512)[b.c.n_tokens] = C.llama_pos(pos)
+	unsafe.Slice(b.c.n_seq_id, 512)[b.c.n_tokens] = C.int(len(seqIds))
+
+	for i, s := range seqIds {
+		unsafe.Slice((unsafe.Slice(b.c.seq_id, 512)[b.c.n_tokens]), C.int(len(seqIds)))[i] = C.int32_t(s)
+	}
+
+	if logits {
+		unsafe.Slice(b.c.logits, 512)[b.c.n_tokens] = 1
+	}
+
+	b.c.n_tokens += 1
+}
+
+func (b *Batch) Clear() {
+	b.c.n_tokens = 0
+}
+
+type Model struct {
+	c *C.struct_llama_model
+}
+
+func (m *Model) TokenToPiece(token Token) string {
+	buf := make([]byte, 12)
+	C.llama_token_to_piece(
+		m.c,
+		C.int32_t(token),
+		(*C.char)(unsafe.Pointer(&buf[0])),
+		C.int32_t(12),
+		C.bool(true),
+	)
+	return strings.TrimRight(string(buf), "\x00")
+}
+
+func (m *Model) Tokenize(text string, maxTokens int, addSpecial bool, parseSpecial bool) ([]Token, error) {
+	cTokens := make([]C.llama_token, maxTokens)
+	cText := C.CString(text)
+	defer C.free(unsafe.Pointer(cText))
+
+	result := C.llama_tokenize(
+		m.c,
+		cText,
+		C.int32_t(len(text)),
+		&cTokens[0],
+		C.int32_t(maxTokens),
+		C.bool(addSpecial),
+		C.bool(parseSpecial),
+	)
+
+	if result < 0 {
+		return nil, fmt.Errorf("tokenization failed, required %d tokens", -result)
+	}
+
+	tokens := make([]Token, result)
+	for i := 0; i < int(result); i++ {
+		tokens[i] = Token(cTokens[i])
+	}
+
+	return tokens, nil
+}

llama/llama.h

@@ -40,7 +40,7 @@
 #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 5
+#define LLAMA_SESSION_VERSION 6
 
 #define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
 #define LLAMA_STATE_SEQ_VERSION 1
@@ -69,6 +69,23 @@ extern "C" {
         LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
     };
 
+    // pre-tokenization types
+    enum llama_vocab_pre_type {
+        LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
+        LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
+        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
+        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
+        LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
+        LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
+        LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
+        LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
+        LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
+        LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
+        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 10,
+        LLAMA_VOCAB_PRE_TYPE_OLMO           = 11,
+        LLAMA_VOCAB_PRE_TYPE_DBRX           = 12,
+    };
+
     // note: these values should be synchronized with ggml_rope
     // TODO: maybe move this enum to ggml.h (ggml_rope_type)
     enum llama_rope_type {
@@ -122,6 +139,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -159,7 +177,7 @@ extern "C" {
         bool sorted;
     } llama_token_data_array;
 
-    typedef bool (*llama_progress_callback)(float progress, void *ctx);
+    typedef bool (*llama_progress_callback)(float progress, void * user_data);
 
     // Input data for llama_decode
     // A llama_batch object can contain input about one or many sequences
@@ -195,15 +213,19 @@ extern "C" {
         LLAMA_KV_OVERRIDE_TYPE_INT,
         LLAMA_KV_OVERRIDE_TYPE_FLOAT,
         LLAMA_KV_OVERRIDE_TYPE_BOOL,
+        LLAMA_KV_OVERRIDE_TYPE_STR,
     };
 
     struct llama_model_kv_override {
-        char key[128];
         enum llama_model_kv_override_type tag;
+
+        char key[128];
+
         union {
-            int64_t int_value;
-            double float_value;
-            bool bool_value;
+            int64_t val_i64;
+            double  val_f64;
+            bool    val_bool;
+            char    val_str[128];
         };
     };
 
@@ -220,6 +242,9 @@ extern "C" {
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         const float * tensor_split;
 
+        // comma separated list of RPC servers to use for offloading
+        const char * rpc_servers;
+
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
         // If the provided progress_callback returns true, model loading continues.
         // If it returns false, model loading is immediately aborted.
@@ -232,9 +257,10 @@ extern "C" {
         const struct llama_model_kv_override * kv_overrides;
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool vocab_only; // only load the vocabulary, no weights
-        bool use_mmap;   // use mmap if possible
-        bool use_mlock;  // force system to keep model in RAM
+        bool vocab_only;    // only load the vocabulary, no weights
+        bool use_mmap;      // use mmap if possible
+        bool use_mlock;     // force system to keep model in RAM
+        bool check_tensors; // validate model tensor data
     };
 
     struct llama_context_params {
@@ -270,6 +296,7 @@ extern "C" {
         bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
         bool embeddings;  // if true, extract embeddings (together with logits)
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
+        bool flash_attn;  // whether to use flash attention
 
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
@@ -288,6 +315,7 @@ extern "C" {
         bool quantize_output_tensor;         // quantize output.weight
         bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
         bool pure;                           // quantize all tensors to the default type
+        bool keep_split;                     // quantize to the same number of shards
         void * imatrix;                      // pointer to importance matrix data
         void * kv_overrides;                 // pointer to vector containing overrides
     } llama_model_quantize_params;
@@ -390,8 +418,10 @@ extern "C" {
     LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
 
-    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
-    LLAMA_API enum llama_rope_type  llama_rope_type (const struct llama_model * model);
+    LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
+
+    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model   * model);
+    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model   * model);
 
     LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
     LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
@@ -522,7 +552,7 @@ extern "C" {
     // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
     LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
 
-    // Clear the KV cache
+    // Clear the KV cache - both cell info is erased and KV data is zeroed
     LLAMA_API void llama_kv_cache_clear(
             struct llama_context * ctx);
 
@@ -783,6 +813,9 @@ extern "C" {
 
     LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
 
+    // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
+    LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
+
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
@@ -796,7 +829,7 @@ extern "C" {
     // Returns -1 if unknown, 1 for true or 0 for false.
     LLAMA_API int32_t         llama_add_eos_token(const struct llama_model * model);
 
-    // codellama infill tokens
+    // Codellama infill tokens
     LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
     LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
     LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
@@ -825,11 +858,13 @@ extern "C" {
     // Uses the vocabulary in the provided context.
     // Does not write null terminator to the buffer.
     // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
+    // @param special If true, special tokens are rendered in the output.
     LLAMA_API int32_t llama_token_to_piece(
               const struct llama_model * model,
                            llama_token   token,
                                   char * buf,
-                               int32_t   length);
+                               int32_t   length,
+                                  bool   special);
 
     /// Apply chat template. Inspired by hf apply_chat_template() on python.
     /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
@@ -982,7 +1017,7 @@ extern "C" {
             struct llama_context * ctx,
           llama_token_data_array * candidates);
 
-    /// @details Randomly selects a token from the candidates based on their probabilities.
+    /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
     LLAMA_API llama_token llama_sample_token(
             struct llama_context * ctx,
           llama_token_data_array * candidates);
@@ -1069,8 +1104,9 @@ extern "C" {
 // Internal API to be implemented by llama.cpp and used by tests/benchmarks only
 #ifdef LLAMA_API_INTERNAL
 
-#include <vector>
+#include <random>
 #include <string>
+#include <vector>
 
 struct ggml_tensor;
 
@@ -1107,6 +1143,10 @@ std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
         const std::string & src,
         llama_partial_utf8   partial_start);
 
+// Randomly selects a token from the candidates based on their probabilities using given std::mt19937.
+// This is a temporary workaround in order to fix race conditions when sampling with multiple sequences.
+llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng);
+
 #endif // LLAMA_API_INTERNAL
 
 #endif // LLAMA_H

llama/metal.sh

@@ -6,5 +6,6 @@ echo "_ggml_metallib_start:"              >> $TEMP_ASSEMBLY
 echo ".incbin \"ggml-metal-embed.metal\"" >> $TEMP_ASSEMBLY
 echo ".globl _ggml_metallib_end"          >> $TEMP_ASSEMBLY
 echo "_ggml_metallib_end:"                >> $TEMP_ASSEMBLY
-as $TEMP_ASSEMBLY -o ggml-metal-embed.o
+as -mmacosx-version-min=11.3 $TEMP_ASSEMBLY -o ggml-metal.o
 rm -f $TEMP_ASSEMBLY
+rm -rf ggml-metal-embed.metal

llama/server/README.md

@@ -0,0 +1,9 @@
+# `llama/server`
+
+```
+./server -model <model binary>
+```
+
+```
+curl POST -H "Content-Type: application/json" -d '{"prompt": "hi"}' http://localhost:8080/
+```

llama/server/main.go

@@ -0,0 +1,126 @@
+package main
+
+import (
+	"encoding/json"
+	"flag"
+	"fmt"
+	"log"
+	"net"
+	"net/http"
+	"sync"
+
+	"github.com/ollama/ollama/llama"
+)
+
+type Request struct {
+	Prompt string `json:"prompt"`
+}
+
+type Response struct {
+	Token string `json:"token"`
+}
+
+type Server struct {
+	model *llama.Model
+	lc    *llama.Context
+	batch *llama.Batch
+}
+
+var mu sync.Mutex
+
+func (s *Server) stream(w http.ResponseWriter, r *http.Request) {
+	var request Request
+	if err := json.NewDecoder(r.Body).Decode(&request); err != nil {
+		http.Error(w, "Bad request", http.StatusBadRequest)
+		return
+	}
+
+	// Set the headers to indicate streaming
+	w.Header().Set("Content-Type", "application/json")
+	w.Header().Set("Transfer-Encoding", "chunked")
+	w.WriteHeader(http.StatusOK)
+
+	enc := json.NewEncoder(w)
+
+	// main loop
+	tokens, err := s.model.Tokenize(request.Prompt, 2048, true, true)
+	if err != nil {
+		panic(err)
+	}
+
+	fmt.Println("tokens", tokens)
+
+	batch := llama.NewBatch(512, 0, 1)
+
+	// prompt eval
+	for i, t := range tokens {
+		batch.Add(t, llama.Pos(i), []llama.SeqId{0}, true)
+	}
+
+	// main loop
+	for n := batch.NumTokens(); n < 2048; n++ {
+		mu.Lock()
+		err = s.lc.Decode(batch)
+		if err != nil {
+			panic("Failed to decode")
+		}
+
+		// sample a token
+		token := s.lc.SampleTokenGreedy(batch)
+		mu.Unlock()
+
+		// if it's an end of sequence token, break
+		if s.model.TokenIsEog(token) {
+			break
+		}
+
+		// print the token
+		str := s.model.TokenToPiece(token)
+
+		if err := enc.Encode(&Response{Token: str}); err != nil {
+			log.Println("Failed to encode result:", err)
+			return
+		}
+		w.(http.Flusher).Flush()
+
+		batch.Clear()
+		batch.Add(token, llama.Pos(n), []llama.SeqId{0}, true)
+	}
+}
+
+func main() {
+	mp := flag.String("model", "", "Path to model binary file")
+	flag.Parse()
+
+	// load the model
+	llama.BackendInit()
+	params := llama.NewModelParams()
+	model := llama.LoadModelFromFile(*mp, params)
+	ctxParams := llama.NewContextParams()
+	lc := llama.NewContextWithModel(model, ctxParams)
+	if lc == nil {
+		panic("Failed to create context")
+	}
+
+	server := &Server{
+		model: model,
+		lc:    lc,
+	}
+
+	addr := "127.0.0.1:8080"
+	listener, err := net.Listen("tcp", addr)
+	if err != nil {
+		fmt.Println("Listen error:", err)
+		return
+	}
+	defer listener.Close()
+
+	httpServer := http.Server{
+		Handler: http.HandlerFunc(server.stream),
+	}
+
+	log.Println("Server listening on", addr)
+	if err := httpServer.Serve(listener); err != nil {
+		log.Fatal("server error:", err)
+	}
+}

llama/sgemm.h

@@ -1,11 +1,13 @@
 #pragma once
+#include <stdint.h>
 #include <stdbool.h>
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-bool llamafile_sgemm(int, int, int, const void *, int, const void *, int,
-                     void *, int, int, int, int, int, int, int);
+bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
+                     const void *, int64_t, void *, int64_t, int, int,
+                     int, int, int, int);
 
 #ifdef __cplusplus
 }

llama/sync.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
 # Set the source directory
-src_dir="../../llama.cpp"
+src_dir=$1
 
 # Set the destination directory (current directory)
 dst_dir="."
@@ -23,7 +23,7 @@ cp $src_dir/ggml-quants.c $dst_dir/ggml-quants.c
 cp $src_dir/ggml-quants.h $dst_dir/ggml-quants.h
 cp $src_dir/ggml-metal.metal $dst_dir/ggml-metal.metal
 cp $src_dir/ggml-metal.h $dst_dir/ggml-metal.h
-cp $src_dir/ggml-metal.m $dst_dir/ggml-metal.m
+cp $src_dir/ggml-metal.m $dst_dir/ggml-metal-darwin_arm64.m
 cp $src_dir/ggml-impl.h $dst_dir/ggml-impl.h
 cp $src_dir/ggml-cuda.h $dst_dir/ggml-cuda.h
 cp $src_dir/ggml-cuda.cu $dst_dir/ggml-cuda.cu
@@ -33,3 +33,12 @@ cp $src_dir/ggml-backend.c $dst_dir/ggml-backend.c
 cp $src_dir/ggml-backend-impl.h $dst_dir/ggml-backend-impl.h
 cp $src_dir/ggml-alloc.h $dst_dir/ggml-alloc.h
 cp $src_dir/ggml-alloc.c $dst_dir/ggml-alloc.c
+
+sed -i 's/extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();/\/\/ extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();/' ggml-cuda.cu
+sed -i '34iGGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);' ggml-cuda.h
+
+
+# ggml-cuda
+mkdir -p $dst_dir/ggml-cuda
+cp $src_dir/ggml-cuda/*.cu $dst_dir/ggml-cuda/
+cp $src_dir/ggml-cuda/*.cuh $dst_dir/ggml-cuda/

llama/unicode-data.cpp

@@ -1,1651 +1,6983 @@
+// generated with scripts/gen-unicode-data.py
+
 #include "unicode-data.h"
 
 #include <cstdint>
-#include <map>
-#include <utility>
 #include <vector>
+#include <unordered_map>
+#include <unordered_set>
 
-const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_digit = {
-{0x00000030, 0x00000039}, {0x000000B2, 0x000000B3}, {0x000000B9, 0x000000B9}, {0x00000660, 0x00000669},
-{0x000006F0, 0x000006F9}, {0x000007C0, 0x000007C9}, {0x00000966, 0x0000096F}, {0x000009E6, 0x000009EF},
-{0x00000A66, 0x00000A6F}, {0x00000AE6, 0x00000AEF}, {0x00000B66, 0x00000B6F}, {0x00000BE6, 0x00000BEF},
-{0x00000C66, 0x00000C6F}, {0x00000CE6, 0x00000CEF}, {0x00000D66, 0x00000D6F}, {0x00000DE6, 0x00000DEF},
-{0x00000E50, 0x00000E59}, {0x00000ED0, 0x00000ED9}, {0x00000F20, 0x00000F29}, {0x00001040, 0x00001049},
-{0x00001090, 0x00001099}, {0x00001369, 0x00001371}, {0x000017E0, 0x000017E9}, {0x00001810, 0x00001819},
-{0x00001946, 0x0000194F}, {0x000019D0, 0x000019DA}, {0x00001A80, 0x00001A89}, {0x00001A90, 0x00001A99},
-{0x00001B50, 0x00001B59}, {0x00001BB0, 0x00001BB9}, {0x00001C40, 0x00001C49}, {0x00001C50, 0x00001C59},
-{0x00002070, 0x00002070}, {0x00002074, 0x00002079}, {0x00002080, 0x00002089}, {0x00002460, 0x00002468},
-{0x00002474, 0x0000247C}, {0x00002488, 0x00002490}, {0x000024EA, 0x000024EA}, {0x000024F5, 0x000024FD},
-{0x000024FF, 0x000024FF}, {0x00002776, 0x0000277E}, {0x00002780, 0x00002788}, {0x0000278A, 0x00002792},
-{0x0000A620, 0x0000A629}, {0x0000A8D0, 0x0000A8D9}, {0x0000A900, 0x0000A909}, {0x0000A9D0, 0x0000A9D9},
-{0x0000A9F0, 0x0000A9F9}, {0x0000AA50, 0x0000AA59}, {0x0000ABF0, 0x0000ABF9}, {0x0000FF10, 0x0000FF19},
-{0x000104A0, 0x000104A9}, {0x00010A40, 0x00010A43}, {0x00010D30, 0x00010D39}, {0x00010E60, 0x00010E68},
-{0x00011052, 0x0001105A}, {0x00011066, 0x0001106F}, {0x000110F0, 0x000110F9}, {0x00011136, 0x0001113F},
-{0x000111D0, 0x000111D9}, {0x000112F0, 0x000112F9}, {0x00011450, 0x00011459}, {0x000114D0, 0x000114D9},
-{0x00011650, 0x00011659}, {0x000116C0, 0x000116C9}, {0x00011730, 0x00011739}, {0x000118E0, 0x000118E9},
-{0x00011950, 0x00011959}, {0x00011C50, 0x00011C59}, {0x00011D50, 0x00011D59}, {0x00011DA0, 0x00011DA9},
-{0x00016A60, 0x00016A69}, {0x00016B50, 0x00016B59}, {0x0001D7CE, 0x0001D7FF}, {0x0001E140, 0x0001E149},
-{0x0001E2F0, 0x0001E2F9}, {0x0001E950, 0x0001E959}, {0x0001F100, 0x0001F10A}, {0x0001FBF0, 0x0001FBF9},
-};
-
-const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter = {
-{0x00000041, 0x0000005A}, {0x00000061, 0x0000007A}, {0x000000AA, 0x000000AA}, {0x000000B5, 0x000000B5},
-{0x000000BA, 0x000000BA}, {0x000000C0, 0x000000D6}, {0x000000D8, 0x000000F6}, {0x000000F8, 0x000002C1},
-{0x000002C6, 0x000002D1}, {0x000002E0, 0x000002E4}, {0x000002EC, 0x000002EC}, {0x000002EE, 0x000002EE},
-{0x00000370, 0x00000374}, {0x00000376, 0x00000377}, {0x0000037A, 0x0000037D}, {0x0000037F, 0x0000037F},
-{0x00000386, 0x00000386}, {0x00000388, 0x0000038A}, {0x0000038C, 0x0000038C}, {0x0000038E, 0x000003A1},
-{0x000003A3, 0x000003F5}, {0x000003F7, 0x00000481}, {0x0000048A, 0x0000052F}, {0x00000531, 0x00000556},
-{0x00000559, 0x00000559}, {0x00000560, 0x00000588}, {0x000005D0, 0x000005EA}, {0x000005EF, 0x000005F2},
-{0x00000620, 0x0000064A}, {0x0000066E, 0x0000066F}, {0x00000671, 0x000006D3}, {0x000006D5, 0x000006D5},
-{0x000006E5, 0x000006E6}, {0x000006EE, 0x000006EF}, {0x000006FA, 0x000006FC}, {0x000006FF, 0x000006FF},
-{0x00000710, 0x00000710}, {0x00000712, 0x0000072F}, {0x0000074D, 0x000007A5}, {0x000007B1, 0x000007B1},
-{0x000007CA, 0x000007EA}, {0x000007F4, 0x000007F5}, {0x000007FA, 0x000007FA}, {0x00000800, 0x00000815},
-{0x0000081A, 0x0000081A}, {0x00000824, 0x00000824}, {0x00000828, 0x00000828}, {0x00000840, 0x00000858},
-{0x00000860, 0x0000086A}, {0x000008A0, 0x000008B4}, {0x000008B6, 0x000008C7}, {0x00000904, 0x00000939},
-{0x0000093D, 0x0000093D}, {0x00000950, 0x00000950}, {0x00000958, 0x00000961}, {0x00000971, 0x00000980},
-{0x00000985, 0x0000098C}, {0x0000098F, 0x00000990}, {0x00000993, 0x000009A8}, {0x000009AA, 0x000009B0},
-{0x000009B2, 0x000009B2}, {0x000009B6, 0x000009B9}, {0x000009BD, 0x000009BD}, {0x000009CE, 0x000009CE},
-{0x000009DC, 0x000009DD}, {0x000009DF, 0x000009E1}, {0x000009F0, 0x000009F1}, {0x000009FC, 0x000009FC},
-{0x00000A05, 0x00000A0A}, {0x00000A0F, 0x00000A10}, {0x00000A13, 0x00000A28}, {0x00000A2A, 0x00000A30},
-{0x00000A32, 0x00000A33}, {0x00000A35, 0x00000A36}, {0x00000A38, 0x00000A39}, {0x00000A59, 0x00000A5C},
-{0x00000A5E, 0x00000A5E}, {0x00000A72, 0x00000A74}, {0x00000A85, 0x00000A8D}, {0x00000A8F, 0x00000A91},
-{0x00000A93, 0x00000AA8}, {0x00000AAA, 0x00000AB0}, {0x00000AB2, 0x00000AB3}, {0x00000AB5, 0x00000AB9},
-{0x00000ABD, 0x00000ABD}, {0x00000AD0, 0x00000AD0}, {0x00000AE0, 0x00000AE1}, {0x00000AF9, 0x00000AF9},
-{0x00000B05, 0x00000B0C}, {0x00000B0F, 0x00000B10}, {0x00000B13, 0x00000B28}, {0x00000B2A, 0x00000B30},
-{0x00000B32, 0x00000B33}, {0x00000B35, 0x00000B39}, {0x00000B3D, 0x00000B3D}, {0x00000B5C, 0x00000B5D},
-{0x00000B5F, 0x00000B61}, {0x00000B71, 0x00000B71}, {0x00000B83, 0x00000B83}, {0x00000B85, 0x00000B8A},
-{0x00000B8E, 0x00000B90}, {0x00000B92, 0x00000B95}, {0x00000B99, 0x00000B9A}, {0x00000B9C, 0x00000B9C},
-{0x00000B9E, 0x00000B9F}, {0x00000BA3, 0x00000BA4}, {0x00000BA8, 0x00000BAA}, {0x00000BAE, 0x00000BB9},
-{0x00000BD0, 0x00000BD0}, {0x00000C05, 0x00000C0C}, {0x00000C0E, 0x00000C10}, {0x00000C12, 0x00000C28},
-{0x00000C2A, 0x00000C39}, {0x00000C3D, 0x00000C3D}, {0x00000C58, 0x00000C5A}, {0x00000C60, 0x00000C61},
-{0x00000C80, 0x00000C80}, {0x00000C85, 0x00000C8C}, {0x00000C8E, 0x00000C90}, {0x00000C92, 0x00000CA8},
-{0x00000CAA, 0x00000CB3}, {0x00000CB5, 0x00000CB9}, {0x00000CBD, 0x00000CBD}, {0x00000CDE, 0x00000CDE},
-{0x00000CE0, 0x00000CE1}, {0x00000CF1, 0x00000CF2}, {0x00000D04, 0x00000D0C}, {0x00000D0E, 0x00000D10},
-{0x00000D12, 0x00000D3A}, {0x00000D3D, 0x00000D3D}, {0x00000D4E, 0x00000D4E}, {0x00000D54, 0x00000D56},
-{0x00000D5F, 0x00000D61}, {0x00000D7A, 0x00000D7F}, {0x00000D85, 0x00000D96}, {0x00000D9A, 0x00000DB1},
-{0x00000DB3, 0x00000DBB}, {0x00000DBD, 0x00000DBD}, {0x00000DC0, 0x00000DC6}, {0x00000E01, 0x00000E30},
-{0x00000E32, 0x00000E33}, {0x00000E40, 0x00000E46}, {0x00000E81, 0x00000E82}, {0x00000E84, 0x00000E84},
-{0x00000E86, 0x00000E8A}, {0x00000E8C, 0x00000EA3}, {0x00000EA5, 0x00000EA5}, {0x00000EA7, 0x00000EB0},
-{0x00000EB2, 0x00000EB3}, {0x00000EBD, 0x00000EBD}, {0x00000EC0, 0x00000EC4}, {0x00000EC6, 0x00000EC6},
-{0x00000EDC, 0x00000EDF}, {0x00000F00, 0x00000F00}, {0x00000F40, 0x00000F47}, {0x00000F49, 0x00000F6C},
-{0x00000F88, 0x00000F8C}, {0x00001000, 0x0000102A}, {0x0000103F, 0x0000103F}, {0x00001050, 0x00001055},
-{0x0000105A, 0x0000105D}, {0x00001061, 0x00001061}, {0x00001065, 0x00001066}, {0x0000106E, 0x00001070},
-{0x00001075, 0x00001081}, {0x0000108E, 0x0000108E}, {0x000010A0, 0x000010C5}, {0x000010C7, 0x000010C7},
-{0x000010CD, 0x000010CD}, {0x000010D0, 0x000010FA}, {0x000010FC, 0x00001248}, {0x0000124A, 0x0000124D},
-{0x00001250, 0x00001256}, {0x00001258, 0x00001258}, {0x0000125A, 0x0000125D}, {0x00001260, 0x00001288},
-{0x0000128A, 0x0000128D}, {0x00001290, 0x000012B0}, {0x000012B2, 0x000012B5}, {0x000012B8, 0x000012BE},
-{0x000012C0, 0x000012C0}, {0x000012C2, 0x000012C5}, {0x000012C8, 0x000012D6}, {0x000012D8, 0x00001310},
-{0x00001312, 0x00001315}, {0x00001318, 0x0000135A}, {0x00001380, 0x0000138F}, {0x000013A0, 0x000013F5},
-{0x000013F8, 0x000013FD}, {0x00001401, 0x0000166C}, {0x0000166F, 0x0000167F}, {0x00001681, 0x0000169A},
-{0x000016A0, 0x000016EA}, {0x000016F1, 0x000016F8}, {0x00001700, 0x0000170C}, {0x0000170E, 0x00001711},
-{0x00001720, 0x00001731}, {0x00001740, 0x00001751}, {0x00001760, 0x0000176C}, {0x0000176E, 0x00001770},
-{0x00001780, 0x000017B3}, {0x000017D7, 0x000017D7}, {0x000017DC, 0x000017DC}, {0x00001820, 0x00001878},
-{0x00001880, 0x00001884}, {0x00001887, 0x000018A8}, {0x000018AA, 0x000018AA}, {0x000018B0, 0x000018F5},
-{0x00001900, 0x0000191E}, {0x00001950, 0x0000196D}, {0x00001970, 0x00001974}, {0x00001980, 0x000019AB},
-{0x000019B0, 0x000019C9}, {0x00001A00, 0x00001A16}, {0x00001A20, 0x00001A54}, {0x00001AA7, 0x00001AA7},
-{0x00001B05, 0x00001B33}, {0x00001B45, 0x00001B4B}, {0x00001B83, 0x00001BA0}, {0x00001BAE, 0x00001BAF},
-{0x00001BBA, 0x00001BE5}, {0x00001C00, 0x00001C23}, {0x00001C4D, 0x00001C4F}, {0x00001C5A, 0x00001C7D},
-{0x00001C80, 0x00001C88}, {0x00001C90, 0x00001CBA}, {0x00001CBD, 0x00001CBF}, {0x00001CE9, 0x00001CEC},
-{0x00001CEE, 0x00001CF3}, {0x00001CF5, 0x00001CF6}, {0x00001CFA, 0x00001CFA}, {0x00001D00, 0x00001DBF},
-{0x00001E00, 0x00001F15}, {0x00001F18, 0x00001F1D}, {0x00001F20, 0x00001F45}, {0x00001F48, 0x00001F4D},
-{0x00001F50, 0x00001F57}, {0x00001F59, 0x00001F59}, {0x00001F5B, 0x00001F5B}, {0x00001F5D, 0x00001F5D},
-{0x00001F5F, 0x00001F7D}, {0x00001F80, 0x00001FB4}, {0x00001FB6, 0x00001FBC}, {0x00001FBE, 0x00001FBE},
-{0x00001FC2, 0x00001FC4}, {0x00001FC6, 0x00001FCC}, {0x00001FD0, 0x00001FD3}, {0x00001FD6, 0x00001FDB},
-{0x00001FE0, 0x00001FEC}, {0x00001FF2, 0x00001FF4}, {0x00001FF6, 0x00001FFC}, {0x00002071, 0x00002071},
-{0x0000207F, 0x0000207F}, {0x00002090, 0x0000209C}, {0x00002102, 0x00002102}, {0x00002107, 0x00002107},
-{0x0000210A, 0x00002113}, {0x00002115, 0x00002115}, {0x00002119, 0x0000211D}, {0x00002124, 0x00002124},
-{0x00002126, 0x00002126}, {0x00002128, 0x00002128}, {0x0000212A, 0x0000212D}, {0x0000212F, 0x00002139},
-{0x0000213C, 0x0000213F}, {0x00002145, 0x00002149}, {0x0000214E, 0x0000214E}, {0x00002183, 0x00002184},
-{0x00002C00, 0x00002C2E}, {0x00002C30, 0x00002C5E}, {0x00002C60, 0x00002CE4}, {0x00002CEB, 0x00002CEE},
-{0x00002CF2, 0x00002CF3}, {0x00002D00, 0x00002D25}, {0x00002D27, 0x00002D27}, {0x00002D2D, 0x00002D2D},
-{0x00002D30, 0x00002D67}, {0x00002D6F, 0x00002D6F}, {0x00002D80, 0x00002D96}, {0x00002DA0, 0x00002DA6},
-{0x00002DA8, 0x00002DAE}, {0x00002DB0, 0x00002DB6}, {0x00002DB8, 0x00002DBE}, {0x00002DC0, 0x00002DC6},
-{0x00002DC8, 0x00002DCE}, {0x00002DD0, 0x00002DD6}, {0x00002DD8, 0x00002DDE}, {0x00002E2F, 0x00002E2F},
-{0x00003005, 0x00003006}, {0x00003031, 0x00003035}, {0x0000303B, 0x0000303C}, {0x00003041, 0x00003096},
-{0x0000309D, 0x0000309F}, {0x000030A1, 0x000030FA}, {0x000030FC, 0x000030FF}, {0x00003105, 0x0000312F},
-{0x00003131, 0x0000318E}, {0x000031A0, 0x000031BF}, {0x000031F0, 0x000031FF}, {0x00003400, 0x00004DBF},
-{0x00004E00, 0x00009FFC}, {0x0000A000, 0x0000A48C}, {0x0000A4D0, 0x0000A4FD}, {0x0000A500, 0x0000A60C},
-{0x0000A610, 0x0000A61F}, {0x0000A62A, 0x0000A62B}, {0x0000A640, 0x0000A66E}, {0x0000A67F, 0x0000A69D},
-{0x0000A6A0, 0x0000A6E5}, {0x0000A717, 0x0000A71F}, {0x0000A722, 0x0000A788}, {0x0000A78B, 0x0000A7BF},
-{0x0000A7C2, 0x0000A7CA}, {0x0000A7F5, 0x0000A801}, {0x0000A803, 0x0000A805}, {0x0000A807, 0x0000A80A},
-{0x0000A80C, 0x0000A822}, {0x0000A840, 0x0000A873}, {0x0000A882, 0x0000A8B3}, {0x0000A8F2, 0x0000A8F7},
-{0x0000A8FB, 0x0000A8FB}, {0x0000A8FD, 0x0000A8FE}, {0x0000A90A, 0x0000A925}, {0x0000A930, 0x0000A946},
-{0x0000A960, 0x0000A97C}, {0x0000A984, 0x0000A9B2}, {0x0000A9CF, 0x0000A9CF}, {0x0000A9E0, 0x0000A9E4},
-{0x0000A9E6, 0x0000A9EF}, {0x0000A9FA, 0x0000A9FE}, {0x0000AA00, 0x0000AA28}, {0x0000AA40, 0x0000AA42},
-{0x0000AA44, 0x0000AA4B}, {0x0000AA60, 0x0000AA76}, {0x0000AA7A, 0x0000AA7A}, {0x0000AA7E, 0x0000AAAF},
-{0x0000AAB1, 0x0000AAB1}, {0x0000AAB5, 0x0000AAB6}, {0x0000AAB9, 0x0000AABD}, {0x0000AAC0, 0x0000AAC0},
-{0x0000AAC2, 0x0000AAC2}, {0x0000AADB, 0x0000AADD}, {0x0000AAE0, 0x0000AAEA}, {0x0000AAF2, 0x0000AAF4},
-{0x0000AB01, 0x0000AB06}, {0x0000AB09, 0x0000AB0E}, {0x0000AB11, 0x0000AB16}, {0x0000AB20, 0x0000AB26},
-{0x0000AB28, 0x0000AB2E}, {0x0000AB30, 0x0000AB5A}, {0x0000AB5C, 0x0000AB69}, {0x0000AB70, 0x0000ABE2},
-{0x0000AC00, 0x0000D7A3}, {0x0000D7B0, 0x0000D7C6}, {0x0000D7CB, 0x0000D7FB}, {0x0000F900, 0x0000FA6D},
-{0x0000FA70, 0x0000FAD9}, {0x0000FB00, 0x0000FB06}, {0x0000FB13, 0x0000FB17}, {0x0000FB1D, 0x0000FB1D},
-{0x0000FB1F, 0x0000FB28}, {0x0000FB2A, 0x0000FB36}, {0x0000FB38, 0x0000FB3C}, {0x0000FB3E, 0x0000FB3E},
-{0x0000FB40, 0x0000FB41}, {0x0000FB43, 0x0000FB44}, {0x0000FB46, 0x0000FBB1}, {0x0000FBD3, 0x0000FD3D},
-{0x0000FD50, 0x0000FD8F}, {0x0000FD92, 0x0000FDC7}, {0x0000FDF0, 0x0000FDFB}, {0x0000FE70, 0x0000FE74},
-{0x0000FE76, 0x0000FEFC}, {0x0000FF21, 0x0000FF3A}, {0x0000FF41, 0x0000FF5A}, {0x0000FF66, 0x0000FFBE},
-{0x0000FFC2, 0x0000FFC7}, {0x0000FFCA, 0x0000FFCF}, {0x0000FFD2, 0x0000FFD7}, {0x0000FFDA, 0x0000FFDC},
-{0x00010000, 0x0001000B}, {0x0001000D, 0x00010026}, {0x00010028, 0x0001003A}, {0x0001003C, 0x0001003D},
-{0x0001003F, 0x0001004D}, {0x00010050, 0x0001005D}, {0x00010080, 0x000100FA}, {0x00010280, 0x0001029C},
-{0x000102A0, 0x000102D0}, {0x00010300, 0x0001031F}, {0x0001032D, 0x00010340}, {0x00010342, 0x00010349},
-{0x00010350, 0x00010375}, {0x00010380, 0x0001039D}, {0x000103A0, 0x000103C3}, {0x000103C8, 0x000103CF},
-{0x00010400, 0x0001049D}, {0x000104B0, 0x000104D3}, {0x000104D8, 0x000104FB}, {0x00010500, 0x00010527},
-{0x00010530, 0x00010563}, {0x00010600, 0x00010736}, {0x00010740, 0x00010755}, {0x00010760, 0x00010767},
-{0x00010800, 0x00010805}, {0x00010808, 0x00010808}, {0x0001080A, 0x00010835}, {0x00010837, 0x00010838},
-{0x0001083C, 0x0001083C}, {0x0001083F, 0x00010855}, {0x00010860, 0x00010876}, {0x00010880, 0x0001089E},
-{0x000108E0, 0x000108F2}, {0x000108F4, 0x000108F5}, {0x00010900, 0x00010915}, {0x00010920, 0x00010939},
-{0x00010980, 0x000109B7}, {0x000109BE, 0x000109BF}, {0x00010A00, 0x00010A00}, {0x00010A10, 0x00010A13},
-{0x00010A15, 0x00010A17}, {0x00010A19, 0x00010A35}, {0x00010A60, 0x00010A7C}, {0x00010A80, 0x00010A9C},
-{0x00010AC0, 0x00010AC7}, {0x00010AC9, 0x00010AE4}, {0x00010B00, 0x00010B35}, {0x00010B40, 0x00010B55},
-{0x00010B60, 0x00010B72}, {0x00010B80, 0x00010B91}, {0x00010C00, 0x00010C48}, {0x00010C80, 0x00010CB2},
-{0x00010CC0, 0x00010CF2}, {0x00010D00, 0x00010D23}, {0x00010E80, 0x00010EA9}, {0x00010EB0, 0x00010EB1},
-{0x00010F00, 0x00010F1C}, {0x00010F27, 0x00010F27}, {0x00010F30, 0x00010F45}, {0x00010FB0, 0x00010FC4},
-{0x00010FE0, 0x00010FF6}, {0x00011003, 0x00011037}, {0x00011083, 0x000110AF}, {0x000110D0, 0x000110E8},
-{0x00011103, 0x00011126}, {0x00011144, 0x00011144}, {0x00011147, 0x00011147}, {0x00011150, 0x00011172},
-{0x00011176, 0x00011176}, {0x00011183, 0x000111B2}, {0x000111C1, 0x000111C4}, {0x000111DA, 0x000111DA},
-{0x000111DC, 0x000111DC}, {0x00011200, 0x00011211}, {0x00011213, 0x0001122B}, {0x00011280, 0x00011286},
-{0x00011288, 0x00011288}, {0x0001128A, 0x0001128D}, {0x0001128F, 0x0001129D}, {0x0001129F, 0x000112A8},
-{0x000112B0, 0x000112DE}, {0x00011305, 0x0001130C}, {0x0001130F, 0x00011310}, {0x00011313, 0x00011328},
-{0x0001132A, 0x00011330}, {0x00011332, 0x00011333}, {0x00011335, 0x00011339}, {0x0001133D, 0x0001133D},
-{0x00011350, 0x00011350}, {0x0001135D, 0x00011361}, {0x00011400, 0x00011434}, {0x00011447, 0x0001144A},
-{0x0001145F, 0x00011461}, {0x00011480, 0x000114AF}, {0x000114C4, 0x000114C5}, {0x000114C7, 0x000114C7},
-{0x00011580, 0x000115AE}, {0x000115D8, 0x000115DB}, {0x00011600, 0x0001162F}, {0x00011644, 0x00011644},
-{0x00011680, 0x000116AA}, {0x000116B8, 0x000116B8}, {0x00011700, 0x0001171A}, {0x00011800, 0x0001182B},
-{0x000118A0, 0x000118DF}, {0x000118FF, 0x00011906}, {0x00011909, 0x00011909}, {0x0001190C, 0x00011913},
-{0x00011915, 0x00011916}, {0x00011918, 0x0001192F}, {0x0001193F, 0x0001193F}, {0x00011941, 0x00011941},
-{0x000119A0, 0x000119A7}, {0x000119AA, 0x000119D0}, {0x000119E1, 0x000119E1}, {0x000119E3, 0x000119E3},
-{0x00011A00, 0x00011A00}, {0x00011A0B, 0x00011A32}, {0x00011A3A, 0x00011A3A}, {0x00011A50, 0x00011A50},
-{0x00011A5C, 0x00011A89}, {0x00011A9D, 0x00011A9D}, {0x00011AC0, 0x00011AF8}, {0x00011C00, 0x00011C08},
-{0x00011C0A, 0x00011C2E}, {0x00011C40, 0x00011C40}, {0x00011C72, 0x00011C8F}, {0x00011D00, 0x00011D06},
-{0x00011D08, 0x00011D09}, {0x00011D0B, 0x00011D30}, {0x00011D46, 0x00011D46}, {0x00011D60, 0x00011D65},
-{0x00011D67, 0x00011D68}, {0x00011D6A, 0x00011D89}, {0x00011D98, 0x00011D98}, {0x00011EE0, 0x00011EF2},
-{0x00011FB0, 0x00011FB0}, {0x00012000, 0x00012399}, {0x00012480, 0x00012543}, {0x00013000, 0x0001342E},
-{0x00014400, 0x00014646}, {0x00016800, 0x00016A38}, {0x00016A40, 0x00016A5E}, {0x00016AD0, 0x00016AED},
-{0x00016B00, 0x00016B2F}, {0x00016B40, 0x00016B43}, {0x00016B63, 0x00016B77}, {0x00016B7D, 0x00016B8F},
-{0x00016E40, 0x00016E7F}, {0x00016F00, 0x00016F4A}, {0x00016F50, 0x00016F50}, {0x00016F93, 0x00016F9F},
-{0x00016FE0, 0x00016FE1}, {0x00016FE3, 0x00016FE3}, {0x00017000, 0x000187F7}, {0x00018800, 0x00018CD5},
-{0x00018D00, 0x00018D08}, {0x0001B000, 0x0001B11E}, {0x0001B150, 0x0001B152}, {0x0001B164, 0x0001B167},
-{0x0001B170, 0x0001B2FB}, {0x0001BC00, 0x0001BC6A}, {0x0001BC70, 0x0001BC7C}, {0x0001BC80, 0x0001BC88},
-{0x0001BC90, 0x0001BC99}, {0x0001D400, 0x0001D454}, {0x0001D456, 0x0001D49C}, {0x0001D49E, 0x0001D49F},
-{0x0001D4A2, 0x0001D4A2}, {0x0001D4A5, 0x0001D4A6}, {0x0001D4A9, 0x0001D4AC}, {0x0001D4AE, 0x0001D4B9},
-{0x0001D4BB, 0x0001D4BB}, {0x0001D4BD, 0x0001D4C3}, {0x0001D4C5, 0x0001D505}, {0x0001D507, 0x0001D50A},
-{0x0001D50D, 0x0001D514}, {0x0001D516, 0x0001D51C}, {0x0001D51E, 0x0001D539}, {0x0001D53B, 0x0001D53E},
-{0x0001D540, 0x0001D544}, {0x0001D546, 0x0001D546}, {0x0001D54A, 0x0001D550}, {0x0001D552, 0x0001D6A5},
-{0x0001D6A8, 0x0001D6C0}, {0x0001D6C2, 0x0001D6DA}, {0x0001D6DC, 0x0001D6FA}, {0x0001D6FC, 0x0001D714},
-{0x0001D716, 0x0001D734}, {0x0001D736, 0x0001D74E}, {0x0001D750, 0x0001D76E}, {0x0001D770, 0x0001D788},
-{0x0001D78A, 0x0001D7A8}, {0x0001D7AA, 0x0001D7C2}, {0x0001D7C4, 0x0001D7CB}, {0x0001E100, 0x0001E12C},
-{0x0001E137, 0x0001E13D}, {0x0001E14E, 0x0001E14E}, {0x0001E2C0, 0x0001E2EB}, {0x0001E800, 0x0001E8C4},
-{0x0001E900, 0x0001E943}, {0x0001E94B, 0x0001E94B}, {0x0001EE00, 0x0001EE03}, {0x0001EE05, 0x0001EE1F},
-{0x0001EE21, 0x0001EE22}, {0x0001EE24, 0x0001EE24}, {0x0001EE27, 0x0001EE27}, {0x0001EE29, 0x0001EE32},
-{0x0001EE34, 0x0001EE37}, {0x0001EE39, 0x0001EE39}, {0x0001EE3B, 0x0001EE3B}, {0x0001EE42, 0x0001EE42},
-{0x0001EE47, 0x0001EE47}, {0x0001EE49, 0x0001EE49}, {0x0001EE4B, 0x0001EE4B}, {0x0001EE4D, 0x0001EE4F},
-{0x0001EE51, 0x0001EE52}, {0x0001EE54, 0x0001EE54}, {0x0001EE57, 0x0001EE57}, {0x0001EE59, 0x0001EE59},
-{0x0001EE5B, 0x0001EE5B}, {0x0001EE5D, 0x0001EE5D}, {0x0001EE5F, 0x0001EE5F}, {0x0001EE61, 0x0001EE62},
-{0x0001EE64, 0x0001EE64}, {0x0001EE67, 0x0001EE6A}, {0x0001EE6C, 0x0001EE72}, {0x0001EE74, 0x0001EE77},
-{0x0001EE79, 0x0001EE7C}, {0x0001EE7E, 0x0001EE7E}, {0x0001EE80, 0x0001EE89}, {0x0001EE8B, 0x0001EE9B},
-{0x0001EEA1, 0x0001EEA3}, {0x0001EEA5, 0x0001EEA9}, {0x0001EEAB, 0x0001EEBB}, {0x00020000, 0x0002A6DD},
-{0x0002A700, 0x0002B734}, {0x0002B740, 0x0002B81D}, {0x0002B820, 0x0002CEA1}, {0x0002CEB0, 0x0002EBE0},
-{0x0002F800, 0x0002FA1D}, {0x00030000, 0x0003134A},
-};
-
-const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace = {
-{0x00000009, 0x0000000D}, {0x0000001C, 0x00000020}, {0x00000085, 0x00000085}, {0x000000A0, 0x000000A0},
-{0x00001680, 0x00001680}, {0x00002000, 0x0000200A}, {0x00002028, 0x00002029}, {0x0000202F, 0x0000202F},
-{0x0000205F, 0x0000205F}, {0x00003000, 0x00003000},
-};
-
-const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark = {
-{0x00000300, 0x0000036F}, {0x00000483, 0x00000489}, {0x00000591, 0x000005BD}, {0x000005BF, 0x000005BF},
-{0x000005C1, 0x000005C2}, {0x000005C4, 0x000005C5}, {0x000005C7, 0x000005C7}, {0x00000610, 0x0000061A},
-{0x0000064B, 0x0000065F}, {0x00000670, 0x00000670}, {0x000006D6, 0x000006DC}, {0x000006DF, 0x000006E4},
-{0x000006E7, 0x000006E8}, {0x000006EA, 0x000006ED}, {0x00000711, 0x00000711}, {0x00000730, 0x0000074A},
-{0x000007A6, 0x000007B0}, {0x000007EB, 0x000007F3}, {0x000007FD, 0x000007FD}, {0x00000816, 0x00000819},
-{0x0000081B, 0x00000823}, {0x00000825, 0x00000827}, {0x00000829, 0x0000082D}, {0x00000859, 0x0000085B},
-{0x000008D3, 0x000008E1}, {0x000008E3, 0x00000903}, {0x0000093A, 0x0000093C}, {0x0000093E, 0x0000094F},
-{0x00000951, 0x00000957}, {0x00000962, 0x00000963}, {0x00000981, 0x00000983}, {0x000009BC, 0x000009BC},
-{0x000009BE, 0x000009C4}, {0x000009C7, 0x000009C8}, {0x000009CB, 0x000009CD}, {0x000009D7, 0x000009D7},
-{0x000009E2, 0x000009E3}, {0x000009FE, 0x000009FE}, {0x00000A01, 0x00000A03}, {0x00000A3C, 0x00000A3C},
-{0x00000A3E, 0x00000A42}, {0x00000A47, 0x00000A48}, {0x00000A4B, 0x00000A4D}, {0x00000A51, 0x00000A51},
-{0x00000A70, 0x00000A71}, {0x00000A75, 0x00000A75}, {0x00000A81, 0x00000A83}, {0x00000ABC, 0x00000ABC},
-{0x00000ABE, 0x00000AC5}, {0x00000AC7, 0x00000AC9}, {0x00000ACB, 0x00000ACD}, {0x00000AE2, 0x00000AE3},
-{0x00000AFA, 0x00000AFF}, {0x00000B01, 0x00000B03}, {0x00000B3C, 0x00000B3C}, {0x00000B3E, 0x00000B44},
-{0x00000B47, 0x00000B48}, {0x00000B4B, 0x00000B4D}, {0x00000B55, 0x00000B57}, {0x00000B62, 0x00000B63},
-{0x00000B82, 0x00000B82}, {0x00000BBE, 0x00000BC2}, {0x00000BC6, 0x00000BC8}, {0x00000BCA, 0x00000BCD},
-{0x00000BD7, 0x00000BD7}, {0x00000C00, 0x00000C04}, {0x00000C3E, 0x00000C44}, {0x00000C46, 0x00000C48},
-{0x00000C4A, 0x00000C4D}, {0x00000C55, 0x00000C56}, {0x00000C62, 0x00000C63}, {0x00000C81, 0x00000C83},
-{0x00000CBC, 0x00000CBC}, {0x00000CBE, 0x00000CC4}, {0x00000CC6, 0x00000CC8}, {0x00000CCA, 0x00000CCD},
-{0x00000CD5, 0x00000CD6}, {0x00000CE2, 0x00000CE3}, {0x00000D00, 0x00000D03}, {0x00000D3B, 0x00000D3C},
-{0x00000D3E, 0x00000D44}, {0x00000D46, 0x00000D48}, {0x00000D4A, 0x00000D4D}, {0x00000D57, 0x00000D57},
-{0x00000D62, 0x00000D63}, {0x00000D81, 0x00000D83}, {0x00000DCA, 0x00000DCA}, {0x00000DCF, 0x00000DD4},
-{0x00000DD6, 0x00000DD6}, {0x00000DD8, 0x00000DDF}, {0x00000DF2, 0x00000DF3}, {0x00000E31, 0x00000E31},
-{0x00000E34, 0x00000E3A}, {0x00000E47, 0x00000E4E}, {0x00000EB1, 0x00000EB1}, {0x00000EB4, 0x00000EBC},
-{0x00000EC8, 0x00000ECD}, {0x00000F18, 0x00000F19}, {0x00000F35, 0x00000F35}, {0x00000F37, 0x00000F37},
-{0x00000F39, 0x00000F39}, {0x00000F3E, 0x00000F3F}, {0x00000F71, 0x00000F84}, {0x00000F86, 0x00000F87},
-{0x00000F8D, 0x00000F97}, {0x00000F99, 0x00000FBC}, {0x00000FC6, 0x00000FC6}, {0x0000102B, 0x0000103E},
-{0x00001056, 0x00001059}, {0x0000105E, 0x00001060}, {0x00001062, 0x00001064}, {0x00001067, 0x0000106D},
-{0x00001071, 0x00001074}, {0x00001082, 0x0000108D}, {0x0000108F, 0x0000108F}, {0x0000109A, 0x0000109D},
-{0x0000135D, 0x0000135F}, {0x00001712, 0x00001714}, {0x00001732, 0x00001734}, {0x00001752, 0x00001753},
-{0x00001772, 0x00001773}, {0x000017B4, 0x000017D3}, {0x000017DD, 0x000017DD}, {0x0000180B, 0x0000180D},
-{0x00001885, 0x00001886}, {0x000018A9, 0x000018A9}, {0x00001920, 0x0000192B}, {0x00001930, 0x0000193B},
-{0x00001A17, 0x00001A1B}, {0x00001A55, 0x00001A5E}, {0x00001A60, 0x00001A7C}, {0x00001A7F, 0x00001A7F},
-{0x00001AB0, 0x00001AC0}, {0x00001B00, 0x00001B04}, {0x00001B34, 0x00001B44}, {0x00001B6B, 0x00001B73},
-{0x00001B80, 0x00001B82}, {0x00001BA1, 0x00001BAD}, {0x00001BE6, 0x00001BF3}, {0x00001C24, 0x00001C37},
-{0x00001CD0, 0x00001CD2}, {0x00001CD4, 0x00001CE8}, {0x00001CED, 0x00001CED}, {0x00001CF4, 0x00001CF4},
-{0x00001CF7, 0x00001CF9}, {0x00001DC0, 0x00001DF9}, {0x00001DFB, 0x00001DFF}, {0x000020D0, 0x000020F0},
-{0x00002CEF, 0x00002CF1}, {0x00002D7F, 0x00002D7F}, {0x00002DE0, 0x00002DFF}, {0x0000302A, 0x0000302F},
-{0x00003099, 0x0000309A}, {0x0000A66F, 0x0000A672}, {0x0000A674, 0x0000A67D}, {0x0000A69E, 0x0000A69F},
-{0x0000A6F0, 0x0000A6F1}, {0x0000A802, 0x0000A802}, {0x0000A806, 0x0000A806}, {0x0000A80B, 0x0000A80B},
-{0x0000A823, 0x0000A827}, {0x0000A82C, 0x0000A82C}, {0x0000A880, 0x0000A881}, {0x0000A8B4, 0x0000A8C5},
-{0x0000A8E0, 0x0000A8F1}, {0x0000A8FF, 0x0000A8FF}, {0x0000A926, 0x0000A92D}, {0x0000A947, 0x0000A953},
-{0x0000A980, 0x0000A983}, {0x0000A9B3, 0x0000A9C0}, {0x0000A9E5, 0x0000A9E5}, {0x0000AA29, 0x0000AA36},
-{0x0000AA43, 0x0000AA43}, {0x0000AA4C, 0x0000AA4D}, {0x0000AA7B, 0x0000AA7D}, {0x0000AAB0, 0x0000AAB0},
-{0x0000AAB2, 0x0000AAB4}, {0x0000AAB7, 0x0000AAB8}, {0x0000AABE, 0x0000AABF}, {0x0000AAC1, 0x0000AAC1},
-{0x0000AAEB, 0x0000AAEF}, {0x0000AAF5, 0x0000AAF6}, {0x0000ABE3, 0x0000ABEA}, {0x0000ABEC, 0x0000ABED},
-{0x0000FB1E, 0x0000FB1E}, {0x0000FE00, 0x0000FE0F}, {0x0000FE20, 0x0000FE2F}, {0x000101FD, 0x000101FD},
-{0x000102E0, 0x000102E0}, {0x00010376, 0x0001037A}, {0x00010A01, 0x00010A03}, {0x00010A05, 0x00010A06},
-{0x00010A0C, 0x00010A0F}, {0x00010A38, 0x00010A3A}, {0x00010A3F, 0x00010A3F}, {0x00010AE5, 0x00010AE6},
-{0x00010D24, 0x00010D27}, {0x00010EAB, 0x00010EAC}, {0x00010F46, 0x00010F50}, {0x00011000, 0x00011002},
-{0x00011038, 0x00011046}, {0x0001107F, 0x00011082}, {0x000110B0, 0x000110BA}, {0x00011100, 0x00011102},
-{0x00011127, 0x00011134}, {0x00011145, 0x00011146}, {0x00011173, 0x00011173}, {0x00011180, 0x00011182},
-{0x000111B3, 0x000111C0}, {0x000111C9, 0x000111CC}, {0x000111CE, 0x000111CF}, {0x0001122C, 0x00011237},
-{0x0001123E, 0x0001123E}, {0x000112DF, 0x000112EA}, {0x00011300, 0x00011303}, {0x0001133B, 0x0001133C},
-{0x0001133E, 0x00011344}, {0x00011347, 0x00011348}, {0x0001134B, 0x0001134D}, {0x00011357, 0x00011357},
-{0x00011362, 0x00011363}, {0x00011366, 0x0001136C}, {0x00011370, 0x00011374}, {0x00011435, 0x00011446},
-{0x0001145E, 0x0001145E}, {0x000114B0, 0x000114C3}, {0x000115AF, 0x000115B5}, {0x000115B8, 0x000115C0},
-{0x000115DC, 0x000115DD}, {0x00011630, 0x00011640}, {0x000116AB, 0x000116B7}, {0x0001171D, 0x0001172B},
-{0x0001182C, 0x0001183A}, {0x00011930, 0x00011935}, {0x00011937, 0x00011938}, {0x0001193B, 0x0001193E},
-{0x00011940, 0x00011940}, {0x00011942, 0x00011943}, {0x000119D1, 0x000119D7}, {0x000119DA, 0x000119E0},
-{0x000119E4, 0x000119E4}, {0x00011A01, 0x00011A0A}, {0x00011A33, 0x00011A39}, {0x00011A3B, 0x00011A3E},
-{0x00011A47, 0x00011A47}, {0x00011A51, 0x00011A5B}, {0x00011A8A, 0x00011A99}, {0x00011C2F, 0x00011C36},
-{0x00011C38, 0x00011C3F}, {0x00011C92, 0x00011CA7}, {0x00011CA9, 0x00011CB6}, {0x00011D31, 0x00011D36},
-{0x00011D3A, 0x00011D3A}, {0x00011D3C, 0x00011D3D}, {0x00011D3F, 0x00011D45}, {0x00011D47, 0x00011D47},
-{0x00011D8A, 0x00011D8E}, {0x00011D90, 0x00011D91}, {0x00011D93, 0x00011D97}, {0x00011EF3, 0x00011EF6},
-{0x00016AF0, 0x00016AF4}, {0x00016B30, 0x00016B36}, {0x00016F4F, 0x00016F4F}, {0x00016F51, 0x00016F87},
-{0x00016F8F, 0x00016F92}, {0x00016FE4, 0x00016FE4}, {0x00016FF0, 0x00016FF1}, {0x0001BC9D, 0x0001BC9E},
-{0x0001D165, 0x0001D169}, {0x0001D16D, 0x0001D172}, {0x0001D17B, 0x0001D182}, {0x0001D185, 0x0001D18B},
-{0x0001D1AA, 0x0001D1AD}, {0x0001D242, 0x0001D244}, {0x0001DA00, 0x0001DA36}, {0x0001DA3B, 0x0001DA6C},
-{0x0001DA75, 0x0001DA75}, {0x0001DA84, 0x0001DA84}, {0x0001DA9B, 0x0001DA9F}, {0x0001DAA1, 0x0001DAAF},
-{0x0001E000, 0x0001E006}, {0x0001E008, 0x0001E018}, {0x0001E01B, 0x0001E021}, {0x0001E023, 0x0001E024},
-{0x0001E026, 0x0001E02A}, {0x0001E130, 0x0001E136}, {0x0001E2EC, 0x0001E2EF}, {0x0001E8D0, 0x0001E8D6},
-{0x0001E944, 0x0001E94A}, {0x000E0100, 0x000E01EF},
+const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags = {  // start, flags // last=next_start-1
+{0x000000, 0x0080},
+{0x000020, 0x0008},
+{0x000021, 0x0020},
+{0x000024, 0x0040},
+{0x000025, 0x0020},
+{0x00002B, 0x0040},
+{0x00002C, 0x0020},
+{0x000030, 0x0002},
+{0x00003A, 0x0020},
+{0x00003C, 0x0040},
+{0x00003F, 0x0020},
+{0x000041, 0x0004},
+{0x00005B, 0x0020},
+{0x00005E, 0x0040},
+{0x00005F, 0x0020},
+{0x000060, 0x0040},
+{0x000061, 0x0004},
+{0x00007B, 0x0020},
+{0x00007C, 0x0040},
+{0x00007D, 0x0020},
+{0x00007E, 0x0040},
+{0x00007F, 0x0080},
+{0x0000A0, 0x0008},
+{0x0000A1, 0x0020},
+{0x0000A2, 0x0040},
+{0x0000A7, 0x0020},
+{0x0000A8, 0x0040},
+{0x0000AA, 0x0004},
+{0x0000AB, 0x0020},
+{0x0000AC, 0x0040},
+{0x0000AD, 0x0080},
+{0x0000AE, 0x0040},
+{0x0000B2, 0x0002},
+{0x0000B4, 0x0040},
+{0x0000B5, 0x0004},
+{0x0000B6, 0x0020},
+{0x0000B8, 0x0040},
+{0x0000B9, 0x0002},
+{0x0000BA, 0x0004},
+{0x0000BB, 0x0020},
+{0x0000BC, 0x0002},
+{0x0000BF, 0x0020},
+{0x0000C0, 0x0004},
+{0x0000D7, 0x0040},
+{0x0000D8, 0x0004},
+{0x0000F7, 0x0040},
+{0x0000F8, 0x0004},
+{0x0002C2, 0x0040},
+{0x0002C6, 0x0004},
+{0x0002D2, 0x0040},
+{0x0002E0, 0x0004},
+{0x0002E5, 0x0040},
+{0x0002EC, 0x0004},
+{0x0002ED, 0x0040},
+{0x0002EE, 0x0004},
+{0x0002EF, 0x0040},
+{0x000300, 0x0010},
+{0x000370, 0x0004},
+{0x000375, 0x0040},
+{0x000376, 0x0004},
+{0x000378, 0x0080},
+{0x00037A, 0x0004},
+{0x00037E, 0x0020},
+{0x00037F, 0x0004},
+{0x000380, 0x0080},
+{0x000384, 0x0040},
+{0x000386, 0x0004},
+{0x000387, 0x0020},
+{0x000388, 0x0004},
+{0x00038B, 0x0080},
+{0x00038C, 0x0004},
+{0x00038D, 0x0080},
+{0x00038E, 0x0004},
+{0x0003A2, 0x0080},
+{0x0003A3, 0x0004},
+{0x0003F6, 0x0040},
+{0x0003F7, 0x0004},
+{0x000482, 0x0040},
+{0x000483, 0x0010},
+{0x00048A, 0x0004},
+{0x000530, 0x0080},
+{0x000531, 0x0004},
+{0x000557, 0x0080},
+{0x000559, 0x0004},
+{0x00055A, 0x0020},
+{0x000560, 0x0004},
+{0x000589, 0x0020},
+{0x00058B, 0x0080},
+{0x00058D, 0x0040},
+{0x000590, 0x0080},
+{0x000591, 0x0010},
+{0x0005BE, 0x0020},
+{0x0005BF, 0x0010},
+{0x0005C0, 0x0020},
+{0x0005C1, 0x0010},
+{0x0005C3, 0x0020},
+{0x0005C4, 0x0010},
+{0x0005C6, 0x0020},
+{0x0005C7, 0x0010},
+{0x0005C8, 0x0080},
+{0x0005D0, 0x0004},
+{0x0005EB, 0x0080},
+{0x0005EF, 0x0004},
+{0x0005F3, 0x0020},
+{0x0005F5, 0x0080},
+{0x000606, 0x0040},
+{0x000609, 0x0020},
+{0x00060B, 0x0040},
+{0x00060C, 0x0020},
+{0x00060E, 0x0040},
+{0x000610, 0x0010},
+{0x00061B, 0x0020},
+{0x00061C, 0x0080},
+{0x00061D, 0x0020},
+{0x000620, 0x0004},
+{0x00064B, 0x0010},
+{0x000660, 0x0002},
+{0x00066A, 0x0020},
+{0x00066E, 0x0004},
+{0x000670, 0x0010},
+{0x000671, 0x0004},
+{0x0006D4, 0x0020},
+{0x0006D5, 0x0004},
+{0x0006D6, 0x0010},
+{0x0006DD, 0x0080},
+{0x0006DE, 0x0040},
+{0x0006DF, 0x0010},
+{0x0006E5, 0x0004},
+{0x0006E7, 0x0010},
+{0x0006E9, 0x0040},
+{0x0006EA, 0x0010},
+{0x0006EE, 0x0004},
+{0x0006F0, 0x0002},
+{0x0006FA, 0x0004},
+{0x0006FD, 0x0040},
+{0x0006FF, 0x0004},
+{0x000700, 0x0020},
+{0x00070E, 0x0080},
+{0x000710, 0x0004},
+{0x000711, 0x0010},
+{0x000712, 0x0004},
+{0x000730, 0x0010},
+{0x00074B, 0x0080},
+{0x00074D, 0x0004},
+{0x0007A6, 0x0010},
+{0x0007B1, 0x0004},
+{0x0007B2, 0x0080},
+{0x0007C0, 0x0002},
+{0x0007CA, 0x0004},
+{0x0007EB, 0x0010},
+{0x0007F4, 0x0004},
+{0x0007F6, 0x0040},
+{0x0007F7, 0x0020},
+{0x0007FA, 0x0004},
+{0x0007FB, 0x0080},
+{0x0007FD, 0x0010},
+{0x0007FE, 0x0040},
+{0x000800, 0x0004},
+{0x000816, 0x0010},
+{0x00081A, 0x0004},
+{0x00081B, 0x0010},
+{0x000824, 0x0004},
+{0x000825, 0x0010},
+{0x000828, 0x0004},
+{0x000829, 0x0010},
+{0x00082E, 0x0080},
+{0x000830, 0x0020},
+{0x00083F, 0x0080},
+{0x000840, 0x0004},
+{0x000859, 0x0010},
+{0x00085C, 0x0080},
+{0x00085E, 0x0020},
+{0x00085F, 0x0080},
+{0x000860, 0x0004},
+{0x00086B, 0x0080},
+{0x000870, 0x0004},
+{0x000888, 0x0040},
+{0x000889, 0x0004},
+{0x00088F, 0x0080},
+{0x000898, 0x0010},
+{0x0008A0, 0x0004},
+{0x0008CA, 0x0010},
+{0x0008E2, 0x0080},
+{0x0008E3, 0x0010},
+{0x000904, 0x0004},
+{0x00093A, 0x0010},
+{0x00093D, 0x0004},
+{0x00093E, 0x0010},
+{0x000950, 0x0004},
+{0x000951, 0x0010},
+{0x000958, 0x0004},
+{0x000962, 0x0010},
+{0x000964, 0x0020},
+{0x000966, 0x0002},
+{0x000970, 0x0020},
+{0x000971, 0x0004},
+{0x000981, 0x0010},
+{0x000984, 0x0080},
+{0x000985, 0x0004},
+{0x00098D, 0x0080},
+{0x00098F, 0x0004},
+{0x000991, 0x0080},
+{0x000993, 0x0004},
+{0x0009A9, 0x0080},
+{0x0009AA, 0x0004},
+{0x0009B1, 0x0080},
+{0x0009B2, 0x0004},
+{0x0009B3, 0x0080},
+{0x0009B6, 0x0004},
+{0x0009BA, 0x0080},
+{0x0009BC, 0x0010},
+{0x0009BD, 0x0004},
+{0x0009BE, 0x0010},
+{0x0009C5, 0x0080},
+{0x0009C7, 0x0010},
+{0x0009C9, 0x0080},
+{0x0009CB, 0x0010},
+{0x0009CE, 0x0004},
+{0x0009CF, 0x0080},
+{0x0009D7, 0x0010},
+{0x0009D8, 0x0080},
+{0x0009DC, 0x0004},
+{0x0009DE, 0x0080},
+{0x0009DF, 0x0004},
+{0x0009E2, 0x0010},
+{0x0009E4, 0x0080},
+{0x0009E6, 0x0002},
+{0x0009F0, 0x0004},
+{0x0009F2, 0x0040},
+{0x0009F4, 0x0002},
+{0x0009FA, 0x0040},
+{0x0009FC, 0x0004},
+{0x0009FD, 0x0020},
+{0x0009FE, 0x0010},
+{0x0009FF, 0x0080},
+{0x000A01, 0x0010},
+{0x000A04, 0x0080},
+{0x000A05, 0x0004},
+{0x000A0B, 0x0080},
+{0x000A0F, 0x0004},
+{0x000A11, 0x0080},
+{0x000A13, 0x0004},
+{0x000A29, 0x0080},
+{0x000A2A, 0x0004},
+{0x000A31, 0x0080},
+{0x000A32, 0x0004},
+{0x000A34, 0x0080},
+{0x000A35, 0x0004},
+{0x000A37, 0x0080},
+{0x000A38, 0x0004},
+{0x000A3A, 0x0080},
+{0x000A3C, 0x0010},
+{0x000A3D, 0x0080},
+{0x000A3E, 0x0010},
+{0x000A43, 0x0080},
+{0x000A47, 0x0010},
+{0x000A49, 0x0080},
+{0x000A4B, 0x0010},
+{0x000A4E, 0x0080},
+{0x000A51, 0x0010},
+{0x000A52, 0x0080},
+{0x000A59, 0x0004},
+{0x000A5D, 0x0080},
+{0x000A5E, 0x0004},
+{0x000A5F, 0x0080},
+{0x000A66, 0x0002},
+{0x000A70, 0x0010},
+{0x000A72, 0x0004},
+{0x000A75, 0x0010},
+{0x000A76, 0x0020},
+{0x000A77, 0x0080},
+{0x000A81, 0x0010},
+{0x000A84, 0x0080},
+{0x000A85, 0x0004},
+{0x000A8E, 0x0080},
+{0x000A8F, 0x0004},
+{0x000A92, 0x0080},
+{0x000A93, 0x0004},
+{0x000AA9, 0x0080},
+{0x000AAA, 0x0004},
+{0x000AB1, 0x0080},
+{0x000AB2, 0x0004},
+{0x000AB4, 0x0080},
+{0x000AB5, 0x0004},
+{0x000ABA, 0x0080},
+{0x000ABC, 0x0010},
+{0x000ABD, 0x0004},
+{0x000ABE, 0x0010},
+{0x000AC6, 0x0080},
+{0x000AC7, 0x0010},
+{0x000ACA, 0x0080},
+{0x000ACB, 0x0010},
+{0x000ACE, 0x0080},
+{0x000AD0, 0x0004},
+{0x000AD1, 0x0080},
+{0x000AE0, 0x0004},
+{0x000AE2, 0x0010},
+{0x000AE4, 0x0080},
+{0x000AE6, 0x0002},
+{0x000AF0, 0x0020},
+{0x000AF1, 0x0040},
+{0x000AF2, 0x0080},
+{0x000AF9, 0x0004},
+{0x000AFA, 0x0010},
+{0x000B00, 0x0080},
+{0x000B01, 0x0010},
+{0x000B04, 0x0080},
+{0x000B05, 0x0004},
+{0x000B0D, 0x0080},
+{0x000B0F, 0x0004},
+{0x000B11, 0x0080},
+{0x000B13, 0x0004},
+{0x000B29, 0x0080},
+{0x000B2A, 0x0004},
+{0x000B31, 0x0080},
+{0x000B32, 0x0004},
+{0x000B34, 0x0080},
+{0x000B35, 0x0004},
+{0x000B3A, 0x0080},
+{0x000B3C, 0x0010},
+{0x000B3D, 0x0004},
+{0x000B3E, 0x0010},
+{0x000B45, 0x0080},
+{0x000B47, 0x0010},
+{0x000B49, 0x0080},
+{0x000B4B, 0x0010},
+{0x000B4E, 0x0080},
+{0x000B55, 0x0010},
+{0x000B58, 0x0080},
+{0x000B5C, 0x0004},
+{0x000B5E, 0x0080},
+{0x000B5F, 0x0004},
+{0x000B62, 0x0010},
+{0x000B64, 0x0080},
+{0x000B66, 0x0002},
+{0x000B70, 0x0040},
+{0x000B71, 0x0004},
+{0x000B72, 0x0002},
+{0x000B78, 0x0080},
+{0x000B82, 0x0010},
+{0x000B83, 0x0004},
+{0x000B84, 0x0080},
+{0x000B85, 0x0004},
+{0x000B8B, 0x0080},
+{0x000B8E, 0x0004},
+{0x000B91, 0x0080},
+{0x000B92, 0x0004},
+{0x000B96, 0x0080},
+{0x000B99, 0x0004},
+{0x000B9B, 0x0080},
+{0x000B9C, 0x0004},
+{0x000B9D, 0x0080},
+{0x000B9E, 0x0004},
+{0x000BA0, 0x0080},
+{0x000BA3, 0x0004},
+{0x000BA5, 0x0080},
+{0x000BA8, 0x0004},
+{0x000BAB, 0x0080},
+{0x000BAE, 0x0004},
+{0x000BBA, 0x0080},
+{0x000BBE, 0x0010},
+{0x000BC3, 0x0080},
+{0x000BC6, 0x0010},
+{0x000BC9, 0x0080},
+{0x000BCA, 0x0010},
+{0x000BCE, 0x0080},
+{0x000BD0, 0x0004},
+{0x000BD1, 0x0080},
+{0x000BD7, 0x0010},
+{0x000BD8, 0x0080},
+{0x000BE6, 0x0002},
+{0x000BF3, 0x0040},
+{0x000BFB, 0x0080},
+{0x000C00, 0x0010},
+{0x000C05, 0x0004},
+{0x000C0D, 0x0080},
+{0x000C0E, 0x0004},
+{0x000C11, 0x0080},
+{0x000C12, 0x0004},
+{0x000C29, 0x0080},
+{0x000C2A, 0x0004},
+{0x000C3A, 0x0080},
+{0x000C3C, 0x0010},
+{0x000C3D, 0x0004},
+{0x000C3E, 0x0010},
+{0x000C45, 0x0080},
+{0x000C46, 0x0010},
+{0x000C49, 0x0080},
+{0x000C4A, 0x0010},
+{0x000C4E, 0x0080},
+{0x000C55, 0x0010},
+{0x000C57, 0x0080},
+{0x000C58, 0x0004},
+{0x000C5B, 0x0080},
+{0x000C5D, 0x0004},
+{0x000C5E, 0x0080},
+{0x000C60, 0x0004},
+{0x000C62, 0x0010},
+{0x000C64, 0x0080},
+{0x000C66, 0x0002},
+{0x000C70, 0x0080},
+{0x000C77, 0x0020},
+{0x000C78, 0x0002},
+{0x000C7F, 0x0040},
+{0x000C80, 0x0004},
+{0x000C81, 0x0010},
+{0x000C84, 0x0020},
+{0x000C85, 0x0004},
+{0x000C8D, 0x0080},
+{0x000C8E, 0x0004},
+{0x000C91, 0x0080},
+{0x000C92, 0x0004},
+{0x000CA9, 0x0080},
+{0x000CAA, 0x0004},
+{0x000CB4, 0x0080},
+{0x000CB5, 0x0004},
+{0x000CBA, 0x0080},
+{0x000CBC, 0x0010},
+{0x000CBD, 0x0004},
+{0x000CBE, 0x0010},
+{0x000CC5, 0x0080},
+{0x000CC6, 0x0010},
+{0x000CC9, 0x0080},
+{0x000CCA, 0x0010},
+{0x000CCE, 0x0080},
+{0x000CD5, 0x0010},
+{0x000CD7, 0x0080},
+{0x000CDD, 0x0004},
+{0x000CDF, 0x0080},
+{0x000CE0, 0x0004},
+{0x000CE2, 0x0010},
+{0x000CE4, 0x0080},
+{0x000CE6, 0x0002},
+{0x000CF0, 0x0080},
+{0x000CF1, 0x0004},
+{0x000CF3, 0x0010},
+{0x000CF4, 0x0080},
+{0x000D00, 0x0010},
+{0x000D04, 0x0004},
+{0x000D0D, 0x0080},
+{0x000D0E, 0x0004},
+{0x000D11, 0x0080},
+{0x000D12, 0x0004},
+{0x000D3B, 0x0010},
+{0x000D3D, 0x0004},
+{0x000D3E, 0x0010},
+{0x000D45, 0x0080},
+{0x000D46, 0x0010},
+{0x000D49, 0x0080},
+{0x000D4A, 0x0010},
+{0x000D4E, 0x0004},
+{0x000D4F, 0x0040},
+{0x000D50, 0x0080},
+{0x000D54, 0x0004},
+{0x000D57, 0x0010},
+{0x000D58, 0x0002},
+{0x000D5F, 0x0004},
+{0x000D62, 0x0010},
+{0x000D64, 0x0080},
+{0x000D66, 0x0002},
+{0x000D79, 0x0040},
+{0x000D7A, 0x0004},
+{0x000D80, 0x0080},
+{0x000D81, 0x0010},
+{0x000D84, 0x0080},
+{0x000D85, 0x0004},
+{0x000D97, 0x0080},
+{0x000D9A, 0x0004},
+{0x000DB2, 0x0080},
+{0x000DB3, 0x0004},
+{0x000DBC, 0x0080},
+{0x000DBD, 0x0004},
+{0x000DBE, 0x0080},
+{0x000DC0, 0x0004},
+{0x000DC7, 0x0080},
+{0x000DCA, 0x0010},
+{0x000DCB, 0x0080},
+{0x000DCF, 0x0010},
+{0x000DD5, 0x0080},
+{0x000DD6, 0x0010},
+{0x000DD7, 0x0080},
+{0x000DD8, 0x0010},
+{0x000DE0, 0x0080},
+{0x000DE6, 0x0002},
+{0x000DF0, 0x0080},
+{0x000DF2, 0x0010},
+{0x000DF4, 0x0020},
+{0x000DF5, 0x0080},
+{0x000E01, 0x0004},
+{0x000E31, 0x0010},
+{0x000E32, 0x0004},
+{0x000E34, 0x0010},
+{0x000E3B, 0x0080},
+{0x000E3F, 0x0040},
+{0x000E40, 0x0004},
+{0x000E47, 0x0010},
+{0x000E4F, 0x0020},
+{0x000E50, 0x0002},
+{0x000E5A, 0x0020},
+{0x000E5C, 0x0080},
+{0x000E81, 0x0004},
+{0x000E83, 0x0080},
+{0x000E84, 0x0004},
+{0x000E85, 0x0080},
+{0x000E86, 0x0004},
+{0x000E8B, 0x0080},
+{0x000E8C, 0x0004},
+{0x000EA4, 0x0080},
+{0x000EA5, 0x0004},
+{0x000EA6, 0x0080},
+{0x000EA7, 0x0004},
+{0x000EB1, 0x0010},
+{0x000EB2, 0x0004},
+{0x000EB4, 0x0010},
+{0x000EBD, 0x0004},
+{0x000EBE, 0x0080},
+{0x000EC0, 0x0004},
+{0x000EC5, 0x0080},
+{0x000EC6, 0x0004},
+{0x000EC7, 0x0080},
+{0x000EC8, 0x0010},
+{0x000ECF, 0x0080},
+{0x000ED0, 0x0002},
+{0x000EDA, 0x0080},
+{0x000EDC, 0x0004},
+{0x000EE0, 0x0080},
+{0x000F00, 0x0004},
+{0x000F01, 0x0040},
+{0x000F04, 0x0020},
+{0x000F13, 0x0040},
+{0x000F14, 0x0020},
+{0x000F15, 0x0040},
+{0x000F18, 0x0010},
+{0x000F1A, 0x0040},
+{0x000F20, 0x0002},
+{0x000F34, 0x0040},
+{0x000F35, 0x0010},
+{0x000F36, 0x0040},
+{0x000F37, 0x0010},
+{0x000F38, 0x0040},
+{0x000F39, 0x0010},
+{0x000F3A, 0x0020},
+{0x000F3E, 0x0010},
+{0x000F40, 0x0004},
+{0x000F48, 0x0080},
+{0x000F49, 0x0004},
+{0x000F6D, 0x0080},
+{0x000F71, 0x0010},
+{0x000F85, 0x0020},
+{0x000F86, 0x0010},
+{0x000F88, 0x0004},
+{0x000F8D, 0x0010},
+{0x000F98, 0x0080},
+{0x000F99, 0x0010},
+{0x000FBD, 0x0080},
+{0x000FBE, 0x0040},
+{0x000FC6, 0x0010},
+{0x000FC7, 0x0040},
+{0x000FCD, 0x0080},
+{0x000FCE, 0x0040},
+{0x000FD0, 0x0020},
+{0x000FD5, 0x0040},
+{0x000FD9, 0x0020},
+{0x000FDB, 0x0080},
+{0x001000, 0x0004},
+{0x00102B, 0x0010},
+{0x00103F, 0x0004},
+{0x001040, 0x0002},
+{0x00104A, 0x0020},
+{0x001050, 0x0004},
+{0x001056, 0x0010},
+{0x00105A, 0x0004},
+{0x00105E, 0x0010},
+{0x001061, 0x0004},
+{0x001062, 0x0010},
+{0x001065, 0x0004},
+{0x001067, 0x0010},
+{0x00106E, 0x0004},
+{0x001071, 0x0010},
+{0x001075, 0x0004},
+{0x001082, 0x0010},
+{0x00108E, 0x0004},
+{0x00108F, 0x0010},
+{0x001090, 0x0002},
+{0x00109A, 0x0010},
+{0x00109E, 0x0040},
+{0x0010A0, 0x0004},
+{0x0010C6, 0x0080},
+{0x0010C7, 0x0004},
+{0x0010C8, 0x0080},
+{0x0010CD, 0x0004},
+{0x0010CE, 0x0080},
+{0x0010D0, 0x0004},
+{0x0010FB, 0x0020},
+{0x0010FC, 0x0004},
+{0x001249, 0x0080},
+{0x00124A, 0x0004},
+{0x00124E, 0x0080},
+{0x001250, 0x0004},
+{0x001257, 0x0080},
+{0x001258, 0x0004},
+{0x001259, 0x0080},
+{0x00125A, 0x0004},
+{0x00125E, 0x0080},
+{0x001260, 0x0004},
+{0x001289, 0x0080},
+{0x00128A, 0x0004},
+{0x00128E, 0x0080},
+{0x001290, 0x0004},
+{0x0012B1, 0x0080},
+{0x0012B2, 0x0004},
+{0x0012B6, 0x0080},
+{0x0012B8, 0x0004},
+{0x0012BF, 0x0080},
+{0x0012C0, 0x0004},
+{0x0012C1, 0x0080},
+{0x0012C2, 0x0004},
+{0x0012C6, 0x0080},
+{0x0012C8, 0x0004},
+{0x0012D7, 0x0080},
+{0x0012D8, 0x0004},
+{0x001311, 0x0080},
+{0x001312, 0x0004},
+{0x001316, 0x0080},
+{0x001318, 0x0004},
+{0x00135B, 0x0080},
+{0x00135D, 0x0010},
+{0x001360, 0x0020},
+{0x001369, 0x0002},
+{0x00137D, 0x0080},
+{0x001380, 0x0004},
+{0x001390, 0x0040},
+{0x00139A, 0x0080},
+{0x0013A0, 0x0004},
+{0x0013F6, 0x0080},
+{0x0013F8, 0x0004},
+{0x0013FE, 0x0080},
+{0x001400, 0x0020},
+{0x001401, 0x0004},
+{0x00166D, 0x0040},
+{0x00166E, 0x0020},
+{0x00166F, 0x0004},
+{0x001680, 0x0008},
+{0x001681, 0x0004},
+{0x00169B, 0x0020},
+{0x00169D, 0x0080},
+{0x0016A0, 0x0004},
+{0x0016EB, 0x0020},
+{0x0016EE, 0x0002},
+{0x0016F1, 0x0004},
+{0x0016F9, 0x0080},
+{0x001700, 0x0004},
+{0x001712, 0x0010},
+{0x001716, 0x0080},
+{0x00171F, 0x0004},
+{0x001732, 0x0010},
+{0x001735, 0x0020},
+{0x001737, 0x0080},
+{0x001740, 0x0004},
+{0x001752, 0x0010},
+{0x001754, 0x0080},
+{0x001760, 0x0004},
+{0x00176D, 0x0080},
+{0x00176E, 0x0004},
+{0x001771, 0x0080},
+{0x001772, 0x0010},
+{0x001774, 0x0080},
+{0x001780, 0x0004},
+{0x0017B4, 0x0010},
+{0x0017D4, 0x0020},
+{0x0017D7, 0x0004},
+{0x0017D8, 0x0020},
+{0x0017DB, 0x0040},
+{0x0017DC, 0x0004},
+{0x0017DD, 0x0010},
+{0x0017DE, 0x0080},
+{0x0017E0, 0x0002},
+{0x0017EA, 0x0080},
+{0x0017F0, 0x0002},
+{0x0017FA, 0x0080},
+{0x001800, 0x0020},
+{0x00180B, 0x0010},
+{0x00180E, 0x0080},
+{0x00180F, 0x0010},
+{0x001810, 0x0002},
+{0x00181A, 0x0080},
+{0x001820, 0x0004},
+{0x001879, 0x0080},
+{0x001880, 0x0004},
+{0x001885, 0x0010},
+{0x001887, 0x0004},
+{0x0018A9, 0x0010},
+{0x0018AA, 0x0004},
+{0x0018AB, 0x0080},
+{0x0018B0, 0x0004},
+{0x0018F6, 0x0080},
+{0x001900, 0x0004},
+{0x00191F, 0x0080},
+{0x001920, 0x0010},
+{0x00192C, 0x0080},
+{0x001930, 0x0010},
+{0x00193C, 0x0080},
+{0x001940, 0x0040},
+{0x001941, 0x0080},
+{0x001944, 0x0020},
+{0x001946, 0x0002},
+{0x001950, 0x0004},
+{0x00196E, 0x0080},
+{0x001970, 0x0004},
+{0x001975, 0x0080},
+{0x001980, 0x0004},
+{0x0019AC, 0x0080},
+{0x0019B0, 0x0004},
+{0x0019CA, 0x0080},
+{0x0019D0, 0x0002},
+{0x0019DB, 0x0080},
+{0x0019DE, 0x0040},
+{0x001A00, 0x0004},
+{0x001A17, 0x0010},
+{0x001A1C, 0x0080},
+{0x001A1E, 0x0020},
+{0x001A20, 0x0004},
+{0x001A55, 0x0010},
+{0x001A5F, 0x0080},
+{0x001A60, 0x0010},
+{0x001A7D, 0x0080},
+{0x001A7F, 0x0010},
+{0x001A80, 0x0002},
+{0x001A8A, 0x0080},
+{0x001A90, 0x0002},
+{0x001A9A, 0x0080},
+{0x001AA0, 0x0020},
+{0x001AA7, 0x0004},
+{0x001AA8, 0x0020},
+{0x001AAE, 0x0080},
+{0x001AB0, 0x0010},
+{0x001ACF, 0x0080},
+{0x001B00, 0x0010},
+{0x001B05, 0x0004},
+{0x001B34, 0x0010},
+{0x001B45, 0x0004},
+{0x001B4D, 0x0080},
+{0x001B50, 0x0002},
+{0x001B5A, 0x0020},
+{0x001B61, 0x0040},
+{0x001B6B, 0x0010},
+{0x001B74, 0x0040},
+{0x001B7D, 0x0020},
+{0x001B7F, 0x0080},
+{0x001B80, 0x0010},
+{0x001B83, 0x0004},
+{0x001BA1, 0x0010},
+{0x001BAE, 0x0004},
+{0x001BB0, 0x0002},
+{0x001BBA, 0x0004},
+{0x001BE6, 0x0010},
+{0x001BF4, 0x0080},
+{0x001BFC, 0x0020},
+{0x001C00, 0x0004},
+{0x001C24, 0x0010},
+{0x001C38, 0x0080},
+{0x001C3B, 0x0020},
+{0x001C40, 0x0002},
+{0x001C4A, 0x0080},
+{0x001C4D, 0x0004},
+{0x001C50, 0x0002},
+{0x001C5A, 0x0004},
+{0x001C7E, 0x0020},
+{0x001C80, 0x0004},
+{0x001C89, 0x0080},
+{0x001C90, 0x0004},
+{0x001CBB, 0x0080},
+{0x001CBD, 0x0004},
+{0x001CC0, 0x0020},
+{0x001CC8, 0x0080},
+{0x001CD0, 0x0010},
+{0x001CD3, 0x0020},
+{0x001CD4, 0x0010},
+{0x001CE9, 0x0004},
+{0x001CED, 0x0010},
+{0x001CEE, 0x0004},
+{0x001CF4, 0x0010},
+{0x001CF5, 0x0004},
+{0x001CF7, 0x0010},
+{0x001CFA, 0x0004},
+{0x001CFB, 0x0080},
+{0x001D00, 0x0004},
+{0x001DC0, 0x0010},
+{0x001E00, 0x0004},
+{0x001F16, 0x0080},
+{0x001F18, 0x0004},
+{0x001F1E, 0x0080},
+{0x001F20, 0x0004},
+{0x001F46, 0x0080},
+{0x001F48, 0x0004},
+{0x001F4E, 0x0080},
+{0x001F50, 0x0004},
+{0x001F58, 0x0080},
+{0x001F59, 0x0004},
+{0x001F5A, 0x0080},
+{0x001F5B, 0x0004},
+{0x001F5C, 0x0080},
+{0x001F5D, 0x0004},
+{0x001F5E, 0x0080},
+{0x001F5F, 0x0004},
+{0x001F7E, 0x0080},
+{0x001F80, 0x0004},
+{0x001FB5, 0x0080},
+{0x001FB6, 0x0004},
+{0x001FBD, 0x0040},
+{0x001FBE, 0x0004},
+{0x001FBF, 0x0040},
+{0x001FC2, 0x0004},
+{0x001FC5, 0x0080},
+{0x001FC6, 0x0004},
+{0x001FCD, 0x0040},
+{0x001FD0, 0x0004},
+{0x001FD4, 0x0080},
+{0x001FD6, 0x0004},
+{0x001FDC, 0x0080},
+{0x001FDD, 0x0040},
+{0x001FE0, 0x0004},
+{0x001FED, 0x0040},
+{0x001FF0, 0x0080},
+{0x001FF2, 0x0004},
+{0x001FF5, 0x0080},
+{0x001FF6, 0x0004},
+{0x001FFD, 0x0040},
+{0x001FFF, 0x0080},
+{0x002000, 0x0008},
+{0x00200B, 0x0080},
+{0x002010, 0x0020},
+{0x002028, 0x0008},
+{0x00202A, 0x0080},
+{0x00202F, 0x0008},
+{0x002030, 0x0020},
+{0x002044, 0x0040},
+{0x002045, 0x0020},
+{0x002052, 0x0040},
+{0x002053, 0x0020},
+{0x00205F, 0x0008},
+{0x002060, 0x0080},
+{0x002070, 0x0002},
+{0x002071, 0x0004},
+{0x002072, 0x0080},
+{0x002074, 0x0002},
+{0x00207A, 0x0040},
+{0x00207D, 0x0020},
+{0x00207F, 0x0004},
+{0x002080, 0x0002},
+{0x00208A, 0x0040},
+{0x00208D, 0x0020},
+{0x00208F, 0x0080},
+{0x002090, 0x0004},
+{0x00209D, 0x0080},
+{0x0020A0, 0x0040},
+{0x0020C1, 0x0080},
+{0x0020D0, 0x0010},
+{0x0020F1, 0x0080},
+{0x002100, 0x0040},
+{0x002102, 0x0004},
+{0x002103, 0x0040},
+{0x002107, 0x0004},
+{0x002108, 0x0040},
+{0x00210A, 0x0004},
+{0x002114, 0x0040},
+{0x002115, 0x0004},
+{0x002116, 0x0040},
+{0x002119, 0x0004},
+{0x00211E, 0x0040},
+{0x002124, 0x0004},
+{0x002125, 0x0040},
+{0x002126, 0x0004},
+{0x002127, 0x0040},
+{0x002128, 0x0004},
+{0x002129, 0x0040},
+{0x00212A, 0x0004},
+{0x00212E, 0x0040},
+{0x00212F, 0x0004},
+{0x00213A, 0x0040},
+{0x00213C, 0x0004},
+{0x002140, 0x0040},
+{0x002145, 0x0004},
+{0x00214A, 0x0040},
+{0x00214E, 0x0004},
+{0x00214F, 0x0040},
+{0x002150, 0x0002},
+{0x002183, 0x0004},
+{0x002185, 0x0002},
+{0x00218A, 0x0040},
+{0x00218C, 0x0080},
+{0x002190, 0x0040},
+{0x002308, 0x0020},
+{0x00230C, 0x0040},
+{0x002329, 0x0020},
+{0x00232B, 0x0040},
+{0x002427, 0x0080},
+{0x002440, 0x0040},
+{0x00244B, 0x0080},
+{0x002460, 0x0002},
+{0x00249C, 0x0040},
+{0x0024EA, 0x0002},
+{0x002500, 0x0040},
+{0x002768, 0x0020},
+{0x002776, 0x0002},
+{0x002794, 0x0040},
+{0x0027C5, 0x0020},
+{0x0027C7, 0x0040},
+{0x0027E6, 0x0020},
+{0x0027F0, 0x0040},
+{0x002983, 0x0020},
+{0x002999, 0x0040},
+{0x0029D8, 0x0020},
+{0x0029DC, 0x0040},
+{0x0029FC, 0x0020},
+{0x0029FE, 0x0040},
+{0x002B74, 0x0080},
+{0x002B76, 0x0040},
+{0x002B96, 0x0080},
+{0x002B97, 0x0040},
+{0x002C00, 0x0004},
+{0x002CE5, 0x0040},
+{0x002CEB, 0x0004},
+{0x002CEF, 0x0010},
+{0x002CF2, 0x0004},
+{0x002CF4, 0x0080},
+{0x002CF9, 0x0020},
+{0x002CFD, 0x0002},
+{0x002CFE, 0x0020},
+{0x002D00, 0x0004},
+{0x002D26, 0x0080},
+{0x002D27, 0x0004},
+{0x002D28, 0x0080},
+{0x002D2D, 0x0004},
+{0x002D2E, 0x0080},
+{0x002D30, 0x0004},
+{0x002D68, 0x0080},
+{0x002D6F, 0x0004},
+{0x002D70, 0x0020},
+{0x002D71, 0x0080},
+{0x002D7F, 0x0010},
+{0x002D80, 0x0004},
+{0x002D97, 0x0080},
+{0x002DA0, 0x0004},
+{0x002DA7, 0x0080},
+{0x002DA8, 0x0004},
+{0x002DAF, 0x0080},
+{0x002DB0, 0x0004},
+{0x002DB7, 0x0080},
+{0x002DB8, 0x0004},
+{0x002DBF, 0x0080},
+{0x002DC0, 0x0004},
+{0x002DC7, 0x0080},
+{0x002DC8, 0x0004},
+{0x002DCF, 0x0080},
+{0x002DD0, 0x0004},
+{0x002DD7, 0x0080},
+{0x002DD8, 0x0004},
+{0x002DDF, 0x0080},
+{0x002DE0, 0x0010},
+{0x002E00, 0x0020},
+{0x002E2F, 0x0004},
+{0x002E30, 0x0020},
+{0x002E50, 0x0040},
+{0x002E52, 0x0020},
+{0x002E5E, 0x0080},
+{0x002E80, 0x0040},
+{0x002E9A, 0x0080},
+{0x002E9B, 0x0040},
+{0x002EF4, 0x0080},
+{0x002F00, 0x0040},
+{0x002FD6, 0x0080},
+{0x002FF0, 0x0040},
+{0x003000, 0x0008},
+{0x003001, 0x0020},
+{0x003004, 0x0040},
+{0x003005, 0x0004},
+{0x003007, 0x0002},
+{0x003008, 0x0020},
+{0x003012, 0x0040},
+{0x003014, 0x0020},
+{0x003020, 0x0040},
+{0x003021, 0x0002},
+{0x00302A, 0x0010},
+{0x003030, 0x0020},
+{0x003031, 0x0004},
+{0x003036, 0x0040},
+{0x003038, 0x0002},
+{0x00303B, 0x0004},
+{0x00303D, 0x0020},
+{0x00303E, 0x0040},
+{0x003040, 0x0080},
+{0x003041, 0x0004},
+{0x003097, 0x0080},
+{0x003099, 0x0010},
+{0x00309B, 0x0040},
+{0x00309D, 0x0004},
+{0x0030A0, 0x0020},
+{0x0030A1, 0x0004},
+{0x0030FB, 0x0020},
+{0x0030FC, 0x0004},
+{0x003100, 0x0080},
+{0x003105, 0x0004},
+{0x003130, 0x0080},
+{0x003131, 0x0004},
+{0x00318F, 0x0080},
+{0x003190, 0x0040},
+{0x003192, 0x0002},
+{0x003196, 0x0040},
+{0x0031A0, 0x0004},
+{0x0031C0, 0x0040},
+{0x0031E4, 0x0080},
+{0x0031EF, 0x0040},
+{0x0031F0, 0x0004},
+{0x003200, 0x0040},
+{0x00321F, 0x0080},
+{0x003220, 0x0002},
+{0x00322A, 0x0040},
+{0x003248, 0x0002},
+{0x003250, 0x0040},
+{0x003251, 0x0002},
+{0x003260, 0x0040},
+{0x003280, 0x0002},
+{0x00328A, 0x0040},
+{0x0032B1, 0x0002},
+{0x0032C0, 0x0040},
+{0x003400, 0x0004},
+{0x004DC0, 0x0040},
+{0x004E00, 0x0004},
+{0x00A48D, 0x0080},
+{0x00A490, 0x0040},
+{0x00A4C7, 0x0080},
+{0x00A4D0, 0x0004},
+{0x00A4FE, 0x0020},
+{0x00A500, 0x0004},
+{0x00A60D, 0x0020},
+{0x00A610, 0x0004},
+{0x00A620, 0x0002},
+{0x00A62A, 0x0004},
+{0x00A62C, 0x0080},
+{0x00A640, 0x0004},
+{0x00A66F, 0x0010},
+{0x00A673, 0x0020},
+{0x00A674, 0x0010},
+{0x00A67E, 0x0020},
+{0x00A67F, 0x0004},
+{0x00A69E, 0x0010},
+{0x00A6A0, 0x0004},
+{0x00A6E6, 0x0002},
+{0x00A6F0, 0x0010},
+{0x00A6F2, 0x0020},
+{0x00A6F8, 0x0080},
+{0x00A700, 0x0040},
+{0x00A717, 0x0004},
+{0x00A720, 0x0040},
+{0x00A722, 0x0004},
+{0x00A789, 0x0040},
+{0x00A78B, 0x0004},
+{0x00A7CB, 0x0080},
+{0x00A7D0, 0x0004},
+{0x00A7D2, 0x0080},
+{0x00A7D3, 0x0004},
+{0x00A7D4, 0x0080},
+{0x00A7D5, 0x0004},
+{0x00A7DA, 0x0080},
+{0x00A7F2, 0x0004},
+{0x00A802, 0x0010},
+{0x00A803, 0x0004},
+{0x00A806, 0x0010},
+{0x00A807, 0x0004},
+{0x00A80B, 0x0010},
+{0x00A80C, 0x0004},
+{0x00A823, 0x0010},
+{0x00A828, 0x0040},
+{0x00A82C, 0x0010},
+{0x00A82D, 0x0080},
+{0x00A830, 0x0002},
+{0x00A836, 0x0040},
+{0x00A83A, 0x0080},
+{0x00A840, 0x0004},
+{0x00A874, 0x0020},
+{0x00A878, 0x0080},
+{0x00A880, 0x0010},
+{0x00A882, 0x0004},
+{0x00A8B4, 0x0010},
+{0x00A8C6, 0x0080},
+{0x00A8CE, 0x0020},
+{0x00A8D0, 0x0002},
+{0x00A8DA, 0x0080},
+{0x00A8E0, 0x0010},
+{0x00A8F2, 0x0004},
+{0x00A8F8, 0x0020},
+{0x00A8FB, 0x0004},
+{0x00A8FC, 0x0020},
+{0x00A8FD, 0x0004},
+{0x00A8FF, 0x0010},
+{0x00A900, 0x0002},
+{0x00A90A, 0x0004},
+{0x00A926, 0x0010},
+{0x00A92E, 0x0020},
+{0x00A930, 0x0004},
+{0x00A947, 0x0010},
+{0x00A954, 0x0080},
+{0x00A95F, 0x0020},
+{0x00A960, 0x0004},
+{0x00A97D, 0x0080},
+{0x00A980, 0x0010},
+{0x00A984, 0x0004},
+{0x00A9B3, 0x0010},
+{0x00A9C1, 0x0020},
+{0x00A9CE, 0x0080},
+{0x00A9CF, 0x0004},
+{0x00A9D0, 0x0002},
+{0x00A9DA, 0x0080},
+{0x00A9DE, 0x0020},
+{0x00A9E0, 0x0004},
+{0x00A9E5, 0x0010},
+{0x00A9E6, 0x0004},
+{0x00A9F0, 0x0002},
+{0x00A9FA, 0x0004},
+{0x00A9FF, 0x0080},
+{0x00AA00, 0x0004},
+{0x00AA29, 0x0010},
+{0x00AA37, 0x0080},
+{0x00AA40, 0x0004},
+{0x00AA43, 0x0010},
+{0x00AA44, 0x0004},
+{0x00AA4C, 0x0010},
+{0x00AA4E, 0x0080},
+{0x00AA50, 0x0002},
+{0x00AA5A, 0x0080},
+{0x00AA5C, 0x0020},
+{0x00AA60, 0x0004},
+{0x00AA77, 0x0040},
+{0x00AA7A, 0x0004},
+{0x00AA7B, 0x0010},
+{0x00AA7E, 0x0004},
+{0x00AAB0, 0x0010},
+{0x00AAB1, 0x0004},
+{0x00AAB2, 0x0010},
+{0x00AAB5, 0x0004},
+{0x00AAB7, 0x0010},
+{0x00AAB9, 0x0004},
+{0x00AABE, 0x0010},
+{0x00AAC0, 0x0004},
+{0x00AAC1, 0x0010},
+{0x00AAC2, 0x0004},
+{0x00AAC3, 0x0080},
+{0x00AADB, 0x0004},
+{0x00AADE, 0x0020},
+{0x00AAE0, 0x0004},
+{0x00AAEB, 0x0010},
+{0x00AAF0, 0x0020},
+{0x00AAF2, 0x0004},
+{0x00AAF5, 0x0010},
+{0x00AAF7, 0x0080},
+{0x00AB01, 0x0004},
+{0x00AB07, 0x0080},
+{0x00AB09, 0x0004},
+{0x00AB0F, 0x0080},
+{0x00AB11, 0x0004},
+{0x00AB17, 0x0080},
+{0x00AB20, 0x0004},
+{0x00AB27, 0x0080},
+{0x00AB28, 0x0004},
+{0x00AB2F, 0x0080},
+{0x00AB30, 0x0004},
+{0x00AB5B, 0x0040},
+{0x00AB5C, 0x0004},
+{0x00AB6A, 0x0040},
+{0x00AB6C, 0x0080},
+{0x00AB70, 0x0004},
+{0x00ABE3, 0x0010},
+{0x00ABEB, 0x0020},
+{0x00ABEC, 0x0010},
+{0x00ABEE, 0x0080},
+{0x00ABF0, 0x0002},
+{0x00ABFA, 0x0080},
+{0x00AC00, 0x0004},
+{0x00D7A4, 0x0080},
+{0x00D7B0, 0x0004},
+{0x00D7C7, 0x0080},
+{0x00D7CB, 0x0004},
+{0x00D7FC, 0x0080},
+{0x00F900, 0x0004},
+{0x00FA6E, 0x0080},
+{0x00FA70, 0x0004},
+{0x00FADA, 0x0080},
+{0x00FB00, 0x0004},
+{0x00FB07, 0x0080},
+{0x00FB13, 0x0004},
+{0x00FB18, 0x0080},
+{0x00FB1D, 0x0004},
+{0x00FB1E, 0x0010},
+{0x00FB1F, 0x0004},
+{0x00FB29, 0x0040},
+{0x00FB2A, 0x0004},
+{0x00FB37, 0x0080},
+{0x00FB38, 0x0004},
+{0x00FB3D, 0x0080},
+{0x00FB3E, 0x0004},
+{0x00FB3F, 0x0080},
+{0x00FB40, 0x0004},
+{0x00FB42, 0x0080},
+{0x00FB43, 0x0004},
+{0x00FB45, 0x0080},
+{0x00FB46, 0x0004},
+{0x00FBB2, 0x0040},
+{0x00FBC3, 0x0080},
+{0x00FBD3, 0x0004},
+{0x00FD3E, 0x0020},
+{0x00FD40, 0x0040},
+{0x00FD50, 0x0004},
+{0x00FD90, 0x0080},
+{0x00FD92, 0x0004},
+{0x00FDC8, 0x0080},
+{0x00FDCF, 0x0040},
+{0x00FDD0, 0x0080},
+{0x00FDF0, 0x0004},
+{0x00FDFC, 0x0040},
+{0x00FE00, 0x0010},
+{0x00FE10, 0x0020},
+{0x00FE1A, 0x0080},
+{0x00FE20, 0x0010},
+{0x00FE30, 0x0020},
+{0x00FE53, 0x0080},
+{0x00FE54, 0x0020},
+{0x00FE62, 0x0040},
+{0x00FE63, 0x0020},
+{0x00FE64, 0x0040},
+{0x00FE67, 0x0080},
+{0x00FE68, 0x0020},
+{0x00FE69, 0x0040},
+{0x00FE6A, 0x0020},
+{0x00FE6C, 0x0080},
+{0x00FE70, 0x0004},
+{0x00FE75, 0x0080},
+{0x00FE76, 0x0004},
+{0x00FEFD, 0x0080},
+{0x00FF01, 0x0020},
+{0x00FF04, 0x0040},
+{0x00FF05, 0x0020},
+{0x00FF0B, 0x0040},
+{0x00FF0C, 0x0020},
+{0x00FF10, 0x0002},
+{0x00FF1A, 0x0020},
+{0x00FF1C, 0x0040},
+{0x00FF1F, 0x0020},
+{0x00FF21, 0x0004},
+{0x00FF3B, 0x0020},
+{0x00FF3E, 0x0040},
+{0x00FF3F, 0x0020},
+{0x00FF40, 0x0040},
+{0x00FF41, 0x0004},
+{0x00FF5B, 0x0020},
+{0x00FF5C, 0x0040},
+{0x00FF5D, 0x0020},
+{0x00FF5E, 0x0040},
+{0x00FF5F, 0x0020},
+{0x00FF66, 0x0004},
+{0x00FFBF, 0x0080},
+{0x00FFC2, 0x0004},
+{0x00FFC8, 0x0080},
+{0x00FFCA, 0x0004},
+{0x00FFD0, 0x0080},
+{0x00FFD2, 0x0004},
+{0x00FFD8, 0x0080},
+{0x00FFDA, 0x0004},
+{0x00FFDD, 0x0080},
+{0x00FFE0, 0x0040},
+{0x00FFE7, 0x0080},
+{0x00FFE8, 0x0040},
+{0x00FFEF, 0x0080},
+{0x00FFFC, 0x0040},
+{0x00FFFE, 0x0080},
+{0x010000, 0x0004},
+{0x01000C, 0x0080},
+{0x01000D, 0x0004},
+{0x010027, 0x0080},
+{0x010028, 0x0004},
+{0x01003B, 0x0080},
+{0x01003C, 0x0004},
+{0x01003E, 0x0080},
+{0x01003F, 0x0004},
+{0x01004E, 0x0080},
+{0x010050, 0x0004},
+{0x01005E, 0x0080},
+{0x010080, 0x0004},
+{0x0100FB, 0x0080},
+{0x010100, 0x0020},
+{0x010103, 0x0080},
+{0x010107, 0x0002},
+{0x010134, 0x0080},
+{0x010137, 0x0040},
+{0x010140, 0x0002},
+{0x010179, 0x0040},
+{0x01018A, 0x0002},
+{0x01018C, 0x0040},
+{0x01018F, 0x0080},
+{0x010190, 0x0040},
+{0x01019D, 0x0080},
+{0x0101A0, 0x0040},
+{0x0101A1, 0x0080},
+{0x0101D0, 0x0040},
+{0x0101FD, 0x0010},
+{0x0101FE, 0x0080},
+{0x010280, 0x0004},
+{0x01029D, 0x0080},
+{0x0102A0, 0x0004},
+{0x0102D1, 0x0080},
+{0x0102E0, 0x0010},
+{0x0102E1, 0x0002},
+{0x0102FC, 0x0080},
+{0x010300, 0x0004},
+{0x010320, 0x0002},
+{0x010324, 0x0080},
+{0x01032D, 0x0004},
+{0x010341, 0x0002},
+{0x010342, 0x0004},
+{0x01034A, 0x0002},
+{0x01034B, 0x0080},
+{0x010350, 0x0004},
+{0x010376, 0x0010},
+{0x01037B, 0x0080},
+{0x010380, 0x0004},
+{0x01039E, 0x0080},
+{0x01039F, 0x0020},
+{0x0103A0, 0x0004},
+{0x0103C4, 0x0080},
+{0x0103C8, 0x0004},
+{0x0103D0, 0x0020},
+{0x0103D1, 0x0002},
+{0x0103D6, 0x0080},
+{0x010400, 0x0004},
+{0x01049E, 0x0080},
+{0x0104A0, 0x0002},
+{0x0104AA, 0x0080},
+{0x0104B0, 0x0004},
+{0x0104D4, 0x0080},
+{0x0104D8, 0x0004},
+{0x0104FC, 0x0080},
+{0x010500, 0x0004},
+{0x010528, 0x0080},
+{0x010530, 0x0004},
+{0x010564, 0x0080},
+{0x01056F, 0x0020},
+{0x010570, 0x0004},
+{0x01057B, 0x0080},
+{0x01057C, 0x0004},
+{0x01058B, 0x0080},
+{0x01058C, 0x0004},
+{0x010593, 0x0080},
+{0x010594, 0x0004},
+{0x010596, 0x0080},
+{0x010597, 0x0004},
+{0x0105A2, 0x0080},
+{0x0105A3, 0x0004},
+{0x0105B2, 0x0080},
+{0x0105B3, 0x0004},
+{0x0105BA, 0x0080},
+{0x0105BB, 0x0004},
+{0x0105BD, 0x0080},
+{0x010600, 0x0004},
+{0x010737, 0x0080},
+{0x010740, 0x0004},
+{0x010756, 0x0080},
+{0x010760, 0x0004},
+{0x010768, 0x0080},
+{0x010780, 0x0004},
+{0x010786, 0x0080},
+{0x010787, 0x0004},
+{0x0107B1, 0x0080},
+{0x0107B2, 0x0004},
+{0x0107BB, 0x0080},
+{0x010800, 0x0004},
+{0x010806, 0x0080},
+{0x010808, 0x0004},
+{0x010809, 0x0080},
+{0x01080A, 0x0004},
+{0x010836, 0x0080},
+{0x010837, 0x0004},
+{0x010839, 0x0080},
+{0x01083C, 0x0004},
+{0x01083D, 0x0080},
+{0x01083F, 0x0004},
+{0x010856, 0x0080},
+{0x010857, 0x0020},
+{0x010858, 0x0002},
+{0x010860, 0x0004},
+{0x010877, 0x0040},
+{0x010879, 0x0002},
+{0x010880, 0x0004},
+{0x01089F, 0x0080},
+{0x0108A7, 0x0002},
+{0x0108B0, 0x0080},
+{0x0108E0, 0x0004},
+{0x0108F3, 0x0080},
+{0x0108F4, 0x0004},
+{0x0108F6, 0x0080},
+{0x0108FB, 0x0002},
+{0x010900, 0x0004},
+{0x010916, 0x0002},
+{0x01091C, 0x0080},
+{0x01091F, 0x0020},
+{0x010920, 0x0004},
+{0x01093A, 0x0080},
+{0x01093F, 0x0020},
+{0x010940, 0x0080},
+{0x010980, 0x0004},
+{0x0109B8, 0x0080},
+{0x0109BC, 0x0002},
+{0x0109BE, 0x0004},
+{0x0109C0, 0x0002},
+{0x0109D0, 0x0080},
+{0x0109D2, 0x0002},
+{0x010A00, 0x0004},
+{0x010A01, 0x0010},
+{0x010A04, 0x0080},
+{0x010A05, 0x0010},
+{0x010A07, 0x0080},
+{0x010A0C, 0x0010},
+{0x010A10, 0x0004},
+{0x010A14, 0x0080},
+{0x010A15, 0x0004},
+{0x010A18, 0x0080},
+{0x010A19, 0x0004},
+{0x010A36, 0x0080},
+{0x010A38, 0x0010},
+{0x010A3B, 0x0080},
+{0x010A3F, 0x0010},
+{0x010A40, 0x0002},
+{0x010A49, 0x0080},
+{0x010A50, 0x0020},
+{0x010A59, 0x0080},
+{0x010A60, 0x0004},
+{0x010A7D, 0x0002},
+{0x010A7F, 0x0020},
+{0x010A80, 0x0004},
+{0x010A9D, 0x0002},
+{0x010AA0, 0x0080},
+{0x010AC0, 0x0004},
+{0x010AC8, 0x0040},
+{0x010AC9, 0x0004},
+{0x010AE5, 0x0010},
+{0x010AE7, 0x0080},
+{0x010AEB, 0x0002},
+{0x010AF0, 0x0020},
+{0x010AF7, 0x0080},
+{0x010B00, 0x0004},
+{0x010B36, 0x0080},
+{0x010B39, 0x0020},
+{0x010B40, 0x0004},
+{0x010B56, 0x0080},
+{0x010B58, 0x0002},
+{0x010B60, 0x0004},
+{0x010B73, 0x0080},
+{0x010B78, 0x0002},
+{0x010B80, 0x0004},
+{0x010B92, 0x0080},
+{0x010B99, 0x0020},
+{0x010B9D, 0x0080},
+{0x010BA9, 0x0002},
+{0x010BB0, 0x0080},
+{0x010C00, 0x0004},
+{0x010C49, 0x0080},
+{0x010C80, 0x0004},
+{0x010CB3, 0x0080},
+{0x010CC0, 0x0004},
+{0x010CF3, 0x0080},
+{0x010CFA, 0x0002},
+{0x010D00, 0x0004},
+{0x010D24, 0x0010},
+{0x010D28, 0x0080},
+{0x010D30, 0x0002},
+{0x010D3A, 0x0080},
+{0x010E60, 0x0002},
+{0x010E7F, 0x0080},
+{0x010E80, 0x0004},
+{0x010EAA, 0x0080},
+{0x010EAB, 0x0010},
+{0x010EAD, 0x0020},
+{0x010EAE, 0x0080},
+{0x010EB0, 0x0004},
+{0x010EB2, 0x0080},
+{0x010EFD, 0x0010},
+{0x010F00, 0x0004},
+{0x010F1D, 0x0002},
+{0x010F27, 0x0004},
+{0x010F28, 0x0080},
+{0x010F30, 0x0004},
+{0x010F46, 0x0010},
+{0x010F51, 0x0002},
+{0x010F55, 0x0020},
+{0x010F5A, 0x0080},
+{0x010F70, 0x0004},
+{0x010F82, 0x0010},
+{0x010F86, 0x0020},
+{0x010F8A, 0x0080},
+{0x010FB0, 0x0004},
+{0x010FC5, 0x0002},
+{0x010FCC, 0x0080},
+{0x010FE0, 0x0004},
+{0x010FF7, 0x0080},
+{0x011000, 0x0010},
+{0x011003, 0x0004},
+{0x011038, 0x0010},
+{0x011047, 0x0020},
+{0x01104E, 0x0080},
+{0x011052, 0x0002},
+{0x011070, 0x0010},
+{0x011071, 0x0004},
+{0x011073, 0x0010},
+{0x011075, 0x0004},
+{0x011076, 0x0080},
+{0x01107F, 0x0010},
+{0x011083, 0x0004},
+{0x0110B0, 0x0010},
+{0x0110BB, 0x0020},
+{0x0110BD, 0x0080},
+{0x0110BE, 0x0020},
+{0x0110C2, 0x0010},
+{0x0110C3, 0x0080},
+{0x0110D0, 0x0004},
+{0x0110E9, 0x0080},
+{0x0110F0, 0x0002},
+{0x0110FA, 0x0080},
+{0x011100, 0x0010},
+{0x011103, 0x0004},
+{0x011127, 0x0010},
+{0x011135, 0x0080},
+{0x011136, 0x0002},
+{0x011140, 0x0020},
+{0x011144, 0x0004},
+{0x011145, 0x0010},
+{0x011147, 0x0004},
+{0x011148, 0x0080},
+{0x011150, 0x0004},
+{0x011173, 0x0010},
+{0x011174, 0x0020},
+{0x011176, 0x0004},
+{0x011177, 0x0080},
+{0x011180, 0x0010},
+{0x011183, 0x0004},
+{0x0111B3, 0x0010},
+{0x0111C1, 0x0004},
+{0x0111C5, 0x0020},
+{0x0111C9, 0x0010},
+{0x0111CD, 0x0020},
+{0x0111CE, 0x0010},
+{0x0111D0, 0x0002},
+{0x0111DA, 0x0004},
+{0x0111DB, 0x0020},
+{0x0111DC, 0x0004},
+{0x0111DD, 0x0020},
+{0x0111E0, 0x0080},
+{0x0111E1, 0x0002},
+{0x0111F5, 0x0080},
+{0x011200, 0x0004},
+{0x011212, 0x0080},
+{0x011213, 0x0004},
+{0x01122C, 0x0010},
+{0x011238, 0x0020},
+{0x01123E, 0x0010},
+{0x01123F, 0x0004},
+{0x011241, 0x0010},
+{0x011242, 0x0080},
+{0x011280, 0x0004},
+{0x011287, 0x0080},
+{0x011288, 0x0004},
+{0x011289, 0x0080},
+{0x01128A, 0x0004},
+{0x01128E, 0x0080},
+{0x01128F, 0x0004},
+{0x01129E, 0x0080},
+{0x01129F, 0x0004},
+{0x0112A9, 0x0020},
+{0x0112AA, 0x0080},
+{0x0112B0, 0x0004},
+{0x0112DF, 0x0010},
+{0x0112EB, 0x0080},
+{0x0112F0, 0x0002},
+{0x0112FA, 0x0080},
+{0x011300, 0x0010},
+{0x011304, 0x0080},
+{0x011305, 0x0004},
+{0x01130D, 0x0080},
+{0x01130F, 0x0004},
+{0x011311, 0x0080},
+{0x011313, 0x0004},
+{0x011329, 0x0080},
+{0x01132A, 0x0004},
+{0x011331, 0x0080},
+{0x011332, 0x0004},
+{0x011334, 0x0080},
+{0x011335, 0x0004},
+{0x01133A, 0x0080},
+{0x01133B, 0x0010},
+{0x01133D, 0x0004},
+{0x01133E, 0x0010},
+{0x011345, 0x0080},
+{0x011347, 0x0010},
+{0x011349, 0x0080},
+{0x01134B, 0x0010},
+{0x01134E, 0x0080},
+{0x011350, 0x0004},
+{0x011351, 0x0080},
+{0x011357, 0x0010},
+{0x011358, 0x0080},
+{0x01135D, 0x0004},
+{0x011362, 0x0010},
+{0x011364, 0x0080},
+{0x011366, 0x0010},
+{0x01136D, 0x0080},
+{0x011370, 0x0010},
+{0x011375, 0x0080},
+{0x011400, 0x0004},
+{0x011435, 0x0010},
+{0x011447, 0x0004},
+{0x01144B, 0x0020},
+{0x011450, 0x0002},
+{0x01145A, 0x0020},
+{0x01145C, 0x0080},
+{0x01145D, 0x0020},
+{0x01145E, 0x0010},
+{0x01145F, 0x0004},
+{0x011462, 0x0080},
+{0x011480, 0x0004},
+{0x0114B0, 0x0010},
+{0x0114C4, 0x0004},
+{0x0114C6, 0x0020},
+{0x0114C7, 0x0004},
+{0x0114C8, 0x0080},
+{0x0114D0, 0x0002},
+{0x0114DA, 0x0080},
+{0x011580, 0x0004},
+{0x0115AF, 0x0010},
+{0x0115B6, 0x0080},
+{0x0115B8, 0x0010},
+{0x0115C1, 0x0020},
+{0x0115D8, 0x0004},
+{0x0115DC, 0x0010},
+{0x0115DE, 0x0080},
+{0x011600, 0x0004},
+{0x011630, 0x0010},
+{0x011641, 0x0020},
+{0x011644, 0x0004},
+{0x011645, 0x0080},
+{0x011650, 0x0002},
+{0x01165A, 0x0080},
+{0x011660, 0x0020},
+{0x01166D, 0x0080},
+{0x011680, 0x0004},
+{0x0116AB, 0x0010},
+{0x0116B8, 0x0004},
+{0x0116B9, 0x0020},
+{0x0116BA, 0x0080},
+{0x0116C0, 0x0002},
+{0x0116CA, 0x0080},
+{0x011700, 0x0004},
+{0x01171B, 0x0080},
+{0x01171D, 0x0010},
+{0x01172C, 0x0080},
+{0x011730, 0x0002},
+{0x01173C, 0x0020},
+{0x01173F, 0x0040},
+{0x011740, 0x0004},
+{0x011747, 0x0080},
+{0x011800, 0x0004},
+{0x01182C, 0x0010},
+{0x01183B, 0x0020},
+{0x01183C, 0x0080},
+{0x0118A0, 0x0004},
+{0x0118E0, 0x0002},
+{0x0118F3, 0x0080},
+{0x0118FF, 0x0004},
+{0x011907, 0x0080},
+{0x011909, 0x0004},
+{0x01190A, 0x0080},
+{0x01190C, 0x0004},
+{0x011914, 0x0080},
+{0x011915, 0x0004},
+{0x011917, 0x0080},
+{0x011918, 0x0004},
+{0x011930, 0x0010},
+{0x011936, 0x0080},
+{0x011937, 0x0010},
+{0x011939, 0x0080},
+{0x01193B, 0x0010},
+{0x01193F, 0x0004},
+{0x011940, 0x0010},
+{0x011941, 0x0004},
+{0x011942, 0x0010},
+{0x011944, 0x0020},
+{0x011947, 0x0080},
+{0x011950, 0x0002},
+{0x01195A, 0x0080},
+{0x0119A0, 0x0004},
+{0x0119A8, 0x0080},
+{0x0119AA, 0x0004},
+{0x0119D1, 0x0010},
+{0x0119D8, 0x0080},
+{0x0119DA, 0x0010},
+{0x0119E1, 0x0004},
+{0x0119E2, 0x0020},
+{0x0119E3, 0x0004},
+{0x0119E4, 0x0010},
+{0x0119E5, 0x0080},
+{0x011A00, 0x0004},
+{0x011A01, 0x0010},
+{0x011A0B, 0x0004},
+{0x011A33, 0x0010},
+{0x011A3A, 0x0004},
+{0x011A3B, 0x0010},
+{0x011A3F, 0x0020},
+{0x011A47, 0x0010},
+{0x011A48, 0x0080},
+{0x011A50, 0x0004},
+{0x011A51, 0x0010},
+{0x011A5C, 0x0004},
+{0x011A8A, 0x0010},
+{0x011A9A, 0x0020},
+{0x011A9D, 0x0004},
+{0x011A9E, 0x0020},
+{0x011AA3, 0x0080},
+{0x011AB0, 0x0004},
+{0x011AF9, 0x0080},
+{0x011B00, 0x0020},
+{0x011B0A, 0x0080},
+{0x011C00, 0x0004},
+{0x011C09, 0x0080},
+{0x011C0A, 0x0004},
+{0x011C2F, 0x0010},
+{0x011C37, 0x0080},
+{0x011C38, 0x0010},
+{0x011C40, 0x0004},
+{0x011C41, 0x0020},
+{0x011C46, 0x0080},
+{0x011C50, 0x0002},
+{0x011C6D, 0x0080},
+{0x011C70, 0x0020},
+{0x011C72, 0x0004},
+{0x011C90, 0x0080},
+{0x011C92, 0x0010},
+{0x011CA8, 0x0080},
+{0x011CA9, 0x0010},
+{0x011CB7, 0x0080},
+{0x011D00, 0x0004},
+{0x011D07, 0x0080},
+{0x011D08, 0x0004},
+{0x011D0A, 0x0080},
+{0x011D0B, 0x0004},
+{0x011D31, 0x0010},
+{0x011D37, 0x0080},
+{0x011D3A, 0x0010},
+{0x011D3B, 0x0080},
+{0x011D3C, 0x0010},
+{0x011D3E, 0x0080},
+{0x011D3F, 0x0010},
+{0x011D46, 0x0004},
+{0x011D47, 0x0010},
+{0x011D48, 0x0080},
+{0x011D50, 0x0002},
+{0x011D5A, 0x0080},
+{0x011D60, 0x0004},
+{0x011D66, 0x0080},
+{0x011D67, 0x0004},
+{0x011D69, 0x0080},
+{0x011D6A, 0x0004},
+{0x011D8A, 0x0010},
+{0x011D8F, 0x0080},
+{0x011D90, 0x0010},
+{0x011D92, 0x0080},
+{0x011D93, 0x0010},
+{0x011D98, 0x0004},
+{0x011D99, 0x0080},
+{0x011DA0, 0x0002},
+{0x011DAA, 0x0080},
+{0x011EE0, 0x0004},
+{0x011EF3, 0x0010},
+{0x011EF7, 0x0020},
+{0x011EF9, 0x0080},
+{0x011F00, 0x0010},
+{0x011F02, 0x0004},
+{0x011F03, 0x0010},
+{0x011F04, 0x0004},
+{0x011F11, 0x0080},
+{0x011F12, 0x0004},
+{0x011F34, 0x0010},
+{0x011F3B, 0x0080},
+{0x011F3E, 0x0010},
+{0x011F43, 0x0020},
+{0x011F50, 0x0002},
+{0x011F5A, 0x0080},
+{0x011FB0, 0x0004},
+{0x011FB1, 0x0080},
+{0x011FC0, 0x0002},
+{0x011FD5, 0x0040},
+{0x011FF2, 0x0080},
+{0x011FFF, 0x0020},
+{0x012000, 0x0004},
+{0x01239A, 0x0080},
+{0x012400, 0x0002},
+{0x01246F, 0x0080},
+{0x012470, 0x0020},
+{0x012475, 0x0080},
+{0x012480, 0x0004},
+{0x012544, 0x0080},
+{0x012F90, 0x0004},
+{0x012FF1, 0x0020},
+{0x012FF3, 0x0080},
+{0x013000, 0x0004},
+{0x013430, 0x0080},
+{0x013440, 0x0010},
+{0x013441, 0x0004},
+{0x013447, 0x0010},
+{0x013456, 0x0080},
+{0x014400, 0x0004},
+{0x014647, 0x0080},
+{0x016800, 0x0004},
+{0x016A39, 0x0080},
+{0x016A40, 0x0004},
+{0x016A5F, 0x0080},
+{0x016A60, 0x0002},
+{0x016A6A, 0x0080},
+{0x016A6E, 0x0020},
+{0x016A70, 0x0004},
+{0x016ABF, 0x0080},
+{0x016AC0, 0x0002},
+{0x016ACA, 0x0080},
+{0x016AD0, 0x0004},
+{0x016AEE, 0x0080},
+{0x016AF0, 0x0010},
+{0x016AF5, 0x0020},
+{0x016AF6, 0x0080},
+{0x016B00, 0x0004},
+{0x016B30, 0x0010},
+{0x016B37, 0x0020},
+{0x016B3C, 0x0040},
+{0x016B40, 0x0004},
+{0x016B44, 0x0020},
+{0x016B45, 0x0040},
+{0x016B46, 0x0080},
+{0x016B50, 0x0002},
+{0x016B5A, 0x0080},
+{0x016B5B, 0x0002},
+{0x016B62, 0x0080},
+{0x016B63, 0x0004},
+{0x016B78, 0x0080},
+{0x016B7D, 0x0004},
+{0x016B90, 0x0080},
+{0x016E40, 0x0004},
+{0x016E80, 0x0002},
+{0x016E97, 0x0020},
+{0x016E9B, 0x0080},
+{0x016F00, 0x0004},
+{0x016F4B, 0x0080},
+{0x016F4F, 0x0010},
+{0x016F50, 0x0004},
+{0x016F51, 0x0010},
+{0x016F88, 0x0080},
+{0x016F8F, 0x0010},
+{0x016F93, 0x0004},
+{0x016FA0, 0x0080},
+{0x016FE0, 0x0004},
+{0x016FE2, 0x0020},
+{0x016FE3, 0x0004},
+{0x016FE4, 0x0010},
+{0x016FE5, 0x0080},
+{0x016FF0, 0x0010},
+{0x016FF2, 0x0080},
+{0x017000, 0x0004},
+{0x0187F8, 0x0080},
+{0x018800, 0x0004},
+{0x018CD6, 0x0080},
+{0x018D00, 0x0004},
+{0x018D09, 0x0080},
+{0x01AFF0, 0x0004},
+{0x01AFF4, 0x0080},
+{0x01AFF5, 0x0004},
+{0x01AFFC, 0x0080},
+{0x01AFFD, 0x0004},
+{0x01AFFF, 0x0080},
+{0x01B000, 0x0004},
+{0x01B123, 0x0080},
+{0x01B132, 0x0004},
+{0x01B133, 0x0080},
+{0x01B150, 0x0004},
+{0x01B153, 0x0080},
+{0x01B155, 0x0004},
+{0x01B156, 0x0080},
+{0x01B164, 0x0004},
+{0x01B168, 0x0080},
+{0x01B170, 0x0004},
+{0x01B2FC, 0x0080},
+{0x01BC00, 0x0004},
+{0x01BC6B, 0x0080},
+{0x01BC70, 0x0004},
+{0x01BC7D, 0x0080},
+{0x01BC80, 0x0004},
+{0x01BC89, 0x0080},
+{0x01BC90, 0x0004},
+{0x01BC9A, 0x0080},
+{0x01BC9C, 0x0040},
+{0x01BC9D, 0x0010},
+{0x01BC9F, 0x0020},
+{0x01BCA0, 0x0080},
+{0x01CF00, 0x0010},
+{0x01CF2E, 0x0080},
+{0x01CF30, 0x0010},
+{0x01CF47, 0x0080},
+{0x01CF50, 0x0040},
+{0x01CFC4, 0x0080},
+{0x01D000, 0x0040},
+{0x01D0F6, 0x0080},
+{0x01D100, 0x0040},
+{0x01D127, 0x0080},
+{0x01D129, 0x0040},
+{0x01D165, 0x0010},
+{0x01D16A, 0x0040},
+{0x01D16D, 0x0010},
+{0x01D173, 0x0080},
+{0x01D17B, 0x0010},
+{0x01D183, 0x0040},
+{0x01D185, 0x0010},
+{0x01D18C, 0x0040},
+{0x01D1AA, 0x0010},
+{0x01D1AE, 0x0040},
+{0x01D1EB, 0x0080},
+{0x01D200, 0x0040},
+{0x01D242, 0x0010},
+{0x01D245, 0x0040},
+{0x01D246, 0x0080},
+{0x01D2C0, 0x0002},
+{0x01D2D4, 0x0080},
+{0x01D2E0, 0x0002},
+{0x01D2F4, 0x0080},
+{0x01D300, 0x0040},
+{0x01D357, 0x0080},
+{0x01D360, 0x0002},
+{0x01D379, 0x0080},
+{0x01D400, 0x0004},
+{0x01D455, 0x0080},
+{0x01D456, 0x0004},
+{0x01D49D, 0x0080},
+{0x01D49E, 0x0004},
+{0x01D4A0, 0x0080},
+{0x01D4A2, 0x0004},
+{0x01D4A3, 0x0080},
+{0x01D4A5, 0x0004},
+{0x01D4A7, 0x0080},
+{0x01D4A9, 0x0004},
+{0x01D4AD, 0x0080},
+{0x01D4AE, 0x0004},
+{0x01D4BA, 0x0080},
+{0x01D4BB, 0x0004},
+{0x01D4BC, 0x0080},
+{0x01D4BD, 0x0004},
+{0x01D4C4, 0x0080},
+{0x01D4C5, 0x0004},
+{0x01D506, 0x0080},
+{0x01D507, 0x0004},
+{0x01D50B, 0x0080},
+{0x01D50D, 0x0004},
+{0x01D515, 0x0080},
+{0x01D516, 0x0004},
+{0x01D51D, 0x0080},
+{0x01D51E, 0x0004},
+{0x01D53A, 0x0080},
+{0x01D53B, 0x0004},
+{0x01D53F, 0x0080},
+{0x01D540, 0x0004},
+{0x01D545, 0x0080},
+{0x01D546, 0x0004},
+{0x01D547, 0x0080},
+{0x01D54A, 0x0004},
+{0x01D551, 0x0080},
+{0x01D552, 0x0004},
+{0x01D6A6, 0x0080},
+{0x01D6A8, 0x0004},
+{0x01D6C1, 0x0040},
+{0x01D6C2, 0x0004},
+{0x01D6DB, 0x0040},
+{0x01D6DC, 0x0004},
+{0x01D6FB, 0x0040},
+{0x01D6FC, 0x0004},
+{0x01D715, 0x0040},
+{0x01D716, 0x0004},
+{0x01D735, 0x0040},
+{0x01D736, 0x0004},
+{0x01D74F, 0x0040},
+{0x01D750, 0x0004},
+{0x01D76F, 0x0040},
+{0x01D770, 0x0004},
+{0x01D789, 0x0040},
+{0x01D78A, 0x0004},
+{0x01D7A9, 0x0040},
+{0x01D7AA, 0x0004},
+{0x01D7C3, 0x0040},
+{0x01D7C4, 0x0004},
+{0x01D7CC, 0x0080},
+{0x01D7CE, 0x0002},
+{0x01D800, 0x0040},
+{0x01DA00, 0x0010},
+{0x01DA37, 0x0040},
+{0x01DA3B, 0x0010},
+{0x01DA6D, 0x0040},
+{0x01DA75, 0x0010},
+{0x01DA76, 0x0040},
+{0x01DA84, 0x0010},
+{0x01DA85, 0x0040},
+{0x01DA87, 0x0020},
+{0x01DA8C, 0x0080},
+{0x01DA9B, 0x0010},
+{0x01DAA0, 0x0080},
+{0x01DAA1, 0x0010},
+{0x01DAB0, 0x0080},
+{0x01DF00, 0x0004},
+{0x01DF1F, 0x0080},
+{0x01DF25, 0x0004},
+{0x01DF2B, 0x0080},
+{0x01E000, 0x0010},
+{0x01E007, 0x0080},
+{0x01E008, 0x0010},
+{0x01E019, 0x0080},
+{0x01E01B, 0x0010},
+{0x01E022, 0x0080},
+{0x01E023, 0x0010},
+{0x01E025, 0x0080},
+{0x01E026, 0x0010},
+{0x01E02B, 0x0080},
+{0x01E030, 0x0004},
+{0x01E06E, 0x0080},
+{0x01E08F, 0x0010},
+{0x01E090, 0x0080},
+{0x01E100, 0x0004},
+{0x01E12D, 0x0080},
+{0x01E130, 0x0010},
+{0x01E137, 0x0004},
+{0x01E13E, 0x0080},
+{0x01E140, 0x0002},
+{0x01E14A, 0x0080},
+{0x01E14E, 0x0004},
+{0x01E14F, 0x0040},
+{0x01E150, 0x0080},
+{0x01E290, 0x0004},
+{0x01E2AE, 0x0010},
+{0x01E2AF, 0x0080},
+{0x01E2C0, 0x0004},
+{0x01E2EC, 0x0010},
+{0x01E2F0, 0x0002},
+{0x01E2FA, 0x0080},
+{0x01E2FF, 0x0040},
+{0x01E300, 0x0080},
+{0x01E4D0, 0x0004},
+{0x01E4EC, 0x0010},
+{0x01E4F0, 0x0002},
+{0x01E4FA, 0x0080},
+{0x01E7E0, 0x0004},
+{0x01E7E7, 0x0080},
+{0x01E7E8, 0x0004},
+{0x01E7EC, 0x0080},
+{0x01E7ED, 0x0004},
+{0x01E7EF, 0x0080},
+{0x01E7F0, 0x0004},
+{0x01E7FF, 0x0080},
+{0x01E800, 0x0004},
+{0x01E8C5, 0x0080},
+{0x01E8C7, 0x0002},
+{0x01E8D0, 0x0010},
+{0x01E8D7, 0x0080},
+{0x01E900, 0x0004},
+{0x01E944, 0x0010},
+{0x01E94B, 0x0004},
+{0x01E94C, 0x0080},
+{0x01E950, 0x0002},
+{0x01E95A, 0x0080},
+{0x01E95E, 0x0020},
+{0x01E960, 0x0080},
+{0x01EC71, 0x0002},
+{0x01ECAC, 0x0040},
+{0x01ECAD, 0x0002},
+{0x01ECB0, 0x0040},
+{0x01ECB1, 0x0002},
+{0x01ECB5, 0x0080},
+{0x01ED01, 0x0002},
+{0x01ED2E, 0x0040},
+{0x01ED2F, 0x0002},
+{0x01ED3E, 0x0080},
+{0x01EE00, 0x0004},
+{0x01EE04, 0x0080},
+{0x01EE05, 0x0004},
+{0x01EE20, 0x0080},
+{0x01EE21, 0x0004},
+{0x01EE23, 0x0080},
+{0x01EE24, 0x0004},
+{0x01EE25, 0x0080},
+{0x01EE27, 0x0004},
+{0x01EE28, 0x0080},
+{0x01EE29, 0x0004},
+{0x01EE33, 0x0080},
+{0x01EE34, 0x0004},
+{0x01EE38, 0x0080},
+{0x01EE39, 0x0004},
+{0x01EE3A, 0x0080},
+{0x01EE3B, 0x0004},
+{0x01EE3C, 0x0080},
+{0x01EE42, 0x0004},
+{0x01EE43, 0x0080},
+{0x01EE47, 0x0004},
+{0x01EE48, 0x0080},
+{0x01EE49, 0x0004},
+{0x01EE4A, 0x0080},
+{0x01EE4B, 0x0004},
+{0x01EE4C, 0x0080},
+{0x01EE4D, 0x0004},
+{0x01EE50, 0x0080},
+{0x01EE51, 0x0004},
+{0x01EE53, 0x0080},
+{0x01EE54, 0x0004},
+{0x01EE55, 0x0080},
+{0x01EE57, 0x0004},
+{0x01EE58, 0x0080},
+{0x01EE59, 0x0004},
+{0x01EE5A, 0x0080},
+{0x01EE5B, 0x0004},
+{0x01EE5C, 0x0080},
+{0x01EE5D, 0x0004},
+{0x01EE5E, 0x0080},
+{0x01EE5F, 0x0004},
+{0x01EE60, 0x0080},
+{0x01EE61, 0x0004},
+{0x01EE63, 0x0080},
+{0x01EE64, 0x0004},
+{0x01EE65, 0x0080},
+{0x01EE67, 0x0004},
+{0x01EE6B, 0x0080},
+{0x01EE6C, 0x0004},
+{0x01EE73, 0x0080},
+{0x01EE74, 0x0004},
+{0x01EE78, 0x0080},
+{0x01EE79, 0x0004},
+{0x01EE7D, 0x0080},
+{0x01EE7E, 0x0004},
+{0x01EE7F, 0x0080},
+{0x01EE80, 0x0004},
+{0x01EE8A, 0x0080},
+{0x01EE8B, 0x0004},
+{0x01EE9C, 0x0080},
+{0x01EEA1, 0x0004},
+{0x01EEA4, 0x0080},
+{0x01EEA5, 0x0004},
+{0x01EEAA, 0x0080},
+{0x01EEAB, 0x0004},
+{0x01EEBC, 0x0080},
+{0x01EEF0, 0x0040},
+{0x01EEF2, 0x0080},
+{0x01F000, 0x0040},
+{0x01F02C, 0x0080},
+{0x01F030, 0x0040},
+{0x01F094, 0x0080},
+{0x01F0A0, 0x0040},
+{0x01F0AF, 0x0080},
+{0x01F0B1, 0x0040},
+{0x01F0C0, 0x0080},
+{0x01F0C1, 0x0040},
+{0x01F0D0, 0x0080},
+{0x01F0D1, 0x0040},
+{0x01F0F6, 0x0080},
+{0x01F100, 0x0002},
+{0x01F10D, 0x0040},
+{0x01F1AE, 0x0080},
+{0x01F1E6, 0x0040},
+{0x01F203, 0x0080},
+{0x01F210, 0x0040},
+{0x01F23C, 0x0080},
+{0x01F240, 0x0040},
+{0x01F249, 0x0080},
+{0x01F250, 0x0040},
+{0x01F252, 0x0080},
+{0x01F260, 0x0040},
+{0x01F266, 0x0080},
+{0x01F300, 0x0040},
+{0x01F6D8, 0x0080},
+{0x01F6DC, 0x0040},
+{0x01F6ED, 0x0080},
+{0x01F6F0, 0x0040},
+{0x01F6FD, 0x0080},
+{0x01F700, 0x0040},
+{0x01F777, 0x0080},
+{0x01F77B, 0x0040},
+{0x01F7DA, 0x0080},
+{0x01F7E0, 0x0040},
+{0x01F7EC, 0x0080},
+{0x01F7F0, 0x0040},
+{0x01F7F1, 0x0080},
+{0x01F800, 0x0040},
+{0x01F80C, 0x0080},
+{0x01F810, 0x0040},
+{0x01F848, 0x0080},
+{0x01F850, 0x0040},
+{0x01F85A, 0x0080},
+{0x01F860, 0x0040},
+{0x01F888, 0x0080},
+{0x01F890, 0x0040},
+{0x01F8AE, 0x0080},
+{0x01F8B0, 0x0040},
+{0x01F8B2, 0x0080},
+{0x01F900, 0x0040},
+{0x01FA54, 0x0080},
+{0x01FA60, 0x0040},
+{0x01FA6E, 0x0080},
+{0x01FA70, 0x0040},
+{0x01FA7D, 0x0080},
+{0x01FA80, 0x0040},
+{0x01FA89, 0x0080},
+{0x01FA90, 0x0040},
+{0x01FABE, 0x0080},
+{0x01FABF, 0x0040},
+{0x01FAC6, 0x0080},
+{0x01FACE, 0x0040},
+{0x01FADC, 0x0080},
+{0x01FAE0, 0x0040},
+{0x01FAE9, 0x0080},
+{0x01FAF0, 0x0040},
+{0x01FAF9, 0x0080},
+{0x01FB00, 0x0040},
+{0x01FB93, 0x0080},
+{0x01FB94, 0x0040},
+{0x01FBCB, 0x0080},
+{0x01FBF0, 0x0002},
+{0x01FBFA, 0x0080},
+{0x020000, 0x0004},
+{0x02A6E0, 0x0080},
+{0x02A700, 0x0004},
+{0x02B73A, 0x0080},
+{0x02B740, 0x0004},
+{0x02B81E, 0x0080},
+{0x02B820, 0x0004},
+{0x02CEA2, 0x0080},
+{0x02CEB0, 0x0004},
+{0x02EBE1, 0x0080},
+{0x02EBF0, 0x0004},
+{0x02EE5E, 0x0080},
+{0x02F800, 0x0004},
+{0x02FA1E, 0x0080},
+{0x030000, 0x0004},
+{0x03134B, 0x0080},
+{0x031350, 0x0004},
+{0x0323B0, 0x0080},
+{0x0E0100, 0x0010},
+{0x0E01F0, 0x0080},
+{0x110000, 0x0000},
 };
 
-const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation = {
-{0x00000021, 0x00000023}, {0x00000025, 0x0000002A}, {0x0000002C, 0x0000002F}, {0x0000003A, 0x0000003B},
-{0x0000003F, 0x00000040}, {0x0000005B, 0x0000005D}, {0x0000005F, 0x0000005F}, {0x0000007B, 0x0000007B},
-{0x0000007D, 0x0000007D}, {0x000000A1, 0x000000A1}, {0x000000A7, 0x000000A7}, {0x000000AB, 0x000000AB},
-{0x000000B6, 0x000000B7}, {0x000000BB, 0x000000BB}, {0x000000BF, 0x000000BF}, {0x0000037E, 0x0000037E},
-{0x00000387, 0x00000387}, {0x0000055A, 0x0000055F}, {0x00000589, 0x0000058A}, {0x000005BE, 0x000005BE},
-{0x000005C0, 0x000005C0}, {0x000005C3, 0x000005C3}, {0x000005C6, 0x000005C6}, {0x000005F3, 0x000005F4},
-{0x00000609, 0x0000060A}, {0x0000060C, 0x0000060D}, {0x0000061B, 0x0000061B}, {0x0000061E, 0x0000061F},
-{0x0000066A, 0x0000066D}, {0x000006D4, 0x000006D4}, {0x00000700, 0x0000070D}, {0x000007F7, 0x000007F9},
-{0x00000830, 0x0000083E}, {0x0000085E, 0x0000085E}, {0x00000964, 0x00000965}, {0x00000970, 0x00000970},
-{0x000009FD, 0x000009FD}, {0x00000A76, 0x00000A76}, {0x00000AF0, 0x00000AF0}, {0x00000C77, 0x00000C77},
-{0x00000C84, 0x00000C84}, {0x00000DF4, 0x00000DF4}, {0x00000E4F, 0x00000E4F}, {0x00000E5A, 0x00000E5B},
-{0x00000F04, 0x00000F12}, {0x00000F14, 0x00000F14}, {0x00000F3A, 0x00000F3D}, {0x00000F85, 0x00000F85},
-{0x00000FD0, 0x00000FD4}, {0x00000FD9, 0x00000FDA}, {0x0000104A, 0x0000104F}, {0x000010FB, 0x000010FB},
-{0x00001360, 0x00001368}, {0x00001400, 0x00001400}, {0x0000166E, 0x0000166E}, {0x0000169B, 0x0000169C},
-{0x000016EB, 0x000016ED}, {0x00001735, 0x00001736}, {0x000017D4, 0x000017D6}, {0x000017D8, 0x000017DA},
-{0x00001800, 0x0000180A}, {0x00001944, 0x00001945}, {0x00001A1E, 0x00001A1F}, {0x00001AA0, 0x00001AA6},
-{0x00001AA8, 0x00001AAD}, {0x00001B5A, 0x00001B60}, {0x00001BFC, 0x00001BFF}, {0x00001C3B, 0x00001C3F},
-{0x00001C7E, 0x00001C7F}, {0x00001CC0, 0x00001CC7}, {0x00001CD3, 0x00001CD3}, {0x00002010, 0x00002027},
-{0x00002030, 0x00002043}, {0x00002045, 0x00002051}, {0x00002053, 0x0000205E}, {0x0000207D, 0x0000207E},
-{0x0000208D, 0x0000208E}, {0x00002308, 0x0000230B}, {0x00002329, 0x0000232A}, {0x00002768, 0x00002775},
-{0x000027C5, 0x000027C6}, {0x000027E6, 0x000027EF}, {0x00002983, 0x00002998}, {0x000029D8, 0x000029DB},
-{0x000029FC, 0x000029FD}, {0x00002CF9, 0x00002CFC}, {0x00002CFE, 0x00002CFF}, {0x00002D70, 0x00002D70},
-{0x00002E00, 0x00002E2E}, {0x00002E30, 0x00002E4F}, {0x00002E52, 0x00002E52}, {0x00003001, 0x00003003},
-{0x00003008, 0x00003011}, {0x00003014, 0x0000301F}, {0x00003030, 0x00003030}, {0x0000303D, 0x0000303D},
-{0x000030A0, 0x000030A0}, {0x000030FB, 0x000030FB}, {0x0000A4FE, 0x0000A4FF}, {0x0000A60D, 0x0000A60F},
-{0x0000A673, 0x0000A673}, {0x0000A67E, 0x0000A67E}, {0x0000A6F2, 0x0000A6F7}, {0x0000A874, 0x0000A877},
-{0x0000A8CE, 0x0000A8CF}, {0x0000A8F8, 0x0000A8FA}, {0x0000A8FC, 0x0000A8FC}, {0x0000A92E, 0x0000A92F},
-{0x0000A95F, 0x0000A95F}, {0x0000A9C1, 0x0000A9CD}, {0x0000A9DE, 0x0000A9DF}, {0x0000AA5C, 0x0000AA5F},
-{0x0000AADE, 0x0000AADF}, {0x0000AAF0, 0x0000AAF1}, {0x0000ABEB, 0x0000ABEB}, {0x0000FD3E, 0x0000FD3F},
-{0x0000FE10, 0x0000FE19}, {0x0000FE30, 0x0000FE52}, {0x0000FE54, 0x0000FE61}, {0x0000FE63, 0x0000FE63},
-{0x0000FE68, 0x0000FE68}, {0x0000FE6A, 0x0000FE6B}, {0x0000FF01, 0x0000FF03}, {0x0000FF05, 0x0000FF0A},
-{0x0000FF0C, 0x0000FF0F}, {0x0000FF1A, 0x0000FF1B}, {0x0000FF1F, 0x0000FF20}, {0x0000FF3B, 0x0000FF3D},
-{0x0000FF3F, 0x0000FF3F}, {0x0000FF5B, 0x0000FF5B}, {0x0000FF5D, 0x0000FF5D}, {0x0000FF5F, 0x0000FF65},
-{0x00010100, 0x00010102}, {0x0001039F, 0x0001039F}, {0x000103D0, 0x000103D0}, {0x0001056F, 0x0001056F},
-{0x00010857, 0x00010857}, {0x0001091F, 0x0001091F}, {0x0001093F, 0x0001093F}, {0x00010A50, 0x00010A58},
-{0x00010A7F, 0x00010A7F}, {0x00010AF0, 0x00010AF6}, {0x00010B39, 0x00010B3F}, {0x00010B99, 0x00010B9C},
-{0x00010EAD, 0x00010EAD}, {0x00010F55, 0x00010F59}, {0x00011047, 0x0001104D}, {0x000110BB, 0x000110BC},
-{0x000110BE, 0x000110C1}, {0x00011140, 0x00011143}, {0x00011174, 0x00011175}, {0x000111C5, 0x000111C8},
-{0x000111CD, 0x000111CD}, {0x000111DB, 0x000111DB}, {0x000111DD, 0x000111DF}, {0x00011238, 0x0001123D},
-{0x000112A9, 0x000112A9}, {0x0001144B, 0x0001144F}, {0x0001145A, 0x0001145B}, {0x0001145D, 0x0001145D},
-{0x000114C6, 0x000114C6}, {0x000115C1, 0x000115D7}, {0x00011641, 0x00011643}, {0x00011660, 0x0001166C},
-{0x0001173C, 0x0001173E}, {0x0001183B, 0x0001183B}, {0x00011944, 0x00011946}, {0x000119E2, 0x000119E2},
-{0x00011A3F, 0x00011A46}, {0x00011A9A, 0x00011A9C}, {0x00011A9E, 0x00011AA2}, {0x00011C41, 0x00011C45},
-{0x00011C70, 0x00011C71}, {0x00011EF7, 0x00011EF8}, {0x00011FFF, 0x00011FFF}, {0x00012470, 0x00012474},
-{0x00016A6E, 0x00016A6F}, {0x00016AF5, 0x00016AF5}, {0x00016B37, 0x00016B3B}, {0x00016B44, 0x00016B44},
-{0x00016E97, 0x00016E9A}, {0x00016FE2, 0x00016FE2}, {0x0001BC9F, 0x0001BC9F}, {0x0001DA87, 0x0001DA8B},
-{0x0001E95E, 0x0001E95F},
+const std::unordered_set<uint32_t> unicode_set_whitespace = {
+0x000009, 0x00000A, 0x00000B, 0x00000C, 0x00000D, 0x000020, 0x000085, 0x0000A0, 0x001680, 0x002000, 0x002001, 0x002002, 0x002003, 0x002004, 0x002005, 0x002006, 0x002007, 0x002008, 0x002009, 0x00200A, 0x002028, 0x002029, 0x00202F, 0x00205F, 0x003000
 };
 
-const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol = {
-{0x00000024, 0x00000024}, {0x0000002B, 0x0000002B}, {0x0000003C, 0x0000003E}, {0x0000005E, 0x0000005E},
-{0x00000060, 0x00000060}, {0x0000007C, 0x0000007C}, {0x0000007E, 0x0000007E}, {0x000000A2, 0x000000A6},
-{0x000000A8, 0x000000A9}, {0x000000AC, 0x000000AC}, {0x000000AE, 0x000000B1}, {0x000000B4, 0x000000B4},
-{0x000000B8, 0x000000B8}, {0x000000D7, 0x000000D7}, {0x000000F7, 0x000000F7}, {0x000002C2, 0x000002C5},
-{0x000002D2, 0x000002DF}, {0x000002E5, 0x000002EB}, {0x000002ED, 0x000002ED}, {0x000002EF, 0x000002FF},
-{0x00000375, 0x00000375}, {0x00000384, 0x00000385}, {0x000003F6, 0x000003F6}, {0x00000482, 0x00000482},
-{0x0000058D, 0x0000058F}, {0x00000606, 0x00000608}, {0x0000060B, 0x0000060B}, {0x0000060E, 0x0000060F},
-{0x000006DE, 0x000006DE}, {0x000006E9, 0x000006E9}, {0x000006FD, 0x000006FE}, {0x000007F6, 0x000007F6},
-{0x000007FE, 0x000007FF}, {0x000009F2, 0x000009F3}, {0x000009FA, 0x000009FB}, {0x00000AF1, 0x00000AF1},
-{0x00000B70, 0x00000B70}, {0x00000BF3, 0x00000BFA}, {0x00000C7F, 0x00000C7F}, {0x00000D4F, 0x00000D4F},
-{0x00000D79, 0x00000D79}, {0x00000E3F, 0x00000E3F}, {0x00000F01, 0x00000F03}, {0x00000F13, 0x00000F13},
-{0x00000F15, 0x00000F17}, {0x00000F1A, 0x00000F1F}, {0x00000F34, 0x00000F34}, {0x00000F36, 0x00000F36},
-{0x00000F38, 0x00000F38}, {0x00000FBE, 0x00000FC5}, {0x00000FC7, 0x00000FCC}, {0x00000FCE, 0x00000FCF},
-{0x00000FD5, 0x00000FD8}, {0x0000109E, 0x0000109F}, {0x00001390, 0x00001399}, {0x0000166D, 0x0000166D},
-{0x000017DB, 0x000017DB}, {0x00001940, 0x00001940}, {0x000019DE, 0x000019FF}, {0x00001B61, 0x00001B6A},
-{0x00001B74, 0x00001B7C}, {0x00001FBD, 0x00001FBD}, {0x00001FBF, 0x00001FC1}, {0x00001FCD, 0x00001FCF},
-{0x00001FDD, 0x00001FDF}, {0x00001FED, 0x00001FEF}, {0x00001FFD, 0x00001FFE}, {0x00002044, 0x00002044},
-{0x00002052, 0x00002052}, {0x0000207A, 0x0000207C}, {0x0000208A, 0x0000208C}, {0x000020A0, 0x000020BF},
-{0x00002100, 0x00002101}, {0x00002103, 0x00002106}, {0x00002108, 0x00002109}, {0x00002114, 0x00002114},
-{0x00002116, 0x00002118}, {0x0000211E, 0x00002123}, {0x00002125, 0x00002125}, {0x00002127, 0x00002127},
-{0x00002129, 0x00002129}, {0x0000212E, 0x0000212E}, {0x0000213A, 0x0000213B}, {0x00002140, 0x00002144},
-{0x0000214A, 0x0000214D}, {0x0000214F, 0x0000214F}, {0x0000218A, 0x0000218B}, {0x00002190, 0x00002307},
-{0x0000230C, 0x00002328}, {0x0000232B, 0x00002426}, {0x00002440, 0x0000244A}, {0x0000249C, 0x000024E9},
-{0x00002500, 0x00002767}, {0x00002794, 0x000027C4}, {0x000027C7, 0x000027E5}, {0x000027F0, 0x00002982},
-{0x00002999, 0x000029D7}, {0x000029DC, 0x000029FB}, {0x000029FE, 0x00002B73}, {0x00002B76, 0x00002B95},
-{0x00002B97, 0x00002BFF}, {0x00002CE5, 0x00002CEA}, {0x00002E50, 0x00002E51}, {0x00002E80, 0x00002E99},
-{0x00002E9B, 0x00002EF3}, {0x00002F00, 0x00002FD5}, {0x00002FF0, 0x00002FFB}, {0x00003004, 0x00003004},
-{0x00003012, 0x00003013}, {0x00003020, 0x00003020}, {0x00003036, 0x00003037}, {0x0000303E, 0x0000303F},
-{0x0000309B, 0x0000309C}, {0x00003190, 0x00003191}, {0x00003196, 0x0000319F}, {0x000031C0, 0x000031E3},
-{0x00003200, 0x0000321E}, {0x0000322A, 0x00003247}, {0x00003250, 0x00003250}, {0x00003260, 0x0000327F},
-{0x0000328A, 0x000032B0}, {0x000032C0, 0x000033FF}, {0x00004DC0, 0x00004DFF}, {0x0000A490, 0x0000A4C6},
-{0x0000A700, 0x0000A716}, {0x0000A720, 0x0000A721}, {0x0000A789, 0x0000A78A}, {0x0000A828, 0x0000A82B},
-{0x0000A836, 0x0000A839}, {0x0000AA77, 0x0000AA79}, {0x0000AB5B, 0x0000AB5B}, {0x0000AB6A, 0x0000AB6B},
-{0x0000FB29, 0x0000FB29}, {0x0000FBB2, 0x0000FBC1}, {0x0000FDFC, 0x0000FDFD}, {0x0000FE62, 0x0000FE62},
-{0x0000FE64, 0x0000FE66}, {0x0000FE69, 0x0000FE69}, {0x0000FF04, 0x0000FF04}, {0x0000FF0B, 0x0000FF0B},
-{0x0000FF1C, 0x0000FF1E}, {0x0000FF3E, 0x0000FF3E}, {0x0000FF40, 0x0000FF40}, {0x0000FF5C, 0x0000FF5C},
-{0x0000FF5E, 0x0000FF5E}, {0x0000FFE0, 0x0000FFE6}, {0x0000FFE8, 0x0000FFEE}, {0x0000FFFC, 0x0000FFFD},
-{0x00010137, 0x0001013F}, {0x00010179, 0x00010189}, {0x0001018C, 0x0001018E}, {0x00010190, 0x0001019C},
-{0x000101A0, 0x000101A0}, {0x000101D0, 0x000101FC}, {0x00010877, 0x00010878}, {0x00010AC8, 0x00010AC8},
-{0x0001173F, 0x0001173F}, {0x00011FD5, 0x00011FF1}, {0x00016B3C, 0x00016B3F}, {0x00016B45, 0x00016B45},
-{0x0001BC9C, 0x0001BC9C}, {0x0001D000, 0x0001D0F5}, {0x0001D100, 0x0001D126}, {0x0001D129, 0x0001D164},
-{0x0001D16A, 0x0001D16C}, {0x0001D183, 0x0001D184}, {0x0001D18C, 0x0001D1A9}, {0x0001D1AE, 0x0001D1E8},
-{0x0001D200, 0x0001D241}, {0x0001D245, 0x0001D245}, {0x0001D300, 0x0001D356}, {0x0001D6C1, 0x0001D6C1},
-{0x0001D6DB, 0x0001D6DB}, {0x0001D6FB, 0x0001D6FB}, {0x0001D715, 0x0001D715}, {0x0001D735, 0x0001D735},
-{0x0001D74F, 0x0001D74F}, {0x0001D76F, 0x0001D76F}, {0x0001D789, 0x0001D789}, {0x0001D7A9, 0x0001D7A9},
-{0x0001D7C3, 0x0001D7C3}, {0x0001D800, 0x0001D9FF}, {0x0001DA37, 0x0001DA3A}, {0x0001DA6D, 0x0001DA74},
-{0x0001DA76, 0x0001DA83}, {0x0001DA85, 0x0001DA86}, {0x0001E14F, 0x0001E14F}, {0x0001E2FF, 0x0001E2FF},
-{0x0001ECAC, 0x0001ECAC}, {0x0001ECB0, 0x0001ECB0}, {0x0001ED2E, 0x0001ED2E}, {0x0001EEF0, 0x0001EEF1},
-{0x0001F000, 0x0001F02B}, {0x0001F030, 0x0001F093}, {0x0001F0A0, 0x0001F0AE}, {0x0001F0B1, 0x0001F0BF},
-{0x0001F0C1, 0x0001F0CF}, {0x0001F0D1, 0x0001F0F5}, {0x0001F10D, 0x0001F1AD}, {0x0001F1E6, 0x0001F202},
-{0x0001F210, 0x0001F23B}, {0x0001F240, 0x0001F248}, {0x0001F250, 0x0001F251}, {0x0001F260, 0x0001F265},
-{0x0001F300, 0x0001F6D7}, {0x0001F6E0, 0x0001F6EC}, {0x0001F6F0, 0x0001F6FC}, {0x0001F700, 0x0001F773},
-{0x0001F780, 0x0001F7D8}, {0x0001F7E0, 0x0001F7EB}, {0x0001F800, 0x0001F80B}, {0x0001F810, 0x0001F847},
-{0x0001F850, 0x0001F859}, {0x0001F860, 0x0001F887}, {0x0001F890, 0x0001F8AD}, {0x0001F8B0, 0x0001F8B1},
-{0x0001F900, 0x0001F978}, {0x0001F97A, 0x0001F9CB}, {0x0001F9CD, 0x0001FA53}, {0x0001FA60, 0x0001FA6D},
-{0x0001FA70, 0x0001FA74}, {0x0001FA78, 0x0001FA7A}, {0x0001FA80, 0x0001FA86}, {0x0001FA90, 0x0001FAA8},
-{0x0001FAB0, 0x0001FAB6}, {0x0001FAC0, 0x0001FAC2}, {0x0001FAD0, 0x0001FAD6}, {0x0001FB00, 0x0001FB92},
-{0x0001FB94, 0x0001FBCA},
+const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase = {
+{0x000041, 0x000061},
+{0x000042, 0x000062},
+{0x000043, 0x000063},
+{0x000044, 0x000064},
+{0x000045, 0x000065},
+{0x000046, 0x000066},
+{0x000047, 0x000067},
+{0x000048, 0x000068},
+{0x000049, 0x000069},
+{0x00004A, 0x00006A},
+{0x00004B, 0x00006B},
+{0x00004C, 0x00006C},
+{0x00004D, 0x00006D},
+{0x00004E, 0x00006E},
+{0x00004F, 0x00006F},
+{0x000050, 0x000070},
+{0x000051, 0x000071},
+{0x000052, 0x000072},
+{0x000053, 0x000073},
+{0x000054, 0x000074},
+{0x000055, 0x000075},
+{0x000056, 0x000076},
+{0x000057, 0x000077},
+{0x000058, 0x000078},
+{0x000059, 0x000079},
+{0x00005A, 0x00007A},
+{0x0000C0, 0x0000E0},
+{0x0000C1, 0x0000E1},
+{0x0000C2, 0x0000E2},
+{0x0000C3, 0x0000E3},
+{0x0000C4, 0x0000E4},
+{0x0000C5, 0x0000E5},
+{0x0000C6, 0x0000E6},
+{0x0000C7, 0x0000E7},
+{0x0000C8, 0x0000E8},
+{0x0000C9, 0x0000E9},
+{0x0000CA, 0x0000EA},
+{0x0000CB, 0x0000EB},
+{0x0000CC, 0x0000EC},
+{0x0000CD, 0x0000ED},
+{0x0000CE, 0x0000EE},
+{0x0000CF, 0x0000EF},
+{0x0000D0, 0x0000F0},
+{0x0000D1, 0x0000F1},
+{0x0000D2, 0x0000F2},
+{0x0000D3, 0x0000F3},
+{0x0000D4, 0x0000F4},
+{0x0000D5, 0x0000F5},
+{0x0000D6, 0x0000F6},
+{0x0000D8, 0x0000F8},
+{0x0000D9, 0x0000F9},
+{0x0000DA, 0x0000FA},
+{0x0000DB, 0x0000FB},
+{0x0000DC, 0x0000FC},
+{0x0000DD, 0x0000FD},
+{0x0000DE, 0x0000FE},
+{0x000100, 0x000101},
+{0x000102, 0x000103},
+{0x000104, 0x000105},
+{0x000106, 0x000107},
+{0x000108, 0x000109},
+{0x00010A, 0x00010B},
+{0x00010C, 0x00010D},
+{0x00010E, 0x00010F},
+{0x000110, 0x000111},
+{0x000112, 0x000113},
+{0x000114, 0x000115},
+{0x000116, 0x000117},
+{0x000118, 0x000119},
+{0x00011A, 0x00011B},
+{0x00011C, 0x00011D},
+{0x00011E, 0x00011F},
+{0x000120, 0x000121},
+{0x000122, 0x000123},
+{0x000124, 0x000125},
+{0x000126, 0x000127},
+{0x000128, 0x000129},
+{0x00012A, 0x00012B},
+{0x00012C, 0x00012D},
+{0x00012E, 0x00012F},
+{0x000130, 0x000069},
+{0x000132, 0x000133},
+{0x000134, 0x000135},
+{0x000136, 0x000137},
+{0x000139, 0x00013A},
+{0x00013B, 0x00013C},
+{0x00013D, 0x00013E},
+{0x00013F, 0x000140},
+{0x000141, 0x000142},
+{0x000143, 0x000144},
+{0x000145, 0x000146},
+{0x000147, 0x000148},
+{0x00014A, 0x00014B},
+{0x00014C, 0x00014D},
+{0x00014E, 0x00014F},
+{0x000150, 0x000151},
+{0x000152, 0x000153},
+{0x000154, 0x000155},
+{0x000156, 0x000157},
+{0x000158, 0x000159},
+{0x00015A, 0x00015B},
+{0x00015C, 0x00015D},
+{0x00015E, 0x00015F},
+{0x000160, 0x000161},
+{0x000162, 0x000163},
+{0x000164, 0x000165},
+{0x000166, 0x000167},
+{0x000168, 0x000169},
+{0x00016A, 0x00016B},
+{0x00016C, 0x00016D},
+{0x00016E, 0x00016F},
+{0x000170, 0x000171},
+{0x000172, 0x000173},
+{0x000174, 0x000175},
+{0x000176, 0x000177},
+{0x000178, 0x0000FF},
+{0x000179, 0x00017A},
+{0x00017B, 0x00017C},
+{0x00017D, 0x00017E},
+{0x000181, 0x000253},
+{0x000182, 0x000183},
+{0x000184, 0x000185},
+{0x000186, 0x000254},
+{0x000187, 0x000188},
+{0x000189, 0x000256},
+{0x00018A, 0x000257},
+{0x00018B, 0x00018C},
+{0x00018E, 0x0001DD},
+{0x00018F, 0x000259},
+{0x000190, 0x00025B},
+{0x000191, 0x000192},
+{0x000193, 0x000260},
+{0x000194, 0x000263},
+{0x000196, 0x000269},
+{0x000197, 0x000268},
+{0x000198, 0x000199},
+{0x00019C, 0x00026F},
+{0x00019D, 0x000272},
+{0x00019F, 0x000275},
+{0x0001A0, 0x0001A1},
+{0x0001A2, 0x0001A3},
+{0x0001A4, 0x0001A5},
+{0x0001A6, 0x000280},
+{0x0001A7, 0x0001A8},
+{0x0001A9, 0x000283},
+{0x0001AC, 0x0001AD},
+{0x0001AE, 0x000288},
+{0x0001AF, 0x0001B0},
+{0x0001B1, 0x00028A},
+{0x0001B2, 0x00028B},
+{0x0001B3, 0x0001B4},
+{0x0001B5, 0x0001B6},
+{0x0001B7, 0x000292},
+{0x0001B8, 0x0001B9},
+{0x0001BC, 0x0001BD},
+{0x0001C4, 0x0001C6},
+{0x0001C5, 0x0001C6},
+{0x0001C7, 0x0001C9},
+{0x0001C8, 0x0001C9},
+{0x0001CA, 0x0001CC},
+{0x0001CB, 0x0001CC},
+{0x0001CD, 0x0001CE},
+{0x0001CF, 0x0001D0},
+{0x0001D1, 0x0001D2},
+{0x0001D3, 0x0001D4},
+{0x0001D5, 0x0001D6},
+{0x0001D7, 0x0001D8},
+{0x0001D9, 0x0001DA},
+{0x0001DB, 0x0001DC},
+{0x0001DE, 0x0001DF},
+{0x0001E0, 0x0001E1},
+{0x0001E2, 0x0001E3},
+{0x0001E4, 0x0001E5},
+{0x0001E6, 0x0001E7},
+{0x0001E8, 0x0001E9},
+{0x0001EA, 0x0001EB},
+{0x0001EC, 0x0001ED},
+{0x0001EE, 0x0001EF},
+{0x0001F1, 0x0001F3},
+{0x0001F2, 0x0001F3},
+{0x0001F4, 0x0001F5},
+{0x0001F6, 0x000195},
+{0x0001F7, 0x0001BF},
+{0x0001F8, 0x0001F9},
+{0x0001FA, 0x0001FB},
+{0x0001FC, 0x0001FD},
+{0x0001FE, 0x0001FF},
+{0x000200, 0x000201},
+{0x000202, 0x000203},
+{0x000204, 0x000205},
+{0x000206, 0x000207},
+{0x000208, 0x000209},
+{0x00020A, 0x00020B},
+{0x00020C, 0x00020D},
+{0x00020E, 0x00020F},
+{0x000210, 0x000211},
+{0x000212, 0x000213},
+{0x000214, 0x000215},
+{0x000216, 0x000217},
+{0x000218, 0x000219},
+{0x00021A, 0x00021B},
+{0x00021C, 0x00021D},
+{0x00021E, 0x00021F},
+{0x000220, 0x00019E},
+{0x000222, 0x000223},
+{0x000224, 0x000225},
+{0x000226, 0x000227},
+{0x000228, 0x000229},
+{0x00022A, 0x00022B},
+{0x00022C, 0x00022D},
+{0x00022E, 0x00022F},
+{0x000230, 0x000231},
+{0x000232, 0x000233},
+{0x00023A, 0x002C65},
+{0x00023B, 0x00023C},
+{0x00023D, 0x00019A},
+{0x00023E, 0x002C66},
+{0x000241, 0x000242},
+{0x000243, 0x000180},
+{0x000244, 0x000289},
+{0x000245, 0x00028C},
+{0x000246, 0x000247},
+{0x000248, 0x000249},
+{0x00024A, 0x00024B},
+{0x00024C, 0x00024D},
+{0x00024E, 0x00024F},
+{0x000370, 0x000371},
+{0x000372, 0x000373},
+{0x000376, 0x000377},
+{0x00037F, 0x0003F3},
+{0x000386, 0x0003AC},
+{0x000388, 0x0003AD},
+{0x000389, 0x0003AE},
+{0x00038A, 0x0003AF},
+{0x00038C, 0x0003CC},
+{0x00038E, 0x0003CD},
+{0x00038F, 0x0003CE},
+{0x000391, 0x0003B1},
+{0x000392, 0x0003B2},
+{0x000393, 0x0003B3},
+{0x000394, 0x0003B4},
+{0x000395, 0x0003B5},
+{0x000396, 0x0003B6},
+{0x000397, 0x0003B7},
+{0x000398, 0x0003B8},
+{0x000399, 0x0003B9},
+{0x00039A, 0x0003BA},
+{0x00039B, 0x0003BB},
+{0x00039C, 0x0003BC},
+{0x00039D, 0x0003BD},
+{0x00039E, 0x0003BE},
+{0x00039F, 0x0003BF},
+{0x0003A0, 0x0003C0},
+{0x0003A1, 0x0003C1},
+{0x0003A3, 0x0003C3},
+{0x0003A4, 0x0003C4},
+{0x0003A5, 0x0003C5},
+{0x0003A6, 0x0003C6},
+{0x0003A7, 0x0003C7},
+{0x0003A8, 0x0003C8},
+{0x0003A9, 0x0003C9},
+{0x0003AA, 0x0003CA},
+{0x0003AB, 0x0003CB},
+{0x0003CF, 0x0003D7},
+{0x0003D8, 0x0003D9},
+{0x0003DA, 0x0003DB},
+{0x0003DC, 0x0003DD},
+{0x0003DE, 0x0003DF},
+{0x0003E0, 0x0003E1},
+{0x0003E2, 0x0003E3},
+{0x0003E4, 0x0003E5},
+{0x0003E6, 0x0003E7},
+{0x0003E8, 0x0003E9},
+{0x0003EA, 0x0003EB},
+{0x0003EC, 0x0003ED},
+{0x0003EE, 0x0003EF},
+{0x0003F4, 0x0003B8},
+{0x0003F7, 0x0003F8},
+{0x0003F9, 0x0003F2},
+{0x0003FA, 0x0003FB},
+{0x0003FD, 0x00037B},
+{0x0003FE, 0x00037C},
+{0x0003FF, 0x00037D},
+{0x000400, 0x000450},
+{0x000401, 0x000451},
+{0x000402, 0x000452},
+{0x000403, 0x000453},
+{0x000404, 0x000454},
+{0x000405, 0x000455},
+{0x000406, 0x000456},
+{0x000407, 0x000457},
+{0x000408, 0x000458},
+{0x000409, 0x000459},
+{0x00040A, 0x00045A},
+{0x00040B, 0x00045B},
+{0x00040C, 0x00045C},
+{0x00040D, 0x00045D},
+{0x00040E, 0x00045E},
+{0x00040F, 0x00045F},
+{0x000410, 0x000430},
+{0x000411, 0x000431},
+{0x000412, 0x000432},
+{0x000413, 0x000433},
+{0x000414, 0x000434},
+{0x000415, 0x000435},
+{0x000416, 0x000436},
+{0x000417, 0x000437},
+{0x000418, 0x000438},
+{0x000419, 0x000439},
+{0x00041A, 0x00043A},
+{0x00041B, 0x00043B},
+{0x00041C, 0x00043C},
+{0x00041D, 0x00043D},
+{0x00041E, 0x00043E},
+{0x00041F, 0x00043F},
+{0x000420, 0x000440},
+{0x000421, 0x000441},
+{0x000422, 0x000442},
+{0x000423, 0x000443},
+{0x000424, 0x000444},
+{0x000425, 0x000445},
+{0x000426, 0x000446},
+{0x000427, 0x000447},
+{0x000428, 0x000448},
+{0x000429, 0x000449},
+{0x00042A, 0x00044A},
+{0x00042B, 0x00044B},
+{0x00042C, 0x00044C},
+{0x00042D, 0x00044D},
+{0x00042E, 0x00044E},
+{0x00042F, 0x00044F},
+{0x000460, 0x000461},
+{0x000462, 0x000463},
+{0x000464, 0x000465},
+{0x000466, 0x000467},
+{0x000468, 0x000469},
+{0x00046A, 0x00046B},
+{0x00046C, 0x00046D},
+{0x00046E, 0x00046F},
+{0x000470, 0x000471},
+{0x000472, 0x000473},
+{0x000474, 0x000475},
+{0x000476, 0x000477},
+{0x000478, 0x000479},
+{0x00047A, 0x00047B},
+{0x00047C, 0x00047D},
+{0x00047E, 0x00047F},
+{0x000480, 0x000481},
+{0x00048A, 0x00048B},
+{0x00048C, 0x00048D},
+{0x00048E, 0x00048F},
+{0x000490, 0x000491},
+{0x000492, 0x000493},
+{0x000494, 0x000495},
+{0x000496, 0x000497},
+{0x000498, 0x000499},
+{0x00049A, 0x00049B},
+{0x00049C, 0x00049D},
+{0x00049E, 0x00049F},
+{0x0004A0, 0x0004A1},
+{0x0004A2, 0x0004A3},
+{0x0004A4, 0x0004A5},
+{0x0004A6, 0x0004A7},
+{0x0004A8, 0x0004A9},
+{0x0004AA, 0x0004AB},
+{0x0004AC, 0x0004AD},
+{0x0004AE, 0x0004AF},
+{0x0004B0, 0x0004B1},
+{0x0004B2, 0x0004B3},
+{0x0004B4, 0x0004B5},
+{0x0004B6, 0x0004B7},
+{0x0004B8, 0x0004B9},
+{0x0004BA, 0x0004BB},
+{0x0004BC, 0x0004BD},
+{0x0004BE, 0x0004BF},
+{0x0004C0, 0x0004CF},
+{0x0004C1, 0x0004C2},
+{0x0004C3, 0x0004C4},
+{0x0004C5, 0x0004C6},
+{0x0004C7, 0x0004C8},
+{0x0004C9, 0x0004CA},
+{0x0004CB, 0x0004CC},
+{0x0004CD, 0x0004CE},
+{0x0004D0, 0x0004D1},
+{0x0004D2, 0x0004D3},
+{0x0004D4, 0x0004D5},
+{0x0004D6, 0x0004D7},
+{0x0004D8, 0x0004D9},
+{0x0004DA, 0x0004DB},
+{0x0004DC, 0x0004DD},
+{0x0004DE, 0x0004DF},
+{0x0004E0, 0x0004E1},
+{0x0004E2, 0x0004E3},
+{0x0004E4, 0x0004E5},
+{0x0004E6, 0x0004E7},
+{0x0004E8, 0x0004E9},
+{0x0004EA, 0x0004EB},
+{0x0004EC, 0x0004ED},
+{0x0004EE, 0x0004EF},
+{0x0004F0, 0x0004F1},
+{0x0004F2, 0x0004F3},
+{0x0004F4, 0x0004F5},
+{0x0004F6, 0x0004F7},
+{0x0004F8, 0x0004F9},
+{0x0004FA, 0x0004FB},
+{0x0004FC, 0x0004FD},
+{0x0004FE, 0x0004FF},
+{0x000500, 0x000501},
+{0x000502, 0x000503},
+{0x000504, 0x000505},
+{0x000506, 0x000507},
+{0x000508, 0x000509},
+{0x00050A, 0x00050B},
+{0x00050C, 0x00050D},
+{0x00050E, 0x00050F},
+{0x000510, 0x000511},
+{0x000512, 0x000513},
+{0x000514, 0x000515},
+{0x000516, 0x000517},
+{0x000518, 0x000519},
+{0x00051A, 0x00051B},
+{0x00051C, 0x00051D},
+{0x00051E, 0x00051F},
+{0x000520, 0x000521},
+{0x000522, 0x000523},
+{0x000524, 0x000525},
+{0x000526, 0x000527},
+{0x000528, 0x000529},
+{0x00052A, 0x00052B},
+{0x00052C, 0x00052D},
+{0x00052E, 0x00052F},
+{0x000531, 0x000561},
+{0x000532, 0x000562},
+{0x000533, 0x000563},
+{0x000534, 0x000564},
+{0x000535, 0x000565},
+{0x000536, 0x000566},
+{0x000537, 0x000567},
+{0x000538, 0x000568},
+{0x000539, 0x000569},
+{0x00053A, 0x00056A},
+{0x00053B, 0x00056B},
+{0x00053C, 0x00056C},
+{0x00053D, 0x00056D},
+{0x00053E, 0x00056E},
+{0x00053F, 0x00056F},
+{0x000540, 0x000570},
+{0x000541, 0x000571},
+{0x000542, 0x000572},
+{0x000543, 0x000573},
+{0x000544, 0x000574},
+{0x000545, 0x000575},
+{0x000546, 0x000576},
+{0x000547, 0x000577},
+{0x000548, 0x000578},
+{0x000549, 0x000579},
+{0x00054A, 0x00057A},
+{0x00054B, 0x00057B},
+{0x00054C, 0x00057C},
+{0x00054D, 0x00057D},
+{0x00054E, 0x00057E},
+{0x00054F, 0x00057F},
+{0x000550, 0x000580},
+{0x000551, 0x000581},
+{0x000552, 0x000582},
+{0x000553, 0x000583},
+{0x000554, 0x000584},
+{0x000555, 0x000585},
+{0x000556, 0x000586},
+{0x0010A0, 0x002D00},
+{0x0010A1, 0x002D01},
+{0x0010A2, 0x002D02},
+{0x0010A3, 0x002D03},
+{0x0010A4, 0x002D04},
+{0x0010A5, 0x002D05},
+{0x0010A6, 0x002D06},
+{0x0010A7, 0x002D07},
+{0x0010A8, 0x002D08},
+{0x0010A9, 0x002D09},
+{0x0010AA, 0x002D0A},
+{0x0010AB, 0x002D0B},
+{0x0010AC, 0x002D0C},
+{0x0010AD, 0x002D0D},
+{0x0010AE, 0x002D0E},
+{0x0010AF, 0x002D0F},
+{0x0010B0, 0x002D10},
+{0x0010B1, 0x002D11},
+{0x0010B2, 0x002D12},
+{0x0010B3, 0x002D13},
+{0x0010B4, 0x002D14},
+{0x0010B5, 0x002D15},
+{0x0010B6, 0x002D16},
+{0x0010B7, 0x002D17},
+{0x0010B8, 0x002D18},
+{0x0010B9, 0x002D19},
+{0x0010BA, 0x002D1A},
+{0x0010BB, 0x002D1B},
+{0x0010BC, 0x002D1C},
+{0x0010BD, 0x002D1D},
+{0x0010BE, 0x002D1E},
+{0x0010BF, 0x002D1F},
+{0x0010C0, 0x002D20},
+{0x0010C1, 0x002D21},
+{0x0010C2, 0x002D22},
+{0x0010C3, 0x002D23},
+{0x0010C4, 0x002D24},
+{0x0010C5, 0x002D25},
+{0x0010C7, 0x002D27},
+{0x0010CD, 0x002D2D},
+{0x0013A0, 0x00AB70},
+{0x0013A1, 0x00AB71},
+{0x0013A2, 0x00AB72},
+{0x0013A3, 0x00AB73},
+{0x0013A4, 0x00AB74},
+{0x0013A5, 0x00AB75},
+{0x0013A6, 0x00AB76},
+{0x0013A7, 0x00AB77},
+{0x0013A8, 0x00AB78},
+{0x0013A9, 0x00AB79},
+{0x0013AA, 0x00AB7A},
+{0x0013AB, 0x00AB7B},
+{0x0013AC, 0x00AB7C},
+{0x0013AD, 0x00AB7D},
+{0x0013AE, 0x00AB7E},
+{0x0013AF, 0x00AB7F},
+{0x0013B0, 0x00AB80},
+{0x0013B1, 0x00AB81},
+{0x0013B2, 0x00AB82},
+{0x0013B3, 0x00AB83},
+{0x0013B4, 0x00AB84},
+{0x0013B5, 0x00AB85},
+{0x0013B6, 0x00AB86},
+{0x0013B7, 0x00AB87},
+{0x0013B8, 0x00AB88},
+{0x0013B9, 0x00AB89},
+{0x0013BA, 0x00AB8A},
+{0x0013BB, 0x00AB8B},
+{0x0013BC, 0x00AB8C},
+{0x0013BD, 0x00AB8D},
+{0x0013BE, 0x00AB8E},
+{0x0013BF, 0x00AB8F},
+{0x0013C0, 0x00AB90},
+{0x0013C1, 0x00AB91},
+{0x0013C2, 0x00AB92},
+{0x0013C3, 0x00AB93},
+{0x0013C4, 0x00AB94},
+{0x0013C5, 0x00AB95},
+{0x0013C6, 0x00AB96},
+{0x0013C7, 0x00AB97},
+{0x0013C8, 0x00AB98},
+{0x0013C9, 0x00AB99},
+{0x0013CA, 0x00AB9A},
+{0x0013CB, 0x00AB9B},
+{0x0013CC, 0x00AB9C},
+{0x0013CD, 0x00AB9D},
+{0x0013CE, 0x00AB9E},
+{0x0013CF, 0x00AB9F},
+{0x0013D0, 0x00ABA0},
+{0x0013D1, 0x00ABA1},
+{0x0013D2, 0x00ABA2},
+{0x0013D3, 0x00ABA3},
+{0x0013D4, 0x00ABA4},
+{0x0013D5, 0x00ABA5},
+{0x0013D6, 0x00ABA6},
+{0x0013D7, 0x00ABA7},
+{0x0013D8, 0x00ABA8},
+{0x0013D9, 0x00ABA9},
+{0x0013DA, 0x00ABAA},
+{0x0013DB, 0x00ABAB},
+{0x0013DC, 0x00ABAC},
+{0x0013DD, 0x00ABAD},
+{0x0013DE, 0x00ABAE},
+{0x0013DF, 0x00ABAF},
+{0x0013E0, 0x00ABB0},
+{0x0013E1, 0x00ABB1},
+{0x0013E2, 0x00ABB2},
+{0x0013E3, 0x00ABB3},
+{0x0013E4, 0x00ABB4},
+{0x0013E5, 0x00ABB5},
+{0x0013E6, 0x00ABB6},
+{0x0013E7, 0x00ABB7},
+{0x0013E8, 0x00ABB8},
+{0x0013E9, 0x00ABB9},
+{0x0013EA, 0x00ABBA},
+{0x0013EB, 0x00ABBB},
+{0x0013EC, 0x00ABBC},
+{0x0013ED, 0x00ABBD},
+{0x0013EE, 0x00ABBE},
+{0x0013EF, 0x00ABBF},
+{0x0013F0, 0x0013F8},
+{0x0013F1, 0x0013F9},
+{0x0013F2, 0x0013FA},
+{0x0013F3, 0x0013FB},
+{0x0013F4, 0x0013FC},
+{0x0013F5, 0x0013FD},
+{0x001C90, 0x0010D0},
+{0x001C91, 0x0010D1},
+{0x001C92, 0x0010D2},
+{0x001C93, 0x0010D3},
+{0x001C94, 0x0010D4},
+{0x001C95, 0x0010D5},
+{0x001C96, 0x0010D6},
+{0x001C97, 0x0010D7},
+{0x001C98, 0x0010D8},
+{0x001C99, 0x0010D9},
+{0x001C9A, 0x0010DA},
+{0x001C9B, 0x0010DB},
+{0x001C9C, 0x0010DC},
+{0x001C9D, 0x0010DD},
+{0x001C9E, 0x0010DE},
+{0x001C9F, 0x0010DF},
+{0x001CA0, 0x0010E0},
+{0x001CA1, 0x0010E1},
+{0x001CA2, 0x0010E2},
+{0x001CA3, 0x0010E3},
+{0x001CA4, 0x0010E4},
+{0x001CA5, 0x0010E5},
+{0x001CA6, 0x0010E6},
+{0x001CA7, 0x0010E7},
+{0x001CA8, 0x0010E8},
+{0x001CA9, 0x0010E9},
+{0x001CAA, 0x0010EA},
+{0x001CAB, 0x0010EB},
+{0x001CAC, 0x0010EC},
+{0x001CAD, 0x0010ED},
+{0x001CAE, 0x0010EE},
+{0x001CAF, 0x0010EF},
+{0x001CB0, 0x0010F0},
+{0x001CB1, 0x0010F1},
+{0x001CB2, 0x0010F2},
+{0x001CB3, 0x0010F3},
+{0x001CB4, 0x0010F4},
+{0x001CB5, 0x0010F5},
+{0x001CB6, 0x0010F6},
+{0x001CB7, 0x0010F7},
+{0x001CB8, 0x0010F8},
+{0x001CB9, 0x0010F9},
+{0x001CBA, 0x0010FA},
+{0x001CBD, 0x0010FD},
+{0x001CBE, 0x0010FE},
+{0x001CBF, 0x0010FF},
+{0x001E00, 0x001E01},
+{0x001E02, 0x001E03},
+{0x001E04, 0x001E05},
+{0x001E06, 0x001E07},
+{0x001E08, 0x001E09},
+{0x001E0A, 0x001E0B},
+{0x001E0C, 0x001E0D},
+{0x001E0E, 0x001E0F},
+{0x001E10, 0x001E11},
+{0x001E12, 0x001E13},
+{0x001E14, 0x001E15},
+{0x001E16, 0x001E17},
+{0x001E18, 0x001E19},
+{0x001E1A, 0x001E1B},
+{0x001E1C, 0x001E1D},
+{0x001E1E, 0x001E1F},
+{0x001E20, 0x001E21},
+{0x001E22, 0x001E23},
+{0x001E24, 0x001E25},
+{0x001E26, 0x001E27},
+{0x001E28, 0x001E29},
+{0x001E2A, 0x001E2B},
+{0x001E2C, 0x001E2D},
+{0x001E2E, 0x001E2F},
+{0x001E30, 0x001E31},
+{0x001E32, 0x001E33},
+{0x001E34, 0x001E35},
+{0x001E36, 0x001E37},
+{0x001E38, 0x001E39},
+{0x001E3A, 0x001E3B},
+{0x001E3C, 0x001E3D},
+{0x001E3E, 0x001E3F},
+{0x001E40, 0x001E41},
+{0x001E42, 0x001E43},
+{0x001E44, 0x001E45},
+{0x001E46, 0x001E47},
+{0x001E48, 0x001E49},
+{0x001E4A, 0x001E4B},
+{0x001E4C, 0x001E4D},
+{0x001E4E, 0x001E4F},
+{0x001E50, 0x001E51},
+{0x001E52, 0x001E53},
+{0x001E54, 0x001E55},
+{0x001E56, 0x001E57},
+{0x001E58, 0x001E59},
+{0x001E5A, 0x001E5B},
+{0x001E5C, 0x001E5D},
+{0x001E5E, 0x001E5F},
+{0x001E60, 0x001E61},
+{0x001E62, 0x001E63},
+{0x001E64, 0x001E65},
+{0x001E66, 0x001E67},
+{0x001E68, 0x001E69},
+{0x001E6A, 0x001E6B},
+{0x001E6C, 0x001E6D},
+{0x001E6E, 0x001E6F},
+{0x001E70, 0x001E71},
+{0x001E72, 0x001E73},
+{0x001E74, 0x001E75},
+{0x001E76, 0x001E77},
+{0x001E78, 0x001E79},
+{0x001E7A, 0x001E7B},
+{0x001E7C, 0x001E7D},
+{0x001E7E, 0x001E7F},
+{0x001E80, 0x001E81},
+{0x001E82, 0x001E83},
+{0x001E84, 0x001E85},
+{0x001E86, 0x001E87},
+{0x001E88, 0x001E89},
+{0x001E8A, 0x001E8B},
+{0x001E8C, 0x001E8D},
+{0x001E8E, 0x001E8F},
+{0x001E90, 0x001E91},
+{0x001E92, 0x001E93},
+{0x001E94, 0x001E95},
+{0x001E9E, 0x0000DF},
+{0x001EA0, 0x001EA1},
+{0x001EA2, 0x001EA3},
+{0x001EA4, 0x001EA5},
+{0x001EA6, 0x001EA7},
+{0x001EA8, 0x001EA9},
+{0x001EAA, 0x001EAB},
+{0x001EAC, 0x001EAD},
+{0x001EAE, 0x001EAF},
+{0x001EB0, 0x001EB1},
+{0x001EB2, 0x001EB3},
+{0x001EB4, 0x001EB5},
+{0x001EB6, 0x001EB7},
+{0x001EB8, 0x001EB9},
+{0x001EBA, 0x001EBB},
+{0x001EBC, 0x001EBD},
+{0x001EBE, 0x001EBF},
+{0x001EC0, 0x001EC1},
+{0x001EC2, 0x001EC3},
+{0x001EC4, 0x001EC5},
+{0x001EC6, 0x001EC7},
+{0x001EC8, 0x001EC9},
+{0x001ECA, 0x001ECB},
+{0x001ECC, 0x001ECD},
+{0x001ECE, 0x001ECF},
+{0x001ED0, 0x001ED1},
+{0x001ED2, 0x001ED3},
+{0x001ED4, 0x001ED5},
+{0x001ED6, 0x001ED7},
+{0x001ED8, 0x001ED9},
+{0x001EDA, 0x001EDB},
+{0x001EDC, 0x001EDD},
+{0x001EDE, 0x001EDF},
+{0x001EE0, 0x001EE1},
+{0x001EE2, 0x001EE3},
+{0x001EE4, 0x001EE5},
+{0x001EE6, 0x001EE7},
+{0x001EE8, 0x001EE9},
+{0x001EEA, 0x001EEB},
+{0x001EEC, 0x001EED},
+{0x001EEE, 0x001EEF},
+{0x001EF0, 0x001EF1},
+{0x001EF2, 0x001EF3},
+{0x001EF4, 0x001EF5},
+{0x001EF6, 0x001EF7},
+{0x001EF8, 0x001EF9},
+{0x001EFA, 0x001EFB},
+{0x001EFC, 0x001EFD},
+{0x001EFE, 0x001EFF},
+{0x001F08, 0x001F00},
+{0x001F09, 0x001F01},
+{0x001F0A, 0x001F02},
+{0x001F0B, 0x001F03},
+{0x001F0C, 0x001F04},
+{0x001F0D, 0x001F05},
+{0x001F0E, 0x001F06},
+{0x001F0F, 0x001F07},
+{0x001F18, 0x001F10},
+{0x001F19, 0x001F11},
+{0x001F1A, 0x001F12},
+{0x001F1B, 0x001F13},
+{0x001F1C, 0x001F14},
+{0x001F1D, 0x001F15},
+{0x001F28, 0x001F20},
+{0x001F29, 0x001F21},
+{0x001F2A, 0x001F22},
+{0x001F2B, 0x001F23},
+{0x001F2C, 0x001F24},
+{0x001F2D, 0x001F25},
+{0x001F2E, 0x001F26},
+{0x001F2F, 0x001F27},
+{0x001F38, 0x001F30},
+{0x001F39, 0x001F31},
+{0x001F3A, 0x001F32},
+{0x001F3B, 0x001F33},
+{0x001F3C, 0x001F34},
+{0x001F3D, 0x001F35},
+{0x001F3E, 0x001F36},
+{0x001F3F, 0x001F37},
+{0x001F48, 0x001F40},
+{0x001F49, 0x001F41},
+{0x001F4A, 0x001F42},
+{0x001F4B, 0x001F43},
+{0x001F4C, 0x001F44},
+{0x001F4D, 0x001F45},
+{0x001F59, 0x001F51},
+{0x001F5B, 0x001F53},
+{0x001F5D, 0x001F55},
+{0x001F5F, 0x001F57},
+{0x001F68, 0x001F60},
+{0x001F69, 0x001F61},
+{0x001F6A, 0x001F62},
+{0x001F6B, 0x001F63},
+{0x001F6C, 0x001F64},
+{0x001F6D, 0x001F65},
+{0x001F6E, 0x001F66},
+{0x001F6F, 0x001F67},
+{0x001F88, 0x001F80},
+{0x001F89, 0x001F81},
+{0x001F8A, 0x001F82},
+{0x001F8B, 0x001F83},
+{0x001F8C, 0x001F84},
+{0x001F8D, 0x001F85},
+{0x001F8E, 0x001F86},
+{0x001F8F, 0x001F87},
+{0x001F98, 0x001F90},
+{0x001F99, 0x001F91},
+{0x001F9A, 0x001F92},
+{0x001F9B, 0x001F93},
+{0x001F9C, 0x001F94},
+{0x001F9D, 0x001F95},
+{0x001F9E, 0x001F96},
+{0x001F9F, 0x001F97},
+{0x001FA8, 0x001FA0},
+{0x001FA9, 0x001FA1},
+{0x001FAA, 0x001FA2},
+{0x001FAB, 0x001FA3},
+{0x001FAC, 0x001FA4},
+{0x001FAD, 0x001FA5},
+{0x001FAE, 0x001FA6},
+{0x001FAF, 0x001FA7},
+{0x001FB8, 0x001FB0},
+{0x001FB9, 0x001FB1},
+{0x001FBA, 0x001F70},
+{0x001FBB, 0x001F71},
+{0x001FBC, 0x001FB3},
+{0x001FC8, 0x001F72},
+{0x001FC9, 0x001F73},
+{0x001FCA, 0x001F74},
+{0x001FCB, 0x001F75},
+{0x001FCC, 0x001FC3},
+{0x001FD8, 0x001FD0},
+{0x001FD9, 0x001FD1},
+{0x001FDA, 0x001F76},
+{0x001FDB, 0x001F77},
+{0x001FE8, 0x001FE0},
+{0x001FE9, 0x001FE1},
+{0x001FEA, 0x001F7A},
+{0x001FEB, 0x001F7B},
+{0x001FEC, 0x001FE5},
+{0x001FF8, 0x001F78},
+{0x001FF9, 0x001F79},
+{0x001FFA, 0x001F7C},
+{0x001FFB, 0x001F7D},
+{0x001FFC, 0x001FF3},
+{0x002126, 0x0003C9},
+{0x00212A, 0x00006B},
+{0x00212B, 0x0000E5},
+{0x002132, 0x00214E},
+{0x002160, 0x002170},
+{0x002161, 0x002171},
+{0x002162, 0x002172},
+{0x002163, 0x002173},
+{0x002164, 0x002174},
+{0x002165, 0x002175},
+{0x002166, 0x002176},
+{0x002167, 0x002177},
+{0x002168, 0x002178},
+{0x002169, 0x002179},
+{0x00216A, 0x00217A},
+{0x00216B, 0x00217B},
+{0x00216C, 0x00217C},
+{0x00216D, 0x00217D},
+{0x00216E, 0x00217E},
+{0x00216F, 0x00217F},
+{0x002183, 0x002184},
+{0x0024B6, 0x0024D0},
+{0x0024B7, 0x0024D1},
+{0x0024B8, 0x0024D2},
+{0x0024B9, 0x0024D3},
+{0x0024BA, 0x0024D4},
+{0x0024BB, 0x0024D5},
+{0x0024BC, 0x0024D6},
+{0x0024BD, 0x0024D7},
+{0x0024BE, 0x0024D8},
+{0x0024BF, 0x0024D9},
+{0x0024C0, 0x0024DA},
+{0x0024C1, 0x0024DB},
+{0x0024C2, 0x0024DC},
+{0x0024C3, 0x0024DD},
+{0x0024C4, 0x0024DE},
+{0x0024C5, 0x0024DF},
+{0x0024C6, 0x0024E0},
+{0x0024C7, 0x0024E1},
+{0x0024C8, 0x0024E2},
+{0x0024C9, 0x0024E3},
+{0x0024CA, 0x0024E4},
+{0x0024CB, 0x0024E5},
+{0x0024CC, 0x0024E6},
+{0x0024CD, 0x0024E7},
+{0x0024CE, 0x0024E8},
+{0x0024CF, 0x0024E9},
+{0x002C00, 0x002C30},
+{0x002C01, 0x002C31},
+{0x002C02, 0x002C32},
+{0x002C03, 0x002C33},
+{0x002C04, 0x002C34},
+{0x002C05, 0x002C35},
+{0x002C06, 0x002C36},
+{0x002C07, 0x002C37},
+{0x002C08, 0x002C38},
+{0x002C09, 0x002C39},
+{0x002C0A, 0x002C3A},
+{0x002C0B, 0x002C3B},
+{0x002C0C, 0x002C3C},
+{0x002C0D, 0x002C3D},
+{0x002C0E, 0x002C3E},
+{0x002C0F, 0x002C3F},
+{0x002C10, 0x002C40},
+{0x002C11, 0x002C41},
+{0x002C12, 0x002C42},
+{0x002C13, 0x002C43},
+{0x002C14, 0x002C44},
+{0x002C15, 0x002C45},
+{0x002C16, 0x002C46},
+{0x002C17, 0x002C47},
+{0x002C18, 0x002C48},
+{0x002C19, 0x002C49},
+{0x002C1A, 0x002C4A},
+{0x002C1B, 0x002C4B},
+{0x002C1C, 0x002C4C},
+{0x002C1D, 0x002C4D},
+{0x002C1E, 0x002C4E},
+{0x002C1F, 0x002C4F},
+{0x002C20, 0x002C50},
+{0x002C21, 0x002C51},
+{0x002C22, 0x002C52},
+{0x002C23, 0x002C53},
+{0x002C24, 0x002C54},
+{0x002C25, 0x002C55},
+{0x002C26, 0x002C56},
+{0x002C27, 0x002C57},
+{0x002C28, 0x002C58},
+{0x002C29, 0x002C59},
+{0x002C2A, 0x002C5A},
+{0x002C2B, 0x002C5B},
+{0x002C2C, 0x002C5C},
+{0x002C2D, 0x002C5D},
+{0x002C2E, 0x002C5E},
+{0x002C60, 0x002C61},
+{0x002C62, 0x00026B},
+{0x002C63, 0x001D7D},
+{0x002C64, 0x00027D},
+{0x002C67, 0x002C68},
+{0x002C69, 0x002C6A},
+{0x002C6B, 0x002C6C},
+{0x002C6D, 0x000251},
+{0x002C6E, 0x000271},
+{0x002C6F, 0x000250},
+{0x002C70, 0x000252},
+{0x002C72, 0x002C73},
+{0x002C75, 0x002C76},
+{0x002C7E, 0x00023F},
+{0x002C7F, 0x000240},
+{0x002C80, 0x002C81},
+{0x002C82, 0x002C83},
+{0x002C84, 0x002C85},
+{0x002C86, 0x002C87},
+{0x002C88, 0x002C89},
+{0x002C8A, 0x002C8B},
+{0x002C8C, 0x002C8D},
+{0x002C8E, 0x002C8F},
+{0x002C90, 0x002C91},
+{0x002C92, 0x002C93},
+{0x002C94, 0x002C95},
+{0x002C96, 0x002C97},
+{0x002C98, 0x002C99},
+{0x002C9A, 0x002C9B},
+{0x002C9C, 0x002C9D},
+{0x002C9E, 0x002C9F},
+{0x002CA0, 0x002CA1},
+{0x002CA2, 0x002CA3},
+{0x002CA4, 0x002CA5},
+{0x002CA6, 0x002CA7},
+{0x002CA8, 0x002CA9},
+{0x002CAA, 0x002CAB},
+{0x002CAC, 0x002CAD},
+{0x002CAE, 0x002CAF},
+{0x002CB0, 0x002CB1},
+{0x002CB2, 0x002CB3},
+{0x002CB4, 0x002CB5},
+{0x002CB6, 0x002CB7},
+{0x002CB8, 0x002CB9},
+{0x002CBA, 0x002CBB},
+{0x002CBC, 0x002CBD},
+{0x002CBE, 0x002CBF},
+{0x002CC0, 0x002CC1},
+{0x002CC2, 0x002CC3},
+{0x002CC4, 0x002CC5},
+{0x002CC6, 0x002CC7},
+{0x002CC8, 0x002CC9},
+{0x002CCA, 0x002CCB},
+{0x002CCC, 0x002CCD},
+{0x002CCE, 0x002CCF},
+{0x002CD0, 0x002CD1},
+{0x002CD2, 0x002CD3},
+{0x002CD4, 0x002CD5},
+{0x002CD6, 0x002CD7},
+{0x002CD8, 0x002CD9},
+{0x002CDA, 0x002CDB},
+{0x002CDC, 0x002CDD},
+{0x002CDE, 0x002CDF},
+{0x002CE0, 0x002CE1},
+{0x002CE2, 0x002CE3},
+{0x002CEB, 0x002CEC},
+{0x002CED, 0x002CEE},
+{0x002CF2, 0x002CF3},
+{0x00A640, 0x00A641},
+{0x00A642, 0x00A643},
+{0x00A644, 0x00A645},
+{0x00A646, 0x00A647},
+{0x00A648, 0x00A649},
+{0x00A64A, 0x00A64B},
+{0x00A64C, 0x00A64D},
+{0x00A64E, 0x00A64F},
+{0x00A650, 0x00A651},
+{0x00A652, 0x00A653},
+{0x00A654, 0x00A655},
+{0x00A656, 0x00A657},
+{0x00A658, 0x00A659},
+{0x00A65A, 0x00A65B},
+{0x00A65C, 0x00A65D},
+{0x00A65E, 0x00A65F},
+{0x00A660, 0x00A661},
+{0x00A662, 0x00A663},
+{0x00A664, 0x00A665},
+{0x00A666, 0x00A667},
+{0x00A668, 0x00A669},
+{0x00A66A, 0x00A66B},
+{0x00A66C, 0x00A66D},
+{0x00A680, 0x00A681},
+{0x00A682, 0x00A683},
+{0x00A684, 0x00A685},
+{0x00A686, 0x00A687},
+{0x00A688, 0x00A689},
+{0x00A68A, 0x00A68B},
+{0x00A68C, 0x00A68D},
+{0x00A68E, 0x00A68F},
+{0x00A690, 0x00A691},
+{0x00A692, 0x00A693},
+{0x00A694, 0x00A695},
+{0x00A696, 0x00A697},
+{0x00A698, 0x00A699},
+{0x00A69A, 0x00A69B},
+{0x00A722, 0x00A723},
+{0x00A724, 0x00A725},
+{0x00A726, 0x00A727},
+{0x00A728, 0x00A729},
+{0x00A72A, 0x00A72B},
+{0x00A72C, 0x00A72D},
+{0x00A72E, 0x00A72F},
+{0x00A732, 0x00A733},
+{0x00A734, 0x00A735},
+{0x00A736, 0x00A737},
+{0x00A738, 0x00A739},
+{0x00A73A, 0x00A73B},
+{0x00A73C, 0x00A73D},
+{0x00A73E, 0x00A73F},
+{0x00A740, 0x00A741},
+{0x00A742, 0x00A743},
+{0x00A744, 0x00A745},
+{0x00A746, 0x00A747},
+{0x00A748, 0x00A749},
+{0x00A74A, 0x00A74B},
+{0x00A74C, 0x00A74D},
+{0x00A74E, 0x00A74F},
+{0x00A750, 0x00A751},
+{0x00A752, 0x00A753},
+{0x00A754, 0x00A755},
+{0x00A756, 0x00A757},
+{0x00A758, 0x00A759},
+{0x00A75A, 0x00A75B},
+{0x00A75C, 0x00A75D},
+{0x00A75E, 0x00A75F},
+{0x00A760, 0x00A761},
+{0x00A762, 0x00A763},
+{0x00A764, 0x00A765},
+{0x00A766, 0x00A767},
+{0x00A768, 0x00A769},
+{0x00A76A, 0x00A76B},
+{0x00A76C, 0x00A76D},
+{0x00A76E, 0x00A76F},
+{0x00A779, 0x00A77A},
+{0x00A77B, 0x00A77C},
+{0x00A77D, 0x001D79},
+{0x00A77E, 0x00A77F},
+{0x00A780, 0x00A781},
+{0x00A782, 0x00A783},
+{0x00A784, 0x00A785},
+{0x00A786, 0x00A787},
+{0x00A78B, 0x00A78C},
+{0x00A78D, 0x000265},
+{0x00A790, 0x00A791},
+{0x00A792, 0x00A793},
+{0x00A796, 0x00A797},
+{0x00A798, 0x00A799},
+{0x00A79A, 0x00A79B},
+{0x00A79C, 0x00A79D},
+{0x00A79E, 0x00A79F},
+{0x00A7A0, 0x00A7A1},
+{0x00A7A2, 0x00A7A3},
+{0x00A7A4, 0x00A7A5},
+{0x00A7A6, 0x00A7A7},
+{0x00A7A8, 0x00A7A9},
+{0x00A7AA, 0x000266},
+{0x00A7AB, 0x00025C},
+{0x00A7AC, 0x000261},
+{0x00A7AD, 0x00026C},
+{0x00A7AE, 0x00026A},
+{0x00A7B0, 0x00029E},
+{0x00A7B1, 0x000287},
+{0x00A7B2, 0x00029D},
+{0x00A7B3, 0x00AB53},
+{0x00A7B4, 0x00A7B5},
+{0x00A7B6, 0x00A7B7},
+{0x00A7B8, 0x00A7B9},
+{0x00A7BA, 0x00A7BB},
+{0x00A7BC, 0x00A7BD},
+{0x00A7BE, 0x00A7BF},
+{0x00A7C2, 0x00A7C3},
+{0x00A7C4, 0x00A794},
+{0x00A7C5, 0x000282},
+{0x00A7C6, 0x001D8E},
+{0x00A7C7, 0x00A7C8},
+{0x00A7C9, 0x00A7CA},
+{0x00A7F5, 0x00A7F6},
+{0x00FF21, 0x00FF41},
+{0x00FF22, 0x00FF42},
+{0x00FF23, 0x00FF43},
+{0x00FF24, 0x00FF44},
+{0x00FF25, 0x00FF45},
+{0x00FF26, 0x00FF46},
+{0x00FF27, 0x00FF47},
+{0x00FF28, 0x00FF48},
+{0x00FF29, 0x00FF49},
+{0x00FF2A, 0x00FF4A},
+{0x00FF2B, 0x00FF4B},
+{0x00FF2C, 0x00FF4C},
+{0x00FF2D, 0x00FF4D},
+{0x00FF2E, 0x00FF4E},
+{0x00FF2F, 0x00FF4F},
+{0x00FF30, 0x00FF50},
+{0x00FF31, 0x00FF51},
+{0x00FF32, 0x00FF52},
+{0x00FF33, 0x00FF53},
+{0x00FF34, 0x00FF54},
+{0x00FF35, 0x00FF55},
+{0x00FF36, 0x00FF56},
+{0x00FF37, 0x00FF57},
+{0x00FF38, 0x00FF58},
+{0x00FF39, 0x00FF59},
+{0x00FF3A, 0x00FF5A},
+{0x010400, 0x010428},
+{0x010401, 0x010429},
+{0x010402, 0x01042A},
+{0x010403, 0x01042B},
+{0x010404, 0x01042C},
+{0x010405, 0x01042D},
+{0x010406, 0x01042E},
+{0x010407, 0x01042F},
+{0x010408, 0x010430},
+{0x010409, 0x010431},
+{0x01040A, 0x010432},
+{0x01040B, 0x010433},
+{0x01040C, 0x010434},
+{0x01040D, 0x010435},
+{0x01040E, 0x010436},
+{0x01040F, 0x010437},
+{0x010410, 0x010438},
+{0x010411, 0x010439},
+{0x010412, 0x01043A},
+{0x010413, 0x01043B},
+{0x010414, 0x01043C},
+{0x010415, 0x01043D},
+{0x010416, 0x01043E},
+{0x010417, 0x01043F},
+{0x010418, 0x010440},
+{0x010419, 0x010441},
+{0x01041A, 0x010442},
+{0x01041B, 0x010443},
+{0x01041C, 0x010444},
+{0x01041D, 0x010445},
+{0x01041E, 0x010446},
+{0x01041F, 0x010447},
+{0x010420, 0x010448},
+{0x010421, 0x010449},
+{0x010422, 0x01044A},
+{0x010423, 0x01044B},
+{0x010424, 0x01044C},
+{0x010425, 0x01044D},
+{0x010426, 0x01044E},
+{0x010427, 0x01044F},
+{0x0104B0, 0x0104D8},
+{0x0104B1, 0x0104D9},
+{0x0104B2, 0x0104DA},
+{0x0104B3, 0x0104DB},
+{0x0104B4, 0x0104DC},
+{0x0104B5, 0x0104DD},
+{0x0104B6, 0x0104DE},
+{0x0104B7, 0x0104DF},
+{0x0104B8, 0x0104E0},
+{0x0104B9, 0x0104E1},
+{0x0104BA, 0x0104E2},
+{0x0104BB, 0x0104E3},
+{0x0104BC, 0x0104E4},
+{0x0104BD, 0x0104E5},
+{0x0104BE, 0x0104E6},
+{0x0104BF, 0x0104E7},
+{0x0104C0, 0x0104E8},
+{0x0104C1, 0x0104E9},
+{0x0104C2, 0x0104EA},
+{0x0104C3, 0x0104EB},
+{0x0104C4, 0x0104EC},
+{0x0104C5, 0x0104ED},
+{0x0104C6, 0x0104EE},
+{0x0104C7, 0x0104EF},
+{0x0104C8, 0x0104F0},
+{0x0104C9, 0x0104F1},
+{0x0104CA, 0x0104F2},
+{0x0104CB, 0x0104F3},
+{0x0104CC, 0x0104F4},
+{0x0104CD, 0x0104F5},
+{0x0104CE, 0x0104F6},
+{0x0104CF, 0x0104F7},
+{0x0104D0, 0x0104F8},
+{0x0104D1, 0x0104F9},
+{0x0104D2, 0x0104FA},
+{0x0104D3, 0x0104FB},
+{0x010C80, 0x010CC0},
+{0x010C81, 0x010CC1},
+{0x010C82, 0x010CC2},
+{0x010C83, 0x010CC3},
+{0x010C84, 0x010CC4},
+{0x010C85, 0x010CC5},
+{0x010C86, 0x010CC6},
+{0x010C87, 0x010CC7},
+{0x010C88, 0x010CC8},
+{0x010C89, 0x010CC9},
+{0x010C8A, 0x010CCA},
+{0x010C8B, 0x010CCB},
+{0x010C8C, 0x010CCC},
+{0x010C8D, 0x010CCD},
+{0x010C8E, 0x010CCE},
+{0x010C8F, 0x010CCF},
+{0x010C90, 0x010CD0},
+{0x010C91, 0x010CD1},
+{0x010C92, 0x010CD2},
+{0x010C93, 0x010CD3},
+{0x010C94, 0x010CD4},
+{0x010C95, 0x010CD5},
+{0x010C96, 0x010CD6},
+{0x010C97, 0x010CD7},
+{0x010C98, 0x010CD8},
+{0x010C99, 0x010CD9},
+{0x010C9A, 0x010CDA},
+{0x010C9B, 0x010CDB},
+{0x010C9C, 0x010CDC},
+{0x010C9D, 0x010CDD},
+{0x010C9E, 0x010CDE},
+{0x010C9F, 0x010CDF},
+{0x010CA0, 0x010CE0},
+{0x010CA1, 0x010CE1},
+{0x010CA2, 0x010CE2},
+{0x010CA3, 0x010CE3},
+{0x010CA4, 0x010CE4},
+{0x010CA5, 0x010CE5},
+{0x010CA6, 0x010CE6},
+{0x010CA7, 0x010CE7},
+{0x010CA8, 0x010CE8},
+{0x010CA9, 0x010CE9},
+{0x010CAA, 0x010CEA},
+{0x010CAB, 0x010CEB},
+{0x010CAC, 0x010CEC},
+{0x010CAD, 0x010CED},
+{0x010CAE, 0x010CEE},
+{0x010CAF, 0x010CEF},
+{0x010CB0, 0x010CF0},
+{0x010CB1, 0x010CF1},
+{0x010CB2, 0x010CF2},
+{0x0118A0, 0x0118C0},
+{0x0118A1, 0x0118C1},
+{0x0118A2, 0x0118C2},
+{0x0118A3, 0x0118C3},
+{0x0118A4, 0x0118C4},
+{0x0118A5, 0x0118C5},
+{0x0118A6, 0x0118C6},
+{0x0118A7, 0x0118C7},
+{0x0118A8, 0x0118C8},
+{0x0118A9, 0x0118C9},
+{0x0118AA, 0x0118CA},
+{0x0118AB, 0x0118CB},
+{0x0118AC, 0x0118CC},
+{0x0118AD, 0x0118CD},
+{0x0118AE, 0x0118CE},
+{0x0118AF, 0x0118CF},
+{0x0118B0, 0x0118D0},
+{0x0118B1, 0x0118D1},
+{0x0118B2, 0x0118D2},
+{0x0118B3, 0x0118D3},
+{0x0118B4, 0x0118D4},
+{0x0118B5, 0x0118D5},
+{0x0118B6, 0x0118D6},
+{0x0118B7, 0x0118D7},
+{0x0118B8, 0x0118D8},
+{0x0118B9, 0x0118D9},
+{0x0118BA, 0x0118DA},
+{0x0118BB, 0x0118DB},
+{0x0118BC, 0x0118DC},
+{0x0118BD, 0x0118DD},
+{0x0118BE, 0x0118DE},
+{0x0118BF, 0x0118DF},
+{0x016E40, 0x016E60},
+{0x016E41, 0x016E61},
+{0x016E42, 0x016E62},
+{0x016E43, 0x016E63},
+{0x016E44, 0x016E64},
+{0x016E45, 0x016E65},
+{0x016E46, 0x016E66},
+{0x016E47, 0x016E67},
+{0x016E48, 0x016E68},
+{0x016E49, 0x016E69},
+{0x016E4A, 0x016E6A},
+{0x016E4B, 0x016E6B},
+{0x016E4C, 0x016E6C},
+{0x016E4D, 0x016E6D},
+{0x016E4E, 0x016E6E},
+{0x016E4F, 0x016E6F},
+{0x016E50, 0x016E70},
+{0x016E51, 0x016E71},
+{0x016E52, 0x016E72},
+{0x016E53, 0x016E73},
+{0x016E54, 0x016E74},
+{0x016E55, 0x016E75},
+{0x016E56, 0x016E76},
+{0x016E57, 0x016E77},
+{0x016E58, 0x016E78},
+{0x016E59, 0x016E79},
+{0x016E5A, 0x016E7A},
+{0x016E5B, 0x016E7B},
+{0x016E5C, 0x016E7C},
+{0x016E5D, 0x016E7D},
+{0x016E5E, 0x016E7E},
+{0x016E5F, 0x016E7F},
+{0x01E900, 0x01E922},
+{0x01E901, 0x01E923},
+{0x01E902, 0x01E924},
+{0x01E903, 0x01E925},
+{0x01E904, 0x01E926},
+{0x01E905, 0x01E927},
+{0x01E906, 0x01E928},
+{0x01E907, 0x01E929},
+{0x01E908, 0x01E92A},
+{0x01E909, 0x01E92B},
+{0x01E90A, 0x01E92C},
+{0x01E90B, 0x01E92D},
+{0x01E90C, 0x01E92E},
+{0x01E90D, 0x01E92F},
+{0x01E90E, 0x01E930},
+{0x01E90F, 0x01E931},
+{0x01E910, 0x01E932},
+{0x01E911, 0x01E933},
+{0x01E912, 0x01E934},
+{0x01E913, 0x01E935},
+{0x01E914, 0x01E936},
+{0x01E915, 0x01E937},
+{0x01E916, 0x01E938},
+{0x01E917, 0x01E939},
+{0x01E918, 0x01E93A},
+{0x01E919, 0x01E93B},
+{0x01E91A, 0x01E93C},
+{0x01E91B, 0x01E93D},
+{0x01E91C, 0x01E93E},
+{0x01E91D, 0x01E93F},
+{0x01E91E, 0x01E940},
+{0x01E91F, 0x01E941},
+{0x01E920, 0x01E942},
+{0x01E921, 0x01E943},
 };
 
-const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control = {
-{0x00000000, 0x00000008}, {0x0000000E, 0x0000001B}, {0x0000007F, 0x00000084}, {0x00000086, 0x0000009F},
-{0x000000AD, 0x000000AD}, {0x00000378, 0x00000379}, {0x00000380, 0x00000383}, {0x0000038B, 0x0000038B},
-{0x0000038D, 0x0000038D}, {0x000003A2, 0x000003A2}, {0x00000530, 0x00000530}, {0x00000557, 0x00000558},
-{0x0000058B, 0x0000058C}, {0x00000590, 0x00000590}, {0x000005C8, 0x000005CF}, {0x000005EB, 0x000005EE},
-{0x000005F5, 0x00000605}, {0x0000061C, 0x0000061D}, {0x000006DD, 0x000006DD}, {0x0000070E, 0x0000070F},
-{0x0000074B, 0x0000074C}, {0x000007B2, 0x000007BF}, {0x000007FB, 0x000007FC}, {0x0000082E, 0x0000082F},
-{0x0000083F, 0x0000083F}, {0x0000085C, 0x0000085D}, {0x0000085F, 0x0000085F}, {0x0000086B, 0x0000089F},
-{0x000008B5, 0x000008B5}, {0x000008C8, 0x000008D2}, {0x000008E2, 0x000008E2}, {0x00000984, 0x00000984},
-{0x0000098D, 0x0000098E}, {0x00000991, 0x00000992}, {0x000009A9, 0x000009A9}, {0x000009B1, 0x000009B1},
-{0x000009B3, 0x000009B5}, {0x000009BA, 0x000009BB}, {0x000009C5, 0x000009C6}, {0x000009C9, 0x000009CA},
-{0x000009CF, 0x000009D6}, {0x000009D8, 0x000009DB}, {0x000009DE, 0x000009DE}, {0x000009E4, 0x000009E5},
-{0x000009FF, 0x00000A00}, {0x00000A04, 0x00000A04}, {0x00000A0B, 0x00000A0E}, {0x00000A11, 0x00000A12},
-{0x00000A29, 0x00000A29}, {0x00000A31, 0x00000A31}, {0x00000A34, 0x00000A34}, {0x00000A37, 0x00000A37},
-{0x00000A3A, 0x00000A3B}, {0x00000A3D, 0x00000A3D}, {0x00000A43, 0x00000A46}, {0x00000A49, 0x00000A4A},
-{0x00000A4E, 0x00000A50}, {0x00000A52, 0x00000A58}, {0x00000A5D, 0x00000A5D}, {0x00000A5F, 0x00000A65},
-{0x00000A77, 0x00000A80}, {0x00000A84, 0x00000A84}, {0x00000A8E, 0x00000A8E}, {0x00000A92, 0x00000A92},
-{0x00000AA9, 0x00000AA9}, {0x00000AB1, 0x00000AB1}, {0x00000AB4, 0x00000AB4}, {0x00000ABA, 0x00000ABB},
-{0x00000AC6, 0x00000AC6}, {0x00000ACA, 0x00000ACA}, {0x00000ACE, 0x00000ACF}, {0x00000AD1, 0x00000ADF},
-{0x00000AE4, 0x00000AE5}, {0x00000AF2, 0x00000AF8}, {0x00000B00, 0x00000B00}, {0x00000B04, 0x00000B04},
-{0x00000B0D, 0x00000B0E}, {0x00000B11, 0x00000B12}, {0x00000B29, 0x00000B29}, {0x00000B31, 0x00000B31},
-{0x00000B34, 0x00000B34}, {0x00000B3A, 0x00000B3B}, {0x00000B45, 0x00000B46}, {0x00000B49, 0x00000B4A},
-{0x00000B4E, 0x00000B54}, {0x00000B58, 0x00000B5B}, {0x00000B5E, 0x00000B5E}, {0x00000B64, 0x00000B65},
-{0x00000B78, 0x00000B81}, {0x00000B84, 0x00000B84}, {0x00000B8B, 0x00000B8D}, {0x00000B91, 0x00000B91},
-{0x00000B96, 0x00000B98}, {0x00000B9B, 0x00000B9B}, {0x00000B9D, 0x00000B9D}, {0x00000BA0, 0x00000BA2},
-{0x00000BA5, 0x00000BA7}, {0x00000BAB, 0x00000BAD}, {0x00000BBA, 0x00000BBD}, {0x00000BC3, 0x00000BC5},
-{0x00000BC9, 0x00000BC9}, {0x00000BCE, 0x00000BCF}, {0x00000BD1, 0x00000BD6}, {0x00000BD8, 0x00000BE5},
-{0x00000BFB, 0x00000BFF}, {0x00000C0D, 0x00000C0D}, {0x00000C11, 0x00000C11}, {0x00000C29, 0x00000C29},
-{0x00000C3A, 0x00000C3C}, {0x00000C45, 0x00000C45}, {0x00000C49, 0x00000C49}, {0x00000C4E, 0x00000C54},
-{0x00000C57, 0x00000C57}, {0x00000C5B, 0x00000C5F}, {0x00000C64, 0x00000C65}, {0x00000C70, 0x00000C76},
-{0x00000C8D, 0x00000C8D}, {0x00000C91, 0x00000C91}, {0x00000CA9, 0x00000CA9}, {0x00000CB4, 0x00000CB4},
-{0x00000CBA, 0x00000CBB}, {0x00000CC5, 0x00000CC5}, {0x00000CC9, 0x00000CC9}, {0x00000CCE, 0x00000CD4},
-{0x00000CD7, 0x00000CDD}, {0x00000CDF, 0x00000CDF}, {0x00000CE4, 0x00000CE5}, {0x00000CF0, 0x00000CF0},
-{0x00000CF3, 0x00000CFF}, {0x00000D0D, 0x00000D0D}, {0x00000D11, 0x00000D11}, {0x00000D45, 0x00000D45},
-{0x00000D49, 0x00000D49}, {0x00000D50, 0x00000D53}, {0x00000D64, 0x00000D65}, {0x00000D80, 0x00000D80},
-{0x00000D84, 0x00000D84}, {0x00000D97, 0x00000D99}, {0x00000DB2, 0x00000DB2}, {0x00000DBC, 0x00000DBC},
-{0x00000DBE, 0x00000DBF}, {0x00000DC7, 0x00000DC9}, {0x00000DCB, 0x00000DCE}, {0x00000DD5, 0x00000DD5},
-{0x00000DD7, 0x00000DD7}, {0x00000DE0, 0x00000DE5}, {0x00000DF0, 0x00000DF1}, {0x00000DF5, 0x00000E00},
-{0x00000E3B, 0x00000E3E}, {0x00000E5C, 0x00000E80}, {0x00000E83, 0x00000E83}, {0x00000E85, 0x00000E85},
-{0x00000E8B, 0x00000E8B}, {0x00000EA4, 0x00000EA4}, {0x00000EA6, 0x00000EA6}, {0x00000EBE, 0x00000EBF},
-{0x00000EC5, 0x00000EC5}, {0x00000EC7, 0x00000EC7}, {0x00000ECE, 0x00000ECF}, {0x00000EDA, 0x00000EDB},
-{0x00000EE0, 0x00000EFF}, {0x00000F48, 0x00000F48}, {0x00000F6D, 0x00000F70}, {0x00000F98, 0x00000F98},
-{0x00000FBD, 0x00000FBD}, {0x00000FCD, 0x00000FCD}, {0x00000FDB, 0x00000FFF}, {0x000010C6, 0x000010C6},
-{0x000010C8, 0x000010CC}, {0x000010CE, 0x000010CF}, {0x00001249, 0x00001249}, {0x0000124E, 0x0000124F},
-{0x00001257, 0x00001257}, {0x00001259, 0x00001259}, {0x0000125E, 0x0000125F}, {0x00001289, 0x00001289},
-{0x0000128E, 0x0000128F}, {0x000012B1, 0x000012B1}, {0x000012B6, 0x000012B7}, {0x000012BF, 0x000012BF},
-{0x000012C1, 0x000012C1}, {0x000012C6, 0x000012C7}, {0x000012D7, 0x000012D7}, {0x00001311, 0x00001311},
-{0x00001316, 0x00001317}, {0x0000135B, 0x0000135C}, {0x0000137D, 0x0000137F}, {0x0000139A, 0x0000139F},
-{0x000013F6, 0x000013F7}, {0x000013FE, 0x000013FF}, {0x0000169D, 0x0000169F}, {0x000016F9, 0x000016FF},
-{0x0000170D, 0x0000170D}, {0x00001715, 0x0000171F}, {0x00001737, 0x0000173F}, {0x00001754, 0x0000175F},
-{0x0000176D, 0x0000176D}, {0x00001771, 0x00001771}, {0x00001774, 0x0000177F}, {0x000017DE, 0x000017DF},
-{0x000017EA, 0x000017EF}, {0x000017FA, 0x000017FF}, {0x0000180E, 0x0000180F}, {0x0000181A, 0x0000181F},
-{0x00001879, 0x0000187F}, {0x000018AB, 0x000018AF}, {0x000018F6, 0x000018FF}, {0x0000191F, 0x0000191F},
-{0x0000192C, 0x0000192F}, {0x0000193C, 0x0000193F}, {0x00001941, 0x00001943}, {0x0000196E, 0x0000196F},
-{0x00001975, 0x0000197F}, {0x000019AC, 0x000019AF}, {0x000019CA, 0x000019CF}, {0x000019DB, 0x000019DD},
-{0x00001A1C, 0x00001A1D}, {0x00001A5F, 0x00001A5F}, {0x00001A7D, 0x00001A7E}, {0x00001A8A, 0x00001A8F},
-{0x00001A9A, 0x00001A9F}, {0x00001AAE, 0x00001AAF}, {0x00001AC1, 0x00001AFF}, {0x00001B4C, 0x00001B4F},
-{0x00001B7D, 0x00001B7F}, {0x00001BF4, 0x00001BFB}, {0x00001C38, 0x00001C3A}, {0x00001C4A, 0x00001C4C},
-{0x00001C89, 0x00001C8F}, {0x00001CBB, 0x00001CBC}, {0x00001CC8, 0x00001CCF}, {0x00001CFB, 0x00001CFF},
-{0x00001DFA, 0x00001DFA}, {0x00001F16, 0x00001F17}, {0x00001F1E, 0x00001F1F}, {0x00001F46, 0x00001F47},
-{0x00001F4E, 0x00001F4F}, {0x00001F58, 0x00001F58}, {0x00001F5A, 0x00001F5A}, {0x00001F5C, 0x00001F5C},
-{0x00001F5E, 0x00001F5E}, {0x00001F7E, 0x00001F7F}, {0x00001FB5, 0x00001FB5}, {0x00001FC5, 0x00001FC5},
-{0x00001FD4, 0x00001FD5}, {0x00001FDC, 0x00001FDC}, {0x00001FF0, 0x00001FF1}, {0x00001FF5, 0x00001FF5},
-{0x00001FFF, 0x00001FFF}, {0x0000200B, 0x0000200F}, {0x0000202A, 0x0000202E}, {0x00002060, 0x0000206F},
-{0x00002072, 0x00002073}, {0x0000208F, 0x0000208F}, {0x0000209D, 0x0000209F}, {0x000020C0, 0x000020CF},
-{0x000020F1, 0x000020FF}, {0x0000218C, 0x0000218F}, {0x00002427, 0x0000243F}, {0x0000244B, 0x0000245F},
-{0x00002B74, 0x00002B75}, {0x00002B96, 0x00002B96}, {0x00002C2F, 0x00002C2F}, {0x00002C5F, 0x00002C5F},
-{0x00002CF4, 0x00002CF8}, {0x00002D26, 0x00002D26}, {0x00002D28, 0x00002D2C}, {0x00002D2E, 0x00002D2F},
-{0x00002D68, 0x00002D6E}, {0x00002D71, 0x00002D7E}, {0x00002D97, 0x00002D9F}, {0x00002DA7, 0x00002DA7},
-{0x00002DAF, 0x00002DAF}, {0x00002DB7, 0x00002DB7}, {0x00002DBF, 0x00002DBF}, {0x00002DC7, 0x00002DC7},
-{0x00002DCF, 0x00002DCF}, {0x00002DD7, 0x00002DD7}, {0x00002DDF, 0x00002DDF}, {0x00002E53, 0x00002E7F},
-{0x00002E9A, 0x00002E9A}, {0x00002EF4, 0x00002EFF}, {0x00002FD6, 0x00002FEF}, {0x00002FFC, 0x00002FFF},
-{0x00003040, 0x00003040}, {0x00003097, 0x00003098}, {0x00003100, 0x00003104}, {0x00003130, 0x00003130},
-{0x0000318F, 0x0000318F}, {0x000031E4, 0x000031EF}, {0x0000321F, 0x0000321F}, {0x00009FFD, 0x00009FFF},
-{0x0000A48D, 0x0000A48F}, {0x0000A4C7, 0x0000A4CF}, {0x0000A62C, 0x0000A63F}, {0x0000A6F8, 0x0000A6FF},
-{0x0000A7C0, 0x0000A7C1}, {0x0000A7CB, 0x0000A7F4}, {0x0000A82D, 0x0000A82F}, {0x0000A83A, 0x0000A83F},
-{0x0000A878, 0x0000A87F}, {0x0000A8C6, 0x0000A8CD}, {0x0000A8DA, 0x0000A8DF}, {0x0000A954, 0x0000A95E},
-{0x0000A97D, 0x0000A97F}, {0x0000A9CE, 0x0000A9CE}, {0x0000A9DA, 0x0000A9DD}, {0x0000A9FF, 0x0000A9FF},
-{0x0000AA37, 0x0000AA3F}, {0x0000AA4E, 0x0000AA4F}, {0x0000AA5A, 0x0000AA5B}, {0x0000AAC3, 0x0000AADA},
-{0x0000AAF7, 0x0000AB00}, {0x0000AB07, 0x0000AB08}, {0x0000AB0F, 0x0000AB10}, {0x0000AB17, 0x0000AB1F},
-{0x0000AB27, 0x0000AB27}, {0x0000AB2F, 0x0000AB2F}, {0x0000AB6C, 0x0000AB6F}, {0x0000ABEE, 0x0000ABEF},
-{0x0000ABFA, 0x0000ABFF}, {0x0000D7A4, 0x0000D7AF}, {0x0000D7C7, 0x0000D7CA}, {0x0000D7FC, 0x0000F8FF},
-{0x0000FA6E, 0x0000FA6F}, {0x0000FADA, 0x0000FAFF}, {0x0000FB07, 0x0000FB12}, {0x0000FB18, 0x0000FB1C},
-{0x0000FB37, 0x0000FB37}, {0x0000FB3D, 0x0000FB3D}, {0x0000FB3F, 0x0000FB3F}, {0x0000FB42, 0x0000FB42},
-{0x0000FB45, 0x0000FB45}, {0x0000FBC2, 0x0000FBD2}, {0x0000FD40, 0x0000FD4F}, {0x0000FD90, 0x0000FD91},
-{0x0000FDC8, 0x0000FDEF}, {0x0000FDFE, 0x0000FDFF}, {0x0000FE1A, 0x0000FE1F}, {0x0000FE53, 0x0000FE53},
-{0x0000FE67, 0x0000FE67}, {0x0000FE6C, 0x0000FE6F}, {0x0000FE75, 0x0000FE75}, {0x0000FEFD, 0x0000FF00},
-{0x0000FFBF, 0x0000FFC1}, {0x0000FFC8, 0x0000FFC9}, {0x0000FFD0, 0x0000FFD1}, {0x0000FFD8, 0x0000FFD9},
-{0x0000FFDD, 0x0000FFDF}, {0x0000FFE7, 0x0000FFE7}, {0x0000FFEF, 0x0000FFFB}, {0x0000FFFE, 0x0000FFFF},
-{0x0001000C, 0x0001000C}, {0x00010027, 0x00010027}, {0x0001003B, 0x0001003B}, {0x0001003E, 0x0001003E},
-{0x0001004E, 0x0001004F}, {0x0001005E, 0x0001007F}, {0x000100FB, 0x000100FF}, {0x00010103, 0x00010106},
-{0x00010134, 0x00010136}, {0x0001018F, 0x0001018F}, {0x0001019D, 0x0001019F}, {0x000101A1, 0x000101CF},
-{0x000101FE, 0x0001027F}, {0x0001029D, 0x0001029F}, {0x000102D1, 0x000102DF}, {0x000102FC, 0x000102FF},
-{0x00010324, 0x0001032C}, {0x0001034B, 0x0001034F}, {0x0001037B, 0x0001037F}, {0x0001039E, 0x0001039E},
-{0x000103C4, 0x000103C7}, {0x000103D6, 0x000103FF}, {0x0001049E, 0x0001049F}, {0x000104AA, 0x000104AF},
-{0x000104D4, 0x000104D7}, {0x000104FC, 0x000104FF}, {0x00010528, 0x0001052F}, {0x00010564, 0x0001056E},
-{0x00010570, 0x000105FF}, {0x00010737, 0x0001073F}, {0x00010756, 0x0001075F}, {0x00010768, 0x000107FF},
-{0x00010806, 0x00010807}, {0x00010809, 0x00010809}, {0x00010836, 0x00010836}, {0x00010839, 0x0001083B},
-{0x0001083D, 0x0001083E}, {0x00010856, 0x00010856}, {0x0001089F, 0x000108A6}, {0x000108B0, 0x000108DF},
-{0x000108F3, 0x000108F3}, {0x000108F6, 0x000108FA}, {0x0001091C, 0x0001091E}, {0x0001093A, 0x0001093E},
-{0x00010940, 0x0001097F}, {0x000109B8, 0x000109BB}, {0x000109D0, 0x000109D1}, {0x00010A04, 0x00010A04},
-{0x00010A07, 0x00010A0B}, {0x00010A14, 0x00010A14}, {0x00010A18, 0x00010A18}, {0x00010A36, 0x00010A37},
-{0x00010A3B, 0x00010A3E}, {0x00010A49, 0x00010A4F}, {0x00010A59, 0x00010A5F}, {0x00010AA0, 0x00010ABF},
-{0x00010AE7, 0x00010AEA}, {0x00010AF7, 0x00010AFF}, {0x00010B36, 0x00010B38}, {0x00010B56, 0x00010B57},
-{0x00010B73, 0x00010B77}, {0x00010B92, 0x00010B98}, {0x00010B9D, 0x00010BA8}, {0x00010BB0, 0x00010BFF},
-{0x00010C49, 0x00010C7F}, {0x00010CB3, 0x00010CBF}, {0x00010CF3, 0x00010CF9}, {0x00010D28, 0x00010D2F},
-{0x00010D3A, 0x00010E5F}, {0x00010E7F, 0x00010E7F}, {0x00010EAA, 0x00010EAA}, {0x00010EAE, 0x00010EAF},
-{0x00010EB2, 0x00010EFF}, {0x00010F28, 0x00010F2F}, {0x00010F5A, 0x00010FAF}, {0x00010FCC, 0x00010FDF},
-{0x00010FF7, 0x00010FFF}, {0x0001104E, 0x00011051}, {0x00011070, 0x0001107E}, {0x000110BD, 0x000110BD},
-{0x000110C2, 0x000110CF}, {0x000110E9, 0x000110EF}, {0x000110FA, 0x000110FF}, {0x00011135, 0x00011135},
-{0x00011148, 0x0001114F}, {0x00011177, 0x0001117F}, {0x000111E0, 0x000111E0}, {0x000111F5, 0x000111FF},
-{0x00011212, 0x00011212}, {0x0001123F, 0x0001127F}, {0x00011287, 0x00011287}, {0x00011289, 0x00011289},
-{0x0001128E, 0x0001128E}, {0x0001129E, 0x0001129E}, {0x000112AA, 0x000112AF}, {0x000112EB, 0x000112EF},
-{0x000112FA, 0x000112FF}, {0x00011304, 0x00011304}, {0x0001130D, 0x0001130E}, {0x00011311, 0x00011312},
-{0x00011329, 0x00011329}, {0x00011331, 0x00011331}, {0x00011334, 0x00011334}, {0x0001133A, 0x0001133A},
-{0x00011345, 0x00011346}, {0x00011349, 0x0001134A}, {0x0001134E, 0x0001134F}, {0x00011351, 0x00011356},
-{0x00011358, 0x0001135C}, {0x00011364, 0x00011365}, {0x0001136D, 0x0001136F}, {0x00011375, 0x000113FF},
-{0x0001145C, 0x0001145C}, {0x00011462, 0x0001147F}, {0x000114C8, 0x000114CF}, {0x000114DA, 0x0001157F},
-{0x000115B6, 0x000115B7}, {0x000115DE, 0x000115FF}, {0x00011645, 0x0001164F}, {0x0001165A, 0x0001165F},
-{0x0001166D, 0x0001167F}, {0x000116B9, 0x000116BF}, {0x000116CA, 0x000116FF}, {0x0001171B, 0x0001171C},
-{0x0001172C, 0x0001172F}, {0x00011740, 0x000117FF}, {0x0001183C, 0x0001189F}, {0x000118F3, 0x000118FE},
-{0x00011907, 0x00011908}, {0x0001190A, 0x0001190B}, {0x00011914, 0x00011914}, {0x00011917, 0x00011917},
-{0x00011936, 0x00011936}, {0x00011939, 0x0001193A}, {0x00011947, 0x0001194F}, {0x0001195A, 0x0001199F},
-{0x000119A8, 0x000119A9}, {0x000119D8, 0x000119D9}, {0x000119E5, 0x000119FF}, {0x00011A48, 0x00011A4F},
-{0x00011AA3, 0x00011ABF}, {0x00011AF9, 0x00011BFF}, {0x00011C09, 0x00011C09}, {0x00011C37, 0x00011C37},
-{0x00011C46, 0x00011C4F}, {0x00011C6D, 0x00011C6F}, {0x00011C90, 0x00011C91}, {0x00011CA8, 0x00011CA8},
-{0x00011CB7, 0x00011CFF}, {0x00011D07, 0x00011D07}, {0x00011D0A, 0x00011D0A}, {0x00011D37, 0x00011D39},
-{0x00011D3B, 0x00011D3B}, {0x00011D3E, 0x00011D3E}, {0x00011D48, 0x00011D4F}, {0x00011D5A, 0x00011D5F},
-{0x00011D66, 0x00011D66}, {0x00011D69, 0x00011D69}, {0x00011D8F, 0x00011D8F}, {0x00011D92, 0x00011D92},
-{0x00011D99, 0x00011D9F}, {0x00011DAA, 0x00011EDF}, {0x00011EF9, 0x00011FAF}, {0x00011FB1, 0x00011FBF},
-{0x00011FF2, 0x00011FFE}, {0x0001239A, 0x000123FF}, {0x0001246F, 0x0001246F}, {0x00012475, 0x0001247F},
-{0x00012544, 0x00012FFF}, {0x0001342F, 0x000143FF}, {0x00014647, 0x000167FF}, {0x00016A39, 0x00016A3F},
-{0x00016A5F, 0x00016A5F}, {0x00016A6A, 0x00016A6D}, {0x00016A70, 0x00016ACF}, {0x00016AEE, 0x00016AEF},
-{0x00016AF6, 0x00016AFF}, {0x00016B46, 0x00016B4F}, {0x00016B5A, 0x00016B5A}, {0x00016B62, 0x00016B62},
-{0x00016B78, 0x00016B7C}, {0x00016B90, 0x00016E3F}, {0x00016E9B, 0x00016EFF}, {0x00016F4B, 0x00016F4E},
-{0x00016F88, 0x00016F8E}, {0x00016FA0, 0x00016FDF}, {0x00016FE5, 0x00016FEF}, {0x00016FF2, 0x00016FFF},
-{0x000187F8, 0x000187FF}, {0x00018CD6, 0x00018CFF}, {0x00018D09, 0x0001AFFF}, {0x0001B11F, 0x0001B14F},
-{0x0001B153, 0x0001B163}, {0x0001B168, 0x0001B16F}, {0x0001B2FC, 0x0001BBFF}, {0x0001BC6B, 0x0001BC6F},
-{0x0001BC7D, 0x0001BC7F}, {0x0001BC89, 0x0001BC8F}, {0x0001BC9A, 0x0001BC9B}, {0x0001BCA0, 0x0001CFFF},
-{0x0001D0F6, 0x0001D0FF}, {0x0001D127, 0x0001D128}, {0x0001D173, 0x0001D17A}, {0x0001D1E9, 0x0001D1FF},
-{0x0001D246, 0x0001D2DF}, {0x0001D2F4, 0x0001D2FF}, {0x0001D357, 0x0001D35F}, {0x0001D379, 0x0001D3FF},
-{0x0001D455, 0x0001D455}, {0x0001D49D, 0x0001D49D}, {0x0001D4A0, 0x0001D4A1}, {0x0001D4A3, 0x0001D4A4},
-{0x0001D4A7, 0x0001D4A8}, {0x0001D4AD, 0x0001D4AD}, {0x0001D4BA, 0x0001D4BA}, {0x0001D4BC, 0x0001D4BC},
-{0x0001D4C4, 0x0001D4C4}, {0x0001D506, 0x0001D506}, {0x0001D50B, 0x0001D50C}, {0x0001D515, 0x0001D515},
-{0x0001D51D, 0x0001D51D}, {0x0001D53A, 0x0001D53A}, {0x0001D53F, 0x0001D53F}, {0x0001D545, 0x0001D545},
-{0x0001D547, 0x0001D549}, {0x0001D551, 0x0001D551}, {0x0001D6A6, 0x0001D6A7}, {0x0001D7CC, 0x0001D7CD},
-{0x0001DA8C, 0x0001DA9A}, {0x0001DAA0, 0x0001DAA0}, {0x0001DAB0, 0x0001DFFF}, {0x0001E007, 0x0001E007},
-{0x0001E019, 0x0001E01A}, {0x0001E022, 0x0001E022}, {0x0001E025, 0x0001E025}, {0x0001E02B, 0x0001E0FF},
-{0x0001E12D, 0x0001E12F}, {0x0001E13E, 0x0001E13F}, {0x0001E14A, 0x0001E14D}, {0x0001E150, 0x0001E2BF},
-{0x0001E2FA, 0x0001E2FE}, {0x0001E300, 0x0001E7FF}, {0x0001E8C5, 0x0001E8C6}, {0x0001E8D7, 0x0001E8FF},
-{0x0001E94C, 0x0001E94F}, {0x0001E95A, 0x0001E95D}, {0x0001E960, 0x0001EC70}, {0x0001ECB5, 0x0001ED00},
-{0x0001ED3E, 0x0001EDFF}, {0x0001EE04, 0x0001EE04}, {0x0001EE20, 0x0001EE20}, {0x0001EE23, 0x0001EE23},
-{0x0001EE25, 0x0001EE26}, {0x0001EE28, 0x0001EE28}, {0x0001EE33, 0x0001EE33}, {0x0001EE38, 0x0001EE38},
-{0x0001EE3A, 0x0001EE3A}, {0x0001EE3C, 0x0001EE41}, {0x0001EE43, 0x0001EE46}, {0x0001EE48, 0x0001EE48},
-{0x0001EE4A, 0x0001EE4A}, {0x0001EE4C, 0x0001EE4C}, {0x0001EE50, 0x0001EE50}, {0x0001EE53, 0x0001EE53},
-{0x0001EE55, 0x0001EE56}, {0x0001EE58, 0x0001EE58}, {0x0001EE5A, 0x0001EE5A}, {0x0001EE5C, 0x0001EE5C},
-{0x0001EE5E, 0x0001EE5E}, {0x0001EE60, 0x0001EE60}, {0x0001EE63, 0x0001EE63}, {0x0001EE65, 0x0001EE66},
-{0x0001EE6B, 0x0001EE6B}, {0x0001EE73, 0x0001EE73}, {0x0001EE78, 0x0001EE78}, {0x0001EE7D, 0x0001EE7D},
-{0x0001EE7F, 0x0001EE7F}, {0x0001EE8A, 0x0001EE8A}, {0x0001EE9C, 0x0001EEA0}, {0x0001EEA4, 0x0001EEA4},
-{0x0001EEAA, 0x0001EEAA}, {0x0001EEBC, 0x0001EEEF}, {0x0001EEF2, 0x0001EFFF}, {0x0001F02C, 0x0001F02F},
-{0x0001F094, 0x0001F09F}, {0x0001F0AF, 0x0001F0B0}, {0x0001F0C0, 0x0001F0C0}, {0x0001F0D0, 0x0001F0D0},
-{0x0001F0F6, 0x0001F0FF}, {0x0001F1AE, 0x0001F1E5}, {0x0001F203, 0x0001F20F}, {0x0001F23C, 0x0001F23F},
-{0x0001F249, 0x0001F24F}, {0x0001F252, 0x0001F25F}, {0x0001F266, 0x0001F2FF}, {0x0001F6D8, 0x0001F6DF},
-{0x0001F6ED, 0x0001F6EF}, {0x0001F6FD, 0x0001F6FF}, {0x0001F774, 0x0001F77F}, {0x0001F7D9, 0x0001F7DF},
-{0x0001F7EC, 0x0001F7FF}, {0x0001F80C, 0x0001F80F}, {0x0001F848, 0x0001F84F}, {0x0001F85A, 0x0001F85F},
-{0x0001F888, 0x0001F88F}, {0x0001F8AE, 0x0001F8AF}, {0x0001F8B2, 0x0001F8FF}, {0x0001F979, 0x0001F979},
-{0x0001F9CC, 0x0001F9CC}, {0x0001FA54, 0x0001FA5F}, {0x0001FA6E, 0x0001FA6F}, {0x0001FA75, 0x0001FA77},
-{0x0001FA7B, 0x0001FA7F}, {0x0001FA87, 0x0001FA8F}, {0x0001FAA9, 0x0001FAAF}, {0x0001FAB7, 0x0001FABF},
-{0x0001FAC3, 0x0001FACF}, {0x0001FAD7, 0x0001FAFF}, {0x0001FB93, 0x0001FB93}, {0x0001FBCB, 0x0001FBEF},
-{0x0001FBFA, 0x0001FFFF}, {0x0002A6DE, 0x0002A6FF}, {0x0002B735, 0x0002B73F}, {0x0002B81E, 0x0002B81F},
-{0x0002CEA2, 0x0002CEAF}, {0x0002EBE1, 0x0002F7FF}, {0x0002FA1E, 0x0002FFFF}, {0x0003134B, 0x000E00FF},
-{0x000E01F0, 0x0010FFFF},
+const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase = {
+{0x000061, 0x000041},
+{0x000062, 0x000042},
+{0x000063, 0x000043},
+{0x000064, 0x000044},
+{0x000065, 0x000045},
+{0x000066, 0x000046},
+{0x000067, 0x000047},
+{0x000068, 0x000048},
+{0x000069, 0x000049},
+{0x00006A, 0x00004A},
+{0x00006B, 0x00004B},
+{0x00006C, 0x00004C},
+{0x00006D, 0x00004D},
+{0x00006E, 0x00004E},
+{0x00006F, 0x00004F},
+{0x000070, 0x000050},
+{0x000071, 0x000051},
+{0x000072, 0x000052},
+{0x000073, 0x000053},
+{0x000074, 0x000054},
+{0x000075, 0x000055},
+{0x000076, 0x000056},
+{0x000077, 0x000057},
+{0x000078, 0x000058},
+{0x000079, 0x000059},
+{0x00007A, 0x00005A},
+{0x0000B5, 0x00039C},
+{0x0000DF, 0x000053},
+{0x0000E0, 0x0000C0},
+{0x0000E1, 0x0000C1},
+{0x0000E2, 0x0000C2},
+{0x0000E3, 0x0000C3},
+{0x0000E4, 0x0000C4},
+{0x0000E5, 0x0000C5},
+{0x0000E6, 0x0000C6},
+{0x0000E7, 0x0000C7},
+{0x0000E8, 0x0000C8},
+{0x0000E9, 0x0000C9},
+{0x0000EA, 0x0000CA},
+{0x0000EB, 0x0000CB},
+{0x0000EC, 0x0000CC},
+{0x0000ED, 0x0000CD},
+{0x0000EE, 0x0000CE},
+{0x0000EF, 0x0000CF},
+{0x0000F0, 0x0000D0},
+{0x0000F1, 0x0000D1},
+{0x0000F2, 0x0000D2},
+{0x0000F3, 0x0000D3},
+{0x0000F4, 0x0000D4},
+{0x0000F5, 0x0000D5},
+{0x0000F6, 0x0000D6},
+{0x0000F8, 0x0000D8},
+{0x0000F9, 0x0000D9},
+{0x0000FA, 0x0000DA},
+{0x0000FB, 0x0000DB},
+{0x0000FC, 0x0000DC},
+{0x0000FD, 0x0000DD},
+{0x0000FE, 0x0000DE},
+{0x0000FF, 0x000178},
+{0x000101, 0x000100},
+{0x000103, 0x000102},
+{0x000105, 0x000104},
+{0x000107, 0x000106},
+{0x000109, 0x000108},
+{0x00010B, 0x00010A},
+{0x00010D, 0x00010C},
+{0x00010F, 0x00010E},
+{0x000111, 0x000110},
+{0x000113, 0x000112},
+{0x000115, 0x000114},
+{0x000117, 0x000116},
+{0x000119, 0x000118},
+{0x00011B, 0x00011A},
+{0x00011D, 0x00011C},
+{0x00011F, 0x00011E},
+{0x000121, 0x000120},
+{0x000123, 0x000122},
+{0x000125, 0x000124},
+{0x000127, 0x000126},
+{0x000129, 0x000128},
+{0x00012B, 0x00012A},
+{0x00012D, 0x00012C},
+{0x00012F, 0x00012E},
+{0x000131, 0x000049},
+{0x000133, 0x000132},
+{0x000135, 0x000134},
+{0x000137, 0x000136},
+{0x00013A, 0x000139},
+{0x00013C, 0x00013B},
+{0x00013E, 0x00013D},
+{0x000140, 0x00013F},
+{0x000142, 0x000141},
+{0x000144, 0x000143},
+{0x000146, 0x000145},
+{0x000148, 0x000147},
+{0x000149, 0x0002BC},
+{0x00014B, 0x00014A},
+{0x00014D, 0x00014C},
+{0x00014F, 0x00014E},
+{0x000151, 0x000150},
+{0x000153, 0x000152},
+{0x000155, 0x000154},
+{0x000157, 0x000156},
+{0x000159, 0x000158},
+{0x00015B, 0x00015A},
+{0x00015D, 0x00015C},
+{0x00015F, 0x00015E},
+{0x000161, 0x000160},
+{0x000163, 0x000162},
+{0x000165, 0x000164},
+{0x000167, 0x000166},
+{0x000169, 0x000168},
+{0x00016B, 0x00016A},
+{0x00016D, 0x00016C},
+{0x00016F, 0x00016E},
+{0x000171, 0x000170},
+{0x000173, 0x000172},
+{0x000175, 0x000174},
+{0x000177, 0x000176},
+{0x00017A, 0x000179},
+{0x00017C, 0x00017B},
+{0x00017E, 0x00017D},
+{0x00017F, 0x000053},
+{0x000180, 0x000243},
+{0x000183, 0x000182},
+{0x000185, 0x000184},
+{0x000188, 0x000187},
+{0x00018C, 0x00018B},
+{0x000192, 0x000191},
+{0x000195, 0x0001F6},
+{0x000199, 0x000198},
+{0x00019A, 0x00023D},
+{0x00019E, 0x000220},
+{0x0001A1, 0x0001A0},
+{0x0001A3, 0x0001A2},
+{0x0001A5, 0x0001A4},
+{0x0001A8, 0x0001A7},
+{0x0001AD, 0x0001AC},
+{0x0001B0, 0x0001AF},
+{0x0001B4, 0x0001B3},
+{0x0001B6, 0x0001B5},
+{0x0001B9, 0x0001B8},
+{0x0001BD, 0x0001BC},
+{0x0001BF, 0x0001F7},
+{0x0001C5, 0x0001C4},
+{0x0001C6, 0x0001C4},
+{0x0001C8, 0x0001C7},
+{0x0001C9, 0x0001C7},
+{0x0001CB, 0x0001CA},
+{0x0001CC, 0x0001CA},
+{0x0001CE, 0x0001CD},
+{0x0001D0, 0x0001CF},
+{0x0001D2, 0x0001D1},
+{0x0001D4, 0x0001D3},
+{0x0001D6, 0x0001D5},
+{0x0001D8, 0x0001D7},
+{0x0001DA, 0x0001D9},
+{0x0001DC, 0x0001DB},
+{0x0001DD, 0x00018E},
+{0x0001DF, 0x0001DE},
+{0x0001E1, 0x0001E0},
+{0x0001E3, 0x0001E2},
+{0x0001E5, 0x0001E4},
+{0x0001E7, 0x0001E6},
+{0x0001E9, 0x0001E8},
+{0x0001EB, 0x0001EA},
+{0x0001ED, 0x0001EC},
+{0x0001EF, 0x0001EE},
+{0x0001F0, 0x00004A},
+{0x0001F2, 0x0001F1},
+{0x0001F3, 0x0001F1},
+{0x0001F5, 0x0001F4},
+{0x0001F9, 0x0001F8},
+{0x0001FB, 0x0001FA},
+{0x0001FD, 0x0001FC},
+{0x0001FF, 0x0001FE},
+{0x000201, 0x000200},
+{0x000203, 0x000202},
+{0x000205, 0x000204},
+{0x000207, 0x000206},
+{0x000209, 0x000208},
+{0x00020B, 0x00020A},
+{0x00020D, 0x00020C},
+{0x00020F, 0x00020E},
+{0x000211, 0x000210},
+{0x000213, 0x000212},
+{0x000215, 0x000214},
+{0x000217, 0x000216},
+{0x000219, 0x000218},
+{0x00021B, 0x00021A},
+{0x00021D, 0x00021C},
+{0x00021F, 0x00021E},
+{0x000223, 0x000222},
+{0x000225, 0x000224},
+{0x000227, 0x000226},
+{0x000229, 0x000228},
+{0x00022B, 0x00022A},
+{0x00022D, 0x00022C},
+{0x00022F, 0x00022E},
+{0x000231, 0x000230},
+{0x000233, 0x000232},
+{0x00023C, 0x00023B},
+{0x00023F, 0x002C7E},
+{0x000240, 0x002C7F},
+{0x000242, 0x000241},
+{0x000247, 0x000246},
+{0x000249, 0x000248},
+{0x00024B, 0x00024A},
+{0x00024D, 0x00024C},
+{0x00024F, 0x00024E},
+{0x000250, 0x002C6F},
+{0x000251, 0x002C6D},
+{0x000252, 0x002C70},
+{0x000253, 0x000181},
+{0x000254, 0x000186},
+{0x000256, 0x000189},
+{0x000257, 0x00018A},
+{0x000259, 0x00018F},
+{0x00025B, 0x000190},
+{0x00025C, 0x00A7AB},
+{0x000260, 0x000193},
+{0x000261, 0x00A7AC},
+{0x000263, 0x000194},
+{0x000265, 0x00A78D},
+{0x000266, 0x00A7AA},
+{0x000268, 0x000197},
+{0x000269, 0x000196},
+{0x00026A, 0x00A7AE},
+{0x00026B, 0x002C62},
+{0x00026C, 0x00A7AD},
+{0x00026F, 0x00019C},
+{0x000271, 0x002C6E},
+{0x000272, 0x00019D},
+{0x000275, 0x00019F},
+{0x00027D, 0x002C64},
+{0x000280, 0x0001A6},
+{0x000282, 0x00A7C5},
+{0x000283, 0x0001A9},
+{0x000287, 0x00A7B1},
+{0x000288, 0x0001AE},
+{0x000289, 0x000244},
+{0x00028A, 0x0001B1},
+{0x00028B, 0x0001B2},
+{0x00028C, 0x000245},
+{0x000292, 0x0001B7},
+{0x00029D, 0x00A7B2},
+{0x00029E, 0x00A7B0},
+{0x000345, 0x000399},
+{0x000371, 0x000370},
+{0x000373, 0x000372},
+{0x000377, 0x000376},
+{0x00037B, 0x0003FD},
+{0x00037C, 0x0003FE},
+{0x00037D, 0x0003FF},
+{0x000390, 0x000399},
+{0x0003AC, 0x000386},
+{0x0003AD, 0x000388},
+{0x0003AE, 0x000389},
+{0x0003AF, 0x00038A},
+{0x0003B0, 0x0003A5},
+{0x0003B1, 0x000391},
+{0x0003B2, 0x000392},
+{0x0003B3, 0x000393},
+{0x0003B4, 0x000394},
+{0x0003B5, 0x000395},
+{0x0003B6, 0x000396},
+{0x0003B7, 0x000397},
+{0x0003B8, 0x000398},
+{0x0003B9, 0x000399},
+{0x0003BA, 0x00039A},
+{0x0003BB, 0x00039B},
+{0x0003BC, 0x00039C},
+{0x0003BD, 0x00039D},
+{0x0003BE, 0x00039E},
+{0x0003BF, 0x00039F},
+{0x0003C0, 0x0003A0},
+{0x0003C1, 0x0003A1},
+{0x0003C2, 0x0003A3},
+{0x0003C3, 0x0003A3},
+{0x0003C4, 0x0003A4},
+{0x0003C5, 0x0003A5},
+{0x0003C6, 0x0003A6},
+{0x0003C7, 0x0003A7},
+{0x0003C8, 0x0003A8},
+{0x0003C9, 0x0003A9},
+{0x0003CA, 0x0003AA},
+{0x0003CB, 0x0003AB},
+{0x0003CC, 0x00038C},
+{0x0003CD, 0x00038E},
+{0x0003CE, 0x00038F},
+{0x0003D0, 0x000392},
+{0x0003D1, 0x000398},
+{0x0003D5, 0x0003A6},
+{0x0003D6, 0x0003A0},
+{0x0003D7, 0x0003CF},
+{0x0003D9, 0x0003D8},
+{0x0003DB, 0x0003DA},
+{0x0003DD, 0x0003DC},
+{0x0003DF, 0x0003DE},
+{0x0003E1, 0x0003E0},
+{0x0003E3, 0x0003E2},
+{0x0003E5, 0x0003E4},
+{0x0003E7, 0x0003E6},
+{0x0003E9, 0x0003E8},
+{0x0003EB, 0x0003EA},
+{0x0003ED, 0x0003EC},
+{0x0003EF, 0x0003EE},
+{0x0003F0, 0x00039A},
+{0x0003F1, 0x0003A1},
+{0x0003F2, 0x0003F9},
+{0x0003F3, 0x00037F},
+{0x0003F5, 0x000395},
+{0x0003F8, 0x0003F7},
+{0x0003FB, 0x0003FA},
+{0x000430, 0x000410},
+{0x000431, 0x000411},
+{0x000432, 0x000412},
+{0x000433, 0x000413},
+{0x000434, 0x000414},
+{0x000435, 0x000415},
+{0x000436, 0x000416},
+{0x000437, 0x000417},
+{0x000438, 0x000418},
+{0x000439, 0x000419},
+{0x00043A, 0x00041A},
+{0x00043B, 0x00041B},
+{0x00043C, 0x00041C},
+{0x00043D, 0x00041D},
+{0x00043E, 0x00041E},
+{0x00043F, 0x00041F},
+{0x000440, 0x000420},
+{0x000441, 0x000421},
+{0x000442, 0x000422},
+{0x000443, 0x000423},
+{0x000444, 0x000424},
+{0x000445, 0x000425},
+{0x000446, 0x000426},
+{0x000447, 0x000427},
+{0x000448, 0x000428},
+{0x000449, 0x000429},
+{0x00044A, 0x00042A},
+{0x00044B, 0x00042B},
+{0x00044C, 0x00042C},
+{0x00044D, 0x00042D},
+{0x00044E, 0x00042E},
+{0x00044F, 0x00042F},
+{0x000450, 0x000400},
+{0x000451, 0x000401},
+{0x000452, 0x000402},
+{0x000453, 0x000403},
+{0x000454, 0x000404},
+{0x000455, 0x000405},
+{0x000456, 0x000406},
+{0x000457, 0x000407},
+{0x000458, 0x000408},
+{0x000459, 0x000409},
+{0x00045A, 0x00040A},
+{0x00045B, 0x00040B},
+{0x00045C, 0x00040C},
+{0x00045D, 0x00040D},
+{0x00045E, 0x00040E},
+{0x00045F, 0x00040F},
+{0x000461, 0x000460},
+{0x000463, 0x000462},
+{0x000465, 0x000464},
+{0x000467, 0x000466},
+{0x000469, 0x000468},
+{0x00046B, 0x00046A},
+{0x00046D, 0x00046C},
+{0x00046F, 0x00046E},
+{0x000471, 0x000470},
+{0x000473, 0x000472},
+{0x000475, 0x000474},
+{0x000477, 0x000476},
+{0x000479, 0x000478},
+{0x00047B, 0x00047A},
+{0x00047D, 0x00047C},
+{0x00047F, 0x00047E},
+{0x000481, 0x000480},
+{0x00048B, 0x00048A},
+{0x00048D, 0x00048C},
+{0x00048F, 0x00048E},
+{0x000491, 0x000490},
+{0x000493, 0x000492},
+{0x000495, 0x000494},
+{0x000497, 0x000496},
+{0x000499, 0x000498},
+{0x00049B, 0x00049A},
+{0x00049D, 0x00049C},
+{0x00049F, 0x00049E},
+{0x0004A1, 0x0004A0},
+{0x0004A3, 0x0004A2},
+{0x0004A5, 0x0004A4},
+{0x0004A7, 0x0004A6},
+{0x0004A9, 0x0004A8},
+{0x0004AB, 0x0004AA},
+{0x0004AD, 0x0004AC},
+{0x0004AF, 0x0004AE},
+{0x0004B1, 0x0004B0},
+{0x0004B3, 0x0004B2},
+{0x0004B5, 0x0004B4},
+{0x0004B7, 0x0004B6},
+{0x0004B9, 0x0004B8},
+{0x0004BB, 0x0004BA},
+{0x0004BD, 0x0004BC},
+{0x0004BF, 0x0004BE},
+{0x0004C2, 0x0004C1},
+{0x0004C4, 0x0004C3},
+{0x0004C6, 0x0004C5},
+{0x0004C8, 0x0004C7},
+{0x0004CA, 0x0004C9},
+{0x0004CC, 0x0004CB},
+{0x0004CE, 0x0004CD},
+{0x0004CF, 0x0004C0},
+{0x0004D1, 0x0004D0},
+{0x0004D3, 0x0004D2},
+{0x0004D5, 0x0004D4},
+{0x0004D7, 0x0004D6},
+{0x0004D9, 0x0004D8},
+{0x0004DB, 0x0004DA},
+{0x0004DD, 0x0004DC},
+{0x0004DF, 0x0004DE},
+{0x0004E1, 0x0004E0},
+{0x0004E3, 0x0004E2},
+{0x0004E5, 0x0004E4},
+{0x0004E7, 0x0004E6},
+{0x0004E9, 0x0004E8},
+{0x0004EB, 0x0004EA},
+{0x0004ED, 0x0004EC},
+{0x0004EF, 0x0004EE},
+{0x0004F1, 0x0004F0},
+{0x0004F3, 0x0004F2},
+{0x0004F5, 0x0004F4},
+{0x0004F7, 0x0004F6},
+{0x0004F9, 0x0004F8},
+{0x0004FB, 0x0004FA},
+{0x0004FD, 0x0004FC},
+{0x0004FF, 0x0004FE},
+{0x000501, 0x000500},
+{0x000503, 0x000502},
+{0x000505, 0x000504},
+{0x000507, 0x000506},
+{0x000509, 0x000508},
+{0x00050B, 0x00050A},
+{0x00050D, 0x00050C},
+{0x00050F, 0x00050E},
+{0x000511, 0x000510},
+{0x000513, 0x000512},
+{0x000515, 0x000514},
+{0x000517, 0x000516},
+{0x000519, 0x000518},
+{0x00051B, 0x00051A},
+{0x00051D, 0x00051C},
+{0x00051F, 0x00051E},
+{0x000521, 0x000520},
+{0x000523, 0x000522},
+{0x000525, 0x000524},
+{0x000527, 0x000526},
+{0x000529, 0x000528},
+{0x00052B, 0x00052A},
+{0x00052D, 0x00052C},
+{0x00052F, 0x00052E},
+{0x000561, 0x000531},
+{0x000562, 0x000532},
+{0x000563, 0x000533},
+{0x000564, 0x000534},
+{0x000565, 0x000535},
+{0x000566, 0x000536},
+{0x000567, 0x000537},
+{0x000568, 0x000538},
+{0x000569, 0x000539},
+{0x00056A, 0x00053A},
+{0x00056B, 0x00053B},
+{0x00056C, 0x00053C},
+{0x00056D, 0x00053D},
+{0x00056E, 0x00053E},
+{0x00056F, 0x00053F},
+{0x000570, 0x000540},
+{0x000571, 0x000541},
+{0x000572, 0x000542},
+{0x000573, 0x000543},
+{0x000574, 0x000544},
+{0x000575, 0x000545},
+{0x000576, 0x000546},
+{0x000577, 0x000547},
+{0x000578, 0x000548},
+{0x000579, 0x000549},
+{0x00057A, 0x00054A},
+{0x00057B, 0x00054B},
+{0x00057C, 0x00054C},
+{0x00057D, 0x00054D},
+{0x00057E, 0x00054E},
+{0x00057F, 0x00054F},
+{0x000580, 0x000550},
+{0x000581, 0x000551},
+{0x000582, 0x000552},
+{0x000583, 0x000553},
+{0x000584, 0x000554},
+{0x000585, 0x000555},
+{0x000586, 0x000556},
+{0x000587, 0x000535},
+{0x0010D0, 0x001C90},
+{0x0010D1, 0x001C91},
+{0x0010D2, 0x001C92},
+{0x0010D3, 0x001C93},
+{0x0010D4, 0x001C94},
+{0x0010D5, 0x001C95},
+{0x0010D6, 0x001C96},
+{0x0010D7, 0x001C97},
+{0x0010D8, 0x001C98},
+{0x0010D9, 0x001C99},
+{0x0010DA, 0x001C9A},
+{0x0010DB, 0x001C9B},
+{0x0010DC, 0x001C9C},
+{0x0010DD, 0x001C9D},
+{0x0010DE, 0x001C9E},
+{0x0010DF, 0x001C9F},
+{0x0010E0, 0x001CA0},
+{0x0010E1, 0x001CA1},
+{0x0010E2, 0x001CA2},
+{0x0010E3, 0x001CA3},
+{0x0010E4, 0x001CA4},
+{0x0010E5, 0x001CA5},
+{0x0010E6, 0x001CA6},
+{0x0010E7, 0x001CA7},
+{0x0010E8, 0x001CA8},
+{0x0010E9, 0x001CA9},
+{0x0010EA, 0x001CAA},
+{0x0010EB, 0x001CAB},
+{0x0010EC, 0x001CAC},
+{0x0010ED, 0x001CAD},
+{0x0010EE, 0x001CAE},
+{0x0010EF, 0x001CAF},
+{0x0010F0, 0x001CB0},
+{0x0010F1, 0x001CB1},
+{0x0010F2, 0x001CB2},
+{0x0010F3, 0x001CB3},
+{0x0010F4, 0x001CB4},
+{0x0010F5, 0x001CB5},
+{0x0010F6, 0x001CB6},
+{0x0010F7, 0x001CB7},
+{0x0010F8, 0x001CB8},
+{0x0010F9, 0x001CB9},
+{0x0010FA, 0x001CBA},
+{0x0010FD, 0x001CBD},
+{0x0010FE, 0x001CBE},
+{0x0010FF, 0x001CBF},
+{0x0013F8, 0x0013F0},
+{0x0013F9, 0x0013F1},
+{0x0013FA, 0x0013F2},
+{0x0013FB, 0x0013F3},
+{0x0013FC, 0x0013F4},
+{0x0013FD, 0x0013F5},
+{0x001C80, 0x000412},
+{0x001C81, 0x000414},
+{0x001C82, 0x00041E},
+{0x001C83, 0x000421},
+{0x001C84, 0x000422},
+{0x001C85, 0x000422},
+{0x001C86, 0x00042A},
+{0x001C87, 0x000462},
+{0x001C88, 0x00A64A},
+{0x001D79, 0x00A77D},
+{0x001D7D, 0x002C63},
+{0x001D8E, 0x00A7C6},
+{0x001E01, 0x001E00},
+{0x001E03, 0x001E02},
+{0x001E05, 0x001E04},
+{0x001E07, 0x001E06},
+{0x001E09, 0x001E08},
+{0x001E0B, 0x001E0A},
+{0x001E0D, 0x001E0C},
+{0x001E0F, 0x001E0E},
+{0x001E11, 0x001E10},
+{0x001E13, 0x001E12},
+{0x001E15, 0x001E14},
+{0x001E17, 0x001E16},
+{0x001E19, 0x001E18},
+{0x001E1B, 0x001E1A},
+{0x001E1D, 0x001E1C},
+{0x001E1F, 0x001E1E},
+{0x001E21, 0x001E20},
+{0x001E23, 0x001E22},
+{0x001E25, 0x001E24},
+{0x001E27, 0x001E26},
+{0x001E29, 0x001E28},
+{0x001E2B, 0x001E2A},
+{0x001E2D, 0x001E2C},
+{0x001E2F, 0x001E2E},
+{0x001E31, 0x001E30},
+{0x001E33, 0x001E32},
+{0x001E35, 0x001E34},
+{0x001E37, 0x001E36},
+{0x001E39, 0x001E38},
+{0x001E3B, 0x001E3A},
+{0x001E3D, 0x001E3C},
+{0x001E3F, 0x001E3E},
+{0x001E41, 0x001E40},
+{0x001E43, 0x001E42},
+{0x001E45, 0x001E44},
+{0x001E47, 0x001E46},
+{0x001E49, 0x001E48},
+{0x001E4B, 0x001E4A},
+{0x001E4D, 0x001E4C},
+{0x001E4F, 0x001E4E},
+{0x001E51, 0x001E50},
+{0x001E53, 0x001E52},
+{0x001E55, 0x001E54},
+{0x001E57, 0x001E56},
+{0x001E59, 0x001E58},
+{0x001E5B, 0x001E5A},
+{0x001E5D, 0x001E5C},
+{0x001E5F, 0x001E5E},
+{0x001E61, 0x001E60},
+{0x001E63, 0x001E62},
+{0x001E65, 0x001E64},
+{0x001E67, 0x001E66},
+{0x001E69, 0x001E68},
+{0x001E6B, 0x001E6A},
+{0x001E6D, 0x001E6C},
+{0x001E6F, 0x001E6E},
+{0x001E71, 0x001E70},
+{0x001E73, 0x001E72},
+{0x001E75, 0x001E74},
+{0x001E77, 0x001E76},
+{0x001E79, 0x001E78},
+{0x001E7B, 0x001E7A},
+{0x001E7D, 0x001E7C},
+{0x001E7F, 0x001E7E},
+{0x001E81, 0x001E80},
+{0x001E83, 0x001E82},
+{0x001E85, 0x001E84},
+{0x001E87, 0x001E86},
+{0x001E89, 0x001E88},
+{0x001E8B, 0x001E8A},
+{0x001E8D, 0x001E8C},
+{0x001E8F, 0x001E8E},
+{0x001E91, 0x001E90},
+{0x001E93, 0x001E92},
+{0x001E95, 0x001E94},
+{0x001E96, 0x000048},
+{0x001E97, 0x000054},
+{0x001E98, 0x000057},
+{0x001E99, 0x000059},
+{0x001E9A, 0x000041},
+{0x001E9B, 0x001E60},
+{0x001EA1, 0x001EA0},
+{0x001EA3, 0x001EA2},
+{0x001EA5, 0x001EA4},
+{0x001EA7, 0x001EA6},
+{0x001EA9, 0x001EA8},
+{0x001EAB, 0x001EAA},
+{0x001EAD, 0x001EAC},
+{0x001EAF, 0x001EAE},
+{0x001EB1, 0x001EB0},
+{0x001EB3, 0x001EB2},
+{0x001EB5, 0x001EB4},
+{0x001EB7, 0x001EB6},
+{0x001EB9, 0x001EB8},
+{0x001EBB, 0x001EBA},
+{0x001EBD, 0x001EBC},
+{0x001EBF, 0x001EBE},
+{0x001EC1, 0x001EC0},
+{0x001EC3, 0x001EC2},
+{0x001EC5, 0x001EC4},
+{0x001EC7, 0x001EC6},
+{0x001EC9, 0x001EC8},
+{0x001ECB, 0x001ECA},
+{0x001ECD, 0x001ECC},
+{0x001ECF, 0x001ECE},
+{0x001ED1, 0x001ED0},
+{0x001ED3, 0x001ED2},
+{0x001ED5, 0x001ED4},
+{0x001ED7, 0x001ED6},
+{0x001ED9, 0x001ED8},
+{0x001EDB, 0x001EDA},
+{0x001EDD, 0x001EDC},
+{0x001EDF, 0x001EDE},
+{0x001EE1, 0x001EE0},
+{0x001EE3, 0x001EE2},
+{0x001EE5, 0x001EE4},
+{0x001EE7, 0x001EE6},
+{0x001EE9, 0x001EE8},
+{0x001EEB, 0x001EEA},
+{0x001EED, 0x001EEC},
+{0x001EEF, 0x001EEE},
+{0x001EF1, 0x001EF0},
+{0x001EF3, 0x001EF2},
+{0x001EF5, 0x001EF4},
+{0x001EF7, 0x001EF6},
+{0x001EF9, 0x001EF8},
+{0x001EFB, 0x001EFA},
+{0x001EFD, 0x001EFC},
+{0x001EFF, 0x001EFE},
+{0x001F00, 0x001F08},
+{0x001F01, 0x001F09},
+{0x001F02, 0x001F0A},
+{0x001F03, 0x001F0B},
+{0x001F04, 0x001F0C},
+{0x001F05, 0x001F0D},
+{0x001F06, 0x001F0E},
+{0x001F07, 0x001F0F},
+{0x001F10, 0x001F18},
+{0x001F11, 0x001F19},
+{0x001F12, 0x001F1A},
+{0x001F13, 0x001F1B},
+{0x001F14, 0x001F1C},
+{0x001F15, 0x001F1D},
+{0x001F20, 0x001F28},
+{0x001F21, 0x001F29},
+{0x001F22, 0x001F2A},
+{0x001F23, 0x001F2B},
+{0x001F24, 0x001F2C},
+{0x001F25, 0x001F2D},
+{0x001F26, 0x001F2E},
+{0x001F27, 0x001F2F},
+{0x001F30, 0x001F38},
+{0x001F31, 0x001F39},
+{0x001F32, 0x001F3A},
+{0x001F33, 0x001F3B},
+{0x001F34, 0x001F3C},
+{0x001F35, 0x001F3D},
+{0x001F36, 0x001F3E},
+{0x001F37, 0x001F3F},
+{0x001F40, 0x001F48},
+{0x001F41, 0x001F49},
+{0x001F42, 0x001F4A},
+{0x001F43, 0x001F4B},
+{0x001F44, 0x001F4C},
+{0x001F45, 0x001F4D},
+{0x001F50, 0x0003A5},
+{0x001F51, 0x001F59},
+{0x001F52, 0x0003A5},
+{0x001F53, 0x001F5B},
+{0x001F54, 0x0003A5},
+{0x001F55, 0x001F5D},
+{0x001F56, 0x0003A5},
+{0x001F57, 0x001F5F},
+{0x001F60, 0x001F68},
+{0x001F61, 0x001F69},
+{0x001F62, 0x001F6A},
+{0x001F63, 0x001F6B},
+{0x001F64, 0x001F6C},
+{0x001F65, 0x001F6D},
+{0x001F66, 0x001F6E},
+{0x001F67, 0x001F6F},
+{0x001F70, 0x001FBA},
+{0x001F71, 0x001FBB},
+{0x001F72, 0x001FC8},
+{0x001F73, 0x001FC9},
+{0x001F74, 0x001FCA},
+{0x001F75, 0x001FCB},
+{0x001F76, 0x001FDA},
+{0x001F77, 0x001FDB},
+{0x001F78, 0x001FF8},
+{0x001F79, 0x001FF9},
+{0x001F7A, 0x001FEA},
+{0x001F7B, 0x001FEB},
+{0x001F7C, 0x001FFA},
+{0x001F7D, 0x001FFB},
+{0x001F80, 0x001F08},
+{0x001F81, 0x001F09},
+{0x001F82, 0x001F0A},
+{0x001F83, 0x001F0B},
+{0x001F84, 0x001F0C},
+{0x001F85, 0x001F0D},
+{0x001F86, 0x001F0E},
+{0x001F87, 0x001F0F},
+{0x001F88, 0x001F08},
+{0x001F89, 0x001F09},
+{0x001F8A, 0x001F0A},
+{0x001F8B, 0x001F0B},
+{0x001F8C, 0x001F0C},
+{0x001F8D, 0x001F0D},
+{0x001F8E, 0x001F0E},
+{0x001F8F, 0x001F0F},
+{0x001F90, 0x001F28},
+{0x001F91, 0x001F29},
+{0x001F92, 0x001F2A},
+{0x001F93, 0x001F2B},
+{0x001F94, 0x001F2C},
+{0x001F95, 0x001F2D},
+{0x001F96, 0x001F2E},
+{0x001F97, 0x001F2F},
+{0x001F98, 0x001F28},
+{0x001F99, 0x001F29},
+{0x001F9A, 0x001F2A},
+{0x001F9B, 0x001F2B},
+{0x001F9C, 0x001F2C},
+{0x001F9D, 0x001F2D},
+{0x001F9E, 0x001F2E},
+{0x001F9F, 0x001F2F},
+{0x001FA0, 0x001F68},
+{0x001FA1, 0x001F69},
+{0x001FA2, 0x001F6A},
+{0x001FA3, 0x001F6B},
+{0x001FA4, 0x001F6C},
+{0x001FA5, 0x001F6D},
+{0x001FA6, 0x001F6E},
+{0x001FA7, 0x001F6F},
+{0x001FA8, 0x001F68},
+{0x001FA9, 0x001F69},
+{0x001FAA, 0x001F6A},
+{0x001FAB, 0x001F6B},
+{0x001FAC, 0x001F6C},
+{0x001FAD, 0x001F6D},
+{0x001FAE, 0x001F6E},
+{0x001FAF, 0x001F6F},
+{0x001FB0, 0x001FB8},
+{0x001FB1, 0x001FB9},
+{0x001FB2, 0x001FBA},
+{0x001FB3, 0x000391},
+{0x001FB4, 0x000386},
+{0x001FB6, 0x000391},
+{0x001FB7, 0x000391},
+{0x001FBC, 0x000391},
+{0x001FBE, 0x000399},
+{0x001FC2, 0x001FCA},
+{0x001FC3, 0x000397},
+{0x001FC4, 0x000389},
+{0x001FC6, 0x000397},
+{0x001FC7, 0x000397},
+{0x001FCC, 0x000397},
+{0x001FD0, 0x001FD8},
+{0x001FD1, 0x001FD9},
+{0x001FD2, 0x000399},
+{0x001FD3, 0x000399},
+{0x001FD6, 0x000399},
+{0x001FD7, 0x000399},
+{0x001FE0, 0x001FE8},
+{0x001FE1, 0x001FE9},
+{0x001FE2, 0x0003A5},
+{0x001FE3, 0x0003A5},
+{0x001FE4, 0x0003A1},
+{0x001FE5, 0x001FEC},
+{0x001FE6, 0x0003A5},
+{0x001FE7, 0x0003A5},
+{0x001FF2, 0x001FFA},
+{0x001FF3, 0x0003A9},
+{0x001FF4, 0x00038F},
+{0x001FF6, 0x0003A9},
+{0x001FF7, 0x0003A9},
+{0x001FFC, 0x0003A9},
+{0x00214E, 0x002132},
+{0x002170, 0x002160},
+{0x002171, 0x002161},
+{0x002172, 0x002162},
+{0x002173, 0x002163},
+{0x002174, 0x002164},
+{0x002175, 0x002165},
+{0x002176, 0x002166},
+{0x002177, 0x002167},
+{0x002178, 0x002168},
+{0x002179, 0x002169},
+{0x00217A, 0x00216A},
+{0x00217B, 0x00216B},
+{0x00217C, 0x00216C},
+{0x00217D, 0x00216D},
+{0x00217E, 0x00216E},
+{0x00217F, 0x00216F},
+{0x002184, 0x002183},
+{0x0024D0, 0x0024B6},
+{0x0024D1, 0x0024B7},
+{0x0024D2, 0x0024B8},
+{0x0024D3, 0x0024B9},
+{0x0024D4, 0x0024BA},
+{0x0024D5, 0x0024BB},
+{0x0024D6, 0x0024BC},
+{0x0024D7, 0x0024BD},
+{0x0024D8, 0x0024BE},
+{0x0024D9, 0x0024BF},
+{0x0024DA, 0x0024C0},
+{0x0024DB, 0x0024C1},
+{0x0024DC, 0x0024C2},
+{0x0024DD, 0x0024C3},
+{0x0024DE, 0x0024C4},
+{0x0024DF, 0x0024C5},
+{0x0024E0, 0x0024C6},
+{0x0024E1, 0x0024C7},
+{0x0024E2, 0x0024C8},
+{0x0024E3, 0x0024C9},
+{0x0024E4, 0x0024CA},
+{0x0024E5, 0x0024CB},
+{0x0024E6, 0x0024CC},
+{0x0024E7, 0x0024CD},
+{0x0024E8, 0x0024CE},
+{0x0024E9, 0x0024CF},
+{0x002C30, 0x002C00},
+{0x002C31, 0x002C01},
+{0x002C32, 0x002C02},
+{0x002C33, 0x002C03},
+{0x002C34, 0x002C04},
+{0x002C35, 0x002C05},
+{0x002C36, 0x002C06},
+{0x002C37, 0x002C07},
+{0x002C38, 0x002C08},
+{0x002C39, 0x002C09},
+{0x002C3A, 0x002C0A},
+{0x002C3B, 0x002C0B},
+{0x002C3C, 0x002C0C},
+{0x002C3D, 0x002C0D},
+{0x002C3E, 0x002C0E},
+{0x002C3F, 0x002C0F},
+{0x002C40, 0x002C10},
+{0x002C41, 0x002C11},
+{0x002C42, 0x002C12},
+{0x002C43, 0x002C13},
+{0x002C44, 0x002C14},
+{0x002C45, 0x002C15},
+{0x002C46, 0x002C16},
+{0x002C47, 0x002C17},
+{0x002C48, 0x002C18},
+{0x002C49, 0x002C19},
+{0x002C4A, 0x002C1A},
+{0x002C4B, 0x002C1B},
+{0x002C4C, 0x002C1C},
+{0x002C4D, 0x002C1D},
+{0x002C4E, 0x002C1E},
+{0x002C4F, 0x002C1F},
+{0x002C50, 0x002C20},
+{0x002C51, 0x002C21},
+{0x002C52, 0x002C22},
+{0x002C53, 0x002C23},
+{0x002C54, 0x002C24},
+{0x002C55, 0x002C25},
+{0x002C56, 0x002C26},
+{0x002C57, 0x002C27},
+{0x002C58, 0x002C28},
+{0x002C59, 0x002C29},
+{0x002C5A, 0x002C2A},
+{0x002C5B, 0x002C2B},
+{0x002C5C, 0x002C2C},
+{0x002C5D, 0x002C2D},
+{0x002C5E, 0x002C2E},
+{0x002C61, 0x002C60},
+{0x002C65, 0x00023A},
+{0x002C66, 0x00023E},
+{0x002C68, 0x002C67},
+{0x002C6A, 0x002C69},
+{0x002C6C, 0x002C6B},
+{0x002C73, 0x002C72},
+{0x002C76, 0x002C75},
+{0x002C81, 0x002C80},
+{0x002C83, 0x002C82},
+{0x002C85, 0x002C84},
+{0x002C87, 0x002C86},
+{0x002C89, 0x002C88},
+{0x002C8B, 0x002C8A},
+{0x002C8D, 0x002C8C},
+{0x002C8F, 0x002C8E},
+{0x002C91, 0x002C90},
+{0x002C93, 0x002C92},
+{0x002C95, 0x002C94},
+{0x002C97, 0x002C96},
+{0x002C99, 0x002C98},
+{0x002C9B, 0x002C9A},
+{0x002C9D, 0x002C9C},
+{0x002C9F, 0x002C9E},
+{0x002CA1, 0x002CA0},
+{0x002CA3, 0x002CA2},
+{0x002CA5, 0x002CA4},
+{0x002CA7, 0x002CA6},
+{0x002CA9, 0x002CA8},
+{0x002CAB, 0x002CAA},
+{0x002CAD, 0x002CAC},
+{0x002CAF, 0x002CAE},
+{0x002CB1, 0x002CB0},
+{0x002CB3, 0x002CB2},
+{0x002CB5, 0x002CB4},
+{0x002CB7, 0x002CB6},
+{0x002CB9, 0x002CB8},
+{0x002CBB, 0x002CBA},
+{0x002CBD, 0x002CBC},
+{0x002CBF, 0x002CBE},
+{0x002CC1, 0x002CC0},
+{0x002CC3, 0x002CC2},
+{0x002CC5, 0x002CC4},
+{0x002CC7, 0x002CC6},
+{0x002CC9, 0x002CC8},
+{0x002CCB, 0x002CCA},
+{0x002CCD, 0x002CCC},
+{0x002CCF, 0x002CCE},
+{0x002CD1, 0x002CD0},
+{0x002CD3, 0x002CD2},
+{0x002CD5, 0x002CD4},
+{0x002CD7, 0x002CD6},
+{0x002CD9, 0x002CD8},
+{0x002CDB, 0x002CDA},
+{0x002CDD, 0x002CDC},
+{0x002CDF, 0x002CDE},
+{0x002CE1, 0x002CE0},
+{0x002CE3, 0x002CE2},
+{0x002CEC, 0x002CEB},
+{0x002CEE, 0x002CED},
+{0x002CF3, 0x002CF2},
+{0x002D00, 0x0010A0},
+{0x002D01, 0x0010A1},
+{0x002D02, 0x0010A2},
+{0x002D03, 0x0010A3},
+{0x002D04, 0x0010A4},
+{0x002D05, 0x0010A5},
+{0x002D06, 0x0010A6},
+{0x002D07, 0x0010A7},
+{0x002D08, 0x0010A8},
+{0x002D09, 0x0010A9},
+{0x002D0A, 0x0010AA},
+{0x002D0B, 0x0010AB},
+{0x002D0C, 0x0010AC},
+{0x002D0D, 0x0010AD},
+{0x002D0E, 0x0010AE},
+{0x002D0F, 0x0010AF},
+{0x002D10, 0x0010B0},
+{0x002D11, 0x0010B1},
+{0x002D12, 0x0010B2},
+{0x002D13, 0x0010B3},
+{0x002D14, 0x0010B4},
+{0x002D15, 0x0010B5},
+{0x002D16, 0x0010B6},
+{0x002D17, 0x0010B7},
+{0x002D18, 0x0010B8},
+{0x002D19, 0x0010B9},
+{0x002D1A, 0x0010BA},
+{0x002D1B, 0x0010BB},
+{0x002D1C, 0x0010BC},
+{0x002D1D, 0x0010BD},
+{0x002D1E, 0x0010BE},
+{0x002D1F, 0x0010BF},
+{0x002D20, 0x0010C0},
+{0x002D21, 0x0010C1},
+{0x002D22, 0x0010C2},
+{0x002D23, 0x0010C3},
+{0x002D24, 0x0010C4},
+{0x002D25, 0x0010C5},
+{0x002D27, 0x0010C7},
+{0x002D2D, 0x0010CD},
+{0x00A641, 0x00A640},
+{0x00A643, 0x00A642},
+{0x00A645, 0x00A644},
+{0x00A647, 0x00A646},
+{0x00A649, 0x00A648},
+{0x00A64B, 0x00A64A},
+{0x00A64D, 0x00A64C},
+{0x00A64F, 0x00A64E},
+{0x00A651, 0x00A650},
+{0x00A653, 0x00A652},
+{0x00A655, 0x00A654},
+{0x00A657, 0x00A656},
+{0x00A659, 0x00A658},
+{0x00A65B, 0x00A65A},
+{0x00A65D, 0x00A65C},
+{0x00A65F, 0x00A65E},
+{0x00A661, 0x00A660},
+{0x00A663, 0x00A662},
+{0x00A665, 0x00A664},
+{0x00A667, 0x00A666},
+{0x00A669, 0x00A668},
+{0x00A66B, 0x00A66A},
+{0x00A66D, 0x00A66C},
+{0x00A681, 0x00A680},
+{0x00A683, 0x00A682},
+{0x00A685, 0x00A684},
+{0x00A687, 0x00A686},
+{0x00A689, 0x00A688},
+{0x00A68B, 0x00A68A},
+{0x00A68D, 0x00A68C},
+{0x00A68F, 0x00A68E},
+{0x00A691, 0x00A690},
+{0x00A693, 0x00A692},
+{0x00A695, 0x00A694},
+{0x00A697, 0x00A696},
+{0x00A699, 0x00A698},
+{0x00A69B, 0x00A69A},
+{0x00A723, 0x00A722},
+{0x00A725, 0x00A724},
+{0x00A727, 0x00A726},
+{0x00A729, 0x00A728},
+{0x00A72B, 0x00A72A},
+{0x00A72D, 0x00A72C},
+{0x00A72F, 0x00A72E},
+{0x00A733, 0x00A732},
+{0x00A735, 0x00A734},
+{0x00A737, 0x00A736},
+{0x00A739, 0x00A738},
+{0x00A73B, 0x00A73A},
+{0x00A73D, 0x00A73C},
+{0x00A73F, 0x00A73E},
+{0x00A741, 0x00A740},
+{0x00A743, 0x00A742},
+{0x00A745, 0x00A744},
+{0x00A747, 0x00A746},
+{0x00A749, 0x00A748},
+{0x00A74B, 0x00A74A},
+{0x00A74D, 0x00A74C},
+{0x00A74F, 0x00A74E},
+{0x00A751, 0x00A750},
+{0x00A753, 0x00A752},
+{0x00A755, 0x00A754},
+{0x00A757, 0x00A756},
+{0x00A759, 0x00A758},
+{0x00A75B, 0x00A75A},
+{0x00A75D, 0x00A75C},
+{0x00A75F, 0x00A75E},
+{0x00A761, 0x00A760},
+{0x00A763, 0x00A762},
+{0x00A765, 0x00A764},
+{0x00A767, 0x00A766},
+{0x00A769, 0x00A768},
+{0x00A76B, 0x00A76A},
+{0x00A76D, 0x00A76C},
+{0x00A76F, 0x00A76E},
+{0x00A77A, 0x00A779},
+{0x00A77C, 0x00A77B},
+{0x00A77F, 0x00A77E},
+{0x00A781, 0x00A780},
+{0x00A783, 0x00A782},
+{0x00A785, 0x00A784},
+{0x00A787, 0x00A786},
+{0x00A78C, 0x00A78B},
+{0x00A791, 0x00A790},
+{0x00A793, 0x00A792},
+{0x00A794, 0x00A7C4},
+{0x00A797, 0x00A796},
+{0x00A799, 0x00A798},
+{0x00A79B, 0x00A79A},
+{0x00A79D, 0x00A79C},
+{0x00A79F, 0x00A79E},
+{0x00A7A1, 0x00A7A0},
+{0x00A7A3, 0x00A7A2},
+{0x00A7A5, 0x00A7A4},
+{0x00A7A7, 0x00A7A6},
+{0x00A7A9, 0x00A7A8},
+{0x00A7B5, 0x00A7B4},
+{0x00A7B7, 0x00A7B6},
+{0x00A7B9, 0x00A7B8},
+{0x00A7BB, 0x00A7BA},
+{0x00A7BD, 0x00A7BC},
+{0x00A7BF, 0x00A7BE},
+{0x00A7C3, 0x00A7C2},
+{0x00A7C8, 0x00A7C7},
+{0x00A7CA, 0x00A7C9},
+{0x00A7F6, 0x00A7F5},
+{0x00AB53, 0x00A7B3},
+{0x00AB70, 0x0013A0},
+{0x00AB71, 0x0013A1},
+{0x00AB72, 0x0013A2},
+{0x00AB73, 0x0013A3},
+{0x00AB74, 0x0013A4},
+{0x00AB75, 0x0013A5},
+{0x00AB76, 0x0013A6},
+{0x00AB77, 0x0013A7},
+{0x00AB78, 0x0013A8},
+{0x00AB79, 0x0013A9},
+{0x00AB7A, 0x0013AA},
+{0x00AB7B, 0x0013AB},
+{0x00AB7C, 0x0013AC},
+{0x00AB7D, 0x0013AD},
+{0x00AB7E, 0x0013AE},
+{0x00AB7F, 0x0013AF},
+{0x00AB80, 0x0013B0},
+{0x00AB81, 0x0013B1},
+{0x00AB82, 0x0013B2},
+{0x00AB83, 0x0013B3},
+{0x00AB84, 0x0013B4},
+{0x00AB85, 0x0013B5},
+{0x00AB86, 0x0013B6},
+{0x00AB87, 0x0013B7},
+{0x00AB88, 0x0013B8},
+{0x00AB89, 0x0013B9},
+{0x00AB8A, 0x0013BA},
+{0x00AB8B, 0x0013BB},
+{0x00AB8C, 0x0013BC},
+{0x00AB8D, 0x0013BD},
+{0x00AB8E, 0x0013BE},
+{0x00AB8F, 0x0013BF},
+{0x00AB90, 0x0013C0},
+{0x00AB91, 0x0013C1},
+{0x00AB92, 0x0013C2},
+{0x00AB93, 0x0013C3},
+{0x00AB94, 0x0013C4},
+{0x00AB95, 0x0013C5},
+{0x00AB96, 0x0013C6},
+{0x00AB97, 0x0013C7},
+{0x00AB98, 0x0013C8},
+{0x00AB99, 0x0013C9},
+{0x00AB9A, 0x0013CA},
+{0x00AB9B, 0x0013CB},
+{0x00AB9C, 0x0013CC},
+{0x00AB9D, 0x0013CD},
+{0x00AB9E, 0x0013CE},
+{0x00AB9F, 0x0013CF},
+{0x00ABA0, 0x0013D0},
+{0x00ABA1, 0x0013D1},
+{0x00ABA2, 0x0013D2},
+{0x00ABA3, 0x0013D3},
+{0x00ABA4, 0x0013D4},
+{0x00ABA5, 0x0013D5},
+{0x00ABA6, 0x0013D6},
+{0x00ABA7, 0x0013D7},
+{0x00ABA8, 0x0013D8},
+{0x00ABA9, 0x0013D9},
+{0x00ABAA, 0x0013DA},
+{0x00ABAB, 0x0013DB},
+{0x00ABAC, 0x0013DC},
+{0x00ABAD, 0x0013DD},
+{0x00ABAE, 0x0013DE},
+{0x00ABAF, 0x0013DF},
+{0x00ABB0, 0x0013E0},
+{0x00ABB1, 0x0013E1},
+{0x00ABB2, 0x0013E2},
+{0x00ABB3, 0x0013E3},
+{0x00ABB4, 0x0013E4},
+{0x00ABB5, 0x0013E5},
+{0x00ABB6, 0x0013E6},
+{0x00ABB7, 0x0013E7},
+{0x00ABB8, 0x0013E8},
+{0x00ABB9, 0x0013E9},
+{0x00ABBA, 0x0013EA},
+{0x00ABBB, 0x0013EB},
+{0x00ABBC, 0x0013EC},
+{0x00ABBD, 0x0013ED},
+{0x00ABBE, 0x0013EE},
+{0x00ABBF, 0x0013EF},
+{0x00FB00, 0x000046},
+{0x00FB01, 0x000046},
+{0x00FB02, 0x000046},
+{0x00FB03, 0x000046},
+{0x00FB04, 0x000046},
+{0x00FB05, 0x000053},
+{0x00FB06, 0x000053},
+{0x00FB13, 0x000544},
+{0x00FB14, 0x000544},
+{0x00FB15, 0x000544},
+{0x00FB16, 0x00054E},
+{0x00FB17, 0x000544},
+{0x00FF41, 0x00FF21},
+{0x00FF42, 0x00FF22},
+{0x00FF43, 0x00FF23},
+{0x00FF44, 0x00FF24},
+{0x00FF45, 0x00FF25},
+{0x00FF46, 0x00FF26},
+{0x00FF47, 0x00FF27},
+{0x00FF48, 0x00FF28},
+{0x00FF49, 0x00FF29},
+{0x00FF4A, 0x00FF2A},
+{0x00FF4B, 0x00FF2B},
+{0x00FF4C, 0x00FF2C},
+{0x00FF4D, 0x00FF2D},
+{0x00FF4E, 0x00FF2E},
+{0x00FF4F, 0x00FF2F},
+{0x00FF50, 0x00FF30},
+{0x00FF51, 0x00FF31},
+{0x00FF52, 0x00FF32},
+{0x00FF53, 0x00FF33},
+{0x00FF54, 0x00FF34},
+{0x00FF55, 0x00FF35},
+{0x00FF56, 0x00FF36},
+{0x00FF57, 0x00FF37},
+{0x00FF58, 0x00FF38},
+{0x00FF59, 0x00FF39},
+{0x00FF5A, 0x00FF3A},
+{0x010428, 0x010400},
+{0x010429, 0x010401},
+{0x01042A, 0x010402},
+{0x01042B, 0x010403},
+{0x01042C, 0x010404},
+{0x01042D, 0x010405},
+{0x01042E, 0x010406},
+{0x01042F, 0x010407},
+{0x010430, 0x010408},
+{0x010431, 0x010409},
+{0x010432, 0x01040A},
+{0x010433, 0x01040B},
+{0x010434, 0x01040C},
+{0x010435, 0x01040D},
+{0x010436, 0x01040E},
+{0x010437, 0x01040F},
+{0x010438, 0x010410},
+{0x010439, 0x010411},
+{0x01043A, 0x010412},
+{0x01043B, 0x010413},
+{0x01043C, 0x010414},
+{0x01043D, 0x010415},
+{0x01043E, 0x010416},
+{0x01043F, 0x010417},
+{0x010440, 0x010418},
+{0x010441, 0x010419},
+{0x010442, 0x01041A},
+{0x010443, 0x01041B},
+{0x010444, 0x01041C},
+{0x010445, 0x01041D},
+{0x010446, 0x01041E},
+{0x010447, 0x01041F},
+{0x010448, 0x010420},
+{0x010449, 0x010421},
+{0x01044A, 0x010422},
+{0x01044B, 0x010423},
+{0x01044C, 0x010424},
+{0x01044D, 0x010425},
+{0x01044E, 0x010426},
+{0x01044F, 0x010427},
+{0x0104D8, 0x0104B0},
+{0x0104D9, 0x0104B1},
+{0x0104DA, 0x0104B2},
+{0x0104DB, 0x0104B3},
+{0x0104DC, 0x0104B4},
+{0x0104DD, 0x0104B5},
+{0x0104DE, 0x0104B6},
+{0x0104DF, 0x0104B7},
+{0x0104E0, 0x0104B8},
+{0x0104E1, 0x0104B9},
+{0x0104E2, 0x0104BA},
+{0x0104E3, 0x0104BB},
+{0x0104E4, 0x0104BC},
+{0x0104E5, 0x0104BD},
+{0x0104E6, 0x0104BE},
+{0x0104E7, 0x0104BF},
+{0x0104E8, 0x0104C0},
+{0x0104E9, 0x0104C1},
+{0x0104EA, 0x0104C2},
+{0x0104EB, 0x0104C3},
+{0x0104EC, 0x0104C4},
+{0x0104ED, 0x0104C5},
+{0x0104EE, 0x0104C6},
+{0x0104EF, 0x0104C7},
+{0x0104F0, 0x0104C8},
+{0x0104F1, 0x0104C9},
+{0x0104F2, 0x0104CA},
+{0x0104F3, 0x0104CB},
+{0x0104F4, 0x0104CC},
+{0x0104F5, 0x0104CD},
+{0x0104F6, 0x0104CE},
+{0x0104F7, 0x0104CF},
+{0x0104F8, 0x0104D0},
+{0x0104F9, 0x0104D1},
+{0x0104FA, 0x0104D2},
+{0x0104FB, 0x0104D3},
+{0x010CC0, 0x010C80},
+{0x010CC1, 0x010C81},
+{0x010CC2, 0x010C82},
+{0x010CC3, 0x010C83},
+{0x010CC4, 0x010C84},
+{0x010CC5, 0x010C85},
+{0x010CC6, 0x010C86},
+{0x010CC7, 0x010C87},
+{0x010CC8, 0x010C88},
+{0x010CC9, 0x010C89},
+{0x010CCA, 0x010C8A},
+{0x010CCB, 0x010C8B},
+{0x010CCC, 0x010C8C},
+{0x010CCD, 0x010C8D},
+{0x010CCE, 0x010C8E},
+{0x010CCF, 0x010C8F},
+{0x010CD0, 0x010C90},
+{0x010CD1, 0x010C91},
+{0x010CD2, 0x010C92},
+{0x010CD3, 0x010C93},
+{0x010CD4, 0x010C94},
+{0x010CD5, 0x010C95},
+{0x010CD6, 0x010C96},
+{0x010CD7, 0x010C97},
+{0x010CD8, 0x010C98},
+{0x010CD9, 0x010C99},
+{0x010CDA, 0x010C9A},
+{0x010CDB, 0x010C9B},
+{0x010CDC, 0x010C9C},
+{0x010CDD, 0x010C9D},
+{0x010CDE, 0x010C9E},
+{0x010CDF, 0x010C9F},
+{0x010CE0, 0x010CA0},
+{0x010CE1, 0x010CA1},
+{0x010CE2, 0x010CA2},
+{0x010CE3, 0x010CA3},
+{0x010CE4, 0x010CA4},
+{0x010CE5, 0x010CA5},
+{0x010CE6, 0x010CA6},
+{0x010CE7, 0x010CA7},
+{0x010CE8, 0x010CA8},
+{0x010CE9, 0x010CA9},
+{0x010CEA, 0x010CAA},
+{0x010CEB, 0x010CAB},
+{0x010CEC, 0x010CAC},
+{0x010CED, 0x010CAD},
+{0x010CEE, 0x010CAE},
+{0x010CEF, 0x010CAF},
+{0x010CF0, 0x010CB0},
+{0x010CF1, 0x010CB1},
+{0x010CF2, 0x010CB2},
+{0x0118C0, 0x0118A0},
+{0x0118C1, 0x0118A1},
+{0x0118C2, 0x0118A2},
+{0x0118C3, 0x0118A3},
+{0x0118C4, 0x0118A4},
+{0x0118C5, 0x0118A5},
+{0x0118C6, 0x0118A6},
+{0x0118C7, 0x0118A7},
+{0x0118C8, 0x0118A8},
+{0x0118C9, 0x0118A9},
+{0x0118CA, 0x0118AA},
+{0x0118CB, 0x0118AB},
+{0x0118CC, 0x0118AC},
+{0x0118CD, 0x0118AD},
+{0x0118CE, 0x0118AE},
+{0x0118CF, 0x0118AF},
+{0x0118D0, 0x0118B0},
+{0x0118D1, 0x0118B1},
+{0x0118D2, 0x0118B2},
+{0x0118D3, 0x0118B3},
+{0x0118D4, 0x0118B4},
+{0x0118D5, 0x0118B5},
+{0x0118D6, 0x0118B6},
+{0x0118D7, 0x0118B7},
+{0x0118D8, 0x0118B8},
+{0x0118D9, 0x0118B9},
+{0x0118DA, 0x0118BA},
+{0x0118DB, 0x0118BB},
+{0x0118DC, 0x0118BC},
+{0x0118DD, 0x0118BD},
+{0x0118DE, 0x0118BE},
+{0x0118DF, 0x0118BF},
+{0x016E60, 0x016E40},
+{0x016E61, 0x016E41},
+{0x016E62, 0x016E42},
+{0x016E63, 0x016E43},
+{0x016E64, 0x016E44},
+{0x016E65, 0x016E45},
+{0x016E66, 0x016E46},
+{0x016E67, 0x016E47},
+{0x016E68, 0x016E48},
+{0x016E69, 0x016E49},
+{0x016E6A, 0x016E4A},
+{0x016E6B, 0x016E4B},
+{0x016E6C, 0x016E4C},
+{0x016E6D, 0x016E4D},
+{0x016E6E, 0x016E4E},
+{0x016E6F, 0x016E4F},
+{0x016E70, 0x016E50},
+{0x016E71, 0x016E51},
+{0x016E72, 0x016E52},
+{0x016E73, 0x016E53},
+{0x016E74, 0x016E54},
+{0x016E75, 0x016E55},
+{0x016E76, 0x016E56},
+{0x016E77, 0x016E57},
+{0x016E78, 0x016E58},
+{0x016E79, 0x016E59},
+{0x016E7A, 0x016E5A},
+{0x016E7B, 0x016E5B},
+{0x016E7C, 0x016E5C},
+{0x016E7D, 0x016E5D},
+{0x016E7E, 0x016E5E},
+{0x016E7F, 0x016E5F},
+{0x01E922, 0x01E900},
+{0x01E923, 0x01E901},
+{0x01E924, 0x01E902},
+{0x01E925, 0x01E903},
+{0x01E926, 0x01E904},
+{0x01E927, 0x01E905},
+{0x01E928, 0x01E906},
+{0x01E929, 0x01E907},
+{0x01E92A, 0x01E908},
+{0x01E92B, 0x01E909},
+{0x01E92C, 0x01E90A},
+{0x01E92D, 0x01E90B},
+{0x01E92E, 0x01E90C},
+{0x01E92F, 0x01E90D},
+{0x01E930, 0x01E90E},
+{0x01E931, 0x01E90F},
+{0x01E932, 0x01E910},
+{0x01E933, 0x01E911},
+{0x01E934, 0x01E912},
+{0x01E935, 0x01E913},
+{0x01E936, 0x01E914},
+{0x01E937, 0x01E915},
+{0x01E938, 0x01E916},
+{0x01E939, 0x01E917},
+{0x01E93A, 0x01E918},
+{0x01E93B, 0x01E919},
+{0x01E93C, 0x01E91A},
+{0x01E93D, 0x01E91B},
+{0x01E93E, 0x01E91C},
+{0x01E93F, 0x01E91D},
+{0x01E940, 0x01E91E},
+{0x01E941, 0x01E91F},
+{0x01E942, 0x01E920},
+{0x01E943, 0x01E921},
 };
 
-const std::multimap<uint32_t, uint32_t> unicode_map_nfd = {
-{0x000000C0, 0x00000041}, {0x000000C0, 0x00000300}, {0x000000C1, 0x00000041}, {0x000000C1, 0x00000301},
-{0x000000C2, 0x00000041}, {0x000000C2, 0x00000302}, {0x000000C3, 0x00000041}, {0x000000C3, 0x00000303},
-{0x000000C4, 0x00000041}, {0x000000C4, 0x00000308}, {0x000000C5, 0x00000041}, {0x000000C5, 0x0000030A},
-{0x000000C7, 0x00000043}, {0x000000C7, 0x00000327}, {0x000000C8, 0x00000045}, {0x000000C8, 0x00000300},
-{0x000000C9, 0x00000045}, {0x000000C9, 0x00000301}, {0x000000CA, 0x00000045}, {0x000000CA, 0x00000302},
-{0x000000CB, 0x00000045}, {0x000000CB, 0x00000308}, {0x000000CC, 0x00000049}, {0x000000CC, 0x00000300},
-{0x000000CD, 0x00000049}, {0x000000CD, 0x00000301}, {0x000000CE, 0x00000049}, {0x000000CE, 0x00000302},
-{0x000000CF, 0x00000049}, {0x000000CF, 0x00000308}, {0x000000D1, 0x0000004E}, {0x000000D1, 0x00000303},
-{0x000000D2, 0x0000004F}, {0x000000D2, 0x00000300}, {0x000000D3, 0x0000004F}, {0x000000D3, 0x00000301},
-{0x000000D4, 0x0000004F}, {0x000000D4, 0x00000302}, {0x000000D5, 0x0000004F}, {0x000000D5, 0x00000303},
-{0x000000D6, 0x0000004F}, {0x000000D6, 0x00000308}, {0x000000D9, 0x00000055}, {0x000000D9, 0x00000300},
-{0x000000DA, 0x00000055}, {0x000000DA, 0x00000301}, {0x000000DB, 0x00000055}, {0x000000DB, 0x00000302},
-{0x000000DC, 0x00000055}, {0x000000DC, 0x00000308}, {0x000000DD, 0x00000059}, {0x000000DD, 0x00000301},
-{0x000000E0, 0x00000061}, {0x000000E0, 0x00000300}, {0x000000E1, 0x00000061}, {0x000000E1, 0x00000301},
-{0x000000E2, 0x00000061}, {0x000000E2, 0x00000302}, {0x000000E3, 0x00000061}, {0x000000E3, 0x00000303},
-{0x000000E4, 0x00000061}, {0x000000E4, 0x00000308}, {0x000000E5, 0x00000061}, {0x000000E5, 0x0000030A},
-{0x000000E7, 0x00000063}, {0x000000E7, 0x00000327}, {0x000000E8, 0x00000065}, {0x000000E8, 0x00000300},
-{0x000000E9, 0x00000065}, {0x000000E9, 0x00000301}, {0x000000EA, 0x00000065}, {0x000000EA, 0x00000302},
-{0x000000EB, 0x00000065}, {0x000000EB, 0x00000308}, {0x000000EC, 0x00000069}, {0x000000EC, 0x00000300},
-{0x000000ED, 0x00000069}, {0x000000ED, 0x00000301}, {0x000000EE, 0x00000069}, {0x000000EE, 0x00000302},
-{0x000000EF, 0x00000069}, {0x000000EF, 0x00000308}, {0x000000F1, 0x0000006E}, {0x000000F1, 0x00000303},
-{0x000000F2, 0x0000006F}, {0x000000F2, 0x00000300}, {0x000000F3, 0x0000006F}, {0x000000F3, 0x00000301},
-{0x000000F4, 0x0000006F}, {0x000000F4, 0x00000302}, {0x000000F5, 0x0000006F}, {0x000000F5, 0x00000303},
-{0x000000F6, 0x0000006F}, {0x000000F6, 0x00000308}, {0x000000F9, 0x00000075}, {0x000000F9, 0x00000300},
-{0x000000FA, 0x00000075}, {0x000000FA, 0x00000301}, {0x000000FB, 0x00000075}, {0x000000FB, 0x00000302},
-{0x000000FC, 0x00000075}, {0x000000FC, 0x00000308}, {0x000000FD, 0x00000079}, {0x000000FD, 0x00000301},
-{0x000000FF, 0x00000079}, {0x000000FF, 0x00000308}, {0x00000100, 0x00000041}, {0x00000100, 0x00000304},
-{0x00000101, 0x00000061}, {0x00000101, 0x00000304}, {0x00000102, 0x00000041}, {0x00000102, 0x00000306},
-{0x00000103, 0x00000061}, {0x00000103, 0x00000306}, {0x00000104, 0x00000041}, {0x00000104, 0x00000328},
-{0x00000105, 0x00000061}, {0x00000105, 0x00000328}, {0x00000106, 0x00000043}, {0x00000106, 0x00000301},
-{0x00000107, 0x00000063}, {0x00000107, 0x00000301}, {0x00000108, 0x00000043}, {0x00000108, 0x00000302},
-{0x00000109, 0x00000063}, {0x00000109, 0x00000302}, {0x0000010A, 0x00000043}, {0x0000010A, 0x00000307},
-{0x0000010B, 0x00000063}, {0x0000010B, 0x00000307}, {0x0000010C, 0x00000043}, {0x0000010C, 0x0000030C},
-{0x0000010D, 0x00000063}, {0x0000010D, 0x0000030C}, {0x0000010E, 0x00000044}, {0x0000010E, 0x0000030C},
-{0x0000010F, 0x00000064}, {0x0000010F, 0x0000030C}, {0x00000112, 0x00000045}, {0x00000112, 0x00000304},
-{0x00000113, 0x00000065}, {0x00000113, 0x00000304}, {0x00000114, 0x00000045}, {0x00000114, 0x00000306},
-{0x00000115, 0x00000065}, {0x00000115, 0x00000306}, {0x00000116, 0x00000045}, {0x00000116, 0x00000307},
-{0x00000117, 0x00000065}, {0x00000117, 0x00000307}, {0x00000118, 0x00000045}, {0x00000118, 0x00000328},
-{0x00000119, 0x00000065}, {0x00000119, 0x00000328}, {0x0000011A, 0x00000045}, {0x0000011A, 0x0000030C},
-{0x0000011B, 0x00000065}, {0x0000011B, 0x0000030C}, {0x0000011C, 0x00000047}, {0x0000011C, 0x00000302},
-{0x0000011D, 0x00000067}, {0x0000011D, 0x00000302}, {0x0000011E, 0x00000047}, {0x0000011E, 0x00000306},
-{0x0000011F, 0x00000067}, {0x0000011F, 0x00000306}, {0x00000120, 0x00000047}, {0x00000120, 0x00000307},
-{0x00000121, 0x00000067}, {0x00000121, 0x00000307}, {0x00000122, 0x00000047}, {0x00000122, 0x00000327},
-{0x00000123, 0x00000067}, {0x00000123, 0x00000327}, {0x00000124, 0x00000048}, {0x00000124, 0x00000302},
-{0x00000125, 0x00000068}, {0x00000125, 0x00000302}, {0x00000128, 0x00000049}, {0x00000128, 0x00000303},
-{0x00000129, 0x00000069}, {0x00000129, 0x00000303}, {0x0000012A, 0x00000049}, {0x0000012A, 0x00000304},
-{0x0000012B, 0x00000069}, {0x0000012B, 0x00000304}, {0x0000012C, 0x00000049}, {0x0000012C, 0x00000306},
-{0x0000012D, 0x00000069}, {0x0000012D, 0x00000306}, {0x0000012E, 0x00000049}, {0x0000012E, 0x00000328},
-{0x0000012F, 0x00000069}, {0x0000012F, 0x00000328}, {0x00000130, 0x00000049}, {0x00000130, 0x00000307},
-{0x00000134, 0x0000004A}, {0x00000134, 0x00000302}, {0x00000135, 0x0000006A}, {0x00000135, 0x00000302},
-{0x00000136, 0x0000004B}, {0x00000136, 0x00000327}, {0x00000137, 0x0000006B}, {0x00000137, 0x00000327},
-{0x00000139, 0x0000004C}, {0x00000139, 0x00000301}, {0x0000013A, 0x0000006C}, {0x0000013A, 0x00000301},
-{0x0000013B, 0x0000004C}, {0x0000013B, 0x00000327}, {0x0000013C, 0x0000006C}, {0x0000013C, 0x00000327},
-{0x0000013D, 0x0000004C}, {0x0000013D, 0x0000030C}, {0x0000013E, 0x0000006C}, {0x0000013E, 0x0000030C},
-{0x00000143, 0x0000004E}, {0x00000143, 0x00000301}, {0x00000144, 0x0000006E}, {0x00000144, 0x00000301},
-{0x00000145, 0x0000004E}, {0x00000145, 0x00000327}, {0x00000146, 0x0000006E}, {0x00000146, 0x00000327},
-{0x00000147, 0x0000004E}, {0x00000147, 0x0000030C}, {0x00000148, 0x0000006E}, {0x00000148, 0x0000030C},
-{0x0000014C, 0x0000004F}, {0x0000014C, 0x00000304}, {0x0000014D, 0x0000006F}, {0x0000014D, 0x00000304},
-{0x0000014E, 0x0000004F}, {0x0000014E, 0x00000306}, {0x0000014F, 0x0000006F}, {0x0000014F, 0x00000306},
-{0x00000150, 0x0000004F}, {0x00000150, 0x0000030B}, {0x00000151, 0x0000006F}, {0x00000151, 0x0000030B},
-{0x00000154, 0x00000052}, {0x00000154, 0x00000301}, {0x00000155, 0x00000072}, {0x00000155, 0x00000301},
-{0x00000156, 0x00000052}, {0x00000156, 0x00000327}, {0x00000157, 0x00000072}, {0x00000157, 0x00000327},
-{0x00000158, 0x00000052}, {0x00000158, 0x0000030C}, {0x00000159, 0x00000072}, {0x00000159, 0x0000030C},
-{0x0000015A, 0x00000053}, {0x0000015A, 0x00000301}, {0x0000015B, 0x00000073}, {0x0000015B, 0x00000301},
-{0x0000015C, 0x00000053}, {0x0000015C, 0x00000302}, {0x0000015D, 0x00000073}, {0x0000015D, 0x00000302},
-{0x0000015E, 0x00000053}, {0x0000015E, 0x00000327}, {0x0000015F, 0x00000073}, {0x0000015F, 0x00000327},
-{0x00000160, 0x00000053}, {0x00000160, 0x0000030C}, {0x00000161, 0x00000073}, {0x00000161, 0x0000030C},
-{0x00000162, 0x00000054}, {0x00000162, 0x00000327}, {0x00000163, 0x00000074}, {0x00000163, 0x00000327},
-{0x00000164, 0x00000054}, {0x00000164, 0x0000030C}, {0x00000165, 0x00000074}, {0x00000165, 0x0000030C},
-{0x00000168, 0x00000055}, {0x00000168, 0x00000303}, {0x00000169, 0x00000075}, {0x00000169, 0x00000303},
-{0x0000016A, 0x00000055}, {0x0000016A, 0x00000304}, {0x0000016B, 0x00000075}, {0x0000016B, 0x00000304},
-{0x0000016C, 0x00000055}, {0x0000016C, 0x00000306}, {0x0000016D, 0x00000075}, {0x0000016D, 0x00000306},
-{0x0000016E, 0x00000055}, {0x0000016E, 0x0000030A}, {0x0000016F, 0x00000075}, {0x0000016F, 0x0000030A},
-{0x00000170, 0x00000055}, {0x00000170, 0x0000030B}, {0x00000171, 0x00000075}, {0x00000171, 0x0000030B},
-{0x00000172, 0x00000055}, {0x00000172, 0x00000328}, {0x00000173, 0x00000075}, {0x00000173, 0x00000328},
-{0x00000174, 0x00000057}, {0x00000174, 0x00000302}, {0x00000175, 0x00000077}, {0x00000175, 0x00000302},
-{0x00000176, 0x00000059}, {0x00000176, 0x00000302}, {0x00000177, 0x00000079}, {0x00000177, 0x00000302},
-{0x00000178, 0x00000059}, {0x00000178, 0x00000308}, {0x00000179, 0x0000005A}, {0x00000179, 0x00000301},
-{0x0000017A, 0x0000007A}, {0x0000017A, 0x00000301}, {0x0000017B, 0x0000005A}, {0x0000017B, 0x00000307},
-{0x0000017C, 0x0000007A}, {0x0000017C, 0x00000307}, {0x0000017D, 0x0000005A}, {0x0000017D, 0x0000030C},
-{0x0000017E, 0x0000007A}, {0x0000017E, 0x0000030C}, {0x000001A0, 0x0000004F}, {0x000001A0, 0x0000031B},
-{0x000001A1, 0x0000006F}, {0x000001A1, 0x0000031B}, {0x000001AF, 0x00000055}, {0x000001AF, 0x0000031B},
-{0x000001B0, 0x00000075}, {0x000001B0, 0x0000031B}, {0x000001CD, 0x00000041}, {0x000001CD, 0x0000030C},
-{0x000001CE, 0x00000061}, {0x000001CE, 0x0000030C}, {0x000001CF, 0x00000049}, {0x000001CF, 0x0000030C},
-{0x000001D0, 0x00000069}, {0x000001D0, 0x0000030C}, {0x000001D1, 0x0000004F}, {0x000001D1, 0x0000030C},
-{0x000001D2, 0x0000006F}, {0x000001D2, 0x0000030C}, {0x000001D3, 0x00000055}, {0x000001D3, 0x0000030C},
-{0x000001D4, 0x00000075}, {0x000001D4, 0x0000030C}, {0x000001D5, 0x00000055}, {0x000001D5, 0x00000308},
-{0x000001D5, 0x00000304}, {0x000001D6, 0x00000075}, {0x000001D6, 0x00000308}, {0x000001D6, 0x00000304},
-{0x000001D7, 0x00000055}, {0x000001D7, 0x00000308}, {0x000001D7, 0x00000301}, {0x000001D8, 0x00000075},
-{0x000001D8, 0x00000308}, {0x000001D8, 0x00000301}, {0x000001D9, 0x00000055}, {0x000001D9, 0x00000308},
-{0x000001D9, 0x0000030C}, {0x000001DA, 0x00000075}, {0x000001DA, 0x00000308}, {0x000001DA, 0x0000030C},
-{0x000001DB, 0x00000055}, {0x000001DB, 0x00000308}, {0x000001DB, 0x00000300}, {0x000001DC, 0x00000075},
-{0x000001DC, 0x00000308}, {0x000001DC, 0x00000300}, {0x000001DE, 0x00000041}, {0x000001DE, 0x00000308},
-{0x000001DE, 0x00000304}, {0x000001DF, 0x00000061}, {0x000001DF, 0x00000308}, {0x000001DF, 0x00000304},
-{0x000001E0, 0x00000041}, {0x000001E0, 0x00000307}, {0x000001E0, 0x00000304}, {0x000001E1, 0x00000061},
-{0x000001E1, 0x00000307}, {0x000001E1, 0x00000304}, {0x000001E2, 0x000000C6}, {0x000001E2, 0x00000304},
-{0x000001E3, 0x000000E6}, {0x000001E3, 0x00000304}, {0x000001E6, 0x00000047}, {0x000001E6, 0x0000030C},
-{0x000001E7, 0x00000067}, {0x000001E7, 0x0000030C}, {0x000001E8, 0x0000004B}, {0x000001E8, 0x0000030C},
-{0x000001E9, 0x0000006B}, {0x000001E9, 0x0000030C}, {0x000001EA, 0x0000004F}, {0x000001EA, 0x00000328},
-{0x000001EB, 0x0000006F}, {0x000001EB, 0x00000328}, {0x000001EC, 0x0000004F}, {0x000001EC, 0x00000328},
-{0x000001EC, 0x00000304}, {0x000001ED, 0x0000006F}, {0x000001ED, 0x00000328}, {0x000001ED, 0x00000304},
-{0x000001EE, 0x000001B7}, {0x000001EE, 0x0000030C}, {0x000001EF, 0x00000292}, {0x000001EF, 0x0000030C},
-{0x000001F0, 0x0000006A}, {0x000001F0, 0x0000030C}, {0x000001F4, 0x00000047}, {0x000001F4, 0x00000301},
-{0x000001F5, 0x00000067}, {0x000001F5, 0x00000301}, {0x000001F8, 0x0000004E}, {0x000001F8, 0x00000300},
-{0x000001F9, 0x0000006E}, {0x000001F9, 0x00000300}, {0x000001FA, 0x00000041}, {0x000001FA, 0x0000030A},
-{0x000001FA, 0x00000301}, {0x000001FB, 0x00000061}, {0x000001FB, 0x0000030A}, {0x000001FB, 0x00000301},
-{0x000001FC, 0x000000C6}, {0x000001FC, 0x00000301}, {0x000001FD, 0x000000E6}, {0x000001FD, 0x00000301},
-{0x000001FE, 0x000000D8}, {0x000001FE, 0x00000301}, {0x000001FF, 0x000000F8}, {0x000001FF, 0x00000301},
-{0x00000200, 0x00000041}, {0x00000200, 0x0000030F}, {0x00000201, 0x00000061}, {0x00000201, 0x0000030F},
-{0x00000202, 0x00000041}, {0x00000202, 0x00000311}, {0x00000203, 0x00000061}, {0x00000203, 0x00000311},
-{0x00000204, 0x00000045}, {0x00000204, 0x0000030F}, {0x00000205, 0x00000065}, {0x00000205, 0x0000030F},
-{0x00000206, 0x00000045}, {0x00000206, 0x00000311}, {0x00000207, 0x00000065}, {0x00000207, 0x00000311},
-{0x00000208, 0x00000049}, {0x00000208, 0x0000030F}, {0x00000209, 0x00000069}, {0x00000209, 0x0000030F},
-{0x0000020A, 0x00000049}, {0x0000020A, 0x00000311}, {0x0000020B, 0x00000069}, {0x0000020B, 0x00000311},
-{0x0000020C, 0x0000004F}, {0x0000020C, 0x0000030F}, {0x0000020D, 0x0000006F}, {0x0000020D, 0x0000030F},
-{0x0000020E, 0x0000004F}, {0x0000020E, 0x00000311}, {0x0000020F, 0x0000006F}, {0x0000020F, 0x00000311},
-{0x00000210, 0x00000052}, {0x00000210, 0x0000030F}, {0x00000211, 0x00000072}, {0x00000211, 0x0000030F},
-{0x00000212, 0x00000052}, {0x00000212, 0x00000311}, {0x00000213, 0x00000072}, {0x00000213, 0x00000311},
-{0x00000214, 0x00000055}, {0x00000214, 0x0000030F}, {0x00000215, 0x00000075}, {0x00000215, 0x0000030F},
-{0x00000216, 0x00000055}, {0x00000216, 0x00000311}, {0x00000217, 0x00000075}, {0x00000217, 0x00000311},
-{0x00000218, 0x00000053}, {0x00000218, 0x00000326}, {0x00000219, 0x00000073}, {0x00000219, 0x00000326},
-{0x0000021A, 0x00000054}, {0x0000021A, 0x00000326}, {0x0000021B, 0x00000074}, {0x0000021B, 0x00000326},
-{0x0000021E, 0x00000048}, {0x0000021E, 0x0000030C}, {0x0000021F, 0x00000068}, {0x0000021F, 0x0000030C},
-{0x00000226, 0x00000041}, {0x00000226, 0x00000307}, {0x00000227, 0x00000061}, {0x00000227, 0x00000307},
-{0x00000228, 0x00000045}, {0x00000228, 0x00000327}, {0x00000229, 0x00000065}, {0x00000229, 0x00000327},
-{0x0000022A, 0x0000004F}, {0x0000022A, 0x00000308}, {0x0000022A, 0x00000304}, {0x0000022B, 0x0000006F},
-{0x0000022B, 0x00000308}, {0x0000022B, 0x00000304}, {0x0000022C, 0x0000004F}, {0x0000022C, 0x00000303},
-{0x0000022C, 0x00000304}, {0x0000022D, 0x0000006F}, {0x0000022D, 0x00000303}, {0x0000022D, 0x00000304},
-{0x0000022E, 0x0000004F}, {0x0000022E, 0x00000307}, {0x0000022F, 0x0000006F}, {0x0000022F, 0x00000307},
-{0x00000230, 0x0000004F}, {0x00000230, 0x00000307}, {0x00000230, 0x00000304}, {0x00000231, 0x0000006F},
-{0x00000231, 0x00000307}, {0x00000231, 0x00000304}, {0x00000232, 0x00000059}, {0x00000232, 0x00000304},
-{0x00000233, 0x00000079}, {0x00000233, 0x00000304}, {0x00000340, 0x00000300}, {0x00000341, 0x00000301},
-{0x00000343, 0x00000313}, {0x00000344, 0x00000308}, {0x00000344, 0x00000301}, {0x00000374, 0x000002B9},
-{0x0000037E, 0x0000003B}, {0x00000385, 0x000000A8}, {0x00000385, 0x00000301}, {0x00000386, 0x00000391},
-{0x00000386, 0x00000301}, {0x00000387, 0x000000B7}, {0x00000388, 0x00000395}, {0x00000388, 0x00000301},
-{0x00000389, 0x00000397}, {0x00000389, 0x00000301}, {0x0000038A, 0x00000399}, {0x0000038A, 0x00000301},
-{0x0000038C, 0x0000039F}, {0x0000038C, 0x00000301}, {0x0000038E, 0x000003A5}, {0x0000038E, 0x00000301},
-{0x0000038F, 0x000003A9}, {0x0000038F, 0x00000301}, {0x00000390, 0x000003B9}, {0x00000390, 0x00000308},
-{0x00000390, 0x00000301}, {0x000003AA, 0x00000399}, {0x000003AA, 0x00000308}, {0x000003AB, 0x000003A5},
-{0x000003AB, 0x00000308}, {0x000003AC, 0x000003B1}, {0x000003AC, 0x00000301}, {0x000003AD, 0x000003B5},
-{0x000003AD, 0x00000301}, {0x000003AE, 0x000003B7}, {0x000003AE, 0x00000301}, {0x000003AF, 0x000003B9},
-{0x000003AF, 0x00000301}, {0x000003B0, 0x000003C5}, {0x000003B0, 0x00000308}, {0x000003B0, 0x00000301},
-{0x000003CA, 0x000003B9}, {0x000003CA, 0x00000308}, {0x000003CB, 0x000003C5}, {0x000003CB, 0x00000308},
-{0x000003CC, 0x000003BF}, {0x000003CC, 0x00000301}, {0x000003CD, 0x000003C5}, {0x000003CD, 0x00000301},
-{0x000003CE, 0x000003C9}, {0x000003CE, 0x00000301}, {0x000003D3, 0x000003D2}, {0x000003D3, 0x00000301},
-{0x000003D4, 0x000003D2}, {0x000003D4, 0x00000308}, {0x00000400, 0x00000415}, {0x00000400, 0x00000300},
-{0x00000401, 0x00000415}, {0x00000401, 0x00000308}, {0x00000403, 0x00000413}, {0x00000403, 0x00000301},
-{0x00000407, 0x00000406}, {0x00000407, 0x00000308}, {0x0000040C, 0x0000041A}, {0x0000040C, 0x00000301},
-{0x0000040D, 0x00000418}, {0x0000040D, 0x00000300}, {0x0000040E, 0x00000423}, {0x0000040E, 0x00000306},
-{0x00000419, 0x00000418}, {0x00000419, 0x00000306}, {0x00000439, 0x00000438}, {0x00000439, 0x00000306},
-{0x00000450, 0x00000435}, {0x00000450, 0x00000300}, {0x00000451, 0x00000435}, {0x00000451, 0x00000308},
-{0x00000453, 0x00000433}, {0x00000453, 0x00000301}, {0x00000457, 0x00000456}, {0x00000457, 0x00000308},
-{0x0000045C, 0x0000043A}, {0x0000045C, 0x00000301}, {0x0000045D, 0x00000438}, {0x0000045D, 0x00000300},
-{0x0000045E, 0x00000443}, {0x0000045E, 0x00000306}, {0x00000476, 0x00000474}, {0x00000476, 0x0000030F},
-{0x00000477, 0x00000475}, {0x00000477, 0x0000030F}, {0x000004C1, 0x00000416}, {0x000004C1, 0x00000306},
-{0x000004C2, 0x00000436}, {0x000004C2, 0x00000306}, {0x000004D0, 0x00000410}, {0x000004D0, 0x00000306},
-{0x000004D1, 0x00000430}, {0x000004D1, 0x00000306}, {0x000004D2, 0x00000410}, {0x000004D2, 0x00000308},
-{0x000004D3, 0x00000430}, {0x000004D3, 0x00000308}, {0x000004D6, 0x00000415}, {0x000004D6, 0x00000306},
-{0x000004D7, 0x00000435}, {0x000004D7, 0x00000306}, {0x000004DA, 0x000004D8}, {0x000004DA, 0x00000308},
-{0x000004DB, 0x000004D9}, {0x000004DB, 0x00000308}, {0x000004DC, 0x00000416}, {0x000004DC, 0x00000308},
-{0x000004DD, 0x00000436}, {0x000004DD, 0x00000308}, {0x000004DE, 0x00000417}, {0x000004DE, 0x00000308},
-{0x000004DF, 0x00000437}, {0x000004DF, 0x00000308}, {0x000004E2, 0x00000418}, {0x000004E2, 0x00000304},
-{0x000004E3, 0x00000438}, {0x000004E3, 0x00000304}, {0x000004E4, 0x00000418}, {0x000004E4, 0x00000308},
-{0x000004E5, 0x00000438}, {0x000004E5, 0x00000308}, {0x000004E6, 0x0000041E}, {0x000004E6, 0x00000308},
-{0x000004E7, 0x0000043E}, {0x000004E7, 0x00000308}, {0x000004EA, 0x000004E8}, {0x000004EA, 0x00000308},
-{0x000004EB, 0x000004E9}, {0x000004EB, 0x00000308}, {0x000004EC, 0x0000042D}, {0x000004EC, 0x00000308},
-{0x000004ED, 0x0000044D}, {0x000004ED, 0x00000308}, {0x000004EE, 0x00000423}, {0x000004EE, 0x00000304},
-{0x000004EF, 0x00000443}, {0x000004EF, 0x00000304}, {0x000004F0, 0x00000423}, {0x000004F0, 0x00000308},
-{0x000004F1, 0x00000443}, {0x000004F1, 0x00000308}, {0x000004F2, 0x00000423}, {0x000004F2, 0x0000030B},
-{0x000004F3, 0x00000443}, {0x000004F3, 0x0000030B}, {0x000004F4, 0x00000427}, {0x000004F4, 0x00000308},
-{0x000004F5, 0x00000447}, {0x000004F5, 0x00000308}, {0x000004F8, 0x0000042B}, {0x000004F8, 0x00000308},
-{0x000004F9, 0x0000044B}, {0x000004F9, 0x00000308}, {0x00000622, 0x00000627}, {0x00000622, 0x00000653},
-{0x00000623, 0x00000627}, {0x00000623, 0x00000654}, {0x00000624, 0x00000648}, {0x00000624, 0x00000654},
-{0x00000625, 0x00000627}, {0x00000625, 0x00000655}, {0x00000626, 0x0000064A}, {0x00000626, 0x00000654},
-{0x000006C0, 0x000006D5}, {0x000006C0, 0x00000654}, {0x000006C2, 0x000006C1}, {0x000006C2, 0x00000654},
-{0x000006D3, 0x000006D2}, {0x000006D3, 0x00000654}, {0x00000929, 0x00000928}, {0x00000929, 0x0000093C},
-{0x00000931, 0x00000930}, {0x00000931, 0x0000093C}, {0x00000934, 0x00000933}, {0x00000934, 0x0000093C},
-{0x00000958, 0x00000915}, {0x00000958, 0x0000093C}, {0x00000959, 0x00000916}, {0x00000959, 0x0000093C},
-{0x0000095A, 0x00000917}, {0x0000095A, 0x0000093C}, {0x0000095B, 0x0000091C}, {0x0000095B, 0x0000093C},
-{0x0000095C, 0x00000921}, {0x0000095C, 0x0000093C}, {0x0000095D, 0x00000922}, {0x0000095D, 0x0000093C},
-{0x0000095E, 0x0000092B}, {0x0000095E, 0x0000093C}, {0x0000095F, 0x0000092F}, {0x0000095F, 0x0000093C},
-{0x000009CB, 0x000009C7}, {0x000009CB, 0x000009BE}, {0x000009CC, 0x000009C7}, {0x000009CC, 0x000009D7},
-{0x000009DC, 0x000009A1}, {0x000009DC, 0x000009BC}, {0x000009DD, 0x000009A2}, {0x000009DD, 0x000009BC},
-{0x000009DF, 0x000009AF}, {0x000009DF, 0x000009BC}, {0x00000A33, 0x00000A32}, {0x00000A33, 0x00000A3C},
-{0x00000A36, 0x00000A38}, {0x00000A36, 0x00000A3C}, {0x00000A59, 0x00000A16}, {0x00000A59, 0x00000A3C},
-{0x00000A5A, 0x00000A17}, {0x00000A5A, 0x00000A3C}, {0x00000A5B, 0x00000A1C}, {0x00000A5B, 0x00000A3C},
-{0x00000A5E, 0x00000A2B}, {0x00000A5E, 0x00000A3C}, {0x00000B48, 0x00000B47}, {0x00000B48, 0x00000B56},
-{0x00000B4B, 0x00000B47}, {0x00000B4B, 0x00000B3E}, {0x00000B4C, 0x00000B47}, {0x00000B4C, 0x00000B57},
-{0x00000B5C, 0x00000B21}, {0x00000B5C, 0x00000B3C}, {0x00000B5D, 0x00000B22}, {0x00000B5D, 0x00000B3C},
-{0x00000B94, 0x00000B92}, {0x00000B94, 0x00000BD7}, {0x00000BCA, 0x00000BC6}, {0x00000BCA, 0x00000BBE},
-{0x00000BCB, 0x00000BC7}, {0x00000BCB, 0x00000BBE}, {0x00000BCC, 0x00000BC6}, {0x00000BCC, 0x00000BD7},
-{0x00000C48, 0x00000C46}, {0x00000C48, 0x00000C56}, {0x00000CC0, 0x00000CBF}, {0x00000CC0, 0x00000CD5},
-{0x00000CC7, 0x00000CC6}, {0x00000CC7, 0x00000CD5}, {0x00000CC8, 0x00000CC6}, {0x00000CC8, 0x00000CD6},
-{0x00000CCA, 0x00000CC6}, {0x00000CCA, 0x00000CC2}, {0x00000CCB, 0x00000CC6}, {0x00000CCB, 0x00000CC2},
-{0x00000CCB, 0x00000CD5}, {0x00000D4A, 0x00000D46}, {0x00000D4A, 0x00000D3E}, {0x00000D4B, 0x00000D47},
-{0x00000D4B, 0x00000D3E}, {0x00000D4C, 0x00000D46}, {0x00000D4C, 0x00000D57}, {0x00000DDA, 0x00000DD9},
-{0x00000DDA, 0x00000DCA}, {0x00000DDC, 0x00000DD9}, {0x00000DDC, 0x00000DCF}, {0x00000DDD, 0x00000DD9},
-{0x00000DDD, 0x00000DCF}, {0x00000DDD, 0x00000DCA}, {0x00000DDE, 0x00000DD9}, {0x00000DDE, 0x00000DDF},
-{0x00000F43, 0x00000F42}, {0x00000F43, 0x00000FB7}, {0x00000F4D, 0x00000F4C}, {0x00000F4D, 0x00000FB7},
-{0x00000F52, 0x00000F51}, {0x00000F52, 0x00000FB7}, {0x00000F57, 0x00000F56}, {0x00000F57, 0x00000FB7},
-{0x00000F5C, 0x00000F5B}, {0x00000F5C, 0x00000FB7}, {0x00000F69, 0x00000F40}, {0x00000F69, 0x00000FB5},
-{0x00000F73, 0x00000F71}, {0x00000F73, 0x00000F72}, {0x00000F75, 0x00000F71}, {0x00000F75, 0x00000F74},
-{0x00000F76, 0x00000FB2}, {0x00000F76, 0x00000F80}, {0x00000F78, 0x00000FB3}, {0x00000F78, 0x00000F80},
-{0x00000F81, 0x00000F71}, {0x00000F81, 0x00000F80}, {0x00000F93, 0x00000F92}, {0x00000F93, 0x00000FB7},
-{0x00000F9D, 0x00000F9C}, {0x00000F9D, 0x00000FB7}, {0x00000FA2, 0x00000FA1}, {0x00000FA2, 0x00000FB7},
-{0x00000FA7, 0x00000FA6}, {0x00000FA7, 0x00000FB7}, {0x00000FAC, 0x00000FAB}, {0x00000FAC, 0x00000FB7},
-{0x00000FB9, 0x00000F90}, {0x00000FB9, 0x00000FB5}, {0x00001026, 0x00001025}, {0x00001026, 0x0000102E},
-{0x00001B06, 0x00001B05}, {0x00001B06, 0x00001B35}, {0x00001B08, 0x00001B07}, {0x00001B08, 0x00001B35},
-{0x00001B0A, 0x00001B09}, {0x00001B0A, 0x00001B35}, {0x00001B0C, 0x00001B0B}, {0x00001B0C, 0x00001B35},
-{0x00001B0E, 0x00001B0D}, {0x00001B0E, 0x00001B35}, {0x00001B12, 0x00001B11}, {0x00001B12, 0x00001B35},
-{0x00001B3B, 0x00001B3A}, {0x00001B3B, 0x00001B35}, {0x00001B3D, 0x00001B3C}, {0x00001B3D, 0x00001B35},
-{0x00001B40, 0x00001B3E}, {0x00001B40, 0x00001B35}, {0x00001B41, 0x00001B3F}, {0x00001B41, 0x00001B35},
-{0x00001B43, 0x00001B42}, {0x00001B43, 0x00001B35}, {0x00001E00, 0x00000041}, {0x00001E00, 0x00000325},
-{0x00001E01, 0x00000061}, {0x00001E01, 0x00000325}, {0x00001E02, 0x00000042}, {0x00001E02, 0x00000307},
-{0x00001E03, 0x00000062}, {0x00001E03, 0x00000307}, {0x00001E04, 0x00000042}, {0x00001E04, 0x00000323},
-{0x00001E05, 0x00000062}, {0x00001E05, 0x00000323}, {0x00001E06, 0x00000042}, {0x00001E06, 0x00000331},
-{0x00001E07, 0x00000062}, {0x00001E07, 0x00000331}, {0x00001E08, 0x00000043}, {0x00001E08, 0x00000327},
-{0x00001E08, 0x00000301}, {0x00001E09, 0x00000063}, {0x00001E09, 0x00000327}, {0x00001E09, 0x00000301},
-{0x00001E0A, 0x00000044}, {0x00001E0A, 0x00000307}, {0x00001E0B, 0x00000064}, {0x00001E0B, 0x00000307},
-{0x00001E0C, 0x00000044}, {0x00001E0C, 0x00000323}, {0x00001E0D, 0x00000064}, {0x00001E0D, 0x00000323},
-{0x00001E0E, 0x00000044}, {0x00001E0E, 0x00000331}, {0x00001E0F, 0x00000064}, {0x00001E0F, 0x00000331},
-{0x00001E10, 0x00000044}, {0x00001E10, 0x00000327}, {0x00001E11, 0x00000064}, {0x00001E11, 0x00000327},
-{0x00001E12, 0x00000044}, {0x00001E12, 0x0000032D}, {0x00001E13, 0x00000064}, {0x00001E13, 0x0000032D},
-{0x00001E14, 0x00000045}, {0x00001E14, 0x00000304}, {0x00001E14, 0x00000300}, {0x00001E15, 0x00000065},
-{0x00001E15, 0x00000304}, {0x00001E15, 0x00000300}, {0x00001E16, 0x00000045}, {0x00001E16, 0x00000304},
-{0x00001E16, 0x00000301}, {0x00001E17, 0x00000065}, {0x00001E17, 0x00000304}, {0x00001E17, 0x00000301},
-{0x00001E18, 0x00000045}, {0x00001E18, 0x0000032D}, {0x00001E19, 0x00000065}, {0x00001E19, 0x0000032D},
-{0x00001E1A, 0x00000045}, {0x00001E1A, 0x00000330}, {0x00001E1B, 0x00000065}, {0x00001E1B, 0x00000330},
-{0x00001E1C, 0x00000045}, {0x00001E1C, 0x00000327}, {0x00001E1C, 0x00000306}, {0x00001E1D, 0x00000065},
-{0x00001E1D, 0x00000327}, {0x00001E1D, 0x00000306}, {0x00001E1E, 0x00000046}, {0x00001E1E, 0x00000307},
-{0x00001E1F, 0x00000066}, {0x00001E1F, 0x00000307}, {0x00001E20, 0x00000047}, {0x00001E20, 0x00000304},
-{0x00001E21, 0x00000067}, {0x00001E21, 0x00000304}, {0x00001E22, 0x00000048}, {0x00001E22, 0x00000307},
-{0x00001E23, 0x00000068}, {0x00001E23, 0x00000307}, {0x00001E24, 0x00000048}, {0x00001E24, 0x00000323},
-{0x00001E25, 0x00000068}, {0x00001E25, 0x00000323}, {0x00001E26, 0x00000048}, {0x00001E26, 0x00000308},
-{0x00001E27, 0x00000068}, {0x00001E27, 0x00000308}, {0x00001E28, 0x00000048}, {0x00001E28, 0x00000327},
-{0x00001E29, 0x00000068}, {0x00001E29, 0x00000327}, {0x00001E2A, 0x00000048}, {0x00001E2A, 0x0000032E},
-{0x00001E2B, 0x00000068}, {0x00001E2B, 0x0000032E}, {0x00001E2C, 0x00000049}, {0x00001E2C, 0x00000330},
-{0x00001E2D, 0x00000069}, {0x00001E2D, 0x00000330}, {0x00001E2E, 0x00000049}, {0x00001E2E, 0x00000308},
-{0x00001E2E, 0x00000301}, {0x00001E2F, 0x00000069}, {0x00001E2F, 0x00000308}, {0x00001E2F, 0x00000301},
-{0x00001E30, 0x0000004B}, {0x00001E30, 0x00000301}, {0x00001E31, 0x0000006B}, {0x00001E31, 0x00000301},
-{0x00001E32, 0x0000004B}, {0x00001E32, 0x00000323}, {0x00001E33, 0x0000006B}, {0x00001E33, 0x00000323},
-{0x00001E34, 0x0000004B}, {0x00001E34, 0x00000331}, {0x00001E35, 0x0000006B}, {0x00001E35, 0x00000331},
-{0x00001E36, 0x0000004C}, {0x00001E36, 0x00000323}, {0x00001E37, 0x0000006C}, {0x00001E37, 0x00000323},
-{0x00001E38, 0x0000004C}, {0x00001E38, 0x00000323}, {0x00001E38, 0x00000304}, {0x00001E39, 0x0000006C},
-{0x00001E39, 0x00000323}, {0x00001E39, 0x00000304}, {0x00001E3A, 0x0000004C}, {0x00001E3A, 0x00000331},
-{0x00001E3B, 0x0000006C}, {0x00001E3B, 0x00000331}, {0x00001E3C, 0x0000004C}, {0x00001E3C, 0x0000032D},
-{0x00001E3D, 0x0000006C}, {0x00001E3D, 0x0000032D}, {0x00001E3E, 0x0000004D}, {0x00001E3E, 0x00000301},
-{0x00001E3F, 0x0000006D}, {0x00001E3F, 0x00000301}, {0x00001E40, 0x0000004D}, {0x00001E40, 0x00000307},
-{0x00001E41, 0x0000006D}, {0x00001E41, 0x00000307}, {0x00001E42, 0x0000004D}, {0x00001E42, 0x00000323},
-{0x00001E43, 0x0000006D}, {0x00001E43, 0x00000323}, {0x00001E44, 0x0000004E}, {0x00001E44, 0x00000307},
-{0x00001E45, 0x0000006E}, {0x00001E45, 0x00000307}, {0x00001E46, 0x0000004E}, {0x00001E46, 0x00000323},
-{0x00001E47, 0x0000006E}, {0x00001E47, 0x00000323}, {0x00001E48, 0x0000004E}, {0x00001E48, 0x00000331},
-{0x00001E49, 0x0000006E}, {0x00001E49, 0x00000331}, {0x00001E4A, 0x0000004E}, {0x00001E4A, 0x0000032D},
-{0x00001E4B, 0x0000006E}, {0x00001E4B, 0x0000032D}, {0x00001E4C, 0x0000004F}, {0x00001E4C, 0x00000303},
-{0x00001E4C, 0x00000301}, {0x00001E4D, 0x0000006F}, {0x00001E4D, 0x00000303}, {0x00001E4D, 0x00000301},
-{0x00001E4E, 0x0000004F}, {0x00001E4E, 0x00000303}, {0x00001E4E, 0x00000308}, {0x00001E4F, 0x0000006F},
-{0x00001E4F, 0x00000303}, {0x00001E4F, 0x00000308}, {0x00001E50, 0x0000004F}, {0x00001E50, 0x00000304},
-{0x00001E50, 0x00000300}, {0x00001E51, 0x0000006F}, {0x00001E51, 0x00000304}, {0x00001E51, 0x00000300},
-{0x00001E52, 0x0000004F}, {0x00001E52, 0x00000304}, {0x00001E52, 0x00000301}, {0x00001E53, 0x0000006F},
-{0x00001E53, 0x00000304}, {0x00001E53, 0x00000301}, {0x00001E54, 0x00000050}, {0x00001E54, 0x00000301},
-{0x00001E55, 0x00000070}, {0x00001E55, 0x00000301}, {0x00001E56, 0x00000050}, {0x00001E56, 0x00000307},
-{0x00001E57, 0x00000070}, {0x00001E57, 0x00000307}, {0x00001E58, 0x00000052}, {0x00001E58, 0x00000307},
-{0x00001E59, 0x00000072}, {0x00001E59, 0x00000307}, {0x00001E5A, 0x00000052}, {0x00001E5A, 0x00000323},
-{0x00001E5B, 0x00000072}, {0x00001E5B, 0x00000323}, {0x00001E5C, 0x00000052}, {0x00001E5C, 0x00000323},
-{0x00001E5C, 0x00000304}, {0x00001E5D, 0x00000072}, {0x00001E5D, 0x00000323}, {0x00001E5D, 0x00000304},
-{0x00001E5E, 0x00000052}, {0x00001E5E, 0x00000331}, {0x00001E5F, 0x00000072}, {0x00001E5F, 0x00000331},
-{0x00001E60, 0x00000053}, {0x00001E60, 0x00000307}, {0x00001E61, 0x00000073}, {0x00001E61, 0x00000307},
-{0x00001E62, 0x00000053}, {0x00001E62, 0x00000323}, {0x00001E63, 0x00000073}, {0x00001E63, 0x00000323},
-{0x00001E64, 0x00000053}, {0x00001E64, 0x00000301}, {0x00001E64, 0x00000307}, {0x00001E65, 0x00000073},
-{0x00001E65, 0x00000301}, {0x00001E65, 0x00000307}, {0x00001E66, 0x00000053}, {0x00001E66, 0x0000030C},
-{0x00001E66, 0x00000307}, {0x00001E67, 0x00000073}, {0x00001E67, 0x0000030C}, {0x00001E67, 0x00000307},
-{0x00001E68, 0x00000053}, {0x00001E68, 0x00000323}, {0x00001E68, 0x00000307}, {0x00001E69, 0x00000073},
-{0x00001E69, 0x00000323}, {0x00001E69, 0x00000307}, {0x00001E6A, 0x00000054}, {0x00001E6A, 0x00000307},
-{0x00001E6B, 0x00000074}, {0x00001E6B, 0x00000307}, {0x00001E6C, 0x00000054}, {0x00001E6C, 0x00000323},
-{0x00001E6D, 0x00000074}, {0x00001E6D, 0x00000323}, {0x00001E6E, 0x00000054}, {0x00001E6E, 0x00000331},
-{0x00001E6F, 0x00000074}, {0x00001E6F, 0x00000331}, {0x00001E70, 0x00000054}, {0x00001E70, 0x0000032D},
-{0x00001E71, 0x00000074}, {0x00001E71, 0x0000032D}, {0x00001E72, 0x00000055}, {0x00001E72, 0x00000324},
-{0x00001E73, 0x00000075}, {0x00001E73, 0x00000324}, {0x00001E74, 0x00000055}, {0x00001E74, 0x00000330},
-{0x00001E75, 0x00000075}, {0x00001E75, 0x00000330}, {0x00001E76, 0x00000055}, {0x00001E76, 0x0000032D},
-{0x00001E77, 0x00000075}, {0x00001E77, 0x0000032D}, {0x00001E78, 0x00000055}, {0x00001E78, 0x00000303},
-{0x00001E78, 0x00000301}, {0x00001E79, 0x00000075}, {0x00001E79, 0x00000303}, {0x00001E79, 0x00000301},
-{0x00001E7A, 0x00000055}, {0x00001E7A, 0x00000304}, {0x00001E7A, 0x00000308}, {0x00001E7B, 0x00000075},
-{0x00001E7B, 0x00000304}, {0x00001E7B, 0x00000308}, {0x00001E7C, 0x00000056}, {0x00001E7C, 0x00000303},
-{0x00001E7D, 0x00000076}, {0x00001E7D, 0x00000303}, {0x00001E7E, 0x00000056}, {0x00001E7E, 0x00000323},
-{0x00001E7F, 0x00000076}, {0x00001E7F, 0x00000323}, {0x00001E80, 0x00000057}, {0x00001E80, 0x00000300},
-{0x00001E81, 0x00000077}, {0x00001E81, 0x00000300}, {0x00001E82, 0x00000057}, {0x00001E82, 0x00000301},
-{0x00001E83, 0x00000077}, {0x00001E83, 0x00000301}, {0x00001E84, 0x00000057}, {0x00001E84, 0x00000308},
-{0x00001E85, 0x00000077}, {0x00001E85, 0x00000308}, {0x00001E86, 0x00000057}, {0x00001E86, 0x00000307},
-{0x00001E87, 0x00000077}, {0x00001E87, 0x00000307}, {0x00001E88, 0x00000057}, {0x00001E88, 0x00000323},
-{0x00001E89, 0x00000077}, {0x00001E89, 0x00000323}, {0x00001E8A, 0x00000058}, {0x00001E8A, 0x00000307},
-{0x00001E8B, 0x00000078}, {0x00001E8B, 0x00000307}, {0x00001E8C, 0x00000058}, {0x00001E8C, 0x00000308},
-{0x00001E8D, 0x00000078}, {0x00001E8D, 0x00000308}, {0x00001E8E, 0x00000059}, {0x00001E8E, 0x00000307},
-{0x00001E8F, 0x00000079}, {0x00001E8F, 0x00000307}, {0x00001E90, 0x0000005A}, {0x00001E90, 0x00000302},
-{0x00001E91, 0x0000007A}, {0x00001E91, 0x00000302}, {0x00001E92, 0x0000005A}, {0x00001E92, 0x00000323},
-{0x00001E93, 0x0000007A}, {0x00001E93, 0x00000323}, {0x00001E94, 0x0000005A}, {0x00001E94, 0x00000331},
-{0x00001E95, 0x0000007A}, {0x00001E95, 0x00000331}, {0x00001E96, 0x00000068}, {0x00001E96, 0x00000331},
-{0x00001E97, 0x00000074}, {0x00001E97, 0x00000308}, {0x00001E98, 0x00000077}, {0x00001E98, 0x0000030A},
-{0x00001E99, 0x00000079}, {0x00001E99, 0x0000030A}, {0x00001E9B, 0x0000017F}, {0x00001E9B, 0x00000307},
-{0x00001EA0, 0x00000041}, {0x00001EA0, 0x00000323}, {0x00001EA1, 0x00000061}, {0x00001EA1, 0x00000323},
-{0x00001EA2, 0x00000041}, {0x00001EA2, 0x00000309}, {0x00001EA3, 0x00000061}, {0x00001EA3, 0x00000309},
-{0x00001EA4, 0x00000041}, {0x00001EA4, 0x00000302}, {0x00001EA4, 0x00000301}, {0x00001EA5, 0x00000061},
-{0x00001EA5, 0x00000302}, {0x00001EA5, 0x00000301}, {0x00001EA6, 0x00000041}, {0x00001EA6, 0x00000302},
-{0x00001EA6, 0x00000300}, {0x00001EA7, 0x00000061}, {0x00001EA7, 0x00000302}, {0x00001EA7, 0x00000300},
-{0x00001EA8, 0x00000041}, {0x00001EA8, 0x00000302}, {0x00001EA8, 0x00000309}, {0x00001EA9, 0x00000061},
-{0x00001EA9, 0x00000302}, {0x00001EA9, 0x00000309}, {0x00001EAA, 0x00000041}, {0x00001EAA, 0x00000302},
-{0x00001EAA, 0x00000303}, {0x00001EAB, 0x00000061}, {0x00001EAB, 0x00000302}, {0x00001EAB, 0x00000303},
-{0x00001EAC, 0x00000041}, {0x00001EAC, 0x00000323}, {0x00001EAC, 0x00000302}, {0x00001EAD, 0x00000061},
-{0x00001EAD, 0x00000323}, {0x00001EAD, 0x00000302}, {0x00001EAE, 0x00000041}, {0x00001EAE, 0x00000306},
-{0x00001EAE, 0x00000301}, {0x00001EAF, 0x00000061}, {0x00001EAF, 0x00000306}, {0x00001EAF, 0x00000301},
-{0x00001EB0, 0x00000041}, {0x00001EB0, 0x00000306}, {0x00001EB0, 0x00000300}, {0x00001EB1, 0x00000061},
-{0x00001EB1, 0x00000306}, {0x00001EB1, 0x00000300}, {0x00001EB2, 0x00000041}, {0x00001EB2, 0x00000306},
-{0x00001EB2, 0x00000309}, {0x00001EB3, 0x00000061}, {0x00001EB3, 0x00000306}, {0x00001EB3, 0x00000309},
-{0x00001EB4, 0x00000041}, {0x00001EB4, 0x00000306}, {0x00001EB4, 0x00000303}, {0x00001EB5, 0x00000061},
-{0x00001EB5, 0x00000306}, {0x00001EB5, 0x00000303}, {0x00001EB6, 0x00000041}, {0x00001EB6, 0x00000323},
-{0x00001EB6, 0x00000306}, {0x00001EB7, 0x00000061}, {0x00001EB7, 0x00000323}, {0x00001EB7, 0x00000306},
-{0x00001EB8, 0x00000045}, {0x00001EB8, 0x00000323}, {0x00001EB9, 0x00000065}, {0x00001EB9, 0x00000323},
-{0x00001EBA, 0x00000045}, {0x00001EBA, 0x00000309}, {0x00001EBB, 0x00000065}, {0x00001EBB, 0x00000309},
-{0x00001EBC, 0x00000045}, {0x00001EBC, 0x00000303}, {0x00001EBD, 0x00000065}, {0x00001EBD, 0x00000303},
-{0x00001EBE, 0x00000045}, {0x00001EBE, 0x00000302}, {0x00001EBE, 0x00000301}, {0x00001EBF, 0x00000065},
-{0x00001EBF, 0x00000302}, {0x00001EBF, 0x00000301}, {0x00001EC0, 0x00000045}, {0x00001EC0, 0x00000302},
-{0x00001EC0, 0x00000300}, {0x00001EC1, 0x00000065}, {0x00001EC1, 0x00000302}, {0x00001EC1, 0x00000300},
-{0x00001EC2, 0x00000045}, {0x00001EC2, 0x00000302}, {0x00001EC2, 0x00000309}, {0x00001EC3, 0x00000065},
-{0x00001EC3, 0x00000302}, {0x00001EC3, 0x00000309}, {0x00001EC4, 0x00000045}, {0x00001EC4, 0x00000302},
-{0x00001EC4, 0x00000303}, {0x00001EC5, 0x00000065}, {0x00001EC5, 0x00000302}, {0x00001EC5, 0x00000303},
-{0x00001EC6, 0x00000045}, {0x00001EC6, 0x00000323}, {0x00001EC6, 0x00000302}, {0x00001EC7, 0x00000065},
-{0x00001EC7, 0x00000323}, {0x00001EC7, 0x00000302}, {0x00001EC8, 0x00000049}, {0x00001EC8, 0x00000309},
-{0x00001EC9, 0x00000069}, {0x00001EC9, 0x00000309}, {0x00001ECA, 0x00000049}, {0x00001ECA, 0x00000323},
-{0x00001ECB, 0x00000069}, {0x00001ECB, 0x00000323}, {0x00001ECC, 0x0000004F}, {0x00001ECC, 0x00000323},
-{0x00001ECD, 0x0000006F}, {0x00001ECD, 0x00000323}, {0x00001ECE, 0x0000004F}, {0x00001ECE, 0x00000309},
-{0x00001ECF, 0x0000006F}, {0x00001ECF, 0x00000309}, {0x00001ED0, 0x0000004F}, {0x00001ED0, 0x00000302},
-{0x00001ED0, 0x00000301}, {0x00001ED1, 0x0000006F}, {0x00001ED1, 0x00000302}, {0x00001ED1, 0x00000301},
-{0x00001ED2, 0x0000004F}, {0x00001ED2, 0x00000302}, {0x00001ED2, 0x00000300}, {0x00001ED3, 0x0000006F},
-{0x00001ED3, 0x00000302}, {0x00001ED3, 0x00000300}, {0x00001ED4, 0x0000004F}, {0x00001ED4, 0x00000302},
-{0x00001ED4, 0x00000309}, {0x00001ED5, 0x0000006F}, {0x00001ED5, 0x00000302}, {0x00001ED5, 0x00000309},
-{0x00001ED6, 0x0000004F}, {0x00001ED6, 0x00000302}, {0x00001ED6, 0x00000303}, {0x00001ED7, 0x0000006F},
-{0x00001ED7, 0x00000302}, {0x00001ED7, 0x00000303}, {0x00001ED8, 0x0000004F}, {0x00001ED8, 0x00000323},
-{0x00001ED8, 0x00000302}, {0x00001ED9, 0x0000006F}, {0x00001ED9, 0x00000323}, {0x00001ED9, 0x00000302},
-{0x00001EDA, 0x0000004F}, {0x00001EDA, 0x0000031B}, {0x00001EDA, 0x00000301}, {0x00001EDB, 0x0000006F},
-{0x00001EDB, 0x0000031B}, {0x00001EDB, 0x00000301}, {0x00001EDC, 0x0000004F}, {0x00001EDC, 0x0000031B},
-{0x00001EDC, 0x00000300}, {0x00001EDD, 0x0000006F}, {0x00001EDD, 0x0000031B}, {0x00001EDD, 0x00000300},
-{0x00001EDE, 0x0000004F}, {0x00001EDE, 0x0000031B}, {0x00001EDE, 0x00000309}, {0x00001EDF, 0x0000006F},
-{0x00001EDF, 0x0000031B}, {0x00001EDF, 0x00000309}, {0x00001EE0, 0x0000004F}, {0x00001EE0, 0x0000031B},
-{0x00001EE0, 0x00000303}, {0x00001EE1, 0x0000006F}, {0x00001EE1, 0x0000031B}, {0x00001EE1, 0x00000303},
-{0x00001EE2, 0x0000004F}, {0x00001EE2, 0x0000031B}, {0x00001EE2, 0x00000323}, {0x00001EE3, 0x0000006F},
-{0x00001EE3, 0x0000031B}, {0x00001EE3, 0x00000323}, {0x00001EE4, 0x00000055}, {0x00001EE4, 0x00000323},
-{0x00001EE5, 0x00000075}, {0x00001EE5, 0x00000323}, {0x00001EE6, 0x00000055}, {0x00001EE6, 0x00000309},
-{0x00001EE7, 0x00000075}, {0x00001EE7, 0x00000309}, {0x00001EE8, 0x00000055}, {0x00001EE8, 0x0000031B},
-{0x00001EE8, 0x00000301}, {0x00001EE9, 0x00000075}, {0x00001EE9, 0x0000031B}, {0x00001EE9, 0x00000301},
-{0x00001EEA, 0x00000055}, {0x00001EEA, 0x0000031B}, {0x00001EEA, 0x00000300}, {0x00001EEB, 0x00000075},
-{0x00001EEB, 0x0000031B}, {0x00001EEB, 0x00000300}, {0x00001EEC, 0x00000055}, {0x00001EEC, 0x0000031B},
-{0x00001EEC, 0x00000309}, {0x00001EED, 0x00000075}, {0x00001EED, 0x0000031B}, {0x00001EED, 0x00000309},
-{0x00001EEE, 0x00000055}, {0x00001EEE, 0x0000031B}, {0x00001EEE, 0x00000303}, {0x00001EEF, 0x00000075},
-{0x00001EEF, 0x0000031B}, {0x00001EEF, 0x00000303}, {0x00001EF0, 0x00000055}, {0x00001EF0, 0x0000031B},
-{0x00001EF0, 0x00000323}, {0x00001EF1, 0x00000075}, {0x00001EF1, 0x0000031B}, {0x00001EF1, 0x00000323},
-{0x00001EF2, 0x00000059}, {0x00001EF2, 0x00000300}, {0x00001EF3, 0x00000079}, {0x00001EF3, 0x00000300},
-{0x00001EF4, 0x00000059}, {0x00001EF4, 0x00000323}, {0x00001EF5, 0x00000079}, {0x00001EF5, 0x00000323},
-{0x00001EF6, 0x00000059}, {0x00001EF6, 0x00000309}, {0x00001EF7, 0x00000079}, {0x00001EF7, 0x00000309},
-{0x00001EF8, 0x00000059}, {0x00001EF8, 0x00000303}, {0x00001EF9, 0x00000079}, {0x00001EF9, 0x00000303},
-{0x00001F00, 0x000003B1}, {0x00001F00, 0x00000313}, {0x00001F01, 0x000003B1}, {0x00001F01, 0x00000314},
-{0x00001F02, 0x000003B1}, {0x00001F02, 0x00000313}, {0x00001F02, 0x00000300}, {0x00001F03, 0x000003B1},
-{0x00001F03, 0x00000314}, {0x00001F03, 0x00000300}, {0x00001F04, 0x000003B1}, {0x00001F04, 0x00000313},
-{0x00001F04, 0x00000301}, {0x00001F05, 0x000003B1}, {0x00001F05, 0x00000314}, {0x00001F05, 0x00000301},
-{0x00001F06, 0x000003B1}, {0x00001F06, 0x00000313}, {0x00001F06, 0x00000342}, {0x00001F07, 0x000003B1},
-{0x00001F07, 0x00000314}, {0x00001F07, 0x00000342}, {0x00001F08, 0x00000391}, {0x00001F08, 0x00000313},
-{0x00001F09, 0x00000391}, {0x00001F09, 0x00000314}, {0x00001F0A, 0x00000391}, {0x00001F0A, 0x00000313},
-{0x00001F0A, 0x00000300}, {0x00001F0B, 0x00000391}, {0x00001F0B, 0x00000314}, {0x00001F0B, 0x00000300},
-{0x00001F0C, 0x00000391}, {0x00001F0C, 0x00000313}, {0x00001F0C, 0x00000301}, {0x00001F0D, 0x00000391},
-{0x00001F0D, 0x00000314}, {0x00001F0D, 0x00000301}, {0x00001F0E, 0x00000391}, {0x00001F0E, 0x00000313},
-{0x00001F0E, 0x00000342}, {0x00001F0F, 0x00000391}, {0x00001F0F, 0x00000314}, {0x00001F0F, 0x00000342},
-{0x00001F10, 0x000003B5}, {0x00001F10, 0x00000313}, {0x00001F11, 0x000003B5}, {0x00001F11, 0x00000314},
-{0x00001F12, 0x000003B5}, {0x00001F12, 0x00000313}, {0x00001F12, 0x00000300}, {0x00001F13, 0x000003B5},
-{0x00001F13, 0x00000314}, {0x00001F13, 0x00000300}, {0x00001F14, 0x000003B5}, {0x00001F14, 0x00000313},
-{0x00001F14, 0x00000301}, {0x00001F15, 0x000003B5}, {0x00001F15, 0x00000314}, {0x00001F15, 0x00000301},
-{0x00001F18, 0x00000395}, {0x00001F18, 0x00000313}, {0x00001F19, 0x00000395}, {0x00001F19, 0x00000314},
-{0x00001F1A, 0x00000395}, {0x00001F1A, 0x00000313}, {0x00001F1A, 0x00000300}, {0x00001F1B, 0x00000395},
-{0x00001F1B, 0x00000314}, {0x00001F1B, 0x00000300}, {0x00001F1C, 0x00000395}, {0x00001F1C, 0x00000313},
-{0x00001F1C, 0x00000301}, {0x00001F1D, 0x00000395}, {0x00001F1D, 0x00000314}, {0x00001F1D, 0x00000301},
-{0x00001F20, 0x000003B7}, {0x00001F20, 0x00000313}, {0x00001F21, 0x000003B7}, {0x00001F21, 0x00000314},
-{0x00001F22, 0x000003B7}, {0x00001F22, 0x00000313}, {0x00001F22, 0x00000300}, {0x00001F23, 0x000003B7},
-{0x00001F23, 0x00000314}, {0x00001F23, 0x00000300}, {0x00001F24, 0x000003B7}, {0x00001F24, 0x00000313},
-{0x00001F24, 0x00000301}, {0x00001F25, 0x000003B7}, {0x00001F25, 0x00000314}, {0x00001F25, 0x00000301},
-{0x00001F26, 0x000003B7}, {0x00001F26, 0x00000313}, {0x00001F26, 0x00000342}, {0x00001F27, 0x000003B7},
-{0x00001F27, 0x00000314}, {0x00001F27, 0x00000342}, {0x00001F28, 0x00000397}, {0x00001F28, 0x00000313},
-{0x00001F29, 0x00000397}, {0x00001F29, 0x00000314}, {0x00001F2A, 0x00000397}, {0x00001F2A, 0x00000313},
-{0x00001F2A, 0x00000300}, {0x00001F2B, 0x00000397}, {0x00001F2B, 0x00000314}, {0x00001F2B, 0x00000300},
-{0x00001F2C, 0x00000397}, {0x00001F2C, 0x00000313}, {0x00001F2C, 0x00000301}, {0x00001F2D, 0x00000397},
-{0x00001F2D, 0x00000314}, {0x00001F2D, 0x00000301}, {0x00001F2E, 0x00000397}, {0x00001F2E, 0x00000313},
-{0x00001F2E, 0x00000342}, {0x00001F2F, 0x00000397}, {0x00001F2F, 0x00000314}, {0x00001F2F, 0x00000342},
-{0x00001F30, 0x000003B9}, {0x00001F30, 0x00000313}, {0x00001F31, 0x000003B9}, {0x00001F31, 0x00000314},
-{0x00001F32, 0x000003B9}, {0x00001F32, 0x00000313}, {0x00001F32, 0x00000300}, {0x00001F33, 0x000003B9},
-{0x00001F33, 0x00000314}, {0x00001F33, 0x00000300}, {0x00001F34, 0x000003B9}, {0x00001F34, 0x00000313},
-{0x00001F34, 0x00000301}, {0x00001F35, 0x000003B9}, {0x00001F35, 0x00000314}, {0x00001F35, 0x00000301},
-{0x00001F36, 0x000003B9}, {0x00001F36, 0x00000313}, {0x00001F36, 0x00000342}, {0x00001F37, 0x000003B9},
-{0x00001F37, 0x00000314}, {0x00001F37, 0x00000342}, {0x00001F38, 0x00000399}, {0x00001F38, 0x00000313},
-{0x00001F39, 0x00000399}, {0x00001F39, 0x00000314}, {0x00001F3A, 0x00000399}, {0x00001F3A, 0x00000313},
-{0x00001F3A, 0x00000300}, {0x00001F3B, 0x00000399}, {0x00001F3B, 0x00000314}, {0x00001F3B, 0x00000300},
-{0x00001F3C, 0x00000399}, {0x00001F3C, 0x00000313}, {0x00001F3C, 0x00000301}, {0x00001F3D, 0x00000399},
-{0x00001F3D, 0x00000314}, {0x00001F3D, 0x00000301}, {0x00001F3E, 0x00000399}, {0x00001F3E, 0x00000313},
-{0x00001F3E, 0x00000342}, {0x00001F3F, 0x00000399}, {0x00001F3F, 0x00000314}, {0x00001F3F, 0x00000342},
-{0x00001F40, 0x000003BF}, {0x00001F40, 0x00000313}, {0x00001F41, 0x000003BF}, {0x00001F41, 0x00000314},
-{0x00001F42, 0x000003BF}, {0x00001F42, 0x00000313}, {0x00001F42, 0x00000300}, {0x00001F43, 0x000003BF},
-{0x00001F43, 0x00000314}, {0x00001F43, 0x00000300}, {0x00001F44, 0x000003BF}, {0x00001F44, 0x00000313},
-{0x00001F44, 0x00000301}, {0x00001F45, 0x000003BF}, {0x00001F45, 0x00000314}, {0x00001F45, 0x00000301},
-{0x00001F48, 0x0000039F}, {0x00001F48, 0x00000313}, {0x00001F49, 0x0000039F}, {0x00001F49, 0x00000314},
-{0x00001F4A, 0x0000039F}, {0x00001F4A, 0x00000313}, {0x00001F4A, 0x00000300}, {0x00001F4B, 0x0000039F},
-{0x00001F4B, 0x00000314}, {0x00001F4B, 0x00000300}, {0x00001F4C, 0x0000039F}, {0x00001F4C, 0x00000313},
-{0x00001F4C, 0x00000301}, {0x00001F4D, 0x0000039F}, {0x00001F4D, 0x00000314}, {0x00001F4D, 0x00000301},
-{0x00001F50, 0x000003C5}, {0x00001F50, 0x00000313}, {0x00001F51, 0x000003C5}, {0x00001F51, 0x00000314},
-{0x00001F52, 0x000003C5}, {0x00001F52, 0x00000313}, {0x00001F52, 0x00000300}, {0x00001F53, 0x000003C5},
-{0x00001F53, 0x00000314}, {0x00001F53, 0x00000300}, {0x00001F54, 0x000003C5}, {0x00001F54, 0x00000313},
-{0x00001F54, 0x00000301}, {0x00001F55, 0x000003C5}, {0x00001F55, 0x00000314}, {0x00001F55, 0x00000301},
-{0x00001F56, 0x000003C5}, {0x00001F56, 0x00000313}, {0x00001F56, 0x00000342}, {0x00001F57, 0x000003C5},
-{0x00001F57, 0x00000314}, {0x00001F57, 0x00000342}, {0x00001F59, 0x000003A5}, {0x00001F59, 0x00000314},
-{0x00001F5B, 0x000003A5}, {0x00001F5B, 0x00000314}, {0x00001F5B, 0x00000300}, {0x00001F5D, 0x000003A5},
-{0x00001F5D, 0x00000314}, {0x00001F5D, 0x00000301}, {0x00001F5F, 0x000003A5}, {0x00001F5F, 0x00000314},
-{0x00001F5F, 0x00000342}, {0x00001F60, 0x000003C9}, {0x00001F60, 0x00000313}, {0x00001F61, 0x000003C9},
-{0x00001F61, 0x00000314}, {0x00001F62, 0x000003C9}, {0x00001F62, 0x00000313}, {0x00001F62, 0x00000300},
-{0x00001F63, 0x000003C9}, {0x00001F63, 0x00000314}, {0x00001F63, 0x00000300}, {0x00001F64, 0x000003C9},
-{0x00001F64, 0x00000313}, {0x00001F64, 0x00000301}, {0x00001F65, 0x000003C9}, {0x00001F65, 0x00000314},
-{0x00001F65, 0x00000301}, {0x00001F66, 0x000003C9}, {0x00001F66, 0x00000313}, {0x00001F66, 0x00000342},
-{0x00001F67, 0x000003C9}, {0x00001F67, 0x00000314}, {0x00001F67, 0x00000342}, {0x00001F68, 0x000003A9},
-{0x00001F68, 0x00000313}, {0x00001F69, 0x000003A9}, {0x00001F69, 0x00000314}, {0x00001F6A, 0x000003A9},
-{0x00001F6A, 0x00000313}, {0x00001F6A, 0x00000300}, {0x00001F6B, 0x000003A9}, {0x00001F6B, 0x00000314},
-{0x00001F6B, 0x00000300}, {0x00001F6C, 0x000003A9}, {0x00001F6C, 0x00000313}, {0x00001F6C, 0x00000301},
-{0x00001F6D, 0x000003A9}, {0x00001F6D, 0x00000314}, {0x00001F6D, 0x00000301}, {0x00001F6E, 0x000003A9},
-{0x00001F6E, 0x00000313}, {0x00001F6E, 0x00000342}, {0x00001F6F, 0x000003A9}, {0x00001F6F, 0x00000314},
-{0x00001F6F, 0x00000342}, {0x00001F70, 0x000003B1}, {0x00001F70, 0x00000300}, {0x00001F71, 0x000003B1},
-{0x00001F71, 0x00000301}, {0x00001F72, 0x000003B5}, {0x00001F72, 0x00000300}, {0x00001F73, 0x000003B5},
-{0x00001F73, 0x00000301}, {0x00001F74, 0x000003B7}, {0x00001F74, 0x00000300}, {0x00001F75, 0x000003B7},
-{0x00001F75, 0x00000301}, {0x00001F76, 0x000003B9}, {0x00001F76, 0x00000300}, {0x00001F77, 0x000003B9},
-{0x00001F77, 0x00000301}, {0x00001F78, 0x000003BF}, {0x00001F78, 0x00000300}, {0x00001F79, 0x000003BF},
-{0x00001F79, 0x00000301}, {0x00001F7A, 0x000003C5}, {0x00001F7A, 0x00000300}, {0x00001F7B, 0x000003C5},
-{0x00001F7B, 0x00000301}, {0x00001F7C, 0x000003C9}, {0x00001F7C, 0x00000300}, {0x00001F7D, 0x000003C9},
-{0x00001F7D, 0x00000301}, {0x00001F80, 0x000003B1}, {0x00001F80, 0x00000313}, {0x00001F80, 0x00000345},
-{0x00001F81, 0x000003B1}, {0x00001F81, 0x00000314}, {0x00001F81, 0x00000345}, {0x00001F82, 0x000003B1},
-{0x00001F82, 0x00000313}, {0x00001F82, 0x00000300}, {0x00001F82, 0x00000345}, {0x00001F83, 0x000003B1},
-{0x00001F83, 0x00000314}, {0x00001F83, 0x00000300}, {0x00001F83, 0x00000345}, {0x00001F84, 0x000003B1},
-{0x00001F84, 0x00000313}, {0x00001F84, 0x00000301}, {0x00001F84, 0x00000345}, {0x00001F85, 0x000003B1},
-{0x00001F85, 0x00000314}, {0x00001F85, 0x00000301}, {0x00001F85, 0x00000345}, {0x00001F86, 0x000003B1},
-{0x00001F86, 0x00000313}, {0x00001F86, 0x00000342}, {0x00001F86, 0x00000345}, {0x00001F87, 0x000003B1},
-{0x00001F87, 0x00000314}, {0x00001F87, 0x00000342}, {0x00001F87, 0x00000345}, {0x00001F88, 0x00000391},
-{0x00001F88, 0x00000313}, {0x00001F88, 0x00000345}, {0x00001F89, 0x00000391}, {0x00001F89, 0x00000314},
-{0x00001F89, 0x00000345}, {0x00001F8A, 0x00000391}, {0x00001F8A, 0x00000313}, {0x00001F8A, 0x00000300},
-{0x00001F8A, 0x00000345}, {0x00001F8B, 0x00000391}, {0x00001F8B, 0x00000314}, {0x00001F8B, 0x00000300},
-{0x00001F8B, 0x00000345}, {0x00001F8C, 0x00000391}, {0x00001F8C, 0x00000313}, {0x00001F8C, 0x00000301},
-{0x00001F8C, 0x00000345}, {0x00001F8D, 0x00000391}, {0x00001F8D, 0x00000314}, {0x00001F8D, 0x00000301},
-{0x00001F8D, 0x00000345}, {0x00001F8E, 0x00000391}, {0x00001F8E, 0x00000313}, {0x00001F8E, 0x00000342},
-{0x00001F8E, 0x00000345}, {0x00001F8F, 0x00000391}, {0x00001F8F, 0x00000314}, {0x00001F8F, 0x00000342},
-{0x00001F8F, 0x00000345}, {0x00001F90, 0x000003B7}, {0x00001F90, 0x00000313}, {0x00001F90, 0x00000345},
-{0x00001F91, 0x000003B7}, {0x00001F91, 0x00000314}, {0x00001F91, 0x00000345}, {0x00001F92, 0x000003B7},
-{0x00001F92, 0x00000313}, {0x00001F92, 0x00000300}, {0x00001F92, 0x00000345}, {0x00001F93, 0x000003B7},
-{0x00001F93, 0x00000314}, {0x00001F93, 0x00000300}, {0x00001F93, 0x00000345}, {0x00001F94, 0x000003B7},
-{0x00001F94, 0x00000313}, {0x00001F94, 0x00000301}, {0x00001F94, 0x00000345}, {0x00001F95, 0x000003B7},
-{0x00001F95, 0x00000314}, {0x00001F95, 0x00000301}, {0x00001F95, 0x00000345}, {0x00001F96, 0x000003B7},
-{0x00001F96, 0x00000313}, {0x00001F96, 0x00000342}, {0x00001F96, 0x00000345}, {0x00001F97, 0x000003B7},
-{0x00001F97, 0x00000314}, {0x00001F97, 0x00000342}, {0x00001F97, 0x00000345}, {0x00001F98, 0x00000397},
-{0x00001F98, 0x00000313}, {0x00001F98, 0x00000345}, {0x00001F99, 0x00000397}, {0x00001F99, 0x00000314},
-{0x00001F99, 0x00000345}, {0x00001F9A, 0x00000397}, {0x00001F9A, 0x00000313}, {0x00001F9A, 0x00000300},
-{0x00001F9A, 0x00000345}, {0x00001F9B, 0x00000397}, {0x00001F9B, 0x00000314}, {0x00001F9B, 0x00000300},
-{0x00001F9B, 0x00000345}, {0x00001F9C, 0x00000397}, {0x00001F9C, 0x00000313}, {0x00001F9C, 0x00000301},
-{0x00001F9C, 0x00000345}, {0x00001F9D, 0x00000397}, {0x00001F9D, 0x00000314}, {0x00001F9D, 0x00000301},
-{0x00001F9D, 0x00000345}, {0x00001F9E, 0x00000397}, {0x00001F9E, 0x00000313}, {0x00001F9E, 0x00000342},
-{0x00001F9E, 0x00000345}, {0x00001F9F, 0x00000397}, {0x00001F9F, 0x00000314}, {0x00001F9F, 0x00000342},
-{0x00001F9F, 0x00000345}, {0x00001FA0, 0x000003C9}, {0x00001FA0, 0x00000313}, {0x00001FA0, 0x00000345},
-{0x00001FA1, 0x000003C9}, {0x00001FA1, 0x00000314}, {0x00001FA1, 0x00000345}, {0x00001FA2, 0x000003C9},
-{0x00001FA2, 0x00000313}, {0x00001FA2, 0x00000300}, {0x00001FA2, 0x00000345}, {0x00001FA3, 0x000003C9},
-{0x00001FA3, 0x00000314}, {0x00001FA3, 0x00000300}, {0x00001FA3, 0x00000345}, {0x00001FA4, 0x000003C9},
-{0x00001FA4, 0x00000313}, {0x00001FA4, 0x00000301}, {0x00001FA4, 0x00000345}, {0x00001FA5, 0x000003C9},
-{0x00001FA5, 0x00000314}, {0x00001FA5, 0x00000301}, {0x00001FA5, 0x00000345}, {0x00001FA6, 0x000003C9},
-{0x00001FA6, 0x00000313}, {0x00001FA6, 0x00000342}, {0x00001FA6, 0x00000345}, {0x00001FA7, 0x000003C9},
-{0x00001FA7, 0x00000314}, {0x00001FA7, 0x00000342}, {0x00001FA7, 0x00000345}, {0x00001FA8, 0x000003A9},
-{0x00001FA8, 0x00000313}, {0x00001FA8, 0x00000345}, {0x00001FA9, 0x000003A9}, {0x00001FA9, 0x00000314},
-{0x00001FA9, 0x00000345}, {0x00001FAA, 0x000003A9}, {0x00001FAA, 0x00000313}, {0x00001FAA, 0x00000300},
-{0x00001FAA, 0x00000345}, {0x00001FAB, 0x000003A9}, {0x00001FAB, 0x00000314}, {0x00001FAB, 0x00000300},
-{0x00001FAB, 0x00000345}, {0x00001FAC, 0x000003A9}, {0x00001FAC, 0x00000313}, {0x00001FAC, 0x00000301},
-{0x00001FAC, 0x00000345}, {0x00001FAD, 0x000003A9}, {0x00001FAD, 0x00000314}, {0x00001FAD, 0x00000301},
-{0x00001FAD, 0x00000345}, {0x00001FAE, 0x000003A9}, {0x00001FAE, 0x00000313}, {0x00001FAE, 0x00000342},
-{0x00001FAE, 0x00000345}, {0x00001FAF, 0x000003A9}, {0x00001FAF, 0x00000314}, {0x00001FAF, 0x00000342},
-{0x00001FAF, 0x00000345}, {0x00001FB0, 0x000003B1}, {0x00001FB0, 0x00000306}, {0x00001FB1, 0x000003B1},
-{0x00001FB1, 0x00000304}, {0x00001FB2, 0x000003B1}, {0x00001FB2, 0x00000300}, {0x00001FB2, 0x00000345},
-{0x00001FB3, 0x000003B1}, {0x00001FB3, 0x00000345}, {0x00001FB4, 0x000003B1}, {0x00001FB4, 0x00000301},
-{0x00001FB4, 0x00000345}, {0x00001FB6, 0x000003B1}, {0x00001FB6, 0x00000342}, {0x00001FB7, 0x000003B1},
-{0x00001FB7, 0x00000342}, {0x00001FB7, 0x00000345}, {0x00001FB8, 0x00000391}, {0x00001FB8, 0x00000306},
-{0x00001FB9, 0x00000391}, {0x00001FB9, 0x00000304}, {0x00001FBA, 0x00000391}, {0x00001FBA, 0x00000300},
-{0x00001FBB, 0x00000391}, {0x00001FBB, 0x00000301}, {0x00001FBC, 0x00000391}, {0x00001FBC, 0x00000345},
-{0x00001FBE, 0x000003B9}, {0x00001FC1, 0x000000A8}, {0x00001FC1, 0x00000342}, {0x00001FC2, 0x000003B7},
-{0x00001FC2, 0x00000300}, {0x00001FC2, 0x00000345}, {0x00001FC3, 0x000003B7}, {0x00001FC3, 0x00000345},
-{0x00001FC4, 0x000003B7}, {0x00001FC4, 0x00000301}, {0x00001FC4, 0x00000345}, {0x00001FC6, 0x000003B7},
-{0x00001FC6, 0x00000342}, {0x00001FC7, 0x000003B7}, {0x00001FC7, 0x00000342}, {0x00001FC7, 0x00000345},
-{0x00001FC8, 0x00000395}, {0x00001FC8, 0x00000300}, {0x00001FC9, 0x00000395}, {0x00001FC9, 0x00000301},
-{0x00001FCA, 0x00000397}, {0x00001FCA, 0x00000300}, {0x00001FCB, 0x00000397}, {0x00001FCB, 0x00000301},
-{0x00001FCC, 0x00000397}, {0x00001FCC, 0x00000345}, {0x00001FCD, 0x00001FBF}, {0x00001FCD, 0x00000300},
-{0x00001FCE, 0x00001FBF}, {0x00001FCE, 0x00000301}, {0x00001FCF, 0x00001FBF}, {0x00001FCF, 0x00000342},
-{0x00001FD0, 0x000003B9}, {0x00001FD0, 0x00000306}, {0x00001FD1, 0x000003B9}, {0x00001FD1, 0x00000304},
-{0x00001FD2, 0x000003B9}, {0x00001FD2, 0x00000308}, {0x00001FD2, 0x00000300}, {0x00001FD3, 0x000003B9},
-{0x00001FD3, 0x00000308}, {0x00001FD3, 0x00000301}, {0x00001FD6, 0x000003B9}, {0x00001FD6, 0x00000342},
-{0x00001FD7, 0x000003B9}, {0x00001FD7, 0x00000308}, {0x00001FD7, 0x00000342}, {0x00001FD8, 0x00000399},
-{0x00001FD8, 0x00000306}, {0x00001FD9, 0x00000399}, {0x00001FD9, 0x00000304}, {0x00001FDA, 0x00000399},
-{0x00001FDA, 0x00000300}, {0x00001FDB, 0x00000399}, {0x00001FDB, 0x00000301}, {0x00001FDD, 0x00001FFE},
-{0x00001FDD, 0x00000300}, {0x00001FDE, 0x00001FFE}, {0x00001FDE, 0x00000301}, {0x00001FDF, 0x00001FFE},
-{0x00001FDF, 0x00000342}, {0x00001FE0, 0x000003C5}, {0x00001FE0, 0x00000306}, {0x00001FE1, 0x000003C5},
-{0x00001FE1, 0x00000304}, {0x00001FE2, 0x000003C5}, {0x00001FE2, 0x00000308}, {0x00001FE2, 0x00000300},
-{0x00001FE3, 0x000003C5}, {0x00001FE3, 0x00000308}, {0x00001FE3, 0x00000301}, {0x00001FE4, 0x000003C1},
-{0x00001FE4, 0x00000313}, {0x00001FE5, 0x000003C1}, {0x00001FE5, 0x00000314}, {0x00001FE6, 0x000003C5},
-{0x00001FE6, 0x00000342}, {0x00001FE7, 0x000003C5}, {0x00001FE7, 0x00000308}, {0x00001FE7, 0x00000342},
-{0x00001FE8, 0x000003A5}, {0x00001FE8, 0x00000306}, {0x00001FE9, 0x000003A5}, {0x00001FE9, 0x00000304},
-{0x00001FEA, 0x000003A5}, {0x00001FEA, 0x00000300}, {0x00001FEB, 0x000003A5}, {0x00001FEB, 0x00000301},
-{0x00001FEC, 0x000003A1}, {0x00001FEC, 0x00000314}, {0x00001FED, 0x000000A8}, {0x00001FED, 0x00000300},
-{0x00001FEE, 0x000000A8}, {0x00001FEE, 0x00000301}, {0x00001FEF, 0x00000060}, {0x00001FF2, 0x000003C9},
-{0x00001FF2, 0x00000300}, {0x00001FF2, 0x00000345}, {0x00001FF3, 0x000003C9}, {0x00001FF3, 0x00000345},
-{0x00001FF4, 0x000003C9}, {0x00001FF4, 0x00000301}, {0x00001FF4, 0x00000345}, {0x00001FF6, 0x000003C9},
-{0x00001FF6, 0x00000342}, {0x00001FF7, 0x000003C9}, {0x00001FF7, 0x00000342}, {0x00001FF7, 0x00000345},
-{0x00001FF8, 0x0000039F}, {0x00001FF8, 0x00000300}, {0x00001FF9, 0x0000039F}, {0x00001FF9, 0x00000301},
-{0x00001FFA, 0x000003A9}, {0x00001FFA, 0x00000300}, {0x00001FFB, 0x000003A9}, {0x00001FFB, 0x00000301},
-{0x00001FFC, 0x000003A9}, {0x00001FFC, 0x00000345}, {0x00001FFD, 0x000000B4}, {0x00002000, 0x00002002},
-{0x00002001, 0x00002003}, {0x00002126, 0x000003A9}, {0x0000212A, 0x0000004B}, {0x0000212B, 0x00000041},
-{0x0000212B, 0x0000030A}, {0x0000219A, 0x00002190}, {0x0000219A, 0x00000338}, {0x0000219B, 0x00002192},
-{0x0000219B, 0x00000338}, {0x000021AE, 0x00002194}, {0x000021AE, 0x00000338}, {0x000021CD, 0x000021D0},
-{0x000021CD, 0x00000338}, {0x000021CE, 0x000021D4}, {0x000021CE, 0x00000338}, {0x000021CF, 0x000021D2},
-{0x000021CF, 0x00000338}, {0x00002204, 0x00002203}, {0x00002204, 0x00000338}, {0x00002209, 0x00002208},
-{0x00002209, 0x00000338}, {0x0000220C, 0x0000220B}, {0x0000220C, 0x00000338}, {0x00002224, 0x00002223},
-{0x00002224, 0x00000338}, {0x00002226, 0x00002225}, {0x00002226, 0x00000338}, {0x00002241, 0x0000223C},
-{0x00002241, 0x00000338}, {0x00002244, 0x00002243}, {0x00002244, 0x00000338}, {0x00002247, 0x00002245},
-{0x00002247, 0x00000338}, {0x00002249, 0x00002248}, {0x00002249, 0x00000338}, {0x00002260, 0x0000003D},
-{0x00002260, 0x00000338}, {0x00002262, 0x00002261}, {0x00002262, 0x00000338}, {0x0000226D, 0x0000224D},
-{0x0000226D, 0x00000338}, {0x0000226E, 0x0000003C}, {0x0000226E, 0x00000338}, {0x0000226F, 0x0000003E},
-{0x0000226F, 0x00000338}, {0x00002270, 0x00002264}, {0x00002270, 0x00000338}, {0x00002271, 0x00002265},
-{0x00002271, 0x00000338}, {0x00002274, 0x00002272}, {0x00002274, 0x00000338}, {0x00002275, 0x00002273},
-{0x00002275, 0x00000338}, {0x00002278, 0x00002276}, {0x00002278, 0x00000338}, {0x00002279, 0x00002277},
-{0x00002279, 0x00000338}, {0x00002280, 0x0000227A}, {0x00002280, 0x00000338}, {0x00002281, 0x0000227B},
-{0x00002281, 0x00000338}, {0x00002284, 0x00002282}, {0x00002284, 0x00000338}, {0x00002285, 0x00002283},
-{0x00002285, 0x00000338}, {0x00002288, 0x00002286}, {0x00002288, 0x00000338}, {0x00002289, 0x00002287},
-{0x00002289, 0x00000338}, {0x000022AC, 0x000022A2}, {0x000022AC, 0x00000338}, {0x000022AD, 0x000022A8},
-{0x000022AD, 0x00000338}, {0x000022AE, 0x000022A9}, {0x000022AE, 0x00000338}, {0x000022AF, 0x000022AB},
-{0x000022AF, 0x00000338}, {0x000022E0, 0x0000227C}, {0x000022E0, 0x00000338}, {0x000022E1, 0x0000227D},
-{0x000022E1, 0x00000338}, {0x000022E2, 0x00002291}, {0x000022E2, 0x00000338}, {0x000022E3, 0x00002292},
-{0x000022E3, 0x00000338}, {0x000022EA, 0x000022B2}, {0x000022EA, 0x00000338}, {0x000022EB, 0x000022B3},
-{0x000022EB, 0x00000338}, {0x000022EC, 0x000022B4}, {0x000022EC, 0x00000338}, {0x000022ED, 0x000022B5},
-{0x000022ED, 0x00000338}, {0x00002329, 0x00003008}, {0x0000232A, 0x00003009}, {0x00002ADC, 0x00002ADD},
-{0x00002ADC, 0x00000338}, {0x0000304C, 0x0000304B}, {0x0000304C, 0x00003099}, {0x0000304E, 0x0000304D},
-{0x0000304E, 0x00003099}, {0x00003050, 0x0000304F}, {0x00003050, 0x00003099}, {0x00003052, 0x00003051},
-{0x00003052, 0x00003099}, {0x00003054, 0x00003053}, {0x00003054, 0x00003099}, {0x00003056, 0x00003055},
-{0x00003056, 0x00003099}, {0x00003058, 0x00003057}, {0x00003058, 0x00003099}, {0x0000305A, 0x00003059},
-{0x0000305A, 0x00003099}, {0x0000305C, 0x0000305B}, {0x0000305C, 0x00003099}, {0x0000305E, 0x0000305D},
-{0x0000305E, 0x00003099}, {0x00003060, 0x0000305F}, {0x00003060, 0x00003099}, {0x00003062, 0x00003061},
-{0x00003062, 0x00003099}, {0x00003065, 0x00003064}, {0x00003065, 0x00003099}, {0x00003067, 0x00003066},
-{0x00003067, 0x00003099}, {0x00003069, 0x00003068}, {0x00003069, 0x00003099}, {0x00003070, 0x0000306F},
-{0x00003070, 0x00003099}, {0x00003071, 0x0000306F}, {0x00003071, 0x0000309A}, {0x00003073, 0x00003072},
-{0x00003073, 0x00003099}, {0x00003074, 0x00003072}, {0x00003074, 0x0000309A}, {0x00003076, 0x00003075},
-{0x00003076, 0x00003099}, {0x00003077, 0x00003075}, {0x00003077, 0x0000309A}, {0x00003079, 0x00003078},
-{0x00003079, 0x00003099}, {0x0000307A, 0x00003078}, {0x0000307A, 0x0000309A}, {0x0000307C, 0x0000307B},
-{0x0000307C, 0x00003099}, {0x0000307D, 0x0000307B}, {0x0000307D, 0x0000309A}, {0x00003094, 0x00003046},
-{0x00003094, 0x00003099}, {0x0000309E, 0x0000309D}, {0x0000309E, 0x00003099}, {0x000030AC, 0x000030AB},
-{0x000030AC, 0x00003099}, {0x000030AE, 0x000030AD}, {0x000030AE, 0x00003099}, {0x000030B0, 0x000030AF},
-{0x000030B0, 0x00003099}, {0x000030B2, 0x000030B1}, {0x000030B2, 0x00003099}, {0x000030B4, 0x000030B3},
-{0x000030B4, 0x00003099}, {0x000030B6, 0x000030B5}, {0x000030B6, 0x00003099}, {0x000030B8, 0x000030B7},
-{0x000030B8, 0x00003099}, {0x000030BA, 0x000030B9}, {0x000030BA, 0x00003099}, {0x000030BC, 0x000030BB},
-{0x000030BC, 0x00003099}, {0x000030BE, 0x000030BD}, {0x000030BE, 0x00003099}, {0x000030C0, 0x000030BF},
-{0x000030C0, 0x00003099}, {0x000030C2, 0x000030C1}, {0x000030C2, 0x00003099}, {0x000030C5, 0x000030C4},
-{0x000030C5, 0x00003099}, {0x000030C7, 0x000030C6}, {0x000030C7, 0x00003099}, {0x000030C9, 0x000030C8},
-{0x000030C9, 0x00003099}, {0x000030D0, 0x000030CF}, {0x000030D0, 0x00003099}, {0x000030D1, 0x000030CF},
-{0x000030D1, 0x0000309A}, {0x000030D3, 0x000030D2}, {0x000030D3, 0x00003099}, {0x000030D4, 0x000030D2},
-{0x000030D4, 0x0000309A}, {0x000030D6, 0x000030D5}, {0x000030D6, 0x00003099}, {0x000030D7, 0x000030D5},
-{0x000030D7, 0x0000309A}, {0x000030D9, 0x000030D8}, {0x000030D9, 0x00003099}, {0x000030DA, 0x000030D8},
-{0x000030DA, 0x0000309A}, {0x000030DC, 0x000030DB}, {0x000030DC, 0x00003099}, {0x000030DD, 0x000030DB},
-{0x000030DD, 0x0000309A}, {0x000030F4, 0x000030A6}, {0x000030F4, 0x00003099}, {0x000030F7, 0x000030EF},
-{0x000030F7, 0x00003099}, {0x000030F8, 0x000030F0}, {0x000030F8, 0x00003099}, {0x000030F9, 0x000030F1},
-{0x000030F9, 0x00003099}, {0x000030FA, 0x000030F2}, {0x000030FA, 0x00003099}, {0x000030FE, 0x000030FD},
-{0x000030FE, 0x00003099}, {0x0000F900, 0x00008C48}, {0x0000F901, 0x000066F4}, {0x0000F902, 0x00008ECA},
-{0x0000F903, 0x00008CC8}, {0x0000F904, 0x00006ED1}, {0x0000F905, 0x00004E32}, {0x0000F906, 0x000053E5},
-{0x0000F907, 0x00009F9C}, {0x0000F908, 0x00009F9C}, {0x0000F909, 0x00005951}, {0x0000F90A, 0x000091D1},
-{0x0000F90B, 0x00005587}, {0x0000F90C, 0x00005948}, {0x0000F90D, 0x000061F6}, {0x0000F90E, 0x00007669},
-{0x0000F90F, 0x00007F85}, {0x0000F910, 0x0000863F}, {0x0000F911, 0x000087BA}, {0x0000F912, 0x000088F8},
-{0x0000F913, 0x0000908F}, {0x0000F914, 0x00006A02}, {0x0000F915, 0x00006D1B}, {0x0000F916, 0x000070D9},
-{0x0000F917, 0x000073DE}, {0x0000F918, 0x0000843D}, {0x0000F919, 0x0000916A}, {0x0000F91A, 0x000099F1},
-{0x0000F91B, 0x00004E82}, {0x0000F91C, 0x00005375}, {0x0000F91D, 0x00006B04}, {0x0000F91E, 0x0000721B},
-{0x0000F91F, 0x0000862D}, {0x0000F920, 0x00009E1E}, {0x0000F921, 0x00005D50}, {0x0000F922, 0x00006FEB},
-{0x0000F923, 0x000085CD}, {0x0000F924, 0x00008964}, {0x0000F925, 0x000062C9}, {0x0000F926, 0x000081D8},
-{0x0000F927, 0x0000881F}, {0x0000F928, 0x00005ECA}, {0x0000F929, 0x00006717}, {0x0000F92A, 0x00006D6A},
-{0x0000F92B, 0x000072FC}, {0x0000F92C, 0x000090CE}, {0x0000F92D, 0x00004F86}, {0x0000F92E, 0x000051B7},
-{0x0000F92F, 0x000052DE}, {0x0000F930, 0x000064C4}, {0x0000F931, 0x00006AD3}, {0x0000F932, 0x00007210},
-{0x0000F933, 0x000076E7}, {0x0000F934, 0x00008001}, {0x0000F935, 0x00008606}, {0x0000F936, 0x0000865C},
-{0x0000F937, 0x00008DEF}, {0x0000F938, 0x00009732}, {0x0000F939, 0x00009B6F}, {0x0000F93A, 0x00009DFA},
-{0x0000F93B, 0x0000788C}, {0x0000F93C, 0x0000797F}, {0x0000F93D, 0x00007DA0}, {0x0000F93E, 0x000083C9},
-{0x0000F93F, 0x00009304}, {0x0000F940, 0x00009E7F}, {0x0000F941, 0x00008AD6}, {0x0000F942, 0x000058DF},
-{0x0000F943, 0x00005F04}, {0x0000F944, 0x00007C60}, {0x0000F945, 0x0000807E}, {0x0000F946, 0x00007262},
-{0x0000F947, 0x000078CA}, {0x0000F948, 0x00008CC2}, {0x0000F949, 0x000096F7}, {0x0000F94A, 0x000058D8},
-{0x0000F94B, 0x00005C62}, {0x0000F94C, 0x00006A13}, {0x0000F94D, 0x00006DDA}, {0x0000F94E, 0x00006F0F},
-{0x0000F94F, 0x00007D2F}, {0x0000F950, 0x00007E37}, {0x0000F951, 0x0000964B}, {0x0000F952, 0x000052D2},
-{0x0000F953, 0x0000808B}, {0x0000F954, 0x000051DC}, {0x0000F955, 0x000051CC}, {0x0000F956, 0x00007A1C},
-{0x0000F957, 0x00007DBE}, {0x0000F958, 0x000083F1}, {0x0000F959, 0x00009675}, {0x0000F95A, 0x00008B80},
-{0x0000F95B, 0x000062CF}, {0x0000F95C, 0x00006A02}, {0x0000F95D, 0x00008AFE}, {0x0000F95E, 0x00004E39},
-{0x0000F95F, 0x00005BE7}, {0x0000F960, 0x00006012}, {0x0000F961, 0x00007387}, {0x0000F962, 0x00007570},
-{0x0000F963, 0x00005317}, {0x0000F964, 0x000078FB}, {0x0000F965, 0x00004FBF}, {0x0000F966, 0x00005FA9},
-{0x0000F967, 0x00004E0D}, {0x0000F968, 0x00006CCC}, {0x0000F969, 0x00006578}, {0x0000F96A, 0x00007D22},
-{0x0000F96B, 0x000053C3}, {0x0000F96C, 0x0000585E}, {0x0000F96D, 0x00007701}, {0x0000F96E, 0x00008449},
-{0x0000F96F, 0x00008AAA}, {0x0000F970, 0x00006BBA}, {0x0000F971, 0x00008FB0}, {0x0000F972, 0x00006C88},
-{0x0000F973, 0x000062FE}, {0x0000F974, 0x000082E5}, {0x0000F975, 0x000063A0}, {0x0000F976, 0x00007565},
-{0x0000F977, 0x00004EAE}, {0x0000F978, 0x00005169}, {0x0000F979, 0x000051C9}, {0x0000F97A, 0x00006881},
-{0x0000F97B, 0x00007CE7}, {0x0000F97C, 0x0000826F}, {0x0000F97D, 0x00008AD2}, {0x0000F97E, 0x000091CF},
-{0x0000F97F, 0x000052F5}, {0x0000F980, 0x00005442}, {0x0000F981, 0x00005973}, {0x0000F982, 0x00005EEC},
-{0x0000F983, 0x000065C5}, {0x0000F984, 0x00006FFE}, {0x0000F985, 0x0000792A}, {0x0000F986, 0x000095AD},
-{0x0000F987, 0x00009A6A}, {0x0000F988, 0x00009E97}, {0x0000F989, 0x00009ECE}, {0x0000F98A, 0x0000529B},
-{0x0000F98B, 0x000066C6}, {0x0000F98C, 0x00006B77}, {0x0000F98D, 0x00008F62}, {0x0000F98E, 0x00005E74},
-{0x0000F98F, 0x00006190}, {0x0000F990, 0x00006200}, {0x0000F991, 0x0000649A}, {0x0000F992, 0x00006F23},
-{0x0000F993, 0x00007149}, {0x0000F994, 0x00007489}, {0x0000F995, 0x000079CA}, {0x0000F996, 0x00007DF4},
-{0x0000F997, 0x0000806F}, {0x0000F998, 0x00008F26}, {0x0000F999, 0x000084EE}, {0x0000F99A, 0x00009023},
-{0x0000F99B, 0x0000934A}, {0x0000F99C, 0x00005217}, {0x0000F99D, 0x000052A3}, {0x0000F99E, 0x000054BD},
-{0x0000F99F, 0x000070C8}, {0x0000F9A0, 0x000088C2}, {0x0000F9A1, 0x00008AAA}, {0x0000F9A2, 0x00005EC9},
-{0x0000F9A3, 0x00005FF5}, {0x0000F9A4, 0x0000637B}, {0x0000F9A5, 0x00006BAE}, {0x0000F9A6, 0x00007C3E},
-{0x0000F9A7, 0x00007375}, {0x0000F9A8, 0x00004EE4}, {0x0000F9A9, 0x000056F9}, {0x0000F9AA, 0x00005BE7},
-{0x0000F9AB, 0x00005DBA}, {0x0000F9AC, 0x0000601C}, {0x0000F9AD, 0x000073B2}, {0x0000F9AE, 0x00007469},
-{0x0000F9AF, 0x00007F9A}, {0x0000F9B0, 0x00008046}, {0x0000F9B1, 0x00009234}, {0x0000F9B2, 0x000096F6},
-{0x0000F9B3, 0x00009748}, {0x0000F9B4, 0x00009818}, {0x0000F9B5, 0x00004F8B}, {0x0000F9B6, 0x000079AE},
-{0x0000F9B7, 0x000091B4}, {0x0000F9B8, 0x000096B8}, {0x0000F9B9, 0x000060E1}, {0x0000F9BA, 0x00004E86},
-{0x0000F9BB, 0x000050DA}, {0x0000F9BC, 0x00005BEE}, {0x0000F9BD, 0x00005C3F}, {0x0000F9BE, 0x00006599},
-{0x0000F9BF, 0x00006A02}, {0x0000F9C0, 0x000071CE}, {0x0000F9C1, 0x00007642}, {0x0000F9C2, 0x000084FC},
-{0x0000F9C3, 0x0000907C}, {0x0000F9C4, 0x00009F8D}, {0x0000F9C5, 0x00006688}, {0x0000F9C6, 0x0000962E},
-{0x0000F9C7, 0x00005289}, {0x0000F9C8, 0x0000677B}, {0x0000F9C9, 0x000067F3}, {0x0000F9CA, 0x00006D41},
-{0x0000F9CB, 0x00006E9C}, {0x0000F9CC, 0x00007409}, {0x0000F9CD, 0x00007559}, {0x0000F9CE, 0x0000786B},
-{0x0000F9CF, 0x00007D10}, {0x0000F9D0, 0x0000985E}, {0x0000F9D1, 0x0000516D}, {0x0000F9D2, 0x0000622E},
-{0x0000F9D3, 0x00009678}, {0x0000F9D4, 0x0000502B}, {0x0000F9D5, 0x00005D19}, {0x0000F9D6, 0x00006DEA},
-{0x0000F9D7, 0x00008F2A}, {0x0000F9D8, 0x00005F8B}, {0x0000F9D9, 0x00006144}, {0x0000F9DA, 0x00006817},
-{0x0000F9DB, 0x00007387}, {0x0000F9DC, 0x00009686}, {0x0000F9DD, 0x00005229}, {0x0000F9DE, 0x0000540F},
-{0x0000F9DF, 0x00005C65}, {0x0000F9E0, 0x00006613}, {0x0000F9E1, 0x0000674E}, {0x0000F9E2, 0x000068A8},
-{0x0000F9E3, 0x00006CE5}, {0x0000F9E4, 0x00007406}, {0x0000F9E5, 0x000075E2}, {0x0000F9E6, 0x00007F79},
-{0x0000F9E7, 0x000088CF}, {0x0000F9E8, 0x000088E1}, {0x0000F9E9, 0x000091CC}, {0x0000F9EA, 0x000096E2},
-{0x0000F9EB, 0x0000533F}, {0x0000F9EC, 0x00006EBA}, {0x0000F9ED, 0x0000541D}, {0x0000F9EE, 0x000071D0},
-{0x0000F9EF, 0x00007498}, {0x0000F9F0, 0x000085FA}, {0x0000F9F1, 0x000096A3}, {0x0000F9F2, 0x00009C57},
-{0x0000F9F3, 0x00009E9F}, {0x0000F9F4, 0x00006797}, {0x0000F9F5, 0x00006DCB}, {0x0000F9F6, 0x000081E8},
-{0x0000F9F7, 0x00007ACB}, {0x0000F9F8, 0x00007B20}, {0x0000F9F9, 0x00007C92}, {0x0000F9FA, 0x000072C0},
-{0x0000F9FB, 0x00007099}, {0x0000F9FC, 0x00008B58}, {0x0000F9FD, 0x00004EC0}, {0x0000F9FE, 0x00008336},
-{0x0000F9FF, 0x0000523A}, {0x0000FA00, 0x00005207}, {0x0000FA01, 0x00005EA6}, {0x0000FA02, 0x000062D3},
-{0x0000FA03, 0x00007CD6}, {0x0000FA04, 0x00005B85}, {0x0000FA05, 0x00006D1E}, {0x0000FA06, 0x000066B4},
-{0x0000FA07, 0x00008F3B}, {0x0000FA08, 0x0000884C}, {0x0000FA09, 0x0000964D}, {0x0000FA0A, 0x0000898B},
-{0x0000FA0B, 0x00005ED3}, {0x0000FA0C, 0x00005140}, {0x0000FA0D, 0x000055C0}, {0x0000FA10, 0x0000585A},
-{0x0000FA12, 0x00006674}, {0x0000FA15, 0x000051DE}, {0x0000FA16, 0x0000732A}, {0x0000FA17, 0x000076CA},
-{0x0000FA18, 0x0000793C}, {0x0000FA19, 0x0000795E}, {0x0000FA1A, 0x00007965}, {0x0000FA1B, 0x0000798F},
-{0x0000FA1C, 0x00009756}, {0x0000FA1D, 0x00007CBE}, {0x0000FA1E, 0x00007FBD}, {0x0000FA20, 0x00008612},
-{0x0000FA22, 0x00008AF8}, {0x0000FA25, 0x00009038}, {0x0000FA26, 0x000090FD}, {0x0000FA2A, 0x000098EF},
-{0x0000FA2B, 0x000098FC}, {0x0000FA2C, 0x00009928}, {0x0000FA2D, 0x00009DB4}, {0x0000FA2E, 0x000090DE},
-{0x0000FA2F, 0x000096B7}, {0x0000FA30, 0x00004FAE}, {0x0000FA31, 0x000050E7}, {0x0000FA32, 0x0000514D},
-{0x0000FA33, 0x000052C9}, {0x0000FA34, 0x000052E4}, {0x0000FA35, 0x00005351}, {0x0000FA36, 0x0000559D},
-{0x0000FA37, 0x00005606}, {0x0000FA38, 0x00005668}, {0x0000FA39, 0x00005840}, {0x0000FA3A, 0x000058A8},
-{0x0000FA3B, 0x00005C64}, {0x0000FA3C, 0x00005C6E}, {0x0000FA3D, 0x00006094}, {0x0000FA3E, 0x00006168},
-{0x0000FA3F, 0x0000618E}, {0x0000FA40, 0x000061F2}, {0x0000FA41, 0x0000654F}, {0x0000FA42, 0x000065E2},
-{0x0000FA43, 0x00006691}, {0x0000FA44, 0x00006885}, {0x0000FA45, 0x00006D77}, {0x0000FA46, 0x00006E1A},
-{0x0000FA47, 0x00006F22}, {0x0000FA48, 0x0000716E}, {0x0000FA49, 0x0000722B}, {0x0000FA4A, 0x00007422},
-{0x0000FA4B, 0x00007891}, {0x0000FA4C, 0x0000793E}, {0x0000FA4D, 0x00007949}, {0x0000FA4E, 0x00007948},
-{0x0000FA4F, 0x00007950}, {0x0000FA50, 0x00007956}, {0x0000FA51, 0x0000795D}, {0x0000FA52, 0x0000798D},
-{0x0000FA53, 0x0000798E}, {0x0000FA54, 0x00007A40}, {0x0000FA55, 0x00007A81}, {0x0000FA56, 0x00007BC0},
-{0x0000FA57, 0x00007DF4}, {0x0000FA58, 0x00007E09}, {0x0000FA59, 0x00007E41}, {0x0000FA5A, 0x00007F72},
-{0x0000FA5B, 0x00008005}, {0x0000FA5C, 0x000081ED}, {0x0000FA5D, 0x00008279}, {0x0000FA5E, 0x00008279},
-{0x0000FA5F, 0x00008457}, {0x0000FA60, 0x00008910}, {0x0000FA61, 0x00008996}, {0x0000FA62, 0x00008B01},
-{0x0000FA63, 0x00008B39}, {0x0000FA64, 0x00008CD3}, {0x0000FA65, 0x00008D08}, {0x0000FA66, 0x00008FB6},
-{0x0000FA67, 0x00009038}, {0x0000FA68, 0x000096E3}, {0x0000FA69, 0x000097FF}, {0x0000FA6A, 0x0000983B},
-{0x0000FA6B, 0x00006075}, {0x0000FA6C, 0x000242EE}, {0x0000FA6D, 0x00008218}, {0x0000FA70, 0x00004E26},
-{0x0000FA71, 0x000051B5}, {0x0000FA72, 0x00005168}, {0x0000FA73, 0x00004F80}, {0x0000FA74, 0x00005145},
-{0x0000FA75, 0x00005180}, {0x0000FA76, 0x000052C7}, {0x0000FA77, 0x000052FA}, {0x0000FA78, 0x0000559D},
-{0x0000FA79, 0x00005555}, {0x0000FA7A, 0x00005599}, {0x0000FA7B, 0x000055E2}, {0x0000FA7C, 0x0000585A},
-{0x0000FA7D, 0x000058B3}, {0x0000FA7E, 0x00005944}, {0x0000FA7F, 0x00005954}, {0x0000FA80, 0x00005A62},
-{0x0000FA81, 0x00005B28}, {0x0000FA82, 0x00005ED2}, {0x0000FA83, 0x00005ED9}, {0x0000FA84, 0x00005F69},
-{0x0000FA85, 0x00005FAD}, {0x0000FA86, 0x000060D8}, {0x0000FA87, 0x0000614E}, {0x0000FA88, 0x00006108},
-{0x0000FA89, 0x0000618E}, {0x0000FA8A, 0x00006160}, {0x0000FA8B, 0x000061F2}, {0x0000FA8C, 0x00006234},
-{0x0000FA8D, 0x000063C4}, {0x0000FA8E, 0x0000641C}, {0x0000FA8F, 0x00006452}, {0x0000FA90, 0x00006556},
-{0x0000FA91, 0x00006674}, {0x0000FA92, 0x00006717}, {0x0000FA93, 0x0000671B}, {0x0000FA94, 0x00006756},
-{0x0000FA95, 0x00006B79}, {0x0000FA96, 0x00006BBA}, {0x0000FA97, 0x00006D41}, {0x0000FA98, 0x00006EDB},
-{0x0000FA99, 0x00006ECB}, {0x0000FA9A, 0x00006F22}, {0x0000FA9B, 0x0000701E}, {0x0000FA9C, 0x0000716E},
-{0x0000FA9D, 0x000077A7}, {0x0000FA9E, 0x00007235}, {0x0000FA9F, 0x000072AF}, {0x0000FAA0, 0x0000732A},
-{0x0000FAA1, 0x00007471}, {0x0000FAA2, 0x00007506}, {0x0000FAA3, 0x0000753B}, {0x0000FAA4, 0x0000761D},
-{0x0000FAA5, 0x0000761F}, {0x0000FAA6, 0x000076CA}, {0x0000FAA7, 0x000076DB}, {0x0000FAA8, 0x000076F4},
-{0x0000FAA9, 0x0000774A}, {0x0000FAAA, 0x00007740}, {0x0000FAAB, 0x000078CC}, {0x0000FAAC, 0x00007AB1},
-{0x0000FAAD, 0x00007BC0}, {0x0000FAAE, 0x00007C7B}, {0x0000FAAF, 0x00007D5B}, {0x0000FAB0, 0x00007DF4},
-{0x0000FAB1, 0x00007F3E}, {0x0000FAB2, 0x00008005}, {0x0000FAB3, 0x00008352}, {0x0000FAB4, 0x000083EF},
-{0x0000FAB5, 0x00008779}, {0x0000FAB6, 0x00008941}, {0x0000FAB7, 0x00008986}, {0x0000FAB8, 0x00008996},
-{0x0000FAB9, 0x00008ABF}, {0x0000FABA, 0x00008AF8}, {0x0000FABB, 0x00008ACB}, {0x0000FABC, 0x00008B01},
-{0x0000FABD, 0x00008AFE}, {0x0000FABE, 0x00008AED}, {0x0000FABF, 0x00008B39}, {0x0000FAC0, 0x00008B8A},
-{0x0000FAC1, 0x00008D08}, {0x0000FAC2, 0x00008F38}, {0x0000FAC3, 0x00009072}, {0x0000FAC4, 0x00009199},
-{0x0000FAC5, 0x00009276}, {0x0000FAC6, 0x0000967C}, {0x0000FAC7, 0x000096E3}, {0x0000FAC8, 0x00009756},
-{0x0000FAC9, 0x000097DB}, {0x0000FACA, 0x000097FF}, {0x0000FACB, 0x0000980B}, {0x0000FACC, 0x0000983B},
-{0x0000FACD, 0x00009B12}, {0x0000FACE, 0x00009F9C}, {0x0000FACF, 0x0002284A}, {0x0000FAD0, 0x00022844},
-{0x0000FAD1, 0x000233D5}, {0x0000FAD2, 0x00003B9D}, {0x0000FAD3, 0x00004018}, {0x0000FAD4, 0x00004039},
-{0x0000FAD5, 0x00025249}, {0x0000FAD6, 0x00025CD0}, {0x0000FAD7, 0x00027ED3}, {0x0000FAD8, 0x00009F43},
-{0x0000FAD9, 0x00009F8E}, {0x0000FB1D, 0x000005D9}, {0x0000FB1D, 0x000005B4}, {0x0000FB1F, 0x000005F2},
-{0x0000FB1F, 0x000005B7}, {0x0000FB2A, 0x000005E9}, {0x0000FB2A, 0x000005C1}, {0x0000FB2B, 0x000005E9},
-{0x0000FB2B, 0x000005C2}, {0x0000FB2C, 0x000005E9}, {0x0000FB2C, 0x000005BC}, {0x0000FB2C, 0x000005C1},
-{0x0000FB2D, 0x000005E9}, {0x0000FB2D, 0x000005BC}, {0x0000FB2D, 0x000005C2}, {0x0000FB2E, 0x000005D0},
-{0x0000FB2E, 0x000005B7}, {0x0000FB2F, 0x000005D0}, {0x0000FB2F, 0x000005B8}, {0x0000FB30, 0x000005D0},
-{0x0000FB30, 0x000005BC}, {0x0000FB31, 0x000005D1}, {0x0000FB31, 0x000005BC}, {0x0000FB32, 0x000005D2},
-{0x0000FB32, 0x000005BC}, {0x0000FB33, 0x000005D3}, {0x0000FB33, 0x000005BC}, {0x0000FB34, 0x000005D4},
-{0x0000FB34, 0x000005BC}, {0x0000FB35, 0x000005D5}, {0x0000FB35, 0x000005BC}, {0x0000FB36, 0x000005D6},
-{0x0000FB36, 0x000005BC}, {0x0000FB38, 0x000005D8}, {0x0000FB38, 0x000005BC}, {0x0000FB39, 0x000005D9},
-{0x0000FB39, 0x000005BC}, {0x0000FB3A, 0x000005DA}, {0x0000FB3A, 0x000005BC}, {0x0000FB3B, 0x000005DB},
-{0x0000FB3B, 0x000005BC}, {0x0000FB3C, 0x000005DC}, {0x0000FB3C, 0x000005BC}, {0x0000FB3E, 0x000005DE},
-{0x0000FB3E, 0x000005BC}, {0x0000FB40, 0x000005E0}, {0x0000FB40, 0x000005BC}, {0x0000FB41, 0x000005E1},
-{0x0000FB41, 0x000005BC}, {0x0000FB43, 0x000005E3}, {0x0000FB43, 0x000005BC}, {0x0000FB44, 0x000005E4},
-{0x0000FB44, 0x000005BC}, {0x0000FB46, 0x000005E6}, {0x0000FB46, 0x000005BC}, {0x0000FB47, 0x000005E7},
-{0x0000FB47, 0x000005BC}, {0x0000FB48, 0x000005E8}, {0x0000FB48, 0x000005BC}, {0x0000FB49, 0x000005E9},
-{0x0000FB49, 0x000005BC}, {0x0000FB4A, 0x000005EA}, {0x0000FB4A, 0x000005BC}, {0x0000FB4B, 0x000005D5},
-{0x0000FB4B, 0x000005B9}, {0x0000FB4C, 0x000005D1}, {0x0000FB4C, 0x000005BF}, {0x0000FB4D, 0x000005DB},
-{0x0000FB4D, 0x000005BF}, {0x0000FB4E, 0x000005E4}, {0x0000FB4E, 0x000005BF}, {0x0001109A, 0x00011099},
-{0x0001109A, 0x000110BA}, {0x0001109C, 0x0001109B}, {0x0001109C, 0x000110BA}, {0x000110AB, 0x000110A5},
-{0x000110AB, 0x000110BA}, {0x0001112E, 0x00011131}, {0x0001112E, 0x00011127}, {0x0001112F, 0x00011132},
-{0x0001112F, 0x00011127}, {0x0001134B, 0x00011347}, {0x0001134B, 0x0001133E}, {0x0001134C, 0x00011347},
-{0x0001134C, 0x00011357}, {0x000114BB, 0x000114B9}, {0x000114BB, 0x000114BA}, {0x000114BC, 0x000114B9},
-{0x000114BC, 0x000114B0}, {0x000114BE, 0x000114B9}, {0x000114BE, 0x000114BD}, {0x000115BA, 0x000115B8},
-{0x000115BA, 0x000115AF}, {0x000115BB, 0x000115B9}, {0x000115BB, 0x000115AF}, {0x0001D15E, 0x0001D157},
-{0x0001D15E, 0x0001D165}, {0x0001D15F, 0x0001D158}, {0x0001D15F, 0x0001D165}, {0x0001D160, 0x0001D158},
-{0x0001D160, 0x0001D165}, {0x0001D160, 0x0001D16E}, {0x0001D161, 0x0001D158}, {0x0001D161, 0x0001D165},
-{0x0001D161, 0x0001D16F}, {0x0001D162, 0x0001D158}, {0x0001D162, 0x0001D165}, {0x0001D162, 0x0001D170},
-{0x0001D163, 0x0001D158}, {0x0001D163, 0x0001D165}, {0x0001D163, 0x0001D171}, {0x0001D164, 0x0001D158},
-{0x0001D164, 0x0001D165}, {0x0001D164, 0x0001D172}, {0x0001D1BB, 0x0001D1B9}, {0x0001D1BB, 0x0001D165},
-{0x0001D1BC, 0x0001D1BA}, {0x0001D1BC, 0x0001D165}, {0x0001D1BD, 0x0001D1B9}, {0x0001D1BD, 0x0001D165},
-{0x0001D1BD, 0x0001D16E}, {0x0001D1BE, 0x0001D1BA}, {0x0001D1BE, 0x0001D165}, {0x0001D1BE, 0x0001D16E},
-{0x0001D1BF, 0x0001D1B9}, {0x0001D1BF, 0x0001D165}, {0x0001D1BF, 0x0001D16F}, {0x0001D1C0, 0x0001D1BA},
-{0x0001D1C0, 0x0001D165}, {0x0001D1C0, 0x0001D16F}, {0x0002F800, 0x00004E3D}, {0x0002F801, 0x00004E38},
-{0x0002F802, 0x00004E41}, {0x0002F803, 0x00020122}, {0x0002F804, 0x00004F60}, {0x0002F805, 0x00004FAE},
-{0x0002F806, 0x00004FBB}, {0x0002F807, 0x00005002}, {0x0002F808, 0x0000507A}, {0x0002F809, 0x00005099},
-{0x0002F80A, 0x000050E7}, {0x0002F80B, 0x000050CF}, {0x0002F80C, 0x0000349E}, {0x0002F80D, 0x0002063A},
-{0x0002F80E, 0x0000514D}, {0x0002F80F, 0x00005154}, {0x0002F810, 0x00005164}, {0x0002F811, 0x00005177},
-{0x0002F812, 0x0002051C}, {0x0002F813, 0x000034B9}, {0x0002F814, 0x00005167}, {0x0002F815, 0x0000518D},
-{0x0002F816, 0x0002054B}, {0x0002F817, 0x00005197}, {0x0002F818, 0x000051A4}, {0x0002F819, 0x00004ECC},
-{0x0002F81A, 0x000051AC}, {0x0002F81B, 0x000051B5}, {0x0002F81C, 0x000291DF}, {0x0002F81D, 0x000051F5},
-{0x0002F81E, 0x00005203}, {0x0002F81F, 0x000034DF}, {0x0002F820, 0x0000523B}, {0x0002F821, 0x00005246},
-{0x0002F822, 0x00005272}, {0x0002F823, 0x00005277}, {0x0002F824, 0x00003515}, {0x0002F825, 0x000052C7},
-{0x0002F826, 0x000052C9}, {0x0002F827, 0x000052E4}, {0x0002F828, 0x000052FA}, {0x0002F829, 0x00005305},
-{0x0002F82A, 0x00005306}, {0x0002F82B, 0x00005317}, {0x0002F82C, 0x00005349}, {0x0002F82D, 0x00005351},
-{0x0002F82E, 0x0000535A}, {0x0002F82F, 0x00005373}, {0x0002F830, 0x0000537D}, {0x0002F831, 0x0000537F},
-{0x0002F832, 0x0000537F}, {0x0002F833, 0x0000537F}, {0x0002F834, 0x00020A2C}, {0x0002F835, 0x00007070},
-{0x0002F836, 0x000053CA}, {0x0002F837, 0x000053DF}, {0x0002F838, 0x00020B63}, {0x0002F839, 0x000053EB},
-{0x0002F83A, 0x000053F1}, {0x0002F83B, 0x00005406}, {0x0002F83C, 0x0000549E}, {0x0002F83D, 0x00005438},
-{0x0002F83E, 0x00005448}, {0x0002F83F, 0x00005468}, {0x0002F840, 0x000054A2}, {0x0002F841, 0x000054F6},
-{0x0002F842, 0x00005510}, {0x0002F843, 0x00005553}, {0x0002F844, 0x00005563}, {0x0002F845, 0x00005584},
-{0x0002F846, 0x00005584}, {0x0002F847, 0x00005599}, {0x0002F848, 0x000055AB}, {0x0002F849, 0x000055B3},
-{0x0002F84A, 0x000055C2}, {0x0002F84B, 0x00005716}, {0x0002F84C, 0x00005606}, {0x0002F84D, 0x00005717},
-{0x0002F84E, 0x00005651}, {0x0002F84F, 0x00005674}, {0x0002F850, 0x00005207}, {0x0002F851, 0x000058EE},
-{0x0002F852, 0x000057CE}, {0x0002F853, 0x000057F4}, {0x0002F854, 0x0000580D}, {0x0002F855, 0x0000578B},
-{0x0002F856, 0x00005832}, {0x0002F857, 0x00005831}, {0x0002F858, 0x000058AC}, {0x0002F859, 0x000214E4},
-{0x0002F85A, 0x000058F2}, {0x0002F85B, 0x000058F7}, {0x0002F85C, 0x00005906}, {0x0002F85D, 0x0000591A},
-{0x0002F85E, 0x00005922}, {0x0002F85F, 0x00005962}, {0x0002F860, 0x000216A8}, {0x0002F861, 0x000216EA},
-{0x0002F862, 0x000059EC}, {0x0002F863, 0x00005A1B}, {0x0002F864, 0x00005A27}, {0x0002F865, 0x000059D8},
-{0x0002F866, 0x00005A66}, {0x0002F867, 0x000036EE}, {0x0002F868, 0x000036FC}, {0x0002F869, 0x00005B08},
-{0x0002F86A, 0x00005B3E}, {0x0002F86B, 0x00005B3E}, {0x0002F86C, 0x000219C8}, {0x0002F86D, 0x00005BC3},
-{0x0002F86E, 0x00005BD8}, {0x0002F86F, 0x00005BE7}, {0x0002F870, 0x00005BF3}, {0x0002F871, 0x00021B18},
-{0x0002F872, 0x00005BFF}, {0x0002F873, 0x00005C06}, {0x0002F874, 0x00005F53}, {0x0002F875, 0x00005C22},
-{0x0002F876, 0x00003781}, {0x0002F877, 0x00005C60}, {0x0002F878, 0x00005C6E}, {0x0002F879, 0x00005CC0},
-{0x0002F87A, 0x00005C8D}, {0x0002F87B, 0x00021DE4}, {0x0002F87C, 0x00005D43}, {0x0002F87D, 0x00021DE6},
-{0x0002F87E, 0x00005D6E}, {0x0002F87F, 0x00005D6B}, {0x0002F880, 0x00005D7C}, {0x0002F881, 0x00005DE1},
-{0x0002F882, 0x00005DE2}, {0x0002F883, 0x0000382F}, {0x0002F884, 0x00005DFD}, {0x0002F885, 0x00005E28},
-{0x0002F886, 0x00005E3D}, {0x0002F887, 0x00005E69}, {0x0002F888, 0x00003862}, {0x0002F889, 0x00022183},
-{0x0002F88A, 0x0000387C}, {0x0002F88B, 0x00005EB0}, {0x0002F88C, 0x00005EB3}, {0x0002F88D, 0x00005EB6},
-{0x0002F88E, 0x00005ECA}, {0x0002F88F, 0x0002A392}, {0x0002F890, 0x00005EFE}, {0x0002F891, 0x00022331},
-{0x0002F892, 0x00022331}, {0x0002F893, 0x00008201}, {0x0002F894, 0x00005F22}, {0x0002F895, 0x00005F22},
-{0x0002F896, 0x000038C7}, {0x0002F897, 0x000232B8}, {0x0002F898, 0x000261DA}, {0x0002F899, 0x00005F62},
-{0x0002F89A, 0x00005F6B}, {0x0002F89B, 0x000038E3}, {0x0002F89C, 0x00005F9A}, {0x0002F89D, 0x00005FCD},
-{0x0002F89E, 0x00005FD7}, {0x0002F89F, 0x00005FF9}, {0x0002F8A0, 0x00006081}, {0x0002F8A1, 0x0000393A},
-{0x0002F8A2, 0x0000391C}, {0x0002F8A3, 0x00006094}, {0x0002F8A4, 0x000226D4}, {0x0002F8A5, 0x000060C7},
-{0x0002F8A6, 0x00006148}, {0x0002F8A7, 0x0000614C}, {0x0002F8A8, 0x0000614E}, {0x0002F8A9, 0x0000614C},
-{0x0002F8AA, 0x0000617A}, {0x0002F8AB, 0x0000618E}, {0x0002F8AC, 0x000061B2}, {0x0002F8AD, 0x000061A4},
-{0x0002F8AE, 0x000061AF}, {0x0002F8AF, 0x000061DE}, {0x0002F8B0, 0x000061F2}, {0x0002F8B1, 0x000061F6},
-{0x0002F8B2, 0x00006210}, {0x0002F8B3, 0x0000621B}, {0x0002F8B4, 0x0000625D}, {0x0002F8B5, 0x000062B1},
-{0x0002F8B6, 0x000062D4}, {0x0002F8B7, 0x00006350}, {0x0002F8B8, 0x00022B0C}, {0x0002F8B9, 0x0000633D},
-{0x0002F8BA, 0x000062FC}, {0x0002F8BB, 0x00006368}, {0x0002F8BC, 0x00006383}, {0x0002F8BD, 0x000063E4},
-{0x0002F8BE, 0x00022BF1}, {0x0002F8BF, 0x00006422}, {0x0002F8C0, 0x000063C5}, {0x0002F8C1, 0x000063A9},
-{0x0002F8C2, 0x00003A2E}, {0x0002F8C3, 0x00006469}, {0x0002F8C4, 0x0000647E}, {0x0002F8C5, 0x0000649D},
-{0x0002F8C6, 0x00006477}, {0x0002F8C7, 0x00003A6C}, {0x0002F8C8, 0x0000654F}, {0x0002F8C9, 0x0000656C},
-{0x0002F8CA, 0x0002300A}, {0x0002F8CB, 0x000065E3}, {0x0002F8CC, 0x000066F8}, {0x0002F8CD, 0x00006649},
-{0x0002F8CE, 0x00003B19}, {0x0002F8CF, 0x00006691}, {0x0002F8D0, 0x00003B08}, {0x0002F8D1, 0x00003AE4},
-{0x0002F8D2, 0x00005192}, {0x0002F8D3, 0x00005195}, {0x0002F8D4, 0x00006700}, {0x0002F8D5, 0x0000669C},
-{0x0002F8D6, 0x000080AD}, {0x0002F8D7, 0x000043D9}, {0x0002F8D8, 0x00006717}, {0x0002F8D9, 0x0000671B},
-{0x0002F8DA, 0x00006721}, {0x0002F8DB, 0x0000675E}, {0x0002F8DC, 0x00006753}, {0x0002F8DD, 0x000233C3},
-{0x0002F8DE, 0x00003B49}, {0x0002F8DF, 0x000067FA}, {0x0002F8E0, 0x00006785}, {0x0002F8E1, 0x00006852},
-{0x0002F8E2, 0x00006885}, {0x0002F8E3, 0x0002346D}, {0x0002F8E4, 0x0000688E}, {0x0002F8E5, 0x0000681F},
-{0x0002F8E6, 0x00006914}, {0x0002F8E7, 0x00003B9D}, {0x0002F8E8, 0x00006942}, {0x0002F8E9, 0x000069A3},
-{0x0002F8EA, 0x000069EA}, {0x0002F8EB, 0x00006AA8}, {0x0002F8EC, 0x000236A3}, {0x0002F8ED, 0x00006ADB},
-{0x0002F8EE, 0x00003C18}, {0x0002F8EF, 0x00006B21}, {0x0002F8F0, 0x000238A7}, {0x0002F8F1, 0x00006B54},
-{0x0002F8F2, 0x00003C4E}, {0x0002F8F3, 0x00006B72}, {0x0002F8F4, 0x00006B9F}, {0x0002F8F5, 0x00006BBA},
-{0x0002F8F6, 0x00006BBB}, {0x0002F8F7, 0x00023A8D}, {0x0002F8F8, 0x00021D0B}, {0x0002F8F9, 0x00023AFA},
-{0x0002F8FA, 0x00006C4E}, {0x0002F8FB, 0x00023CBC}, {0x0002F8FC, 0x00006CBF}, {0x0002F8FD, 0x00006CCD},
-{0x0002F8FE, 0x00006C67}, {0x0002F8FF, 0x00006D16}, {0x0002F900, 0x00006D3E}, {0x0002F901, 0x00006D77},
-{0x0002F902, 0x00006D41}, {0x0002F903, 0x00006D69}, {0x0002F904, 0x00006D78}, {0x0002F905, 0x00006D85},
-{0x0002F906, 0x00023D1E}, {0x0002F907, 0x00006D34}, {0x0002F908, 0x00006E2F}, {0x0002F909, 0x00006E6E},
-{0x0002F90A, 0x00003D33}, {0x0002F90B, 0x00006ECB}, {0x0002F90C, 0x00006EC7}, {0x0002F90D, 0x00023ED1},
-{0x0002F90E, 0x00006DF9}, {0x0002F90F, 0x00006F6E}, {0x0002F910, 0x00023F5E}, {0x0002F911, 0x00023F8E},
-{0x0002F912, 0x00006FC6}, {0x0002F913, 0x00007039}, {0x0002F914, 0x0000701E}, {0x0002F915, 0x0000701B},
-{0x0002F916, 0x00003D96}, {0x0002F917, 0x0000704A}, {0x0002F918, 0x0000707D}, {0x0002F919, 0x00007077},
-{0x0002F91A, 0x000070AD}, {0x0002F91B, 0x00020525}, {0x0002F91C, 0x00007145}, {0x0002F91D, 0x00024263},
-{0x0002F91E, 0x0000719C}, {0x0002F91F, 0x000243AB}, {0x0002F920, 0x00007228}, {0x0002F921, 0x00007235},
-{0x0002F922, 0x00007250}, {0x0002F923, 0x00024608}, {0x0002F924, 0x00007280}, {0x0002F925, 0x00007295},
-{0x0002F926, 0x00024735}, {0x0002F927, 0x00024814}, {0x0002F928, 0x0000737A}, {0x0002F929, 0x0000738B},
-{0x0002F92A, 0x00003EAC}, {0x0002F92B, 0x000073A5}, {0x0002F92C, 0x00003EB8}, {0x0002F92D, 0x00003EB8},
-{0x0002F92E, 0x00007447}, {0x0002F92F, 0x0000745C}, {0x0002F930, 0x00007471}, {0x0002F931, 0x00007485},
-{0x0002F932, 0x000074CA}, {0x0002F933, 0x00003F1B}, {0x0002F934, 0x00007524}, {0x0002F935, 0x00024C36},
-{0x0002F936, 0x0000753E}, {0x0002F937, 0x00024C92}, {0x0002F938, 0x00007570}, {0x0002F939, 0x0002219F},
-{0x0002F93A, 0x00007610}, {0x0002F93B, 0x00024FA1}, {0x0002F93C, 0x00024FB8}, {0x0002F93D, 0x00025044},
-{0x0002F93E, 0x00003FFC}, {0x0002F93F, 0x00004008}, {0x0002F940, 0x000076F4}, {0x0002F941, 0x000250F3},
-{0x0002F942, 0x000250F2}, {0x0002F943, 0x00025119}, {0x0002F944, 0x00025133}, {0x0002F945, 0x0000771E},
-{0x0002F946, 0x0000771F}, {0x0002F947, 0x0000771F}, {0x0002F948, 0x0000774A}, {0x0002F949, 0x00004039},
-{0x0002F94A, 0x0000778B}, {0x0002F94B, 0x00004046}, {0x0002F94C, 0x00004096}, {0x0002F94D, 0x0002541D},
-{0x0002F94E, 0x0000784E}, {0x0002F94F, 0x0000788C}, {0x0002F950, 0x000078CC}, {0x0002F951, 0x000040E3},
-{0x0002F952, 0x00025626}, {0x0002F953, 0x00007956}, {0x0002F954, 0x0002569A}, {0x0002F955, 0x000256C5},
-{0x0002F956, 0x0000798F}, {0x0002F957, 0x000079EB}, {0x0002F958, 0x0000412F}, {0x0002F959, 0x00007A40},
-{0x0002F95A, 0x00007A4A}, {0x0002F95B, 0x00007A4F}, {0x0002F95C, 0x0002597C}, {0x0002F95D, 0x00025AA7},
-{0x0002F95E, 0x00025AA7}, {0x0002F95F, 0x00007AEE}, {0x0002F960, 0x00004202}, {0x0002F961, 0x00025BAB},
-{0x0002F962, 0x00007BC6}, {0x0002F963, 0x00007BC9}, {0x0002F964, 0x00004227}, {0x0002F965, 0x00025C80},
-{0x0002F966, 0x00007CD2}, {0x0002F967, 0x000042A0}, {0x0002F968, 0x00007CE8}, {0x0002F969, 0x00007CE3},
-{0x0002F96A, 0x00007D00}, {0x0002F96B, 0x00025F86}, {0x0002F96C, 0x00007D63}, {0x0002F96D, 0x00004301},
-{0x0002F96E, 0x00007DC7}, {0x0002F96F, 0x00007E02}, {0x0002F970, 0x00007E45}, {0x0002F971, 0x00004334},
-{0x0002F972, 0x00026228}, {0x0002F973, 0x00026247}, {0x0002F974, 0x00004359}, {0x0002F975, 0x000262D9},
-{0x0002F976, 0x00007F7A}, {0x0002F977, 0x0002633E}, {0x0002F978, 0x00007F95}, {0x0002F979, 0x00007FFA},
-{0x0002F97A, 0x00008005}, {0x0002F97B, 0x000264DA}, {0x0002F97C, 0x00026523}, {0x0002F97D, 0x00008060},
-{0x0002F97E, 0x000265A8}, {0x0002F97F, 0x00008070}, {0x0002F980, 0x0002335F}, {0x0002F981, 0x000043D5},
-{0x0002F982, 0x000080B2}, {0x0002F983, 0x00008103}, {0x0002F984, 0x0000440B}, {0x0002F985, 0x0000813E},
-{0x0002F986, 0x00005AB5}, {0x0002F987, 0x000267A7}, {0x0002F988, 0x000267B5}, {0x0002F989, 0x00023393},
-{0x0002F98A, 0x0002339C}, {0x0002F98B, 0x00008201}, {0x0002F98C, 0x00008204}, {0x0002F98D, 0x00008F9E},
-{0x0002F98E, 0x0000446B}, {0x0002F98F, 0x00008291}, {0x0002F990, 0x0000828B}, {0x0002F991, 0x0000829D},
-{0x0002F992, 0x000052B3}, {0x0002F993, 0x000082B1}, {0x0002F994, 0x000082B3}, {0x0002F995, 0x000082BD},
-{0x0002F996, 0x000082E6}, {0x0002F997, 0x00026B3C}, {0x0002F998, 0x000082E5}, {0x0002F999, 0x0000831D},
-{0x0002F99A, 0x00008363}, {0x0002F99B, 0x000083AD}, {0x0002F99C, 0x00008323}, {0x0002F99D, 0x000083BD},
-{0x0002F99E, 0x000083E7}, {0x0002F99F, 0x00008457}, {0x0002F9A0, 0x00008353}, {0x0002F9A1, 0x000083CA},
-{0x0002F9A2, 0x000083CC}, {0x0002F9A3, 0x000083DC}, {0x0002F9A4, 0x00026C36}, {0x0002F9A5, 0x00026D6B},
-{0x0002F9A6, 0x00026CD5}, {0x0002F9A7, 0x0000452B}, {0x0002F9A8, 0x000084F1}, {0x0002F9A9, 0x000084F3},
-{0x0002F9AA, 0x00008516}, {0x0002F9AB, 0x000273CA}, {0x0002F9AC, 0x00008564}, {0x0002F9AD, 0x00026F2C},
-{0x0002F9AE, 0x0000455D}, {0x0002F9AF, 0x00004561}, {0x0002F9B0, 0x00026FB1}, {0x0002F9B1, 0x000270D2},
-{0x0002F9B2, 0x0000456B}, {0x0002F9B3, 0x00008650}, {0x0002F9B4, 0x0000865C}, {0x0002F9B5, 0x00008667},
-{0x0002F9B6, 0x00008669}, {0x0002F9B7, 0x000086A9}, {0x0002F9B8, 0x00008688}, {0x0002F9B9, 0x0000870E},
-{0x0002F9BA, 0x000086E2}, {0x0002F9BB, 0x00008779}, {0x0002F9BC, 0x00008728}, {0x0002F9BD, 0x0000876B},
-{0x0002F9BE, 0x00008786}, {0x0002F9BF, 0x000045D7}, {0x0002F9C0, 0x000087E1}, {0x0002F9C1, 0x00008801},
-{0x0002F9C2, 0x000045F9}, {0x0002F9C3, 0x00008860}, {0x0002F9C4, 0x00008863}, {0x0002F9C5, 0x00027667},
-{0x0002F9C6, 0x000088D7}, {0x0002F9C7, 0x000088DE}, {0x0002F9C8, 0x00004635}, {0x0002F9C9, 0x000088FA},
-{0x0002F9CA, 0x000034BB}, {0x0002F9CB, 0x000278AE}, {0x0002F9CC, 0x00027966}, {0x0002F9CD, 0x000046BE},
-{0x0002F9CE, 0x000046C7}, {0x0002F9CF, 0x00008AA0}, {0x0002F9D0, 0x00008AED}, {0x0002F9D1, 0x00008B8A},
-{0x0002F9D2, 0x00008C55}, {0x0002F9D3, 0x00027CA8}, {0x0002F9D4, 0x00008CAB}, {0x0002F9D5, 0x00008CC1},
-{0x0002F9D6, 0x00008D1B}, {0x0002F9D7, 0x00008D77}, {0x0002F9D8, 0x00027F2F}, {0x0002F9D9, 0x00020804},
-{0x0002F9DA, 0x00008DCB}, {0x0002F9DB, 0x00008DBC}, {0x0002F9DC, 0x00008DF0}, {0x0002F9DD, 0x000208DE},
-{0x0002F9DE, 0x00008ED4}, {0x0002F9DF, 0x00008F38}, {0x0002F9E0, 0x000285D2}, {0x0002F9E1, 0x000285ED},
-{0x0002F9E2, 0x00009094}, {0x0002F9E3, 0x000090F1}, {0x0002F9E4, 0x00009111}, {0x0002F9E5, 0x0002872E},
-{0x0002F9E6, 0x0000911B}, {0x0002F9E7, 0x00009238}, {0x0002F9E8, 0x000092D7}, {0x0002F9E9, 0x000092D8},
-{0x0002F9EA, 0x0000927C}, {0x0002F9EB, 0x000093F9}, {0x0002F9EC, 0x00009415}, {0x0002F9ED, 0x00028BFA},
-{0x0002F9EE, 0x0000958B}, {0x0002F9EF, 0x00004995}, {0x0002F9F0, 0x000095B7}, {0x0002F9F1, 0x00028D77},
-{0x0002F9F2, 0x000049E6}, {0x0002F9F3, 0x000096C3}, {0x0002F9F4, 0x00005DB2}, {0x0002F9F5, 0x00009723},
-{0x0002F9F6, 0x00029145}, {0x0002F9F7, 0x0002921A}, {0x0002F9F8, 0x00004A6E}, {0x0002F9F9, 0x00004A76},
-{0x0002F9FA, 0x000097E0}, {0x0002F9FB, 0x0002940A}, {0x0002F9FC, 0x00004AB2}, {0x0002F9FD, 0x00029496},
-{0x0002F9FE, 0x0000980B}, {0x0002F9FF, 0x0000980B}, {0x0002FA00, 0x00009829}, {0x0002FA01, 0x000295B6},
-{0x0002FA02, 0x000098E2}, {0x0002FA03, 0x00004B33}, {0x0002FA04, 0x00009929}, {0x0002FA05, 0x000099A7},
-{0x0002FA06, 0x000099C2}, {0x0002FA07, 0x000099FE}, {0x0002FA08, 0x00004BCE}, {0x0002FA09, 0x00029B30},
-{0x0002FA0A, 0x00009B12}, {0x0002FA0B, 0x00009C40}, {0x0002FA0C, 0x00009CFD}, {0x0002FA0D, 0x00004CCE},
-{0x0002FA0E, 0x00004CED}, {0x0002FA0F, 0x00009D67}, {0x0002FA10, 0x0002A0CE}, {0x0002FA11, 0x00004CF8},
-{0x0002FA12, 0x0002A105}, {0x0002FA13, 0x0002A20E}, {0x0002FA14, 0x0002A291}, {0x0002FA15, 0x00009EBB},
-{0x0002FA16, 0x00004D56}, {0x0002FA17, 0x00009EF9}, {0x0002FA18, 0x00009EFE}, {0x0002FA19, 0x00009F05},
-{0x0002FA1A, 0x00009F0F}, {0x0002FA1B, 0x00009F16}, {0x0002FA1D, 0x0002A600},
+const std::vector<range_nfd> unicode_ranges_nfd = {  // start, last, nfd
+{0x000000, 0x000000, 0x000000},
+{0x0000C0, 0x0000C5, 0x000041},
+{0x0000C7, 0x0000C7, 0x000043},
+{0x0000C8, 0x0000CB, 0x000045},
+{0x0000CC, 0x0000CF, 0x000049},
+{0x0000D1, 0x0000D1, 0x00004E},
+{0x0000D2, 0x0000D6, 0x00004F},
+{0x0000D9, 0x0000DC, 0x000055},
+{0x0000DD, 0x0000DD, 0x000059},
+{0x0000E0, 0x0000E5, 0x000061},
+{0x0000E7, 0x0000E7, 0x000063},
+{0x0000E8, 0x0000EB, 0x000065},
+{0x0000EC, 0x0000EF, 0x000069},
+{0x0000F1, 0x0000F1, 0x00006E},
+{0x0000F2, 0x0000F6, 0x00006F},
+{0x0000F9, 0x0000FC, 0x000075},
+{0x0000FD, 0x0000FD, 0x000079},
+{0x0000FF, 0x0000FF, 0x000079},
+{0x000100, 0x000100, 0x000041},
+{0x000101, 0x000101, 0x000061},
+{0x000102, 0x000102, 0x000041},
+{0x000103, 0x000103, 0x000061},
+{0x000104, 0x000104, 0x000041},
+{0x000105, 0x000105, 0x000061},
+{0x000106, 0x000106, 0x000043},
+{0x000107, 0x000107, 0x000063},
+{0x000108, 0x000108, 0x000043},
+{0x000109, 0x000109, 0x000063},
+{0x00010A, 0x00010A, 0x000043},
+{0x00010B, 0x00010B, 0x000063},
+{0x00010C, 0x00010C, 0x000043},
+{0x00010D, 0x00010D, 0x000063},
+{0x00010E, 0x00010E, 0x000044},
+{0x00010F, 0x00010F, 0x000064},
+{0x000112, 0x000112, 0x000045},
+{0x000113, 0x000113, 0x000065},
+{0x000114, 0x000114, 0x000045},
+{0x000115, 0x000115, 0x000065},
+{0x000116, 0x000116, 0x000045},
+{0x000117, 0x000117, 0x000065},
+{0x000118, 0x000118, 0x000045},
+{0x000119, 0x000119, 0x000065},
+{0x00011A, 0x00011A, 0x000045},
+{0x00011B, 0x00011B, 0x000065},
+{0x00011C, 0x00011C, 0x000047},
+{0x00011D, 0x00011D, 0x000067},
+{0x00011E, 0x00011E, 0x000047},
+{0x00011F, 0x00011F, 0x000067},
+{0x000120, 0x000120, 0x000047},
+{0x000121, 0x000121, 0x000067},
+{0x000122, 0x000122, 0x000047},
+{0x000123, 0x000123, 0x000067},
+{0x000124, 0x000124, 0x000048},
+{0x000125, 0x000125, 0x000068},
+{0x000128, 0x000128, 0x000049},
+{0x000129, 0x000129, 0x000069},
+{0x00012A, 0x00012A, 0x000049},
+{0x00012B, 0x00012B, 0x000069},
+{0x00012C, 0x00012C, 0x000049},
+{0x00012D, 0x00012D, 0x000069},
+{0x00012E, 0x00012E, 0x000049},
+{0x00012F, 0x00012F, 0x000069},
+{0x000130, 0x000130, 0x000049},
+{0x000134, 0x000134, 0x00004A},
+{0x000135, 0x000135, 0x00006A},
+{0x000136, 0x000136, 0x00004B},
+{0x000137, 0x000137, 0x00006B},
+{0x000139, 0x000139, 0x00004C},
+{0x00013A, 0x00013A, 0x00006C},
+{0x00013B, 0x00013B, 0x00004C},
+{0x00013C, 0x00013C, 0x00006C},
+{0x00013D, 0x00013D, 0x00004C},
+{0x00013E, 0x00013E, 0x00006C},
+{0x000143, 0x000143, 0x00004E},
+{0x000144, 0x000144, 0x00006E},
+{0x000145, 0x000145, 0x00004E},
+{0x000146, 0x000146, 0x00006E},
+{0x000147, 0x000147, 0x00004E},
+{0x000148, 0x000148, 0x00006E},
+{0x00014C, 0x00014C, 0x00004F},
+{0x00014D, 0x00014D, 0x00006F},
+{0x00014E, 0x00014E, 0x00004F},
+{0x00014F, 0x00014F, 0x00006F},
+{0x000150, 0x000150, 0x00004F},
+{0x000151, 0x000151, 0x00006F},
+{0x000154, 0x000154, 0x000052},
+{0x000155, 0x000155, 0x000072},
+{0x000156, 0x000156, 0x000052},
+{0x000157, 0x000157, 0x000072},
+{0x000158, 0x000158, 0x000052},
+{0x000159, 0x000159, 0x000072},
+{0x00015A, 0x00015A, 0x000053},
+{0x00015B, 0x00015B, 0x000073},
+{0x00015C, 0x00015C, 0x000053},
+{0x00015D, 0x00015D, 0x000073},
+{0x00015E, 0x00015E, 0x000053},
+{0x00015F, 0x00015F, 0x000073},
+{0x000160, 0x000160, 0x000053},
+{0x000161, 0x000161, 0x000073},
+{0x000162, 0x000162, 0x000054},
+{0x000163, 0x000163, 0x000074},
+{0x000164, 0x000164, 0x000054},
+{0x000165, 0x000165, 0x000074},
+{0x000168, 0x000168, 0x000055},
+{0x000169, 0x000169, 0x000075},
+{0x00016A, 0x00016A, 0x000055},
+{0x00016B, 0x00016B, 0x000075},
+{0x00016C, 0x00016C, 0x000055},
+{0x00016D, 0x00016D, 0x000075},
+{0x00016E, 0x00016E, 0x000055},
+{0x00016F, 0x00016F, 0x000075},
+{0x000170, 0x000170, 0x000055},
+{0x000171, 0x000171, 0x000075},
+{0x000172, 0x000172, 0x000055},
+{0x000173, 0x000173, 0x000075},
+{0x000174, 0x000174, 0x000057},
+{0x000175, 0x000175, 0x000077},
+{0x000176, 0x000176, 0x000059},
+{0x000177, 0x000177, 0x000079},
+{0x000178, 0x000178, 0x000059},
+{0x000179, 0x000179, 0x00005A},
+{0x00017A, 0x00017A, 0x00007A},
+{0x00017B, 0x00017B, 0x00005A},
+{0x00017C, 0x00017C, 0x00007A},
+{0x00017D, 0x00017D, 0x00005A},
+{0x00017E, 0x00017E, 0x00007A},
+{0x0001A0, 0x0001A0, 0x00004F},
+{0x0001A1, 0x0001A1, 0x00006F},
+{0x0001AF, 0x0001AF, 0x000055},
+{0x0001B0, 0x0001B0, 0x000075},
+{0x0001CD, 0x0001CD, 0x000041},
+{0x0001CE, 0x0001CE, 0x000061},
+{0x0001CF, 0x0001CF, 0x000049},
+{0x0001D0, 0x0001D0, 0x000069},
+{0x0001D1, 0x0001D1, 0x00004F},
+{0x0001D2, 0x0001D2, 0x00006F},
+{0x0001D3, 0x0001D3, 0x000055},
+{0x0001D4, 0x0001D4, 0x000075},
+{0x0001D5, 0x0001D5, 0x000055},
+{0x0001D6, 0x0001D6, 0x000075},
+{0x0001D7, 0x0001D7, 0x000055},
+{0x0001D8, 0x0001D8, 0x000075},
+{0x0001D9, 0x0001D9, 0x000055},
+{0x0001DA, 0x0001DA, 0x000075},
+{0x0001DB, 0x0001DB, 0x000055},
+{0x0001DC, 0x0001DC, 0x000075},
+{0x0001DE, 0x0001DE, 0x000041},
+{0x0001DF, 0x0001DF, 0x000061},
+{0x0001E0, 0x0001E0, 0x000041},
+{0x0001E1, 0x0001E1, 0x000061},
+{0x0001E2, 0x0001E2, 0x0000C6},
+{0x0001E3, 0x0001E3, 0x0000E6},
+{0x0001E6, 0x0001E6, 0x000047},
+{0x0001E7, 0x0001E7, 0x000067},
+{0x0001E8, 0x0001E8, 0x00004B},
+{0x0001E9, 0x0001E9, 0x00006B},
+{0x0001EA, 0x0001EA, 0x00004F},
+{0x0001EB, 0x0001EB, 0x00006F},
+{0x0001EC, 0x0001EC, 0x00004F},
+{0x0001ED, 0x0001ED, 0x00006F},
+{0x0001EE, 0x0001EE, 0x0001B7},
+{0x0001EF, 0x0001EF, 0x000292},
+{0x0001F0, 0x0001F0, 0x00006A},
+{0x0001F4, 0x0001F4, 0x000047},
+{0x0001F5, 0x0001F5, 0x000067},
+{0x0001F8, 0x0001F8, 0x00004E},
+{0x0001F9, 0x0001F9, 0x00006E},
+{0x0001FA, 0x0001FA, 0x000041},
+{0x0001FB, 0x0001FB, 0x000061},
+{0x0001FC, 0x0001FC, 0x0000C6},
+{0x0001FD, 0x0001FD, 0x0000E6},
+{0x0001FE, 0x0001FE, 0x0000D8},
+{0x0001FF, 0x0001FF, 0x0000F8},
+{0x000200, 0x000200, 0x000041},
+{0x000201, 0x000201, 0x000061},
+{0x000202, 0x000202, 0x000041},
+{0x000203, 0x000203, 0x000061},
+{0x000204, 0x000204, 0x000045},
+{0x000205, 0x000205, 0x000065},
+{0x000206, 0x000206, 0x000045},
+{0x000207, 0x000207, 0x000065},
+{0x000208, 0x000208, 0x000049},
+{0x000209, 0x000209, 0x000069},
+{0x00020A, 0x00020A, 0x000049},
+{0x00020B, 0x00020B, 0x000069},
+{0x00020C, 0x00020C, 0x00004F},
+{0x00020D, 0x00020D, 0x00006F},
+{0x00020E, 0x00020E, 0x00004F},
+{0x00020F, 0x00020F, 0x00006F},
+{0x000210, 0x000210, 0x000052},
+{0x000211, 0x000211, 0x000072},
+{0x000212, 0x000212, 0x000052},
+{0x000213, 0x000213, 0x000072},
+{0x000214, 0x000214, 0x000055},
+{0x000215, 0x000215, 0x000075},
+{0x000216, 0x000216, 0x000055},
+{0x000217, 0x000217, 0x000075},
+{0x000218, 0x000218, 0x000053},
+{0x000219, 0x000219, 0x000073},
+{0x00021A, 0x00021A, 0x000054},
+{0x00021B, 0x00021B, 0x000074},
+{0x00021E, 0x00021E, 0x000048},
+{0x00021F, 0x00021F, 0x000068},
+{0x000226, 0x000226, 0x000041},
+{0x000227, 0x000227, 0x000061},
+{0x000228, 0x000228, 0x000045},
+{0x000229, 0x000229, 0x000065},
+{0x00022A, 0x00022A, 0x00004F},
+{0x00022B, 0x00022B, 0x00006F},
+{0x00022C, 0x00022C, 0x00004F},
+{0x00022D, 0x00022D, 0x00006F},
+{0x00022E, 0x00022E, 0x00004F},
+{0x00022F, 0x00022F, 0x00006F},
+{0x000230, 0x000230, 0x00004F},
+{0x000231, 0x000231, 0x00006F},
+{0x000232, 0x000232, 0x000059},
+{0x000233, 0x000233, 0x000079},
+{0x000340, 0x000340, 0x000300},
+{0x000341, 0x000341, 0x000301},
+{0x000343, 0x000343, 0x000313},
+{0x000344, 0x000344, 0x000308},
+{0x000374, 0x000374, 0x0002B9},
+{0x00037E, 0x00037E, 0x00003B},
+{0x000385, 0x000385, 0x0000A8},
+{0x000386, 0x000386, 0x000391},
+{0x000387, 0x000387, 0x0000B7},
+{0x000388, 0x000388, 0x000395},
+{0x000389, 0x000389, 0x000397},
+{0x00038A, 0x00038A, 0x000399},
+{0x00038C, 0x00038C, 0x00039F},
+{0x00038E, 0x00038E, 0x0003A5},
+{0x00038F, 0x00038F, 0x0003A9},
+{0x000390, 0x000390, 0x0003B9},
+{0x0003AA, 0x0003AA, 0x000399},
+{0x0003AB, 0x0003AB, 0x0003A5},
+{0x0003AC, 0x0003AC, 0x0003B1},
+{0x0003AD, 0x0003AD, 0x0003B5},
+{0x0003AE, 0x0003AE, 0x0003B7},
+{0x0003AF, 0x0003AF, 0x0003B9},
+{0x0003B0, 0x0003B0, 0x0003C5},
+{0x0003CA, 0x0003CA, 0x0003B9},
+{0x0003CB, 0x0003CB, 0x0003C5},
+{0x0003CC, 0x0003CC, 0x0003BF},
+{0x0003CD, 0x0003CD, 0x0003C5},
+{0x0003CE, 0x0003CE, 0x0003C9},
+{0x0003D3, 0x0003D4, 0x0003D2},
+{0x000400, 0x000401, 0x000415},
+{0x000403, 0x000403, 0x000413},
+{0x000407, 0x000407, 0x000406},
+{0x00040C, 0x00040C, 0x00041A},
+{0x00040D, 0x00040D, 0x000418},
+{0x00040E, 0x00040E, 0x000423},
+{0x000419, 0x000419, 0x000418},
+{0x000439, 0x000439, 0x000438},
+{0x000450, 0x000451, 0x000435},
+{0x000453, 0x000453, 0x000433},
+{0x000457, 0x000457, 0x000456},
+{0x00045C, 0x00045C, 0x00043A},
+{0x00045D, 0x00045D, 0x000438},
+{0x00045E, 0x00045E, 0x000443},
+{0x000476, 0x000476, 0x000474},
+{0x000477, 0x000477, 0x000475},
+{0x0004C1, 0x0004C1, 0x000416},
+{0x0004C2, 0x0004C2, 0x000436},
+{0x0004D0, 0x0004D0, 0x000410},
+{0x0004D1, 0x0004D1, 0x000430},
+{0x0004D2, 0x0004D2, 0x000410},
+{0x0004D3, 0x0004D3, 0x000430},
+{0x0004D6, 0x0004D6, 0x000415},
+{0x0004D7, 0x0004D7, 0x000435},
+{0x0004DA, 0x0004DA, 0x0004D8},
+{0x0004DB, 0x0004DB, 0x0004D9},
+{0x0004DC, 0x0004DC, 0x000416},
+{0x0004DD, 0x0004DD, 0x000436},
+{0x0004DE, 0x0004DE, 0x000417},
+{0x0004DF, 0x0004DF, 0x000437},
+{0x0004E2, 0x0004E2, 0x000418},
+{0x0004E3, 0x0004E3, 0x000438},
+{0x0004E4, 0x0004E4, 0x000418},
+{0x0004E5, 0x0004E5, 0x000438},
+{0x0004E6, 0x0004E6, 0x00041E},
+{0x0004E7, 0x0004E7, 0x00043E},
+{0x0004EA, 0x0004EA, 0x0004E8},
+{0x0004EB, 0x0004EB, 0x0004E9},
+{0x0004EC, 0x0004EC, 0x00042D},
+{0x0004ED, 0x0004ED, 0x00044D},
+{0x0004EE, 0x0004EE, 0x000423},
+{0x0004EF, 0x0004EF, 0x000443},
+{0x0004F0, 0x0004F0, 0x000423},
+{0x0004F1, 0x0004F1, 0x000443},
+{0x0004F2, 0x0004F2, 0x000423},
+{0x0004F3, 0x0004F3, 0x000443},
+{0x0004F4, 0x0004F4, 0x000427},
+{0x0004F5, 0x0004F5, 0x000447},
+{0x0004F8, 0x0004F8, 0x00042B},
+{0x0004F9, 0x0004F9, 0x00044B},
+{0x000622, 0x000623, 0x000627},
+{0x000624, 0x000624, 0x000648},
+{0x000625, 0x000625, 0x000627},
+{0x000626, 0x000626, 0x00064A},
+{0x0006C0, 0x0006C0, 0x0006D5},
+{0x0006C2, 0x0006C2, 0x0006C1},
+{0x0006D3, 0x0006D3, 0x0006D2},
+{0x000929, 0x000929, 0x000928},
+{0x000931, 0x000931, 0x000930},
+{0x000934, 0x000934, 0x000933},
+{0x000958, 0x000958, 0x000915},
+{0x000959, 0x000959, 0x000916},
+{0x00095A, 0x00095A, 0x000917},
+{0x00095B, 0x00095B, 0x00091C},
+{0x00095C, 0x00095C, 0x000921},
+{0x00095D, 0x00095D, 0x000922},
+{0x00095E, 0x00095E, 0x00092B},
+{0x00095F, 0x00095F, 0x00092F},
+{0x0009CB, 0x0009CC, 0x0009C7},
+{0x0009DC, 0x0009DC, 0x0009A1},
+{0x0009DD, 0x0009DD, 0x0009A2},
+{0x0009DF, 0x0009DF, 0x0009AF},
+{0x000A33, 0x000A33, 0x000A32},
+{0x000A36, 0x000A36, 0x000A38},
+{0x000A59, 0x000A59, 0x000A16},
+{0x000A5A, 0x000A5A, 0x000A17},
+{0x000A5B, 0x000A5B, 0x000A1C},
+{0x000A5E, 0x000A5E, 0x000A2B},
+{0x000B48, 0x000B48, 0x000B47},
+{0x000B4B, 0x000B4C, 0x000B47},
+{0x000B5C, 0x000B5C, 0x000B21},
+{0x000B5D, 0x000B5D, 0x000B22},
+{0x000B94, 0x000B94, 0x000B92},
+{0x000BCA, 0x000BCA, 0x000BC6},
+{0x000BCB, 0x000BCB, 0x000BC7},
+{0x000BCC, 0x000BCC, 0x000BC6},
+{0x000C48, 0x000C48, 0x000C46},
+{0x000CC0, 0x000CC0, 0x000CBF},
+{0x000CC7, 0x000CC8, 0x000CC6},
+{0x000CCA, 0x000CCB, 0x000CC6},
+{0x000D4A, 0x000D4A, 0x000D46},
+{0x000D4B, 0x000D4B, 0x000D47},
+{0x000D4C, 0x000D4C, 0x000D46},
+{0x000DDA, 0x000DDA, 0x000DD9},
+{0x000DDC, 0x000DDE, 0x000DD9},
+{0x000F43, 0x000F43, 0x000F42},
+{0x000F4D, 0x000F4D, 0x000F4C},
+{0x000F52, 0x000F52, 0x000F51},
+{0x000F57, 0x000F57, 0x000F56},
+{0x000F5C, 0x000F5C, 0x000F5B},
+{0x000F69, 0x000F69, 0x000F40},
+{0x000F73, 0x000F73, 0x000F71},
+{0x000F75, 0x000F75, 0x000F71},
+{0x000F76, 0x000F76, 0x000FB2},
+{0x000F78, 0x000F78, 0x000FB3},
+{0x000F81, 0x000F81, 0x000F71},
+{0x000F93, 0x000F93, 0x000F92},
+{0x000F9D, 0x000F9D, 0x000F9C},
+{0x000FA2, 0x000FA2, 0x000FA1},
+{0x000FA7, 0x000FA7, 0x000FA6},
+{0x000FAC, 0x000FAC, 0x000FAB},
+{0x000FB9, 0x000FB9, 0x000F90},
+{0x001026, 0x001026, 0x001025},
+{0x001B06, 0x001B06, 0x001B05},
+{0x001B08, 0x001B08, 0x001B07},
+{0x001B0A, 0x001B0A, 0x001B09},
+{0x001B0C, 0x001B0C, 0x001B0B},
+{0x001B0E, 0x001B0E, 0x001B0D},
+{0x001B12, 0x001B12, 0x001B11},
+{0x001B3B, 0x001B3B, 0x001B3A},
+{0x001B3D, 0x001B3D, 0x001B3C},
+{0x001B40, 0x001B40, 0x001B3E},
+{0x001B41, 0x001B41, 0x001B3F},
+{0x001B43, 0x001B43, 0x001B42},
+{0x001E00, 0x001E00, 0x000041},
+{0x001E01, 0x001E01, 0x000061},
+{0x001E02, 0x001E02, 0x000042},
+{0x001E03, 0x001E03, 0x000062},
+{0x001E04, 0x001E04, 0x000042},
+{0x001E05, 0x001E05, 0x000062},
+{0x001E06, 0x001E06, 0x000042},
+{0x001E07, 0x001E07, 0x000062},
+{0x001E08, 0x001E08, 0x000043},
+{0x001E09, 0x001E09, 0x000063},
+{0x001E0A, 0x001E0A, 0x000044},
+{0x001E0B, 0x001E0B, 0x000064},
+{0x001E0C, 0x001E0C, 0x000044},
+{0x001E0D, 0x001E0D, 0x000064},
+{0x001E0E, 0x001E0E, 0x000044},
+{0x001E0F, 0x001E0F, 0x000064},
+{0x001E10, 0x001E10, 0x000044},
+{0x001E11, 0x001E11, 0x000064},
+{0x001E12, 0x001E12, 0x000044},
+{0x001E13, 0x001E13, 0x000064},
+{0x001E14, 0x001E14, 0x000045},
+{0x001E15, 0x001E15, 0x000065},
+{0x001E16, 0x001E16, 0x000045},
+{0x001E17, 0x001E17, 0x000065},
+{0x001E18, 0x001E18, 0x000045},
+{0x001E19, 0x001E19, 0x000065},
+{0x001E1A, 0x001E1A, 0x000045},
+{0x001E1B, 0x001E1B, 0x000065},
+{0x001E1C, 0x001E1C, 0x000045},
+{0x001E1D, 0x001E1D, 0x000065},
+{0x001E1E, 0x001E1E, 0x000046},
+{0x001E1F, 0x001E1F, 0x000066},
+{0x001E20, 0x001E20, 0x000047},
+{0x001E21, 0x001E21, 0x000067},
+{0x001E22, 0x001E22, 0x000048},
+{0x001E23, 0x001E23, 0x000068},
+{0x001E24, 0x001E24, 0x000048},
+{0x001E25, 0x001E25, 0x000068},
+{0x001E26, 0x001E26, 0x000048},
+{0x001E27, 0x001E27, 0x000068},
+{0x001E28, 0x001E28, 0x000048},
+{0x001E29, 0x001E29, 0x000068},
+{0x001E2A, 0x001E2A, 0x000048},
+{0x001E2B, 0x001E2B, 0x000068},
+{0x001E2C, 0x001E2C, 0x000049},
+{0x001E2D, 0x001E2D, 0x000069},
+{0x001E2E, 0x001E2E, 0x000049},
+{0x001E2F, 0x001E2F, 0x000069},
+{0x001E30, 0x001E30, 0x00004B},
+{0x001E31, 0x001E31, 0x00006B},
+{0x001E32, 0x001E32, 0x00004B},
+{0x001E33, 0x001E33, 0x00006B},
+{0x001E34, 0x001E34, 0x00004B},
+{0x001E35, 0x001E35, 0x00006B},
+{0x001E36, 0x001E36, 0x00004C},
+{0x001E37, 0x001E37, 0x00006C},
+{0x001E38, 0x001E38, 0x00004C},
+{0x001E39, 0x001E39, 0x00006C},
+{0x001E3A, 0x001E3A, 0x00004C},
+{0x001E3B, 0x001E3B, 0x00006C},
+{0x001E3C, 0x001E3C, 0x00004C},
+{0x001E3D, 0x001E3D, 0x00006C},
+{0x001E3E, 0x001E3E, 0x00004D},
+{0x001E3F, 0x001E3F, 0x00006D},
+{0x001E40, 0x001E40, 0x00004D},
+{0x001E41, 0x001E41, 0x00006D},
+{0x001E42, 0x001E42, 0x00004D},
+{0x001E43, 0x001E43, 0x00006D},
+{0x001E44, 0x001E44, 0x00004E},
+{0x001E45, 0x001E45, 0x00006E},
+{0x001E46, 0x001E46, 0x00004E},
+{0x001E47, 0x001E47, 0x00006E},
+{0x001E48, 0x001E48, 0x00004E},
+{0x001E49, 0x001E49, 0x00006E},
+{0x001E4A, 0x001E4A, 0x00004E},
+{0x001E4B, 0x001E4B, 0x00006E},
+{0x001E4C, 0x001E4C, 0x00004F},
+{0x001E4D, 0x001E4D, 0x00006F},
+{0x001E4E, 0x001E4E, 0x00004F},
+{0x001E4F, 0x001E4F, 0x00006F},
+{0x001E50, 0x001E50, 0x00004F},
+{0x001E51, 0x001E51, 0x00006F},
+{0x001E52, 0x001E52, 0x00004F},
+{0x001E53, 0x001E53, 0x00006F},
+{0x001E54, 0x001E54, 0x000050},
+{0x001E55, 0x001E55, 0x000070},
+{0x001E56, 0x001E56, 0x000050},
+{0x001E57, 0x001E57, 0x000070},
+{0x001E58, 0x001E58, 0x000052},
+{0x001E59, 0x001E59, 0x000072},
+{0x001E5A, 0x001E5A, 0x000052},
+{0x001E5B, 0x001E5B, 0x000072},
+{0x001E5C, 0x001E5C, 0x000052},
+{0x001E5D, 0x001E5D, 0x000072},
+{0x001E5E, 0x001E5E, 0x000052},
+{0x001E5F, 0x001E5F, 0x000072},
+{0x001E60, 0x001E60, 0x000053},
+{0x001E61, 0x001E61, 0x000073},
+{0x001E62, 0x001E62, 0x000053},
+{0x001E63, 0x001E63, 0x000073},
+{0x001E64, 0x001E64, 0x000053},
+{0x001E65, 0x001E65, 0x000073},
+{0x001E66, 0x001E66, 0x000053},
+{0x001E67, 0x001E67, 0x000073},
+{0x001E68, 0x001E68, 0x000053},
+{0x001E69, 0x001E69, 0x000073},
+{0x001E6A, 0x001E6A, 0x000054},
+{0x001E6B, 0x001E6B, 0x000074},
+{0x001E6C, 0x001E6C, 0x000054},
+{0x001E6D, 0x001E6D, 0x000074},
+{0x001E6E, 0x001E6E, 0x000054},
+{0x001E6F, 0x001E6F, 0x000074},
+{0x001E70, 0x001E70, 0x000054},
+{0x001E71, 0x001E71, 0x000074},
+{0x001E72, 0x001E72, 0x000055},
+{0x001E73, 0x001E73, 0x000075},
+{0x001E74, 0x001E74, 0x000055},
+{0x001E75, 0x001E75, 0x000075},
+{0x001E76, 0x001E76, 0x000055},
+{0x001E77, 0x001E77, 0x000075},
+{0x001E78, 0x001E78, 0x000055},
+{0x001E79, 0x001E79, 0x000075},
+{0x001E7A, 0x001E7A, 0x000055},
+{0x001E7B, 0x001E7B, 0x000075},
+{0x001E7C, 0x001E7C, 0x000056},
+{0x001E7D, 0x001E7D, 0x000076},
+{0x001E7E, 0x001E7E, 0x000056},
+{0x001E7F, 0x001E7F, 0x000076},
+{0x001E80, 0x001E80, 0x000057},
+{0x001E81, 0x001E81, 0x000077},
+{0x001E82, 0x001E82, 0x000057},
+{0x001E83, 0x001E83, 0x000077},
+{0x001E84, 0x001E84, 0x000057},
+{0x001E85, 0x001E85, 0x000077},
+{0x001E86, 0x001E86, 0x000057},
+{0x001E87, 0x001E87, 0x000077},
+{0x001E88, 0x001E88, 0x000057},
+{0x001E89, 0x001E89, 0x000077},
+{0x001E8A, 0x001E8A, 0x000058},
+{0x001E8B, 0x001E8B, 0x000078},
+{0x001E8C, 0x001E8C, 0x000058},
+{0x001E8D, 0x001E8D, 0x000078},
+{0x001E8E, 0x001E8E, 0x000059},
+{0x001E8F, 0x001E8F, 0x000079},
+{0x001E90, 0x001E90, 0x00005A},
+{0x001E91, 0x001E91, 0x00007A},
+{0x001E92, 0x001E92, 0x00005A},
+{0x001E93, 0x001E93, 0x00007A},
+{0x001E94, 0x001E94, 0x00005A},
+{0x001E95, 0x001E95, 0x00007A},
+{0x001E96, 0x001E96, 0x000068},
+{0x001E97, 0x001E97, 0x000074},
+{0x001E98, 0x001E98, 0x000077},
+{0x001E99, 0x001E99, 0x000079},
+{0x001E9B, 0x001E9B, 0x00017F},
+{0x001EA0, 0x001EA0, 0x000041},
+{0x001EA1, 0x001EA1, 0x000061},
+{0x001EA2, 0x001EA2, 0x000041},
+{0x001EA3, 0x001EA3, 0x000061},
+{0x001EA4, 0x001EA4, 0x000041},
+{0x001EA5, 0x001EA5, 0x000061},
+{0x001EA6, 0x001EA6, 0x000041},
+{0x001EA7, 0x001EA7, 0x000061},
+{0x001EA8, 0x001EA8, 0x000041},
+{0x001EA9, 0x001EA9, 0x000061},
+{0x001EAA, 0x001EAA, 0x000041},
+{0x001EAB, 0x001EAB, 0x000061},
+{0x001EAC, 0x001EAC, 0x000041},
+{0x001EAD, 0x001EAD, 0x000061},
+{0x001EAE, 0x001EAE, 0x000041},
+{0x001EAF, 0x001EAF, 0x000061},
+{0x001EB0, 0x001EB0, 0x000041},
+{0x001EB1, 0x001EB1, 0x000061},
+{0x001EB2, 0x001EB2, 0x000041},
+{0x001EB3, 0x001EB3, 0x000061},
+{0x001EB4, 0x001EB4, 0x000041},
+{0x001EB5, 0x001EB5, 0x000061},
+{0x001EB6, 0x001EB6, 0x000041},
+{0x001EB7, 0x001EB7, 0x000061},
+{0x001EB8, 0x001EB8, 0x000045},
+{0x001EB9, 0x001EB9, 0x000065},
+{0x001EBA, 0x001EBA, 0x000045},
+{0x001EBB, 0x001EBB, 0x000065},
+{0x001EBC, 0x001EBC, 0x000045},
+{0x001EBD, 0x001EBD, 0x000065},
+{0x001EBE, 0x001EBE, 0x000045},
+{0x001EBF, 0x001EBF, 0x000065},
+{0x001EC0, 0x001EC0, 0x000045},
+{0x001EC1, 0x001EC1, 0x000065},
+{0x001EC2, 0x001EC2, 0x000045},
+{0x001EC3, 0x001EC3, 0x000065},
+{0x001EC4, 0x001EC4, 0x000045},
+{0x001EC5, 0x001EC5, 0x000065},
+{0x001EC6, 0x001EC6, 0x000045},
+{0x001EC7, 0x001EC7, 0x000065},
+{0x001EC8, 0x001EC8, 0x000049},
+{0x001EC9, 0x001EC9, 0x000069},
+{0x001ECA, 0x001ECA, 0x000049},
+{0x001ECB, 0x001ECB, 0x000069},
+{0x001ECC, 0x001ECC, 0x00004F},
+{0x001ECD, 0x001ECD, 0x00006F},
+{0x001ECE, 0x001ECE, 0x00004F},
+{0x001ECF, 0x001ECF, 0x00006F},
+{0x001ED0, 0x001ED0, 0x00004F},
+{0x001ED1, 0x001ED1, 0x00006F},
+{0x001ED2, 0x001ED2, 0x00004F},
+{0x001ED3, 0x001ED3, 0x00006F},
+{0x001ED4, 0x001ED4, 0x00004F},
+{0x001ED5, 0x001ED5, 0x00006F},
+{0x001ED6, 0x001ED6, 0x00004F},
+{0x001ED7, 0x001ED7, 0x00006F},
+{0x001ED8, 0x001ED8, 0x00004F},
+{0x001ED9, 0x001ED9, 0x00006F},
+{0x001EDA, 0x001EDA, 0x00004F},
+{0x001EDB, 0x001EDB, 0x00006F},
+{0x001EDC, 0x001EDC, 0x00004F},
+{0x001EDD, 0x001EDD, 0x00006F},
+{0x001EDE, 0x001EDE, 0x00004F},
+{0x001EDF, 0x001EDF, 0x00006F},
+{0x001EE0, 0x001EE0, 0x00004F},
+{0x001EE1, 0x001EE1, 0x00006F},
+{0x001EE2, 0x001EE2, 0x00004F},
+{0x001EE3, 0x001EE3, 0x00006F},
+{0x001EE4, 0x001EE4, 0x000055},
+{0x001EE5, 0x001EE5, 0x000075},
+{0x001EE6, 0x001EE6, 0x000055},
+{0x001EE7, 0x001EE7, 0x000075},
+{0x001EE8, 0x001EE8, 0x000055},
+{0x001EE9, 0x001EE9, 0x000075},
+{0x001EEA, 0x001EEA, 0x000055},
+{0x001EEB, 0x001EEB, 0x000075},
+{0x001EEC, 0x001EEC, 0x000055},
+{0x001EED, 0x001EED, 0x000075},
+{0x001EEE, 0x001EEE, 0x000055},
+{0x001EEF, 0x001EEF, 0x000075},
+{0x001EF0, 0x001EF0, 0x000055},
+{0x001EF1, 0x001EF1, 0x000075},
+{0x001EF2, 0x001EF2, 0x000059},
+{0x001EF3, 0x001EF3, 0x000079},
+{0x001EF4, 0x001EF4, 0x000059},
+{0x001EF5, 0x001EF5, 0x000079},
+{0x001EF6, 0x001EF6, 0x000059},
+{0x001EF7, 0x001EF7, 0x000079},
+{0x001EF8, 0x001EF8, 0x000059},
+{0x001EF9, 0x001EF9, 0x000079},
+{0x001F00, 0x001F07, 0x0003B1},
+{0x001F08, 0x001F0F, 0x000391},
+{0x001F10, 0x001F15, 0x0003B5},
+{0x001F18, 0x001F1D, 0x000395},
+{0x001F20, 0x001F27, 0x0003B7},
+{0x001F28, 0x001F2F, 0x000397},
+{0x001F30, 0x001F37, 0x0003B9},
+{0x001F38, 0x001F3F, 0x000399},
+{0x001F40, 0x001F45, 0x0003BF},
+{0x001F48, 0x001F4D, 0x00039F},
+{0x001F50, 0x001F57, 0x0003C5},
+{0x001F59, 0x001F59, 0x0003A5},
+{0x001F5B, 0x001F5B, 0x0003A5},
+{0x001F5D, 0x001F5D, 0x0003A5},
+{0x001F5F, 0x001F5F, 0x0003A5},
+{0x001F60, 0x001F67, 0x0003C9},
+{0x001F68, 0x001F6F, 0x0003A9},
+{0x001F70, 0x001F71, 0x0003B1},
+{0x001F72, 0x001F73, 0x0003B5},
+{0x001F74, 0x001F75, 0x0003B7},
+{0x001F76, 0x001F77, 0x0003B9},
+{0x001F78, 0x001F79, 0x0003BF},
+{0x001F7A, 0x001F7B, 0x0003C5},
+{0x001F7C, 0x001F7D, 0x0003C9},
+{0x001F80, 0x001F87, 0x0003B1},
+{0x001F88, 0x001F8F, 0x000391},
+{0x001F90, 0x001F97, 0x0003B7},
+{0x001F98, 0x001F9F, 0x000397},
+{0x001FA0, 0x001FA7, 0x0003C9},
+{0x001FA8, 0x001FAF, 0x0003A9},
+{0x001FB0, 0x001FB4, 0x0003B1},
+{0x001FB6, 0x001FB7, 0x0003B1},
+{0x001FB8, 0x001FBC, 0x000391},
+{0x001FBE, 0x001FBE, 0x0003B9},
+{0x001FC1, 0x001FC1, 0x0000A8},
+{0x001FC2, 0x001FC4, 0x0003B7},
+{0x001FC6, 0x001FC7, 0x0003B7},
+{0x001FC8, 0x001FC9, 0x000395},
+{0x001FCA, 0x001FCC, 0x000397},
+{0x001FCD, 0x001FCF, 0x001FBF},
+{0x001FD0, 0x001FD3, 0x0003B9},
+{0x001FD6, 0x001FD7, 0x0003B9},
+{0x001FD8, 0x001FDB, 0x000399},
+{0x001FDD, 0x001FDF, 0x001FFE},
+{0x001FE0, 0x001FE3, 0x0003C5},
+{0x001FE4, 0x001FE5, 0x0003C1},
+{0x001FE6, 0x001FE7, 0x0003C5},
+{0x001FE8, 0x001FEB, 0x0003A5},
+{0x001FEC, 0x001FEC, 0x0003A1},
+{0x001FED, 0x001FEE, 0x0000A8},
+{0x001FEF, 0x001FEF, 0x000060},
+{0x001FF2, 0x001FF4, 0x0003C9},
+{0x001FF6, 0x001FF7, 0x0003C9},
+{0x001FF8, 0x001FF9, 0x00039F},
+{0x001FFA, 0x001FFC, 0x0003A9},
+{0x001FFD, 0x001FFD, 0x0000B4},
+{0x002000, 0x002000, 0x002002},
+{0x002001, 0x002001, 0x002003},
+{0x002126, 0x002126, 0x0003A9},
+{0x00212A, 0x00212A, 0x00004B},
+{0x00212B, 0x00212B, 0x000041},
+{0x00219A, 0x00219A, 0x002190},
+{0x00219B, 0x00219B, 0x002192},
+{0x0021AE, 0x0021AE, 0x002194},
+{0x0021CD, 0x0021CD, 0x0021D0},
+{0x0021CE, 0x0021CE, 0x0021D4},
+{0x0021CF, 0x0021CF, 0x0021D2},
+{0x002204, 0x002204, 0x002203},
+{0x002209, 0x002209, 0x002208},
+{0x00220C, 0x00220C, 0x00220B},
+{0x002224, 0x002224, 0x002223},
+{0x002226, 0x002226, 0x002225},
+{0x002241, 0x002241, 0x00223C},
+{0x002244, 0x002244, 0x002243},
+{0x002247, 0x002247, 0x002245},
+{0x002249, 0x002249, 0x002248},
+{0x002260, 0x002260, 0x00003D},
+{0x002262, 0x002262, 0x002261},
+{0x00226D, 0x00226D, 0x00224D},
+{0x00226E, 0x00226E, 0x00003C},
+{0x00226F, 0x00226F, 0x00003E},
+{0x002270, 0x002270, 0x002264},
+{0x002271, 0x002271, 0x002265},
+{0x002274, 0x002274, 0x002272},
+{0x002275, 0x002275, 0x002273},
+{0x002278, 0x002278, 0x002276},
+{0x002279, 0x002279, 0x002277},
+{0x002280, 0x002280, 0x00227A},
+{0x002281, 0x002281, 0x00227B},
+{0x002284, 0x002284, 0x002282},
+{0x002285, 0x002285, 0x002283},
+{0x002288, 0x002288, 0x002286},
+{0x002289, 0x002289, 0x002287},
+{0x0022AC, 0x0022AC, 0x0022A2},
+{0x0022AD, 0x0022AD, 0x0022A8},
+{0x0022AE, 0x0022AE, 0x0022A9},
+{0x0022AF, 0x0022AF, 0x0022AB},
+{0x0022E0, 0x0022E0, 0x00227C},
+{0x0022E1, 0x0022E1, 0x00227D},
+{0x0022E2, 0x0022E2, 0x002291},
+{0x0022E3, 0x0022E3, 0x002292},
+{0x0022EA, 0x0022EA, 0x0022B2},
+{0x0022EB, 0x0022EB, 0x0022B3},
+{0x0022EC, 0x0022EC, 0x0022B4},
+{0x0022ED, 0x0022ED, 0x0022B5},
+{0x002329, 0x002329, 0x003008},
+{0x00232A, 0x00232A, 0x003009},
+{0x002ADC, 0x002ADC, 0x002ADD},
+{0x00304C, 0x00304C, 0x00304B},
+{0x00304E, 0x00304E, 0x00304D},
+{0x003050, 0x003050, 0x00304F},
+{0x003052, 0x003052, 0x003051},
+{0x003054, 0x003054, 0x003053},
+{0x003056, 0x003056, 0x003055},
+{0x003058, 0x003058, 0x003057},
+{0x00305A, 0x00305A, 0x003059},
+{0x00305C, 0x00305C, 0x00305B},
+{0x00305E, 0x00305E, 0x00305D},
+{0x003060, 0x003060, 0x00305F},
+{0x003062, 0x003062, 0x003061},
+{0x003065, 0x003065, 0x003064},
+{0x003067, 0x003067, 0x003066},
+{0x003069, 0x003069, 0x003068},
+{0x003070, 0x003071, 0x00306F},
+{0x003073, 0x003074, 0x003072},
+{0x003076, 0x003077, 0x003075},
+{0x003079, 0x00307A, 0x003078},
+{0x00307C, 0x00307D, 0x00307B},
+{0x003094, 0x003094, 0x003046},
+{0x00309E, 0x00309E, 0x00309D},
+{0x0030AC, 0x0030AC, 0x0030AB},
+{0x0030AE, 0x0030AE, 0x0030AD},
+{0x0030B0, 0x0030B0, 0x0030AF},
+{0x0030B2, 0x0030B2, 0x0030B1},
+{0x0030B4, 0x0030B4, 0x0030B3},
+{0x0030B6, 0x0030B6, 0x0030B5},
+{0x0030B8, 0x0030B8, 0x0030B7},
+{0x0030BA, 0x0030BA, 0x0030B9},
+{0x0030BC, 0x0030BC, 0x0030BB},
+{0x0030BE, 0x0030BE, 0x0030BD},
+{0x0030C0, 0x0030C0, 0x0030BF},
+{0x0030C2, 0x0030C2, 0x0030C1},
+{0x0030C5, 0x0030C5, 0x0030C4},
+{0x0030C7, 0x0030C7, 0x0030C6},
+{0x0030C9, 0x0030C9, 0x0030C8},
+{0x0030D0, 0x0030D1, 0x0030CF},
+{0x0030D3, 0x0030D4, 0x0030D2},
+{0x0030D6, 0x0030D7, 0x0030D5},
+{0x0030D9, 0x0030DA, 0x0030D8},
+{0x0030DC, 0x0030DD, 0x0030DB},
+{0x0030F4, 0x0030F4, 0x0030A6},
+{0x0030F7, 0x0030F7, 0x0030EF},
+{0x0030F8, 0x0030F8, 0x0030F0},
+{0x0030F9, 0x0030F9, 0x0030F1},
+{0x0030FA, 0x0030FA, 0x0030F2},
+{0x0030FE, 0x0030FE, 0x0030FD},
+{0x00AC00, 0x00AE4B, 0x001100},
+{0x00AE4C, 0x00B097, 0x001101},
+{0x00B098, 0x00B2E3, 0x001102},
+{0x00B2E4, 0x00B52F, 0x001103},
+{0x00B530, 0x00B77B, 0x001104},
+{0x00B77C, 0x00B9C7, 0x001105},
+{0x00B9C8, 0x00BC13, 0x001106},
+{0x00BC14, 0x00BE5F, 0x001107},
+{0x00BE60, 0x00C0AB, 0x001108},
+{0x00C0AC, 0x00C2F7, 0x001109},
+{0x00C2F8, 0x00C543, 0x00110A},
+{0x00C544, 0x00C78F, 0x00110B},
+{0x00C790, 0x00C9DB, 0x00110C},
+{0x00C9DC, 0x00CC27, 0x00110D},
+{0x00CC28, 0x00CE73, 0x00110E},
+{0x00CE74, 0x00D0BF, 0x00110F},
+{0x00D0C0, 0x00D30B, 0x001110},
+{0x00D30C, 0x00D557, 0x001111},
+{0x00D558, 0x00D7A3, 0x001112},
+{0x00F900, 0x00F900, 0x008C48},
+{0x00F901, 0x00F901, 0x0066F4},
+{0x00F902, 0x00F902, 0x008ECA},
+{0x00F903, 0x00F903, 0x008CC8},
+{0x00F904, 0x00F904, 0x006ED1},
+{0x00F905, 0x00F905, 0x004E32},
+{0x00F906, 0x00F906, 0x0053E5},
+{0x00F907, 0x00F908, 0x009F9C},
+{0x00F909, 0x00F909, 0x005951},
+{0x00F90A, 0x00F90A, 0x0091D1},
+{0x00F90B, 0x00F90B, 0x005587},
+{0x00F90C, 0x00F90C, 0x005948},
+{0x00F90D, 0x00F90D, 0x0061F6},
+{0x00F90E, 0x00F90E, 0x007669},
+{0x00F90F, 0x00F90F, 0x007F85},
+{0x00F910, 0x00F910, 0x00863F},
+{0x00F911, 0x00F911, 0x0087BA},
+{0x00F912, 0x00F912, 0x0088F8},
+{0x00F913, 0x00F913, 0x00908F},
+{0x00F914, 0x00F914, 0x006A02},
+{0x00F915, 0x00F915, 0x006D1B},
+{0x00F916, 0x00F916, 0x0070D9},
+{0x00F917, 0x00F917, 0x0073DE},
+{0x00F918, 0x00F918, 0x00843D},
+{0x00F919, 0x00F919, 0x00916A},
+{0x00F91A, 0x00F91A, 0x0099F1},
+{0x00F91B, 0x00F91B, 0x004E82},
+{0x00F91C, 0x00F91C, 0x005375},
+{0x00F91D, 0x00F91D, 0x006B04},
+{0x00F91E, 0x00F91E, 0x00721B},
+{0x00F91F, 0x00F91F, 0x00862D},
+{0x00F920, 0x00F920, 0x009E1E},
+{0x00F921, 0x00F921, 0x005D50},
+{0x00F922, 0x00F922, 0x006FEB},
+{0x00F923, 0x00F923, 0x0085CD},
+{0x00F924, 0x00F924, 0x008964},
+{0x00F925, 0x00F925, 0x0062C9},
+{0x00F926, 0x00F926, 0x0081D8},
+{0x00F927, 0x00F927, 0x00881F},
+{0x00F928, 0x00F928, 0x005ECA},
+{0x00F929, 0x00F929, 0x006717},
+{0x00F92A, 0x00F92A, 0x006D6A},
+{0x00F92B, 0x00F92B, 0x0072FC},
+{0x00F92C, 0x00F92C, 0x0090CE},
+{0x00F92D, 0x00F92D, 0x004F86},
+{0x00F92E, 0x00F92E, 0x0051B7},
+{0x00F92F, 0x00F92F, 0x0052DE},
+{0x00F930, 0x00F930, 0x0064C4},
+{0x00F931, 0x00F931, 0x006AD3},
+{0x00F932, 0x00F932, 0x007210},
+{0x00F933, 0x00F933, 0x0076E7},
+{0x00F934, 0x00F934, 0x008001},
+{0x00F935, 0x00F935, 0x008606},
+{0x00F936, 0x00F936, 0x00865C},
+{0x00F937, 0x00F937, 0x008DEF},
+{0x00F938, 0x00F938, 0x009732},
+{0x00F939, 0x00F939, 0x009B6F},
+{0x00F93A, 0x00F93A, 0x009DFA},
+{0x00F93B, 0x00F93B, 0x00788C},
+{0x00F93C, 0x00F93C, 0x00797F},
+{0x00F93D, 0x00F93D, 0x007DA0},
+{0x00F93E, 0x00F93E, 0x0083C9},
+{0x00F93F, 0x00F93F, 0x009304},
+{0x00F940, 0x00F940, 0x009E7F},
+{0x00F941, 0x00F941, 0x008AD6},
+{0x00F942, 0x00F942, 0x0058DF},
+{0x00F943, 0x00F943, 0x005F04},
+{0x00F944, 0x00F944, 0x007C60},
+{0x00F945, 0x00F945, 0x00807E},
+{0x00F946, 0x00F946, 0x007262},
+{0x00F947, 0x00F947, 0x0078CA},
+{0x00F948, 0x00F948, 0x008CC2},
+{0x00F949, 0x00F949, 0x0096F7},
+{0x00F94A, 0x00F94A, 0x0058D8},
+{0x00F94B, 0x00F94B, 0x005C62},
+{0x00F94C, 0x00F94C, 0x006A13},
+{0x00F94D, 0x00F94D, 0x006DDA},
+{0x00F94E, 0x00F94E, 0x006F0F},
+{0x00F94F, 0x00F94F, 0x007D2F},
+{0x00F950, 0x00F950, 0x007E37},
+{0x00F951, 0x00F951, 0x00964B},
+{0x00F952, 0x00F952, 0x0052D2},
+{0x00F953, 0x00F953, 0x00808B},
+{0x00F954, 0x00F954, 0x0051DC},
+{0x00F955, 0x00F955, 0x0051CC},
+{0x00F956, 0x00F956, 0x007A1C},
+{0x00F957, 0x00F957, 0x007DBE},
+{0x00F958, 0x00F958, 0x0083F1},
+{0x00F959, 0x00F959, 0x009675},
+{0x00F95A, 0x00F95A, 0x008B80},
+{0x00F95B, 0x00F95B, 0x0062CF},
+{0x00F95C, 0x00F95C, 0x006A02},
+{0x00F95D, 0x00F95D, 0x008AFE},
+{0x00F95E, 0x00F95E, 0x004E39},
+{0x00F95F, 0x00F95F, 0x005BE7},
+{0x00F960, 0x00F960, 0x006012},
+{0x00F961, 0x00F961, 0x007387},
+{0x00F962, 0x00F962, 0x007570},
+{0x00F963, 0x00F963, 0x005317},
+{0x00F964, 0x00F964, 0x0078FB},
+{0x00F965, 0x00F965, 0x004FBF},
+{0x00F966, 0x00F966, 0x005FA9},
+{0x00F967, 0x00F967, 0x004E0D},
+{0x00F968, 0x00F968, 0x006CCC},
+{0x00F969, 0x00F969, 0x006578},
+{0x00F96A, 0x00F96A, 0x007D22},
+{0x00F96B, 0x00F96B, 0x0053C3},
+{0x00F96C, 0x00F96C, 0x00585E},
+{0x00F96D, 0x00F96D, 0x007701},
+{0x00F96E, 0x00F96E, 0x008449},
+{0x00F96F, 0x00F96F, 0x008AAA},
+{0x00F970, 0x00F970, 0x006BBA},
+{0x00F971, 0x00F971, 0x008FB0},
+{0x00F972, 0x00F972, 0x006C88},
+{0x00F973, 0x00F973, 0x0062FE},
+{0x00F974, 0x00F974, 0x0082E5},
+{0x00F975, 0x00F975, 0x0063A0},
+{0x00F976, 0x00F976, 0x007565},
+{0x00F977, 0x00F977, 0x004EAE},
+{0x00F978, 0x00F978, 0x005169},
+{0x00F979, 0x00F979, 0x0051C9},
+{0x00F97A, 0x00F97A, 0x006881},
+{0x00F97B, 0x00F97B, 0x007CE7},
+{0x00F97C, 0x00F97C, 0x00826F},
+{0x00F97D, 0x00F97D, 0x008AD2},
+{0x00F97E, 0x00F97E, 0x0091CF},
+{0x00F97F, 0x00F97F, 0x0052F5},
+{0x00F980, 0x00F980, 0x005442},
+{0x00F981, 0x00F981, 0x005973},
+{0x00F982, 0x00F982, 0x005EEC},
+{0x00F983, 0x00F983, 0x0065C5},
+{0x00F984, 0x00F984, 0x006FFE},
+{0x00F985, 0x00F985, 0x00792A},
+{0x00F986, 0x00F986, 0x0095AD},
+{0x00F987, 0x00F987, 0x009A6A},
+{0x00F988, 0x00F988, 0x009E97},
+{0x00F989, 0x00F989, 0x009ECE},
+{0x00F98A, 0x00F98A, 0x00529B},
+{0x00F98B, 0x00F98B, 0x0066C6},
+{0x00F98C, 0x00F98C, 0x006B77},
+{0x00F98D, 0x00F98D, 0x008F62},
+{0x00F98E, 0x00F98E, 0x005E74},
+{0x00F98F, 0x00F98F, 0x006190},
+{0x00F990, 0x00F990, 0x006200},
+{0x00F991, 0x00F991, 0x00649A},
+{0x00F992, 0x00F992, 0x006F23},
+{0x00F993, 0x00F993, 0x007149},
+{0x00F994, 0x00F994, 0x007489},
+{0x00F995, 0x00F995, 0x0079CA},
+{0x00F996, 0x00F996, 0x007DF4},
+{0x00F997, 0x00F997, 0x00806F},
+{0x00F998, 0x00F998, 0x008F26},
+{0x00F999, 0x00F999, 0x0084EE},
+{0x00F99A, 0x00F99A, 0x009023},
+{0x00F99B, 0x00F99B, 0x00934A},
+{0x00F99C, 0x00F99C, 0x005217},
+{0x00F99D, 0x00F99D, 0x0052A3},
+{0x00F99E, 0x00F99E, 0x0054BD},
+{0x00F99F, 0x00F99F, 0x0070C8},
+{0x00F9A0, 0x00F9A0, 0x0088C2},
+{0x00F9A1, 0x00F9A1, 0x008AAA},
+{0x00F9A2, 0x00F9A2, 0x005EC9},
+{0x00F9A3, 0x00F9A3, 0x005FF5},
+{0x00F9A4, 0x00F9A4, 0x00637B},
+{0x00F9A5, 0x00F9A5, 0x006BAE},
+{0x00F9A6, 0x00F9A6, 0x007C3E},
+{0x00F9A7, 0x00F9A7, 0x007375},
+{0x00F9A8, 0x00F9A8, 0x004EE4},
+{0x00F9A9, 0x00F9A9, 0x0056F9},
+{0x00F9AA, 0x00F9AA, 0x005BE7},
+{0x00F9AB, 0x00F9AB, 0x005DBA},
+{0x00F9AC, 0x00F9AC, 0x00601C},
+{0x00F9AD, 0x00F9AD, 0x0073B2},
+{0x00F9AE, 0x00F9AE, 0x007469},
+{0x00F9AF, 0x00F9AF, 0x007F9A},
+{0x00F9B0, 0x00F9B0, 0x008046},
+{0x00F9B1, 0x00F9B1, 0x009234},
+{0x00F9B2, 0x00F9B2, 0x0096F6},
+{0x00F9B3, 0x00F9B3, 0x009748},
+{0x00F9B4, 0x00F9B4, 0x009818},
+{0x00F9B5, 0x00F9B5, 0x004F8B},
+{0x00F9B6, 0x00F9B6, 0x0079AE},
+{0x00F9B7, 0x00F9B7, 0x0091B4},
+{0x00F9B8, 0x00F9B8, 0x0096B8},
+{0x00F9B9, 0x00F9B9, 0x0060E1},
+{0x00F9BA, 0x00F9BA, 0x004E86},
+{0x00F9BB, 0x00F9BB, 0x0050DA},
+{0x00F9BC, 0x00F9BC, 0x005BEE},
+{0x00F9BD, 0x00F9BD, 0x005C3F},
+{0x00F9BE, 0x00F9BE, 0x006599},
+{0x00F9BF, 0x00F9BF, 0x006A02},
+{0x00F9C0, 0x00F9C0, 0x0071CE},
+{0x00F9C1, 0x00F9C1, 0x007642},
+{0x00F9C2, 0x00F9C2, 0x0084FC},
+{0x00F9C3, 0x00F9C3, 0x00907C},
+{0x00F9C4, 0x00F9C4, 0x009F8D},
+{0x00F9C5, 0x00F9C5, 0x006688},
+{0x00F9C6, 0x00F9C6, 0x00962E},
+{0x00F9C7, 0x00F9C7, 0x005289},
+{0x00F9C8, 0x00F9C8, 0x00677B},
+{0x00F9C9, 0x00F9C9, 0x0067F3},
+{0x00F9CA, 0x00F9CA, 0x006D41},
+{0x00F9CB, 0x00F9CB, 0x006E9C},
+{0x00F9CC, 0x00F9CC, 0x007409},
+{0x00F9CD, 0x00F9CD, 0x007559},
+{0x00F9CE, 0x00F9CE, 0x00786B},
+{0x00F9CF, 0x00F9CF, 0x007D10},
+{0x00F9D0, 0x00F9D0, 0x00985E},
+{0x00F9D1, 0x00F9D1, 0x00516D},
+{0x00F9D2, 0x00F9D2, 0x00622E},
+{0x00F9D3, 0x00F9D3, 0x009678},
+{0x00F9D4, 0x00F9D4, 0x00502B},
+{0x00F9D5, 0x00F9D5, 0x005D19},
+{0x00F9D6, 0x00F9D6, 0x006DEA},
+{0x00F9D7, 0x00F9D7, 0x008F2A},
+{0x00F9D8, 0x00F9D8, 0x005F8B},
+{0x00F9D9, 0x00F9D9, 0x006144},
+{0x00F9DA, 0x00F9DA, 0x006817},
+{0x00F9DB, 0x00F9DB, 0x007387},
+{0x00F9DC, 0x00F9DC, 0x009686},
+{0x00F9DD, 0x00F9DD, 0x005229},
+{0x00F9DE, 0x00F9DE, 0x00540F},
+{0x00F9DF, 0x00F9DF, 0x005C65},
+{0x00F9E0, 0x00F9E0, 0x006613},
+{0x00F9E1, 0x00F9E1, 0x00674E},
+{0x00F9E2, 0x00F9E2, 0x0068A8},
+{0x00F9E3, 0x00F9E3, 0x006CE5},
+{0x00F9E4, 0x00F9E4, 0x007406},
+{0x00F9E5, 0x00F9E5, 0x0075E2},
+{0x00F9E6, 0x00F9E6, 0x007F79},
+{0x00F9E7, 0x00F9E7, 0x0088CF},
+{0x00F9E8, 0x00F9E8, 0x0088E1},
+{0x00F9E9, 0x00F9E9, 0x0091CC},
+{0x00F9EA, 0x00F9EA, 0x0096E2},
+{0x00F9EB, 0x00F9EB, 0x00533F},
+{0x00F9EC, 0x00F9EC, 0x006EBA},
+{0x00F9ED, 0x00F9ED, 0x00541D},
+{0x00F9EE, 0x00F9EE, 0x0071D0},
+{0x00F9EF, 0x00F9EF, 0x007498},
+{0x00F9F0, 0x00F9F0, 0x0085FA},
+{0x00F9F1, 0x00F9F1, 0x0096A3},
+{0x00F9F2, 0x00F9F2, 0x009C57},
+{0x00F9F3, 0x00F9F3, 0x009E9F},
+{0x00F9F4, 0x00F9F4, 0x006797},
+{0x00F9F5, 0x00F9F5, 0x006DCB},
+{0x00F9F6, 0x00F9F6, 0x0081E8},
+{0x00F9F7, 0x00F9F7, 0x007ACB},
+{0x00F9F8, 0x00F9F8, 0x007B20},
+{0x00F9F9, 0x00F9F9, 0x007C92},
+{0x00F9FA, 0x00F9FA, 0x0072C0},
+{0x00F9FB, 0x00F9FB, 0x007099},
+{0x00F9FC, 0x00F9FC, 0x008B58},
+{0x00F9FD, 0x00F9FD, 0x004EC0},
+{0x00F9FE, 0x00F9FE, 0x008336},
+{0x00F9FF, 0x00F9FF, 0x00523A},
+{0x00FA00, 0x00FA00, 0x005207},
+{0x00FA01, 0x00FA01, 0x005EA6},
+{0x00FA02, 0x00FA02, 0x0062D3},
+{0x00FA03, 0x00FA03, 0x007CD6},
+{0x00FA04, 0x00FA04, 0x005B85},
+{0x00FA05, 0x00FA05, 0x006D1E},
+{0x00FA06, 0x00FA06, 0x0066B4},
+{0x00FA07, 0x00FA07, 0x008F3B},
+{0x00FA08, 0x00FA08, 0x00884C},
+{0x00FA09, 0x00FA09, 0x00964D},
+{0x00FA0A, 0x00FA0A, 0x00898B},
+{0x00FA0B, 0x00FA0B, 0x005ED3},
+{0x00FA0C, 0x00FA0C, 0x005140},
+{0x00FA0D, 0x00FA0D, 0x0055C0},
+{0x00FA10, 0x00FA10, 0x00585A},
+{0x00FA12, 0x00FA12, 0x006674},
+{0x00FA15, 0x00FA15, 0x0051DE},
+{0x00FA16, 0x00FA16, 0x00732A},
+{0x00FA17, 0x00FA17, 0x0076CA},
+{0x00FA18, 0x00FA18, 0x00793C},
+{0x00FA19, 0x00FA19, 0x00795E},
+{0x00FA1A, 0x00FA1A, 0x007965},
+{0x00FA1B, 0x00FA1B, 0x00798F},
+{0x00FA1C, 0x00FA1C, 0x009756},
+{0x00FA1D, 0x00FA1D, 0x007CBE},
+{0x00FA1E, 0x00FA1E, 0x007FBD},
+{0x00FA20, 0x00FA20, 0x008612},
+{0x00FA22, 0x00FA22, 0x008AF8},
+{0x00FA25, 0x00FA25, 0x009038},
+{0x00FA26, 0x00FA26, 0x0090FD},
+{0x00FA2A, 0x00FA2A, 0x0098EF},
+{0x00FA2B, 0x00FA2B, 0x0098FC},
+{0x00FA2C, 0x00FA2C, 0x009928},
+{0x00FA2D, 0x00FA2D, 0x009DB4},
+{0x00FA2E, 0x00FA2E, 0x0090DE},
+{0x00FA2F, 0x00FA2F, 0x0096B7},
+{0x00FA30, 0x00FA30, 0x004FAE},
+{0x00FA31, 0x00FA31, 0x0050E7},
+{0x00FA32, 0x00FA32, 0x00514D},
+{0x00FA33, 0x00FA33, 0x0052C9},
+{0x00FA34, 0x00FA34, 0x0052E4},
+{0x00FA35, 0x00FA35, 0x005351},
+{0x00FA36, 0x00FA36, 0x00559D},
+{0x00FA37, 0x00FA37, 0x005606},
+{0x00FA38, 0x00FA38, 0x005668},
+{0x00FA39, 0x00FA39, 0x005840},
+{0x00FA3A, 0x00FA3A, 0x0058A8},
+{0x00FA3B, 0x00FA3B, 0x005C64},
+{0x00FA3C, 0x00FA3C, 0x005C6E},
+{0x00FA3D, 0x00FA3D, 0x006094},
+{0x00FA3E, 0x00FA3E, 0x006168},
+{0x00FA3F, 0x00FA3F, 0x00618E},
+{0x00FA40, 0x00FA40, 0x0061F2},
+{0x00FA41, 0x00FA41, 0x00654F},
+{0x00FA42, 0x00FA42, 0x0065E2},
+{0x00FA43, 0x00FA43, 0x006691},
+{0x00FA44, 0x00FA44, 0x006885},
+{0x00FA45, 0x00FA45, 0x006D77},
+{0x00FA46, 0x00FA46, 0x006E1A},
+{0x00FA47, 0x00FA47, 0x006F22},
+{0x00FA48, 0x00FA48, 0x00716E},
+{0x00FA49, 0x00FA49, 0x00722B},
+{0x00FA4A, 0x00FA4A, 0x007422},
+{0x00FA4B, 0x00FA4B, 0x007891},
+{0x00FA4C, 0x00FA4C, 0x00793E},
+{0x00FA4D, 0x00FA4D, 0x007949},
+{0x00FA4E, 0x00FA4E, 0x007948},
+{0x00FA4F, 0x00FA4F, 0x007950},
+{0x00FA50, 0x00FA50, 0x007956},
+{0x00FA51, 0x00FA51, 0x00795D},
+{0x00FA52, 0x00FA52, 0x00798D},
+{0x00FA53, 0x00FA53, 0x00798E},
+{0x00FA54, 0x00FA54, 0x007A40},
+{0x00FA55, 0x00FA55, 0x007A81},
+{0x00FA56, 0x00FA56, 0x007BC0},
+{0x00FA57, 0x00FA57, 0x007DF4},
+{0x00FA58, 0x00FA58, 0x007E09},
+{0x00FA59, 0x00FA59, 0x007E41},
+{0x00FA5A, 0x00FA5A, 0x007F72},
+{0x00FA5B, 0x00FA5B, 0x008005},
+{0x00FA5C, 0x00FA5C, 0x0081ED},
+{0x00FA5D, 0x00FA5E, 0x008279},
+{0x00FA5F, 0x00FA5F, 0x008457},
+{0x00FA60, 0x00FA60, 0x008910},
+{0x00FA61, 0x00FA61, 0x008996},
+{0x00FA62, 0x00FA62, 0x008B01},
+{0x00FA63, 0x00FA63, 0x008B39},
+{0x00FA64, 0x00FA64, 0x008CD3},
+{0x00FA65, 0x00FA65, 0x008D08},
+{0x00FA66, 0x00FA66, 0x008FB6},
+{0x00FA67, 0x00FA67, 0x009038},
+{0x00FA68, 0x00FA68, 0x0096E3},
+{0x00FA69, 0x00FA69, 0x0097FF},
+{0x00FA6A, 0x00FA6A, 0x00983B},
+{0x00FA6B, 0x00FA6B, 0x006075},
+{0x00FA6C, 0x00FA6C, 0x0242EE},
+{0x00FA6D, 0x00FA6D, 0x008218},
+{0x00FA70, 0x00FA70, 0x004E26},
+{0x00FA71, 0x00FA71, 0x0051B5},
+{0x00FA72, 0x00FA72, 0x005168},
+{0x00FA73, 0x00FA73, 0x004F80},
+{0x00FA74, 0x00FA74, 0x005145},
+{0x00FA75, 0x00FA75, 0x005180},
+{0x00FA76, 0x00FA76, 0x0052C7},
+{0x00FA77, 0x00FA77, 0x0052FA},
+{0x00FA78, 0x00FA78, 0x00559D},
+{0x00FA79, 0x00FA79, 0x005555},
+{0x00FA7A, 0x00FA7A, 0x005599},
+{0x00FA7B, 0x00FA7B, 0x0055E2},
+{0x00FA7C, 0x00FA7C, 0x00585A},
+{0x00FA7D, 0x00FA7D, 0x0058B3},
+{0x00FA7E, 0x00FA7E, 0x005944},
+{0x00FA7F, 0x00FA7F, 0x005954},
+{0x00FA80, 0x00FA80, 0x005A62},
+{0x00FA81, 0x00FA81, 0x005B28},
+{0x00FA82, 0x00FA82, 0x005ED2},
+{0x00FA83, 0x00FA83, 0x005ED9},
+{0x00FA84, 0x00FA84, 0x005F69},
+{0x00FA85, 0x00FA85, 0x005FAD},
+{0x00FA86, 0x00FA86, 0x0060D8},
+{0x00FA87, 0x00FA87, 0x00614E},
+{0x00FA88, 0x00FA88, 0x006108},
+{0x00FA89, 0x00FA89, 0x00618E},
+{0x00FA8A, 0x00FA8A, 0x006160},
+{0x00FA8B, 0x00FA8B, 0x0061F2},
+{0x00FA8C, 0x00FA8C, 0x006234},
+{0x00FA8D, 0x00FA8D, 0x0063C4},
+{0x00FA8E, 0x00FA8E, 0x00641C},
+{0x00FA8F, 0x00FA8F, 0x006452},
+{0x00FA90, 0x00FA90, 0x006556},
+{0x00FA91, 0x00FA91, 0x006674},
+{0x00FA92, 0x00FA92, 0x006717},
+{0x00FA93, 0x00FA93, 0x00671B},
+{0x00FA94, 0x00FA94, 0x006756},
+{0x00FA95, 0x00FA95, 0x006B79},
+{0x00FA96, 0x00FA96, 0x006BBA},
+{0x00FA97, 0x00FA97, 0x006D41},
+{0x00FA98, 0x00FA98, 0x006EDB},
+{0x00FA99, 0x00FA99, 0x006ECB},
+{0x00FA9A, 0x00FA9A, 0x006F22},
+{0x00FA9B, 0x00FA9B, 0x00701E},
+{0x00FA9C, 0x00FA9C, 0x00716E},
+{0x00FA9D, 0x00FA9D, 0x0077A7},
+{0x00FA9E, 0x00FA9E, 0x007235},
+{0x00FA9F, 0x00FA9F, 0x0072AF},
+{0x00FAA0, 0x00FAA0, 0x00732A},
+{0x00FAA1, 0x00FAA1, 0x007471},
+{0x00FAA2, 0x00FAA2, 0x007506},
+{0x00FAA3, 0x00FAA3, 0x00753B},
+{0x00FAA4, 0x00FAA4, 0x00761D},
+{0x00FAA5, 0x00FAA5, 0x00761F},
+{0x00FAA6, 0x00FAA6, 0x0076CA},
+{0x00FAA7, 0x00FAA7, 0x0076DB},
+{0x00FAA8, 0x00FAA8, 0x0076F4},
+{0x00FAA9, 0x00FAA9, 0x00774A},
+{0x00FAAA, 0x00FAAA, 0x007740},
+{0x00FAAB, 0x00FAAB, 0x0078CC},
+{0x00FAAC, 0x00FAAC, 0x007AB1},
+{0x00FAAD, 0x00FAAD, 0x007BC0},
+{0x00FAAE, 0x00FAAE, 0x007C7B},
+{0x00FAAF, 0x00FAAF, 0x007D5B},
+{0x00FAB0, 0x00FAB0, 0x007DF4},
+{0x00FAB1, 0x00FAB1, 0x007F3E},
+{0x00FAB2, 0x00FAB2, 0x008005},
+{0x00FAB3, 0x00FAB3, 0x008352},
+{0x00FAB4, 0x00FAB4, 0x0083EF},
+{0x00FAB5, 0x00FAB5, 0x008779},
+{0x00FAB6, 0x00FAB6, 0x008941},
+{0x00FAB7, 0x00FAB7, 0x008986},
+{0x00FAB8, 0x00FAB8, 0x008996},
+{0x00FAB9, 0x00FAB9, 0x008ABF},
+{0x00FABA, 0x00FABA, 0x008AF8},
+{0x00FABB, 0x00FABB, 0x008ACB},
+{0x00FABC, 0x00FABC, 0x008B01},
+{0x00FABD, 0x00FABD, 0x008AFE},
+{0x00FABE, 0x00FABE, 0x008AED},
+{0x00FABF, 0x00FABF, 0x008B39},
+{0x00FAC0, 0x00FAC0, 0x008B8A},
+{0x00FAC1, 0x00FAC1, 0x008D08},
+{0x00FAC2, 0x00FAC2, 0x008F38},
+{0x00FAC3, 0x00FAC3, 0x009072},
+{0x00FAC4, 0x00FAC4, 0x009199},
+{0x00FAC5, 0x00FAC5, 0x009276},
+{0x00FAC6, 0x00FAC6, 0x00967C},
+{0x00FAC7, 0x00FAC7, 0x0096E3},
+{0x00FAC8, 0x00FAC8, 0x009756},
+{0x00FAC9, 0x00FAC9, 0x0097DB},
+{0x00FACA, 0x00FACA, 0x0097FF},
+{0x00FACB, 0x00FACB, 0x00980B},
+{0x00FACC, 0x00FACC, 0x00983B},
+{0x00FACD, 0x00FACD, 0x009B12},
+{0x00FACE, 0x00FACE, 0x009F9C},
+{0x00FACF, 0x00FACF, 0x02284A},
+{0x00FAD0, 0x00FAD0, 0x022844},
+{0x00FAD1, 0x00FAD1, 0x0233D5},
+{0x00FAD2, 0x00FAD2, 0x003B9D},
+{0x00FAD3, 0x00FAD3, 0x004018},
+{0x00FAD4, 0x00FAD4, 0x004039},
+{0x00FAD5, 0x00FAD5, 0x025249},
+{0x00FAD6, 0x00FAD6, 0x025CD0},
+{0x00FAD7, 0x00FAD7, 0x027ED3},
+{0x00FAD8, 0x00FAD8, 0x009F43},
+{0x00FAD9, 0x00FAD9, 0x009F8E},
+{0x00FB1D, 0x00FB1D, 0x0005D9},
+{0x00FB1F, 0x00FB1F, 0x0005F2},
+{0x00FB2A, 0x00FB2D, 0x0005E9},
+{0x00FB2E, 0x00FB30, 0x0005D0},
+{0x00FB31, 0x00FB31, 0x0005D1},
+{0x00FB32, 0x00FB32, 0x0005D2},
+{0x00FB33, 0x00FB33, 0x0005D3},
+{0x00FB34, 0x00FB34, 0x0005D4},
+{0x00FB35, 0x00FB35, 0x0005D5},
+{0x00FB36, 0x00FB36, 0x0005D6},
+{0x00FB38, 0x00FB38, 0x0005D8},
+{0x00FB39, 0x00FB39, 0x0005D9},
+{0x00FB3A, 0x00FB3A, 0x0005DA},
+{0x00FB3B, 0x00FB3B, 0x0005DB},
+{0x00FB3C, 0x00FB3C, 0x0005DC},
+{0x00FB3E, 0x00FB3E, 0x0005DE},
+{0x00FB40, 0x00FB40, 0x0005E0},
+{0x00FB41, 0x00FB41, 0x0005E1},
+{0x00FB43, 0x00FB43, 0x0005E3},
+{0x00FB44, 0x00FB44, 0x0005E4},
+{0x00FB46, 0x00FB46, 0x0005E6},
+{0x00FB47, 0x00FB47, 0x0005E7},
+{0x00FB48, 0x00FB48, 0x0005E8},
+{0x00FB49, 0x00FB49, 0x0005E9},
+{0x00FB4A, 0x00FB4A, 0x0005EA},
+{0x00FB4B, 0x00FB4B, 0x0005D5},
+{0x00FB4C, 0x00FB4C, 0x0005D1},
+{0x00FB4D, 0x00FB4D, 0x0005DB},
+{0x00FB4E, 0x00FB4E, 0x0005E4},
+{0x01109A, 0x01109A, 0x011099},
+{0x01109C, 0x01109C, 0x01109B},
+{0x0110AB, 0x0110AB, 0x0110A5},
+{0x01112E, 0x01112E, 0x011131},
+{0x01112F, 0x01112F, 0x011132},
+{0x01134B, 0x01134C, 0x011347},
+{0x0114BB, 0x0114BC, 0x0114B9},
+{0x0114BE, 0x0114BE, 0x0114B9},
+{0x0115BA, 0x0115BA, 0x0115B8},
+{0x0115BB, 0x0115BB, 0x0115B9},
+{0x011938, 0x011938, 0x011935},
+{0x01D15E, 0x01D15E, 0x01D157},
+{0x01D15F, 0x01D164, 0x01D158},
+{0x01D1BB, 0x01D1BB, 0x01D1B9},
+{0x01D1BC, 0x01D1BC, 0x01D1BA},
+{0x01D1BD, 0x01D1BD, 0x01D1B9},
+{0x01D1BE, 0x01D1BE, 0x01D1BA},
+{0x01D1BF, 0x01D1BF, 0x01D1B9},
+{0x01D1C0, 0x01D1C0, 0x01D1BA},
+{0x02F800, 0x02F800, 0x004E3D},
+{0x02F801, 0x02F801, 0x004E38},
+{0x02F802, 0x02F802, 0x004E41},
+{0x02F803, 0x02F803, 0x020122},
+{0x02F804, 0x02F804, 0x004F60},
+{0x02F805, 0x02F805, 0x004FAE},
+{0x02F806, 0x02F806, 0x004FBB},
+{0x02F807, 0x02F807, 0x005002},
+{0x02F808, 0x02F808, 0x00507A},
+{0x02F809, 0x02F809, 0x005099},
+{0x02F80A, 0x02F80A, 0x0050E7},
+{0x02F80B, 0x02F80B, 0x0050CF},
+{0x02F80C, 0x02F80C, 0x00349E},
+{0x02F80D, 0x02F80D, 0x02063A},
+{0x02F80E, 0x02F80E, 0x00514D},
+{0x02F80F, 0x02F80F, 0x005154},
+{0x02F810, 0x02F810, 0x005164},
+{0x02F811, 0x02F811, 0x005177},
+{0x02F812, 0x02F812, 0x02051C},
+{0x02F813, 0x02F813, 0x0034B9},
+{0x02F814, 0x02F814, 0x005167},
+{0x02F815, 0x02F815, 0x00518D},
+{0x02F816, 0x02F816, 0x02054B},
+{0x02F817, 0x02F817, 0x005197},
+{0x02F818, 0x02F818, 0x0051A4},
+{0x02F819, 0x02F819, 0x004ECC},
+{0x02F81A, 0x02F81A, 0x0051AC},
+{0x02F81B, 0x02F81B, 0x0051B5},
+{0x02F81C, 0x02F81C, 0x0291DF},
+{0x02F81D, 0x02F81D, 0x0051F5},
+{0x02F81E, 0x02F81E, 0x005203},
+{0x02F81F, 0x02F81F, 0x0034DF},
+{0x02F820, 0x02F820, 0x00523B},
+{0x02F821, 0x02F821, 0x005246},
+{0x02F822, 0x02F822, 0x005272},
+{0x02F823, 0x02F823, 0x005277},
+{0x02F824, 0x02F824, 0x003515},
+{0x02F825, 0x02F825, 0x0052C7},
+{0x02F826, 0x02F826, 0x0052C9},
+{0x02F827, 0x02F827, 0x0052E4},
+{0x02F828, 0x02F828, 0x0052FA},
+{0x02F829, 0x02F829, 0x005305},
+{0x02F82A, 0x02F82A, 0x005306},
+{0x02F82B, 0x02F82B, 0x005317},
+{0x02F82C, 0x02F82C, 0x005349},
+{0x02F82D, 0x02F82D, 0x005351},
+{0x02F82E, 0x02F82E, 0x00535A},
+{0x02F82F, 0x02F82F, 0x005373},
+{0x02F830, 0x02F830, 0x00537D},
+{0x02F831, 0x02F833, 0x00537F},
+{0x02F834, 0x02F834, 0x020A2C},
+{0x02F835, 0x02F835, 0x007070},
+{0x02F836, 0x02F836, 0x0053CA},
+{0x02F837, 0x02F837, 0x0053DF},
+{0x02F838, 0x02F838, 0x020B63},
+{0x02F839, 0x02F839, 0x0053EB},
+{0x02F83A, 0x02F83A, 0x0053F1},
+{0x02F83B, 0x02F83B, 0x005406},
+{0x02F83C, 0x02F83C, 0x00549E},
+{0x02F83D, 0x02F83D, 0x005438},
+{0x02F83E, 0x02F83E, 0x005448},
+{0x02F83F, 0x02F83F, 0x005468},
+{0x02F840, 0x02F840, 0x0054A2},
+{0x02F841, 0x02F841, 0x0054F6},
+{0x02F842, 0x02F842, 0x005510},
+{0x02F843, 0x02F843, 0x005553},
+{0x02F844, 0x02F844, 0x005563},
+{0x02F845, 0x02F846, 0x005584},
+{0x02F847, 0x02F847, 0x005599},
+{0x02F848, 0x02F848, 0x0055AB},
+{0x02F849, 0x02F849, 0x0055B3},
+{0x02F84A, 0x02F84A, 0x0055C2},
+{0x02F84B, 0x02F84B, 0x005716},
+{0x02F84C, 0x02F84C, 0x005606},
+{0x02F84D, 0x02F84D, 0x005717},
+{0x02F84E, 0x02F84E, 0x005651},
+{0x02F84F, 0x02F84F, 0x005674},
+{0x02F850, 0x02F850, 0x005207},
+{0x02F851, 0x02F851, 0x0058EE},
+{0x02F852, 0x02F852, 0x0057CE},
+{0x02F853, 0x02F853, 0x0057F4},
+{0x02F854, 0x02F854, 0x00580D},
+{0x02F855, 0x02F855, 0x00578B},
+{0x02F856, 0x02F856, 0x005832},
+{0x02F857, 0x02F857, 0x005831},
+{0x02F858, 0x02F858, 0x0058AC},
+{0x02F859, 0x02F859, 0x0214E4},
+{0x02F85A, 0x02F85A, 0x0058F2},
+{0x02F85B, 0x02F85B, 0x0058F7},
+{0x02F85C, 0x02F85C, 0x005906},
+{0x02F85D, 0x02F85D, 0x00591A},
+{0x02F85E, 0x02F85E, 0x005922},
+{0x02F85F, 0x02F85F, 0x005962},
+{0x02F860, 0x02F860, 0x0216A8},
+{0x02F861, 0x02F861, 0x0216EA},
+{0x02F862, 0x02F862, 0x0059EC},
+{0x02F863, 0x02F863, 0x005A1B},
+{0x02F864, 0x02F864, 0x005A27},
+{0x02F865, 0x02F865, 0x0059D8},
+{0x02F866, 0x02F866, 0x005A66},
+{0x02F867, 0x02F867, 0x0036EE},
+{0x02F868, 0x02F868, 0x0036FC},
+{0x02F869, 0x02F869, 0x005B08},
+{0x02F86A, 0x02F86B, 0x005B3E},
+{0x02F86C, 0x02F86C, 0x0219C8},
+{0x02F86D, 0x02F86D, 0x005BC3},
+{0x02F86E, 0x02F86E, 0x005BD8},
+{0x02F86F, 0x02F86F, 0x005BE7},
+{0x02F870, 0x02F870, 0x005BF3},
+{0x02F871, 0x02F871, 0x021B18},
+{0x02F872, 0x02F872, 0x005BFF},
+{0x02F873, 0x02F873, 0x005C06},
+{0x02F874, 0x02F874, 0x005F53},
+{0x02F875, 0x02F875, 0x005C22},
+{0x02F876, 0x02F876, 0x003781},
+{0x02F877, 0x02F877, 0x005C60},
+{0x02F878, 0x02F878, 0x005C6E},
+{0x02F879, 0x02F879, 0x005CC0},
+{0x02F87A, 0x02F87A, 0x005C8D},
+{0x02F87B, 0x02F87B, 0x021DE4},
+{0x02F87C, 0x02F87C, 0x005D43},
+{0x02F87D, 0x02F87D, 0x021DE6},
+{0x02F87E, 0x02F87E, 0x005D6E},
+{0x02F87F, 0x02F87F, 0x005D6B},
+{0x02F880, 0x02F880, 0x005D7C},
+{0x02F881, 0x02F881, 0x005DE1},
+{0x02F882, 0x02F882, 0x005DE2},
+{0x02F883, 0x02F883, 0x00382F},
+{0x02F884, 0x02F884, 0x005DFD},
+{0x02F885, 0x02F885, 0x005E28},
+{0x02F886, 0x02F886, 0x005E3D},
+{0x02F887, 0x02F887, 0x005E69},
+{0x02F888, 0x02F888, 0x003862},
+{0x02F889, 0x02F889, 0x022183},
+{0x02F88A, 0x02F88A, 0x00387C},
+{0x02F88B, 0x02F88B, 0x005EB0},
+{0x02F88C, 0x02F88C, 0x005EB3},
+{0x02F88D, 0x02F88D, 0x005EB6},
+{0x02F88E, 0x02F88E, 0x005ECA},
+{0x02F88F, 0x02F88F, 0x02A392},
+{0x02F890, 0x02F890, 0x005EFE},
+{0x02F891, 0x02F892, 0x022331},
+{0x02F893, 0x02F893, 0x008201},
+{0x02F894, 0x02F895, 0x005F22},
+{0x02F896, 0x02F896, 0x0038C7},
+{0x02F897, 0x02F897, 0x0232B8},
+{0x02F898, 0x02F898, 0x0261DA},
+{0x02F899, 0x02F899, 0x005F62},
+{0x02F89A, 0x02F89A, 0x005F6B},
+{0x02F89B, 0x02F89B, 0x0038E3},
+{0x02F89C, 0x02F89C, 0x005F9A},
+{0x02F89D, 0x02F89D, 0x005FCD},
+{0x02F89E, 0x02F89E, 0x005FD7},
+{0x02F89F, 0x02F89F, 0x005FF9},
+{0x02F8A0, 0x02F8A0, 0x006081},
+{0x02F8A1, 0x02F8A1, 0x00393A},
+{0x02F8A2, 0x02F8A2, 0x00391C},
+{0x02F8A3, 0x02F8A3, 0x006094},
+{0x02F8A4, 0x02F8A4, 0x0226D4},
+{0x02F8A5, 0x02F8A5, 0x0060C7},
+{0x02F8A6, 0x02F8A6, 0x006148},
+{0x02F8A7, 0x02F8A7, 0x00614C},
+{0x02F8A8, 0x02F8A8, 0x00614E},
+{0x02F8A9, 0x02F8A9, 0x00614C},
+{0x02F8AA, 0x02F8AA, 0x00617A},
+{0x02F8AB, 0x02F8AB, 0x00618E},
+{0x02F8AC, 0x02F8AC, 0x0061B2},
+{0x02F8AD, 0x02F8AD, 0x0061A4},
+{0x02F8AE, 0x02F8AE, 0x0061AF},
+{0x02F8AF, 0x02F8AF, 0x0061DE},
+{0x02F8B0, 0x02F8B0, 0x0061F2},
+{0x02F8B1, 0x02F8B1, 0x0061F6},
+{0x02F8B2, 0x02F8B2, 0x006210},
+{0x02F8B3, 0x02F8B3, 0x00621B},
+{0x02F8B4, 0x02F8B4, 0x00625D},
+{0x02F8B5, 0x02F8B5, 0x0062B1},
+{0x02F8B6, 0x02F8B6, 0x0062D4},
+{0x02F8B7, 0x02F8B7, 0x006350},
+{0x02F8B8, 0x02F8B8, 0x022B0C},
+{0x02F8B9, 0x02F8B9, 0x00633D},
+{0x02F8BA, 0x02F8BA, 0x0062FC},
+{0x02F8BB, 0x02F8BB, 0x006368},
+{0x02F8BC, 0x02F8BC, 0x006383},
+{0x02F8BD, 0x02F8BD, 0x0063E4},
+{0x02F8BE, 0x02F8BE, 0x022BF1},
+{0x02F8BF, 0x02F8BF, 0x006422},
+{0x02F8C0, 0x02F8C0, 0x0063C5},
+{0x02F8C1, 0x02F8C1, 0x0063A9},
+{0x02F8C2, 0x02F8C2, 0x003A2E},
+{0x02F8C3, 0x02F8C3, 0x006469},
+{0x02F8C4, 0x02F8C4, 0x00647E},
+{0x02F8C5, 0x02F8C5, 0x00649D},
+{0x02F8C6, 0x02F8C6, 0x006477},
+{0x02F8C7, 0x02F8C7, 0x003A6C},
+{0x02F8C8, 0x02F8C8, 0x00654F},
+{0x02F8C9, 0x02F8C9, 0x00656C},
+{0x02F8CA, 0x02F8CA, 0x02300A},
+{0x02F8CB, 0x02F8CB, 0x0065E3},
+{0x02F8CC, 0x02F8CC, 0x0066F8},
+{0x02F8CD, 0x02F8CD, 0x006649},
+{0x02F8CE, 0x02F8CE, 0x003B19},
+{0x02F8CF, 0x02F8CF, 0x006691},
+{0x02F8D0, 0x02F8D0, 0x003B08},
+{0x02F8D1, 0x02F8D1, 0x003AE4},
+{0x02F8D2, 0x02F8D2, 0x005192},
+{0x02F8D3, 0x02F8D3, 0x005195},
+{0x02F8D4, 0x02F8D4, 0x006700},
+{0x02F8D5, 0x02F8D5, 0x00669C},
+{0x02F8D6, 0x02F8D6, 0x0080AD},
+{0x02F8D7, 0x02F8D7, 0x0043D9},
+{0x02F8D8, 0x02F8D8, 0x006717},
+{0x02F8D9, 0x02F8D9, 0x00671B},
+{0x02F8DA, 0x02F8DA, 0x006721},
+{0x02F8DB, 0x02F8DB, 0x00675E},
+{0x02F8DC, 0x02F8DC, 0x006753},
+{0x02F8DD, 0x02F8DD, 0x0233C3},
+{0x02F8DE, 0x02F8DE, 0x003B49},
+{0x02F8DF, 0x02F8DF, 0x0067FA},
+{0x02F8E0, 0x02F8E0, 0x006785},
+{0x02F8E1, 0x02F8E1, 0x006852},
+{0x02F8E2, 0x02F8E2, 0x006885},
+{0x02F8E3, 0x02F8E3, 0x02346D},
+{0x02F8E4, 0x02F8E4, 0x00688E},
+{0x02F8E5, 0x02F8E5, 0x00681F},
+{0x02F8E6, 0x02F8E6, 0x006914},
+{0x02F8E7, 0x02F8E7, 0x003B9D},
+{0x02F8E8, 0x02F8E8, 0x006942},
+{0x02F8E9, 0x02F8E9, 0x0069A3},
+{0x02F8EA, 0x02F8EA, 0x0069EA},
+{0x02F8EB, 0x02F8EB, 0x006AA8},
+{0x02F8EC, 0x02F8EC, 0x0236A3},
+{0x02F8ED, 0x02F8ED, 0x006ADB},
+{0x02F8EE, 0x02F8EE, 0x003C18},
+{0x02F8EF, 0x02F8EF, 0x006B21},
+{0x02F8F0, 0x02F8F0, 0x0238A7},
+{0x02F8F1, 0x02F8F1, 0x006B54},
+{0x02F8F2, 0x02F8F2, 0x003C4E},
+{0x02F8F3, 0x02F8F3, 0x006B72},
+{0x02F8F4, 0x02F8F4, 0x006B9F},
+{0x02F8F5, 0x02F8F5, 0x006BBA},
+{0x02F8F6, 0x02F8F6, 0x006BBB},
+{0x02F8F7, 0x02F8F7, 0x023A8D},
+{0x02F8F8, 0x02F8F8, 0x021D0B},
+{0x02F8F9, 0x02F8F9, 0x023AFA},
+{0x02F8FA, 0x02F8FA, 0x006C4E},
+{0x02F8FB, 0x02F8FB, 0x023CBC},
+{0x02F8FC, 0x02F8FC, 0x006CBF},
+{0x02F8FD, 0x02F8FD, 0x006CCD},
+{0x02F8FE, 0x02F8FE, 0x006C67},
+{0x02F8FF, 0x02F8FF, 0x006D16},
+{0x02F900, 0x02F900, 0x006D3E},
+{0x02F901, 0x02F901, 0x006D77},
+{0x02F902, 0x02F902, 0x006D41},
+{0x02F903, 0x02F903, 0x006D69},
+{0x02F904, 0x02F904, 0x006D78},
+{0x02F905, 0x02F905, 0x006D85},
+{0x02F906, 0x02F906, 0x023D1E},
+{0x02F907, 0x02F907, 0x006D34},
+{0x02F908, 0x02F908, 0x006E2F},
+{0x02F909, 0x02F909, 0x006E6E},
+{0x02F90A, 0x02F90A, 0x003D33},
+{0x02F90B, 0x02F90B, 0x006ECB},
+{0x02F90C, 0x02F90C, 0x006EC7},
+{0x02F90D, 0x02F90D, 0x023ED1},
+{0x02F90E, 0x02F90E, 0x006DF9},
+{0x02F90F, 0x02F90F, 0x006F6E},
+{0x02F910, 0x02F910, 0x023F5E},
+{0x02F911, 0x02F911, 0x023F8E},
+{0x02F912, 0x02F912, 0x006FC6},
+{0x02F913, 0x02F913, 0x007039},
+{0x02F914, 0x02F914, 0x00701E},
+{0x02F915, 0x02F915, 0x00701B},
+{0x02F916, 0x02F916, 0x003D96},
+{0x02F917, 0x02F917, 0x00704A},
+{0x02F918, 0x02F918, 0x00707D},
+{0x02F919, 0x02F919, 0x007077},
+{0x02F91A, 0x02F91A, 0x0070AD},
+{0x02F91B, 0x02F91B, 0x020525},
+{0x02F91C, 0x02F91C, 0x007145},
+{0x02F91D, 0x02F91D, 0x024263},
+{0x02F91E, 0x02F91E, 0x00719C},
+{0x02F91F, 0x02F91F, 0x0243AB},
+{0x02F920, 0x02F920, 0x007228},
+{0x02F921, 0x02F921, 0x007235},
+{0x02F922, 0x02F922, 0x007250},
+{0x02F923, 0x02F923, 0x024608},
+{0x02F924, 0x02F924, 0x007280},
+{0x02F925, 0x02F925, 0x007295},
+{0x02F926, 0x02F926, 0x024735},
+{0x02F927, 0x02F927, 0x024814},
+{0x02F928, 0x02F928, 0x00737A},
+{0x02F929, 0x02F929, 0x00738B},
+{0x02F92A, 0x02F92A, 0x003EAC},
+{0x02F92B, 0x02F92B, 0x0073A5},
+{0x02F92C, 0x02F92D, 0x003EB8},
+{0x02F92E, 0x02F92E, 0x007447},
+{0x02F92F, 0x02F92F, 0x00745C},
+{0x02F930, 0x02F930, 0x007471},
+{0x02F931, 0x02F931, 0x007485},
+{0x02F932, 0x02F932, 0x0074CA},
+{0x02F933, 0x02F933, 0x003F1B},
+{0x02F934, 0x02F934, 0x007524},
+{0x02F935, 0x02F935, 0x024C36},
+{0x02F936, 0x02F936, 0x00753E},
+{0x02F937, 0x02F937, 0x024C92},
+{0x02F938, 0x02F938, 0x007570},
+{0x02F939, 0x02F939, 0x02219F},
+{0x02F93A, 0x02F93A, 0x007610},
+{0x02F93B, 0x02F93B, 0x024FA1},
+{0x02F93C, 0x02F93C, 0x024FB8},
+{0x02F93D, 0x02F93D, 0x025044},
+{0x02F93E, 0x02F93E, 0x003FFC},
+{0x02F93F, 0x02F93F, 0x004008},
+{0x02F940, 0x02F940, 0x0076F4},
+{0x02F941, 0x02F941, 0x0250F3},
+{0x02F942, 0x02F942, 0x0250F2},
+{0x02F943, 0x02F943, 0x025119},
+{0x02F944, 0x02F944, 0x025133},
+{0x02F945, 0x02F945, 0x00771E},
+{0x02F946, 0x02F947, 0x00771F},
+{0x02F948, 0x02F948, 0x00774A},
+{0x02F949, 0x02F949, 0x004039},
+{0x02F94A, 0x02F94A, 0x00778B},
+{0x02F94B, 0x02F94B, 0x004046},
+{0x02F94C, 0x02F94C, 0x004096},
+{0x02F94D, 0x02F94D, 0x02541D},
+{0x02F94E, 0x02F94E, 0x00784E},
+{0x02F94F, 0x02F94F, 0x00788C},
+{0x02F950, 0x02F950, 0x0078CC},
+{0x02F951, 0x02F951, 0x0040E3},
+{0x02F952, 0x02F952, 0x025626},
+{0x02F953, 0x02F953, 0x007956},
+{0x02F954, 0x02F954, 0x02569A},
+{0x02F955, 0x02F955, 0x0256C5},
+{0x02F956, 0x02F956, 0x00798F},
+{0x02F957, 0x02F957, 0x0079EB},
+{0x02F958, 0x02F958, 0x00412F},
+{0x02F959, 0x02F959, 0x007A40},
+{0x02F95A, 0x02F95A, 0x007A4A},
+{0x02F95B, 0x02F95B, 0x007A4F},
+{0x02F95C, 0x02F95C, 0x02597C},
+{0x02F95D, 0x02F95E, 0x025AA7},
+{0x02F95F, 0x02F95F, 0x007AEE},
+{0x02F960, 0x02F960, 0x004202},
+{0x02F961, 0x02F961, 0x025BAB},
+{0x02F962, 0x02F962, 0x007BC6},
+{0x02F963, 0x02F963, 0x007BC9},
+{0x02F964, 0x02F964, 0x004227},
+{0x02F965, 0x02F965, 0x025C80},
+{0x02F966, 0x02F966, 0x007CD2},
+{0x02F967, 0x02F967, 0x0042A0},
+{0x02F968, 0x02F968, 0x007CE8},
+{0x02F969, 0x02F969, 0x007CE3},
+{0x02F96A, 0x02F96A, 0x007D00},
+{0x02F96B, 0x02F96B, 0x025F86},
+{0x02F96C, 0x02F96C, 0x007D63},
+{0x02F96D, 0x02F96D, 0x004301},
+{0x02F96E, 0x02F96E, 0x007DC7},
+{0x02F96F, 0x02F96F, 0x007E02},
+{0x02F970, 0x02F970, 0x007E45},
+{0x02F971, 0x02F971, 0x004334},
+{0x02F972, 0x02F972, 0x026228},
+{0x02F973, 0x02F973, 0x026247},
+{0x02F974, 0x02F974, 0x004359},
+{0x02F975, 0x02F975, 0x0262D9},
+{0x02F976, 0x02F976, 0x007F7A},
+{0x02F977, 0x02F977, 0x02633E},
+{0x02F978, 0x02F978, 0x007F95},
+{0x02F979, 0x02F979, 0x007FFA},
+{0x02F97A, 0x02F97A, 0x008005},
+{0x02F97B, 0x02F97B, 0x0264DA},
+{0x02F97C, 0x02F97C, 0x026523},
+{0x02F97D, 0x02F97D, 0x008060},
+{0x02F97E, 0x02F97E, 0x0265A8},
+{0x02F97F, 0x02F97F, 0x008070},
+{0x02F980, 0x02F980, 0x02335F},
+{0x02F981, 0x02F981, 0x0043D5},
+{0x02F982, 0x02F982, 0x0080B2},
+{0x02F983, 0x02F983, 0x008103},
+{0x02F984, 0x02F984, 0x00440B},
+{0x02F985, 0x02F985, 0x00813E},
+{0x02F986, 0x02F986, 0x005AB5},
+{0x02F987, 0x02F987, 0x0267A7},
+{0x02F988, 0x02F988, 0x0267B5},
+{0x02F989, 0x02F989, 0x023393},
+{0x02F98A, 0x02F98A, 0x02339C},
+{0x02F98B, 0x02F98B, 0x008201},
+{0x02F98C, 0x02F98C, 0x008204},
+{0x02F98D, 0x02F98D, 0x008F9E},
+{0x02F98E, 0x02F98E, 0x00446B},
+{0x02F98F, 0x02F98F, 0x008291},
+{0x02F990, 0x02F990, 0x00828B},
+{0x02F991, 0x02F991, 0x00829D},
+{0x02F992, 0x02F992, 0x0052B3},
+{0x02F993, 0x02F993, 0x0082B1},
+{0x02F994, 0x02F994, 0x0082B3},
+{0x02F995, 0x02F995, 0x0082BD},
+{0x02F996, 0x02F996, 0x0082E6},
+{0x02F997, 0x02F997, 0x026B3C},
+{0x02F998, 0x02F998, 0x0082E5},
+{0x02F999, 0x02F999, 0x00831D},
+{0x02F99A, 0x02F99A, 0x008363},
+{0x02F99B, 0x02F99B, 0x0083AD},
+{0x02F99C, 0x02F99C, 0x008323},
+{0x02F99D, 0x02F99D, 0x0083BD},
+{0x02F99E, 0x02F99E, 0x0083E7},
+{0x02F99F, 0x02F99F, 0x008457},
+{0x02F9A0, 0x02F9A0, 0x008353},
+{0x02F9A1, 0x02F9A1, 0x0083CA},
+{0x02F9A2, 0x02F9A2, 0x0083CC},
+{0x02F9A3, 0x02F9A3, 0x0083DC},
+{0x02F9A4, 0x02F9A4, 0x026C36},
+{0x02F9A5, 0x02F9A5, 0x026D6B},
+{0x02F9A6, 0x02F9A6, 0x026CD5},
+{0x02F9A7, 0x02F9A7, 0x00452B},
+{0x02F9A8, 0x02F9A8, 0x0084F1},
+{0x02F9A9, 0x02F9A9, 0x0084F3},
+{0x02F9AA, 0x02F9AA, 0x008516},
+{0x02F9AB, 0x02F9AB, 0x0273CA},
+{0x02F9AC, 0x02F9AC, 0x008564},
+{0x02F9AD, 0x02F9AD, 0x026F2C},
+{0x02F9AE, 0x02F9AE, 0x00455D},
+{0x02F9AF, 0x02F9AF, 0x004561},
+{0x02F9B0, 0x02F9B0, 0x026FB1},
+{0x02F9B1, 0x02F9B1, 0x0270D2},
+{0x02F9B2, 0x02F9B2, 0x00456B},
+{0x02F9B3, 0x02F9B3, 0x008650},
+{0x02F9B4, 0x02F9B4, 0x00865C},
+{0x02F9B5, 0x02F9B5, 0x008667},
+{0x02F9B6, 0x02F9B6, 0x008669},
+{0x02F9B7, 0x02F9B7, 0x0086A9},
+{0x02F9B8, 0x02F9B8, 0x008688},
+{0x02F9B9, 0x02F9B9, 0x00870E},
+{0x02F9BA, 0x02F9BA, 0x0086E2},
+{0x02F9BB, 0x02F9BB, 0x008779},
+{0x02F9BC, 0x02F9BC, 0x008728},
+{0x02F9BD, 0x02F9BD, 0x00876B},
+{0x02F9BE, 0x02F9BE, 0x008786},
+{0x02F9BF, 0x02F9BF, 0x0045D7},
+{0x02F9C0, 0x02F9C0, 0x0087E1},
+{0x02F9C1, 0x02F9C1, 0x008801},
+{0x02F9C2, 0x02F9C2, 0x0045F9},
+{0x02F9C3, 0x02F9C3, 0x008860},
+{0x02F9C4, 0x02F9C4, 0x008863},
+{0x02F9C5, 0x02F9C5, 0x027667},
+{0x02F9C6, 0x02F9C6, 0x0088D7},
+{0x02F9C7, 0x02F9C7, 0x0088DE},
+{0x02F9C8, 0x02F9C8, 0x004635},
+{0x02F9C9, 0x02F9C9, 0x0088FA},
+{0x02F9CA, 0x02F9CA, 0x0034BB},
+{0x02F9CB, 0x02F9CB, 0x0278AE},
+{0x02F9CC, 0x02F9CC, 0x027966},
+{0x02F9CD, 0x02F9CD, 0x0046BE},
+{0x02F9CE, 0x02F9CE, 0x0046C7},
+{0x02F9CF, 0x02F9CF, 0x008AA0},
+{0x02F9D0, 0x02F9D0, 0x008AED},
+{0x02F9D1, 0x02F9D1, 0x008B8A},
+{0x02F9D2, 0x02F9D2, 0x008C55},
+{0x02F9D3, 0x02F9D3, 0x027CA8},
+{0x02F9D4, 0x02F9D4, 0x008CAB},
+{0x02F9D5, 0x02F9D5, 0x008CC1},
+{0x02F9D6, 0x02F9D6, 0x008D1B},
+{0x02F9D7, 0x02F9D7, 0x008D77},
+{0x02F9D8, 0x02F9D8, 0x027F2F},
+{0x02F9D9, 0x02F9D9, 0x020804},
+{0x02F9DA, 0x02F9DA, 0x008DCB},
+{0x02F9DB, 0x02F9DB, 0x008DBC},
+{0x02F9DC, 0x02F9DC, 0x008DF0},
+{0x02F9DD, 0x02F9DD, 0x0208DE},
+{0x02F9DE, 0x02F9DE, 0x008ED4},
+{0x02F9DF, 0x02F9DF, 0x008F38},
+{0x02F9E0, 0x02F9E0, 0x0285D2},
+{0x02F9E1, 0x02F9E1, 0x0285ED},
+{0x02F9E2, 0x02F9E2, 0x009094},
+{0x02F9E3, 0x02F9E3, 0x0090F1},
+{0x02F9E4, 0x02F9E4, 0x009111},
+{0x02F9E5, 0x02F9E5, 0x02872E},
+{0x02F9E6, 0x02F9E6, 0x00911B},
+{0x02F9E7, 0x02F9E7, 0x009238},
+{0x02F9E8, 0x02F9E8, 0x0092D7},
+{0x02F9E9, 0x02F9E9, 0x0092D8},
+{0x02F9EA, 0x02F9EA, 0x00927C},
+{0x02F9EB, 0x02F9EB, 0x0093F9},
+{0x02F9EC, 0x02F9EC, 0x009415},
+{0x02F9ED, 0x02F9ED, 0x028BFA},
+{0x02F9EE, 0x02F9EE, 0x00958B},
+{0x02F9EF, 0x02F9EF, 0x004995},
+{0x02F9F0, 0x02F9F0, 0x0095B7},
+{0x02F9F1, 0x02F9F1, 0x028D77},
+{0x02F9F2, 0x02F9F2, 0x0049E6},
+{0x02F9F3, 0x02F9F3, 0x0096C3},
+{0x02F9F4, 0x02F9F4, 0x005DB2},
+{0x02F9F5, 0x02F9F5, 0x009723},
+{0x02F9F6, 0x02F9F6, 0x029145},
+{0x02F9F7, 0x02F9F7, 0x02921A},
+{0x02F9F8, 0x02F9F8, 0x004A6E},
+{0x02F9F9, 0x02F9F9, 0x004A76},
+{0x02F9FA, 0x02F9FA, 0x0097E0},
+{0x02F9FB, 0x02F9FB, 0x02940A},
+{0x02F9FC, 0x02F9FC, 0x004AB2},
+{0x02F9FD, 0x02F9FD, 0x029496},
+{0x02F9FE, 0x02F9FF, 0x00980B},
+{0x02FA00, 0x02FA00, 0x009829},
+{0x02FA01, 0x02FA01, 0x0295B6},
+{0x02FA02, 0x02FA02, 0x0098E2},
+{0x02FA03, 0x02FA03, 0x004B33},
+{0x02FA04, 0x02FA04, 0x009929},
+{0x02FA05, 0x02FA05, 0x0099A7},
+{0x02FA06, 0x02FA06, 0x0099C2},
+{0x02FA07, 0x02FA07, 0x0099FE},
+{0x02FA08, 0x02FA08, 0x004BCE},
+{0x02FA09, 0x02FA09, 0x029B30},
+{0x02FA0A, 0x02FA0A, 0x009B12},
+{0x02FA0B, 0x02FA0B, 0x009C40},
+{0x02FA0C, 0x02FA0C, 0x009CFD},
+{0x02FA0D, 0x02FA0D, 0x004CCE},
+{0x02FA0E, 0x02FA0E, 0x004CED},
+{0x02FA0F, 0x02FA0F, 0x009D67},
+{0x02FA10, 0x02FA10, 0x02A0CE},
+{0x02FA11, 0x02FA11, 0x004CF8},
+{0x02FA12, 0x02FA12, 0x02A105},
+{0x02FA13, 0x02FA13, 0x02A20E},
+{0x02FA14, 0x02FA14, 0x02A291},
+{0x02FA15, 0x02FA15, 0x009EBB},
+{0x02FA16, 0x02FA16, 0x004D56},
+{0x02FA17, 0x02FA17, 0x009EF9},
+{0x02FA18, 0x02FA18, 0x009EFE},
+{0x02FA19, 0x02FA19, 0x009F05},
+{0x02FA1A, 0x02FA1A, 0x009F0F},
+{0x02FA1B, 0x02FA1B, 0x009F16},
+{0x02FA1C, 0x02FA1C, 0x009F3B},
+{0x02FA1D, 0x02FA1D, 0x02A600},
 };
 
-const std::map<char32_t, char32_t> unicode_map_lowercase = {
-{0x00041, 0x00061}, {0x00042, 0x00062}, {0x00043, 0x00063}, {0x00044, 0x00064}, {0x00045, 0x00065}, {0x00046, 0x00066},
-{0x00047, 0x00067}, {0x00048, 0x00068}, {0x00049, 0x00069}, {0x0004A, 0x0006A}, {0x0004B, 0x0006B}, {0x0004C, 0x0006C},
-{0x0004D, 0x0006D}, {0x0004E, 0x0006E}, {0x0004F, 0x0006F}, {0x00050, 0x00070}, {0x00051, 0x00071}, {0x00052, 0x00072},
-{0x00053, 0x00073}, {0x00054, 0x00074}, {0x00055, 0x00075}, {0x00056, 0x00076}, {0x00057, 0x00077}, {0x00058, 0x00078},
-{0x00059, 0x00079}, {0x0005A, 0x0007A}, {0x000C0, 0x000E0}, {0x000C1, 0x000E1}, {0x000C2, 0x000E2}, {0x000C3, 0x000E3},
-{0x000C4, 0x000E4}, {0x000C5, 0x000E5}, {0x000C6, 0x000E6}, {0x000C7, 0x000E7}, {0x000C8, 0x000E8}, {0x000C9, 0x000E9},
-{0x000CA, 0x000EA}, {0x000CB, 0x000EB}, {0x000CC, 0x000EC}, {0x000CD, 0x000ED}, {0x000CE, 0x000EE}, {0x000CF, 0x000EF},
-{0x000D0, 0x000F0}, {0x000D1, 0x000F1}, {0x000D2, 0x000F2}, {0x000D3, 0x000F3}, {0x000D4, 0x000F4}, {0x000D5, 0x000F5},
-{0x000D6, 0x000F6}, {0x000D8, 0x000F8}, {0x000D9, 0x000F9}, {0x000DA, 0x000FA}, {0x000DB, 0x000FB}, {0x000DC, 0x000FC},
-{0x000DD, 0x000FD}, {0x000DE, 0x000FE}, {0x00100, 0x00101}, {0x00102, 0x00103}, {0x00104, 0x00105}, {0x00106, 0x00107},
-{0x00108, 0x00109}, {0x0010A, 0x0010B}, {0x0010C, 0x0010D}, {0x0010E, 0x0010F}, {0x00110, 0x00111}, {0x00112, 0x00113},
-{0x00114, 0x00115}, {0x00116, 0x00117}, {0x00118, 0x00119}, {0x0011A, 0x0011B}, {0x0011C, 0x0011D}, {0x0011E, 0x0011F},
-{0x00120, 0x00121}, {0x00122, 0x00123}, {0x00124, 0x00125}, {0x00126, 0x00127}, {0x00128, 0x00129}, {0x0012A, 0x0012B},
-{0x0012C, 0x0012D}, {0x0012E, 0x0012F}, {0x00130, 0x00069}, {0x00132, 0x00133}, {0x00134, 0x00135}, {0x00136, 0x00137},
-{0x00139, 0x0013A}, {0x0013B, 0x0013C}, {0x0013D, 0x0013E}, {0x0013F, 0x00140}, {0x00141, 0x00142}, {0x00143, 0x00144},
-{0x00145, 0x00146}, {0x00147, 0x00148}, {0x0014A, 0x0014B}, {0x0014C, 0x0014D}, {0x0014E, 0x0014F}, {0x00150, 0x00151},
-{0x00152, 0x00153}, {0x00154, 0x00155}, {0x00156, 0x00157}, {0x00158, 0x00159}, {0x0015A, 0x0015B}, {0x0015C, 0x0015D},
-{0x0015E, 0x0015F}, {0x00160, 0x00161}, {0x00162, 0x00163}, {0x00164, 0x00165}, {0x00166, 0x00167}, {0x00168, 0x00169},
-{0x0016A, 0x0016B}, {0x0016C, 0x0016D}, {0x0016E, 0x0016F}, {0x00170, 0x00171}, {0x00172, 0x00173}, {0x00174, 0x00175},
-{0x00176, 0x00177}, {0x00178, 0x000FF}, {0x00179, 0x0017A}, {0x0017B, 0x0017C}, {0x0017D, 0x0017E}, {0x00181, 0x00253},
-{0x00182, 0x00183}, {0x00184, 0x00185}, {0x00186, 0x00254}, {0x00187, 0x00188}, {0x00189, 0x00256}, {0x0018A, 0x00257},
-{0x0018B, 0x0018C}, {0x0018E, 0x001DD}, {0x0018F, 0x00259}, {0x00190, 0x0025B}, {0x00191, 0x00192}, {0x00193, 0x00260},
-{0x00194, 0x00263}, {0x00196, 0x00269}, {0x00197, 0x00268}, {0x00198, 0x00199}, {0x0019C, 0x0026F}, {0x0019D, 0x00272},
-{0x0019F, 0x00275}, {0x001A0, 0x001A1}, {0x001A2, 0x001A3}, {0x001A4, 0x001A5}, {0x001A6, 0x00280}, {0x001A7, 0x001A8},
-{0x001A9, 0x00283}, {0x001AC, 0x001AD}, {0x001AE, 0x00288}, {0x001AF, 0x001B0}, {0x001B1, 0x0028A}, {0x001B2, 0x0028B},
-{0x001B3, 0x001B4}, {0x001B5, 0x001B6}, {0x001B7, 0x00292}, {0x001B8, 0x001B9}, {0x001BC, 0x001BD}, {0x001C4, 0x001C6},
-{0x001C5, 0x001C6}, {0x001C7, 0x001C9}, {0x001C8, 0x001C9}, {0x001CA, 0x001CC}, {0x001CB, 0x001CC}, {0x001CD, 0x001CE},
-{0x001CF, 0x001D0}, {0x001D1, 0x001D2}, {0x001D3, 0x001D4}, {0x001D5, 0x001D6}, {0x001D7, 0x001D8}, {0x001D9, 0x001DA},
-{0x001DB, 0x001DC}, {0x001DE, 0x001DF}, {0x001E0, 0x001E1}, {0x001E2, 0x001E3}, {0x001E4, 0x001E5}, {0x001E6, 0x001E7},
-{0x001E8, 0x001E9}, {0x001EA, 0x001EB}, {0x001EC, 0x001ED}, {0x001EE, 0x001EF}, {0x001F1, 0x001F3}, {0x001F2, 0x001F3},
-{0x001F4, 0x001F5}, {0x001F6, 0x00195}, {0x001F7, 0x001BF}, {0x001F8, 0x001F9}, {0x001FA, 0x001FB}, {0x001FC, 0x001FD},
-{0x001FE, 0x001FF}, {0x00200, 0x00201}, {0x00202, 0x00203}, {0x00204, 0x00205}, {0x00206, 0x00207}, {0x00208, 0x00209},
-{0x0020A, 0x0020B}, {0x0020C, 0x0020D}, {0x0020E, 0x0020F}, {0x00210, 0x00211}, {0x00212, 0x00213}, {0x00214, 0x00215},
-{0x00216, 0x00217}, {0x00218, 0x00219}, {0x0021A, 0x0021B}, {0x0021C, 0x0021D}, {0x0021E, 0x0021F}, {0x00220, 0x0019E},
-{0x00222, 0x00223}, {0x00224, 0x00225}, {0x00226, 0x00227}, {0x00228, 0x00229}, {0x0022A, 0x0022B}, {0x0022C, 0x0022D},
-{0x0022E, 0x0022F}, {0x00230, 0x00231}, {0x00232, 0x00233}, {0x0023A, 0x02C65}, {0x0023B, 0x0023C}, {0x0023D, 0x0019A},
-{0x0023E, 0x02C66}, {0x00241, 0x00242}, {0x00243, 0x00180}, {0x00244, 0x00289}, {0x00245, 0x0028C}, {0x00246, 0x00247},
-{0x00248, 0x00249}, {0x0024A, 0x0024B}, {0x0024C, 0x0024D}, {0x0024E, 0x0024F}, {0x00370, 0x00371}, {0x00372, 0x00373},
-{0x00376, 0x00377}, {0x0037F, 0x003F3}, {0x00386, 0x003AC}, {0x00388, 0x003AD}, {0x00389, 0x003AE}, {0x0038A, 0x003AF},
-{0x0038C, 0x003CC}, {0x0038E, 0x003CD}, {0x0038F, 0x003CE}, {0x00391, 0x003B1}, {0x00392, 0x003B2}, {0x00393, 0x003B3},
-{0x00394, 0x003B4}, {0x00395, 0x003B5}, {0x00396, 0x003B6}, {0x00397, 0x003B7}, {0x00398, 0x003B8}, {0x00399, 0x003B9},
-{0x0039A, 0x003BA}, {0x0039B, 0x003BB}, {0x0039C, 0x003BC}, {0x0039D, 0x003BD}, {0x0039E, 0x003BE}, {0x0039F, 0x003BF},
-{0x003A0, 0x003C0}, {0x003A1, 0x003C1}, {0x003A3, 0x003C3}, {0x003A4, 0x003C4}, {0x003A5, 0x003C5}, {0x003A6, 0x003C6},
-{0x003A7, 0x003C7}, {0x003A8, 0x003C8}, {0x003A9, 0x003C9}, {0x003AA, 0x003CA}, {0x003AB, 0x003CB}, {0x003CF, 0x003D7},
-{0x003D8, 0x003D9}, {0x003DA, 0x003DB}, {0x003DC, 0x003DD}, {0x003DE, 0x003DF}, {0x003E0, 0x003E1}, {0x003E2, 0x003E3},
-{0x003E4, 0x003E5}, {0x003E6, 0x003E7}, {0x003E8, 0x003E9}, {0x003EA, 0x003EB}, {0x003EC, 0x003ED}, {0x003EE, 0x003EF},
-{0x003F4, 0x003B8}, {0x003F7, 0x003F8}, {0x003F9, 0x003F2}, {0x003FA, 0x003FB}, {0x003FD, 0x0037B}, {0x003FE, 0x0037C},
-{0x003FF, 0x0037D}, {0x00400, 0x00450}, {0x00401, 0x00451}, {0x00402, 0x00452}, {0x00403, 0x00453}, {0x00404, 0x00454},
-{0x00405, 0x00455}, {0x00406, 0x00456}, {0x00407, 0x00457}, {0x00408, 0x00458}, {0x00409, 0x00459}, {0x0040A, 0x0045A},
-{0x0040B, 0x0045B}, {0x0040C, 0x0045C}, {0x0040D, 0x0045D}, {0x0040E, 0x0045E}, {0x0040F, 0x0045F}, {0x00410, 0x00430},
-{0x00411, 0x00431}, {0x00412, 0x00432}, {0x00413, 0x00433}, {0x00414, 0x00434}, {0x00415, 0x00435}, {0x00416, 0x00436},
-{0x00417, 0x00437}, {0x00418, 0x00438}, {0x00419, 0x00439}, {0x0041A, 0x0043A}, {0x0041B, 0x0043B}, {0x0041C, 0x0043C},
-{0x0041D, 0x0043D}, {0x0041E, 0x0043E}, {0x0041F, 0x0043F}, {0x00420, 0x00440}, {0x00421, 0x00441}, {0x00422, 0x00442},
-{0x00423, 0x00443}, {0x00424, 0x00444}, {0x00425, 0x00445}, {0x00426, 0x00446}, {0x00427, 0x00447}, {0x00428, 0x00448},
-{0x00429, 0x00449}, {0x0042A, 0x0044A}, {0x0042B, 0x0044B}, {0x0042C, 0x0044C}, {0x0042D, 0x0044D}, {0x0042E, 0x0044E},
-{0x0042F, 0x0044F}, {0x00460, 0x00461}, {0x00462, 0x00463}, {0x00464, 0x00465}, {0x00466, 0x00467}, {0x00468, 0x00469},
-{0x0046A, 0x0046B}, {0x0046C, 0x0046D}, {0x0046E, 0x0046F}, {0x00470, 0x00471}, {0x00472, 0x00473}, {0x00474, 0x00475},
-{0x00476, 0x00477}, {0x00478, 0x00479}, {0x0047A, 0x0047B}, {0x0047C, 0x0047D}, {0x0047E, 0x0047F}, {0x00480, 0x00481},
-{0x0048A, 0x0048B}, {0x0048C, 0x0048D}, {0x0048E, 0x0048F}, {0x00490, 0x00491}, {0x00492, 0x00493}, {0x00494, 0x00495},
-{0x00496, 0x00497}, {0x00498, 0x00499}, {0x0049A, 0x0049B}, {0x0049C, 0x0049D}, {0x0049E, 0x0049F}, {0x004A0, 0x004A1},
-{0x004A2, 0x004A3}, {0x004A4, 0x004A5}, {0x004A6, 0x004A7}, {0x004A8, 0x004A9}, {0x004AA, 0x004AB}, {0x004AC, 0x004AD},
-{0x004AE, 0x004AF}, {0x004B0, 0x004B1}, {0x004B2, 0x004B3}, {0x004B4, 0x004B5}, {0x004B6, 0x004B7}, {0x004B8, 0x004B9},
-{0x004BA, 0x004BB}, {0x004BC, 0x004BD}, {0x004BE, 0x004BF}, {0x004C0, 0x004CF}, {0x004C1, 0x004C2}, {0x004C3, 0x004C4},
-{0x004C5, 0x004C6}, {0x004C7, 0x004C8}, {0x004C9, 0x004CA}, {0x004CB, 0x004CC}, {0x004CD, 0x004CE}, {0x004D0, 0x004D1},
-{0x004D2, 0x004D3}, {0x004D4, 0x004D5}, {0x004D6, 0x004D7}, {0x004D8, 0x004D9}, {0x004DA, 0x004DB}, {0x004DC, 0x004DD},
-{0x004DE, 0x004DF}, {0x004E0, 0x004E1}, {0x004E2, 0x004E3}, {0x004E4, 0x004E5}, {0x004E6, 0x004E7}, {0x004E8, 0x004E9},
-{0x004EA, 0x004EB}, {0x004EC, 0x004ED}, {0x004EE, 0x004EF}, {0x004F0, 0x004F1}, {0x004F2, 0x004F3}, {0x004F4, 0x004F5},
-{0x004F6, 0x004F7}, {0x004F8, 0x004F9}, {0x004FA, 0x004FB}, {0x004FC, 0x004FD}, {0x004FE, 0x004FF}, {0x00500, 0x00501},
-{0x00502, 0x00503}, {0x00504, 0x00505}, {0x00506, 0x00507}, {0x00508, 0x00509}, {0x0050A, 0x0050B}, {0x0050C, 0x0050D},
-{0x0050E, 0x0050F}, {0x00510, 0x00511}, {0x00512, 0x00513}, {0x00514, 0x00515}, {0x00516, 0x00517}, {0x00518, 0x00519},
-{0x0051A, 0x0051B}, {0x0051C, 0x0051D}, {0x0051E, 0x0051F}, {0x00520, 0x00521}, {0x00522, 0x00523}, {0x00524, 0x00525},
-{0x00526, 0x00527}, {0x00528, 0x00529}, {0x0052A, 0x0052B}, {0x0052C, 0x0052D}, {0x0052E, 0x0052F}, {0x00531, 0x00561},
-{0x00532, 0x00562}, {0x00533, 0x00563}, {0x00534, 0x00564}, {0x00535, 0x00565}, {0x00536, 0x00566}, {0x00537, 0x00567},
-{0x00538, 0x00568}, {0x00539, 0x00569}, {0x0053A, 0x0056A}, {0x0053B, 0x0056B}, {0x0053C, 0x0056C}, {0x0053D, 0x0056D},
-{0x0053E, 0x0056E}, {0x0053F, 0x0056F}, {0x00540, 0x00570}, {0x00541, 0x00571}, {0x00542, 0x00572}, {0x00543, 0x00573},
-{0x00544, 0x00574}, {0x00545, 0x00575}, {0x00546, 0x00576}, {0x00547, 0x00577}, {0x00548, 0x00578}, {0x00549, 0x00579},
-{0x0054A, 0x0057A}, {0x0054B, 0x0057B}, {0x0054C, 0x0057C}, {0x0054D, 0x0057D}, {0x0054E, 0x0057E}, {0x0054F, 0x0057F},
-{0x00550, 0x00580}, {0x00551, 0x00581}, {0x00552, 0x00582}, {0x00553, 0x00583}, {0x00554, 0x00584}, {0x00555, 0x00585},
-{0x00556, 0x00586}, {0x010A0, 0x02D00}, {0x010A1, 0x02D01}, {0x010A2, 0x02D02}, {0x010A3, 0x02D03}, {0x010A4, 0x02D04},
-{0x010A5, 0x02D05}, {0x010A6, 0x02D06}, {0x010A7, 0x02D07}, {0x010A8, 0x02D08}, {0x010A9, 0x02D09}, {0x010AA, 0x02D0A},
-{0x010AB, 0x02D0B}, {0x010AC, 0x02D0C}, {0x010AD, 0x02D0D}, {0x010AE, 0x02D0E}, {0x010AF, 0x02D0F}, {0x010B0, 0x02D10},
-{0x010B1, 0x02D11}, {0x010B2, 0x02D12}, {0x010B3, 0x02D13}, {0x010B4, 0x02D14}, {0x010B5, 0x02D15}, {0x010B6, 0x02D16},
-{0x010B7, 0x02D17}, {0x010B8, 0x02D18}, {0x010B9, 0x02D19}, {0x010BA, 0x02D1A}, {0x010BB, 0x02D1B}, {0x010BC, 0x02D1C},
-{0x010BD, 0x02D1D}, {0x010BE, 0x02D1E}, {0x010BF, 0x02D1F}, {0x010C0, 0x02D20}, {0x010C1, 0x02D21}, {0x010C2, 0x02D22},
-{0x010C3, 0x02D23}, {0x010C4, 0x02D24}, {0x010C5, 0x02D25}, {0x010C7, 0x02D27}, {0x010CD, 0x02D2D}, {0x013A0, 0x0AB70},
-{0x013A1, 0x0AB71}, {0x013A2, 0x0AB72}, {0x013A3, 0x0AB73}, {0x013A4, 0x0AB74}, {0x013A5, 0x0AB75}, {0x013A6, 0x0AB76},
-{0x013A7, 0x0AB77}, {0x013A8, 0x0AB78}, {0x013A9, 0x0AB79}, {0x013AA, 0x0AB7A}, {0x013AB, 0x0AB7B}, {0x013AC, 0x0AB7C},
-{0x013AD, 0x0AB7D}, {0x013AE, 0x0AB7E}, {0x013AF, 0x0AB7F}, {0x013B0, 0x0AB80}, {0x013B1, 0x0AB81}, {0x013B2, 0x0AB82},
-{0x013B3, 0x0AB83}, {0x013B4, 0x0AB84}, {0x013B5, 0x0AB85}, {0x013B6, 0x0AB86}, {0x013B7, 0x0AB87}, {0x013B8, 0x0AB88},
-{0x013B9, 0x0AB89}, {0x013BA, 0x0AB8A}, {0x013BB, 0x0AB8B}, {0x013BC, 0x0AB8C}, {0x013BD, 0x0AB8D}, {0x013BE, 0x0AB8E},
-{0x013BF, 0x0AB8F}, {0x013C0, 0x0AB90}, {0x013C1, 0x0AB91}, {0x013C2, 0x0AB92}, {0x013C3, 0x0AB93}, {0x013C4, 0x0AB94},
-{0x013C5, 0x0AB95}, {0x013C6, 0x0AB96}, {0x013C7, 0x0AB97}, {0x013C8, 0x0AB98}, {0x013C9, 0x0AB99}, {0x013CA, 0x0AB9A},
-{0x013CB, 0x0AB9B}, {0x013CC, 0x0AB9C}, {0x013CD, 0x0AB9D}, {0x013CE, 0x0AB9E}, {0x013CF, 0x0AB9F}, {0x013D0, 0x0ABA0},
-{0x013D1, 0x0ABA1}, {0x013D2, 0x0ABA2}, {0x013D3, 0x0ABA3}, {0x013D4, 0x0ABA4}, {0x013D5, 0x0ABA5}, {0x013D6, 0x0ABA6},
-{0x013D7, 0x0ABA7}, {0x013D8, 0x0ABA8}, {0x013D9, 0x0ABA9}, {0x013DA, 0x0ABAA}, {0x013DB, 0x0ABAB}, {0x013DC, 0x0ABAC},
-{0x013DD, 0x0ABAD}, {0x013DE, 0x0ABAE}, {0x013DF, 0x0ABAF}, {0x013E0, 0x0ABB0}, {0x013E1, 0x0ABB1}, {0x013E2, 0x0ABB2},
-{0x013E3, 0x0ABB3}, {0x013E4, 0x0ABB4}, {0x013E5, 0x0ABB5}, {0x013E6, 0x0ABB6}, {0x013E7, 0x0ABB7}, {0x013E8, 0x0ABB8},
-{0x013E9, 0x0ABB9}, {0x013EA, 0x0ABBA}, {0x013EB, 0x0ABBB}, {0x013EC, 0x0ABBC}, {0x013ED, 0x0ABBD}, {0x013EE, 0x0ABBE},
-{0x013EF, 0x0ABBF}, {0x013F0, 0x013F8}, {0x013F1, 0x013F9}, {0x013F2, 0x013FA}, {0x013F3, 0x013FB}, {0x013F4, 0x013FC},
-{0x013F5, 0x013FD}, {0x01C90, 0x010D0}, {0x01C91, 0x010D1}, {0x01C92, 0x010D2}, {0x01C93, 0x010D3}, {0x01C94, 0x010D4},
-{0x01C95, 0x010D5}, {0x01C96, 0x010D6}, {0x01C97, 0x010D7}, {0x01C98, 0x010D8}, {0x01C99, 0x010D9}, {0x01C9A, 0x010DA},
-{0x01C9B, 0x010DB}, {0x01C9C, 0x010DC}, {0x01C9D, 0x010DD}, {0x01C9E, 0x010DE}, {0x01C9F, 0x010DF}, {0x01CA0, 0x010E0},
-{0x01CA1, 0x010E1}, {0x01CA2, 0x010E2}, {0x01CA3, 0x010E3}, {0x01CA4, 0x010E4}, {0x01CA5, 0x010E5}, {0x01CA6, 0x010E6},
-{0x01CA7, 0x010E7}, {0x01CA8, 0x010E8}, {0x01CA9, 0x010E9}, {0x01CAA, 0x010EA}, {0x01CAB, 0x010EB}, {0x01CAC, 0x010EC},
-{0x01CAD, 0x010ED}, {0x01CAE, 0x010EE}, {0x01CAF, 0x010EF}, {0x01CB0, 0x010F0}, {0x01CB1, 0x010F1}, {0x01CB2, 0x010F2},
-{0x01CB3, 0x010F3}, {0x01CB4, 0x010F4}, {0x01CB5, 0x010F5}, {0x01CB6, 0x010F6}, {0x01CB7, 0x010F7}, {0x01CB8, 0x010F8},
-{0x01CB9, 0x010F9}, {0x01CBA, 0x010FA}, {0x01CBD, 0x010FD}, {0x01CBE, 0x010FE}, {0x01CBF, 0x010FF}, {0x01E00, 0x01E01},
-{0x01E02, 0x01E03}, {0x01E04, 0x01E05}, {0x01E06, 0x01E07}, {0x01E08, 0x01E09}, {0x01E0A, 0x01E0B}, {0x01E0C, 0x01E0D},
-{0x01E0E, 0x01E0F}, {0x01E10, 0x01E11}, {0x01E12, 0x01E13}, {0x01E14, 0x01E15}, {0x01E16, 0x01E17}, {0x01E18, 0x01E19},
-{0x01E1A, 0x01E1B}, {0x01E1C, 0x01E1D}, {0x01E1E, 0x01E1F}, {0x01E20, 0x01E21}, {0x01E22, 0x01E23}, {0x01E24, 0x01E25},
-{0x01E26, 0x01E27}, {0x01E28, 0x01E29}, {0x01E2A, 0x01E2B}, {0x01E2C, 0x01E2D}, {0x01E2E, 0x01E2F}, {0x01E30, 0x01E31},
-{0x01E32, 0x01E33}, {0x01E34, 0x01E35}, {0x01E36, 0x01E37}, {0x01E38, 0x01E39}, {0x01E3A, 0x01E3B}, {0x01E3C, 0x01E3D},
-{0x01E3E, 0x01E3F}, {0x01E40, 0x01E41}, {0x01E42, 0x01E43}, {0x01E44, 0x01E45}, {0x01E46, 0x01E47}, {0x01E48, 0x01E49},
-{0x01E4A, 0x01E4B}, {0x01E4C, 0x01E4D}, {0x01E4E, 0x01E4F}, {0x01E50, 0x01E51}, {0x01E52, 0x01E53}, {0x01E54, 0x01E55},
-{0x01E56, 0x01E57}, {0x01E58, 0x01E59}, {0x01E5A, 0x01E5B}, {0x01E5C, 0x01E5D}, {0x01E5E, 0x01E5F}, {0x01E60, 0x01E61},
-{0x01E62, 0x01E63}, {0x01E64, 0x01E65}, {0x01E66, 0x01E67}, {0x01E68, 0x01E69}, {0x01E6A, 0x01E6B}, {0x01E6C, 0x01E6D},
-{0x01E6E, 0x01E6F}, {0x01E70, 0x01E71}, {0x01E72, 0x01E73}, {0x01E74, 0x01E75}, {0x01E76, 0x01E77}, {0x01E78, 0x01E79},
-{0x01E7A, 0x01E7B}, {0x01E7C, 0x01E7D}, {0x01E7E, 0x01E7F}, {0x01E80, 0x01E81}, {0x01E82, 0x01E83}, {0x01E84, 0x01E85},
-{0x01E86, 0x01E87}, {0x01E88, 0x01E89}, {0x01E8A, 0x01E8B}, {0x01E8C, 0x01E8D}, {0x01E8E, 0x01E8F}, {0x01E90, 0x01E91},
-{0x01E92, 0x01E93}, {0x01E94, 0x01E95}, {0x01E9E, 0x000DF}, {0x01EA0, 0x01EA1}, {0x01EA2, 0x01EA3}, {0x01EA4, 0x01EA5},
-{0x01EA6, 0x01EA7}, {0x01EA8, 0x01EA9}, {0x01EAA, 0x01EAB}, {0x01EAC, 0x01EAD}, {0x01EAE, 0x01EAF}, {0x01EB0, 0x01EB1},
-{0x01EB2, 0x01EB3}, {0x01EB4, 0x01EB5}, {0x01EB6, 0x01EB7}, {0x01EB8, 0x01EB9}, {0x01EBA, 0x01EBB}, {0x01EBC, 0x01EBD},
-{0x01EBE, 0x01EBF}, {0x01EC0, 0x01EC1}, {0x01EC2, 0x01EC3}, {0x01EC4, 0x01EC5}, {0x01EC6, 0x01EC7}, {0x01EC8, 0x01EC9},
-{0x01ECA, 0x01ECB}, {0x01ECC, 0x01ECD}, {0x01ECE, 0x01ECF}, {0x01ED0, 0x01ED1}, {0x01ED2, 0x01ED3}, {0x01ED4, 0x01ED5},
-{0x01ED6, 0x01ED7}, {0x01ED8, 0x01ED9}, {0x01EDA, 0x01EDB}, {0x01EDC, 0x01EDD}, {0x01EDE, 0x01EDF}, {0x01EE0, 0x01EE1},
-{0x01EE2, 0x01EE3}, {0x01EE4, 0x01EE5}, {0x01EE6, 0x01EE7}, {0x01EE8, 0x01EE9}, {0x01EEA, 0x01EEB}, {0x01EEC, 0x01EED},
-{0x01EEE, 0x01EEF}, {0x01EF0, 0x01EF1}, {0x01EF2, 0x01EF3}, {0x01EF4, 0x01EF5}, {0x01EF6, 0x01EF7}, {0x01EF8, 0x01EF9},
-{0x01EFA, 0x01EFB}, {0x01EFC, 0x01EFD}, {0x01EFE, 0x01EFF}, {0x01F08, 0x01F00}, {0x01F09, 0x01F01}, {0x01F0A, 0x01F02},
-{0x01F0B, 0x01F03}, {0x01F0C, 0x01F04}, {0x01F0D, 0x01F05}, {0x01F0E, 0x01F06}, {0x01F0F, 0x01F07}, {0x01F18, 0x01F10},
-{0x01F19, 0x01F11}, {0x01F1A, 0x01F12}, {0x01F1B, 0x01F13}, {0x01F1C, 0x01F14}, {0x01F1D, 0x01F15}, {0x01F28, 0x01F20},
-{0x01F29, 0x01F21}, {0x01F2A, 0x01F22}, {0x01F2B, 0x01F23}, {0x01F2C, 0x01F24}, {0x01F2D, 0x01F25}, {0x01F2E, 0x01F26},
-{0x01F2F, 0x01F27}, {0x01F38, 0x01F30}, {0x01F39, 0x01F31}, {0x01F3A, 0x01F32}, {0x01F3B, 0x01F33}, {0x01F3C, 0x01F34},
-{0x01F3D, 0x01F35}, {0x01F3E, 0x01F36}, {0x01F3F, 0x01F37}, {0x01F48, 0x01F40}, {0x01F49, 0x01F41}, {0x01F4A, 0x01F42},
-{0x01F4B, 0x01F43}, {0x01F4C, 0x01F44}, {0x01F4D, 0x01F45}, {0x01F59, 0x01F51}, {0x01F5B, 0x01F53}, {0x01F5D, 0x01F55},
-{0x01F5F, 0x01F57}, {0x01F68, 0x01F60}, {0x01F69, 0x01F61}, {0x01F6A, 0x01F62}, {0x01F6B, 0x01F63}, {0x01F6C, 0x01F64},
-{0x01F6D, 0x01F65}, {0x01F6E, 0x01F66}, {0x01F6F, 0x01F67}, {0x01F88, 0x01F80}, {0x01F89, 0x01F81}, {0x01F8A, 0x01F82},
-{0x01F8B, 0x01F83}, {0x01F8C, 0x01F84}, {0x01F8D, 0x01F85}, {0x01F8E, 0x01F86}, {0x01F8F, 0x01F87}, {0x01F98, 0x01F90},
-{0x01F99, 0x01F91}, {0x01F9A, 0x01F92}, {0x01F9B, 0x01F93}, {0x01F9C, 0x01F94}, {0x01F9D, 0x01F95}, {0x01F9E, 0x01F96},
-{0x01F9F, 0x01F97}, {0x01FA8, 0x01FA0}, {0x01FA9, 0x01FA1}, {0x01FAA, 0x01FA2}, {0x01FAB, 0x01FA3}, {0x01FAC, 0x01FA4},
-{0x01FAD, 0x01FA5}, {0x01FAE, 0x01FA6}, {0x01FAF, 0x01FA7}, {0x01FB8, 0x01FB0}, {0x01FB9, 0x01FB1}, {0x01FBA, 0x01F70},
-{0x01FBB, 0x01F71}, {0x01FBC, 0x01FB3}, {0x01FC8, 0x01F72}, {0x01FC9, 0x01F73}, {0x01FCA, 0x01F74}, {0x01FCB, 0x01F75},
-{0x01FCC, 0x01FC3}, {0x01FD8, 0x01FD0}, {0x01FD9, 0x01FD1}, {0x01FDA, 0x01F76}, {0x01FDB, 0x01F77}, {0x01FE8, 0x01FE0},
-{0x01FE9, 0x01FE1}, {0x01FEA, 0x01F7A}, {0x01FEB, 0x01F7B}, {0x01FEC, 0x01FE5}, {0x01FF8, 0x01F78}, {0x01FF9, 0x01F79},
-{0x01FFA, 0x01F7C}, {0x01FFB, 0x01F7D}, {0x01FFC, 0x01FF3}, {0x02126, 0x003C9}, {0x0212A, 0x0006B}, {0x0212B, 0x000E5},
-{0x02132, 0x0214E}, {0x02160, 0x02170}, {0x02161, 0x02171}, {0x02162, 0x02172}, {0x02163, 0x02173}, {0x02164, 0x02174},
-{0x02165, 0x02175}, {0x02166, 0x02176}, {0x02167, 0x02177}, {0x02168, 0x02178}, {0x02169, 0x02179}, {0x0216A, 0x0217A},
-{0x0216B, 0x0217B}, {0x0216C, 0x0217C}, {0x0216D, 0x0217D}, {0x0216E, 0x0217E}, {0x0216F, 0x0217F}, {0x02183, 0x02184},
-{0x024B6, 0x024D0}, {0x024B7, 0x024D1}, {0x024B8, 0x024D2}, {0x024B9, 0x024D3}, {0x024BA, 0x024D4}, {0x024BB, 0x024D5},
-{0x024BC, 0x024D6}, {0x024BD, 0x024D7}, {0x024BE, 0x024D8}, {0x024BF, 0x024D9}, {0x024C0, 0x024DA}, {0x024C1, 0x024DB},
-{0x024C2, 0x024DC}, {0x024C3, 0x024DD}, {0x024C4, 0x024DE}, {0x024C5, 0x024DF}, {0x024C6, 0x024E0}, {0x024C7, 0x024E1},
-{0x024C8, 0x024E2}, {0x024C9, 0x024E3}, {0x024CA, 0x024E4}, {0x024CB, 0x024E5}, {0x024CC, 0x024E6}, {0x024CD, 0x024E7},
-{0x024CE, 0x024E8}, {0x024CF, 0x024E9}, {0x02C00, 0x02C30}, {0x02C01, 0x02C31}, {0x02C02, 0x02C32}, {0x02C03, 0x02C33},
-{0x02C04, 0x02C34}, {0x02C05, 0x02C35}, {0x02C06, 0x02C36}, {0x02C07, 0x02C37}, {0x02C08, 0x02C38}, {0x02C09, 0x02C39},
-{0x02C0A, 0x02C3A}, {0x02C0B, 0x02C3B}, {0x02C0C, 0x02C3C}, {0x02C0D, 0x02C3D}, {0x02C0E, 0x02C3E}, {0x02C0F, 0x02C3F},
-{0x02C10, 0x02C40}, {0x02C11, 0x02C41}, {0x02C12, 0x02C42}, {0x02C13, 0x02C43}, {0x02C14, 0x02C44}, {0x02C15, 0x02C45},
-{0x02C16, 0x02C46}, {0x02C17, 0x02C47}, {0x02C18, 0x02C48}, {0x02C19, 0x02C49}, {0x02C1A, 0x02C4A}, {0x02C1B, 0x02C4B},
-{0x02C1C, 0x02C4C}, {0x02C1D, 0x02C4D}, {0x02C1E, 0x02C4E}, {0x02C1F, 0x02C4F}, {0x02C20, 0x02C50}, {0x02C21, 0x02C51},
-{0x02C22, 0x02C52}, {0x02C23, 0x02C53}, {0x02C24, 0x02C54}, {0x02C25, 0x02C55}, {0x02C26, 0x02C56}, {0x02C27, 0x02C57},
-{0x02C28, 0x02C58}, {0x02C29, 0x02C59}, {0x02C2A, 0x02C5A}, {0x02C2B, 0x02C5B}, {0x02C2C, 0x02C5C}, {0x02C2D, 0x02C5D},
-{0x02C2E, 0x02C5E}, {0x02C2F, 0x02C5F}, {0x02C60, 0x02C61}, {0x02C62, 0x0026B}, {0x02C63, 0x01D7D}, {0x02C64, 0x0027D},
-{0x02C67, 0x02C68}, {0x02C69, 0x02C6A}, {0x02C6B, 0x02C6C}, {0x02C6D, 0x00251}, {0x02C6E, 0x00271}, {0x02C6F, 0x00250},
-{0x02C70, 0x00252}, {0x02C72, 0x02C73}, {0x02C75, 0x02C76}, {0x02C7E, 0x0023F}, {0x02C7F, 0x00240}, {0x02C80, 0x02C81},
-{0x02C82, 0x02C83}, {0x02C84, 0x02C85}, {0x02C86, 0x02C87}, {0x02C88, 0x02C89}, {0x02C8A, 0x02C8B}, {0x02C8C, 0x02C8D},
-{0x02C8E, 0x02C8F}, {0x02C90, 0x02C91}, {0x02C92, 0x02C93}, {0x02C94, 0x02C95}, {0x02C96, 0x02C97}, {0x02C98, 0x02C99},
-{0x02C9A, 0x02C9B}, {0x02C9C, 0x02C9D}, {0x02C9E, 0x02C9F}, {0x02CA0, 0x02CA1}, {0x02CA2, 0x02CA3}, {0x02CA4, 0x02CA5},
-{0x02CA6, 0x02CA7}, {0x02CA8, 0x02CA9}, {0x02CAA, 0x02CAB}, {0x02CAC, 0x02CAD}, {0x02CAE, 0x02CAF}, {0x02CB0, 0x02CB1},
-{0x02CB2, 0x02CB3}, {0x02CB4, 0x02CB5}, {0x02CB6, 0x02CB7}, {0x02CB8, 0x02CB9}, {0x02CBA, 0x02CBB}, {0x02CBC, 0x02CBD},
-{0x02CBE, 0x02CBF}, {0x02CC0, 0x02CC1}, {0x02CC2, 0x02CC3}, {0x02CC4, 0x02CC5}, {0x02CC6, 0x02CC7}, {0x02CC8, 0x02CC9},
-{0x02CCA, 0x02CCB}, {0x02CCC, 0x02CCD}, {0x02CCE, 0x02CCF}, {0x02CD0, 0x02CD1}, {0x02CD2, 0x02CD3}, {0x02CD4, 0x02CD5},
-{0x02CD6, 0x02CD7}, {0x02CD8, 0x02CD9}, {0x02CDA, 0x02CDB}, {0x02CDC, 0x02CDD}, {0x02CDE, 0x02CDF}, {0x02CE0, 0x02CE1},
-{0x02CE2, 0x02CE3}, {0x02CEB, 0x02CEC}, {0x02CED, 0x02CEE}, {0x02CF2, 0x02CF3}, {0x0A640, 0x0A641}, {0x0A642, 0x0A643},
-{0x0A644, 0x0A645}, {0x0A646, 0x0A647}, {0x0A648, 0x0A649}, {0x0A64A, 0x0A64B}, {0x0A64C, 0x0A64D}, {0x0A64E, 0x0A64F},
-{0x0A650, 0x0A651}, {0x0A652, 0x0A653}, {0x0A654, 0x0A655}, {0x0A656, 0x0A657}, {0x0A658, 0x0A659}, {0x0A65A, 0x0A65B},
-{0x0A65C, 0x0A65D}, {0x0A65E, 0x0A65F}, {0x0A660, 0x0A661}, {0x0A662, 0x0A663}, {0x0A664, 0x0A665}, {0x0A666, 0x0A667},
-{0x0A668, 0x0A669}, {0x0A66A, 0x0A66B}, {0x0A66C, 0x0A66D}, {0x0A680, 0x0A681}, {0x0A682, 0x0A683}, {0x0A684, 0x0A685},
-{0x0A686, 0x0A687}, {0x0A688, 0x0A689}, {0x0A68A, 0x0A68B}, {0x0A68C, 0x0A68D}, {0x0A68E, 0x0A68F}, {0x0A690, 0x0A691},
-{0x0A692, 0x0A693}, {0x0A694, 0x0A695}, {0x0A696, 0x0A697}, {0x0A698, 0x0A699}, {0x0A69A, 0x0A69B}, {0x0A722, 0x0A723},
-{0x0A724, 0x0A725}, {0x0A726, 0x0A727}, {0x0A728, 0x0A729}, {0x0A72A, 0x0A72B}, {0x0A72C, 0x0A72D}, {0x0A72E, 0x0A72F},
-{0x0A732, 0x0A733}, {0x0A734, 0x0A735}, {0x0A736, 0x0A737}, {0x0A738, 0x0A739}, {0x0A73A, 0x0A73B}, {0x0A73C, 0x0A73D},
-{0x0A73E, 0x0A73F}, {0x0A740, 0x0A741}, {0x0A742, 0x0A743}, {0x0A744, 0x0A745}, {0x0A746, 0x0A747}, {0x0A748, 0x0A749},
-{0x0A74A, 0x0A74B}, {0x0A74C, 0x0A74D}, {0x0A74E, 0x0A74F}, {0x0A750, 0x0A751}, {0x0A752, 0x0A753}, {0x0A754, 0x0A755},
-{0x0A756, 0x0A757}, {0x0A758, 0x0A759}, {0x0A75A, 0x0A75B}, {0x0A75C, 0x0A75D}, {0x0A75E, 0x0A75F}, {0x0A760, 0x0A761},
-{0x0A762, 0x0A763}, {0x0A764, 0x0A765}, {0x0A766, 0x0A767}, {0x0A768, 0x0A769}, {0x0A76A, 0x0A76B}, {0x0A76C, 0x0A76D},
-{0x0A76E, 0x0A76F}, {0x0A779, 0x0A77A}, {0x0A77B, 0x0A77C}, {0x0A77D, 0x01D79}, {0x0A77E, 0x0A77F}, {0x0A780, 0x0A781},
-{0x0A782, 0x0A783}, {0x0A784, 0x0A785}, {0x0A786, 0x0A787}, {0x0A78B, 0x0A78C}, {0x0A78D, 0x00265}, {0x0A790, 0x0A791},
-{0x0A792, 0x0A793}, {0x0A796, 0x0A797}, {0x0A798, 0x0A799}, {0x0A79A, 0x0A79B}, {0x0A79C, 0x0A79D}, {0x0A79E, 0x0A79F},
-{0x0A7A0, 0x0A7A1}, {0x0A7A2, 0x0A7A3}, {0x0A7A4, 0x0A7A5}, {0x0A7A6, 0x0A7A7}, {0x0A7A8, 0x0A7A9}, {0x0A7AA, 0x00266},
-{0x0A7AB, 0x0025C}, {0x0A7AC, 0x00261}, {0x0A7AD, 0x0026C}, {0x0A7AE, 0x0026A}, {0x0A7B0, 0x0029E}, {0x0A7B1, 0x00287},
-{0x0A7B2, 0x0029D}, {0x0A7B3, 0x0AB53}, {0x0A7B4, 0x0A7B5}, {0x0A7B6, 0x0A7B7}, {0x0A7B8, 0x0A7B9}, {0x0A7BA, 0x0A7BB},
-{0x0A7BC, 0x0A7BD}, {0x0A7BE, 0x0A7BF}, {0x0A7C0, 0x0A7C1}, {0x0A7C2, 0x0A7C3}, {0x0A7C4, 0x0A794}, {0x0A7C5, 0x00282},
-{0x0A7C6, 0x01D8E}, {0x0A7C7, 0x0A7C8}, {0x0A7C9, 0x0A7CA}, {0x0A7D0, 0x0A7D1}, {0x0A7D6, 0x0A7D7}, {0x0A7D8, 0x0A7D9},
-{0x0A7F5, 0x0A7F6}, {0x0FF21, 0x0FF41}, {0x0FF22, 0x0FF42}, {0x0FF23, 0x0FF43}, {0x0FF24, 0x0FF44}, {0x0FF25, 0x0FF45},
-{0x0FF26, 0x0FF46}, {0x0FF27, 0x0FF47}, {0x0FF28, 0x0FF48}, {0x0FF29, 0x0FF49}, {0x0FF2A, 0x0FF4A}, {0x0FF2B, 0x0FF4B},
-{0x0FF2C, 0x0FF4C}, {0x0FF2D, 0x0FF4D}, {0x0FF2E, 0x0FF4E}, {0x0FF2F, 0x0FF4F}, {0x0FF30, 0x0FF50}, {0x0FF31, 0x0FF51},
-{0x0FF32, 0x0FF52}, {0x0FF33, 0x0FF53}, {0x0FF34, 0x0FF54}, {0x0FF35, 0x0FF55}, {0x0FF36, 0x0FF56}, {0x0FF37, 0x0FF57},
-{0x0FF38, 0x0FF58}, {0x0FF39, 0x0FF59}, {0x0FF3A, 0x0FF5A}, {0x10400, 0x10428}, {0x10401, 0x10429}, {0x10402, 0x1042A},
-{0x10403, 0x1042B}, {0x10404, 0x1042C}, {0x10405, 0x1042D}, {0x10406, 0x1042E}, {0x10407, 0x1042F}, {0x10408, 0x10430},
-{0x10409, 0x10431}, {0x1040A, 0x10432}, {0x1040B, 0x10433}, {0x1040C, 0x10434}, {0x1040D, 0x10435}, {0x1040E, 0x10436},
-{0x1040F, 0x10437}, {0x10410, 0x10438}, {0x10411, 0x10439}, {0x10412, 0x1043A}, {0x10413, 0x1043B}, {0x10414, 0x1043C},
-{0x10415, 0x1043D}, {0x10416, 0x1043E}, {0x10417, 0x1043F}, {0x10418, 0x10440}, {0x10419, 0x10441}, {0x1041A, 0x10442},
-{0x1041B, 0x10443}, {0x1041C, 0x10444}, {0x1041D, 0x10445}, {0x1041E, 0x10446}, {0x1041F, 0x10447}, {0x10420, 0x10448},
-{0x10421, 0x10449}, {0x10422, 0x1044A}, {0x10423, 0x1044B}, {0x10424, 0x1044C}, {0x10425, 0x1044D}, {0x10426, 0x1044E},
-{0x10427, 0x1044F}, {0x104B0, 0x104D8}, {0x104B1, 0x104D9}, {0x104B2, 0x104DA}, {0x104B3, 0x104DB}, {0x104B4, 0x104DC},
-{0x104B5, 0x104DD}, {0x104B6, 0x104DE}, {0x104B7, 0x104DF}, {0x104B8, 0x104E0}, {0x104B9, 0x104E1}, {0x104BA, 0x104E2},
-{0x104BB, 0x104E3}, {0x104BC, 0x104E4}, {0x104BD, 0x104E5}, {0x104BE, 0x104E6}, {0x104BF, 0x104E7}, {0x104C0, 0x104E8},
-{0x104C1, 0x104E9}, {0x104C2, 0x104EA}, {0x104C3, 0x104EB}, {0x104C4, 0x104EC}, {0x104C5, 0x104ED}, {0x104C6, 0x104EE},
-{0x104C7, 0x104EF}, {0x104C8, 0x104F0}, {0x104C9, 0x104F1}, {0x104CA, 0x104F2}, {0x104CB, 0x104F3}, {0x104CC, 0x104F4},
-{0x104CD, 0x104F5}, {0x104CE, 0x104F6}, {0x104CF, 0x104F7}, {0x104D0, 0x104F8}, {0x104D1, 0x104F9}, {0x104D2, 0x104FA},
-{0x104D3, 0x104FB}, {0x10570, 0x10597}, {0x10571, 0x10598}, {0x10572, 0x10599}, {0x10573, 0x1059A}, {0x10574, 0x1059B},
-{0x10575, 0x1059C}, {0x10576, 0x1059D}, {0x10577, 0x1059E}, {0x10578, 0x1059F}, {0x10579, 0x105A0}, {0x1057A, 0x105A1},
-{0x1057C, 0x105A3}, {0x1057D, 0x105A4}, {0x1057E, 0x105A5}, {0x1057F, 0x105A6}, {0x10580, 0x105A7}, {0x10581, 0x105A8},
-{0x10582, 0x105A9}, {0x10583, 0x105AA}, {0x10584, 0x105AB}, {0x10585, 0x105AC}, {0x10586, 0x105AD}, {0x10587, 0x105AE},
-{0x10588, 0x105AF}, {0x10589, 0x105B0}, {0x1058A, 0x105B1}, {0x1058C, 0x105B3}, {0x1058D, 0x105B4}, {0x1058E, 0x105B5},
-{0x1058F, 0x105B6}, {0x10590, 0x105B7}, {0x10591, 0x105B8}, {0x10592, 0x105B9}, {0x10594, 0x105BB}, {0x10595, 0x105BC},
-{0x10C80, 0x10CC0}, {0x10C81, 0x10CC1}, {0x10C82, 0x10CC2}, {0x10C83, 0x10CC3}, {0x10C84, 0x10CC4}, {0x10C85, 0x10CC5},
-{0x10C86, 0x10CC6}, {0x10C87, 0x10CC7}, {0x10C88, 0x10CC8}, {0x10C89, 0x10CC9}, {0x10C8A, 0x10CCA}, {0x10C8B, 0x10CCB},
-{0x10C8C, 0x10CCC}, {0x10C8D, 0x10CCD}, {0x10C8E, 0x10CCE}, {0x10C8F, 0x10CCF}, {0x10C90, 0x10CD0}, {0x10C91, 0x10CD1},
-{0x10C92, 0x10CD2}, {0x10C93, 0x10CD3}, {0x10C94, 0x10CD4}, {0x10C95, 0x10CD5}, {0x10C96, 0x10CD6}, {0x10C97, 0x10CD7},
-{0x10C98, 0x10CD8}, {0x10C99, 0x10CD9}, {0x10C9A, 0x10CDA}, {0x10C9B, 0x10CDB}, {0x10C9C, 0x10CDC}, {0x10C9D, 0x10CDD},
-{0x10C9E, 0x10CDE}, {0x10C9F, 0x10CDF}, {0x10CA0, 0x10CE0}, {0x10CA1, 0x10CE1}, {0x10CA2, 0x10CE2}, {0x10CA3, 0x10CE3},
-{0x10CA4, 0x10CE4}, {0x10CA5, 0x10CE5}, {0x10CA6, 0x10CE6}, {0x10CA7, 0x10CE7}, {0x10CA8, 0x10CE8}, {0x10CA9, 0x10CE9},
-{0x10CAA, 0x10CEA}, {0x10CAB, 0x10CEB}, {0x10CAC, 0x10CEC}, {0x10CAD, 0x10CED}, {0x10CAE, 0x10CEE}, {0x10CAF, 0x10CEF},
-{0x10CB0, 0x10CF0}, {0x10CB1, 0x10CF1}, {0x10CB2, 0x10CF2}, {0x118A0, 0x118C0}, {0x118A1, 0x118C1}, {0x118A2, 0x118C2},
-{0x118A3, 0x118C3}, {0x118A4, 0x118C4}, {0x118A5, 0x118C5}, {0x118A6, 0x118C6}, {0x118A7, 0x118C7}, {0x118A8, 0x118C8},
-{0x118A9, 0x118C9}, {0x118AA, 0x118CA}, {0x118AB, 0x118CB}, {0x118AC, 0x118CC}, {0x118AD, 0x118CD}, {0x118AE, 0x118CE},
-{0x118AF, 0x118CF}, {0x118B0, 0x118D0}, {0x118B1, 0x118D1}, {0x118B2, 0x118D2}, {0x118B3, 0x118D3}, {0x118B4, 0x118D4},
-{0x118B5, 0x118D5}, {0x118B6, 0x118D6}, {0x118B7, 0x118D7}, {0x118B8, 0x118D8}, {0x118B9, 0x118D9}, {0x118BA, 0x118DA},
-{0x118BB, 0x118DB}, {0x118BC, 0x118DC}, {0x118BD, 0x118DD}, {0x118BE, 0x118DE}, {0x118BF, 0x118DF}, {0x16E40, 0x16E60},
-{0x16E41, 0x16E61}, {0x16E42, 0x16E62}, {0x16E43, 0x16E63}, {0x16E44, 0x16E64}, {0x16E45, 0x16E65}, {0x16E46, 0x16E66},
-{0x16E47, 0x16E67}, {0x16E48, 0x16E68}, {0x16E49, 0x16E69}, {0x16E4A, 0x16E6A}, {0x16E4B, 0x16E6B}, {0x16E4C, 0x16E6C},
-{0x16E4D, 0x16E6D}, {0x16E4E, 0x16E6E}, {0x16E4F, 0x16E6F}, {0x16E50, 0x16E70}, {0x16E51, 0x16E71}, {0x16E52, 0x16E72},
-{0x16E53, 0x16E73}, {0x16E54, 0x16E74}, {0x16E55, 0x16E75}, {0x16E56, 0x16E76}, {0x16E57, 0x16E77}, {0x16E58, 0x16E78},
-{0x16E59, 0x16E79}, {0x16E5A, 0x16E7A}, {0x16E5B, 0x16E7B}, {0x16E5C, 0x16E7C}, {0x16E5D, 0x16E7D}, {0x16E5E, 0x16E7E},
-{0x16E5F, 0x16E7F}, {0x1E900, 0x1E922}, {0x1E901, 0x1E923}, {0x1E902, 0x1E924}, {0x1E903, 0x1E925}, {0x1E904, 0x1E926},
-{0x1E905, 0x1E927}, {0x1E906, 0x1E928}, {0x1E907, 0x1E929}, {0x1E908, 0x1E92A}, {0x1E909, 0x1E92B}, {0x1E90A, 0x1E92C},
-{0x1E90B, 0x1E92D}, {0x1E90C, 0x1E92E}, {0x1E90D, 0x1E92F}, {0x1E90E, 0x1E930}, {0x1E90F, 0x1E931}, {0x1E910, 0x1E932},
-{0x1E911, 0x1E933}, {0x1E912, 0x1E934}, {0x1E913, 0x1E935}, {0x1E914, 0x1E936}, {0x1E915, 0x1E937}, {0x1E916, 0x1E938},
-{0x1E917, 0x1E939}, {0x1E918, 0x1E93A}, {0x1E919, 0x1E93B}, {0x1E91A, 0x1E93C}, {0x1E91B, 0x1E93D}, {0x1E91C, 0x1E93E},
-{0x1E91D, 0x1E93F}, {0x1E91E, 0x1E940}, {0x1E91F, 0x1E941}, {0x1E920, 0x1E942}, {0x1E921, 0x1E943},
-};

llama/unicode-data.h

@@ -1,16 +1,20 @@
 #pragma once
 
 #include <cstdint>
-#include <map>
-#include <utility>
 #include <vector>
+#include <unordered_map>
+#include <unordered_set>
 
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_digit;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control;
-extern const std::multimap<uint32_t, uint32_t> unicode_map_nfd;
-extern const std::map<char32_t, char32_t> unicode_map_lowercase;
+struct range_nfd {
+    uint32_t first;
+    uint32_t last;
+    uint32_t nfd;
+};
+
+static const uint32_t MAX_CODEPOINTS = 0x110000;
+
+extern const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags;
+extern const std::unordered_set<uint32_t> unicode_set_whitespace;
+extern const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase;
+extern const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase;
+extern const std::vector<range_nfd> unicode_ranges_nfd;