llama: add clip dependencies

llama/clip.cpp

@@ -0,0 +1,2104 @@
+/**
+ * llama.cpp - git 059031b8c40e1f4ba60586842c5b1ed3ddf61842
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// NOTE: This is modified from clip.cpp only for LLaVA,
+// so there might be still unnecessary artifacts hanging around
+// I'll gradually clean and extend it
+// Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
+#include "clip.h"
+#include "log.h"
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb_image.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <regex>
+#include <stdexcept>
+#include <vector>
+#include <sstream>
+#include <cinttypes>
+#include <limits>
+
+//#define CLIP_DEBUG_FUNCTIONS
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+static std::string format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), buf.size());
+}
+
+//
+// key constants
+//
+
+#define KEY_FTYPE          "general.file_type"
+#define KEY_NAME           "general.name"
+#define KEY_DESCRIPTION    "general.description"
+#define KEY_HAS_TEXT_ENC   "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC    "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
+#define KEY_USE_GELU       "clip.use_gelu"
+#define KEY_N_EMBD         "clip.%s.embedding_length"
+#define KEY_N_FF           "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK        "clip.%s.block_count"
+#define KEY_N_HEAD         "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM       "clip.%s.projection_dim"
+#define KEY_TOKENS         "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS    "clip.text.context_length"
+#define KEY_IMAGE_SIZE     "clip.vision.image_size"
+#define KEY_PATCH_SIZE     "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN     "clip.vision.image_mean"
+#define KEY_IMAGE_STD      "clip.vision.image_std"
+#define KEY_PROJ_TYPE      "clip.projector_type"
+
+#define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
+#define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
+#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
+
+
+//
+// tensor name constants
+//
+
+#define TN_TOKEN_EMBD      "%s.token_embd.weight"
+#define TN_POS_EMBD        "%s.position_embd.weight"
+#define TN_CLASS_EMBD      "v.class_embd"
+#define TN_PATCH_EMBD      "v.patch_embd.weight"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
+#define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
+#define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
+#define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
+#define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
+#define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_LN_1            "%s.blk.%d.ln1.%s"
+#define TN_LN_2            "%s.blk.%d.ln2.%s"
+#define TN_LN_PRE          "%s.pre_ln.%s"
+#define TN_LN_POST         "%s.post_ln.%s"
+#define TN_TEXT_PROJ       "text_projection.weight"
+#define TN_VIS_PROJ        "visual_projection.weight"
+#define TN_LLAVA_PROJ      "mm.%d.%s"
+#define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
+#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
+#define TN_IMAGE_NEWLINE   "model.image_newline"
+
+
+enum projector_type {
+    PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
+    PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_UNKNOWN,
+};
+
+static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
+    { PROJECTOR_TYPE_MLP, "mlp" },
+    { PROJECTOR_TYPE_LDP, "ldp" },
+    { PROJECTOR_TYPE_LDPV2, "ldpv2"},
+};
+
+
+//
+// utilities to get data from a gguf file
+//
+
+static int get_key_idx(const gguf_context * ctx, const char * key) {
+    int i = gguf_find_key(ctx, key);
+    if (i == -1) {
+        LOG_TEE("key %s not found in file\n", key);
+        throw std::runtime_error(format("Missing required key: %s", key));
+    }
+
+    return i;
+}
+
+static uint32_t get_u32(const gguf_context * ctx, const std::string & key) {
+    const int i = get_key_idx(ctx, key.c_str());
+
+    return gguf_get_val_u32(ctx, i);
+}
+
+static float get_f32(const gguf_context * ctx, const std::string & key) {
+    const int i = get_key_idx(ctx, key.c_str());
+
+    return gguf_get_val_f32(ctx, i);
+}
+
+static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::string & name) {
+    struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
+    if (!cur) {
+        throw std::runtime_error(format("%s: unable to find tensor %s\n", __func__, name.c_str()));
+    }
+
+    return cur;
+}
+
+static std::string get_ftype(int ftype) {
+    return ggml_type_name(static_cast<ggml_type>(ftype));
+}
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    std::string result;
+    for (size_t pos = 0; ; pos += search.length()) {
+        auto new_pos = s.find(search, pos);
+        if (new_pos == std::string::npos) {
+            result += s.substr(pos, s.size() - pos);
+            break;
+        }
+        result += s.substr(pos, new_pos - pos) + replace;
+        pos = new_pos;
+    }
+    s = std::move(result);
+}
+
+static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}
+
+static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") {
+    size_t tensor_size = ggml_nbytes(tensor);
+    LOG_TEE("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
+            prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
+            tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
+}
+
+static projector_type clip_projector_type_from_string(const std::string & name) {
+    for (const auto & kv : PROJECTOR_TYPE_NAMES) { // NOLINT
+        if (kv.second == name) {
+            return kv.first;
+        }
+    }
+    return PROJECTOR_TYPE_UNKNOWN;
+}
+
+#ifdef CLIP_DEBUG_FUNCTIONS
+static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
+    std::ofstream file(filename, std::ios::binary);
+    if (!file.is_open()) {
+        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        return;
+    }
+
+    // PPM header: P6 format, width, height, and max color value
+    file << "P6\n" << img.nx << " " << img.ny << "\n255\n";
+
+    // Write pixel data
+    for (size_t i = 0; i < img.buf.size(); i += 3) {
+        // PPM expects binary data in RGB format, which matches our image buffer
+        file.write(reinterpret_cast<const char*>(&img.buf[i]), 3);
+    }
+
+    file.close();
+}
+
+static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
+    std::ofstream file(filename, std::ios::binary);
+    if (!file.is_open()) {
+        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        return;
+    }
+
+    int fileSize = 54 + 3 * img.nx * img.ny; // File header + info header + pixel data
+    int bytesPerPixel = 3;
+    int widthInBytes = img.nx * bytesPerPixel;
+    int paddingAmount = (4 - (widthInBytes % 4)) % 4;
+    int stride = widthInBytes + paddingAmount;
+
+    // Bitmap file header
+    unsigned char fileHeader[14] = {
+        'B','M',     // Signature
+        0,0,0,0,    // Image file size in bytes
+        0,0,0,0,    // Reserved
+        54,0,0,0    // Start of pixel array
+    };
+
+    // Total file size
+    fileSize = 54 + (stride * img.ny);
+    fileHeader[2] = (unsigned char)(fileSize);
+    fileHeader[3] = (unsigned char)(fileSize >> 8);
+    fileHeader[4] = (unsigned char)(fileSize >> 16);
+    fileHeader[5] = (unsigned char)(fileSize >> 24);
+
+    // Bitmap information header (BITMAPINFOHEADER)
+    unsigned char infoHeader[40] = {
+        40,0,0,0,   // Size of this header (40 bytes)
+        0,0,0,0,    // Image width
+        0,0,0,0,    // Image height
+        1,0,        // Number of color planes
+        24,0,       // Bits per pixel
+        0,0,0,0,    // No compression
+        0,0,0,0,    // Image size (can be 0 for no compression)
+        0,0,0,0,    // X pixels per meter (not specified)
+        0,0,0,0,    // Y pixels per meter (not specified)
+        0,0,0,0,    // Total colors (color table not used)
+        0,0,0,0     // Important colors (all are important)
+    };
+
+    // Width and height in the information header
+    infoHeader[4] = (unsigned char)(img.nx);
+    infoHeader[5] = (unsigned char)(img.nx >> 8);
+    infoHeader[6] = (unsigned char)(img.nx >> 16);
+    infoHeader[7] = (unsigned char)(img.nx >> 24);
+    infoHeader[8] = (unsigned char)(img.ny);
+    infoHeader[9] = (unsigned char)(img.ny >> 8);
+    infoHeader[10] = (unsigned char)(img.ny >> 16);
+    infoHeader[11] = (unsigned char)(img.ny >> 24);
+
+    // Write file headers
+    file.write(reinterpret_cast<char*>(fileHeader), sizeof(fileHeader));
+    file.write(reinterpret_cast<char*>(infoHeader), sizeof(infoHeader));
+
+    // Pixel data
+    std::vector<unsigned char> padding(3, 0); // Max padding size to be added to each row
+    for (int y = img.ny - 1; y >= 0; --y) { // BMP files are stored bottom-to-top
+        for (int x = 0; x < img.nx; ++x) {
+            // Each pixel
+            size_t pixelIndex = (y * img.nx + x) * 3;
+            unsigned char pixel[3] = {
+                img.buf[pixelIndex + 2], // BMP stores pixels in BGR format
+                img.buf[pixelIndex + 1],
+                img.buf[pixelIndex]
+            };
+            file.write(reinterpret_cast<char*>(pixel), 3);
+        }
+        // Write padding for the row
+        file.write(reinterpret_cast<char*>(padding.data()), paddingAmount);
+    }
+
+    file.close();
+}
+
+// debug function to convert f32 to u8
+static void clip_image_convert_f32_to_u8(const clip_image_f32& src, clip_image_u8& dst) {
+    dst.nx = src.nx;
+    dst.ny = src.ny;
+    dst.buf.resize(3 * src.nx * src.ny);
+    for (size_t i = 0; i < src.buf.size(); ++i) {
+        dst.buf[i] = static_cast<uint8_t>(std::min(std::max(int(src.buf[i] * 255.0f), 0), 255));
+    }
+}
+#endif
+
+
+//
+// clip layers
+//
+
+struct clip_hparams {
+    int32_t image_size;
+    int32_t patch_size;
+    int32_t hidden_size;
+    int32_t n_intermediate;
+    int32_t projection_dim;
+    int32_t n_head;
+    int32_t n_layer;
+
+    float eps;
+
+    char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default)
+
+    int32_t image_grid_pinpoints[32];
+    int32_t image_crop_resolution;
+};
+
+struct clip_layer {
+    // attention
+    struct ggml_tensor * k_w;
+    struct ggml_tensor * k_b;
+    struct ggml_tensor * q_w;
+    struct ggml_tensor * q_b;
+    struct ggml_tensor * v_w;
+    struct ggml_tensor * v_b;
+
+    struct ggml_tensor * o_w;
+    struct ggml_tensor * o_b;
+
+    // layernorm 1
+    struct ggml_tensor * ln_1_w;
+    struct ggml_tensor * ln_1_b;
+
+    // ff
+    struct ggml_tensor * ff_i_w;
+    struct ggml_tensor * ff_i_b;
+
+    struct ggml_tensor * ff_o_w;
+    struct ggml_tensor * ff_o_b;
+
+    // layernorm 2
+    struct ggml_tensor * ln_2_w;
+    struct ggml_tensor * ln_2_b;
+};
+
+struct clip_vision_model {
+    struct clip_hparams hparams;
+
+    // embeddings
+    struct ggml_tensor * class_embedding;
+    struct ggml_tensor * patch_embeddings;
+    struct ggml_tensor * patch_bias;
+    struct ggml_tensor * position_embeddings;
+
+    struct ggml_tensor * pre_ln_w;
+    struct ggml_tensor * pre_ln_b;
+
+    std::vector<clip_layer> layers;
+
+    struct ggml_tensor * post_ln_w;
+    struct ggml_tensor * post_ln_b;
+
+    struct ggml_tensor * projection;
+
+    // LLaVA projection
+    struct ggml_tensor * mm_0_w = NULL;
+    struct ggml_tensor * mm_0_b = NULL;
+    struct ggml_tensor * mm_2_w = NULL;
+    struct ggml_tensor * mm_2_b = NULL;
+
+    struct ggml_tensor * image_newline = NULL;
+
+    // Yi type models with mlp+normalization projection
+    struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4
+    struct ggml_tensor * mm_1_b = NULL;
+    struct ggml_tensor * mm_3_w = NULL;
+    struct ggml_tensor * mm_3_b = NULL;
+    struct ggml_tensor * mm_4_w = NULL;
+    struct ggml_tensor * mm_4_b = NULL;
+
+    // MobileVLM projection
+    struct ggml_tensor * mm_model_mlp_1_w;
+    struct ggml_tensor * mm_model_mlp_1_b;
+    struct ggml_tensor * mm_model_mlp_3_w;
+    struct ggml_tensor * mm_model_mlp_3_b;
+    struct ggml_tensor * mm_model_block_1_block_0_0_w;
+    struct ggml_tensor * mm_model_block_1_block_0_1_w;
+    struct ggml_tensor * mm_model_block_1_block_0_1_b;
+    struct ggml_tensor * mm_model_block_1_block_1_fc1_w;
+    struct ggml_tensor * mm_model_block_1_block_1_fc1_b;
+    struct ggml_tensor * mm_model_block_1_block_1_fc2_w;
+    struct ggml_tensor * mm_model_block_1_block_1_fc2_b;
+    struct ggml_tensor * mm_model_block_1_block_2_0_w;
+    struct ggml_tensor * mm_model_block_1_block_2_1_w;
+    struct ggml_tensor * mm_model_block_1_block_2_1_b;
+    struct ggml_tensor * mm_model_block_2_block_0_0_w;
+    struct ggml_tensor * mm_model_block_2_block_0_1_w;
+    struct ggml_tensor * mm_model_block_2_block_0_1_b;
+    struct ggml_tensor * mm_model_block_2_block_1_fc1_w;
+    struct ggml_tensor * mm_model_block_2_block_1_fc1_b;
+    struct ggml_tensor * mm_model_block_2_block_1_fc2_w;
+    struct ggml_tensor * mm_model_block_2_block_1_fc2_b;
+    struct ggml_tensor * mm_model_block_2_block_2_0_w;
+    struct ggml_tensor * mm_model_block_2_block_2_1_w;
+    struct ggml_tensor * mm_model_block_2_block_2_1_b;
+
+    // MobileVLM_V2 projection
+    struct ggml_tensor * mm_model_mlp_0_w;
+    struct ggml_tensor * mm_model_mlp_0_b;
+    struct ggml_tensor * mm_model_mlp_2_w;
+    struct ggml_tensor * mm_model_mlp_2_b;
+    struct ggml_tensor * mm_model_peg_0_w;
+    struct ggml_tensor * mm_model_peg_0_b;
+};
+
+struct clip_ctx {
+    bool has_text_encoder    = false;
+    bool has_vision_encoder  = false;
+    bool has_llava_projector = false;
+
+    struct clip_vision_model vision_model;
+    projector_type proj_type = PROJECTOR_TYPE_MLP;
+
+    float image_mean[3];
+    float image_std[3];
+    bool use_gelu = false;
+    int32_t ftype = 1;
+
+    bool has_class_embedding = true;
+    bool has_pre_norm = true;
+    bool has_post_norm = false;
+    bool has_patch_bias = false;
+
+    struct gguf_context * ctx_gguf;
+    struct ggml_context * ctx_data;
+
+    std::vector<uint8_t> buf_compute_meta;
+
+    // memory buffers to evaluate the model
+    ggml_backend_buffer_t params_buffer  = NULL;
+
+    ggml_backend_t backend       = NULL;
+    ggml_gallocr_t compute_alloc = NULL;
+};
+
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return nullptr;
+    }
+
+    const auto & model = ctx->vision_model;
+    const auto & hparams = model.hparams;
+
+    const int image_size           = hparams.image_size;
+    const int patch_size           = hparams.patch_size;
+    const int num_patches          = ((image_size / patch_size) * (image_size / patch_size));
+    const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
+    const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    const int hidden_size          = hparams.hidden_size;
+    const int n_head               = hparams.n_head;
+    const int d_head               = hidden_size / n_head;
+    const int n_layer              = hparams.n_layer;
+    const float eps                = hparams.eps;
+
+    const int batch_size = imgs->size;
+
+    if (ctx->has_llava_projector) {
+        GGML_ASSERT(batch_size == 1);
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ ctx->buf_compute_meta.size(),
+        /*.mem_buffer =*/ ctx->buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
+    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+
+    if (ctx->has_patch_bias) {
+        // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+    }
+
+    // concat class_embeddings and patch_embeddings
+    struct ggml_tensor * embeddings = inp;
+    if (ctx->has_class_embedding) {
+        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+        ggml_set_name(embeddings, "embeddings");
+        ggml_set_input(embeddings);
+        embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+        embeddings = ggml_acc(ctx0, embeddings, inp,
+                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    }
+
+
+    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    embeddings =
+        ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
+
+    // pre-layernorm
+    if (ctx->has_pre_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "pre_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
+    }
+
+    // loop over layers
+    for (int il = 0; il < n_layer - 1; il++) {
+        struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
+
+        //const size_t nb_q_w = model.layers[il].q_w->nb[0];
+
+        // layernorm1
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w),
+                           model.layers[il].ln_1_b);
+        }
+
+        // self-attention
+        {
+
+            struct ggml_tensor * Q =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
+
+            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+            Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
+            Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+            Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
+
+            struct ggml_tensor * K =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
+
+            K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+            K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+            K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
+
+            struct ggml_tensor * V =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b);
+
+            V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
+            V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+            V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            KQ = ggml_soft_max_inplace(ctx0, KQ);
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+            KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size);
+            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size);
+        }
+
+        // attention output
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b);
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, embeddings);
+
+        embeddings = cur; // embeddings = residual, cur = hidden_states
+
+        // layernorm2
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b);
+        }
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b);
+
+        if (ctx->use_gelu) {
+            cur = ggml_gelu_inplace(ctx0, cur);
+        } else {
+            cur = ggml_gelu_quick_inplace(ctx0, cur);
+        }
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b);
+
+        // residual 2
+        cur = ggml_add(ctx0, embeddings, cur);
+
+        embeddings = cur;
+    }
+
+    // post-layernorm
+    if (ctx->has_post_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "post_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
+    }
+
+    // llava projector
+    {
+        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
+
+        struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
+        ggml_set_name(patches, "patches");
+        ggml_set_input(patches);
+
+        // shape [1, 576, 1024]
+        // ne is whcn, ne = [1024, 576, 1, 1]
+        embeddings = ggml_get_rows(ctx0, embeddings, patches);
+
+        // print_tensor_info(embeddings, "embeddings");
+
+        // llava projector
+        if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+
+            embeddings = ggml_gelu(ctx0, embeddings);
+            embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+
+        } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projector
+            int n_patch = 24;
+            struct ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings);
+            mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b);
+            mlp_1 = ggml_gelu(ctx0, mlp_1);
+            struct ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1);
+            mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b);
+            // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1]
+
+            // block 1
+            struct ggml_tensor * block_1 = nullptr;
+            {
+                // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
+                mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
+                mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
+                // stride = 1, padding = 1, bias is nullptr
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
+
+                // layer norm
+                // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+
+                // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // hardswish
+                struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+                // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // residual
+                block_1 = ggml_add(ctx0, mlp_3, block_1);
+            }
+
+            // block_2
+            {
+                // stride = 2
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
+
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // layer norm
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // hardswish
+                struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                // not sure the parameters is right for globalAvgPooling
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+
+                // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b);
+                block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]);
+                // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1]
+            }
+            embeddings = block_1;
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            struct ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            struct ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
+            // mlp_2 ne = [2048, 576, 1, 1]
+            // // AVG Pool Layer 2*2, strides = 2
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
+            // mlp_2 ne = [576, 2048, 1, 1]
+            mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+            // mlp_2 ne [24, 24, 2048, 1]
+            mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
+            // weight ne = [3, 3, 2048, 1]
+            struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
+        else {
+            GGML_ASSERT(false);
+        }
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+// read and create ggml_context containing the tensors and their data
+struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+    struct ggml_context * meta = NULL;
+
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ &meta,
+    };
+
+    struct gguf_context * ctx = gguf_init_from_file(fname, params);
+    if (!ctx) {
+        throw std::runtime_error(format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname));
+    }
+
+    if (verbosity >= 1) {
+        const int n_tensors = gguf_get_n_tensors(ctx);
+        const int n_kv = gguf_get_n_kv(ctx);
+        const int ftype = get_u32(ctx, KEY_FTYPE);
+        const std::string ftype_str = get_ftype(ftype);
+        const int idx_desc = get_key_idx(ctx, KEY_DESCRIPTION);
+        const std::string description = gguf_get_val_str(ctx, idx_desc);
+        const int idx_name = gguf_find_key(ctx, KEY_NAME);
+        if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
+            const std::string name = gguf_get_val_str(ctx, idx_name);
+            LOG_TEE("%s: model name:   %s\n", __func__, name.c_str());
+        }
+        LOG_TEE("%s: description:  %s\n", __func__, description.c_str());
+        LOG_TEE("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
+        LOG_TEE("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
+        LOG_TEE("%s: n_tensors:    %d\n", __func__, n_tensors);
+        LOG_TEE("%s: n_kv:         %d\n", __func__, n_kv);
+        LOG_TEE("%s: ftype:        %s\n", __func__, ftype_str.c_str());
+        LOG_TEE("\n");
+    }
+    const int n_tensors = gguf_get_n_tensors(ctx);
+
+    // kv
+    const int n_kv = gguf_get_n_kv(ctx);
+    LOG_TEE("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
+        __func__, n_kv, n_tensors, fname);
+    {
+        std::map<enum ggml_type, uint32_t> n_type;
+
+        for (int i = 0; i < n_tensors; i++) {
+            enum ggml_type type = gguf_get_tensor_type(ctx, i);
+
+            n_type[type]++;
+        }
+
+        LOG_TEE("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
+        for (int i = 0; i < n_kv; i++) {
+            const char * name           = gguf_get_key(ctx, i);
+            const enum gguf_type type   = gguf_get_kv_type(ctx, i);
+            const std::string type_name =
+                type == GGUF_TYPE_ARRAY
+                ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx, i)), gguf_get_arr_n(ctx, i))
+                : gguf_type_name(type);
+
+            std::string value          = gguf_kv_to_str(ctx, i);
+            const size_t MAX_VALUE_LEN = 40;
+            if (value.size() > MAX_VALUE_LEN) {
+                value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str());
+            }
+            replace_all(value, "\n", "\\n");
+
+            LOG_TEE("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
+        }
+
+        // print type counts
+        for (auto & kv : n_type) {
+            if (kv.second == 0) {
+                continue;
+            }
+
+            LOG_TEE("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
+        }
+    }
+
+    // data
+    size_t model_size = 0;
+    {
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            const size_t offset = gguf_get_tensor_offset(ctx, i);
+            enum ggml_type type = gguf_get_tensor_type(ctx, i);
+            struct ggml_tensor * cur = ggml_get_tensor(meta, name);
+            size_t tensor_size = ggml_nbytes(cur);
+            model_size += tensor_size;
+            if (verbosity >= 3) {
+                LOG_TEE("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
+                       __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
+            }
+        }
+    }
+
+    clip_ctx * new_clip = new clip_ctx;
+
+    // update projector type
+    {
+        int idx = gguf_find_key(ctx, KEY_PROJ_TYPE);
+        if (idx != -1) {
+            const std::string proj_type = gguf_get_val_str(ctx, idx);
+            new_clip->proj_type = clip_projector_type_from_string(proj_type);
+        } else {
+            new_clip->proj_type = PROJECTOR_TYPE_MLP;
+        }
+
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
+            if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) {
+                new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM;
+            }
+        }
+    }
+
+#ifdef GGML_USE_CUDA
+    new_clip->backend = ggml_backend_cuda_init(0);
+    LOG_TEE("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+    new_clip->backend = ggml_backend_metal_init();
+    LOG_TEE("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+
+    if (!new_clip->backend) {
+        new_clip->backend = ggml_backend_cpu_init();
+        LOG_TEE("%s: CLIP using CPU backend\n", __func__);
+    }
+
+    // model size and capabilities
+    {
+        int idx = get_key_idx(ctx, KEY_HAS_TEXT_ENC);
+        new_clip->has_text_encoder = gguf_get_val_bool(ctx, idx);
+
+        idx = get_key_idx(ctx, KEY_HAS_VIS_ENC);
+        new_clip->has_vision_encoder = gguf_get_val_bool(ctx, idx);
+
+        idx = gguf_find_key(ctx, KEY_HAS_LLAVA_PROJ);
+        if (idx != -1) {
+            new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx);
+        }
+
+        GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+        GGML_ASSERT(new_clip->has_vision_encoder);
+        GGML_ASSERT(!new_clip->has_text_encoder);
+
+        idx = get_key_idx(ctx, KEY_USE_GELU);
+        new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
+
+        if (verbosity >= 1) {
+            LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
+            LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
+            LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
+            LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
+        }
+    }
+
+    LOG_TEE("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
+
+    // load tensors
+    {
+        std::vector<uint8_t> read_buf;
+        struct ggml_init_params params = {
+            /*.mem_size =*/ (n_tensors + 1) * ggml_tensor_overhead(),
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc =*/ true,
+        };
+
+        new_clip->ctx_data = ggml_init(params);
+        if (!new_clip->ctx_data) {
+            LOG_TEE("%s: ggml_init() failed\n", __func__);
+            clip_free(new_clip);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+        auto fin = std::ifstream(fname, std::ios::binary);
+        if (!fin) {
+            LOG_TEE("cannot open model file for loading tensors\n");
+            clip_free(new_clip);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+        // add tensors to context
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor * t = ggml_get_tensor(meta, name);
+            struct ggml_tensor * cur = ggml_dup_tensor(new_clip->ctx_data, t);
+            ggml_set_name(cur, name);
+        }
+
+        // alloc memory and offload data
+        new_clip->params_buffer = ggml_backend_alloc_ctx_tensors(new_clip->ctx_data, new_clip->backend);
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name);
+            const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
+            fin.seekg(offset, std::ios::beg);
+            if (!fin) {
+                LOG_TEE("%s: failed to seek for tensor %s\n", __func__, name);
+                clip_free(new_clip);
+                gguf_free(ctx);
+                return nullptr;
+            }
+            int num_bytes = ggml_nbytes(cur);
+            if (ggml_backend_buffer_is_host(new_clip->params_buffer)) {
+                // for the CPU and Metal backend, we can read directly into the tensor
+                fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+            } else {
+                // read into a temporary buffer first, then copy to device memory
+                read_buf.resize(num_bytes);
+                fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            }
+        }
+        fin.close();
+    }
+
+    // vision model
+    if (new_clip->has_vision_encoder) {
+        // load vision model
+        auto & vision_model = new_clip->vision_model;
+        auto & hparams = vision_model.hparams;
+        hparams.hidden_size    = get_u32(ctx, format(KEY_N_EMBD, "vision"));
+        hparams.n_head         = get_u32(ctx, format(KEY_N_HEAD, "vision"));
+        hparams.n_intermediate = get_u32(ctx, format(KEY_N_FF, "vision"));
+        hparams.n_layer        = get_u32(ctx, format(KEY_N_BLOCK, "vision"));
+        hparams.image_size     = get_u32(ctx, KEY_IMAGE_SIZE);
+        hparams.patch_size     = get_u32(ctx, KEY_PATCH_SIZE);
+        hparams.projection_dim = get_u32(ctx, format(KEY_PROJ_DIM, "vision"));
+        hparams.eps            = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision"));
+
+        try {
+            int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS);
+            int n = gguf_get_arr_n(ctx, idx);
+            const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx);
+            for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) {
+                hparams.image_grid_pinpoints[i] = pinpoints[i];
+            }
+            if (n < 32)
+                hparams.image_grid_pinpoints[n] = 0;
+        } catch (std::runtime_error & e) {
+            hparams.image_grid_pinpoints[0]=0;
+        }
+
+        try {
+            int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE);
+            strcpy(hparams.mm_patch_merge_type, gguf_get_val_str(ctx, idx));
+        } catch (std::runtime_error & e) {
+            strcpy(hparams.mm_patch_merge_type, "flat");
+        }
+
+        try {
+            hparams.image_crop_resolution = get_u32(ctx, KEY_IMAGE_CROP_RESOLUTION); // llava-1.6
+        } catch(const std::exception& e) {
+            hparams.image_crop_resolution = hparams.image_size;
+        }
+
+        int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
+        int idx_std  = get_key_idx(ctx, KEY_IMAGE_STD);
+
+        const float * mean_data = (const float *)gguf_get_arr_data(ctx, idx_mean);
+        const float * std_data  = (const float *)gguf_get_arr_data(ctx, idx_std);
+
+        for (int i = 0; i < 3; ++i) {
+            new_clip->image_mean[i] = mean_data[i];
+            new_clip->image_std[i]  = std_data[i];
+        }
+
+        if (verbosity >= 2) {
+            LOG_TEE("\n%s: vision model hparams\n", __func__);
+            LOG_TEE("image_size         %d\n", hparams.image_size);
+            LOG_TEE("patch_size         %d\n", hparams.patch_size);
+            LOG_TEE("v_hidden_size      %d\n", hparams.hidden_size);
+            LOG_TEE("v_n_intermediate   %d\n", hparams.n_intermediate);
+            LOG_TEE("v_projection_dim   %d\n", hparams.projection_dim);
+            LOG_TEE("v_n_head           %d\n", hparams.n_head);
+            LOG_TEE("v_n_layer          %d\n", hparams.n_layer);
+            LOG_TEE("v_eps              %f\n", hparams.eps);
+            LOG_TEE("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
+            LOG_TEE("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
+            LOG_TEE("v_image_grid_pinpoints: ");
+            for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
+                LOG_TEE("%d ", hparams.image_grid_pinpoints[i]);
+            }
+            LOG_TEE("\n");
+            LOG_TEE("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
+
+        }
+
+        try {
+            vision_model.class_embedding  = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
+            new_clip->has_class_embedding = true;
+        } catch (const std::exception& e) {
+            new_clip->has_class_embedding = false;
+        }
+
+        try {
+            vision_model.pre_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
+            vision_model.pre_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+            new_clip->has_pre_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_pre_norm = false;
+        }
+
+        try {
+            vision_model.post_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight"));
+            vision_model.post_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias"));
+            new_clip->has_post_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_post_norm = false;
+        }
+
+        try {
+            vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS);
+            new_clip->has_patch_bias = true;
+        } catch (std::exception & e) {
+            new_clip->has_patch_bias = false;
+        }
+
+        try {
+            vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+            vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
+        } catch(const std::exception& e) {
+            LOG_TEE("%s: failed to load vision model tensors\n", __func__);
+        }
+
+        // LLaVA projection
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            vision_model.mm_0_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
+            vision_model.mm_0_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
+            try {
+                // Yi-type llava
+                vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight"));
+                vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // missing in Yi-type llava
+                vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+                vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight"));
+                vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight"));
+                vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
+                // LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__);
+            } catch (std::runtime_error & e) {  }
+        } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projection
+            vision_model.mm_model_mlp_1_w               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight"));
+            vision_model.mm_model_mlp_1_b               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias"));
+            vision_model.mm_model_mlp_3_w               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight"));
+            vision_model.mm_model_mlp_3_b               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias"));
+            vision_model.mm_model_block_1_block_0_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight"));
+            vision_model.mm_model_block_1_block_0_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight"));
+            vision_model.mm_model_block_1_block_0_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias"));
+            vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight"));
+            vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias"));
+            vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight"));
+            vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias"));
+            vision_model.mm_model_block_1_block_2_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight"));
+            vision_model.mm_model_block_1_block_2_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight"));
+            vision_model.mm_model_block_1_block_2_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias"));
+            vision_model.mm_model_block_2_block_0_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight"));
+            vision_model.mm_model_block_2_block_0_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight"));
+            vision_model.mm_model_block_2_block_0_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias"));
+            vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight"));
+            vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias"));
+            vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight"));
+            vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias"));
+            vision_model.mm_model_block_2_block_2_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
+            vision_model.mm_model_block_2_block_2_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
+            vision_model.mm_model_block_2_block_2_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
+        }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            // MobilVLM_V2 projection
+            vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "weight"));
+            vision_model.mm_model_mlp_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "bias"));
+            vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "weight"));
+            vision_model.mm_model_mlp_2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "bias"));
+            vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
+            vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
+        }
+        else {
+            std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
+            throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
+        }
+
+        vision_model.layers.resize(hparams.n_layer);
+
+        for (int il = 0; il < hparams.n_layer; ++il) {
+            auto & layer = vision_model.layers[il];
+            layer.k_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "weight"));
+            layer.q_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "weight"));
+            layer.v_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "weight"));
+            layer.o_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "weight"));
+            layer.ln_1_w = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "weight"));
+            layer.ln_2_w = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "weight"));
+            layer.ff_i_w = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "weight"));
+            layer.ff_o_w = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "weight"));
+            layer.k_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "bias"));
+            layer.q_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "bias"));
+            layer.v_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "bias"));
+            layer.o_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "bias"));
+            layer.ln_1_b = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "bias"));
+            layer.ln_2_b = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "bias"));
+            layer.ff_i_b = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "bias"));
+            layer.ff_o_b = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "bias"));
+        }
+    }
+
+    ggml_free(meta);
+
+    new_clip->ctx_gguf = ctx;
+
+    // measure mem requirement and allocate
+    {
+        new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+        new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
+        clip_image_f32_batch batch;
+        batch.size = 1;
+        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
+        ggml_gallocr_reserve(new_clip->compute_alloc, gf);
+        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
+        LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+    }
+
+    return new_clip;
+}
+
+struct clip_image_u8 * clip_image_u8_init() {
+    return new clip_image_u8();
+}
+
+struct clip_image_f32 * clip_image_f32_init() {
+    return new clip_image_f32();
+}
+
+void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
+void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
+void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
+    }
+}
+void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
+    }
+}
+
+static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
+    img->nx = nx;
+    img->ny = ny;
+    img->buf.resize(3 * nx * ny);
+    memcpy(img->buf.data(), data, img->buf.size());
+}
+
+bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
+    if (!data) {
+        LOG_TEE("%s: failed to load image '%s'\n", __func__, fname);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
+    stbi_image_free(data);
+    return true;
+}
+
+bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
+    if (!data) {
+        LOG_TEE("%s: failed to decode image bytes\n", __func__);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
+    stbi_image_free(data);
+    return true;
+}
+
+// Linear interpolation between two points
+inline float clip_lerp(float s, float e, float t) {
+    return s + (e - s) * t;
+}
+// Bilinear resize function
+static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int target_width, int target_height) {
+    dst.nx = target_width;
+    dst.ny = target_height;
+    dst.buf.resize(3 * target_width * target_height);
+
+    float x_ratio = static_cast<float>(src.nx - 1) / target_width;
+    float y_ratio = static_cast<float>(src.ny - 1) / target_height;
+
+    for (int y = 0; y < target_height; y++) {
+        for (int x = 0; x < target_width; x++) {
+            float px = x_ratio * x;
+            float py = y_ratio * y;
+            int x_floor = static_cast<int>(px);
+            int y_floor = static_cast<int>(py);
+            float x_lerp = px - x_floor;
+            float y_lerp = py - y_floor;
+
+            for (int c = 0; c < 3; c++) {
+                float top = clip_lerp(
+                    static_cast<float>(src.buf[3 * (y_floor * src.nx + x_floor) + c]),
+                    static_cast<float>(src.buf[3 * (y_floor * src.nx + (x_floor + 1)) + c]),
+                    x_lerp
+                );
+                float bottom = clip_lerp(
+                    static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + x_floor) + c]),
+                    static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + (x_floor + 1)) + c]),
+                    x_lerp
+                );
+                dst.buf[3 * (y * target_width + x) + c] = static_cast<uint8_t>(clip_lerp(top, bottom, y_lerp));
+            }
+        }
+    }
+}
+
+// Normalize image to float32 - careful with pytorch .to(model.device, dtype=torch.float16) - this sometimes reduces precision (32>16>32), sometimes not
+static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32* dst, const float mean[3], const float std[3]) {
+    dst->nx = src->nx;
+    dst->ny = src->ny;
+    dst->buf.resize(src->buf.size());
+
+    for (size_t i = 0; i < src->buf.size(); ++i) {
+        int c = i % 3; // rgb
+        dst->buf[i] = (static_cast<float>(src->buf[i]) / 255.0f - mean[c]) / std[c];
+    }
+}
+
+inline float clip(float x, float lower, float upper) {
+    return std::max(lower, std::min(x, upper));
+}
+
+static bool bicubic_resize(const clip_image_u8 &img, clip_image_u8 &dst, int target_width, int target_height) {
+    const int nx = img.nx;
+    const int ny = img.ny;
+
+    dst.nx = target_width;
+    dst.ny = target_height;
+    dst.buf.resize(3 * target_width * target_height);
+
+    float Cc;
+    float C[5];
+    float d0, d2, d3, a0, a1, a2, a3;
+    int i, j, k, jj;
+    int x, y;
+    float dx, dy;
+    float tx, ty;
+
+    tx = (float)nx / (float)target_width;
+    ty = (float)ny / (float)target_height;
+
+    // Bicubic interpolation; adapted from ViT.cpp, inspired from :
+    //    -> https://github.com/yglukhov/bicubic-interpolation-image-processing/blob/master/libimage.c#L36
+    //    -> https://en.wikipedia.org/wiki/Bicubic_interpolation
+
+    for (i = 0; i < target_height; i++) {
+        for (j = 0; j < target_width; j++) {
+            x = (int)(tx * j);
+            y = (int)(ty * i);
+
+            dx = tx * j - x;
+            dy = ty * i - y;
+
+            for (k = 0; k < 3; k++) {
+                for (jj = 0; jj <= 3; jj++) {
+                    d0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x - 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+                    d2 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+                    d3 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 2, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+                    a0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+
+                    a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
+                    a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
+                    a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;
+
+                    C[jj] = a0 + a1 * dx + a2 * dx * dx + a3 * dx * dx * dx;
+
+                    d0 = C[0] - C[1];
+                    d2 = C[2] - C[1];
+                    d3 = C[3] - C[1];
+                    a0 = C[1];
+                    a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
+                    a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
+                    a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;
+                    Cc = a0 + a1 * dy + a2 * dy * dy + a3 * dy * dy * dy;
+
+                    const uint8_t Cc2 = std::min(std::max(std::round(Cc), 0.0f), 255.0f);
+                    dst.buf[(i * target_width + j) * 3 + k] = float(Cc2);
+                }
+            }
+        }
+    }
+
+    return true;
+}
+
+// llava-1.6 type of resize_and_pad (black)
+static void resize_and_pad_image(const clip_image_u8& image, clip_image_u8 &image_output, const std::pair<int, int>& target_resolution) {
+    int target_width = target_resolution.first;
+    int target_height = target_resolution.second;
+
+    float scale_w = static_cast<float>(target_width) / image.nx;
+    float scale_h = static_cast<float>(target_height) / image.ny;
+
+    int new_width, new_height;
+
+    if (scale_w < scale_h) {
+        new_width = target_width;
+        new_height = std::min(static_cast<int>(std::ceil(image.ny * scale_w)), target_height);
+    } else {
+        new_height = target_height;
+        new_width = std::min(static_cast<int>(std::ceil(image.nx * scale_h)), target_width);
+    }
+
+    clip_image_u8 resized_image;
+    // bilinear_resize(image, resized_image, new_width, new_height);
+    bicubic_resize(image, resized_image, new_width, new_height);
+
+    clip_image_u8 padded_image;
+    padded_image.nx = target_width;
+    padded_image.ny = target_height;
+    padded_image.buf.resize(3 * target_width * target_height, 0); // Initialize with black
+
+    // Calculate padding offsets
+    int pad_x = (target_width - new_width) / 2;
+    int pad_y = (target_height - new_height) / 2;
+
+    // Copy the resized image into the center of the padded buffer
+    for (int y = 0; y < new_height; ++y) {
+        for (int x = 0; x < new_width; ++x) {
+            for (int c = 0; c < 3; ++c) {
+                padded_image.buf[3 * ((y + pad_y) * target_width + (x + pad_x)) + c] = resized_image.buf[3 * (y * new_width + x) + c];
+            }
+        }
+    }
+    image_output = std::move(padded_image);
+}
+
+/**
+ * Selects the best resolution from a list of possible resolutions based on the original size.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int> & original_size, const std::vector<std::pair<int, int>> & possible_resolutions) {
+    int original_width = original_size.first;
+    int original_height = original_size.second;
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & image, int patch_size) {
+    std::vector<clip_image_u8*> patches;
+    int width = image.nx;
+    int height = image.ny;
+    for (int i = 0; i < height; i += patch_size) {
+        for (int j = 0; j < width; j += patch_size) {
+            clip_image_u8 *patch = clip_image_u8_init();
+            patch->nx = std::min(patch_size, width - j);
+            patch->ny = std::min(patch_size, height - i);
+            patch->buf.resize(3 * patch->nx * patch->ny);
+            for (int y = 0; y < patch->ny; ++y) {
+                for (int x = 0; x < patch->nx; ++x) {
+                    for (int c = 0; c < 3; ++c) {
+                        patch->buf[3 * (y * patch->nx + x) + c] = image.buf[3 * ((i + y) * width + (j + x)) + c];
+                    }
+                }
+            }
+            patches.push_back(patch);
+        }
+    }
+    return patches;
+}
+
+// returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
+// res_imgs memory is being allocated here, previous allocations will be freed if found
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
+    bool pad_to_square = true;
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return false;
+    }
+    auto & params = ctx->vision_model.hparams;
+    // The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing
+    if (strcmp(params.mm_patch_merge_type, "spatial_unpad") == 0) {
+        pad_to_square = false;
+    }
+    // free the previous res_imgs if any set
+    if (res_imgs->size > 0) {
+        clip_image_f32_batch_free(res_imgs);
+    }
+    res_imgs->data = nullptr;
+    res_imgs->size = 0;
+
+    // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
+    // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
+
+    clip_image_u8 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily
+    if (pad_to_square && img->nx != img->ny) {
+        int longer_side = std::max(img->nx, img->ny);
+        temp->nx = longer_side;
+        temp->ny = longer_side;
+        temp->buf.resize(3 * longer_side * longer_side);
+        const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA (this is the mean rgb color * 255)
+
+        // fill with background color
+        for (size_t i = 0; i < temp->buf.size(); i++) {
+            temp->buf[i] = bc[i % 3];
+        }
+
+        // copy from the input image
+        for (int y = 0; y < img->ny; y++) {
+            for (int x = 0; x < img->nx; x++) {
+                const int i = 3 * (y * img->nx + x);
+                const int j = 3 * (y * temp->nx + x);
+                temp->buf[j]   = img->buf[i];
+                temp->buf[j+1] = img->buf[i+1];
+                temp->buf[j+2] = img->buf[i+2];
+            }
+        }
+    } else {
+        if (params.image_grid_pinpoints[0] != 0) {
+            // "spatial_unpad" with "anyres" processing for llava-1.6
+            std::vector<std::pair<int, int>> possible_resolutions;
+            for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+                possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+            }
+            std::pair<int, int> best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions);
+            // clip_image_save_to_bmp(*img, "input.bmp");
+            resize_and_pad_image(*img, *temp, best_resolution);  // we do not pad with mean-bg color anymore in llava-1.6
+            // clip_image_save_to_bmp(*temp, "resized.bmp");
+            // visually verify normalized image:
+            // normalize_image_u8_to_f32(*temp, *res, ctx->image_mean, ctx->image_std);
+            // {
+            //     clip_image_u8 * temp2 = clip_image_u8_init();
+            //     clip_image_convert_f32_to_u8(*res, *temp2);
+            //     clip_image_save_to_bmp(*temp2, "resized_normalized_f32.bmp");
+            //     clip_image_u8_free(temp2);
+            // }
+
+            std::vector<clip_image_u8 *> patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6)
+
+            clip_image_u8 *image_original_resize = clip_image_u8_init();
+            // bilinear_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
+            bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
+            patches.insert(patches.begin(), image_original_resize);
+            // clip_image_f32_batch_init(patches.size());
+            res_imgs->size = patches.size();
+            res_imgs->data = new clip_image_f32[res_imgs->size];
+            int num=0;
+            for (auto& patch : patches) {
+                normalize_image_u8_to_f32(patch, &res_imgs->data[num], ctx->image_mean, ctx->image_std);
+                num++;
+            }
+
+            for (size_t i = 0; i < patches.size(); i++) {
+                // LOG_TEE("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
+                clip_image_u8_free(patches[i]);
+            }
+
+            clip_image_u8_free(temp);
+
+            return true;
+        } else {
+            temp->nx = img->nx;
+            temp->ny = img->ny;
+            temp->buf.resize(img->buf.size());
+            memcpy(temp->buf.data(), img->buf.data(), temp->buf.size());
+        }
+    }
+
+    const int nx = temp->nx;
+    const int ny = temp->ny;
+    // clip_image_save_to_bmp(*temp, "resized_vanilla.bmp");
+
+    const int nx2 = ctx->vision_model.hparams.image_size;
+    const int ny2 = ctx->vision_model.hparams.image_size;
+    clip_image_f32 * res = clip_image_f32_init();
+    res->nx = nx2;
+    res->ny = ny2;
+    res->buf.resize(3 * nx2 * ny2);
+
+    const float scale = std::max(nx, ny) / (float)ctx->vision_model.hparams.image_size;
+
+    const int nx3 = int(nx / scale + 0.5f);
+    const int ny3 = int(ny / scale + 0.5f);
+
+    const auto & m3 = ctx->image_mean; // {0.48145466f, 0.4578275f, 0.40821073f};
+    const auto & s3 = ctx->image_std;  // {0.26862954f, 0.26130258f, 0.27577711f};
+
+    for (int y = 0; y < ny3; y++) {
+        for (int x = 0; x < nx3; x++) {
+            for (int c = 0; c < 3; c++) {
+                // linear interpolation
+                const float sx = (x + 0.5f) * scale - 0.5f;
+                const float sy = (y + 0.5f) * scale - 0.5f;
+
+                const int x0 = std::max(0, (int)std::floor(sx));
+                const int y0 = std::max(0, (int)std::floor(sy));
+
+                const int x1 = std::min(x0 + 1, nx - 1);
+                const int y1 = std::min(y0 + 1, ny - 1);
+
+                const float dx = sx - x0;
+                const float dy = sy - y0;
+
+                const int j00 = 3 * (y0 * nx + x0) + c;
+                const int j01 = 3 * (y0 * nx + x1) + c;
+                const int j10 = 3 * (y1 * nx + x0) + c;
+                const int j11 = 3 * (y1 * nx + x1) + c;
+
+                const float v00 = temp->buf[j00];
+                const float v01 = temp->buf[j01];
+                const float v10 = temp->buf[j10];
+                const float v11 = temp->buf[j11];
+
+                const float v0 = v00 * (1.0f - dx) + v01 * dx;
+                const float v1 = v10 * (1.0f - dx) + v11 * dx;
+
+                const float v = v0 * (1.0f - dy) + v1 * dy;
+
+                const uint8_t v2 = std::min(std::max(std::round(v), 0.0f), 255.0f);
+
+                const int i = 3 * (y * nx3 + x) + c;
+
+                res->buf[i] = ((float(v2) / 255.0f) - m3[c]) / s3[c];
+            }
+        }
+    }
+    clip_image_u8_free(temp);
+
+    // {
+    //     clip_image_u8 * temp2 = clip_image_u8_init();
+    //     clip_image_convert_f32_to_u8(*res, *temp2);
+    //     clip_image_save_to_bmp(*temp2, "resized_normalized_f32_vanilla.bmp");
+    //     clip_image_u8_free(temp2);
+    // }
+    // res_imgs.push_back(res);
+
+    res_imgs->size = 1;
+    res_imgs->data = new clip_image_f32[res_imgs->size];
+    res_imgs->data[0] = *res;
+    clip_image_f32_free(res);
+
+    return true;
+}
+
+ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
+    return ctx->vision_model.image_newline;
+}
+
+void clip_free(clip_ctx * ctx) {
+    ggml_free(ctx->ctx_data);
+    gguf_free(ctx->ctx_gguf);
+
+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
+    delete ctx;
+}
+
+size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
+    return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
+}
+
+int32_t clip_image_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.image_size;
+}
+
+int32_t clip_patch_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.patch_size;
+}
+
+int32_t clip_hidden_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.hidden_size;
+}
+
+const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.mm_patch_merge_type;
+}
+
+const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.image_grid_pinpoints;
+}
+
+int clip_n_patches(const struct clip_ctx * ctx) {
+    const auto & params = ctx->vision_model.hparams;
+
+    int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
+
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+        n_patches /= 4;
+    }
+
+    return n_patches;
+}
+
+bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return false;
+    }
+
+    clip_image_f32_batch imgs{};
+    imgs.size = 1;
+    imgs.data = img;
+    return clip_image_batch_encode(ctx, n_threads, &imgs, vec);
+}
+
+bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return false;
+    }
+
+    int batch_size = imgs->size;
+    if (ctx->has_llava_projector) {
+        GGML_ASSERT(batch_size == 1); // TODO: support multiple images
+    }
+
+    // build the inference graph
+    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
+    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+
+    // set inputs
+    const auto & model = ctx->vision_model;
+    const auto & hparams = model.hparams;
+
+    const int image_size    = hparams.image_size;
+    const int patch_size    = hparams.patch_size;
+    const int num_patches   = ((image_size / patch_size) * (image_size / patch_size));
+    const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
+
+    {
+        struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
+        float * data = (float *)malloc(ggml_nbytes(inp_raw));
+
+        for (size_t i = 0; i < imgs->size; i++) {
+            const int nx = imgs->data[i].nx;
+            const int ny = imgs->data[i].ny;
+            GGML_ASSERT(nx == image_size && ny == image_size);
+
+            const int n = nx * ny;
+
+            for (int b = 0; b < batch_size; b++) {
+                for (int k = 0; k < 3; k++) {
+                    for (int y = 0; y < ny; y++) {
+                        for (int x = 0; x < nx; x++) {
+                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k];
+                        }
+                    }
+                }
+            }
+        }
+        ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
+        free(data);
+    }
+
+    {
+        if (ctx->has_class_embedding) {
+            struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
+
+            void* zero_mem = malloc(ggml_nbytes(embeddings));
+            memset(zero_mem, 0, ggml_nbytes(embeddings));
+            ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+            free(zero_mem);
+        }
+    }
+
+    {
+        struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+
+        int* positions_data = (int*)malloc(ggml_nbytes(positions));
+        for (int i = 0; i < num_positions; i++) {
+            positions_data[i] = i;
+        }
+        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+        free(positions_data);
+    }
+
+    {
+        struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+        int* patches_data = (int*)malloc(ggml_nbytes(patches));
+        for (int i = 0; i < num_patches; i++) {
+            patches_data[i] = i + 1;
+        }
+        ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+        free(patches_data);
+    }
+
+    if (ggml_backend_is_cpu(ctx->backend)) {
+        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
+    }
+
+#ifdef GGML_USE_METAL
+    if (ggml_backend_is_metal(ctx->backend)) {
+        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
+    }
+#endif
+
+    ggml_backend_graph_compute(ctx->backend, gf);
+
+    // the last node is the embedding tensor
+    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
+
+    // copy the embeddings to the location passed by the user
+    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
+
+    return true;
+}
+
+bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
+    assert(itype < GGML_TYPE_COUNT);
+    type = static_cast<ggml_type>(itype);
+
+    auto * ctx_clip = clip_model_load(fname_inp, 2);
+
+    const auto & ctx_src = ctx_clip->ctx_gguf;
+    const auto & ctx_data = ctx_clip->ctx_data;
+
+    auto * ctx_out = gguf_init_empty();
+    gguf_set_kv(ctx_out, ctx_src);
+    gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
+    gguf_set_val_u32(ctx_out, "general.file_type", itype);
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+
+    const int n_tensors = gguf_get_n_tensors(ctx_src);
+
+    for (int i = 0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(ctx_src, i);
+        struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
+        gguf_add_tensor(ctx_out, cur);
+    }
+
+    const size_t meta_size = gguf_get_meta_size(ctx_out);
+    for (size_t i = 0; i < meta_size; ++i) {
+        fout.put(0);
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> k_names = {
+        ".*weight",
+    };
+
+    std::vector<uint8_t> work(512);
+    std::vector<float> conv_buf(512);
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    for (int i = 0; i < n_tensors; ++i) {
+        const std::string name = gguf_get_tensor_name(ctx_src, i);
+        struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name.c_str());
+
+        enum ggml_type new_type;
+        void * new_data;
+        size_t new_size;
+
+        bool quantize = false;
+        for (const auto & s : k_names) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = true;
+                break;
+            }
+        }
+
+        // quantize only 2D tensors
+        quantize &= (ggml_n_dims(cur) == 2);
+
+        if (quantize) {
+            new_type = type;
+            if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
+                new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
+                // LOG_TEE("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+            }
+            const size_t n_elms = ggml_nelements(cur);
+            float * f32_data;
+
+            switch (cur->type) {
+            case GGML_TYPE_F32:
+                f32_data = (float *)cur->data;
+                break;
+            case GGML_TYPE_F16:
+                if (conv_buf.size() < n_elms) {
+                    conv_buf.resize(n_elms);
+                }
+                for (size_t j = 0; j < n_elms; ++j) {
+                    conv_buf[j] = ggml_fp16_to_fp32(((ggml_fp16_t *)cur->data)[j]);
+                }
+                f32_data = (float *)conv_buf.data();
+                break;
+            default:
+                LOG_TEE("Please use an input file in f32 or f16\n");
+                gguf_free(ctx_out);
+                return false;
+            }
+
+            if (work.size() < n_elms * 4) {
+                work.resize(n_elms * 4);
+            }
+            new_data = work.data();
+
+            new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr);
+        } else {
+            new_type = cur->type;
+            new_data = cur->data;
+            new_size = ggml_nbytes(cur);
+        }
+        const size_t orig_size = ggml_nbytes(cur);
+        total_size_org += orig_size;
+        total_size_new += new_size;
+        gguf_set_tensor_type(ctx_out, name.c_str(), new_type);
+        gguf_set_tensor_data(ctx_out, name.c_str(), new_data, new_size);
+        fout.write((const char *)new_data, new_size);
+        size_t pad = GGML_PAD(new_size, gguf_get_alignment(ctx_out)) - new_size;
+        for (size_t j = 0; j < pad; ++j) {
+            fout.put(0);
+        }
+
+        LOG_TEE("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
+               orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+    }
+
+    // go back to beginning of file and write the updated metadata
+    fout.seekp(0, std::ios::beg);
+    std::vector<uint8_t> meta(meta_size);
+    gguf_get_meta_data(ctx_out, meta.data());
+    fout.write((const char *)meta.data(), meta_size);
+
+    fout.close();
+
+    clip_free(ctx_clip);
+    gguf_free(ctx_out);
+
+    {
+        LOG_TEE("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+        LOG_TEE("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+    }
+
+    return true;
+}
+
+int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+        return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0];
+    }
+    if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+        return ctx->vision_model.mm_model_peg_0_b->ne[0];
+    }
+    if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
+        return ctx->vision_model.mm_2_b->ne[0];
+    }
+    if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        return ctx->vision_model.mm_3_b->ne[0];
+    }
+
+    std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
+    throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
+}

llama/clip.h

@@ -0,0 +1,111 @@
+/**
+ * llama.cpp - git 059031b8c40e1f4ba60586842c5b1ed3ddf61842
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef CLIP_H
+#define CLIP_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define CLIP_API __declspec(dllexport)
+#        else
+#            define CLIP_API __declspec(dllimport)
+#        endif
+#    else
+#        define CLIP_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define CLIP_API
+#endif
+
+struct clip_ctx;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct clip_ctx;
+
+struct clip_image_u8_batch {
+    struct clip_image_u8 * data;
+    size_t size;
+};
+
+struct clip_image_f32_batch {
+    struct clip_image_f32 * data;
+    size_t size;
+};
+
+CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+
+CLIP_API void clip_free(struct clip_ctx * ctx);
+
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+
+CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+
+// TODO: should be enum, not string
+CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
+
+CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
+CLIP_API struct clip_image_u8  * clip_image_u8_init ();
+CLIP_API struct clip_image_f32 * clip_image_f32_init();
+
+CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
+CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
+
+CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+
+/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
+CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
+
+/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
+
+CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
+CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
+
+CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // CLIP_H

llama/log.h

@@ -0,0 +1,750 @@
+/**
+ * llama.cpp - git 059031b8c40e1f4ba60586842c5b1ed3ddf61842
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include <chrono>
+#include <cstring>
+#include <sstream>
+#include <iostream>
+#include <thread>
+#include <vector>
+#include <algorithm>
+#include <cinttypes>
+
+// --------------------------------
+//
+// Basic usage:
+//
+// --------
+//
+//  The LOG() and LOG_TEE() macros are ready to go by default
+//   they do not require any initialization.
+//
+//  LOGLN() and LOG_TEELN() are variants which automatically
+//   include \n character at the end of the log string.
+//
+//  LOG() behaves exactly like printf, by default writing to a logfile.
+//  LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
+//
+//  Default logfile is named
+//   "llama.<threadID>.log"
+//  Default LOG_TEE() secondary output target is
+//   stderr
+//
+//  Logs can be dynamically disabled or enabled using functions:
+//   log_disable()
+//  and
+//   log_enable()
+//
+//  A log target can be changed with:
+//   log_set_target( string )
+//    creating and opening, or re-opening a file by string filename
+//  or
+//   log_set_target( FILE* )
+//    allowing to point at stderr, stdout, or any valid FILE* file handler.
+//
+// --------
+//
+// End of Basic usage.
+//
+// --------------------------------
+
+// Specifies a log target.
+//  default uses log_handler() with "llama.log" log file
+//  this can be changed, by defining LOG_TARGET
+//  like so:
+//
+//  #define LOG_TARGET (a valid FILE*)
+//  #include "log.h"
+//
+//  or it can be simply redirected to stdout or stderr
+//  like so:
+//
+//  #define LOG_TARGET stderr
+//  #include "log.h"
+//
+//  The log target can also be redirected to a different function
+//  like so:
+//
+//  #define LOG_TARGET log_handler_different()
+//  #include "log.h"
+//
+//  FILE* log_handler_different()
+//  {
+//      return stderr;
+//  }
+//
+//  or:
+//
+//  #define LOG_TARGET log_handler_another_one("somelog.log")
+//  #include "log.h"
+//
+//  FILE* log_handler_another_one(char*filename)
+//  {
+//      static FILE* logfile = nullptr;
+//      (...)
+//      if( !logfile )
+//      {
+//          fopen(...)
+//      }
+//      (...)
+//      return logfile
+//  }
+//
+#ifndef LOG_TARGET
+    #define LOG_TARGET log_handler()
+#endif
+
+#ifndef LOG_TEE_TARGET
+    #define LOG_TEE_TARGET stderr
+#endif
+
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
+// Utility to obtain "pid" like unique process id and use it when creating log files.
+inline std::string log_get_pid()
+{
+   static std::string pid;
+   if (pid.empty())
+   {
+       // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+       //  it's not the same as "pid" but is unique enough to solve multiple instances
+       //  trying to write to the same log.
+       std::stringstream ss;
+       ss << std::this_thread::get_id();
+       pid = ss.str();
+   }
+
+   return pid;
+}
+
+// Utility function for generating log file names with unique id based on thread id.
+//  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
+//  where the number is a runtime id of the current thread.
+
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
+
+// INTERNAL, DO NOT USE
+inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
+{
+    static bool _multilog = false;
+
+    if (multilog != LogTriStateSame)
+    {
+        _multilog = multilog == LogTriStateTrue;
+    }
+
+    std::stringstream buf;
+
+    buf << log_file_basename;
+    if (_multilog)
+    {
+        buf << ".";
+        buf << log_get_pid();
+    }
+    buf << ".";
+    buf << log_file_extension;
+
+    return buf.str();
+}
+
+#ifndef LOG_DEFAULT_FILE_NAME
+    #define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
+#endif
+
+// Utility for turning #define values into string literals
+//  so we can have a define for stderr and
+//  we can print "stderr" instead of literal stderr, etc.
+#define LOG_STRINGIZE1(s) #s
+#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
+
+#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
+
+// Allows disabling timestamps.
+//  in order to disable, define LOG_NO_TIMESTAMPS
+//  like so:
+//
+//  #define LOG_NO_TIMESTAMPS
+//  #include "log.h"
+//
+#ifndef LOG_NO_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TIMESTAMP_FMT "%s"
+    #define LOG_TIMESTAMP_VAL ,""
+#endif
+
+#ifdef LOG_TEE_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TEE_TIMESTAMP_FMT "%s"
+    #define LOG_TEE_TIMESTAMP_VAL ,""
+#endif
+
+// Allows disabling file/line/function prefix
+//  in order to disable, define LOG_NO_FILE_LINE_FUNCTION
+//  like so:
+//
+//  #define LOG_NO_FILE_LINE_FUNCTION
+//  #include "log.h"
+//
+#ifndef LOG_NO_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_FLF_FMT "%s"
+    #define LOG_FLF_VAL ,""
+#endif
+
+#ifdef LOG_TEE_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_TEE_FLF_FMT "%s"
+    #define LOG_TEE_FLF_VAL ,""
+#endif
+
+// INTERNAL, DO NOT USE
+//  USE LOG() INSTEAD
+//
+#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
+    #define LOG_IMPL(str, ...)                                                                                      \
+    do {                                                                                                            \
+        if (LOG_TARGET != nullptr)                                                                                  \
+        {                                                                                                           \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                     \
+        }                                                                                                           \
+    } while (0)
+#else
+    #define LOG_IMPL(str, ...)                                                                                           \
+    do {                                                                                                                 \
+        if (LOG_TARGET != nullptr)                                                                                       \
+        {                                                                                                                \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                          \
+        }                                                                                                                \
+    } while (0)
+#endif
+
+// INTERNAL, DO NOT USE
+//  USE LOG_TEE() INSTEAD
+//
+#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
+    #define LOG_TEE_IMPL(str, ...)                                                                                                      \
+    do {                                                                                                                                \
+        if (LOG_TARGET != nullptr)                                                                                                      \
+        {                                                                                                                               \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                         \
+        }                                                                                                                               \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                         \
+        {                                                                                                                               \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                     \
+        }                                                                                                                               \
+    } while (0)
+#else
+    #define LOG_TEE_IMPL(str, ...)                                                                                                           \
+    do {                                                                                                                                     \
+        if (LOG_TARGET != nullptr)                                                                                                           \
+        {                                                                                                                                    \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                              \
+        }                                                                                                                                    \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                              \
+        {                                                                                                                                    \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                          \
+        }                                                                                                                                    \
+    } while (0)
+#endif
+
+// The '\0' as a last argument, is a trick to bypass the silly
+//  "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
+//  so we can have a single macro which can be called just like printf.
+
+// Main LOG macro.
+//  behaves like printf, and supports arguments the exact same way.
+//
+#if !defined(_MSC_VER) || defined(__clang__)
+    #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
+#endif
+
+// Main TEE macro.
+//  does the same as LOG
+//  and
+//  simultaneously writes stderr.
+//
+// Secondary target can be changed just like LOG_TARGET
+//  by defining LOG_TEE_TARGET
+//
+#if !defined(_MSC_VER) || defined(__clang__)
+    #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
+#endif
+
+// LOG macro variants with auto endline.
+#if !defined(_MSC_VER) || defined(__clang__)
+    #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
+    #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
+#else
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
+#endif
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+{
+    static bool _initialized = false;
+    static bool _append = false;
+    static bool _disabled = filename.empty() && target == nullptr;
+    static std::string log_current_filename{filename};
+    static FILE *log_current_target{target};
+    static FILE *logfile = nullptr;
+
+    if (change)
+    {
+        if (append != LogTriStateSame)
+        {
+            _append = append == LogTriStateTrue;
+            return logfile;
+        }
+
+        if (disable == LogTriStateTrue)
+        {
+            // Disable primary target
+            _disabled = true;
+        }
+        // If previously disabled, only enable, and keep previous target
+        else if (disable == LogTriStateFalse)
+        {
+            _disabled = false;
+        }
+        // Otherwise, process the arguments
+        else if (log_current_filename != filename || log_current_target != target)
+        {
+            _initialized = false;
+        }
+    }
+
+    if (_disabled)
+    {
+        // Log is disabled
+        return nullptr;
+    }
+
+    if (_initialized)
+    {
+        // with fallback in case something went wrong
+        return logfile ? logfile : stderr;
+    }
+
+    // do the (re)initialization
+    if (target != nullptr)
+    {
+        if (logfile != nullptr && logfile != stdout && logfile != stderr)
+        {
+            fclose(logfile);
+        }
+
+        log_current_filename = LOG_DEFAULT_FILE_NAME;
+        log_current_target = target;
+
+        logfile = target;
+    }
+    else
+    {
+        if (log_current_filename != filename)
+        {
+            if (logfile != nullptr && logfile != stdout && logfile != stderr)
+            {
+                fclose(logfile);
+            }
+        }
+
+        logfile = fopen(filename.c_str(), _append ? "a" : "w");
+    }
+
+    if (!logfile)
+    {
+        //  Verify whether the file was opened, otherwise fallback to stderr
+        logfile = stderr;
+
+        fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
+        fflush(stderr);
+
+        // At this point we let the init flag be to true below, and let the target fallback to stderr
+        //  otherwise we would repeatedly fopen() which was already unsuccessful
+    }
+
+    _initialized = true;
+
+    return logfile ? logfile : stderr;
+}
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+{
+    return log_handler1_impl(change, append, disable, filename, target);
+}
+
+// Disables logs entirely at runtime.
+//  Makes LOG() and LOG_TEE() produce no output,
+//  until enabled back.
+#define log_disable() log_disable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_disable_impl()
+{
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
+}
+
+// Enables logs at runtime.
+#define log_enable() log_enable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_enable_impl()
+{
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
+}
+
+// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
+#define log_set_target(target) log_set_target_impl(target)
+
+// INTERNAL, DO NOT USE
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler() { return log_handler1_impl(); }
+
+// Enable or disable creating separate log files for each run.
+//  can ONLY be invoked BEFORE first log use.
+#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
+// Enable or disable append mode for log file.
+//  can ONLY be invoked BEFORE first log use.
+#define log_append(enable) log_append_impl(enable)
+// INTERNAL, DO NOT USE
+inline FILE *log_append_impl(bool enable)
+{
+    return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
+}
+
+inline void log_test()
+{
+    log_disable();
+    LOG("01 Hello World to nobody, because logs are disabled!\n");
+    log_enable();
+    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET));
+    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n");
+    log_set_target(stderr);
+    LOG("04 Hello World to stderr!\n");
+    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("06 Hello World to default log file!\n");
+    log_set_target(stdout);
+    LOG("07 Hello World to stdout!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("08 Hello World to default log file again!\n");
+    log_disable();
+    LOG("09 Hello World _1_ into the void!\n");
+    log_enable();
+    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n");
+    log_disable();
+    log_set_target("llama.anotherlog.log");
+    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n");
+    log_enable();
+    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n");
+    log_set_target("llama.yetanotherlog.log");
+    LOG("13 Hello World this time in yet new file?\n");
+    log_set_target(log_filename_generator("llama_autonamed", "log"));
+    LOG("14 Hello World in log with generated filename!\n");
+#ifdef _MSC_VER
+    LOG_TEE("15 Hello msvc TEE without arguments\n");
+    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test");
+    LOG_TEELN("17 Hello msvc TEELN without arguments\n");
+    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test");
+    LOG("19 Hello msvc LOG without arguments\n");
+    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test");
+    LOGLN("21 Hello msvc LOGLN without arguments\n");
+    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test");
+#endif
+}
+
+inline bool log_param_single_parse(const std::string & param)
+{
+    if ( param == "--log-test")
+    {
+        log_test();
+        return true;
+    }
+
+    if ( param == "--log-disable")
+    {
+        log_disable();
+        return true;
+    }
+
+    if ( param == "--log-enable")
+    {
+        log_enable();
+        return true;
+    }
+
+    if (param == "--log-new")
+    {
+        log_multilog(true);
+        return true;
+    }
+
+    if (param == "--log-append")
+    {
+        log_append(true);
+        return true;
+    }
+
+    return false;
+}
+
+inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
+{
+    if ( param == "--log-file")
+    {
+        if (!check_but_dont_parse)
+        {
+            log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
+        }
+
+        return true;
+    }
+
+    return false;
+}
+
+inline void log_print_usage()
+{
+    printf("log options:\n");
+    /* format
+    printf("  -h, --help            show this help message and exit\n");*/
+    /* spacing
+    printf("__-param----------------Description\n");*/
+    printf("  --log-test            Run simple logging test\n");
+    printf("  --log-disable         Disable trace logs\n");
+    printf("  --log-enable          Enable trace logs\n");
+    printf("  --log-file            Specify a log filename (without extension)\n");
+    printf("  --log-new             Create a separate new log file on start. "
+                                   "Each log file will have unique name: \"<name>.<ID>.log\"\n");
+    printf("  --log-append          Don't truncate the old log file.\n");
+    printf("\n");
+}
+
+#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
+
+// INTERNAL, DO NOT USE
+inline void log_dump_cmdline_impl(int argc, char **argv)
+{
+    std::stringstream buf;
+    for (int i = 0; i < argc; ++i)
+    {
+        if (std::string(argv[i]).find(' ') != std::string::npos)
+        {
+            buf << " \"" << argv[i] <<"\"";
+        }
+        else
+        {
+            buf << " " << argv[i];
+        }
+    }
+    LOGLN("Cmd:%s", buf.str().c_str());
+}
+
+#define log_tostr(var) log_var_to_string_impl(var).c_str()
+
+inline std::string log_var_to_string_impl(bool var)
+{
+    return var ? "true" : "false";
+}
+
+inline std::string log_var_to_string_impl(std::string var)
+{
+    return var;
+}
+
+inline std::string log_var_to_string_impl(const std::vector<int> & var)
+{
+    std::stringstream buf;
+    buf << "[ ";
+    bool first = true;
+    for (auto e : var)
+    {
+        if (first)
+        {
+            first = false;
+        }
+        else
+        {
+            buf << ", ";
+        }
+        buf << std::to_string(e);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+template <typename C, typename T>
+inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
+{
+    std::stringstream buf;
+    buf << "[ ";
+
+    bool first = true;
+    for (const auto &token : tokens)
+    {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = llama_token_to_piece(ctx, token);
+
+        detokenized.erase(
+            std::remove_if(
+                detokenized.begin(),
+                detokenized.end(),
+                [](const unsigned char c) { return !std::isprint(c); }),
+            detokenized.end());
+
+        buf
+            << "'" << detokenized << "'"
+            << ":" << std::to_string(token);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+template <typename C, typename B>
+inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
+{
+    std::stringstream buf;
+    buf << "[ ";
+
+    bool first = true;
+    for (int i = 0; i < batch.n_tokens; ++i)
+    {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
+
+        detokenized.erase(
+            std::remove_if(
+                detokenized.begin(),
+                detokenized.end(),
+                [](const unsigned char c) { return !std::isprint(c); }),
+            detokenized.end());
+
+        buf
+            << "\n" << std::to_string(i)
+            << ":token '" << detokenized << "'"
+            << ":pos " << std::to_string(batch.pos[i])
+            << ":n_seq_id  " << std::to_string(batch.n_seq_id[i])
+            << ":seq_id " << std::to_string(batch.seq_id[i][0])
+            << ":logits " << std::to_string(batch.logits[i]);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+#ifdef LOG_DISABLE_LOGS
+
+#undef LOG
+#define LOG(...) // dummy stub
+#undef LOGLN
+#define LOGLN(...) // dummy stub
+
+#undef LOG_TEE
+#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_TEELN
+#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_DISABLE
+#define LOG_DISABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_SET_TARGET
+#define LOG_SET_TARGET(...) // dummy stub
+
+#undef LOG_DUMP_CMDLINE
+#define LOG_DUMP_CMDLINE(...) // dummy stub
+
+#endif // LOG_DISABLE_LOGS

llama/stb_image.h

@@ -0,0 +1,8422 @@
+/**
+ * llama.cpp - git 059031b8c40e1f4ba60586842c5b1ed3ddf61842
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+/* stb_image - v2.28 - public domain image loader - http://nothings.org/stb
+                                  no warranty implied; use at your own risk
+
+   Do this:
+      #define STB_IMAGE_IMPLEMENTATION
+   before you include this file in *one* C or C++ file to create the implementation.
+
+   // i.e. it should look like this:
+   #include ...
+   #include ...
+   #include ...
+   #define STB_IMAGE_IMPLEMENTATION
+   #include "stb_image.h"
+
+   You can #define STBI_ASSERT(x) before the #include to avoid using assert.h.
+   And #define STBI_MALLOC, STBI_REALLOC, and STBI_FREE to avoid using malloc,realloc,free
+
+
+   QUICK NOTES:
+      Primarily of interest to game developers and other people who can
+          avoid problematic images and only need the trivial interface
+
+      JPEG baseline & progressive (12 bpc/arithmetic not supported, same as stock IJG lib)
+      PNG 1/2/4/8/16-bit-per-channel
+
+      TGA (not sure what subset, if a subset)
+      BMP non-1bpp, non-RLE
+      PSD (composited view only, no extra channels, 8/16 bit-per-channel)
+
+      GIF (*comp always reports as 4-channel)
+      HDR (radiance rgbE format)
+      PIC (Softimage PIC)
+      PNM (PPM and PGM binary only)
+
+      Animated GIF still needs a proper API, but here's one way to do it:
+          http://gist.github.com/urraka/685d9a6340b26b830d49
+
+      - decode from memory or through FILE (define STBI_NO_STDIO to remove code)
+      - decode from arbitrary I/O callbacks
+      - SIMD acceleration on x86/x64 (SSE2) and ARM (NEON)
+
+   Full documentation under "DOCUMENTATION" below.
+
+
+LICENSE
+
+  See end of file for license information.
+
+RECENT REVISION HISTORY:
+
+      2.28  (2023-01-29) many error fixes, security errors, just tons of stuff
+      2.27  (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes
+      2.26  (2020-07-13) many minor fixes
+      2.25  (2020-02-02) fix warnings
+      2.24  (2020-02-02) fix warnings; thread-local failure_reason and flip_vertically
+      2.23  (2019-08-11) fix clang static analysis warning
+      2.22  (2019-03-04) gif fixes, fix warnings
+      2.21  (2019-02-25) fix typo in comment
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) bugfix, 1-bit BMP, 16-bitness query, fix warnings
+      2.16  (2017-07-23) all functions have 16-bit variants; optimizations; bugfixes
+      2.15  (2017-03-18) fix png-1,2,4; all Imagenet JPGs; no runtime SSE detection on GCC
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-12-04) experimental 16-bit API, only for PNG so far; fixes
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) 16-bit PNGS; enable SSE2 in non-gcc x64
+                         RGB-format JPEG; remove white matting in PSD;
+                         allocate large structures on the stack;
+                         correct channel count for PNG & BMP
+      2.10  (2016-01-22) avoid warning introduced in 2.09
+      2.09  (2016-01-16) 16-bit TGA; comments in PNM files; STBI_REALLOC_SIZED
+
+   See end of file for full revision history.
+
+
+ ============================    Contributors    =========================
+
+ Image formats                          Extensions, features
+    Sean Barrett (jpeg, png, bmp)          Jetro Lauha (stbi_info)
+    Nicolas Schulz (hdr, psd)              Martin "SpartanJ" Golini (stbi_info)
+    Jonathan Dummer (tga)                  James "moose2000" Brown (iPhone PNG)
+    Jean-Marc Lienher (gif)                Ben "Disch" Wenger (io callbacks)
+    Tom Seddon (pic)                       Omar Cornut (1/2/4-bit PNG)
+    Thatcher Ulrich (psd)                  Nicolas Guillemot (vertical flip)
+    Ken Miller (pgm, ppm)                  Richard Mitton (16-bit PSD)
+    github:urraka (animated gif)           Junggon Kim (PNM comments)
+    Christopher Forseth (animated gif)     Daniel Gibson (16-bit TGA)
+                                           socks-the-fox (16-bit PNG)
+                                           Jeremy Sawicki (handle all ImageNet JPGs)
+ Optimizations & bugfixes                  Mikhail Morozov (1-bit BMP)
+    Fabian "ryg" Giesen                    Anael Seghezzi (is-16-bit query)
+    Arseny Kapoulkine                      Simon Breuss (16-bit PNM)
+    John-Mark Allen
+    Carmelo J Fdez-Aguera
+
+ Bug & warning fixes
+    Marc LeBlanc            David Woo          Guillaume George     Martins Mozeiko
+    Christpher Lloyd        Jerry Jansson      Joseph Thomson       Blazej Dariusz Roszkowski
+    Phil Jordan                                Dave Moore           Roy Eltham
+    Hayaki Saito            Nathan Reed        Won Chun
+    Luke Graham             Johan Duparc       Nick Verigakis       the Horde3D community
+    Thomas Ruf              Ronny Chevalier                         github:rlyeh
+    Janez Zemva             John Bartholomew   Michal Cichon        github:romigrou
+    Jonathan Blow           Ken Hamada         Tero Hanninen        github:svdijk
+    Eugene Golushkov        Laurent Gomila     Cort Stratton        github:snagar
+    Aruelien Pocheville     Sergio Gonzalez    Thibault Reuille     github:Zelex
+    Cass Everitt            Ryamond Barbiero                        github:grim210
+    Paul Du Bois            Engin Manap        Aldo Culquicondor    github:sammyhw
+    Philipp Wiesemann       Dale Weiler        Oriol Ferrer Mesia   github:phprus
+    Josh Tobin              Neil Bickford      Matthew Gregan       github:poppolopoppo
+    Julian Raschke          Gregory Mullen     Christian Floisand   github:darealshinji
+    Baldur Karlsson         Kevin Schmidt      JR Smith             github:Michaelangel007
+                            Brad Weinberger    Matvey Cherevko      github:mosra
+    Luca Sas                Alexander Veselov  Zack Middleton       [reserved]
+    Ryan C. Gordon          [reserved]                              [reserved]
+                     DO NOT ADD YOUR NAME HERE
+
+                     Jacko Dirks
+
+  To add your name to the credits, pick a random blank space in the middle and fill it.
+  80% of merge conflicts on stb PRs are due to people adding their name at the end
+  of the credits.
+*/
+
+#ifndef STBI_INCLUDE_STB_IMAGE_H
+#define STBI_INCLUDE_STB_IMAGE_H
+
+// DOCUMENTATION
+//
+// Limitations:
+//    - no 12-bit-per-channel JPEG
+//    - no JPEGs with arithmetic coding
+//    - GIF always returns *comp=4
+//
+// Basic usage (see HDR discussion below for HDR usage):
+//    int x,y,n;
+//    unsigned char *data = stbi_load(filename, &x, &y, &n, 0);
+//    // ... process data if not NULL ...
+//    // ... x = width, y = height, n = # 8-bit components per pixel ...
+//    // ... replace '0' with '1'..'4' to force that many components per pixel
+//    // ... but 'n' will always be the number that it would have been if you said 0
+//    stbi_image_free(data);
+//
+// Standard parameters:
+//    int *x                 -- outputs image width in pixels
+//    int *y                 -- outputs image height in pixels
+//    int *channels_in_file  -- outputs # of image components in image file
+//    int desired_channels   -- if non-zero, # of image components requested in result
+//
+// The return value from an image loader is an 'unsigned char *' which points
+// to the pixel data, or NULL on an allocation failure or if the image is
+// corrupt or invalid. The pixel data consists of *y scanlines of *x pixels,
+// with each pixel consisting of N interleaved 8-bit components; the first
+// pixel pointed to is top-left-most in the image. There is no padding between
+// image scanlines or between pixels, regardless of format. The number of
+// components N is 'desired_channels' if desired_channels is non-zero, or
+// *channels_in_file otherwise. If desired_channels is non-zero,
+// *channels_in_file has the number of components that _would_ have been
+// output otherwise. E.g. if you set desired_channels to 4, you will always
+// get RGBA output, but you can check *channels_in_file to see if it's trivially
+// opaque because e.g. there were only 3 channels in the source image.
+//
+// An output image with N components has the following components interleaved
+// in this order in each pixel:
+//
+//     N=#comp     components
+//       1           grey
+//       2           grey, alpha
+//       3           red, green, blue
+//       4           red, green, blue, alpha
+//
+// If image loading fails for any reason, the return value will be NULL,
+// and *x, *y, *channels_in_file will be unchanged. The function
+// stbi_failure_reason() can be queried for an extremely brief, end-user
+// unfriendly explanation of why the load failed. Define STBI_NO_FAILURE_STRINGS
+// to avoid compiling these strings at all, and STBI_FAILURE_USERMSG to get slightly
+// more user-friendly ones.
+//
+// Paletted PNG, BMP, GIF, and PIC images are automatically depalettized.
+//
+// To query the width, height and component count of an image without having to
+// decode the full file, you can use the stbi_info family of functions:
+//
+//   int x,y,n,ok;
+//   ok = stbi_info(filename, &x, &y, &n);
+//   // returns ok=1 and sets x, y, n if image is a supported format,
+//   // 0 otherwise.
+//
+// Note that stb_image pervasively uses ints in its public API for sizes,
+// including sizes of memory buffers. This is now part of the API and thus
+// hard to change without causing breakage. As a result, the various image
+// loaders all have certain limits on image size; these differ somewhat
+// by format but generally boil down to either just under 2GB or just under
+// 1GB. When the decoded image would be larger than this, stb_image decoding
+// will fail.
+//
+// Additionally, stb_image will reject image files that have any of their
+// dimensions set to a larger value than the configurable STBI_MAX_DIMENSIONS,
+// which defaults to 2**24 = 16777216 pixels. Due to the above memory limit,
+// the only way to have an image with such dimensions load correctly
+// is for it to have a rather extreme aspect ratio. Either way, the
+// assumption here is that such larger images are likely to be malformed
+// or malicious. If you do need to load an image with individual dimensions
+// larger than that, and it still fits in the overall size limit, you can
+// #define STBI_MAX_DIMENSIONS on your own to be something larger.
+//
+// ===========================================================================
+//
+// UNICODE:
+//
+//   If compiling for Windows and you wish to use Unicode filenames, compile
+//   with
+//       #define STBI_WINDOWS_UTF8
+//   and pass utf8-encoded filenames. Call stbi_convert_wchar_to_utf8 to convert
+//   Windows wchar_t filenames to utf8.
+//
+// ===========================================================================
+//
+// Philosophy
+//
+// stb libraries are designed with the following priorities:
+//
+//    1. easy to use
+//    2. easy to maintain
+//    3. good performance
+//
+// Sometimes I let "good performance" creep up in priority over "easy to maintain",
+// and for best performance I may provide less-easy-to-use APIs that give higher
+// performance, in addition to the easy-to-use ones. Nevertheless, it's important
+// to keep in mind that from the standpoint of you, a client of this library,
+// all you care about is #1 and #3, and stb libraries DO NOT emphasize #3 above all.
+//
+// Some secondary priorities arise directly from the first two, some of which
+// provide more explicit reasons why performance can't be emphasized.
+//
+//    - Portable ("ease of use")
+//    - Small source code footprint ("easy to maintain")
+//    - No dependencies ("ease of use")
+//
+// ===========================================================================
+//
+// I/O callbacks
+//
+// I/O callbacks allow you to read from arbitrary sources, like packaged
+// files or some other source. Data read from callbacks are processed
+// through a small internal buffer (currently 128 bytes) to try to reduce
+// overhead.
+//
+// The three functions you must define are "read" (reads some bytes of data),
+// "skip" (skips some bytes of data), "eof" (reports if the stream is at the end).
+//
+// ===========================================================================
+//
+// SIMD support
+//
+// The JPEG decoder will try to automatically use SIMD kernels on x86 when
+// supported by the compiler. For ARM Neon support, you must explicitly
+// request it.
+//
+// (The old do-it-yourself SIMD API is no longer supported in the current
+// code.)
+//
+// On x86, SSE2 will automatically be used when available based on a run-time
+// test; if not, the generic C versions are used as a fall-back. On ARM targets,
+// the typical path is to have separate builds for NEON and non-NEON devices
+// (at least this is true for iOS and Android). Therefore, the NEON support is
+// toggled by a build flag: define STBI_NEON to get NEON loops.
+//
+// If for some reason you do not want to use any of SIMD code, or if
+// you have issues compiling it, you can disable it entirely by
+// defining STBI_NO_SIMD.
+//
+// ===========================================================================
+//
+// HDR image support   (disable by defining STBI_NO_HDR)
+//
+// stb_image supports loading HDR images in general, and currently the Radiance
+// .HDR file format specifically. You can still load any file through the existing
+// interface; if you attempt to load an HDR file, it will be automatically remapped
+// to LDR, assuming gamma 2.2 and an arbitrary scale factor defaulting to 1;
+// both of these constants can be reconfigured through this interface:
+//
+//     stbi_hdr_to_ldr_gamma(2.2f);
+//     stbi_hdr_to_ldr_scale(1.0f);
+//
+// (note, do not use _inverse_ constants; stbi_image will invert them
+// appropriately).
+//
+// Additionally, there is a new, parallel interface for loading files as
+// (linear) floats to preserve the full dynamic range:
+//
+//    float *data = stbi_loadf(filename, &x, &y, &n, 0);
+//
+// If you load LDR images through this interface, those images will
+// be promoted to floating point values, run through the inverse of
+// constants corresponding to the above:
+//
+//     stbi_ldr_to_hdr_scale(1.0f);
+//     stbi_ldr_to_hdr_gamma(2.2f);
+//
+// Finally, given a filename (or an open file or memory block--see header
+// file for details) containing image data, you can query for the "most
+// appropriate" interface to use (that is, whether the image is HDR or
+// not), using:
+//
+//     stbi_is_hdr(char *filename);
+//
+// ===========================================================================
+//
+// iPhone PNG support:
+//
+// We optionally support converting iPhone-formatted PNGs (which store
+// premultiplied BGRA) back to RGB, even though they're internally encoded
+// differently. To enable this conversion, call
+// stbi_convert_iphone_png_to_rgb(1).
+//
+// Call stbi_set_unpremultiply_on_load(1) as well to force a divide per
+// pixel to remove any premultiplied alpha *only* if the image file explicitly
+// says there's premultiplied data (currently only happens in iPhone images,
+// and only if iPhone convert-to-rgb processing is on).
+//
+// ===========================================================================
+//
+// ADDITIONAL CONFIGURATION
+//
+//  - You can suppress implementation of any of the decoders to reduce
+//    your code footprint by #defining one or more of the following
+//    symbols before creating the implementation.
+//
+//        STBI_NO_JPEG
+//        STBI_NO_PNG
+//        STBI_NO_BMP
+//        STBI_NO_PSD
+//        STBI_NO_TGA
+//        STBI_NO_GIF
+//        STBI_NO_HDR
+//        STBI_NO_PIC
+//        STBI_NO_PNM   (.ppm and .pgm)
+//
+//  - You can request *only* certain decoders and suppress all other ones
+//    (this will be more forward-compatible, as addition of new decoders
+//    doesn't require you to disable them explicitly):
+//
+//        STBI_ONLY_JPEG
+//        STBI_ONLY_PNG
+//        STBI_ONLY_BMP
+//        STBI_ONLY_PSD
+//        STBI_ONLY_TGA
+//        STBI_ONLY_GIF
+//        STBI_ONLY_HDR
+//        STBI_ONLY_PIC
+//        STBI_ONLY_PNM   (.ppm and .pgm)
+//
+//   - If you use STBI_NO_PNG (or _ONLY_ without PNG), and you still
+//     want the zlib decoder to be available, #define STBI_SUPPORT_ZLIB
+//
+//  - If you define STBI_MAX_DIMENSIONS, stb_image will reject images greater
+//    than that size (in either width or height) without further processing.
+//    This is to let programs in the wild set an upper bound to prevent
+//    denial-of-service attacks on untrusted data, as one could generate a
+//    valid image of gigantic dimensions and force stb_image to allocate a
+//    huge block of memory and spend disproportionate time decoding it. By
+//    default this is set to (1 << 24), which is 16777216, but that's still
+//    very big.
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif // STBI_NO_STDIO
+
+#define STBI_VERSION 1
+
+enum {
+    STBI_default = 0, // only used for desired_channels
+
+    STBI_grey = 1,
+    STBI_grey_alpha = 2,
+    STBI_rgb = 3,
+    STBI_rgb_alpha = 4
+};
+
+#include <stdlib.h>
+typedef unsigned char stbi_uc;
+typedef unsigned short stbi_us;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef STBIDEF
+#ifdef STB_IMAGE_STATIC
+#define STBIDEF static
+#else
+#define STBIDEF extern
+#endif
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// PRIMARY API - works on images of any type
+//
+
+//
+// load image by filename, open file, or memory buffer
+//
+
+typedef struct {
+    int (*read)(void * user, char * data,
+                int size);            // fill 'data' with 'size' bytes.  return number of bytes actually read
+    void (*skip)(void * user, int n); // skip the next 'n' bytes, or 'unget' the last -n bytes if negative
+    int (*eof)(void * user);          // returns nonzero if we are at end of file/data
+} stbi_io_callbacks;
+
+////////////////////////////////////
+//
+// 8-bits-per-channel interface
+//
+
+STBIDEF stbi_uc * stbi_load_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file,
+                                        int desired_channels);
+STBIDEF stbi_uc * stbi_load_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y,
+                                           int * channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_uc * stbi_load(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels);
+STBIDEF stbi_uc * stbi_load_from_file(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels);
+// for stbi_load_from_file, file pointer is left pointing immediately after image
+#endif
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc * stbi_load_gif_from_memory(stbi_uc const * buffer, int len, int ** delays, int * x, int * y, int * z,
+                                            int * comp, int req_comp);
+#endif
+
+#ifdef STBI_WINDOWS_UTF8
+STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wchar_t * input);
+#endif
+
+////////////////////////////////////
+//
+// 16-bits-per-channel interface
+//
+
+STBIDEF stbi_us * stbi_load_16_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file,
+                                           int desired_channels);
+STBIDEF stbi_us * stbi_load_16_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y,
+                                              int * channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_us * stbi_load_16(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels);
+STBIDEF stbi_us * stbi_load_from_file_16(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels);
+#endif
+
+////////////////////////////////////
+//
+// float-per-channel interface
+//
+#ifndef STBI_NO_LINEAR
+STBIDEF float * stbi_loadf_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file,
+                                       int desired_channels);
+STBIDEF float * stbi_loadf_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * channels_in_file,
+                                          int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF float * stbi_loadf(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels);
+STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels);
+#endif
+#endif
+
+#ifndef STBI_NO_HDR
+STBIDEF void stbi_hdr_to_ldr_gamma(float gamma);
+STBIDEF void stbi_hdr_to_ldr_scale(float scale);
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_LINEAR
+STBIDEF void stbi_ldr_to_hdr_gamma(float gamma);
+STBIDEF void stbi_ldr_to_hdr_scale(float scale);
+#endif // STBI_NO_LINEAR
+
+// stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR
+STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const * clbk, void * user);
+STBIDEF int stbi_is_hdr_from_memory(stbi_uc const * buffer, int len);
+#ifndef STBI_NO_STDIO
+STBIDEF int stbi_is_hdr(char const * filename);
+STBIDEF int stbi_is_hdr_from_file(FILE * f);
+#endif // STBI_NO_STDIO
+
+// get a VERY brief reason for failure
+// on most compilers (and ALL modern mainstream compilers) this is threadsafe
+STBIDEF const char * stbi_failure_reason(void);
+
+// free the loaded image -- this is just free()
+STBIDEF void stbi_image_free(void * retval_from_stbi_load);
+
+// get image dimensions & components without fully decoding
+STBIDEF int stbi_info_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp);
+STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp);
+STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const * buffer, int len);
+STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * clbk, void * user);
+
+#ifndef STBI_NO_STDIO
+STBIDEF int stbi_info(char const * filename, int * x, int * y, int * comp);
+STBIDEF int stbi_info_from_file(FILE * f, int * x, int * y, int * comp);
+STBIDEF int stbi_is_16_bit(char const * filename);
+STBIDEF int stbi_is_16_bit_from_file(FILE * f);
+#endif
+
+// for image formats that explicitly notate that they have premultiplied alpha,
+// we just return the colors as stored in the file. set this flag to force
+// unpremultiplication. results are undefined if the unpremultiply overflow.
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply);
+
+// indicate whether we should process iphone images back to canonical format,
+// or just pass them through "as-is"
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert);
+
+// flip the image vertically, so the first pixel in the output array is the bottom left
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip);
+
+// as above, but only applies to images loaded on the thread that calls the function
+// this function is only available if your compiler supports thread-local variables;
+// calling it will fail to link if your compiler doesn't
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply);
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert);
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip);
+
+// ZLIB client - used by PNG, available for other purposes
+
+STBIDEF char * stbi_zlib_decode_malloc_guesssize(const char * buffer, int len, int initial_size, int * outlen);
+STBIDEF char * stbi_zlib_decode_malloc_guesssize_headerflag(const char * buffer, int len, int initial_size, int * outlen,
+                                                            int parse_header);
+STBIDEF char * stbi_zlib_decode_malloc(const char * buffer, int len, int * outlen);
+STBIDEF int stbi_zlib_decode_buffer(char * obuffer, int olen, const char * ibuffer, int ilen);
+
+STBIDEF char * stbi_zlib_decode_noheader_malloc(const char * buffer, int len, int * outlen);
+STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const char * ibuffer, int ilen);
+
+#ifdef __cplusplus
+}
+#endif
+
+//
+//
+////   end header file   /////////////////////////////////////////////////////
+#endif // STBI_INCLUDE_STB_IMAGE_H
+
+#ifdef STB_IMAGE_IMPLEMENTATION
+
+#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) || defined(STBI_ONLY_TGA) ||                   \
+    defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) ||                    \
+    defined(STBI_ONLY_PNM) || defined(STBI_ONLY_ZLIB)
+#ifndef STBI_ONLY_JPEG
+#define STBI_NO_JPEG
+#endif
+#ifndef STBI_ONLY_PNG
+#define STBI_NO_PNG
+#endif
+#ifndef STBI_ONLY_BMP
+#define STBI_NO_BMP
+#endif
+#ifndef STBI_ONLY_PSD
+#define STBI_NO_PSD
+#endif
+#ifndef STBI_ONLY_TGA
+#define STBI_NO_TGA
+#endif
+#ifndef STBI_ONLY_GIF
+#define STBI_NO_GIF
+#endif
+#ifndef STBI_ONLY_HDR
+#define STBI_NO_HDR
+#endif
+#ifndef STBI_ONLY_PIC
+#define STBI_NO_PIC
+#endif
+#ifndef STBI_ONLY_PNM
+#define STBI_NO_PNM
+#endif
+#endif
+
+#if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB)
+#define STBI_NO_ZLIB
+#endif
+
+#include <limits.h>
+#include <stdarg.h>
+#include <stddef.h> // ptrdiff_t on osx
+#include <stdlib.h>
+#include <string.h>
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
+#include <math.h> // ldexp, pow
+#endif
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif
+
+#ifndef STBI_ASSERT
+#include <assert.h>
+#define STBI_ASSERT(x) assert(x)
+#endif
+
+#ifdef __cplusplus
+#define STBI_EXTERN extern "C"
+#else
+#define STBI_EXTERN extern
+#endif
+
+#ifndef _MSC_VER
+#ifdef __cplusplus
+#define stbi_inline inline
+#else
+#define stbi_inline
+#endif
+#else
+#define stbi_inline __forceinline
+#endif
+
+#ifndef STBI_NO_THREAD_LOCALS
+#if defined(__cplusplus) && __cplusplus >= 201103L
+#define STBI_THREAD_LOCAL thread_local
+#elif defined(__GNUC__) && __GNUC__ < 5
+#define STBI_THREAD_LOCAL __thread
+#elif defined(_MSC_VER)
+#define STBI_THREAD_LOCAL __declspec(thread)
+#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__)
+#define STBI_THREAD_LOCAL _Thread_local
+#endif
+
+#ifndef STBI_THREAD_LOCAL
+#if defined(__GNUC__)
+#define STBI_THREAD_LOCAL __thread
+#endif
+#endif
+#endif
+
+#if defined(_MSC_VER) || defined(__SYMBIAN32__)
+typedef unsigned short stbi__uint16;
+typedef signed short stbi__int16;
+typedef unsigned int stbi__uint32;
+typedef signed int stbi__int32;
+#else
+#include <stdint.h>
+typedef uint16_t stbi__uint16;
+typedef int16_t stbi__int16;
+typedef uint32_t stbi__uint32;
+typedef int32_t stbi__int32;
+#endif
+
+// should produce compiler error if size is wrong
+typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1];
+
+#ifdef _MSC_VER
+#define STBI_NOTUSED(v) (void)(v)
+#else
+#define STBI_NOTUSED(v) (void)sizeof(v)
+#endif
+
+#ifdef _MSC_VER
+#define STBI_HAS_LROTL
+#endif
+
+#ifdef STBI_HAS_LROTL
+#define stbi_lrot(x, y) _lrotl(x, y)
+#else
+#define stbi_lrot(x, y) (((x) << (y)) | ((x) >> (-(y)&31)))
+#endif
+
+#if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED))
+// ok
+#elif !defined(STBI_MALLOC) && !defined(STBI_FREE) && !defined(STBI_REALLOC) && !defined(STBI_REALLOC_SIZED)
+// ok
+#else
+#error "Must define all or none of STBI_MALLOC, STBI_FREE, and STBI_REALLOC (or STBI_REALLOC_SIZED)."
+#endif
+
+#ifndef STBI_MALLOC
+#define STBI_MALLOC(sz) malloc(sz)
+#define STBI_REALLOC(p, newsz) realloc(p, newsz)
+#define STBI_FREE(p) free(p)
+#endif
+
+#ifndef STBI_REALLOC_SIZED
+#define STBI_REALLOC_SIZED(p, oldsz, newsz) STBI_REALLOC(p, newsz)
+#endif
+
+// x86/x64 detection
+#if defined(__x86_64__) || defined(_M_X64)
+#define STBI__X64_TARGET
+#elif defined(__i386) || defined(_M_IX86)
+#define STBI__X86_TARGET
+#endif
+
+#if defined(__GNUC__) && defined(STBI__X86_TARGET) && !defined(__SSE2__) && !defined(STBI_NO_SIMD)
+// gcc doesn't support sse2 intrinsics unless you compile with -msse2,
+// which in turn means it gets to use SSE2 everywhere. This is unfortunate,
+// but previous attempts to provide the SSE2 functions with runtime
+// detection caused numerous issues. The way architecture extensions are
+// exposed in GCC/Clang is, sadly, not really suited for one-file libs.
+// New behavior: if compiled with -msse2, we use SSE2 without any
+// detection; if not, we don't use it at all.
+#define STBI_NO_SIMD
+#endif
+
+#if defined(__MINGW32__) && defined(STBI__X86_TARGET) && !defined(STBI_MINGW_ENABLE_SSE2) && !defined(STBI_NO_SIMD)
+// Note that __MINGW32__ doesn't actually mean 32-bit, so we have to avoid STBI__X64_TARGET
+//
+// 32-bit MinGW wants ESP to be 16-byte aligned, but this is not in the
+// Windows ABI and VC++ as well as Windows DLLs don't maintain that invariant.
+// As a result, enabling SSE2 on 32-bit MinGW is dangerous when not
+// simultaneously enabling "-mstackrealign".
+//
+// See https://github.com/nothings/stb/issues/81 for more information.
+//
+// So default to no SSE2 on 32-bit MinGW. If you've read this far and added
+// -mstackrealign to your build settings, feel free to #define STBI_MINGW_ENABLE_SSE2.
+#define STBI_NO_SIMD
+#endif
+
+#if !defined(STBI_NO_SIMD) && (defined(STBI__X86_TARGET) || defined(STBI__X64_TARGET))
+#define STBI_SSE2
+#include <emmintrin.h>
+
+#ifdef _MSC_VER
+
+#if _MSC_VER >= 1400 // not VC6
+#include <intrin.h>  // __cpuid
+static int stbi__cpuid3(void) {
+    int info[4];
+    __cpuid(info, 1);
+    return info[3];
+}
+#else
+static int stbi__cpuid3(void) {
+    int res;
+    __asm {
+      mov  eax,1
+      cpuid
+      mov  res,edx
+    }
+    return res;
+}
+#endif
+
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void) {
+    int info3 = stbi__cpuid3();
+    return ((info3 >> 26) & 1) != 0;
+}
+#endif
+
+#else // assume GCC-style if not VC++
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void) {
+    // If we're even attempting to compile this on GCC/Clang, that means
+    // -msse2 is on, which means the compiler is allowed to use SSE2
+    // instructions at will, and so are we.
+    return 1;
+}
+#endif
+
+#endif
+#endif
+
+// ARM NEON
+#if defined(STBI_NO_SIMD) && defined(STBI_NEON)
+#undef STBI_NEON
+#endif
+
+#ifdef STBI_NEON
+#include <arm_neon.h>
+#ifdef _MSC_VER
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+#else
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+#endif
+#endif
+
+#ifndef STBI_SIMD_ALIGN
+#define STBI_SIMD_ALIGN(type, name) type name
+#endif
+
+#ifndef STBI_MAX_DIMENSIONS
+#define STBI_MAX_DIMENSIONS (1 << 24)
+#endif
+
+///////////////////////////////////////////////
+//
+//  stbi__context struct and start_xxx functions
+
+// stbi__context structure is our basic context used by all images, so it
+// contains all the IO context, plus some basic image information
+typedef struct {
+    stbi__uint32 img_x, img_y;
+    int img_n, img_out_n;
+
+    stbi_io_callbacks io;
+    void * io_user_data;
+
+    int read_from_callbacks;
+    int buflen;
+    stbi_uc buffer_start[128];
+    int callback_already_read;
+
+    stbi_uc *img_buffer, *img_buffer_end;
+    stbi_uc *img_buffer_original, *img_buffer_original_end;
+} stbi__context;
+
+static void stbi__refill_buffer(stbi__context * s);
+
+// initialize a memory-decode context
+static void stbi__start_mem(stbi__context * s, stbi_uc const * buffer, int len) {
+    s->io.read = NULL;
+    s->read_from_callbacks = 0;
+    s->callback_already_read = 0;
+    s->img_buffer = s->img_buffer_original = (stbi_uc *)buffer;
+    s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *)buffer + len;
+}
+
+// initialize a callback-based context
+static void stbi__start_callbacks(stbi__context * s, stbi_io_callbacks * c, void * user) {
+    s->io = *c;
+    s->io_user_data = user;
+    s->buflen = sizeof(s->buffer_start);
+    s->read_from_callbacks = 1;
+    s->callback_already_read = 0;
+    s->img_buffer = s->img_buffer_original = s->buffer_start;
+    stbi__refill_buffer(s);
+    s->img_buffer_original_end = s->img_buffer_end;
+}
+
+#ifndef STBI_NO_STDIO
+
+static int stbi__stdio_read(void * user, char * data, int size) { return (int)fread(data, 1, size, (FILE *)user); }
+
+static void stbi__stdio_skip(void * user, int n) {
+    int ch;
+    fseek((FILE *)user, n, SEEK_CUR);
+    ch = fgetc((FILE *)user); /* have to read a byte to reset feof()'s flag */
+    if (ch != EOF) {
+        ungetc(ch, (FILE *)user); /* push byte back onto stream if valid. */
+    }
+}
+
+static int stbi__stdio_eof(void * user) { return feof((FILE *)user) || ferror((FILE *)user); }
+
+static stbi_io_callbacks stbi__stdio_callbacks = {
+    stbi__stdio_read,
+    stbi__stdio_skip,
+    stbi__stdio_eof,
+};
+
+static void stbi__start_file(stbi__context * s, FILE * f) { stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *)f); }
+
+// static void stop_file(stbi__context *s) { }
+
+#endif // !STBI_NO_STDIO
+
+static void stbi__rewind(stbi__context * s) {
+    // conceptually rewind SHOULD rewind to the beginning of the stream,
+    // but we just rewind to the beginning of the initial buffer, because
+    // we only use it after doing 'test', which only ever looks at at most 92 bytes
+    s->img_buffer = s->img_buffer_original;
+    s->img_buffer_end = s->img_buffer_original_end;
+}
+
+enum { STBI_ORDER_RGB, STBI_ORDER_BGR };
+
+typedef struct {
+    int bits_per_channel;
+    int num_channels;
+    int channel_order;
+} stbi__result_info;
+
+#ifndef STBI_NO_JPEG
+static int stbi__jpeg_test(stbi__context * s);
+static void * stbi__jpeg_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp);
+#endif
+
+#ifndef STBI_NO_PNG
+static int stbi__png_test(stbi__context * s);
+static void * stbi__png_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static int stbi__png_info(stbi__context * s, int * x, int * y, int * comp);
+static int stbi__png_is16(stbi__context * s);
+#endif
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_test(stbi__context * s);
+static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static int stbi__bmp_info(stbi__context * s, int * x, int * y, int * comp);
+#endif
+
+#ifndef STBI_NO_TGA
+static int stbi__tga_test(stbi__context * s);
+static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp);
+#endif
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_test(stbi__context * s);
+static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc);
+static int stbi__psd_info(stbi__context * s, int * x, int * y, int * comp);
+static int stbi__psd_is16(stbi__context * s);
+#endif
+
+#ifndef STBI_NO_HDR
+static int stbi__hdr_test(stbi__context * s);
+static float * stbi__hdr_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static int stbi__hdr_info(stbi__context * s, int * x, int * y, int * comp);
+#endif
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_test(stbi__context * s);
+static void * stbi__pic_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp);
+#endif
+
+#ifndef STBI_NO_GIF
+static int stbi__gif_test(stbi__context * s);
+static void * stbi__gif_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static void * stbi__load_gif_main(stbi__context * s, int ** delays, int * x, int * y, int * z, int * comp, int req_comp);
+static int stbi__gif_info(stbi__context * s, int * x, int * y, int * comp);
+#endif
+
+#ifndef STBI_NO_PNM
+static int stbi__pnm_test(stbi__context * s);
+static void * stbi__pnm_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri);
+static int stbi__pnm_info(stbi__context * s, int * x, int * y, int * comp);
+static int stbi__pnm_is16(stbi__context * s);
+#endif
+
+static
+#ifdef STBI_THREAD_LOCAL
+    STBI_THREAD_LOCAL
+#endif
+    const char * stbi__g_failure_reason;
+
+STBIDEF const char * stbi_failure_reason(void) { return stbi__g_failure_reason; }
+
+#ifndef STBI_NO_FAILURE_STRINGS
+static int stbi__err(const char * str) {
+    stbi__g_failure_reason = str;
+    return 0;
+}
+#endif
+
+static void * stbi__malloc(size_t size) { return STBI_MALLOC(size); }
+
+// stb_image uses ints pervasively, including for offset calculations.
+// therefore the largest decoded image size we can support with the
+// current code, even on 64-bit targets, is INT_MAX. this is not a
+// significant limitation for the intended use case.
+//
+// we do, however, need to make sure our size calculations don't
+// overflow. hence a few helper functions for size calculations that
+// multiply integers together, making sure that they're non-negative
+// and no overflow occurs.
+
+// return 1 if the sum is valid, 0 on overflow.
+// negative terms are considered invalid.
+static int stbi__addsizes_valid(int a, int b) {
+    if (b < 0)
+        return 0;
+    // now 0 <= b <= INT_MAX, hence also
+    // 0 <= INT_MAX - b <= INTMAX.
+    // And "a + b <= INT_MAX" (which might overflow) is the
+    // same as a <= INT_MAX - b (no overflow)
+    return a <= INT_MAX - b;
+}
+
+// returns 1 if the product is valid, 0 on overflow.
+// negative factors are considered invalid.
+static int stbi__mul2sizes_valid(int a, int b) {
+    if (a < 0 || b < 0)
+        return 0;
+    if (b == 0)
+        return 1; // mul-by-0 is always safe
+    // portable way to check for no overflows in a*b
+    return a <= INT_MAX / b;
+}
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad2sizes_valid(int a, int b, int add) {
+    return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a * b, add);
+}
+#endif
+
+// returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad3sizes_valid(int a, int b, int c, int add) {
+    return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a * b, c) && stbi__addsizes_valid(a * b * c, add);
+}
+
+// returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add) {
+    return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a * b, c) && stbi__mul2sizes_valid(a * b * c, d) &&
+           stbi__addsizes_valid(a * b * c * d, add);
+}
+#endif
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// mallocs with size overflow checking
+static void * stbi__malloc_mad2(int a, int b, int add) {
+    if (!stbi__mad2sizes_valid(a, b, add))
+        return NULL;
+    return stbi__malloc(a * b + add);
+}
+#endif
+
+static void * stbi__malloc_mad3(int a, int b, int c, int add) {
+    if (!stbi__mad3sizes_valid(a, b, c, add))
+        return NULL;
+    return stbi__malloc(a * b * c + add);
+}
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static void * stbi__malloc_mad4(int a, int b, int c, int d, int add) {
+    if (!stbi__mad4sizes_valid(a, b, c, d, add))
+        return NULL;
+    return stbi__malloc(a * b * c * d + add);
+}
+#endif
+
+// returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow.
+static int stbi__addints_valid(int a, int b) {
+    if ((a >= 0) != (b >= 0))
+        return 1; // a and b have different signs, so no overflow
+    if (a < 0 && b < 0)
+        return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0.
+    return a <= INT_MAX - b;
+}
+
+// returns 1 if the product of two signed shorts is valid, 0 on overflow.
+static int stbi__mul2shorts_valid(short a, short b) {
+    if (b == 0 || b == -1)
+        return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow
+    if ((a >= 0) == (b >= 0))
+        return a <= SHRT_MAX / b; // product is positive, so similar to mul2sizes_valid
+    if (b < 0)
+        return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN
+    return a >= SHRT_MIN / b;
+}
+
+// stbi__err - error
+// stbi__errpf - error returning pointer to float
+// stbi__errpuc - error returning pointer to unsigned char
+
+#ifdef STBI_NO_FAILURE_STRINGS
+#define stbi__err(x, y) 0
+#elif defined(STBI_FAILURE_USERMSG)
+#define stbi__err(x, y) stbi__err(y)
+#else
+#define stbi__err(x, y) stbi__err(x)
+#endif
+
+#define stbi__errpf(x, y) ((float *)(size_t)(stbi__err(x, y) ? NULL : NULL))
+#define stbi__errpuc(x, y) ((unsigned char *)(size_t)(stbi__err(x, y) ? NULL : NULL))
+
+STBIDEF void stbi_image_free(void * retval_from_stbi_load) { STBI_FREE(retval_from_stbi_load); }
+
+#ifndef STBI_NO_LINEAR
+static float * stbi__ldr_to_hdr(stbi_uc * data, int x, int y, int comp);
+#endif
+
+#ifndef STBI_NO_HDR
+static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp);
+#endif
+
+static int stbi__vertically_flip_on_load_global = 0;
+
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip) {
+    stbi__vertically_flip_on_load_global = flag_true_if_should_flip;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__vertically_flip_on_load stbi__vertically_flip_on_load_global
+#else
+static STBI_THREAD_LOCAL int stbi__vertically_flip_on_load_local, stbi__vertically_flip_on_load_set;
+
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip) {
+    stbi__vertically_flip_on_load_local = flag_true_if_should_flip;
+    stbi__vertically_flip_on_load_set = 1;
+}
+
+#define stbi__vertically_flip_on_load                                                                                          \
+    (stbi__vertically_flip_on_load_set ? stbi__vertically_flip_on_load_local : stbi__vertically_flip_on_load_global)
+#endif // STBI_THREAD_LOCAL
+
+static void * stbi__load_main(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc) {
+    memset(ri, 0, sizeof(*ri));         // make sure it's initialized if we add new fields
+    ri->bits_per_channel = 8;           // default is 8 so most paths don't have to be changed
+    ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order
+    ri->num_channels = 0;
+
+// test the formats with a very explicit header first (at least a FOURCC
+// or distinctive magic number first)
+#ifndef STBI_NO_PNG
+    if (stbi__png_test(s))
+        return stbi__png_load(s, x, y, comp, req_comp, ri);
+#endif
+#ifndef STBI_NO_BMP
+    if (stbi__bmp_test(s))
+        return stbi__bmp_load(s, x, y, comp, req_comp, ri);
+#endif
+#ifndef STBI_NO_GIF
+    if (stbi__gif_test(s))
+        return stbi__gif_load(s, x, y, comp, req_comp, ri);
+#endif
+#ifndef STBI_NO_PSD
+    if (stbi__psd_test(s))
+        return stbi__psd_load(s, x, y, comp, req_comp, ri, bpc);
+#else
+    STBI_NOTUSED(bpc);
+#endif
+#ifndef STBI_NO_PIC
+    if (stbi__pic_test(s))
+        return stbi__pic_load(s, x, y, comp, req_comp, ri);
+#endif
+
+// then the formats that can end up attempting to load with just 1 or 2
+// bytes matching expectations; these are prone to false positives, so
+// try them later
+#ifndef STBI_NO_JPEG
+    if (stbi__jpeg_test(s))
+        return stbi__jpeg_load(s, x, y, comp, req_comp, ri);
+#endif
+#ifndef STBI_NO_PNM
+    if (stbi__pnm_test(s))
+        return stbi__pnm_load(s, x, y, comp, req_comp, ri);
+#endif
+
+#ifndef STBI_NO_HDR
+    if (stbi__hdr_test(s)) {
+        float * hdr = stbi__hdr_load(s, x, y, comp, req_comp, ri);
+        return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp);
+    }
+#endif
+
+#ifndef STBI_NO_TGA
+    // test tga last because it's a crappy test!
+    if (stbi__tga_test(s))
+        return stbi__tga_load(s, x, y, comp, req_comp, ri);
+#endif
+
+    return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static stbi_uc * stbi__convert_16_to_8(stbi__uint16 * orig, int w, int h, int channels) {
+    int i;
+    int img_len = w * h * channels;
+    stbi_uc * reduced;
+
+    reduced = (stbi_uc *)stbi__malloc(img_len);
+    if (reduced == NULL)
+        return stbi__errpuc("outofmem", "Out of memory");
+
+    for (i = 0; i < img_len; ++i)
+        reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling
+
+    STBI_FREE(orig);
+    return reduced;
+}
+
+static stbi__uint16 * stbi__convert_8_to_16(stbi_uc * orig, int w, int h, int channels) {
+    int i;
+    int img_len = w * h * channels;
+    stbi__uint16 * enlarged;
+
+    enlarged = (stbi__uint16 *)stbi__malloc(img_len * 2);
+    if (enlarged == NULL)
+        return (stbi__uint16 *)stbi__errpuc("outofmem", "Out of memory");
+
+    for (i = 0; i < img_len; ++i)
+        enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff
+
+    STBI_FREE(orig);
+    return enlarged;
+}
+
+static void stbi__vertical_flip(void * image, int w, int h, int bytes_per_pixel) {
+    int row;
+    size_t bytes_per_row = (size_t)w * bytes_per_pixel;
+    stbi_uc temp[2048];
+    stbi_uc * bytes = (stbi_uc *)image;
+
+    for (row = 0; row < (h >> 1); row++) {
+        stbi_uc * row0 = bytes + row * bytes_per_row;
+        stbi_uc * row1 = bytes + (h - row - 1) * bytes_per_row;
+        // swap row0 with row1
+        size_t bytes_left = bytes_per_row;
+        while (bytes_left) {
+            size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp);
+            memcpy(temp, row0, bytes_copy);
+            memcpy(row0, row1, bytes_copy);
+            memcpy(row1, temp, bytes_copy);
+            row0 += bytes_copy;
+            row1 += bytes_copy;
+            bytes_left -= bytes_copy;
+        }
+    }
+}
+
+#ifndef STBI_NO_GIF
+static void stbi__vertical_flip_slices(void * image, int w, int h, int z, int bytes_per_pixel) {
+    int slice;
+    int slice_size = w * h * bytes_per_pixel;
+
+    stbi_uc * bytes = (stbi_uc *)image;
+    for (slice = 0; slice < z; ++slice) {
+        stbi__vertical_flip(bytes, w, h, bytes_per_pixel);
+        bytes += slice_size;
+    }
+}
+#endif
+
+static unsigned char * stbi__load_and_postprocess_8bit(stbi__context * s, int * x, int * y, int * comp, int req_comp) {
+    stbi__result_info ri;
+    void * result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8);
+
+    if (result == NULL)
+        return NULL;
+
+    // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+    STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+    if (ri.bits_per_channel != 8) {
+        result = stbi__convert_16_to_8((stbi__uint16 *)result, *x, *y, req_comp == 0 ? *comp : req_comp);
+        ri.bits_per_channel = 8;
+    }
+
+    // @TODO: move stbi__convert_format to here
+
+    if (stbi__vertically_flip_on_load) {
+        int channels = req_comp ? req_comp : *comp;
+        stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc));
+    }
+
+    return (unsigned char *)result;
+}
+
+static stbi__uint16 * stbi__load_and_postprocess_16bit(stbi__context * s, int * x, int * y, int * comp, int req_comp) {
+    stbi__result_info ri;
+    void * result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16);
+
+    if (result == NULL)
+        return NULL;
+
+    // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+    STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+    if (ri.bits_per_channel != 16) {
+        result = stbi__convert_8_to_16((stbi_uc *)result, *x, *y, req_comp == 0 ? *comp : req_comp);
+        ri.bits_per_channel = 16;
+    }
+
+    // @TODO: move stbi__convert_format16 to here
+    // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision
+
+    if (stbi__vertically_flip_on_load) {
+        int channels = req_comp ? req_comp : *comp;
+        stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16));
+    }
+
+    return (stbi__uint16 *)result;
+}
+
+#if !defined(STBI_NO_HDR) && !defined(STBI_NO_LINEAR)
+static void stbi__float_postprocess(float * result, int * x, int * y, int * comp, int req_comp) {
+    if (stbi__vertically_flip_on_load && result != NULL) {
+        int channels = req_comp ? req_comp : *comp;
+        stbi__vertical_flip(result, *x, *y, channels * sizeof(float));
+    }
+}
+#endif
+
+#ifndef STBI_NO_STDIO
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char * str,
+                                                                    int cbmb, wchar_t * widestr, int cchwide);
+STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags,
+                                                                    const wchar_t * widestr, int cchwide, char * str, int cbmb,
+                                                                    const char * defchar, int * used_default);
+#endif
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wchar_t * input) {
+    return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int)bufferlen, NULL, NULL);
+}
+#endif
+
+static FILE * stbi__fopen(char const * filename, char const * mode) {
+    FILE * f;
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+    wchar_t wMode[64];
+    wchar_t wFilename[1024];
+    if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename) / sizeof(*wFilename)))
+        return 0;
+
+    if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode) / sizeof(*wMode)))
+        return 0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+    if (0 != _wfopen_s(&f, wFilename, wMode))
+        f = 0;
+#else
+    f = _wfopen(wFilename, wMode);
+#endif
+
+#elif defined(_MSC_VER) && _MSC_VER >= 1400
+    if (0 != fopen_s(&f, filename, mode))
+        f = 0;
+#else
+    f = fopen(filename, mode);
+#endif
+    return f;
+}
+
+STBIDEF stbi_uc * stbi_load(char const * filename, int * x, int * y, int * comp, int req_comp) {
+    FILE * f = stbi__fopen(filename, "rb");
+    unsigned char * result;
+    if (!f)
+        return stbi__errpuc("can't fopen", "Unable to open file");
+    result = stbi_load_from_file(f, x, y, comp, req_comp);
+    fclose(f);
+    return result;
+}
+
+STBIDEF stbi_uc * stbi_load_from_file(FILE * f, int * x, int * y, int * comp, int req_comp) {
+    unsigned char * result;
+    stbi__context s;
+    stbi__start_file(&s, f);
+    result = stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp);
+    if (result) {
+        // need to 'unget' all the characters in the IO buffer
+        fseek(f, -(int)(s.img_buffer_end - s.img_buffer), SEEK_CUR);
+    }
+    return result;
+}
+
+STBIDEF stbi__uint16 * stbi_load_from_file_16(FILE * f, int * x, int * y, int * comp, int req_comp) {
+    stbi__uint16 * result;
+    stbi__context s;
+    stbi__start_file(&s, f);
+    result = stbi__load_and_postprocess_16bit(&s, x, y, comp, req_comp);
+    if (result) {
+        // need to 'unget' all the characters in the IO buffer
+        fseek(f, -(int)(s.img_buffer_end - s.img_buffer), SEEK_CUR);
+    }
+    return result;
+}
+
+STBIDEF stbi_us * stbi_load_16(char const * filename, int * x, int * y, int * comp, int req_comp) {
+    FILE * f = stbi__fopen(filename, "rb");
+    stbi__uint16 * result;
+    if (!f)
+        return (stbi_us *)stbi__errpuc("can't fopen", "Unable to open file");
+    result = stbi_load_from_file_16(f, x, y, comp, req_comp);
+    fclose(f);
+    return result;
+}
+
+#endif //! STBI_NO_STDIO
+
+STBIDEF stbi_us * stbi_load_16_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file,
+                                           int desired_channels) {
+    stbi__context s;
+    stbi__start_mem(&s, buffer, len);
+    return stbi__load_and_postprocess_16bit(&s, x, y, channels_in_file, desired_channels);
+}
+
+STBIDEF stbi_us * stbi_load_16_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y,
+                                              int * channels_in_file, int desired_channels) {
+    stbi__context s;
+    stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
+    return stbi__load_and_postprocess_16bit(&s, x, y, channels_in_file, desired_channels);
+}
+
+STBIDEF stbi_uc * stbi_load_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp, int req_comp) {
+    stbi__context s;
+    stbi__start_mem(&s, buffer, len);
+    return stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp);
+}
+
+STBIDEF stbi_uc * stbi_load_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp,
+                                           int req_comp) {
+    stbi__context s;
+    stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
+    return stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp);
+}
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc * stbi_load_gif_from_memory(stbi_uc const * buffer, int len, int ** delays, int * x, int * y, int * z,
+                                            int * comp, int req_comp) {
+    unsigned char * result;
+    stbi__context s;
+    stbi__start_mem(&s, buffer, len);
+
+    result = (unsigned char *)stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp);
+    if (stbi__vertically_flip_on_load) {
+        stbi__vertical_flip_slices(result, *x, *y, *z, *comp);
+    }
+
+    return result;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float * stbi__loadf_main(stbi__context * s, int * x, int * y, int * comp, int req_comp) {
+    unsigned char * data;
+#ifndef STBI_NO_HDR
+    if (stbi__hdr_test(s)) {
+        stbi__result_info ri;
+        float * hdr_data = stbi__hdr_load(s, x, y, comp, req_comp, &ri);
+        if (hdr_data)
+            stbi__float_postprocess(hdr_data, x, y, comp, req_comp);
+        return hdr_data;
+    }
+#endif
+    data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp);
+    if (data)
+        return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp);
+    return stbi__errpf("unknown image type", "Image not of any known type, or corrupt");
+}
+
+STBIDEF float * stbi_loadf_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp, int req_comp) {
+    stbi__context s;
+    stbi__start_mem(&s, buffer, len);
+    return stbi__loadf_main(&s, x, y, comp, req_comp);
+}
+
+STBIDEF float * stbi_loadf_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp,
+                                          int req_comp) {
+    stbi__context s;
+    stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
+    return stbi__loadf_main(&s, x, y, comp, req_comp);
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF float * stbi_loadf(char const * filename, int * x, int * y, int * comp, int req_comp) {
+    float * result;
+    FILE * f = stbi__fopen(filename, "rb");
+    if (!f)
+        return stbi__errpf("can't fopen", "Unable to open file");
+    result = stbi_loadf_from_file(f, x, y, comp, req_comp);
+    fclose(f);
+    return result;
+}
+
+STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * comp, int req_comp) {
+    stbi__context s;
+    stbi__start_file(&s, f);
+    return stbi__loadf_main(&s, x, y, comp, req_comp);
+}
+#endif // !STBI_NO_STDIO
+
+#endif // !STBI_NO_LINEAR
+
+// these is-hdr-or-not is defined independent of whether STBI_NO_LINEAR is
+// defined, for API simplicity; if STBI_NO_LINEAR is defined, it always
+// reports false!
+
+STBIDEF int stbi_is_hdr_from_memory(stbi_uc const * buffer, int len) {
+#ifndef STBI_NO_HDR
+    stbi__context s;
+    stbi__start_mem(&s, buffer, len);
+    return stbi__hdr_test(&s);
+#else
+    STBI_NOTUSED(buffer);
+    STBI_NOTUSED(len);
+    return 0;
+#endif
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int stbi_is_hdr(char const * filename) {
+    FILE * f = stbi__fopen(filename, "rb");
+    int result = 0;
+    if (f) {
+        result = stbi_is_hdr_from_file(f);
+        fclose(f);
+    }
+    return result;
+}
+
+STBIDEF int stbi_is_hdr_from_file(FILE * f) {
+#ifndef STBI_NO_HDR
+    long pos = ftell(f);
+    int res;
+    stbi__context s;
+    stbi__start_file(&s, f);
+    res = stbi__hdr_test(&s);
+    fseek(f, pos, SEEK_SET);
+    return res;
+#else
+    STBI_NOTUSED(f);
+    return 0;
+#endif
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const * clbk, void * user) {
+#ifndef STBI_NO_HDR
+    stbi__context s;
+    stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
+    return stbi__hdr_test(&s);
+#else
+    STBI_NOTUSED(clbk);
+    STBI_NOTUSED(user);
+    return 0;
+#endif
+}
+
+#ifndef STBI_NO_LINEAR
+static float stbi__l2h_gamma = 2.2f, stbi__l2h_scale = 1.0f;
+
+STBIDEF void stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; }
+STBIDEF void stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; }
+#endif
+
+static float stbi__h2l_gamma_i = 1.0f / 2.2f, stbi__h2l_scale_i = 1.0f;
+
+STBIDEF void stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1 / gamma; }
+STBIDEF void stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1 / scale; }
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// Common code used by all image loaders
+//
+
+enum { STBI__SCAN_load = 0, STBI__SCAN_type, STBI__SCAN_header };
+
+static void stbi__refill_buffer(stbi__context * s) {
+    int n = (s->io.read)(s->io_user_data, (char *)s->buffer_start, s->buflen);
+    s->callback_already_read += (int)(s->img_buffer - s->img_buffer_original);
+    if (n == 0) {
+        // at end of file, treat same as if from memory, but need to handle case
+        // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file
+        s->read_from_callbacks = 0;
+        s->img_buffer = s->buffer_start;
+        s->img_buffer_end = s->buffer_start + 1;
+        *s->img_buffer = 0;
+    } else {
+        s->img_buffer = s->buffer_start;
+        s->img_buffer_end = s->buffer_start + n;
+    }
+}
+
+stbi_inline static stbi_uc stbi__get8(stbi__context * s) {
+    if (s->img_buffer < s->img_buffer_end)
+        return *s->img_buffer++;
+    if (s->read_from_callbacks) {
+        stbi__refill_buffer(s);
+        return *s->img_buffer++;
+    }
+    return 0;
+}
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_HDR) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+stbi_inline static int stbi__at_eof(stbi__context * s) {
+    if (s->io.read) {
+        if (!(s->io.eof)(s->io_user_data))
+            return 0;
+        // if feof() is true, check if buffer = end
+        // special case: we've only got the special 0 character at the end
+        if (s->read_from_callbacks == 0)
+            return 1;
+    }
+
+    return s->img_buffer >= s->img_buffer_end;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) &&   \
+    defined(STBI_NO_GIF) && defined(STBI_NO_PIC)
+// nothing
+#else
+static void stbi__skip(stbi__context * s, int n) {
+    if (n == 0)
+        return; // already there!
+    if (n < 0) {
+        s->img_buffer = s->img_buffer_end;
+        return;
+    }
+    if (s->io.read) {
+        int blen = (int)(s->img_buffer_end - s->img_buffer);
+        if (blen < n) {
+            s->img_buffer = s->img_buffer_end;
+            (s->io.skip)(s->io_user_data, n - blen);
+            return;
+        }
+    }
+    s->img_buffer += n;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_TGA) && defined(STBI_NO_HDR) && defined(STBI_NO_PNM)
+// nothing
+#else
+static int stbi__getn(stbi__context * s, stbi_uc * buffer, int n) {
+    if (s->io.read) {
+        int blen = (int)(s->img_buffer_end - s->img_buffer);
+        if (blen < n) {
+            int res, count;
+
+            memcpy(buffer, s->img_buffer, blen);
+
+            count = (s->io.read)(s->io_user_data, (char *)buffer + blen, n - blen);
+            res = (count == (n - blen));
+            s->img_buffer = s->img_buffer_end;
+            return res;
+        }
+    }
+
+    if (s->img_buffer + n <= s->img_buffer_end) {
+        memcpy(buffer, s->img_buffer, n);
+        s->img_buffer += n;
+        return 1;
+    } else
+        return 0;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static int stbi__get16be(stbi__context * s) {
+    int z = stbi__get8(s);
+    return (z << 8) + stbi__get8(s);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static stbi__uint32 stbi__get32be(stbi__context * s) {
+    stbi__uint32 z = stbi__get16be(s);
+    return (z << 16) + stbi__get16be(s);
+}
+#endif
+
+#if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF)
+// nothing
+#else
+static int stbi__get16le(stbi__context * s) {
+    int z = stbi__get8(s);
+    return z + (stbi__get8(s) << 8);
+}
+#endif
+
+#ifndef STBI_NO_BMP
+static stbi__uint32 stbi__get32le(stbi__context * s) {
+    stbi__uint32 z = stbi__get16le(s);
+    z += (stbi__uint32)stbi__get16le(s) << 16;
+    return z;
+}
+#endif
+
+#define STBI__BYTECAST(x) ((stbi_uc)((x)&255)) // truncate int to byte without warnings
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) &&   \
+    defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+//////////////////////////////////////////////////////////////////////////////
+//
+//  generic converter from built-in img_n to req_comp
+//    individual types do this automatically as much as possible (e.g. jpeg
+//    does all cases internally since it needs to colorspace convert anyway,
+//    and it never has alpha, so very few cases ). png can automatically
+//    interleave an alpha=255 channel, but falls back to this for other cases
+//
+//  assume data buffer is malloced, so malloc a new one and free that one
+//  only failure mode is malloc failing
+
+static stbi_uc stbi__compute_y(int r, int g, int b) { return (stbi_uc)(((r * 77) + (g * 150) + (29 * b)) >> 8); }
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) &&    \
+    defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+static unsigned char * stbi__convert_format(unsigned char * data, int img_n, int req_comp, unsigned int x, unsigned int y) {
+    int i, j;
+    unsigned char * good;
+
+    if (req_comp == img_n)
+        return data;
+    STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+    good = (unsigned char *)stbi__malloc_mad3(req_comp, x, y, 0);
+    if (good == NULL) {
+        STBI_FREE(data);
+        return stbi__errpuc("outofmem", "Out of memory");
+    }
+
+    for (j = 0; j < (int)y; ++j) {
+        unsigned char * src = data + j * x * img_n;
+        unsigned char * dest = good + j * x * req_comp;
+
+#define STBI__COMBO(a, b) ((a)*8 + (b))
+#define STBI__CASE(a, b)                                                                                                       \
+    case STBI__COMBO(a, b):                                                                                                    \
+        for (i = x - 1; i >= 0; --i, src += a, dest += b)
+        // convert source image with img_n components to one with req_comp components;
+        // avoid switch per pixel, so use switch per scanline and massive macros
+        switch (STBI__COMBO(img_n, req_comp)) {
+            STBI__CASE(1, 2) {
+                dest[0] = src[0];
+                dest[1] = 255;
+            }
+            break;
+            STBI__CASE(1, 3) { dest[0] = dest[1] = dest[2] = src[0]; }
+            break;
+            STBI__CASE(1, 4) {
+                dest[0] = dest[1] = dest[2] = src[0];
+                dest[3] = 255;
+            }
+            break;
+            STBI__CASE(2, 1) { dest[0] = src[0]; }
+            break;
+            STBI__CASE(2, 3) { dest[0] = dest[1] = dest[2] = src[0]; }
+            break;
+            STBI__CASE(2, 4) {
+                dest[0] = dest[1] = dest[2] = src[0];
+                dest[3] = src[1];
+            }
+            break;
+            STBI__CASE(3, 4) {
+                dest[0] = src[0];
+                dest[1] = src[1];
+                dest[2] = src[2];
+                dest[3] = 255;
+            }
+            break;
+            STBI__CASE(3, 1) { dest[0] = stbi__compute_y(src[0], src[1], src[2]); }
+            break;
+            STBI__CASE(3, 2) {
+                dest[0] = stbi__compute_y(src[0], src[1], src[2]);
+                dest[1] = 255;
+            }
+            break;
+            STBI__CASE(4, 1) { dest[0] = stbi__compute_y(src[0], src[1], src[2]); }
+            break;
+            STBI__CASE(4, 2) {
+                dest[0] = stbi__compute_y(src[0], src[1], src[2]);
+                dest[1] = src[3];
+            }
+            break;
+            STBI__CASE(4, 3) {
+                dest[0] = src[0];
+                dest[1] = src[1];
+                dest[2] = src[2];
+            }
+            break;
+        default:
+            STBI_ASSERT(0);
+            STBI_FREE(data);
+            STBI_FREE(good);
+            return stbi__errpuc("unsupported", "Unsupported format conversion");
+        }
+#undef STBI__CASE
+    }
+
+    STBI_FREE(data);
+    return good;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 stbi__compute_y_16(int r, int g, int b) { return (stbi__uint16)(((r * 77) + (g * 150) + (29 * b)) >> 8); }
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 * stbi__convert_format16(stbi__uint16 * data, int img_n, int req_comp, unsigned int x, unsigned int y) {
+    int i, j;
+    stbi__uint16 * good;
+
+    if (req_comp == img_n)
+        return data;
+    STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+    good = (stbi__uint16 *)stbi__malloc(req_comp * x * y * 2);
+    if (good == NULL) {
+        STBI_FREE(data);
+        return (stbi__uint16 *)stbi__errpuc("outofmem", "Out of memory");
+    }
+
+    for (j = 0; j < (int)y; ++j) {
+        stbi__uint16 * src = data + j * x * img_n;
+        stbi__uint16 * dest = good + j * x * req_comp;
+
+#define STBI__COMBO(a, b) ((a)*8 + (b))
+#define STBI__CASE(a, b)                                                                                                       \
+    case STBI__COMBO(a, b):                                                                                                    \
+        for (i = x - 1; i >= 0; --i, src += a, dest += b)
+        // convert source image with img_n components to one with req_comp components;
+        // avoid switch per pixel, so use switch per scanline and massive macros
+        switch (STBI__COMBO(img_n, req_comp)) {
+            STBI__CASE(1, 2) {
+                dest[0] = src[0];
+                dest[1] = 0xffff;
+            }
+            break;
+            STBI__CASE(1, 3) { dest[0] = dest[1] = dest[2] = src[0]; }
+            break;
+            STBI__CASE(1, 4) {
+                dest[0] = dest[1] = dest[2] = src[0];
+                dest[3] = 0xffff;
+            }
+            break;
+            STBI__CASE(2, 1) { dest[0] = src[0]; }
+            break;
+            STBI__CASE(2, 3) { dest[0] = dest[1] = dest[2] = src[0]; }
+            break;
+            STBI__CASE(2, 4) {
+                dest[0] = dest[1] = dest[2] = src[0];
+                dest[3] = src[1];
+            }
+            break;
+            STBI__CASE(3, 4) {
+                dest[0] = src[0];
+                dest[1] = src[1];
+                dest[2] = src[2];
+                dest[3] = 0xffff;
+            }
+            break;
+            STBI__CASE(3, 1) { dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); }
+            break;
+            STBI__CASE(3, 2) {
+                dest[0] = stbi__compute_y_16(src[0], src[1], src[2]);
+                dest[1] = 0xffff;
+            }
+            break;
+            STBI__CASE(4, 1) { dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); }
+            break;
+            STBI__CASE(4, 2) {
+                dest[0] = stbi__compute_y_16(src[0], src[1], src[2]);
+                dest[1] = src[3];
+            }
+            break;
+            STBI__CASE(4, 3) {
+                dest[0] = src[0];
+                dest[1] = src[1];
+                dest[2] = src[2];
+            }
+            break;
+        default:
+            STBI_ASSERT(0);
+            STBI_FREE(data);
+            STBI_FREE(good);
+            return (stbi__uint16 *)stbi__errpuc("unsupported", "Unsupported format conversion");
+        }
+#undef STBI__CASE
+    }
+
+    STBI_FREE(data);
+    return good;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float * stbi__ldr_to_hdr(stbi_uc * data, int x, int y, int comp) {
+    int i, k, n;
+    float * output;
+    if (!data)
+        return NULL;
+    output = (float *)stbi__malloc_mad4(x, y, comp, sizeof(float), 0);
+    if (output == NULL) {
+        STBI_FREE(data);
+        return stbi__errpf("outofmem", "Out of memory");
+    }
+    // compute number of non-alpha components
+    if (comp & 1)
+        n = comp;
+    else
+        n = comp - 1;
+    for (i = 0; i < x * y; ++i) {
+        for (k = 0; k < n; ++k) {
+            output[i * comp + k] = (float)(pow(data[i * comp + k] / 255.0f, stbi__l2h_gamma) * stbi__l2h_scale);
+        }
+    }
+    if (n < comp) {
+        for (i = 0; i < x * y; ++i) {
+            output[i * comp + n] = data[i * comp + n] / 255.0f;
+        }
+    }
+    STBI_FREE(data);
+    return output;
+}
+#endif
+
+#ifndef STBI_NO_HDR
+#define stbi__float2int(x) ((int)(x))
+static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp) {
+    int i, k, n;
+    stbi_uc * output;
+    if (!data)
+        return NULL;
+    output = (stbi_uc *)stbi__malloc_mad3(x, y, comp, 0);
+    if (output == NULL) {
+        STBI_FREE(data);
+        return stbi__errpuc("outofmem", "Out of memory");
+    }
+    // compute number of non-alpha components
+    if (comp & 1)
+        n = comp;
+    else
+        n = comp - 1;
+    for (i = 0; i < x * y; ++i) {
+        for (k = 0; k < n; ++k) {
+            float z = (float)pow(data[i * comp + k] * stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f;
+            if (z < 0)
+                z = 0;
+            if (z > 255)
+                z = 255;
+            output[i * comp + k] = (stbi_uc)stbi__float2int(z);
+        }
+        if (k < comp) {
+            float z = data[i * comp + k] * 255 + 0.5f;
+            if (z < 0)
+                z = 0;
+            if (z > 255)
+                z = 255;
+            output[i * comp + k] = (stbi_uc)stbi__float2int(z);
+        }
+    }
+    STBI_FREE(data);
+    return output;
+}
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//  "baseline" JPEG/JFIF decoder
+//
+//    simple implementation
+//      - doesn't support delayed output of y-dimension
+//      - simple interface (only one output format: 8-bit interleaved RGB)
+//      - doesn't try to recover corrupt jpegs
+//      - doesn't allow partial loading, loading multiple at once
+//      - still fast on x86 (copying globals into locals doesn't help x86)
+//      - allocates lots of intermediate memory (full size of all components)
+//        - non-interleaved case requires this anyway
+//        - allows good upsampling (see next)
+//    high-quality
+//      - upsampled channels are bilinearly interpolated, even across blocks
+//      - quality integer IDCT derived from IJG's 'slow'
+//    performance
+//      - fast huffman; reasonable integer IDCT
+//      - some SIMD kernels for common paths on targets with SSE2/NEON
+//      - uses a lot of intermediate memory, could cache poorly
+
+#ifndef STBI_NO_JPEG
+
+// huffman decoding acceleration
+#define FAST_BITS 9 // larger handles more cases; smaller stomps less cache
+
+typedef struct {
+    stbi_uc fast[1 << FAST_BITS];
+    // weirdly, repacking this into AoS is a 10% speed loss, instead of a win
+    stbi__uint16 code[256];
+    stbi_uc values[256];
+    stbi_uc size[257];
+    unsigned int maxcode[18];
+    int delta[17]; // old 'firstsymbol' - old 'firstcode'
+} stbi__huffman;
+
+typedef struct {
+    stbi__context * s;
+    stbi__huffman huff_dc[4];
+    stbi__huffman huff_ac[4];
+    stbi__uint16 dequant[4][64];
+    stbi__int16 fast_ac[4][1 << FAST_BITS];
+
+    // sizes for components, interleaved MCUs
+    int img_h_max, img_v_max;
+    int img_mcu_x, img_mcu_y;
+    int img_mcu_w, img_mcu_h;
+
+    // definition of jpeg image component
+    struct {
+        int id;
+        int h, v;
+        int tq;
+        int hd, ha;
+        int dc_pred;
+
+        int x, y, w2, h2;
+        stbi_uc * data;
+        void *raw_data, *raw_coeff;
+        stbi_uc * linebuf;
+        short * coeff;        // progressive only
+        int coeff_w, coeff_h; // number of 8x8 coefficient blocks
+    } img_comp[4];
+
+    stbi__uint32 code_buffer; // jpeg entropy-coded buffer
+    int code_bits;            // number of valid bits
+    unsigned char marker;     // marker seen while filling entropy buffer
+    int nomore;               // flag if we saw a marker so must stop
+
+    int progressive;
+    int spec_start;
+    int spec_end;
+    int succ_high;
+    int succ_low;
+    int eob_run;
+    int jfif;
+    int app14_color_transform; // Adobe APP14 tag
+    int rgb;
+
+    int scan_n, order[4];
+    int restart_interval, todo;
+
+    // kernels
+    void (*idct_block_kernel)(stbi_uc * out, int out_stride, short data[64]);
+    void (*YCbCr_to_RGB_kernel)(stbi_uc * out, const stbi_uc * y, const stbi_uc * pcb, const stbi_uc * pcr, int count,
+                                int step);
+    stbi_uc * (*resample_row_hv_2_kernel)(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs);
+} stbi__jpeg;
+
+static int stbi__build_huffman(stbi__huffman * h, int * count) {
+    int i, j, k = 0;
+    unsigned int code;
+    // build size list for each symbol (from JPEG spec)
+    for (i = 0; i < 16; ++i) {
+        for (j = 0; j < count[i]; ++j) {
+            h->size[k++] = (stbi_uc)(i + 1);
+            if (k >= 257)
+                return stbi__err("bad size list", "Corrupt JPEG");
+        }
+    }
+    h->size[k] = 0;
+
+    // compute actual symbols (from jpeg spec)
+    code = 0;
+    k = 0;
+    for (j = 1; j <= 16; ++j) {
+        // compute delta to add to code to compute symbol id
+        h->delta[j] = k - code;
+        if (h->size[k] == j) {
+            while (h->size[k] == j)
+                h->code[k++] = (stbi__uint16)(code++);
+            if (code - 1 >= (1u << j))
+                return stbi__err("bad code lengths", "Corrupt JPEG");
+        }
+        // compute largest code + 1 for this size, preshifted as needed later
+        h->maxcode[j] = code << (16 - j);
+        code <<= 1;
+    }
+    h->maxcode[j] = 0xffffffff;
+
+    // build non-spec acceleration table; 255 is flag for not-accelerated
+    memset(h->fast, 255, 1 << FAST_BITS);
+    for (i = 0; i < k; ++i) {
+        int s = h->size[i];
+        if (s <= FAST_BITS) {
+            int c = h->code[i] << (FAST_BITS - s);
+            int m = 1 << (FAST_BITS - s);
+            for (j = 0; j < m; ++j) {
+                h->fast[c + j] = (stbi_uc)i;
+            }
+        }
+    }
+    return 1;
+}
+
+// build a table that decodes both magnitude and value of small ACs in
+// one go.
+static void stbi__build_fast_ac(stbi__int16 * fast_ac, stbi__huffman * h) {
+    int i;
+    for (i = 0; i < (1 << FAST_BITS); ++i) {
+        stbi_uc fast = h->fast[i];
+        fast_ac[i] = 0;
+        if (fast < 255) {
+            int rs = h->values[fast];
+            int run = (rs >> 4) & 15;
+            int magbits = rs & 15;
+            int len = h->size[fast];
+
+            if (magbits && len + magbits <= FAST_BITS) {
+                // magnitude code followed by receive_extend code
+                int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits);
+                int m = 1 << (magbits - 1);
+                if (k < m)
+                    k += (~0U << magbits) + 1;
+                // if the result is small enough, we can fit it in fast_ac table
+                if (k >= -128 && k <= 127)
+                    fast_ac[i] = (stbi__int16)((k * 256) + (run * 16) + (len + magbits));
+            }
+        }
+    }
+}
+
+static void stbi__grow_buffer_unsafe(stbi__jpeg * j) {
+    do {
+        unsigned int b = j->nomore ? 0 : stbi__get8(j->s);
+        if (b == 0xff) {
+            int c = stbi__get8(j->s);
+            while (c == 0xff)
+                c = stbi__get8(j->s); // consume fill bytes
+            if (c != 0) {
+                j->marker = (unsigned char)c;
+                j->nomore = 1;
+                return;
+            }
+        }
+        j->code_buffer |= b << (24 - j->code_bits);
+        j->code_bits += 8;
+    } while (j->code_bits <= 24);
+}
+
+// (1 << n) - 1
+static const stbi__uint32 stbi__bmask[17] = {0,   1,    3,    7,    15,   31,    63,    127,  255,
+                                             511, 1023, 2047, 4095, 8191, 16383, 32767, 65535};
+
+// decode a jpeg huffman value from the bitstream
+stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg * j, stbi__huffman * h) {
+    unsigned int temp;
+    int c, k;
+
+    if (j->code_bits < 16)
+        stbi__grow_buffer_unsafe(j);
+
+    // look at the top FAST_BITS and determine what symbol ID it is,
+    // if the code is <= FAST_BITS
+    c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1);
+    k = h->fast[c];
+    if (k < 255) {
+        int s = h->size[k];
+        if (s > j->code_bits)
+            return -1;
+        j->code_buffer <<= s;
+        j->code_bits -= s;
+        return h->values[k];
+    }
+
+    // naive test is to shift the code_buffer down so k bits are
+    // valid, then test against maxcode. To speed this up, we've
+    // preshifted maxcode left so that it has (16-k) 0s at the
+    // end; in other words, regardless of the number of bits, it
+    // wants to be compared against something shifted to have 16;
+    // that way we don't need to shift inside the loop.
+    temp = j->code_buffer >> 16;
+    for (k = FAST_BITS + 1;; ++k)
+        if (temp < h->maxcode[k])
+            break;
+    if (k == 17) {
+        // error! code not found
+        j->code_bits -= 16;
+        return -1;
+    }
+
+    if (k > j->code_bits)
+        return -1;
+
+    // convert the huffman code to the symbol id
+    c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k];
+    if (c < 0 || c >= 256) // symbol id out of bounds!
+        return -1;
+    STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]);
+
+    // convert the id to a symbol
+    j->code_bits -= k;
+    j->code_buffer <<= k;
+    return h->values[c];
+}
+
+// bias[n] = (-1<<n) + 1
+static const int stbi__jbias[16] = {0, -1, -3, -7, -15, -31, -63, -127, -255, -511, -1023, -2047, -4095, -8191, -16383, -32767};
+
+// combined JPEG 'receive' and JPEG 'extend', since baseline
+// always extends everything it receives.
+stbi_inline static int stbi__extend_receive(stbi__jpeg * j, int n) {
+    unsigned int k;
+    int sgn;
+    if (j->code_bits < n)
+        stbi__grow_buffer_unsafe(j);
+    if (j->code_bits < n)
+        return 0; // ran out of bits from stream, return 0s intead of continuing
+
+    sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative)
+    k = stbi_lrot(j->code_buffer, n);
+    j->code_buffer = k & ~stbi__bmask[n];
+    k &= stbi__bmask[n];
+    j->code_bits -= n;
+    return k + (stbi__jbias[n] & (sgn - 1));
+}
+
+// get some unsigned bits
+stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg * j, int n) {
+    unsigned int k;
+    if (j->code_bits < n)
+        stbi__grow_buffer_unsafe(j);
+    if (j->code_bits < n)
+        return 0; // ran out of bits from stream, return 0s intead of continuing
+    k = stbi_lrot(j->code_buffer, n);
+    j->code_buffer = k & ~stbi__bmask[n];
+    k &= stbi__bmask[n];
+    j->code_bits -= n;
+    return k;
+}
+
+stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg * j) {
+    unsigned int k;
+    if (j->code_bits < 1)
+        stbi__grow_buffer_unsafe(j);
+    if (j->code_bits < 1)
+        return 0; // ran out of bits from stream, return 0s intead of continuing
+    k = j->code_buffer;
+    j->code_buffer <<= 1;
+    --j->code_bits;
+    return k & 0x80000000;
+}
+
+// given a value that's at position X in the zigzag stream,
+// where does it appear in the 8x8 matrix coded as row-major?
+static const stbi_uc stbi__jpeg_dezigzag[64 + 15] = {
+    0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, 27, 20, 13, 6, 7, 14, 21, 28, 35,
+    42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, 53, 60, 61, 54, 47, 55, 62, 63,
+    // let corrupt input sample past end
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63};
+
+// decode one 64-entry block--
+static int stbi__jpeg_decode_block(stbi__jpeg * j, short data[64], stbi__huffman * hdc, stbi__huffman * hac, stbi__int16 * fac,
+                                   int b, stbi__uint16 * dequant) {
+    int diff, dc, k;
+    int t;
+
+    if (j->code_bits < 16)
+        stbi__grow_buffer_unsafe(j);
+    t = stbi__jpeg_huff_decode(j, hdc);
+    if (t < 0 || t > 15)
+        return stbi__err("bad huffman code", "Corrupt JPEG");
+
+    // 0 all the ac values now so we can do it 32-bits at a time
+    memset(data, 0, 64 * sizeof(data[0]));
+
+    diff = t ? stbi__extend_receive(j, t) : 0;
+    if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff))
+        return stbi__err("bad delta", "Corrupt JPEG");
+    dc = j->img_comp[b].dc_pred + diff;
+    j->img_comp[b].dc_pred = dc;
+    if (!stbi__mul2shorts_valid(dc, dequant[0]))
+        return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+    data[0] = (short)(dc * dequant[0]);
+
+    // decode AC components, see JPEG spec
+    k = 1;
+    do {
+        unsigned int zig;
+        int c, r, s;
+        if (j->code_bits < 16)
+            stbi__grow_buffer_unsafe(j);
+        c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1);
+        r = fac[c];
+        if (r) {                // fast-AC path
+            k += (r >> 4) & 15; // run
+            s = r & 15;         // combined length
+            if (s > j->code_bits)
+                return stbi__err("bad huffman code", "Combined length longer than code bits available");
+            j->code_buffer <<= s;
+            j->code_bits -= s;
+            // decode into unzigzag'd location
+            zig = stbi__jpeg_dezigzag[k++];
+            data[zig] = (short)((r >> 8) * dequant[zig]);
+        } else {
+            int rs = stbi__jpeg_huff_decode(j, hac);
+            if (rs < 0)
+                return stbi__err("bad huffman code", "Corrupt JPEG");
+            s = rs & 15;
+            r = rs >> 4;
+            if (s == 0) {
+                if (rs != 0xf0)
+                    break; // end block
+                k += 16;
+            } else {
+                k += r;
+                // decode into unzigzag'd location
+                zig = stbi__jpeg_dezigzag[k++];
+                data[zig] = (short)(stbi__extend_receive(j, s) * dequant[zig]);
+            }
+        }
+    } while (k < 64);
+    return 1;
+}
+
+static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg * j, short data[64], stbi__huffman * hdc, int b) {
+    int diff, dc;
+    int t;
+    if (j->spec_end != 0)
+        return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+    if (j->code_bits < 16)
+        stbi__grow_buffer_unsafe(j);
+
+    if (j->succ_high == 0) {
+        // first scan for DC coefficient, must be first
+        memset(data, 0, 64 * sizeof(data[0])); // 0 all the ac values now
+        t = stbi__jpeg_huff_decode(j, hdc);
+        if (t < 0 || t > 15)
+            return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+        diff = t ? stbi__extend_receive(j, t) : 0;
+
+        if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff))
+            return stbi__err("bad delta", "Corrupt JPEG");
+        dc = j->img_comp[b].dc_pred + diff;
+        j->img_comp[b].dc_pred = dc;
+        if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low))
+            return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+        data[0] = (short)(dc * (1 << j->succ_low));
+    } else {
+        // refinement scan for DC coefficient
+        if (stbi__jpeg_get_bit(j))
+            data[0] += (short)(1 << j->succ_low);
+    }
+    return 1;
+}
+
+// @OPTIMIZE: store non-zigzagged during the decode passes,
+// and only de-zigzag when dequantizing
+static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg * j, short data[64], stbi__huffman * hac, stbi__int16 * fac) {
+    int k;
+    if (j->spec_start == 0)
+        return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+    if (j->succ_high == 0) {
+        int shift = j->succ_low;
+
+        if (j->eob_run) {
+            --j->eob_run;
+            return 1;
+        }
+
+        k = j->spec_start;
+        do {
+            unsigned int zig;
+            int c, r, s;
+            if (j->code_bits < 16)
+                stbi__grow_buffer_unsafe(j);
+            c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1);
+            r = fac[c];
+            if (r) {                // fast-AC path
+                k += (r >> 4) & 15; // run
+                s = r & 15;         // combined length
+                if (s > j->code_bits)
+                    return stbi__err("bad huffman code", "Combined length longer than code bits available");
+                j->code_buffer <<= s;
+                j->code_bits -= s;
+                zig = stbi__jpeg_dezigzag[k++];
+                data[zig] = (short)((r >> 8) * (1 << shift));
+            } else {
+                int rs = stbi__jpeg_huff_decode(j, hac);
+                if (rs < 0)
+                    return stbi__err("bad huffman code", "Corrupt JPEG");
+                s = rs & 15;
+                r = rs >> 4;
+                if (s == 0) {
+                    if (r < 15) {
+                        j->eob_run = (1 << r);
+                        if (r)
+                            j->eob_run += stbi__jpeg_get_bits(j, r);
+                        --j->eob_run;
+                        break;
+                    }
+                    k += 16;
+                } else {
+                    k += r;
+                    zig = stbi__jpeg_dezigzag[k++];
+                    data[zig] = (short)(stbi__extend_receive(j, s) * (1 << shift));
+                }
+            }
+        } while (k <= j->spec_end);
+    } else {
+        // refinement scan for these AC coefficients
+
+        short bit = (short)(1 << j->succ_low);
+
+        if (j->eob_run) {
+            --j->eob_run;
+            for (k = j->spec_start; k <= j->spec_end; ++k) {
+                short * p = &data[stbi__jpeg_dezigzag[k]];
+                if (*p != 0)
+                    if (stbi__jpeg_get_bit(j))
+                        if ((*p & bit) == 0) {
+                            if (*p > 0)
+                                *p += bit;
+                            else
+                                *p -= bit;
+                        }
+            }
+        } else {
+            k = j->spec_start;
+            do {
+                int r, s;
+                int rs = stbi__jpeg_huff_decode(
+                    j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh
+                if (rs < 0)
+                    return stbi__err("bad huffman code", "Corrupt JPEG");
+                s = rs & 15;
+                r = rs >> 4;
+                if (s == 0) {
+                    if (r < 15) {
+                        j->eob_run = (1 << r) - 1;
+                        if (r)
+                            j->eob_run += stbi__jpeg_get_bits(j, r);
+                        r = 64; // force end of block
+                    } else {
+                        // r=15 s=0 should write 16 0s, so we just do
+                        // a run of 15 0s and then write s (which is 0),
+                        // so we don't have to do anything special here
+                    }
+                } else {
+                    if (s != 1)
+                        return stbi__err("bad huffman code", "Corrupt JPEG");
+                    // sign bit
+                    if (stbi__jpeg_get_bit(j))
+                        s = bit;
+                    else
+                        s = -bit;
+                }
+
+                // advance by r
+                while (k <= j->spec_end) {
+                    short * p = &data[stbi__jpeg_dezigzag[k++]];
+                    if (*p != 0) {
+                        if (stbi__jpeg_get_bit(j))
+                            if ((*p & bit) == 0) {
+                                if (*p > 0)
+                                    *p += bit;
+                                else
+                                    *p -= bit;
+                            }
+                    } else {
+                        if (r == 0) {
+                            *p = (short)s;
+                            break;
+                        }
+                        --r;
+                    }
+                }
+            } while (k <= j->spec_end);
+        }
+    }
+    return 1;
+}
+
+// take a -128..127 value and stbi__clamp it and convert to 0..255
+stbi_inline static stbi_uc stbi__clamp(int x) {
+    // trick to use a single test to catch both cases
+    if ((unsigned int)x > 255) {
+        if (x < 0)
+            return 0;
+        if (x > 255)
+            return 255;
+    }
+    return (stbi_uc)x;
+}
+
+#define stbi__f2f(x) ((int)(((x)*4096 + 0.5)))
+#define stbi__fsh(x) ((x)*4096)
+
+// derived from jidctint -- DCT_ISLOW
+#define STBI__IDCT_1D(s0, s1, s2, s3, s4, s5, s6, s7)                                                                          \
+    int t0, t1, t2, t3, p1, p2, p3, p4, p5, x0, x1, x2, x3;                                                                    \
+    p2 = s2;                                                                                                                   \
+    p3 = s6;                                                                                                                   \
+    p1 = (p2 + p3) * stbi__f2f(0.5411961f);                                                                                    \
+    t2 = p1 + p3 * stbi__f2f(-1.847759065f);                                                                                   \
+    t3 = p1 + p2 * stbi__f2f(0.765366865f);                                                                                    \
+    p2 = s0;                                                                                                                   \
+    p3 = s4;                                                                                                                   \
+    t0 = stbi__fsh(p2 + p3);                                                                                                   \
+    t1 = stbi__fsh(p2 - p3);                                                                                                   \
+    x0 = t0 + t3;                                                                                                              \
+    x3 = t0 - t3;                                                                                                              \
+    x1 = t1 + t2;                                                                                                              \
+    x2 = t1 - t2;                                                                                                              \
+    t0 = s7;                                                                                                                   \
+    t1 = s5;                                                                                                                   \
+    t2 = s3;                                                                                                                   \
+    t3 = s1;                                                                                                                   \
+    p3 = t0 + t2;                                                                                                              \
+    p4 = t1 + t3;                                                                                                              \
+    p1 = t0 + t3;                                                                                                              \
+    p2 = t1 + t2;                                                                                                              \
+    p5 = (p3 + p4) * stbi__f2f(1.175875602f);                                                                                  \
+    t0 = t0 * stbi__f2f(0.298631336f);                                                                                         \
+    t1 = t1 * stbi__f2f(2.053119869f);                                                                                         \
+    t2 = t2 * stbi__f2f(3.072711026f);                                                                                         \
+    t3 = t3 * stbi__f2f(1.501321110f);                                                                                         \
+    p1 = p5 + p1 * stbi__f2f(-0.899976223f);                                                                                   \
+    p2 = p5 + p2 * stbi__f2f(-2.562915447f);                                                                                   \
+    p3 = p3 * stbi__f2f(-1.961570560f);                                                                                        \
+    p4 = p4 * stbi__f2f(-0.390180644f);                                                                                        \
+    t3 += p1 + p4;                                                                                                             \
+    t2 += p2 + p3;                                                                                                             \
+    t1 += p2 + p4;                                                                                                             \
+    t0 += p1 + p3;
+
+static void stbi__idct_block(stbi_uc * out, int out_stride, short data[64]) {
+    int i, val[64], *v = val;
+    stbi_uc * o;
+    short * d = data;
+
+    // columns
+    for (i = 0; i < 8; ++i, ++d, ++v) {
+        // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing
+        if (d[8] == 0 && d[16] == 0 && d[24] == 0 && d[32] == 0 && d[40] == 0 && d[48] == 0 && d[56] == 0) {
+            //    no shortcut                 0     seconds
+            //    (1|2|3|4|5|6|7)==0          0     seconds
+            //    all separate               -0.047 seconds
+            //    1 && 2|3 && 4|5 && 6|7:    -0.047 seconds
+            int dcterm = d[0] * 4;
+            v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm;
+        } else {
+            STBI__IDCT_1D(d[0], d[8], d[16], d[24], d[32], d[40], d[48], d[56])
+            // constants scaled things up by 1<<12; let's bring them back
+            // down, but keep 2 extra bits of precision
+            x0 += 512;
+            x1 += 512;
+            x2 += 512;
+            x3 += 512;
+            v[0] = (x0 + t3) >> 10;
+            v[56] = (x0 - t3) >> 10;
+            v[8] = (x1 + t2) >> 10;
+            v[48] = (x1 - t2) >> 10;
+            v[16] = (x2 + t1) >> 10;
+            v[40] = (x2 - t1) >> 10;
+            v[24] = (x3 + t0) >> 10;
+            v[32] = (x3 - t0) >> 10;
+        }
+    }
+
+    for (i = 0, v = val, o = out; i < 8; ++i, v += 8, o += out_stride) {
+        // no fast case since the first 1D IDCT spread components out
+        STBI__IDCT_1D(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7])
+        // constants scaled things up by 1<<12, plus we had 1<<2 from first
+        // loop, plus horizontal and vertical each scale by sqrt(8) so together
+        // we've got an extra 1<<3, so 1<<17 total we need to remove.
+        // so we want to round that, which means adding 0.5 * 1<<17,
+        // aka 65536. Also, we'll end up with -128 to 127 that we want
+        // to encode as 0..255 by adding 128, so we'll add that before the shift
+        x0 += 65536 + (128 << 17);
+        x1 += 65536 + (128 << 17);
+        x2 += 65536 + (128 << 17);
+        x3 += 65536 + (128 << 17);
+        // tried computing the shifts into temps, or'ing the temps to see
+        // if any were out of range, but that was slower
+        o[0] = stbi__clamp((x0 + t3) >> 17);
+        o[7] = stbi__clamp((x0 - t3) >> 17);
+        o[1] = stbi__clamp((x1 + t2) >> 17);
+        o[6] = stbi__clamp((x1 - t2) >> 17);
+        o[2] = stbi__clamp((x2 + t1) >> 17);
+        o[5] = stbi__clamp((x2 - t1) >> 17);
+        o[3] = stbi__clamp((x3 + t0) >> 17);
+        o[4] = stbi__clamp((x3 - t0) >> 17);
+    }
+}
+
+#ifdef STBI_SSE2
+// sse2 integer IDCT. not the fastest possible implementation but it
+// produces bit-identical results to the generic C version so it's
+// fully "transparent".
+static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) {
+    // This is constructed to match our regular (generic) integer IDCT exactly.
+    __m128i row0, row1, row2, row3, row4, row5, row6, row7;
+    __m128i tmp;
+
+// dot product constant: even elems=x, odd elems=y
+#define dct_const(x, y) _mm_setr_epi16((x), (y), (x), (y), (x), (y), (x), (y))
+
+// out(0) = c0[even]*x + c0[odd]*y   (c0, x, y 16-bit, out 32-bit)
+// out(1) = c1[even]*x + c1[odd]*y
+#define dct_rot(out0, out1, x, y, c0, c1)                                                                                      \
+    __m128i c0##lo = _mm_unpacklo_epi16((x), (y));                                                                             \
+    __m128i c0##hi = _mm_unpackhi_epi16((x), (y));                                                                             \
+    __m128i out0##_l = _mm_madd_epi16(c0##lo, c0);                                                                             \
+    __m128i out0##_h = _mm_madd_epi16(c0##hi, c0);                                                                             \
+    __m128i out1##_l = _mm_madd_epi16(c0##lo, c1);                                                                             \
+    __m128i out1##_h = _mm_madd_epi16(c0##hi, c1)
+
+// out = in << 12  (in 16-bit, out 32-bit)
+#define dct_widen(out, in)                                                                                                     \
+    __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4);                                        \
+    __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4)
+
+// wide add
+#define dct_wadd(out, a, b)                                                                                                    \
+    __m128i out##_l = _mm_add_epi32(a##_l, b##_l);                                                                             \
+    __m128i out##_h = _mm_add_epi32(a##_h, b##_h)
+
+// wide sub
+#define dct_wsub(out, a, b)                                                                                                    \
+    __m128i out##_l = _mm_sub_epi32(a##_l, b##_l);                                                                             \
+    __m128i out##_h = _mm_sub_epi32(a##_h, b##_h)
+
+// butterfly a/b, add bias, then shift by "s" and pack
+#define dct_bfly32o(out0, out1, a, b, bias, s)                                                                                 \
+    {                                                                                                                          \
+        __m128i abiased_l = _mm_add_epi32(a##_l, bias);                                                                        \
+        __m128i abiased_h = _mm_add_epi32(a##_h, bias);                                                                        \
+        dct_wadd(sum, abiased, b);                                                                                             \
+        dct_wsub(dif, abiased, b);                                                                                             \
+        out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s));                                            \
+        out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s));                                            \
+    }
+
+// 8-bit interleave step (for transposes)
+#define dct_interleave8(a, b)                                                                                                  \
+    tmp = a;                                                                                                                   \
+    a = _mm_unpacklo_epi8(a, b);                                                                                               \
+    b = _mm_unpackhi_epi8(tmp, b)
+
+// 16-bit interleave step (for transposes)
+#define dct_interleave16(a, b)                                                                                                 \
+    tmp = a;                                                                                                                   \
+    a = _mm_unpacklo_epi16(a, b);                                                                                              \
+    b = _mm_unpackhi_epi16(tmp, b)
+
+#define dct_pass(bias, shift)                                                                                                  \
+    {                                                                                                                          \
+        /* even part */                                                                                                        \
+        dct_rot(t2e, t3e, row2, row6, rot0_0, rot0_1);                                                                         \
+        __m128i sum04 = _mm_add_epi16(row0, row4);                                                                             \
+        __m128i dif04 = _mm_sub_epi16(row0, row4);                                                                             \
+        dct_widen(t0e, sum04);                                                                                                 \
+        dct_widen(t1e, dif04);                                                                                                 \
+        dct_wadd(x0, t0e, t3e);                                                                                                \
+        dct_wsub(x3, t0e, t3e);                                                                                                \
+        dct_wadd(x1, t1e, t2e);                                                                                                \
+        dct_wsub(x2, t1e, t2e);                                                                                                \
+        /* odd part */                                                                                                         \
+        dct_rot(y0o, y2o, row7, row3, rot2_0, rot2_1);                                                                         \
+        dct_rot(y1o, y3o, row5, row1, rot3_0, rot3_1);                                                                         \
+        __m128i sum17 = _mm_add_epi16(row1, row7);                                                                             \
+        __m128i sum35 = _mm_add_epi16(row3, row5);                                                                             \
+        dct_rot(y4o, y5o, sum17, sum35, rot1_0, rot1_1);                                                                       \
+        dct_wadd(x4, y0o, y4o);                                                                                                \
+        dct_wadd(x5, y1o, y5o);                                                                                                \
+        dct_wadd(x6, y2o, y5o);                                                                                                \
+        dct_wadd(x7, y3o, y4o);                                                                                                \
+        dct_bfly32o(row0, row7, x0, x7, bias, shift);                                                                          \
+        dct_bfly32o(row1, row6, x1, x6, bias, shift);                                                                          \
+        dct_bfly32o(row2, row5, x2, x5, bias, shift);                                                                          \
+        dct_bfly32o(row3, row4, x3, x4, bias, shift);                                                                          \
+    }
+
+    __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f));
+    __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f(0.765366865f), stbi__f2f(0.5411961f));
+    __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f));
+    __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f));
+    __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f(0.298631336f), stbi__f2f(-1.961570560f));
+    __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f(3.072711026f));
+    __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f(2.053119869f), stbi__f2f(-0.390180644f));
+    __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f(1.501321110f));
+
+    // rounding biases in column/row passes, see stbi__idct_block for explanation.
+    __m128i bias_0 = _mm_set1_epi32(512);
+    __m128i bias_1 = _mm_set1_epi32(65536 + (128 << 17));
+
+    // load
+    row0 = _mm_load_si128((const __m128i *)(data + 0 * 8));
+    row1 = _mm_load_si128((const __m128i *)(data + 1 * 8));
+    row2 = _mm_load_si128((const __m128i *)(data + 2 * 8));
+    row3 = _mm_load_si128((const __m128i *)(data + 3 * 8));
+    row4 = _mm_load_si128((const __m128i *)(data + 4 * 8));
+    row5 = _mm_load_si128((const __m128i *)(data + 5 * 8));
+    row6 = _mm_load_si128((const __m128i *)(data + 6 * 8));
+    row7 = _mm_load_si128((const __m128i *)(data + 7 * 8));
+
+    // column pass
+    dct_pass(bias_0, 10);
+
+    {
+        // 16bit 8x8 transpose pass 1
+        dct_interleave16(row0, row4);
+        dct_interleave16(row1, row5);
+        dct_interleave16(row2, row6);
+        dct_interleave16(row3, row7);
+
+        // transpose pass 2
+        dct_interleave16(row0, row2);
+        dct_interleave16(row1, row3);
+        dct_interleave16(row4, row6);
+        dct_interleave16(row5, row7);
+
+        // transpose pass 3
+        dct_interleave16(row0, row1);
+        dct_interleave16(row2, row3);
+        dct_interleave16(row4, row5);
+        dct_interleave16(row6, row7);
+    }
+
+    // row pass
+    dct_pass(bias_1, 17);
+
+    {
+        // pack
+        __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7
+        __m128i p1 = _mm_packus_epi16(row2, row3);
+        __m128i p2 = _mm_packus_epi16(row4, row5);
+        __m128i p3 = _mm_packus_epi16(row6, row7);
+
+        // 8bit 8x8 transpose pass 1
+        dct_interleave8(p0, p2); // a0e0a1e1...
+        dct_interleave8(p1, p3); // c0g0c1g1...
+
+        // transpose pass 2
+        dct_interleave8(p0, p1); // a0c0e0g0...
+        dct_interleave8(p2, p3); // b0d0f0h0...
+
+        // transpose pass 3
+        dct_interleave8(p0, p2); // a0b0c0d0...
+        dct_interleave8(p1, p3); // a4b4c4d4...
+
+        // store
+        _mm_storel_epi64((__m128i *)out, p0);
+        out += out_stride;
+        _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p0, 0x4e));
+        out += out_stride;
+        _mm_storel_epi64((__m128i *)out, p2);
+        out += out_stride;
+        _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p2, 0x4e));
+        out += out_stride;
+        _mm_storel_epi64((__m128i *)out, p1);
+        out += out_stride;
+        _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p1, 0x4e));
+        out += out_stride;
+        _mm_storel_epi64((__m128i *)out, p3);
+        out += out_stride;
+        _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p3, 0x4e));
+    }
+
+#undef dct_const
+#undef dct_rot
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_interleave8
+#undef dct_interleave16
+#undef dct_pass
+}
+
+#endif // STBI_SSE2
+
+#ifdef STBI_NEON
+
+// NEON integer IDCT. should produce bit-identical
+// results to the generic C version.
+static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) {
+    int16x8_t row0, row1, row2, row3, row4, row5, row6, row7;
+
+    int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f));
+    int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f));
+    int16x4_t rot0_2 = vdup_n_s16(stbi__f2f(0.765366865f));
+    int16x4_t rot1_0 = vdup_n_s16(stbi__f2f(1.175875602f));
+    int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f));
+    int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f));
+    int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f));
+    int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f));
+    int16x4_t rot3_0 = vdup_n_s16(stbi__f2f(0.298631336f));
+    int16x4_t rot3_1 = vdup_n_s16(stbi__f2f(2.053119869f));
+    int16x4_t rot3_2 = vdup_n_s16(stbi__f2f(3.072711026f));
+    int16x4_t rot3_3 = vdup_n_s16(stbi__f2f(1.501321110f));
+
+#define dct_long_mul(out, inq, coeff)                                                                                          \
+    int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff);                                                                   \
+    int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff)
+
+#define dct_long_mac(out, acc, inq, coeff)                                                                                     \
+    int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff);                                                          \
+    int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff)
+
+#define dct_widen(out, inq)                                                                                                    \
+    int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12);                                                                    \
+    int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12)
+
+// wide add
+#define dct_wadd(out, a, b)                                                                                                    \
+    int32x4_t out##_l = vaddq_s32(a##_l, b##_l);                                                                               \
+    int32x4_t out##_h = vaddq_s32(a##_h, b##_h)
+
+// wide sub
+#define dct_wsub(out, a, b)                                                                                                    \
+    int32x4_t out##_l = vsubq_s32(a##_l, b##_l);                                                                               \
+    int32x4_t out##_h = vsubq_s32(a##_h, b##_h)
+
+// butterfly a/b, then shift using "shiftop" by "s" and pack
+#define dct_bfly32o(out0, out1, a, b, shiftop, s)                                                                              \
+    {                                                                                                                          \
+        dct_wadd(sum, a, b);                                                                                                   \
+        dct_wsub(dif, a, b);                                                                                                   \
+        out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s));                                                             \
+        out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s));                                                             \
+    }
+
+#define dct_pass(shiftop, shift)                                                                                               \
+    {                                                                                                                          \
+        /* even part */                                                                                                        \
+        int16x8_t sum26 = vaddq_s16(row2, row6);                                                                               \
+        dct_long_mul(p1e, sum26, rot0_0);                                                                                      \
+        dct_long_mac(t2e, p1e, row6, rot0_1);                                                                                  \
+        dct_long_mac(t3e, p1e, row2, rot0_2);                                                                                  \
+        int16x8_t sum04 = vaddq_s16(row0, row4);                                                                               \
+        int16x8_t dif04 = vsubq_s16(row0, row4);                                                                               \
+        dct_widen(t0e, sum04);                                                                                                 \
+        dct_widen(t1e, dif04);                                                                                                 \
+        dct_wadd(x0, t0e, t3e);                                                                                                \
+        dct_wsub(x3, t0e, t3e);                                                                                                \
+        dct_wadd(x1, t1e, t2e);                                                                                                \
+        dct_wsub(x2, t1e, t2e);                                                                                                \
+        /* odd part */                                                                                                         \
+        int16x8_t sum15 = vaddq_s16(row1, row5);                                                                               \
+        int16x8_t sum17 = vaddq_s16(row1, row7);                                                                               \
+        int16x8_t sum35 = vaddq_s16(row3, row5);                                                                               \
+        int16x8_t sum37 = vaddq_s16(row3, row7);                                                                               \
+        int16x8_t sumodd = vaddq_s16(sum17, sum35);                                                                            \
+        dct_long_mul(p5o, sumodd, rot1_0);                                                                                     \
+        dct_long_mac(p1o, p5o, sum17, rot1_1);                                                                                 \
+        dct_long_mac(p2o, p5o, sum35, rot1_2);                                                                                 \
+        dct_long_mul(p3o, sum37, rot2_0);                                                                                      \
+        dct_long_mul(p4o, sum15, rot2_1);                                                                                      \
+        dct_wadd(sump13o, p1o, p3o);                                                                                           \
+        dct_wadd(sump24o, p2o, p4o);                                                                                           \
+        dct_wadd(sump23o, p2o, p3o);                                                                                           \
+        dct_wadd(sump14o, p1o, p4o);                                                                                           \
+        dct_long_mac(x4, sump13o, row7, rot3_0);                                                                               \
+        dct_long_mac(x5, sump24o, row5, rot3_1);                                                                               \
+        dct_long_mac(x6, sump23o, row3, rot3_2);                                                                               \
+        dct_long_mac(x7, sump14o, row1, rot3_3);                                                                               \
+        dct_bfly32o(row0, row7, x0, x7, shiftop, shift);                                                                       \
+        dct_bfly32o(row1, row6, x1, x6, shiftop, shift);                                                                       \
+        dct_bfly32o(row2, row5, x2, x5, shiftop, shift);                                                                       \
+        dct_bfly32o(row3, row4, x3, x4, shiftop, shift);                                                                       \
+    }
+
+    // load
+    row0 = vld1q_s16(data + 0 * 8);
+    row1 = vld1q_s16(data + 1 * 8);
+    row2 = vld1q_s16(data + 2 * 8);
+    row3 = vld1q_s16(data + 3 * 8);
+    row4 = vld1q_s16(data + 4 * 8);
+    row5 = vld1q_s16(data + 5 * 8);
+    row6 = vld1q_s16(data + 6 * 8);
+    row7 = vld1q_s16(data + 7 * 8);
+
+    // add DC bias
+    row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0));
+
+    // column pass
+    dct_pass(vrshrn_n_s32, 10);
+
+    // 16bit 8x8 transpose
+    {
+// these three map to a single VTRN.16, VTRN.32, and VSWP, respectively.
+// whether compilers actually get this is another story, sadly.
+#define dct_trn16(x, y)                                                                                                        \
+    {                                                                                                                          \
+        int16x8x2_t t = vtrnq_s16(x, y);                                                                                       \
+        x = t.val[0];                                                                                                          \
+        y = t.val[1];                                                                                                          \
+    }
+#define dct_trn32(x, y)                                                                                                        \
+    {                                                                                                                          \
+        int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y));                                         \
+        x = vreinterpretq_s16_s32(t.val[0]);                                                                                   \
+        y = vreinterpretq_s16_s32(t.val[1]);                                                                                   \
+    }
+#define dct_trn64(x, y)                                                                                                        \
+    {                                                                                                                          \
+        int16x8_t x0 = x;                                                                                                      \
+        int16x8_t y0 = y;                                                                                                      \
+        x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0));                                                                  \
+        y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0));                                                                \
+    }
+
+        // pass 1
+        dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6
+        dct_trn16(row2, row3);
+        dct_trn16(row4, row5);
+        dct_trn16(row6, row7);
+
+        // pass 2
+        dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4
+        dct_trn32(row1, row3);
+        dct_trn32(row4, row6);
+        dct_trn32(row5, row7);
+
+        // pass 3
+        dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0
+        dct_trn64(row1, row5);
+        dct_trn64(row2, row6);
+        dct_trn64(row3, row7);
+
+#undef dct_trn16
+#undef dct_trn32
+#undef dct_trn64
+    }
+
+    // row pass
+    // vrshrn_n_s32 only supports shifts up to 16, we need
+    // 17. so do a non-rounding shift of 16 first then follow
+    // up with a rounding shift by 1.
+    dct_pass(vshrn_n_s32, 16);
+
+    {
+        // pack and round
+        uint8x8_t p0 = vqrshrun_n_s16(row0, 1);
+        uint8x8_t p1 = vqrshrun_n_s16(row1, 1);
+        uint8x8_t p2 = vqrshrun_n_s16(row2, 1);
+        uint8x8_t p3 = vqrshrun_n_s16(row3, 1);
+        uint8x8_t p4 = vqrshrun_n_s16(row4, 1);
+        uint8x8_t p5 = vqrshrun_n_s16(row5, 1);
+        uint8x8_t p6 = vqrshrun_n_s16(row6, 1);
+        uint8x8_t p7 = vqrshrun_n_s16(row7, 1);
+
+        // again, these can translate into one instruction, but often don't.
+#define dct_trn8_8(x, y)                                                                                                       \
+    {                                                                                                                          \
+        uint8x8x2_t t = vtrn_u8(x, y);                                                                                         \
+        x = t.val[0];                                                                                                          \
+        y = t.val[1];                                                                                                          \
+    }
+#define dct_trn8_16(x, y)                                                                                                      \
+    {                                                                                                                          \
+        uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y));                                             \
+        x = vreinterpret_u8_u16(t.val[0]);                                                                                     \
+        y = vreinterpret_u8_u16(t.val[1]);                                                                                     \
+    }
+#define dct_trn8_32(x, y)                                                                                                      \
+    {                                                                                                                          \
+        uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y));                                             \
+        x = vreinterpret_u8_u32(t.val[0]);                                                                                     \
+        y = vreinterpret_u8_u32(t.val[1]);                                                                                     \
+    }
+
+        // sadly can't use interleaved stores here since we only write
+        // 8 bytes to each scan line!
+
+        // 8x8 8-bit transpose pass 1
+        dct_trn8_8(p0, p1);
+        dct_trn8_8(p2, p3);
+        dct_trn8_8(p4, p5);
+        dct_trn8_8(p6, p7);
+
+        // pass 2
+        dct_trn8_16(p0, p2);
+        dct_trn8_16(p1, p3);
+        dct_trn8_16(p4, p6);
+        dct_trn8_16(p5, p7);
+
+        // pass 3
+        dct_trn8_32(p0, p4);
+        dct_trn8_32(p1, p5);
+        dct_trn8_32(p2, p6);
+        dct_trn8_32(p3, p7);
+
+        // store
+        vst1_u8(out, p0);
+        out += out_stride;
+        vst1_u8(out, p1);
+        out += out_stride;
+        vst1_u8(out, p2);
+        out += out_stride;
+        vst1_u8(out, p3);
+        out += out_stride;
+        vst1_u8(out, p4);
+        out += out_stride;
+        vst1_u8(out, p5);
+        out += out_stride;
+        vst1_u8(out, p6);
+        out += out_stride;
+        vst1_u8(out, p7);
+
+#undef dct_trn8_8
+#undef dct_trn8_16
+#undef dct_trn8_32
+    }
+
+#undef dct_long_mul
+#undef dct_long_mac
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_pass
+}
+
+#endif // STBI_NEON
+
+#define STBI__MARKER_none 0xff
+// if there's a pending marker from the entropy stream, return that
+// otherwise, fetch from the stream and get a marker. if there's no
+// marker, return 0xff, which is never a valid marker value
+static stbi_uc stbi__get_marker(stbi__jpeg * j) {
+    stbi_uc x;
+    if (j->marker != STBI__MARKER_none) {
+        x = j->marker;
+        j->marker = STBI__MARKER_none;
+        return x;
+    }
+    x = stbi__get8(j->s);
+    if (x != 0xff)
+        return STBI__MARKER_none;
+    while (x == 0xff)
+        x = stbi__get8(j->s); // consume repeated 0xff fill bytes
+    return x;
+}
+
+// in each scan, we'll have scan_n components, and the order
+// of the components is specified by order[]
+#define STBI__RESTART(x) ((x) >= 0xd0 && (x) <= 0xd7)
+
+// after a restart interval, stbi__jpeg_reset the entropy decoder and
+// the dc prediction
+static void stbi__jpeg_reset(stbi__jpeg * j) {
+    j->code_bits = 0;
+    j->code_buffer = 0;
+    j->nomore = 0;
+    j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0;
+    j->marker = STBI__MARKER_none;
+    j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff;
+    j->eob_run = 0;
+    // no more than 1<<31 MCUs if no restart_interal? that's plenty safe,
+    // since we don't even allow 1<<30 pixels
+}
+
+static int stbi__parse_entropy_coded_data(stbi__jpeg * z) {
+    stbi__jpeg_reset(z);
+    if (!z->progressive) {
+        if (z->scan_n == 1) {
+            int i, j;
+            STBI_SIMD_ALIGN(short, data[64]);
+            int n = z->order[0];
+            // non-interleaved data, we just need to process one block at a time,
+            // in trivial scanline order
+            // number of blocks to do just depends on how many actual "pixels" this
+            // component has, independent of interleaved MCU blocking and such
+            int w = (z->img_comp[n].x + 7) >> 3;
+            int h = (z->img_comp[n].y + 7) >> 3;
+            for (j = 0; j < h; ++j) {
+                for (i = 0; i < w; ++i) {
+                    int ha = z->img_comp[n].ha;
+                    if (!stbi__jpeg_decode_block(z, data, z->huff_dc + z->img_comp[n].hd, z->huff_ac + ha, z->fast_ac[ha], n,
+                                                 z->dequant[z->img_comp[n].tq]))
+                        return 0;
+                    z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * j * 8 + i * 8, z->img_comp[n].w2, data);
+                    // every data block is an MCU, so countdown the restart interval
+                    if (--z->todo <= 0) {
+                        if (z->code_bits < 24)
+                            stbi__grow_buffer_unsafe(z);
+                        // if it's NOT a restart, then just bail, so we get corrupt data
+                        // rather than no data
+                        if (!STBI__RESTART(z->marker))
+                            return 1;
+                        stbi__jpeg_reset(z);
+                    }
+                }
+            }
+            return 1;
+        } else { // interleaved
+            int i, j, k, x, y;
+            STBI_SIMD_ALIGN(short, data[64]);
+            for (j = 0; j < z->img_mcu_y; ++j) {
+                for (i = 0; i < z->img_mcu_x; ++i) {
+                    // scan an interleaved mcu... process scan_n components in order
+                    for (k = 0; k < z->scan_n; ++k) {
+                        int n = z->order[k];
+                        // scan out an mcu's worth of this component; that's just determined
+                        // by the basic H and V specified for the component
+                        for (y = 0; y < z->img_comp[n].v; ++y) {
+                            for (x = 0; x < z->img_comp[n].h; ++x) {
+                                int x2 = (i * z->img_comp[n].h + x) * 8;
+                                int y2 = (j * z->img_comp[n].v + y) * 8;
+                                int ha = z->img_comp[n].ha;
+                                if (!stbi__jpeg_decode_block(z, data, z->huff_dc + z->img_comp[n].hd, z->huff_ac + ha,
+                                                             z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq]))
+                                    return 0;
+                                z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * y2 + x2, z->img_comp[n].w2,
+                                                     data);
+                            }
+                        }
+                    }
+                    // after all interleaved components, that's an interleaved MCU,
+                    // so now count down the restart interval
+                    if (--z->todo <= 0) {
+                        if (z->code_bits < 24)
+                            stbi__grow_buffer_unsafe(z);
+                        if (!STBI__RESTART(z->marker))
+                            return 1;
+                        stbi__jpeg_reset(z);
+                    }
+                }
+            }
+            return 1;
+        }
+    } else {
+        if (z->scan_n == 1) {
+            int i, j;
+            int n = z->order[0];
+            // non-interleaved data, we just need to process one block at a time,
+            // in trivial scanline order
+            // number of blocks to do just depends on how many actual "pixels" this
+            // component has, independent of interleaved MCU blocking and such
+            int w = (z->img_comp[n].x + 7) >> 3;
+            int h = (z->img_comp[n].y + 7) >> 3;
+            for (j = 0; j < h; ++j) {
+                for (i = 0; i < w; ++i) {
+                    short * data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+                    if (z->spec_start == 0) {
+                        if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                            return 0;
+                    } else {
+                        int ha = z->img_comp[n].ha;
+                        if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha]))
+                            return 0;
+                    }
+                    // every data block is an MCU, so countdown the restart interval
+                    if (--z->todo <= 0) {
+                        if (z->code_bits < 24)
+                            stbi__grow_buffer_unsafe(z);
+                        if (!STBI__RESTART(z->marker))
+                            return 1;
+                        stbi__jpeg_reset(z);
+                    }
+                }
+            }
+            return 1;
+        } else { // interleaved
+            int i, j, k, x, y;
+            for (j = 0; j < z->img_mcu_y; ++j) {
+                for (i = 0; i < z->img_mcu_x; ++i) {
+                    // scan an interleaved mcu... process scan_n components in order
+                    for (k = 0; k < z->scan_n; ++k) {
+                        int n = z->order[k];
+                        // scan out an mcu's worth of this component; that's just determined
+                        // by the basic H and V specified for the component
+                        for (y = 0; y < z->img_comp[n].v; ++y) {
+                            for (x = 0; x < z->img_comp[n].h; ++x) {
+                                int x2 = (i * z->img_comp[n].h + x);
+                                int y2 = (j * z->img_comp[n].v + y);
+                                short * data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w);
+                                if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                                    return 0;
+                            }
+                        }
+                    }
+                    // after all interleaved components, that's an interleaved MCU,
+                    // so now count down the restart interval
+                    if (--z->todo <= 0) {
+                        if (z->code_bits < 24)
+                            stbi__grow_buffer_unsafe(z);
+                        if (!STBI__RESTART(z->marker))
+                            return 1;
+                        stbi__jpeg_reset(z);
+                    }
+                }
+            }
+            return 1;
+        }
+    }
+}
+
+static void stbi__jpeg_dequantize(short * data, stbi__uint16 * dequant) {
+    int i;
+    for (i = 0; i < 64; ++i)
+        data[i] *= dequant[i];
+}
+
+static void stbi__jpeg_finish(stbi__jpeg * z) {
+    if (z->progressive) {
+        // dequantize and idct the data
+        int i, j, n;
+        for (n = 0; n < z->s->img_n; ++n) {
+            int w = (z->img_comp[n].x + 7) >> 3;
+            int h = (z->img_comp[n].y + 7) >> 3;
+            for (j = 0; j < h; ++j) {
+                for (i = 0; i < w; ++i) {
+                    short * data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+                    stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]);
+                    z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * j * 8 + i * 8, z->img_comp[n].w2, data);
+                }
+            }
+        }
+    }
+}
+
+static int stbi__process_marker(stbi__jpeg * z, int m) {
+    int L;
+    switch (m) {
+    case STBI__MARKER_none: // no marker found
+        return stbi__err("expected marker", "Corrupt JPEG");
+
+    case 0xDD: // DRI - specify restart interval
+        if (stbi__get16be(z->s) != 4)
+            return stbi__err("bad DRI len", "Corrupt JPEG");
+        z->restart_interval = stbi__get16be(z->s);
+        return 1;
+
+    case 0xDB: // DQT - define quantization table
+        L = stbi__get16be(z->s) - 2;
+        while (L > 0) {
+            int q = stbi__get8(z->s);
+            int p = q >> 4, sixteen = (p != 0);
+            int t = q & 15, i;
+            if (p != 0 && p != 1)
+                return stbi__err("bad DQT type", "Corrupt JPEG");
+            if (t > 3)
+                return stbi__err("bad DQT table", "Corrupt JPEG");
+
+            for (i = 0; i < 64; ++i)
+                z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s));
+            L -= (sixteen ? 129 : 65);
+        }
+        return L == 0;
+
+    case 0xC4: // DHT - define huffman table
+        L = stbi__get16be(z->s) - 2;
+        while (L > 0) {
+            stbi_uc * v;
+            int sizes[16], i, n = 0;
+            int q = stbi__get8(z->s);
+            int tc = q >> 4;
+            int th = q & 15;
+            if (tc > 1 || th > 3)
+                return stbi__err("bad DHT header", "Corrupt JPEG");
+            for (i = 0; i < 16; ++i) {
+                sizes[i] = stbi__get8(z->s);
+                n += sizes[i];
+            }
+            if (n > 256)
+                return stbi__err("bad DHT header", "Corrupt JPEG"); // Loop over i < n would write past end of values!
+            L -= 17;
+            if (tc == 0) {
+                if (!stbi__build_huffman(z->huff_dc + th, sizes))
+                    return 0;
+                v = z->huff_dc[th].values;
+            } else {
+                if (!stbi__build_huffman(z->huff_ac + th, sizes))
+                    return 0;
+                v = z->huff_ac[th].values;
+            }
+            for (i = 0; i < n; ++i)
+                v[i] = stbi__get8(z->s);
+            if (tc != 0)
+                stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th);
+            L -= n;
+        }
+        return L == 0;
+    }
+
+    // check for comment block or APP blocks
+    if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) {
+        L = stbi__get16be(z->s);
+        if (L < 2) {
+            if (m == 0xFE)
+                return stbi__err("bad COM len", "Corrupt JPEG");
+            else
+                return stbi__err("bad APP len", "Corrupt JPEG");
+        }
+        L -= 2;
+
+        if (m == 0xE0 && L >= 5) { // JFIF APP0 segment
+            static const unsigned char tag[5] = {'J', 'F', 'I', 'F', '\0'};
+            int ok = 1;
+            int i;
+            for (i = 0; i < 5; ++i)
+                if (stbi__get8(z->s) != tag[i])
+                    ok = 0;
+            L -= 5;
+            if (ok)
+                z->jfif = 1;
+        } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment
+            static const unsigned char tag[6] = {'A', 'd', 'o', 'b', 'e', '\0'};
+            int ok = 1;
+            int i;
+            for (i = 0; i < 6; ++i)
+                if (stbi__get8(z->s) != tag[i])
+                    ok = 0;
+            L -= 6;
+            if (ok) {
+                stbi__get8(z->s);                            // version
+                stbi__get16be(z->s);                         // flags0
+                stbi__get16be(z->s);                         // flags1
+                z->app14_color_transform = stbi__get8(z->s); // color transform
+                L -= 6;
+            }
+        }
+
+        stbi__skip(z->s, L);
+        return 1;
+    }
+
+    return stbi__err("unknown marker", "Corrupt JPEG");
+}
+
+// after we see SOS
+static int stbi__process_scan_header(stbi__jpeg * z) {
+    int i;
+    int Ls = stbi__get16be(z->s);
+    z->scan_n = stbi__get8(z->s);
+    if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int)z->s->img_n)
+        return stbi__err("bad SOS component count", "Corrupt JPEG");
+    if (Ls != 6 + 2 * z->scan_n)
+        return stbi__err("bad SOS len", "Corrupt JPEG");
+    for (i = 0; i < z->scan_n; ++i) {
+        int id = stbi__get8(z->s), which;
+        int q = stbi__get8(z->s);
+        for (which = 0; which < z->s->img_n; ++which)
+            if (z->img_comp[which].id == id)
+                break;
+        if (which == z->s->img_n)
+            return 0; // no match
+        z->img_comp[which].hd = q >> 4;
+        if (z->img_comp[which].hd > 3)
+            return stbi__err("bad DC huff", "Corrupt JPEG");
+        z->img_comp[which].ha = q & 15;
+        if (z->img_comp[which].ha > 3)
+            return stbi__err("bad AC huff", "Corrupt JPEG");
+        z->order[i] = which;
+    }
+
+    {
+        int aa;
+        z->spec_start = stbi__get8(z->s);
+        z->spec_end = stbi__get8(z->s); // should be 63, but might be 0
+        aa = stbi__get8(z->s);
+        z->succ_high = (aa >> 4);
+        z->succ_low = (aa & 15);
+        if (z->progressive) {
+            if (z->spec_start > 63 || z->spec_end > 63 || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13)
+                return stbi__err("bad SOS", "Corrupt JPEG");
+        } else {
+            if (z->spec_start != 0)
+                return stbi__err("bad SOS", "Corrupt JPEG");
+            if (z->succ_high != 0 || z->succ_low != 0)
+                return stbi__err("bad SOS", "Corrupt JPEG");
+            z->spec_end = 63;
+        }
+    }
+
+    return 1;
+}
+
+static int stbi__free_jpeg_components(stbi__jpeg * z, int ncomp, int why) {
+    int i;
+    for (i = 0; i < ncomp; ++i) {
+        if (z->img_comp[i].raw_data) {
+            STBI_FREE(z->img_comp[i].raw_data);
+            z->img_comp[i].raw_data = NULL;
+            z->img_comp[i].data = NULL;
+        }
+        if (z->img_comp[i].raw_coeff) {
+            STBI_FREE(z->img_comp[i].raw_coeff);
+            z->img_comp[i].raw_coeff = 0;
+            z->img_comp[i].coeff = 0;
+        }
+        if (z->img_comp[i].linebuf) {
+            STBI_FREE(z->img_comp[i].linebuf);
+            z->img_comp[i].linebuf = NULL;
+        }
+    }
+    return why;
+}
+
+static int stbi__process_frame_header(stbi__jpeg * z, int scan) {
+    stbi__context * s = z->s;
+    int Lf, p, i, q, h_max = 1, v_max = 1, c;
+    Lf = stbi__get16be(s);
+    if (Lf < 11)
+        return stbi__err("bad SOF len", "Corrupt JPEG"); // JPEG
+    p = stbi__get8(s);
+    if (p != 8)
+        return stbi__err("only 8-bit", "JPEG format not supported: 8-bit only"); // JPEG baseline
+    s->img_y = stbi__get16be(s);
+    if (s->img_y == 0)
+        return stbi__err("no header height",
+                         "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG
+    s->img_x = stbi__get16be(s);
+    if (s->img_x == 0)
+        return stbi__err("0 width", "Corrupt JPEG"); // JPEG requires
+    if (s->img_y > STBI_MAX_DIMENSIONS)
+        return stbi__err("too large", "Very large image (corrupt?)");
+    if (s->img_x > STBI_MAX_DIMENSIONS)
+        return stbi__err("too large", "Very large image (corrupt?)");
+    c = stbi__get8(s);
+    if (c != 3 && c != 1 && c != 4)
+        return stbi__err("bad component count", "Corrupt JPEG");
+    s->img_n = c;
+    for (i = 0; i < c; ++i) {
+        z->img_comp[i].data = NULL;
+        z->img_comp[i].linebuf = NULL;
+    }
+
+    if (Lf != 8 + 3 * s->img_n)
+        return stbi__err("bad SOF len", "Corrupt JPEG");
+
+    z->rgb = 0;
+    for (i = 0; i < s->img_n; ++i) {
+        static const unsigned char rgb[3] = {'R', 'G', 'B'};
+        z->img_comp[i].id = stbi__get8(s);
+        if (s->img_n == 3 && z->img_comp[i].id == rgb[i])
+            ++z->rgb;
+        q = stbi__get8(s);
+        z->img_comp[i].h = (q >> 4);
+        if (!z->img_comp[i].h || z->img_comp[i].h > 4)
+            return stbi__err("bad H", "Corrupt JPEG");
+        z->img_comp[i].v = q & 15;
+        if (!z->img_comp[i].v || z->img_comp[i].v > 4)
+            return stbi__err("bad V", "Corrupt JPEG");
+        z->img_comp[i].tq = stbi__get8(s);
+        if (z->img_comp[i].tq > 3)
+            return stbi__err("bad TQ", "Corrupt JPEG");
+    }
+
+    if (scan != STBI__SCAN_load)
+        return 1;
+
+    if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0))
+        return stbi__err("too large", "Image too large to decode");
+
+    for (i = 0; i < s->img_n; ++i) {
+        if (z->img_comp[i].h > h_max)
+            h_max = z->img_comp[i].h;
+        if (z->img_comp[i].v > v_max)
+            v_max = z->img_comp[i].v;
+    }
+
+    // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios
+    // and I've never seen a non-corrupted JPEG file actually use them
+    for (i = 0; i < s->img_n; ++i) {
+        if (h_max % z->img_comp[i].h != 0)
+            return stbi__err("bad H", "Corrupt JPEG");
+        if (v_max % z->img_comp[i].v != 0)
+            return stbi__err("bad V", "Corrupt JPEG");
+    }
+
+    // compute interleaved mcu info
+    z->img_h_max = h_max;
+    z->img_v_max = v_max;
+    z->img_mcu_w = h_max * 8;
+    z->img_mcu_h = v_max * 8;
+    // these sizes can't be more than 17 bits
+    z->img_mcu_x = (s->img_x + z->img_mcu_w - 1) / z->img_mcu_w;
+    z->img_mcu_y = (s->img_y + z->img_mcu_h - 1) / z->img_mcu_h;
+
+    for (i = 0; i < s->img_n; ++i) {
+        // number of effective pixels (e.g. for non-interleaved MCU)
+        z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max - 1) / h_max;
+        z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max - 1) / v_max;
+        // to simplify generation, we'll allocate enough memory to decode
+        // the bogus oversized data from using interleaved MCUs and their
+        // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't
+        // discard the extra data until colorspace conversion
+        //
+        // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier)
+        // so these muls can't overflow with 32-bit ints (which we require)
+        z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8;
+        z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8;
+        z->img_comp[i].coeff = 0;
+        z->img_comp[i].raw_coeff = 0;
+        z->img_comp[i].linebuf = NULL;
+        z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15);
+        if (z->img_comp[i].raw_data == NULL)
+            return stbi__free_jpeg_components(z, i + 1, stbi__err("outofmem", "Out of memory"));
+        // align blocks for idct using mmx/sse
+        z->img_comp[i].data = (stbi_uc *)(((size_t)z->img_comp[i].raw_data + 15) & ~15);
+        if (z->progressive) {
+            // w2, h2 are multiples of 8 (see above)
+            z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8;
+            z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8;
+            z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15);
+            if (z->img_comp[i].raw_coeff == NULL)
+                return stbi__free_jpeg_components(z, i + 1, stbi__err("outofmem", "Out of memory"));
+            z->img_comp[i].coeff = (short *)(((size_t)z->img_comp[i].raw_coeff + 15) & ~15);
+        }
+    }
+
+    return 1;
+}
+
+// use comparisons since in some cases we handle more than one case (e.g. SOF)
+#define stbi__DNL(x) ((x) == 0xdc)
+#define stbi__SOI(x) ((x) == 0xd8)
+#define stbi__EOI(x) ((x) == 0xd9)
+#define stbi__SOF(x) ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2)
+#define stbi__SOS(x) ((x) == 0xda)
+
+#define stbi__SOF_progressive(x) ((x) == 0xc2)
+
+static int stbi__decode_jpeg_header(stbi__jpeg * z, int scan) {
+    int m;
+    z->jfif = 0;
+    z->app14_color_transform = -1; // valid values are 0,1,2
+    z->marker = STBI__MARKER_none; // initialize cached marker to empty
+    m = stbi__get_marker(z);
+    if (!stbi__SOI(m))
+        return stbi__err("no SOI", "Corrupt JPEG");
+    if (scan == STBI__SCAN_type)
+        return 1;
+    m = stbi__get_marker(z);
+    while (!stbi__SOF(m)) {
+        if (!stbi__process_marker(z, m))
+            return 0;
+        m = stbi__get_marker(z);
+        while (m == STBI__MARKER_none) {
+            // some files have extra padding after their blocks, so ok, we'll scan
+            if (stbi__at_eof(z->s))
+                return stbi__err("no SOF", "Corrupt JPEG");
+            m = stbi__get_marker(z);
+        }
+    }
+    z->progressive = stbi__SOF_progressive(m);
+    if (!stbi__process_frame_header(z, scan))
+        return 0;
+    return 1;
+}
+
+static int stbi__skip_jpeg_junk_at_end(stbi__jpeg * j) {
+    // some JPEGs have junk at end, skip over it but if we find what looks
+    // like a valid marker, resume there
+    while (!stbi__at_eof(j->s)) {
+        int x = stbi__get8(j->s);
+        while (x == 255) { // might be a marker
+            if (stbi__at_eof(j->s))
+                return STBI__MARKER_none;
+            x = stbi__get8(j->s);
+            if (x != 0x00 && x != 0xff) {
+                // not a stuffed zero or lead-in to another marker, looks
+                // like an actual marker, return it
+                return x;
+            }
+            // stuffed zero has x=0 now which ends the loop, meaning we go
+            // back to regular scan loop.
+            // repeated 0xff keeps trying to read the next byte of the marker.
+        }
+    }
+    return STBI__MARKER_none;
+}
+
+// decode image to YCbCr format
+static int stbi__decode_jpeg_image(stbi__jpeg * j) {
+    int m;
+    for (m = 0; m < 4; m++) {
+        j->img_comp[m].raw_data = NULL;
+        j->img_comp[m].raw_coeff = NULL;
+    }
+    j->restart_interval = 0;
+    if (!stbi__decode_jpeg_header(j, STBI__SCAN_load))
+        return 0;
+    m = stbi__get_marker(j);
+    while (!stbi__EOI(m)) {
+        if (stbi__SOS(m)) {
+            if (!stbi__process_scan_header(j))
+                return 0;
+            if (!stbi__parse_entropy_coded_data(j))
+                return 0;
+            if (j->marker == STBI__MARKER_none) {
+                j->marker = stbi__skip_jpeg_junk_at_end(j);
+                // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0
+            }
+            m = stbi__get_marker(j);
+            if (STBI__RESTART(m))
+                m = stbi__get_marker(j);
+        } else if (stbi__DNL(m)) {
+            int Ld = stbi__get16be(j->s);
+            stbi__uint32 NL = stbi__get16be(j->s);
+            if (Ld != 4)
+                return stbi__err("bad DNL len", "Corrupt JPEG");
+            if (NL != j->s->img_y)
+                return stbi__err("bad DNL height", "Corrupt JPEG");
+            m = stbi__get_marker(j);
+        } else {
+            if (!stbi__process_marker(j, m))
+                return 1;
+            m = stbi__get_marker(j);
+        }
+    }
+    if (j->progressive)
+        stbi__jpeg_finish(j);
+    return 1;
+}
+
+// static jfif-centered resampling (across block boundaries)
+
+typedef stbi_uc * (*resample_row_func)(stbi_uc * out, stbi_uc * in0, stbi_uc * in1, int w, int hs);
+
+#define stbi__div4(x) ((stbi_uc)((x) >> 2))
+
+static stbi_uc * resample_row_1(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) {
+    STBI_NOTUSED(out);
+    STBI_NOTUSED(in_far);
+    STBI_NOTUSED(w);
+    STBI_NOTUSED(hs);
+    return in_near;
+}
+
+static stbi_uc * stbi__resample_row_v_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) {
+    // need to generate two samples vertically for every one in input
+    int i;
+    STBI_NOTUSED(hs);
+    for (i = 0; i < w; ++i)
+        out[i] = stbi__div4(3 * in_near[i] + in_far[i] + 2);
+    return out;
+}
+
+static stbi_uc * stbi__resample_row_h_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) {
+    // need to generate two samples horizontally for every one in input
+    int i;
+    stbi_uc * input = in_near;
+
+    if (w == 1) {
+        // if only one sample, can't do any interpolation
+        out[0] = out[1] = input[0];
+        return out;
+    }
+
+    out[0] = input[0];
+    out[1] = stbi__div4(input[0] * 3 + input[1] + 2);
+    for (i = 1; i < w - 1; ++i) {
+        int n = 3 * input[i] + 2;
+        out[i * 2 + 0] = stbi__div4(n + input[i - 1]);
+        out[i * 2 + 1] = stbi__div4(n + input[i + 1]);
+    }
+    out[i * 2 + 0] = stbi__div4(input[w - 2] * 3 + input[w - 1] + 2);
+    out[i * 2 + 1] = input[w - 1];
+
+    STBI_NOTUSED(in_far);
+    STBI_NOTUSED(hs);
+
+    return out;
+}
+
+#define stbi__div16(x) ((stbi_uc)((x) >> 4))
+
+static stbi_uc * stbi__resample_row_hv_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) {
+    // need to generate 2x2 samples for every one in input
+    int i, t0, t1;
+    if (w == 1) {
+        out[0] = out[1] = stbi__div4(3 * in_near[0] + in_far[0] + 2);
+        return out;
+    }
+
+    t1 = 3 * in_near[0] + in_far[0];
+    out[0] = stbi__div4(t1 + 2);
+    for (i = 1; i < w; ++i) {
+        t0 = t1;
+        t1 = 3 * in_near[i] + in_far[i];
+        out[i * 2 - 1] = stbi__div16(3 * t0 + t1 + 8);
+        out[i * 2] = stbi__div16(3 * t1 + t0 + 8);
+    }
+    out[w * 2 - 1] = stbi__div4(t1 + 2);
+
+    STBI_NOTUSED(hs);
+
+    return out;
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static stbi_uc * stbi__resample_row_hv_2_simd(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) {
+    // need to generate 2x2 samples for every one in input
+    int i = 0, t0, t1;
+
+    if (w == 1) {
+        out[0] = out[1] = stbi__div4(3 * in_near[0] + in_far[0] + 2);
+        return out;
+    }
+
+    t1 = 3 * in_near[0] + in_far[0];
+    // process groups of 8 pixels for as long as we can.
+    // note we can't handle the last pixel in a row in this loop
+    // because we need to handle the filter boundary conditions.
+    for (; i < ((w - 1) & ~7); i += 8) {
+#if defined(STBI_SSE2)
+        // load and perform the vertical filtering pass
+        // this uses 3*x + y = 4*x + (y - x)
+        __m128i zero = _mm_setzero_si128();
+        __m128i farb = _mm_loadl_epi64((__m128i *)(in_far + i));
+        __m128i nearb = _mm_loadl_epi64((__m128i *)(in_near + i));
+        __m128i farw = _mm_unpacklo_epi8(farb, zero);
+        __m128i nearw = _mm_unpacklo_epi8(nearb, zero);
+        __m128i diff = _mm_sub_epi16(farw, nearw);
+        __m128i nears = _mm_slli_epi16(nearw, 2);
+        __m128i curr = _mm_add_epi16(nears, diff); // current row
+
+        // horizontal filter works the same based on shifted vers of current
+        // row. "prev" is current row shifted right by 1 pixel; we need to
+        // insert the previous pixel value (from t1).
+        // "next" is current row shifted left by 1 pixel, with first pixel
+        // of next block of 8 pixels added in.
+        __m128i prv0 = _mm_slli_si128(curr, 2);
+        __m128i nxt0 = _mm_srli_si128(curr, 2);
+        __m128i prev = _mm_insert_epi16(prv0, t1, 0);
+        __m128i next = _mm_insert_epi16(nxt0, 3 * in_near[i + 8] + in_far[i + 8], 7);
+
+        // horizontal filter, polyphase implementation since it's convenient:
+        // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+        // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+        // note the shared term.
+        __m128i bias = _mm_set1_epi16(8);
+        __m128i curs = _mm_slli_epi16(curr, 2);
+        __m128i prvd = _mm_sub_epi16(prev, curr);
+        __m128i nxtd = _mm_sub_epi16(next, curr);
+        __m128i curb = _mm_add_epi16(curs, bias);
+        __m128i even = _mm_add_epi16(prvd, curb);
+        __m128i odd = _mm_add_epi16(nxtd, curb);
+
+        // interleave even and odd pixels, then undo scaling.
+        __m128i int0 = _mm_unpacklo_epi16(even, odd);
+        __m128i int1 = _mm_unpackhi_epi16(even, odd);
+        __m128i de0 = _mm_srli_epi16(int0, 4);
+        __m128i de1 = _mm_srli_epi16(int1, 4);
+
+        // pack and write output
+        __m128i outv = _mm_packus_epi16(de0, de1);
+        _mm_storeu_si128((__m128i *)(out + i * 2), outv);
+#elif defined(STBI_NEON)
+        // load and perform the vertical filtering pass
+        // this uses 3*x + y = 4*x + (y - x)
+        uint8x8_t farb = vld1_u8(in_far + i);
+        uint8x8_t nearb = vld1_u8(in_near + i);
+        int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(farb, nearb));
+        int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2));
+        int16x8_t curr = vaddq_s16(nears, diff); // current row
+
+        // horizontal filter works the same based on shifted vers of current
+        // row. "prev" is current row shifted right by 1 pixel; we need to
+        // insert the previous pixel value (from t1).
+        // "next" is current row shifted left by 1 pixel, with first pixel
+        // of next block of 8 pixels added in.
+        int16x8_t prv0 = vextq_s16(curr, curr, 7);
+        int16x8_t nxt0 = vextq_s16(curr, curr, 1);
+        int16x8_t prev = vsetq_lane_s16(t1, prv0, 0);
+        int16x8_t next = vsetq_lane_s16(3 * in_near[i + 8] + in_far[i + 8], nxt0, 7);
+
+        // horizontal filter, polyphase implementation since it's convenient:
+        // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+        // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+        // note the shared term.
+        int16x8_t curs = vshlq_n_s16(curr, 2);
+        int16x8_t prvd = vsubq_s16(prev, curr);
+        int16x8_t nxtd = vsubq_s16(next, curr);
+        int16x8_t even = vaddq_s16(curs, prvd);
+        int16x8_t odd = vaddq_s16(curs, nxtd);
+
+        // undo scaling and round, then store with even/odd phases interleaved
+        uint8x8x2_t o;
+        o.val[0] = vqrshrun_n_s16(even, 4);
+        o.val[1] = vqrshrun_n_s16(odd, 4);
+        vst2_u8(out + i * 2, o);
+#endif
+
+        // "previous" value for next iter
+        t1 = 3 * in_near[i + 7] + in_far[i + 7];
+    }
+
+    t0 = t1;
+    t1 = 3 * in_near[i] + in_far[i];
+    out[i * 2] = stbi__div16(3 * t1 + t0 + 8);
+
+    for (++i; i < w; ++i) {
+        t0 = t1;
+        t1 = 3 * in_near[i] + in_far[i];
+        out[i * 2 - 1] = stbi__div16(3 * t0 + t1 + 8);
+        out[i * 2] = stbi__div16(3 * t1 + t0 + 8);
+    }
+    out[w * 2 - 1] = stbi__div4(t1 + 2);
+
+    STBI_NOTUSED(hs);
+
+    return out;
+}
+#endif
+
+static stbi_uc * stbi__resample_row_generic(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) {
+    // resample with nearest-neighbor
+    int i, j;
+    STBI_NOTUSED(in_far);
+    for (i = 0; i < w; ++i)
+        for (j = 0; j < hs; ++j)
+            out[i * hs + j] = in_near[i];
+    return out;
+}
+
+// this is a reduced-precision calculation of YCbCr-to-RGB introduced
+// to make sure the code produces the same results in both SIMD and scalar
+#define stbi__float2fixed(x) (((int)((x)*4096.0f + 0.5f)) << 8)
+static void stbi__YCbCr_to_RGB_row(stbi_uc * out, const stbi_uc * y, const stbi_uc * pcb, const stbi_uc * pcr, int count,
+                                   int step) {
+    int i;
+    for (i = 0; i < count; ++i) {
+        int y_fixed = (y[i] << 20) + (1 << 19); // rounding
+        int r, g, b;
+        int cr = pcr[i] - 128;
+        int cb = pcb[i] - 128;
+        r = y_fixed + cr * stbi__float2fixed(1.40200f);
+        g = y_fixed + (cr * -stbi__float2fixed(0.71414f)) + ((cb * -stbi__float2fixed(0.34414f)) & 0xffff0000);
+        b = y_fixed + cb * stbi__float2fixed(1.77200f);
+        r >>= 20;
+        g >>= 20;
+        b >>= 20;
+        if ((unsigned)r > 255) {
+            if (r < 0)
+                r = 0;
+            else
+                r = 255;
+        }
+        if ((unsigned)g > 255) {
+            if (g < 0)
+                g = 0;
+            else
+                g = 255;
+        }
+        if ((unsigned)b > 255) {
+            if (b < 0)
+                b = 0;
+            else
+                b = 255;
+        }
+        out[0] = (stbi_uc)r;
+        out[1] = (stbi_uc)g;
+        out[2] = (stbi_uc)b;
+        out[3] = 255;
+        out += step;
+    }
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static void stbi__YCbCr_to_RGB_simd(stbi_uc * out, stbi_uc const * y, stbi_uc const * pcb, stbi_uc const * pcr, int count,
+                                    int step) {
+    int i = 0;
+
+#ifdef STBI_SSE2
+    // step == 3 is pretty ugly on the final interleave, and i'm not convinced
+    // it's useful in practice (you wouldn't use it for textures, for example).
+    // so just accelerate step == 4 case.
+    if (step == 4) {
+        // this is a fairly straightforward implementation and not super-optimized.
+        __m128i signflip = _mm_set1_epi8(-0x80);
+        __m128i cr_const0 = _mm_set1_epi16((short)(1.40200f * 4096.0f + 0.5f));
+        __m128i cr_const1 = _mm_set1_epi16(-(short)(0.71414f * 4096.0f + 0.5f));
+        __m128i cb_const0 = _mm_set1_epi16(-(short)(0.34414f * 4096.0f + 0.5f));
+        __m128i cb_const1 = _mm_set1_epi16((short)(1.77200f * 4096.0f + 0.5f));
+        __m128i y_bias = _mm_set1_epi8((char)(unsigned char)128);
+        __m128i xw = _mm_set1_epi16(255); // alpha channel
+
+        for (; i + 7 < count; i += 8) {
+            // load
+            __m128i y_bytes = _mm_loadl_epi64((__m128i *)(y + i));
+            __m128i cr_bytes = _mm_loadl_epi64((__m128i *)(pcr + i));
+            __m128i cb_bytes = _mm_loadl_epi64((__m128i *)(pcb + i));
+            __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128
+            __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128
+
+            // unpack to short (and left-shift cr, cb by 8)
+            __m128i yw = _mm_unpacklo_epi8(y_bias, y_bytes);
+            __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased);
+            __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased);
+
+            // color transform
+            __m128i yws = _mm_srli_epi16(yw, 4);
+            __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw);
+            __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw);
+            __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1);
+            __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1);
+            __m128i rws = _mm_add_epi16(cr0, yws);
+            __m128i gwt = _mm_add_epi16(cb0, yws);
+            __m128i bws = _mm_add_epi16(yws, cb1);
+            __m128i gws = _mm_add_epi16(gwt, cr1);
+
+            // descale
+            __m128i rw = _mm_srai_epi16(rws, 4);
+            __m128i bw = _mm_srai_epi16(bws, 4);
+            __m128i gw = _mm_srai_epi16(gws, 4);
+
+            // back to byte, set up for transpose
+            __m128i brb = _mm_packus_epi16(rw, bw);
+            __m128i gxb = _mm_packus_epi16(gw, xw);
+
+            // transpose to interleave channels
+            __m128i t0 = _mm_unpacklo_epi8(brb, gxb);
+            __m128i t1 = _mm_unpackhi_epi8(brb, gxb);
+            __m128i o0 = _mm_unpacklo_epi16(t0, t1);
+            __m128i o1 = _mm_unpackhi_epi16(t0, t1);
+
+            // store
+            _mm_storeu_si128((__m128i *)(out + 0), o0);
+            _mm_storeu_si128((__m128i *)(out + 16), o1);
+            out += 32;
+        }
+    }
+#endif
+
+#ifdef STBI_NEON
+    // in this version, step=3 support would be easy to add. but is there demand?
+    if (step == 4) {
+        // this is a fairly straightforward implementation and not super-optimized.
+        uint8x8_t signflip = vdup_n_u8(0x80);
+        int16x8_t cr_const0 = vdupq_n_s16((short)(1.40200f * 4096.0f + 0.5f));
+        int16x8_t cr_const1 = vdupq_n_s16(-(short)(0.71414f * 4096.0f + 0.5f));
+        int16x8_t cb_const0 = vdupq_n_s16(-(short)(0.34414f * 4096.0f + 0.5f));
+        int16x8_t cb_const1 = vdupq_n_s16((short)(1.77200f * 4096.0f + 0.5f));
+
+        for (; i + 7 < count; i += 8) {
+            // load
+            uint8x8_t y_bytes = vld1_u8(y + i);
+            uint8x8_t cr_bytes = vld1_u8(pcr + i);
+            uint8x8_t cb_bytes = vld1_u8(pcb + i);
+            int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip));
+            int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip));
+
+            // expand to s16
+            int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4));
+            int16x8_t crw = vshll_n_s8(cr_biased, 7);
+            int16x8_t cbw = vshll_n_s8(cb_biased, 7);
+
+            // color transform
+            int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0);
+            int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0);
+            int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1);
+            int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1);
+            int16x8_t rws = vaddq_s16(yws, cr0);
+            int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1);
+            int16x8_t bws = vaddq_s16(yws, cb1);
+
+            // undo scaling, round, convert to byte
+            uint8x8x4_t o;
+            o.val[0] = vqrshrun_n_s16(rws, 4);
+            o.val[1] = vqrshrun_n_s16(gws, 4);
+            o.val[2] = vqrshrun_n_s16(bws, 4);
+            o.val[3] = vdup_n_u8(255);
+
+            // store, interleaving r/g/b/a
+            vst4_u8(out, o);
+            out += 8 * 4;
+        }
+    }
+#endif
+
+    for (; i < count; ++i) {
+        int y_fixed = (y[i] << 20) + (1 << 19); // rounding
+        int r, g, b;
+        int cr = pcr[i] - 128;
+        int cb = pcb[i] - 128;
+        r = y_fixed + cr * stbi__float2fixed(1.40200f);
+        g = y_fixed + cr * -stbi__float2fixed(0.71414f) + ((cb * -stbi__float2fixed(0.34414f)) & 0xffff0000);
+        b = y_fixed + cb * stbi__float2fixed(1.77200f);
+        r >>= 20;
+        g >>= 20;
+        b >>= 20;
+        if ((unsigned)r > 255) {
+            if (r < 0)
+                r = 0;
+            else
+                r = 255;
+        }
+        if ((unsigned)g > 255) {
+            if (g < 0)
+                g = 0;
+            else
+                g = 255;
+        }
+        if ((unsigned)b > 255) {
+            if (b < 0)
+                b = 0;
+            else
+                b = 255;
+        }
+        out[0] = (stbi_uc)r;
+        out[1] = (stbi_uc)g;
+        out[2] = (stbi_uc)b;
+        out[3] = 255;
+        out += step;
+    }
+}
+#endif
+
+// set up the kernels
+static void stbi__setup_jpeg(stbi__jpeg * j) {
+    j->idct_block_kernel = stbi__idct_block;
+    j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row;
+    j->resample_row_hv_2_kernel = stbi__resample_row_hv_2;
+
+#ifdef STBI_SSE2
+    if (stbi__sse2_available()) {
+        j->idct_block_kernel = stbi__idct_simd;
+        j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+        j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+    }
+#endif
+
+#ifdef STBI_NEON
+    j->idct_block_kernel = stbi__idct_simd;
+    j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+    j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+#endif
+}
+
+// clean up the temporary component buffers
+static void stbi__cleanup_jpeg(stbi__jpeg * j) { stbi__free_jpeg_components(j, j->s->img_n, 0); }
+
+typedef struct {
+    resample_row_func resample;
+    stbi_uc *line0, *line1;
+    int hs, vs;  // expansion factor in each axis
+    int w_lores; // horizontal pixels pre-expansion
+    int ystep;   // how far through vertical expansion we are
+    int ypos;    // which pre-expansion row we're on
+} stbi__resample;
+
+// fast 0..255 * 0..255 => 0..255 rounded multiplication
+static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y) {
+    unsigned int t = x * y + 128;
+    return (stbi_uc)((t + (t >> 8)) >> 8);
+}
+
+static stbi_uc * load_jpeg_image(stbi__jpeg * z, int * out_x, int * out_y, int * comp, int req_comp) {
+    int n, decode_n, is_rgb;
+    z->s->img_n = 0; // make stbi__cleanup_jpeg safe
+
+    // validate req_comp
+    if (req_comp < 0 || req_comp > 4)
+        return stbi__errpuc("bad req_comp", "Internal error");
+
+    // load a jpeg image from whichever source, but leave in YCbCr format
+    if (!stbi__decode_jpeg_image(z)) {
+        stbi__cleanup_jpeg(z);
+        return NULL;
+    }
+
+    // determine actual number of components to generate
+    n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1;
+
+    is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif));
+
+    if (z->s->img_n == 3 && n < 3 && !is_rgb)
+        decode_n = 1;
+    else
+        decode_n = z->s->img_n;
+
+    // nothing to do if no components requested; check this now to avoid
+    // accessing uninitialized coutput[0] later
+    if (decode_n <= 0) {
+        stbi__cleanup_jpeg(z);
+        return NULL;
+    }
+
+    // resample and color-convert
+    {
+        int k;
+        unsigned int i, j;
+        stbi_uc * output;
+        stbi_uc * coutput[4] = {NULL, NULL, NULL, NULL};
+
+        stbi__resample res_comp[4];
+
+        for (k = 0; k < decode_n; ++k) {
+            stbi__resample * r = &res_comp[k];
+
+            // allocate line buffer big enough for upsampling off the edges
+            // with upsample factor of 4
+            z->img_comp[k].linebuf = (stbi_uc *)stbi__malloc(z->s->img_x + 3);
+            if (!z->img_comp[k].linebuf) {
+                stbi__cleanup_jpeg(z);
+                return stbi__errpuc("outofmem", "Out of memory");
+            }
+
+            r->hs = z->img_h_max / z->img_comp[k].h;
+            r->vs = z->img_v_max / z->img_comp[k].v;
+            r->ystep = r->vs >> 1;
+            r->w_lores = (z->s->img_x + r->hs - 1) / r->hs;
+            r->ypos = 0;
+            r->line0 = r->line1 = z->img_comp[k].data;
+
+            if (r->hs == 1 && r->vs == 1)
+                r->resample = resample_row_1;
+            else if (r->hs == 1 && r->vs == 2)
+                r->resample = stbi__resample_row_v_2;
+            else if (r->hs == 2 && r->vs == 1)
+                r->resample = stbi__resample_row_h_2;
+            else if (r->hs == 2 && r->vs == 2)
+                r->resample = z->resample_row_hv_2_kernel;
+            else
+                r->resample = stbi__resample_row_generic;
+        }
+
+        // can't error after this so, this is safe
+        output = (stbi_uc *)stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1);
+        if (!output) {
+            stbi__cleanup_jpeg(z);
+            return stbi__errpuc("outofmem", "Out of memory");
+        }
+
+        // now go ahead and resample
+        for (j = 0; j < z->s->img_y; ++j) {
+            stbi_uc * out = output + n * z->s->img_x * j;
+            for (k = 0; k < decode_n; ++k) {
+                stbi__resample * r = &res_comp[k];
+                int y_bot = r->ystep >= (r->vs >> 1);
+                coutput[k] = r->resample(z->img_comp[k].linebuf, y_bot ? r->line1 : r->line0, y_bot ? r->line0 : r->line1,
+                                         r->w_lores, r->hs);
+                if (++r->ystep >= r->vs) {
+                    r->ystep = 0;
+                    r->line0 = r->line1;
+                    if (++r->ypos < z->img_comp[k].y)
+                        r->line1 += z->img_comp[k].w2;
+                }
+            }
+            if (n >= 3) {
+                stbi_uc * y = coutput[0];
+                if (z->s->img_n == 3) {
+                    if (is_rgb) {
+                        for (i = 0; i < z->s->img_x; ++i) {
+                            out[0] = y[i];
+                            out[1] = coutput[1][i];
+                            out[2] = coutput[2][i];
+                            out[3] = 255;
+                            out += n;
+                        }
+                    } else {
+                        z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+                    }
+                } else if (z->s->img_n == 4) {
+                    if (z->app14_color_transform == 0) { // CMYK
+                        for (i = 0; i < z->s->img_x; ++i) {
+                            stbi_uc m = coutput[3][i];
+                            out[0] = stbi__blinn_8x8(coutput[0][i], m);
+                            out[1] = stbi__blinn_8x8(coutput[1][i], m);
+                            out[2] = stbi__blinn_8x8(coutput[2][i], m);
+                            out[3] = 255;
+                            out += n;
+                        }
+                    } else if (z->app14_color_transform == 2) { // YCCK
+                        z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+                        for (i = 0; i < z->s->img_x; ++i) {
+                            stbi_uc m = coutput[3][i];
+                            out[0] = stbi__blinn_8x8(255 - out[0], m);
+                            out[1] = stbi__blinn_8x8(255 - out[1], m);
+                            out[2] = stbi__blinn_8x8(255 - out[2], m);
+                            out += n;
+                        }
+                    } else { // YCbCr + alpha?  Ignore the fourth channel for now
+                        z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+                    }
+                } else
+                    for (i = 0; i < z->s->img_x; ++i) {
+                        out[0] = out[1] = out[2] = y[i];
+                        out[3] = 255; // not used if n==3
+                        out += n;
+                    }
+            } else {
+                if (is_rgb) {
+                    if (n == 1)
+                        for (i = 0; i < z->s->img_x; ++i)
+                            *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+                    else {
+                        for (i = 0; i < z->s->img_x; ++i, out += 2) {
+                            out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+                            out[1] = 255;
+                        }
+                    }
+                } else if (z->s->img_n == 4 && z->app14_color_transform == 0) {
+                    for (i = 0; i < z->s->img_x; ++i) {
+                        stbi_uc m = coutput[3][i];
+                        stbi_uc r = stbi__blinn_8x8(coutput[0][i], m);
+                        stbi_uc g = stbi__blinn_8x8(coutput[1][i], m);
+                        stbi_uc b = stbi__blinn_8x8(coutput[2][i], m);
+                        out[0] = stbi__compute_y(r, g, b);
+                        out[1] = 255;
+                        out += n;
+                    }
+                } else if (z->s->img_n == 4 && z->app14_color_transform == 2) {
+                    for (i = 0; i < z->s->img_x; ++i) {
+                        out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]);
+                        out[1] = 255;
+                        out += n;
+                    }
+                } else {
+                    stbi_uc * y = coutput[0];
+                    if (n == 1)
+                        for (i = 0; i < z->s->img_x; ++i)
+                            out[i] = y[i];
+                    else
+                        for (i = 0; i < z->s->img_x; ++i) {
+                            *out++ = y[i];
+                            *out++ = 255;
+                        }
+                }
+            }
+        }
+        stbi__cleanup_jpeg(z);
+        *out_x = z->s->img_x;
+        *out_y = z->s->img_y;
+        if (comp)
+            *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output
+        return output;
+    }
+}
+
+static void * stbi__jpeg_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) {
+    unsigned char * result;
+    stbi__jpeg * j = (stbi__jpeg *)stbi__malloc(sizeof(stbi__jpeg));
+    if (!j)
+        return stbi__errpuc("outofmem", "Out of memory");
+    memset(j, 0, sizeof(stbi__jpeg));
+    STBI_NOTUSED(ri);
+    j->s = s;
+    stbi__setup_jpeg(j);
+    result = load_jpeg_image(j, x, y, comp, req_comp);
+    STBI_FREE(j);
+    return result;
+}
+
+static int stbi__jpeg_test(stbi__context * s) {
+    int r;
+    stbi__jpeg * j = (stbi__jpeg *)stbi__malloc(sizeof(stbi__jpeg));
+    if (!j)
+        return stbi__err("outofmem", "Out of memory");
+    memset(j, 0, sizeof(stbi__jpeg));
+    j->s = s;
+    stbi__setup_jpeg(j);
+    r = stbi__decode_jpeg_header(j, STBI__SCAN_type);
+    stbi__rewind(s);
+    STBI_FREE(j);
+    return r;
+}
+
+static int stbi__jpeg_info_raw(stbi__jpeg * j, int * x, int * y, int * comp) {
+    if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) {
+        stbi__rewind(j->s);
+        return 0;
+    }
+    if (x)
+        *x = j->s->img_x;
+    if (y)
+        *y = j->s->img_y;
+    if (comp)
+        *comp = j->s->img_n >= 3 ? 3 : 1;
+    return 1;
+}
+
+static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp) {
+    int result;
+    stbi__jpeg * j = (stbi__jpeg *)(stbi__malloc(sizeof(stbi__jpeg)));
+    if (!j)
+        return stbi__err("outofmem", "Out of memory");
+    memset(j, 0, sizeof(stbi__jpeg));
+    j->s = s;
+    result = stbi__jpeg_info_raw(j, x, y, comp);
+    STBI_FREE(j);
+    return result;
+}
+#endif
+
+// public domain zlib decode    v0.2  Sean Barrett 2006-11-18
+//    simple implementation
+//      - all input must be provided in an upfront buffer
+//      - all output is written to a single output buffer (can malloc/realloc)
+//    performance
+//      - fast huffman
+
+#ifndef STBI_NO_ZLIB
+
+// fast-way is faster to check than jpeg huffman, but slow way is slower
+#define STBI__ZFAST_BITS 9 // accelerate all cases in default tables
+#define STBI__ZFAST_MASK ((1 << STBI__ZFAST_BITS) - 1)
+#define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet
+
+// zlib-style huffman encoding
+// (jpegs packs from left, zlib from right, so can't share code)
+typedef struct {
+    stbi__uint16 fast[1 << STBI__ZFAST_BITS];
+    stbi__uint16 firstcode[16];
+    int maxcode[17];
+    stbi__uint16 firstsymbol[16];
+    stbi_uc size[STBI__ZNSYMS];
+    stbi__uint16 value[STBI__ZNSYMS];
+} stbi__zhuffman;
+
+stbi_inline static int stbi__bitreverse16(int n) {
+    n = ((n & 0xAAAA) >> 1) | ((n & 0x5555) << 1);
+    n = ((n & 0xCCCC) >> 2) | ((n & 0x3333) << 2);
+    n = ((n & 0xF0F0) >> 4) | ((n & 0x0F0F) << 4);
+    n = ((n & 0xFF00) >> 8) | ((n & 0x00FF) << 8);
+    return n;
+}
+
+stbi_inline static int stbi__bit_reverse(int v, int bits) {
+    STBI_ASSERT(bits <= 16);
+    // to bit reverse n bits, reverse 16 and shift
+    // e.g. 11 bits, bit reverse and shift away 5
+    return stbi__bitreverse16(v) >> (16 - bits);
+}
+
+static int stbi__zbuild_huffman(stbi__zhuffman * z, const stbi_uc * sizelist, int num) {
+    int i, k = 0;
+    int code, next_code[16], sizes[17];
+
+    // DEFLATE spec for generating codes
+    memset(sizes, 0, sizeof(sizes));
+    memset(z->fast, 0, sizeof(z->fast));
+    for (i = 0; i < num; ++i)
+        ++sizes[sizelist[i]];
+    sizes[0] = 0;
+    for (i = 1; i < 16; ++i)
+        if (sizes[i] > (1 << i))
+            return stbi__err("bad sizes", "Corrupt PNG");
+    code = 0;
+    for (i = 1; i < 16; ++i) {
+        next_code[i] = code;
+        z->firstcode[i] = (stbi__uint16)code;
+        z->firstsymbol[i] = (stbi__uint16)k;
+        code = (code + sizes[i]);
+        if (sizes[i])
+            if (code - 1 >= (1 << i))
+                return stbi__err("bad codelengths", "Corrupt PNG");
+        z->maxcode[i] = code << (16 - i); // preshift for inner loop
+        code <<= 1;
+        k += sizes[i];
+    }
+    z->maxcode[16] = 0x10000; // sentinel
+    for (i = 0; i < num; ++i) {
+        int s = sizelist[i];
+        if (s) {
+            int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s];
+            stbi__uint16 fastv = (stbi__uint16)((s << 9) | i);
+            z->size[c] = (stbi_uc)s;
+            z->value[c] = (stbi__uint16)i;
+            if (s <= STBI__ZFAST_BITS) {
+                int j = stbi__bit_reverse(next_code[s], s);
+                while (j < (1 << STBI__ZFAST_BITS)) {
+                    z->fast[j] = fastv;
+                    j += (1 << s);
+                }
+            }
+            ++next_code[s];
+        }
+    }
+    return 1;
+}
+
+// zlib-from-memory implementation for PNG reading
+//    because PNG allows splitting the zlib stream arbitrarily,
+//    and it's annoying structurally to have PNG call ZLIB call PNG,
+//    we require PNG read all the IDATs and combine them into a single
+//    memory buffer
+
+typedef struct {
+    stbi_uc *zbuffer, *zbuffer_end;
+    int num_bits;
+    stbi__uint32 code_buffer;
+
+    char * zout;
+    char * zout_start;
+    char * zout_end;
+    int z_expandable;
+
+    stbi__zhuffman z_length, z_distance;
+} stbi__zbuf;
+
+stbi_inline static int stbi__zeof(stbi__zbuf * z) { return (z->zbuffer >= z->zbuffer_end); }
+
+stbi_inline static stbi_uc stbi__zget8(stbi__zbuf * z) { return stbi__zeof(z) ? 0 : *z->zbuffer++; }
+
+static void stbi__fill_bits(stbi__zbuf * z) {
+    do {
+        if (z->code_buffer >= (1U << z->num_bits)) {
+            z->zbuffer = z->zbuffer_end; /* treat this as EOF so we fail. */
+            return;
+        }
+        z->code_buffer |= (unsigned int)stbi__zget8(z) << z->num_bits;
+        z->num_bits += 8;
+    } while (z->num_bits <= 24);
+}
+
+stbi_inline static unsigned int stbi__zreceive(stbi__zbuf * z, int n) {
+    unsigned int k;
+    if (z->num_bits < n)
+        stbi__fill_bits(z);
+    k = z->code_buffer & ((1 << n) - 1);
+    z->code_buffer >>= n;
+    z->num_bits -= n;
+    return k;
+}
+
+static int stbi__zhuffman_decode_slowpath(stbi__zbuf * a, stbi__zhuffman * z) {
+    int b, s, k;
+    // not resolved by fast table, so compute it the slow way
+    // use jpeg approach, which requires MSbits at top
+    k = stbi__bit_reverse(a->code_buffer, 16);
+    for (s = STBI__ZFAST_BITS + 1;; ++s)
+        if (k < z->maxcode[s])
+            break;
+    if (s >= 16)
+        return -1; // invalid code!
+    // code size is s, so:
+    b = (k >> (16 - s)) - z->firstcode[s] + z->firstsymbol[s];
+    if (b >= STBI__ZNSYMS)
+        return -1; // some data was corrupt somewhere!
+    if (z->size[b] != s)
+        return -1; // was originally an assert, but report failure instead.
+    a->code_buffer >>= s;
+    a->num_bits -= s;
+    return z->value[b];
+}
+
+stbi_inline static int stbi__zhuffman_decode(stbi__zbuf * a, stbi__zhuffman * z) {
+    int b, s;
+    if (a->num_bits < 16) {
+        if (stbi__zeof(a)) {
+            return -1; /* report error for unexpected end of data. */
+        }
+        stbi__fill_bits(a);
+    }
+    b = z->fast[a->code_buffer & STBI__ZFAST_MASK];
+    if (b) {
+        s = b >> 9;
+        a->code_buffer >>= s;
+        a->num_bits -= s;
+        return b & 511;
+    }
+    return stbi__zhuffman_decode_slowpath(a, z);
+}
+
+static int stbi__zexpand(stbi__zbuf * z, char * zout, int n) // need to make room for n bytes
+{
+    char * q;
+    unsigned int cur, limit, old_limit;
+    z->zout = zout;
+    if (!z->z_expandable)
+        return stbi__err("output buffer limit", "Corrupt PNG");
+    cur = (unsigned int)(z->zout - z->zout_start);
+    limit = old_limit = (unsigned)(z->zout_end - z->zout_start);
+    if (UINT_MAX - cur < (unsigned)n)
+        return stbi__err("outofmem", "Out of memory");
+    while (cur + n > limit) {
+        if (limit > UINT_MAX / 2)
+            return stbi__err("outofmem", "Out of memory");
+        limit *= 2;
+    }
+    q = (char *)STBI_REALLOC_SIZED(z->zout_start, old_limit, limit);
+    STBI_NOTUSED(old_limit);
+    if (q == NULL)
+        return stbi__err("outofmem", "Out of memory");
+    z->zout_start = q;
+    z->zout = q + cur;
+    z->zout_end = q + limit;
+    return 1;
+}
+
+static const int stbi__zlength_base[31] = {3,  4,  5,  6,  7,  8,  9,  10,  11,  13,  15,  17,  19,  23, 27, 31,
+                                           35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258, 0,  0};
+
+static const int stbi__zlength_extra[31] = {0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2,
+                                            3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0, 0, 0};
+
+static const int stbi__zdist_base[32] = {1,    2,    3,    4,    5,    7,     9,     13,    17,  25,   33,
+                                         49,   65,   97,   129,  193,  257,   385,   513,   769, 1025, 1537,
+                                         2049, 3073, 4097, 6145, 8193, 12289, 16385, 24577, 0,   0};
+
+static const int stbi__zdist_extra[32] = {0, 0, 0, 0, 1, 1, 2, 2,  3,  3,  4,  4,  5,  5,  6,
+                                          6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13};
+
+static int stbi__parse_huffman_block(stbi__zbuf * a) {
+    char * zout = a->zout;
+    for (;;) {
+        int z = stbi__zhuffman_decode(a, &a->z_length);
+        if (z < 256) {
+            if (z < 0)
+                return stbi__err("bad huffman code", "Corrupt PNG"); // error in huffman codes
+            if (zout >= a->zout_end) {
+                if (!stbi__zexpand(a, zout, 1))
+                    return 0;
+                zout = a->zout;
+            }
+            *zout++ = (char)z;
+        } else {
+            stbi_uc * p;
+            int len, dist;
+            if (z == 256) {
+                a->zout = zout;
+                return 1;
+            }
+            if (z >= 286)
+                return stbi__err("bad huffman code",
+                                 "Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data
+            z -= 257;
+            len = stbi__zlength_base[z];
+            if (stbi__zlength_extra[z])
+                len += stbi__zreceive(a, stbi__zlength_extra[z]);
+            z = stbi__zhuffman_decode(a, &a->z_distance);
+            if (z < 0 || z >= 30)
+                return stbi__err("bad huffman code",
+                                 "Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data
+            dist = stbi__zdist_base[z];
+            if (stbi__zdist_extra[z])
+                dist += stbi__zreceive(a, stbi__zdist_extra[z]);
+            if (zout - a->zout_start < dist)
+                return stbi__err("bad dist", "Corrupt PNG");
+            if (zout + len > a->zout_end) {
+                if (!stbi__zexpand(a, zout, len))
+                    return 0;
+                zout = a->zout;
+            }
+            p = (stbi_uc *)(zout - dist);
+            if (dist == 1) { // run of one byte; common in images.
+                stbi_uc v = *p;
+                if (len) {
+                    do
+                        *zout++ = v;
+                    while (--len);
+                }
+            } else {
+                if (len) {
+                    do
+                        *zout++ = *p++;
+                    while (--len);
+                }
+            }
+        }
+    }
+}
+
+static int stbi__compute_huffman_codes(stbi__zbuf * a) {
+    static const stbi_uc length_dezigzag[19] = {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15};
+    stbi__zhuffman z_codelength;
+    stbi_uc lencodes[286 + 32 + 137]; // padding for maximum single op
+    stbi_uc codelength_sizes[19];
+    int i, n;
+
+    int hlit = stbi__zreceive(a, 5) + 257;
+    int hdist = stbi__zreceive(a, 5) + 1;
+    int hclen = stbi__zreceive(a, 4) + 4;
+    int ntot = hlit + hdist;
+
+    memset(codelength_sizes, 0, sizeof(codelength_sizes));
+    for (i = 0; i < hclen; ++i) {
+        int s = stbi__zreceive(a, 3);
+        codelength_sizes[length_dezigzag[i]] = (stbi_uc)s;
+    }
+    if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19))
+        return 0;
+
+    n = 0;
+    while (n < ntot) {
+        int c = stbi__zhuffman_decode(a, &z_codelength);
+        if (c < 0 || c >= 19)
+            return stbi__err("bad codelengths", "Corrupt PNG");
+        if (c < 16)
+            lencodes[n++] = (stbi_uc)c;
+        else {
+            stbi_uc fill = 0;
+            if (c == 16) {
+                c = stbi__zreceive(a, 2) + 3;
+                if (n == 0)
+                    return stbi__err("bad codelengths", "Corrupt PNG");
+                fill = lencodes[n - 1];
+            } else if (c == 17) {
+                c = stbi__zreceive(a, 3) + 3;
+            } else if (c == 18) {
+                c = stbi__zreceive(a, 7) + 11;
+            } else {
+                return stbi__err("bad codelengths", "Corrupt PNG");
+            }
+            if (ntot - n < c)
+                return stbi__err("bad codelengths", "Corrupt PNG");
+            memset(lencodes + n, fill, c);
+            n += c;
+        }
+    }
+    if (n != ntot)
+        return stbi__err("bad codelengths", "Corrupt PNG");
+    if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit))
+        return 0;
+    if (!stbi__zbuild_huffman(&a->z_distance, lencodes + hlit, hdist))
+        return 0;
+    return 1;
+}
+
+static int stbi__parse_uncompressed_block(stbi__zbuf * a) {
+    stbi_uc header[4];
+    int len, nlen, k;
+    if (a->num_bits & 7)
+        stbi__zreceive(a, a->num_bits & 7); // discard
+    // drain the bit-packed data into header
+    k = 0;
+    while (a->num_bits > 0) {
+        header[k++] = (stbi_uc)(a->code_buffer & 255); // suppress MSVC run-time check
+        a->code_buffer >>= 8;
+        a->num_bits -= 8;
+    }
+    if (a->num_bits < 0)
+        return stbi__err("zlib corrupt", "Corrupt PNG");
+    // now fill header the normal way
+    while (k < 4)
+        header[k++] = stbi__zget8(a);
+    len = header[1] * 256 + header[0];
+    nlen = header[3] * 256 + header[2];
+    if (nlen != (len ^ 0xffff))
+        return stbi__err("zlib corrupt", "Corrupt PNG");
+    if (a->zbuffer + len > a->zbuffer_end)
+        return stbi__err("read past buffer", "Corrupt PNG");
+    if (a->zout + len > a->zout_end)
+        if (!stbi__zexpand(a, a->zout, len))
+            return 0;
+    memcpy(a->zout, a->zbuffer, len);
+    a->zbuffer += len;
+    a->zout += len;
+    return 1;
+}
+
+static int stbi__parse_zlib_header(stbi__zbuf * a) {
+    int cmf = stbi__zget8(a);
+    int cm = cmf & 15;
+    /* int cinfo = cmf >> 4; */
+    int flg = stbi__zget8(a);
+    if (stbi__zeof(a))
+        return stbi__err("bad zlib header", "Corrupt PNG"); // zlib spec
+    if ((cmf * 256 + flg) % 31 != 0)
+        return stbi__err("bad zlib header", "Corrupt PNG"); // zlib spec
+    if (flg & 32)
+        return stbi__err("no preset dict", "Corrupt PNG"); // preset dictionary not allowed in png
+    if (cm != 8)
+        return stbi__err("bad compression", "Corrupt PNG"); // DEFLATE required for png
+    // window = 1 << (8 + cinfo)... but who cares, we fully buffer output
+    return 1;
+}
+
+static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] = {
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+    9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    9, 9, 9, 9, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8};
+static const stbi_uc stbi__zdefault_distance[32] = {5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+                                                    5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5};
+/*
+Init algorithm:
+{
+   int i;   // use <= to match clearly with spec
+   for (i=0; i <= 143; ++i)     stbi__zdefault_length[i]   = 8;
+   for (   ; i <= 255; ++i)     stbi__zdefault_length[i]   = 9;
+   for (   ; i <= 279; ++i)     stbi__zdefault_length[i]   = 7;
+   for (   ; i <= 287; ++i)     stbi__zdefault_length[i]   = 8;
+
+   for (i=0; i <=  31; ++i)     stbi__zdefault_distance[i] = 5;
+}
+*/
+
+static int stbi__parse_zlib(stbi__zbuf * a, int parse_header) {
+    int final, type;
+    if (parse_header)
+        if (!stbi__parse_zlib_header(a))
+            return 0;
+    a->num_bits = 0;
+    a->code_buffer = 0;
+    do {
+        final = stbi__zreceive(a, 1);
+        type = stbi__zreceive(a, 2);
+        if (type == 0) {
+            if (!stbi__parse_uncompressed_block(a))
+                return 0;
+        } else if (type == 3) {
+            return 0;
+        } else {
+            if (type == 1) {
+                // use fixed code lengths
+                if (!stbi__zbuild_huffman(&a->z_length, stbi__zdefault_length, STBI__ZNSYMS))
+                    return 0;
+                if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance, 32))
+                    return 0;
+            } else {
+                if (!stbi__compute_huffman_codes(a))
+                    return 0;
+            }
+            if (!stbi__parse_huffman_block(a))
+                return 0;
+        }
+    } while (!final);
+    return 1;
+}
+
+static int stbi__do_zlib(stbi__zbuf * a, char * obuf, int olen, int exp, int parse_header) {
+    a->zout_start = obuf;
+    a->zout = obuf;
+    a->zout_end = obuf + olen;
+    a->z_expandable = exp;
+
+    return stbi__parse_zlib(a, parse_header);
+}
+
+STBIDEF char * stbi_zlib_decode_malloc_guesssize(const char * buffer, int len, int initial_size, int * outlen) {
+    stbi__zbuf a;
+    char * p = (char *)stbi__malloc(initial_size);
+    if (p == NULL)
+        return NULL;
+    a.zbuffer = (stbi_uc *)buffer;
+    a.zbuffer_end = (stbi_uc *)buffer + len;
+    if (stbi__do_zlib(&a, p, initial_size, 1, 1)) {
+        if (outlen)
+            *outlen = (int)(a.zout - a.zout_start);
+        return a.zout_start;
+    } else {
+        STBI_FREE(a.zout_start);
+        return NULL;
+    }
+}
+
+STBIDEF char * stbi_zlib_decode_malloc(char const * buffer, int len, int * outlen) {
+    return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen);
+}
+
+STBIDEF char * stbi_zlib_decode_malloc_guesssize_headerflag(const char * buffer, int len, int initial_size, int * outlen,
+                                                            int parse_header) {
+    stbi__zbuf a;
+    char * p = (char *)stbi__malloc(initial_size);
+    if (p == NULL)
+        return NULL;
+    a.zbuffer = (stbi_uc *)buffer;
+    a.zbuffer_end = (stbi_uc *)buffer + len;
+    if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) {
+        if (outlen)
+            *outlen = (int)(a.zout - a.zout_start);
+        return a.zout_start;
+    } else {
+        STBI_FREE(a.zout_start);
+        return NULL;
+    }
+}
+
+STBIDEF int stbi_zlib_decode_buffer(char * obuffer, int olen, char const * ibuffer, int ilen) {
+    stbi__zbuf a;
+    a.zbuffer = (stbi_uc *)ibuffer;
+    a.zbuffer_end = (stbi_uc *)ibuffer + ilen;
+    if (stbi__do_zlib(&a, obuffer, olen, 0, 1))
+        return (int)(a.zout - a.zout_start);
+    else
+        return -1;
+}
+
+STBIDEF char * stbi_zlib_decode_noheader_malloc(char const * buffer, int len, int * outlen) {
+    stbi__zbuf a;
+    char * p = (char *)stbi__malloc(16384);
+    if (p == NULL)
+        return NULL;
+    a.zbuffer = (stbi_uc *)buffer;
+    a.zbuffer_end = (stbi_uc *)buffer + len;
+    if (stbi__do_zlib(&a, p, 16384, 1, 0)) {
+        if (outlen)
+            *outlen = (int)(a.zout - a.zout_start);
+        return a.zout_start;
+    } else {
+        STBI_FREE(a.zout_start);
+        return NULL;
+    }
+}
+
+STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const char * ibuffer, int ilen) {
+    stbi__zbuf a;
+    a.zbuffer = (stbi_uc *)ibuffer;
+    a.zbuffer_end = (stbi_uc *)ibuffer + ilen;
+    if (stbi__do_zlib(&a, obuffer, olen, 0, 0))
+        return (int)(a.zout - a.zout_start);
+    else
+        return -1;
+}
+#endif
+
+// public domain "baseline" PNG decoder   v0.10  Sean Barrett 2006-11-18
+//    simple implementation
+//      - only 8-bit samples
+//      - no CRC checking
+//      - allocates lots of intermediate memory
+//        - avoids problem of streaming data between subsystems
+//        - avoids explicit window management
+//    performance
+//      - uses stb_zlib, a PD zlib implementation with fast huffman decoding
+
+#ifndef STBI_NO_PNG
+typedef struct {
+    stbi__uint32 length;
+    stbi__uint32 type;
+} stbi__pngchunk;
+
+static stbi__pngchunk stbi__get_chunk_header(stbi__context * s) {
+    stbi__pngchunk c;
+    c.length = stbi__get32be(s);
+    c.type = stbi__get32be(s);
+    return c;
+}
+
+static int stbi__check_png_header(stbi__context * s) {
+    static const stbi_uc png_sig[8] = {137, 80, 78, 71, 13, 10, 26, 10};
+    int i;
+    for (i = 0; i < 8; ++i)
+        if (stbi__get8(s) != png_sig[i])
+            return stbi__err("bad png sig", "Not a PNG");
+    return 1;
+}
+
+typedef struct {
+    stbi__context * s;
+    stbi_uc *idata, *expanded, *out;
+    int depth;
+} stbi__png;
+
+enum {
+    STBI__F_none = 0,
+    STBI__F_sub = 1,
+    STBI__F_up = 2,
+    STBI__F_avg = 3,
+    STBI__F_paeth = 4,
+    // synthetic filters used for first scanline to avoid needing a dummy row of 0s
+    STBI__F_avg_first,
+    STBI__F_paeth_first
+};
+
+static stbi_uc first_row_filter[5] = {STBI__F_none, STBI__F_sub, STBI__F_none, STBI__F_avg_first, STBI__F_paeth_first};
+
+static int stbi__paeth(int a, int b, int c) {
+    int p = a + b - c;
+    int pa = abs(p - a);
+    int pb = abs(p - b);
+    int pc = abs(p - c);
+    if (pa <= pb && pa <= pc)
+        return a;
+    if (pb <= pc)
+        return b;
+    return c;
+}
+
+static const stbi_uc stbi__depth_scale_table[9] = {0, 0xff, 0x55, 0, 0x11, 0, 0, 0, 0x01};
+
+// create the png data from post-deflated data
+static int stbi__create_png_image_raw(stbi__png * a, stbi_uc * raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x,
+                                      stbi__uint32 y, int depth, int color) {
+    int bytes = (depth == 16 ? 2 : 1);
+    stbi__context * s = a->s;
+    stbi__uint32 i, j, stride = x * out_n * bytes;
+    stbi__uint32 img_len, img_width_bytes;
+    int k;
+    int img_n = s->img_n; // copy it into a local for later
+
+    int output_bytes = out_n * bytes;
+    int filter_bytes = img_n * bytes;
+    int width = x;
+
+    STBI_ASSERT(out_n == s->img_n || out_n == s->img_n + 1);
+    a->out = (stbi_uc *)stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into
+    if (!a->out)
+        return stbi__err("outofmem", "Out of memory");
+
+    if (!stbi__mad3sizes_valid(img_n, x, depth, 7))
+        return stbi__err("too large", "Corrupt PNG");
+    img_width_bytes = (((img_n * x * depth) + 7) >> 3);
+    img_len = (img_width_bytes + 1) * y;
+
+    // we used to check for exact match between raw_len and img_len on non-interlaced PNGs,
+    // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros),
+    // so just check for raw_len < img_len always.
+    if (raw_len < img_len)
+        return stbi__err("not enough pixels", "Corrupt PNG");
+
+    for (j = 0; j < y; ++j) {
+        stbi_uc * cur = a->out + stride * j;
+        stbi_uc * prior;
+        int filter = *raw++;
+
+        if (filter > 4)
+            return stbi__err("invalid filter", "Corrupt PNG");
+
+        if (depth < 8) {
+            if (img_width_bytes > x)
+                return stbi__err("invalid width", "Corrupt PNG");
+            cur += x * out_n - img_width_bytes; // store output to the rightmost img_len bytes, so we can decode in place
+            filter_bytes = 1;
+            width = img_width_bytes;
+        }
+        prior = cur - stride; // bugfix: need to compute this after 'cur +=' computation above
+
+        // if first row, use special filter that doesn't sample previous row
+        if (j == 0)
+            filter = first_row_filter[filter];
+
+        // handle first byte explicitly
+        for (k = 0; k < filter_bytes; ++k) {
+            switch (filter) {
+            case STBI__F_none:
+                cur[k] = raw[k];
+                break;
+            case STBI__F_sub:
+                cur[k] = raw[k];
+                break;
+            case STBI__F_up:
+                cur[k] = STBI__BYTECAST(raw[k] + prior[k]);
+                break;
+            case STBI__F_avg:
+                cur[k] = STBI__BYTECAST(raw[k] + (prior[k] >> 1));
+                break;
+            case STBI__F_paeth:
+                cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(0, prior[k], 0));
+                break;
+            case STBI__F_avg_first:
+                cur[k] = raw[k];
+                break;
+            case STBI__F_paeth_first:
+                cur[k] = raw[k];
+                break;
+            }
+        }
+
+        if (depth == 8) {
+            if (img_n != out_n)
+                cur[img_n] = 255; // first pixel
+            raw += img_n;
+            cur += out_n;
+            prior += out_n;
+        } else if (depth == 16) {
+            if (img_n != out_n) {
+                cur[filter_bytes] = 255;     // first pixel top byte
+                cur[filter_bytes + 1] = 255; // first pixel bottom byte
+            }
+            raw += filter_bytes;
+            cur += output_bytes;
+            prior += output_bytes;
+        } else {
+            raw += 1;
+            cur += 1;
+            prior += 1;
+        }
+
+        // this is a little gross, so that we don't switch per-pixel or per-component
+        if (depth < 8 || img_n == out_n) {
+            int nk = (width - 1) * filter_bytes;
+#define STBI__CASE(f)                                                                                                          \
+    case f:                                                                                                                    \
+        for (k = 0; k < nk; ++k)
+            switch (filter) {
+            // "none" filter turns into a memcpy here; make that explicit.
+            case STBI__F_none:
+                memcpy(cur, raw, nk);
+                break;
+                STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k - filter_bytes]); }
+                break;
+                STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); }
+                break;
+                STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k - filter_bytes]) >> 1)); }
+                break;
+                STBI__CASE(STBI__F_paeth) {
+                    cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - filter_bytes], prior[k], prior[k - filter_bytes]));
+                }
+                break;
+                STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k - filter_bytes] >> 1)); }
+                break;
+                STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - filter_bytes], 0, 0)); }
+                break;
+            }
+#undef STBI__CASE
+            raw += nk;
+        } else {
+            STBI_ASSERT(img_n + 1 == out_n);
+#define STBI__CASE(f)                                                                                                          \
+    case f:                                                                                                                    \
+        for (i = x - 1; i >= 1; --i, cur[filter_bytes] = 255, raw += filter_bytes, cur += output_bytes, prior += output_bytes) \
+            for (k = 0; k < filter_bytes; ++k)
+            switch (filter) {
+                STBI__CASE(STBI__F_none) { cur[k] = raw[k]; }
+                break;
+                STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k - output_bytes]); }
+                break;
+                STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); }
+                break;
+                STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k - output_bytes]) >> 1)); }
+                break;
+                STBI__CASE(STBI__F_paeth) {
+                    cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - output_bytes], prior[k], prior[k - output_bytes]));
+                }
+                break;
+                STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k - output_bytes] >> 1)); }
+                break;
+                STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - output_bytes], 0, 0)); }
+                break;
+            }
+#undef STBI__CASE
+
+            // the loop above sets the high byte of the pixels' alpha, but for
+            // 16 bit png files we also need the low byte set. we'll do that here.
+            if (depth == 16) {
+                cur = a->out + stride * j; // start at the beginning of the row again
+                for (i = 0; i < x; ++i, cur += output_bytes) {
+                    cur[filter_bytes + 1] = 255;
+                }
+            }
+        }
+    }
+
+    // we make a separate pass to expand bits to pixels; for performance,
+    // this could run two scanlines behind the above code, so it won't
+    // intefere with filtering but will still be in the cache.
+    if (depth < 8) {
+        for (j = 0; j < y; ++j) {
+            stbi_uc * cur = a->out + stride * j;
+            stbi_uc * in = a->out + stride * j + x * out_n - img_width_bytes;
+            // unpack 1/2/4-bit into a 8-bit buffer. allows us to keep the common 8-bit path optimal at minimal cost for
+            // 1/2/4-bit png guarante byte alignment, if width is not multiple of 8/4/2 we'll decode dummy trailing data that
+            // will be skipped in the later loop
+            stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range
+
+            // note that the final byte might overshoot and write more data than desired.
+            // we can allocate enough data that this never writes out of memory, but it
+            // could also overwrite the next scanline. can it overwrite non-empty data
+            // on the next scanline? yes, consider 1-pixel-wide scanlines with 1-bit-per-pixel.
+            // so we need to explicitly clamp the final ones
+
+            if (depth == 4) {
+                for (k = x * img_n; k >= 2; k -= 2, ++in) {
+                    *cur++ = scale * ((*in >> 4));
+                    *cur++ = scale * ((*in) & 0x0f);
+                }
+                if (k > 0)
+                    *cur++ = scale * ((*in >> 4));
+            } else if (depth == 2) {
+                for (k = x * img_n; k >= 4; k -= 4, ++in) {
+                    *cur++ = scale * ((*in >> 6));
+                    *cur++ = scale * ((*in >> 4) & 0x03);
+                    *cur++ = scale * ((*in >> 2) & 0x03);
+                    *cur++ = scale * ((*in) & 0x03);
+                }
+                if (k > 0)
+                    *cur++ = scale * ((*in >> 6));
+                if (k > 1)
+                    *cur++ = scale * ((*in >> 4) & 0x03);
+                if (k > 2)
+                    *cur++ = scale * ((*in >> 2) & 0x03);
+            } else if (depth == 1) {
+                for (k = x * img_n; k >= 8; k -= 8, ++in) {
+                    *cur++ = scale * ((*in >> 7));
+                    *cur++ = scale * ((*in >> 6) & 0x01);
+                    *cur++ = scale * ((*in >> 5) & 0x01);
+                    *cur++ = scale * ((*in >> 4) & 0x01);
+                    *cur++ = scale * ((*in >> 3) & 0x01);
+                    *cur++ = scale * ((*in >> 2) & 0x01);
+                    *cur++ = scale * ((*in >> 1) & 0x01);
+                    *cur++ = scale * ((*in) & 0x01);
+                }
+                if (k > 0)
+                    *cur++ = scale * ((*in >> 7));
+                if (k > 1)
+                    *cur++ = scale * ((*in >> 6) & 0x01);
+                if (k > 2)
+                    *cur++ = scale * ((*in >> 5) & 0x01);
+                if (k > 3)
+                    *cur++ = scale * ((*in >> 4) & 0x01);
+                if (k > 4)
+                    *cur++ = scale * ((*in >> 3) & 0x01);
+                if (k > 5)
+                    *cur++ = scale * ((*in >> 2) & 0x01);
+                if (k > 6)
+                    *cur++ = scale * ((*in >> 1) & 0x01);
+            }
+            if (img_n != out_n) {
+                int q;
+                // insert alpha = 255
+                cur = a->out + stride * j;
+                if (img_n == 1) {
+                    for (q = x - 1; q >= 0; --q) {
+                        cur[q * 2 + 1] = 255;
+                        cur[q * 2 + 0] = cur[q];
+                    }
+                } else {
+                    STBI_ASSERT(img_n == 3);
+                    for (q = x - 1; q >= 0; --q) {
+                        cur[q * 4 + 3] = 255;
+                        cur[q * 4 + 2] = cur[q * 3 + 2];
+                        cur[q * 4 + 1] = cur[q * 3 + 1];
+                        cur[q * 4 + 0] = cur[q * 3 + 0];
+                    }
+                }
+            }
+        }
+    } else if (depth == 16) {
+        // force the image data from big-endian to platform-native.
+        // this is done in a separate pass due to the decoding relying
+        // on the data being untouched, but could probably be done
+        // per-line during decode if care is taken.
+        stbi_uc * cur = a->out;
+        stbi__uint16 * cur16 = (stbi__uint16 *)cur;
+
+        for (i = 0; i < x * y * out_n; ++i, cur16++, cur += 2) {
+            *cur16 = (cur[0] << 8) | cur[1];
+        }
+    }
+
+    return 1;
+}
+
+static int stbi__create_png_image(stbi__png * a, stbi_uc * image_data, stbi__uint32 image_data_len, int out_n, int depth,
+                                  int color, int interlaced) {
+    int bytes = (depth == 16 ? 2 : 1);
+    int out_bytes = out_n * bytes;
+    stbi_uc * final;
+    int p;
+    if (!interlaced)
+        return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color);
+
+    // de-interlacing
+    final = (stbi_uc *)stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0);
+    if (!final)
+        return stbi__err("outofmem", "Out of memory");
+    for (p = 0; p < 7; ++p) {
+        int xorig[] = {0, 4, 0, 2, 0, 1, 0};
+        int yorig[] = {0, 0, 4, 0, 2, 0, 1};
+        int xspc[] = {8, 8, 4, 4, 2, 2, 1};
+        int yspc[] = {8, 8, 8, 4, 4, 2, 2};
+        int i, j, x, y;
+        // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1
+        x = (a->s->img_x - xorig[p] + xspc[p] - 1) / xspc[p];
+        y = (a->s->img_y - yorig[p] + yspc[p] - 1) / yspc[p];
+        if (x && y) {
+            stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y;
+            if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) {
+                STBI_FREE(final);
+                return 0;
+            }
+            for (j = 0; j < y; ++j) {
+                for (i = 0; i < x; ++i) {
+                    int out_y = j * yspc[p] + yorig[p];
+                    int out_x = i * xspc[p] + xorig[p];
+                    memcpy(final + out_y * a->s->img_x * out_bytes + out_x * out_bytes, a->out + (j * x + i) * out_bytes,
+                           out_bytes);
+                }
+            }
+            STBI_FREE(a->out);
+            image_data += img_len;
+            image_data_len -= img_len;
+        }
+    }
+    a->out = final;
+
+    return 1;
+}
+
+static int stbi__compute_transparency(stbi__png * z, stbi_uc tc[3], int out_n) {
+    stbi__context * s = z->s;
+    stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+    stbi_uc * p = z->out;
+
+    // compute color-based transparency, assuming we've
+    // already got 255 as the alpha value in the output
+    STBI_ASSERT(out_n == 2 || out_n == 4);
+
+    if (out_n == 2) {
+        for (i = 0; i < pixel_count; ++i) {
+            p[1] = (p[0] == tc[0] ? 0 : 255);
+            p += 2;
+        }
+    } else {
+        for (i = 0; i < pixel_count; ++i) {
+            if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+                p[3] = 0;
+            p += 4;
+        }
+    }
+    return 1;
+}
+
+static int stbi__compute_transparency16(stbi__png * z, stbi__uint16 tc[3], int out_n) {
+    stbi__context * s = z->s;
+    stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+    stbi__uint16 * p = (stbi__uint16 *)z->out;
+
+    // compute color-based transparency, assuming we've
+    // already got 65535 as the alpha value in the output
+    STBI_ASSERT(out_n == 2 || out_n == 4);
+
+    if (out_n == 2) {
+        for (i = 0; i < pixel_count; ++i) {
+            p[1] = (p[0] == tc[0] ? 0 : 65535);
+            p += 2;
+        }
+    } else {
+        for (i = 0; i < pixel_count; ++i) {
+            if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+                p[3] = 0;
+            p += 4;
+        }
+    }
+    return 1;
+}
+
+static int stbi__expand_png_palette(stbi__png * a, stbi_uc * palette, int len, int pal_img_n) {
+    stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y;
+    stbi_uc *p, *temp_out, *orig = a->out;
+
+    p = (stbi_uc *)stbi__malloc_mad2(pixel_count, pal_img_n, 0);
+    if (p == NULL)
+        return stbi__err("outofmem", "Out of memory");
+
+    // between here and free(out) below, exitting would leak
+    temp_out = p;
+
+    if (pal_img_n == 3) {
+        for (i = 0; i < pixel_count; ++i) {
+            int n = orig[i] * 4;
+            p[0] = palette[n];
+            p[1] = palette[n + 1];
+            p[2] = palette[n + 2];
+            p += 3;
+        }
+    } else {
+        for (i = 0; i < pixel_count; ++i) {
+            int n = orig[i] * 4;
+            p[0] = palette[n];
+            p[1] = palette[n + 1];
+            p[2] = palette[n + 2];
+            p[3] = palette[n + 3];
+            p += 4;
+        }
+    }
+    STBI_FREE(a->out);
+    a->out = temp_out;
+
+    STBI_NOTUSED(len);
+
+    return 1;
+}
+
+static int stbi__unpremultiply_on_load_global = 0;
+static int stbi__de_iphone_flag_global = 0;
+
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply) {
+    stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert) {
+    stbi__de_iphone_flag_global = flag_true_if_should_convert;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__unpremultiply_on_load stbi__unpremultiply_on_load_global
+#define stbi__de_iphone_flag stbi__de_iphone_flag_global
+#else
+static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set;
+static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set;
+
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply) {
+    stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply;
+    stbi__unpremultiply_on_load_set = 1;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert) {
+    stbi__de_iphone_flag_local = flag_true_if_should_convert;
+    stbi__de_iphone_flag_set = 1;
+}
+
+#define stbi__unpremultiply_on_load                                                                                            \
+    (stbi__unpremultiply_on_load_set ? stbi__unpremultiply_on_load_local : stbi__unpremultiply_on_load_global)
+#define stbi__de_iphone_flag (stbi__de_iphone_flag_set ? stbi__de_iphone_flag_local : stbi__de_iphone_flag_global)
+#endif // STBI_THREAD_LOCAL
+
+static void stbi__de_iphone(stbi__png * z) {
+    stbi__context * s = z->s;
+    stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+    stbi_uc * p = z->out;
+
+    if (s->img_out_n == 3) { // convert bgr to rgb
+        for (i = 0; i < pixel_count; ++i) {
+            stbi_uc t = p[0];
+            p[0] = p[2];
+            p[2] = t;
+            p += 3;
+        }
+    } else {
+        STBI_ASSERT(s->img_out_n == 4);
+        if (stbi__unpremultiply_on_load) {
+            // convert bgr to rgb and unpremultiply
+            for (i = 0; i < pixel_count; ++i) {
+                stbi_uc a = p[3];
+                stbi_uc t = p[0];
+                if (a) {
+                    stbi_uc half = a / 2;
+                    p[0] = (p[2] * 255 + half) / a;
+                    p[1] = (p[1] * 255 + half) / a;
+                    p[2] = (t * 255 + half) / a;
+                } else {
+                    p[0] = p[2];
+                    p[2] = t;
+                }
+                p += 4;
+            }
+        } else {
+            // convert bgr to rgb
+            for (i = 0; i < pixel_count; ++i) {
+                stbi_uc t = p[0];
+                p[0] = p[2];
+                p[2] = t;
+                p += 4;
+            }
+        }
+    }
+}
+
+#define STBI__PNG_TYPE(a, b, c, d) (((unsigned)(a) << 24) + ((unsigned)(b) << 16) + ((unsigned)(c) << 8) + (unsigned)(d))
+
+static int stbi__parse_png_file(stbi__png * z, int scan, int req_comp) {
+    stbi_uc palette[1024], pal_img_n = 0;
+    stbi_uc has_trans = 0, tc[3] = {0};
+    stbi__uint16 tc16[3];
+    stbi__uint32 ioff = 0, idata_limit = 0, i, pal_len = 0;
+    int first = 1, k, interlace = 0, color = 0, is_iphone = 0;
+    stbi__context * s = z->s;
+
+    z->expanded = NULL;
+    z->idata = NULL;
+    z->out = NULL;
+
+    if (!stbi__check_png_header(s))
+        return 0;
+
+    if (scan == STBI__SCAN_type)
+        return 1;
+
+    for (;;) {
+        stbi__pngchunk c = stbi__get_chunk_header(s);
+        switch (c.type) {
+        case STBI__PNG_TYPE('C', 'g', 'B', 'I'):
+            is_iphone = 1;
+            stbi__skip(s, c.length);
+            break;
+        case STBI__PNG_TYPE('I', 'H', 'D', 'R'): {
+            int comp, filter;
+            if (!first)
+                return stbi__err("multiple IHDR", "Corrupt PNG");
+            first = 0;
+            if (c.length != 13)
+                return stbi__err("bad IHDR len", "Corrupt PNG");
+            s->img_x = stbi__get32be(s);
+            s->img_y = stbi__get32be(s);
+            if (s->img_y > STBI_MAX_DIMENSIONS)
+                return stbi__err("too large", "Very large image (corrupt?)");
+            if (s->img_x > STBI_MAX_DIMENSIONS)
+                return stbi__err("too large", "Very large image (corrupt?)");
+            z->depth = stbi__get8(s);
+            if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16)
+                return stbi__err("1/2/4/8/16-bit only", "PNG not supported: 1/2/4/8/16-bit only");
+            color = stbi__get8(s);
+            if (color > 6)
+                return stbi__err("bad ctype", "Corrupt PNG");
+            if (color == 3 && z->depth == 16)
+                return stbi__err("bad ctype", "Corrupt PNG");
+            if (color == 3)
+                pal_img_n = 3;
+            else if (color & 1)
+                return stbi__err("bad ctype", "Corrupt PNG");
+            comp = stbi__get8(s);
+            if (comp)
+                return stbi__err("bad comp method", "Corrupt PNG");
+            filter = stbi__get8(s);
+            if (filter)
+                return stbi__err("bad filter method", "Corrupt PNG");
+            interlace = stbi__get8(s);
+            if (interlace > 1)
+                return stbi__err("bad interlace method", "Corrupt PNG");
+            if (!s->img_x || !s->img_y)
+                return stbi__err("0-pixel image", "Corrupt PNG");
+            if (!pal_img_n) {
+                s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0);
+                if ((1 << 30) / s->img_x / s->img_n < s->img_y)
+                    return stbi__err("too large", "Image too large to decode");
+            } else {
+                // if paletted, then pal_n is our final components, and
+                // img_n is # components to decompress/filter.
+                s->img_n = 1;
+                if ((1 << 30) / s->img_x / 4 < s->img_y)
+                    return stbi__err("too large", "Corrupt PNG");
+            }
+            // even with SCAN_header, have to scan to see if we have a tRNS
+            break;
+        }
+
+        case STBI__PNG_TYPE('P', 'L', 'T', 'E'): {
+            if (first)
+                return stbi__err("first not IHDR", "Corrupt PNG");
+            if (c.length > 256 * 3)
+                return stbi__err("invalid PLTE", "Corrupt PNG");
+            pal_len = c.length / 3;
+            if (pal_len * 3 != c.length)
+                return stbi__err("invalid PLTE", "Corrupt PNG");
+            for (i = 0; i < pal_len; ++i) {
+                palette[i * 4 + 0] = stbi__get8(s);
+                palette[i * 4 + 1] = stbi__get8(s);
+                palette[i * 4 + 2] = stbi__get8(s);
+                palette[i * 4 + 3] = 255;
+            }
+            break;
+        }
+
+        case STBI__PNG_TYPE('t', 'R', 'N', 'S'): {
+            if (first)
+                return stbi__err("first not IHDR", "Corrupt PNG");
+            if (z->idata)
+                return stbi__err("tRNS after IDAT", "Corrupt PNG");
+            if (pal_img_n) {
+                if (scan == STBI__SCAN_header) {
+                    s->img_n = 4;
+                    return 1;
+                }
+                if (pal_len == 0)
+                    return stbi__err("tRNS before PLTE", "Corrupt PNG");
+                if (c.length > pal_len)
+                    return stbi__err("bad tRNS len", "Corrupt PNG");
+                pal_img_n = 4;
+                for (i = 0; i < c.length; ++i)
+                    palette[i * 4 + 3] = stbi__get8(s);
+            } else {
+                if (!(s->img_n & 1))
+                    return stbi__err("tRNS with alpha", "Corrupt PNG");
+                if (c.length != (stbi__uint32)s->img_n * 2)
+                    return stbi__err("bad tRNS len", "Corrupt PNG");
+                has_trans = 1;
+                // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now.
+                if (scan == STBI__SCAN_header) {
+                    ++s->img_n;
+                    return 1;
+                }
+                if (z->depth == 16) {
+                    for (k = 0; k < s->img_n; ++k)
+                        tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is
+                } else {
+                    for (k = 0; k < s->img_n; ++k)
+                        tc[k] = (stbi_uc)(stbi__get16be(s) & 255) *
+                                stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger
+                }
+            }
+            break;
+        }
+
+        case STBI__PNG_TYPE('I', 'D', 'A', 'T'): {
+            if (first)
+                return stbi__err("first not IHDR", "Corrupt PNG");
+            if (pal_img_n && !pal_len)
+                return stbi__err("no PLTE", "Corrupt PNG");
+            if (scan == STBI__SCAN_header) {
+                // header scan definitely stops at first IDAT
+                if (pal_img_n)
+                    s->img_n = pal_img_n;
+                return 1;
+            }
+            if (c.length > (1u << 30))
+                return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes");
+            if ((int)(ioff + c.length) < (int)ioff)
+                return 0;
+            if (ioff + c.length > idata_limit) {
+                stbi__uint32 idata_limit_old = idata_limit;
+                stbi_uc * p;
+                if (idata_limit == 0)
+                    idata_limit = c.length > 4096 ? c.length : 4096;
+                while (ioff + c.length > idata_limit)
+                    idata_limit *= 2;
+                STBI_NOTUSED(idata_limit_old);
+                p = (stbi_uc *)STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit);
+                if (p == NULL)
+                    return stbi__err("outofmem", "Out of memory");
+                z->idata = p;
+            }
+            if (!stbi__getn(s, z->idata + ioff, c.length))
+                return stbi__err("outofdata", "Corrupt PNG");
+            ioff += c.length;
+            break;
+        }
+
+        case STBI__PNG_TYPE('I', 'E', 'N', 'D'): {
+            stbi__uint32 raw_len, bpl;
+            if (first)
+                return stbi__err("first not IHDR", "Corrupt PNG");
+            if (scan != STBI__SCAN_load)
+                return 1;
+            if (z->idata == NULL)
+                return stbi__err("no IDAT", "Corrupt PNG");
+            // initial guess for decoded data size to avoid unnecessary reallocs
+            bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component
+            raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */;
+            z->expanded = (stbi_uc *)stbi_zlib_decode_malloc_guesssize_headerflag((char *)z->idata, ioff, raw_len,
+                                                                                  (int *)&raw_len, !is_iphone);
+            if (z->expanded == NULL)
+                return 0; // zlib should set error
+            STBI_FREE(z->idata);
+            z->idata = NULL;
+            if ((req_comp == s->img_n + 1 && req_comp != 3 && !pal_img_n) || has_trans)
+                s->img_out_n = s->img_n + 1;
+            else
+                s->img_out_n = s->img_n;
+            if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace))
+                return 0;
+            if (has_trans) {
+                if (z->depth == 16) {
+                    if (!stbi__compute_transparency16(z, tc16, s->img_out_n))
+                        return 0;
+                } else {
+                    if (!stbi__compute_transparency(z, tc, s->img_out_n))
+                        return 0;
+                }
+            }
+            if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2)
+                stbi__de_iphone(z);
+            if (pal_img_n) {
+                // pal_img_n == 3 or 4
+                s->img_n = pal_img_n; // record the actual colors we had
+                s->img_out_n = pal_img_n;
+                if (req_comp >= 3)
+                    s->img_out_n = req_comp;
+                if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n))
+                    return 0;
+            } else if (has_trans) {
+                // non-paletted image with tRNS -> source image has (constant) alpha
+                ++s->img_n;
+            }
+            STBI_FREE(z->expanded);
+            z->expanded = NULL;
+            // end of PNG chunk, read and skip CRC
+            stbi__get32be(s);
+            return 1;
+        }
+
+        default:
+            // if critical, fail
+            if (first)
+                return stbi__err("first not IHDR", "Corrupt PNG");
+            if ((c.type & (1 << 29)) == 0) {
+#ifndef STBI_NO_FAILURE_STRINGS
+                // not threadsafe
+                static char invalid_chunk[] = "XXXX PNG chunk not known";
+                invalid_chunk[0] = STBI__BYTECAST(c.type >> 24);
+                invalid_chunk[1] = STBI__BYTECAST(c.type >> 16);
+                invalid_chunk[2] = STBI__BYTECAST(c.type >> 8);
+                invalid_chunk[3] = STBI__BYTECAST(c.type >> 0);
+#endif
+                return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type");
+            }
+            stbi__skip(s, c.length);
+            break;
+        }
+        // end of PNG chunk, read and skip CRC
+        stbi__get32be(s);
+    }
+}
+
+static void * stbi__do_png(stbi__png * p, int * x, int * y, int * n, int req_comp, stbi__result_info * ri) {
+    void * result = NULL;
+    if (req_comp < 0 || req_comp > 4)
+        return stbi__errpuc("bad req_comp", "Internal error");
+    if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) {
+        if (p->depth <= 8)
+            ri->bits_per_channel = 8;
+        else if (p->depth == 16)
+            ri->bits_per_channel = 16;
+        else
+            return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth");
+        result = p->out;
+        p->out = NULL;
+        if (req_comp && req_comp != p->s->img_out_n) {
+            if (ri->bits_per_channel == 8)
+                result = stbi__convert_format((unsigned char *)result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+            else
+                result = stbi__convert_format16((stbi__uint16 *)result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+            p->s->img_out_n = req_comp;
+            if (result == NULL)
+                return result;
+        }
+        *x = p->s->img_x;
+        *y = p->s->img_y;
+        if (n)
+            *n = p->s->img_n;
+    }
+    STBI_FREE(p->out);
+    p->out = NULL;
+    STBI_FREE(p->expanded);
+    p->expanded = NULL;
+    STBI_FREE(p->idata);
+    p->idata = NULL;
+
+    return result;
+}
+
+static void * stbi__png_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) {
+    stbi__png p;
+    p.s = s;
+    return stbi__do_png(&p, x, y, comp, req_comp, ri);
+}
+
+static int stbi__png_test(stbi__context * s) {
+    int r;
+    r = stbi__check_png_header(s);
+    stbi__rewind(s);
+    return r;
+}
+
+static int stbi__png_info_raw(stbi__png * p, int * x, int * y, int * comp) {
+    if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) {
+        stbi__rewind(p->s);
+        return 0;
+    }
+    if (x)
+        *x = p->s->img_x;
+    if (y)
+        *y = p->s->img_y;
+    if (comp)
+        *comp = p->s->img_n;
+    return 1;
+}
+
+static int stbi__png_info(stbi__context * s, int * x, int * y, int * comp) {
+    stbi__png p;
+    p.s = s;
+    return stbi__png_info_raw(&p, x, y, comp);
+}
+
+static int stbi__png_is16(stbi__context * s) {
+    stbi__png p;
+    p.s = s;
+    if (!stbi__png_info_raw(&p, NULL, NULL, NULL))
+        return 0;
+    if (p.depth != 16) {
+        stbi__rewind(p.s);
+        return 0;
+    }
+    return 1;
+}
+#endif
+
+// Microsoft/Windows BMP image
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_test_raw(stbi__context * s) {
+    int r;
+    int sz;
+    if (stbi__get8(s) != 'B')
+        return 0;
+    if (stbi__get8(s) != 'M')
+        return 0;
+    stbi__get32le(s); // discard filesize
+    stbi__get16le(s); // discard reserved
+    stbi__get16le(s); // discard reserved
+    stbi__get32le(s); // discard data offset
+    sz = stbi__get32le(s);
+    r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124);
+    return r;
+}
+
+static int stbi__bmp_test(stbi__context * s) {
+    int r = stbi__bmp_test_raw(s);
+    stbi__rewind(s);
+    return r;
+}
+
+// returns 0..31 for the highest set bit
+static int stbi__high_bit(unsigned int z) {
+    int n = 0;
+    if (z == 0)
+        return -1;
+    if (z >= 0x10000) {
+        n += 16;
+        z >>= 16;
+    }
+    if (z >= 0x00100) {
+        n += 8;
+        z >>= 8;
+    }
+    if (z >= 0x00010) {
+        n += 4;
+        z >>= 4;
+    }
+    if (z >= 0x00004) {
+        n += 2;
+        z >>= 2;
+    }
+    if (z >= 0x00002) {
+        n += 1; /* >>=  1;*/
+    }
+    return n;
+}
+
+static int stbi__bitcount(unsigned int a) {
+    a = (a & 0x55555555) + ((a >> 1) & 0x55555555); // max 2
+    a = (a & 0x33333333) + ((a >> 2) & 0x33333333); // max 4
+    a = (a + (a >> 4)) & 0x0f0f0f0f;                // max 8 per 4, now 8 bits
+    a = (a + (a >> 8));                             // max 16 per 8 bits
+    a = (a + (a >> 16));                            // max 32 per 8 bits
+    return a & 0xff;
+}
+
+// extract an arbitrarily-aligned N-bit value (N=bits)
+// from v, and then make it 8-bits long and fractionally
+// extend it to full full range.
+static int stbi__shiftsigned(unsigned int v, int shift, int bits) {
+    static unsigned int mul_table[9] = {
+        0,
+        0xff /*0b11111111*/,
+        0x55 /*0b01010101*/,
+        0x49 /*0b01001001*/,
+        0x11 /*0b00010001*/,
+        0x21 /*0b00100001*/,
+        0x41 /*0b01000001*/,
+        0x81 /*0b10000001*/,
+        0x01 /*0b00000001*/,
+    };
+    static unsigned int shift_table[9] = {
+        0, 0, 0, 1, 0, 2, 4, 6, 0,
+    };
+    if (shift < 0)
+        v <<= -shift;
+    else
+        v >>= shift;
+    STBI_ASSERT(v < 256);
+    v >>= (8 - bits);
+    STBI_ASSERT(bits >= 0 && bits <= 8);
+    return (int)((unsigned)v * mul_table[bits]) >> shift_table[bits];
+}
+
+typedef struct {
+    int bpp, offset, hsz;
+    unsigned int mr, mg, mb, ma, all_a;
+    int extra_read;
+} stbi__bmp_data;
+
+static int stbi__bmp_set_mask_defaults(stbi__bmp_data * info, int compress) {
+    // BI_BITFIELDS specifies masks explicitly, don't override
+    if (compress == 3)
+        return 1;
+
+    if (compress == 0) {
+        if (info->bpp == 16) {
+            info->mr = 31u << 10;
+            info->mg = 31u << 5;
+            info->mb = 31u << 0;
+        } else if (info->bpp == 32) {
+            info->mr = 0xffu << 16;
+            info->mg = 0xffu << 8;
+            info->mb = 0xffu << 0;
+            info->ma = 0xffu << 24;
+            info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0
+        } else {
+            // otherwise, use defaults, which is all-0
+            info->mr = info->mg = info->mb = info->ma = 0;
+        }
+        return 1;
+    }
+    return 0; // error
+}
+
+static void * stbi__bmp_parse_header(stbi__context * s, stbi__bmp_data * info) {
+    int hsz;
+    if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M')
+        return stbi__errpuc("not BMP", "Corrupt BMP");
+    stbi__get32le(s); // discard filesize
+    stbi__get16le(s); // discard reserved
+    stbi__get16le(s); // discard reserved
+    info->offset = stbi__get32le(s);
+    info->hsz = hsz = stbi__get32le(s);
+    info->mr = info->mg = info->mb = info->ma = 0;
+    info->extra_read = 14;
+
+    if (info->offset < 0)
+        return stbi__errpuc("bad BMP", "bad BMP");
+
+    if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124)
+        return stbi__errpuc("unknown BMP", "BMP type not supported: unknown");
+    if (hsz == 12) {
+        s->img_x = stbi__get16le(s);
+        s->img_y = stbi__get16le(s);
+    } else {
+        s->img_x = stbi__get32le(s);
+        s->img_y = stbi__get32le(s);
+    }
+    if (stbi__get16le(s) != 1)
+        return stbi__errpuc("bad BMP", "bad BMP");
+    info->bpp = stbi__get16le(s);
+    if (hsz != 12) {
+        int compress = stbi__get32le(s);
+        if (compress == 1 || compress == 2)
+            return stbi__errpuc("BMP RLE", "BMP type not supported: RLE");
+        if (compress >= 4)
+            return stbi__errpuc("BMP JPEG/PNG",
+                                "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes
+        if (compress == 3 && info->bpp != 16 && info->bpp != 32)
+            return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel
+        stbi__get32le(s);                              // discard sizeof
+        stbi__get32le(s);                              // discard hres
+        stbi__get32le(s);                              // discard vres
+        stbi__get32le(s);                              // discard colorsused
+        stbi__get32le(s);                              // discard max important
+        if (hsz == 40 || hsz == 56) {
+            if (hsz == 56) {
+                stbi__get32le(s);
+                stbi__get32le(s);
+                stbi__get32le(s);
+                stbi__get32le(s);
+            }
+            if (info->bpp == 16 || info->bpp == 32) {
+                if (compress == 0) {
+                    stbi__bmp_set_mask_defaults(info, compress);
+                } else if (compress == 3) {
+                    info->mr = stbi__get32le(s);
+                    info->mg = stbi__get32le(s);
+                    info->mb = stbi__get32le(s);
+                    info->extra_read += 12;
+                    // not documented, but generated by photoshop and handled by mspaint
+                    if (info->mr == info->mg && info->mg == info->mb) {
+                        // ?!?!?
+                        return stbi__errpuc("bad BMP", "bad BMP");
+                    }
+                } else
+                    return stbi__errpuc("bad BMP", "bad BMP");
+            }
+        } else {
+            // V4/V5 header
+            int i;
+            if (hsz != 108 && hsz != 124)
+                return stbi__errpuc("bad BMP", "bad BMP");
+            info->mr = stbi__get32le(s);
+            info->mg = stbi__get32le(s);
+            info->mb = stbi__get32le(s);
+            info->ma = stbi__get32le(s);
+            if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs
+                stbi__bmp_set_mask_defaults(info, compress);
+            stbi__get32le(s); // discard color space
+            for (i = 0; i < 12; ++i)
+                stbi__get32le(s); // discard color space parameters
+            if (hsz == 124) {
+                stbi__get32le(s); // discard rendering intent
+                stbi__get32le(s); // discard offset of profile data
+                stbi__get32le(s); // discard size of profile data
+                stbi__get32le(s); // discard reserved
+            }
+        }
+    }
+    return (void *)1;
+}
+
+static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) {
+    stbi_uc * out;
+    unsigned int mr = 0, mg = 0, mb = 0, ma = 0, all_a;
+    stbi_uc pal[256][4];
+    int psize = 0, i, j, width;
+    int flip_vertically, pad, target;
+    stbi__bmp_data info;
+    STBI_NOTUSED(ri);
+
+    info.all_a = 255;
+    if (stbi__bmp_parse_header(s, &info) == NULL)
+        return NULL; // error code already set
+
+    flip_vertically = ((int)s->img_y) > 0;
+    s->img_y = abs((int)s->img_y);
+
+    if (s->img_y > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+    if (s->img_x > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+
+    mr = info.mr;
+    mg = info.mg;
+    mb = info.mb;
+    ma = info.ma;
+    all_a = info.all_a;
+
+    if (info.hsz == 12) {
+        if (info.bpp < 24)
+            psize = (info.offset - info.extra_read - 24) / 3;
+    } else {
+        if (info.bpp < 16)
+            psize = (info.offset - info.extra_read - info.hsz) >> 2;
+    }
+    if (psize == 0) {
+        // accept some number of extra bytes after the header, but if the offset points either to before
+        // the header ends or implies a large amount of extra data, reject the file as malformed
+        int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original);
+        int header_limit = 1024;        // max we actually read is below 256 bytes currently.
+        int extra_data_limit = 256 * 4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size.
+        if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) {
+            return stbi__errpuc("bad header", "Corrupt BMP");
+        }
+        // we established that bytes_read_so_far is positive and sensible.
+        // the first half of this test rejects offsets that are either too small positives, or
+        // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn
+        // ensures the number computed in the second half of the test can't overflow.
+        if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) {
+            return stbi__errpuc("bad offset", "Corrupt BMP");
+        } else {
+            stbi__skip(s, info.offset - bytes_read_so_far);
+        }
+    }
+
+    if (info.bpp == 24 && ma == 0xff000000)
+        s->img_n = 3;
+    else
+        s->img_n = ma ? 4 : 3;
+    if (req_comp && req_comp >= 3) // we can directly decode 3 or 4
+        target = req_comp;
+    else
+        target = s->img_n; // if they want monochrome, we'll post-convert
+
+    // sanity-check size
+    if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0))
+        return stbi__errpuc("too large", "Corrupt BMP");
+
+    out = (stbi_uc *)stbi__malloc_mad3(target, s->img_x, s->img_y, 0);
+    if (!out)
+        return stbi__errpuc("outofmem", "Out of memory");
+    if (info.bpp < 16) {
+        int z = 0;
+        if (psize == 0 || psize > 256) {
+            STBI_FREE(out);
+            return stbi__errpuc("invalid", "Corrupt BMP");
+        }
+        for (i = 0; i < psize; ++i) {
+            pal[i][2] = stbi__get8(s);
+            pal[i][1] = stbi__get8(s);
+            pal[i][0] = stbi__get8(s);
+            if (info.hsz != 12)
+                stbi__get8(s);
+            pal[i][3] = 255;
+        }
+        stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4));
+        if (info.bpp == 1)
+            width = (s->img_x + 7) >> 3;
+        else if (info.bpp == 4)
+            width = (s->img_x + 1) >> 1;
+        else if (info.bpp == 8)
+            width = s->img_x;
+        else {
+            STBI_FREE(out);
+            return stbi__errpuc("bad bpp", "Corrupt BMP");
+        }
+        pad = (-width) & 3;
+        if (info.bpp == 1) {
+            for (j = 0; j < (int)s->img_y; ++j) {
+                int bit_offset = 7, v = stbi__get8(s);
+                for (i = 0; i < (int)s->img_x; ++i) {
+                    int color = (v >> bit_offset) & 0x1;
+                    out[z++] = pal[color][0];
+                    out[z++] = pal[color][1];
+                    out[z++] = pal[color][2];
+                    if (target == 4)
+                        out[z++] = 255;
+                    if (i + 1 == (int)s->img_x)
+                        break;
+                    if ((--bit_offset) < 0) {
+                        bit_offset = 7;
+                        v = stbi__get8(s);
+                    }
+                }
+                stbi__skip(s, pad);
+            }
+        } else {
+            for (j = 0; j < (int)s->img_y; ++j) {
+                for (i = 0; i < (int)s->img_x; i += 2) {
+                    int v = stbi__get8(s), v2 = 0;
+                    if (info.bpp == 4) {
+                        v2 = v & 15;
+                        v >>= 4;
+                    }
+                    out[z++] = pal[v][0];
+                    out[z++] = pal[v][1];
+                    out[z++] = pal[v][2];
+                    if (target == 4)
+                        out[z++] = 255;
+                    if (i + 1 == (int)s->img_x)
+                        break;
+                    v = (info.bpp == 8) ? stbi__get8(s) : v2;
+                    out[z++] = pal[v][0];
+                    out[z++] = pal[v][1];
+                    out[z++] = pal[v][2];
+                    if (target == 4)
+                        out[z++] = 255;
+                }
+                stbi__skip(s, pad);
+            }
+        }
+    } else {
+        int rshift = 0, gshift = 0, bshift = 0, ashift = 0, rcount = 0, gcount = 0, bcount = 0, acount = 0;
+        int z = 0;
+        int easy = 0;
+        stbi__skip(s, info.offset - info.extra_read - info.hsz);
+        if (info.bpp == 24)
+            width = 3 * s->img_x;
+        else if (info.bpp == 16)
+            width = 2 * s->img_x;
+        else /* bpp = 32 and pad = 0 */
+            width = 0;
+        pad = (-width) & 3;
+        if (info.bpp == 24) {
+            easy = 1;
+        } else if (info.bpp == 32) {
+            if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000)
+                easy = 2;
+        }
+        if (!easy) {
+            if (!mr || !mg || !mb) {
+                STBI_FREE(out);
+                return stbi__errpuc("bad masks", "Corrupt BMP");
+            }
+            // right shift amt to put high bit in position #7
+            rshift = stbi__high_bit(mr) - 7;
+            rcount = stbi__bitcount(mr);
+            gshift = stbi__high_bit(mg) - 7;
+            gcount = stbi__bitcount(mg);
+            bshift = stbi__high_bit(mb) - 7;
+            bcount = stbi__bitcount(mb);
+            ashift = stbi__high_bit(ma) - 7;
+            acount = stbi__bitcount(ma);
+            if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) {
+                STBI_FREE(out);
+                return stbi__errpuc("bad masks", "Corrupt BMP");
+            }
+        }
+        for (j = 0; j < (int)s->img_y; ++j) {
+            if (easy) {
+                for (i = 0; i < (int)s->img_x; ++i) {
+                    unsigned char a;
+                    out[z + 2] = stbi__get8(s);
+                    out[z + 1] = stbi__get8(s);
+                    out[z + 0] = stbi__get8(s);
+                    z += 3;
+                    a = (easy == 2 ? stbi__get8(s) : 255);
+                    all_a |= a;
+                    if (target == 4)
+                        out[z++] = a;
+                }
+            } else {
+                int bpp = info.bpp;
+                for (i = 0; i < (int)s->img_x; ++i) {
+                    stbi__uint32 v = (bpp == 16 ? (stbi__uint32)stbi__get16le(s) : stbi__get32le(s));
+                    unsigned int a;
+                    out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount));
+                    out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount));
+                    out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount));
+                    a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255);
+                    all_a |= a;
+                    if (target == 4)
+                        out[z++] = STBI__BYTECAST(a);
+                }
+            }
+            stbi__skip(s, pad);
+        }
+    }
+
+    // if alpha channel is all 0s, replace with all 255s
+    if (target == 4 && all_a == 0)
+        for (i = 4 * s->img_x * s->img_y - 1; i >= 0; i -= 4)
+            out[i] = 255;
+
+    if (flip_vertically) {
+        stbi_uc t;
+        for (j = 0; j < (int)s->img_y >> 1; ++j) {
+            stbi_uc * p1 = out + j * s->img_x * target;
+            stbi_uc * p2 = out + (s->img_y - 1 - j) * s->img_x * target;
+            for (i = 0; i < (int)s->img_x * target; ++i) {
+                t = p1[i];
+                p1[i] = p2[i];
+                p2[i] = t;
+            }
+        }
+    }
+
+    if (req_comp && req_comp != target) {
+        out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y);
+        if (out == NULL)
+            return out; // stbi__convert_format frees input on failure
+    }
+
+    *x = s->img_x;
+    *y = s->img_y;
+    if (comp)
+        *comp = s->img_n;
+    return out;
+}
+#endif
+
+// Targa Truevision - TGA
+// by Jonathan Dummer
+#ifndef STBI_NO_TGA
+// returns STBI_rgb or whatever, 0 on error
+static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int * is_rgb16) {
+    // only RGB or RGBA (incl. 16bit) or grey allowed
+    if (is_rgb16)
+        *is_rgb16 = 0;
+    switch (bits_per_pixel) {
+    case 8:
+        return STBI_grey;
+    case 16:
+        if (is_grey)
+            return STBI_grey_alpha;
+        // fallthrough
+    case 15:
+        if (is_rgb16)
+            *is_rgb16 = 1;
+        return STBI_rgb;
+    case 24: // fallthrough
+    case 32:
+        return bits_per_pixel / 8;
+    default:
+        return 0;
+    }
+}
+
+static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp) {
+    int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp;
+    int sz, tga_colormap_type;
+    stbi__get8(s);                     // discard Offset
+    tga_colormap_type = stbi__get8(s); // colormap type
+    if (tga_colormap_type > 1) {
+        stbi__rewind(s);
+        return 0; // only RGB or indexed allowed
+    }
+    tga_image_type = stbi__get8(s); // image type
+    if (tga_colormap_type == 1) {   // colormapped (paletted) image
+        if (tga_image_type != 1 && tga_image_type != 9) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s, 4);   // skip index of first colormap entry and number of entries
+        sz = stbi__get8(s); //   check bits per palette color entry
+        if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32)) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s, 4); // skip image x and y origin
+        tga_colormap_bpp = sz;
+    } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE
+        if ((tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11)) {
+            stbi__rewind(s);
+            return 0; // only RGB or grey allowed, +/- RLE
+        }
+        stbi__skip(s, 9); // skip colormap specification and image x/y origin
+        tga_colormap_bpp = 0;
+    }
+    tga_w = stbi__get16le(s);
+    if (tga_w < 1) {
+        stbi__rewind(s);
+        return 0; // test width
+    }
+    tga_h = stbi__get16le(s);
+    if (tga_h < 1) {
+        stbi__rewind(s);
+        return 0; // test height
+    }
+    tga_bits_per_pixel = stbi__get8(s); // bits per pixel
+    stbi__get8(s);                      // ignore alpha bits
+    if (tga_colormap_bpp != 0) {
+        if ((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) {
+            // when using a colormap, tga_bits_per_pixel is the size of the indexes
+            // I don't think anything but 8 or 16bit indexes makes sense
+            stbi__rewind(s);
+            return 0;
+        }
+        tga_comp = stbi__tga_get_comp(tga_colormap_bpp, 0, NULL);
+    } else {
+        tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL);
+    }
+    if (!tga_comp) {
+        stbi__rewind(s);
+        return 0;
+    }
+    if (x)
+        *x = tga_w;
+    if (y)
+        *y = tga_h;
+    if (comp)
+        *comp = tga_comp;
+    return 1; // seems to have passed everything
+}
+
+static int stbi__tga_test(stbi__context * s) {
+    int res = 0;
+    int sz, tga_color_type;
+    stbi__get8(s);                  //   discard Offset
+    tga_color_type = stbi__get8(s); //   color type
+    if (tga_color_type > 1)
+        goto errorEnd;         //   only RGB or indexed allowed
+    sz = stbi__get8(s);        //   image type
+    if (tga_color_type == 1) { // colormapped (paletted) image
+        if (sz != 1 && sz != 9)
+            goto errorEnd;  // colortype 1 demands image type 1 or 9
+        stbi__skip(s, 4);   // skip index of first colormap entry and number of entries
+        sz = stbi__get8(s); //   check bits per palette color entry
+        if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32))
+            goto errorEnd;
+        stbi__skip(s, 4); // skip image x and y origin
+    } else {              // "normal" image w/o colormap
+        if ((sz != 2) && (sz != 3) && (sz != 10) && (sz != 11))
+            goto errorEnd; // only RGB or grey allowed, +/- RLE
+        stbi__skip(s, 9);  // skip colormap specification and image x/y origin
+    }
+    if (stbi__get16le(s) < 1)
+        goto errorEnd; //   test width
+    if (stbi__get16le(s) < 1)
+        goto errorEnd;  //   test height
+    sz = stbi__get8(s); //   bits per pixel
+    if ((tga_color_type == 1) && (sz != 8) && (sz != 16))
+        goto errorEnd; // for colormapped images, bpp is size of an index
+    if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32))
+        goto errorEnd;
+
+    res = 1; // if we got this far, everything's good and we can return 1 instead of 0
+
+errorEnd:
+    stbi__rewind(s);
+    return res;
+}
+
+// read 16bit value and convert to 24bit RGB
+static void stbi__tga_read_rgb16(stbi__context * s, stbi_uc * out) {
+    stbi__uint16 px = (stbi__uint16)stbi__get16le(s);
+    stbi__uint16 fiveBitMask = 31;
+    // we have 3 channels with 5bits each
+    int r = (px >> 10) & fiveBitMask;
+    int g = (px >> 5) & fiveBitMask;
+    int b = px & fiveBitMask;
+    // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later
+    out[0] = (stbi_uc)((r * 255) / 31);
+    out[1] = (stbi_uc)((g * 255) / 31);
+    out[2] = (stbi_uc)((b * 255) / 31);
+
+    // some people claim that the most significant bit might be used for alpha
+    // (possibly if an alpha-bit is set in the "image descriptor byte")
+    // but that only made 16bit test images completely translucent..
+    // so let's treat all 15 and 16bit TGAs as RGB with no alpha.
+}
+
+static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) {
+    //   read in the TGA header stuff
+    int tga_offset = stbi__get8(s);
+    int tga_indexed = stbi__get8(s);
+    int tga_image_type = stbi__get8(s);
+    int tga_is_RLE = 0;
+    int tga_palette_start = stbi__get16le(s);
+    int tga_palette_len = stbi__get16le(s);
+    int tga_palette_bits = stbi__get8(s);
+    int tga_x_origin = stbi__get16le(s);
+    int tga_y_origin = stbi__get16le(s);
+    int tga_width = stbi__get16le(s);
+    int tga_height = stbi__get16le(s);
+    int tga_bits_per_pixel = stbi__get8(s);
+    int tga_comp, tga_rgb16 = 0;
+    int tga_inverted = stbi__get8(s);
+    // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?)
+    //   image data
+    unsigned char * tga_data;
+    unsigned char * tga_palette = NULL;
+    int i, j;
+    unsigned char raw_data[4] = {0};
+    int RLE_count = 0;
+    int RLE_repeating = 0;
+    int read_next_pixel = 1;
+    STBI_NOTUSED(ri);
+    STBI_NOTUSED(tga_x_origin); // @TODO
+    STBI_NOTUSED(tga_y_origin); // @TODO
+
+    if (tga_height > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+    if (tga_width > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+
+    //   do a tiny bit of precessing
+    if (tga_image_type >= 8) {
+        tga_image_type -= 8;
+        tga_is_RLE = 1;
+    }
+    tga_inverted = 1 - ((tga_inverted >> 5) & 1);
+
+    //   If I'm paletted, then I'll use the number of bits from the palette
+    if (tga_indexed)
+        tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16);
+    else
+        tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16);
+
+    if (!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency
+        return stbi__errpuc("bad format", "Can't find out TGA pixelformat");
+
+    //   tga info
+    *x = tga_width;
+    *y = tga_height;
+    if (comp)
+        *comp = tga_comp;
+
+    if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0))
+        return stbi__errpuc("too large", "Corrupt TGA");
+
+    tga_data = (unsigned char *)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0);
+    if (!tga_data)
+        return stbi__errpuc("outofmem", "Out of memory");
+
+    // skip to the data's starting position (offset usually = 0)
+    stbi__skip(s, tga_offset);
+
+    if (!tga_indexed && !tga_is_RLE && !tga_rgb16) {
+        for (i = 0; i < tga_height; ++i) {
+            int row = tga_inverted ? tga_height - i - 1 : i;
+            stbi_uc * tga_row = tga_data + row * tga_width * tga_comp;
+            stbi__getn(s, tga_row, tga_width * tga_comp);
+        }
+    } else {
+        //   do I need to load a palette?
+        if (tga_indexed) {
+            if (tga_palette_len == 0) { /* you have to have at least one entry! */
+                STBI_FREE(tga_data);
+                return stbi__errpuc("bad palette", "Corrupt TGA");
+            }
+
+            //   any data to skip? (offset usually = 0)
+            stbi__skip(s, tga_palette_start);
+            //   load the palette
+            tga_palette = (unsigned char *)stbi__malloc_mad2(tga_palette_len, tga_comp, 0);
+            if (!tga_palette) {
+                STBI_FREE(tga_data);
+                return stbi__errpuc("outofmem", "Out of memory");
+            }
+            if (tga_rgb16) {
+                stbi_uc * pal_entry = tga_palette;
+                STBI_ASSERT(tga_comp == STBI_rgb);
+                for (i = 0; i < tga_palette_len; ++i) {
+                    stbi__tga_read_rgb16(s, pal_entry);
+                    pal_entry += tga_comp;
+                }
+            } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) {
+                STBI_FREE(tga_data);
+                STBI_FREE(tga_palette);
+                return stbi__errpuc("bad palette", "Corrupt TGA");
+            }
+        }
+        //   load the data
+        for (i = 0; i < tga_width * tga_height; ++i) {
+            //   if I'm in RLE mode, do I need to get a RLE stbi__pngchunk?
+            if (tga_is_RLE) {
+                if (RLE_count == 0) {
+                    //   yep, get the next byte as a RLE command
+                    int RLE_cmd = stbi__get8(s);
+                    RLE_count = 1 + (RLE_cmd & 127);
+                    RLE_repeating = RLE_cmd >> 7;
+                    read_next_pixel = 1;
+                } else if (!RLE_repeating) {
+                    read_next_pixel = 1;
+                }
+            } else {
+                read_next_pixel = 1;
+            }
+            //   OK, if I need to read a pixel, do it now
+            if (read_next_pixel) {
+                //   load however much data we did have
+                if (tga_indexed) {
+                    // read in index, then perform the lookup
+                    int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s);
+                    if (pal_idx >= tga_palette_len) {
+                        // invalid index
+                        pal_idx = 0;
+                    }
+                    pal_idx *= tga_comp;
+                    for (j = 0; j < tga_comp; ++j) {
+                        raw_data[j] = tga_palette[pal_idx + j];
+                    }
+                } else if (tga_rgb16) {
+                    STBI_ASSERT(tga_comp == STBI_rgb);
+                    stbi__tga_read_rgb16(s, raw_data);
+                } else {
+                    //   read in the data raw
+                    for (j = 0; j < tga_comp; ++j) {
+                        raw_data[j] = stbi__get8(s);
+                    }
+                }
+                //   clear the reading flag for the next pixel
+                read_next_pixel = 0;
+            } // end of reading a pixel
+
+            // copy data
+            for (j = 0; j < tga_comp; ++j)
+                tga_data[i * tga_comp + j] = raw_data[j];
+
+            //   in case we're in RLE mode, keep counting down
+            --RLE_count;
+        }
+        //   do I need to invert the image?
+        if (tga_inverted) {
+            for (j = 0; j * 2 < tga_height; ++j) {
+                int index1 = j * tga_width * tga_comp;
+                int index2 = (tga_height - 1 - j) * tga_width * tga_comp;
+                for (i = tga_width * tga_comp; i > 0; --i) {
+                    unsigned char temp = tga_data[index1];
+                    tga_data[index1] = tga_data[index2];
+                    tga_data[index2] = temp;
+                    ++index1;
+                    ++index2;
+                }
+            }
+        }
+        //   clear my palette, if I had one
+        if (tga_palette != NULL) {
+            STBI_FREE(tga_palette);
+        }
+    }
+
+    // swap RGB - if the source data was RGB16, it already is in the right order
+    if (tga_comp >= 3 && !tga_rgb16) {
+        unsigned char * tga_pixel = tga_data;
+        for (i = 0; i < tga_width * tga_height; ++i) {
+            unsigned char temp = tga_pixel[0];
+            tga_pixel[0] = tga_pixel[2];
+            tga_pixel[2] = temp;
+            tga_pixel += tga_comp;
+        }
+    }
+
+    // convert to target component count
+    if (req_comp && req_comp != tga_comp)
+        tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height);
+
+    //   the things I do to get rid of an error message, and yet keep
+    //   Microsoft's C compilers happy... [8^(
+    tga_palette_start = tga_palette_len = tga_palette_bits = tga_x_origin = tga_y_origin = 0;
+    STBI_NOTUSED(tga_palette_start);
+    //   OK, done
+    return tga_data;
+}
+#endif
+
+// *************************************************************************************************
+// Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_test(stbi__context * s) {
+    int r = (stbi__get32be(s) == 0x38425053);
+    stbi__rewind(s);
+    return r;
+}
+
+static int stbi__psd_decode_rle(stbi__context * s, stbi_uc * p, int pixelCount) {
+    int count, nleft, len;
+
+    count = 0;
+    while ((nleft = pixelCount - count) > 0) {
+        len = stbi__get8(s);
+        if (len == 128) {
+            // No-op.
+        } else if (len < 128) {
+            // Copy next len+1 bytes literally.
+            len++;
+            if (len > nleft)
+                return 0; // corrupt data
+            count += len;
+            while (len) {
+                *p = stbi__get8(s);
+                p += 4;
+                len--;
+            }
+        } else if (len > 128) {
+            stbi_uc val;
+            // Next -len+1 bytes in the dest are replicated from next source byte.
+            // (Interpret len as a negative 8-bit int.)
+            len = 257 - len;
+            if (len > nleft)
+                return 0; // corrupt data
+            val = stbi__get8(s);
+            count += len;
+            while (len) {
+                *p = val;
+                p += 4;
+                len--;
+            }
+        }
+    }
+
+    return 1;
+}
+
+static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc) {
+    int pixelCount;
+    int channelCount, compression;
+    int channel, i;
+    int bitdepth;
+    int w, h;
+    stbi_uc * out;
+    STBI_NOTUSED(ri);
+
+    // Check identifier
+    if (stbi__get32be(s) != 0x38425053) // "8BPS"
+        return stbi__errpuc("not PSD", "Corrupt PSD image");
+
+    // Check file type version.
+    if (stbi__get16be(s) != 1)
+        return stbi__errpuc("wrong version", "Unsupported version of PSD image");
+
+    // Skip 6 reserved bytes.
+    stbi__skip(s, 6);
+
+    // Read the number of channels (R, G, B, A, etc).
+    channelCount = stbi__get16be(s);
+    if (channelCount < 0 || channelCount > 16)
+        return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image");
+
+    // Read the rows and columns of the image.
+    h = stbi__get32be(s);
+    w = stbi__get32be(s);
+
+    if (h > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+    if (w > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+
+    // Make sure the depth is 8 bits.
+    bitdepth = stbi__get16be(s);
+    if (bitdepth != 8 && bitdepth != 16)
+        return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit");
+
+    // Make sure the color mode is RGB.
+    // Valid options are:
+    //   0: Bitmap
+    //   1: Grayscale
+    //   2: Indexed color
+    //   3: RGB color
+    //   4: CMYK color
+    //   7: Multichannel
+    //   8: Duotone
+    //   9: Lab color
+    if (stbi__get16be(s) != 3)
+        return stbi__errpuc("wrong color format", "PSD is not in RGB color format");
+
+    // Skip the Mode Data.  (It's the palette for indexed color; other info for other modes.)
+    stbi__skip(s, stbi__get32be(s));
+
+    // Skip the image resources.  (resolution, pen tool paths, etc)
+    stbi__skip(s, stbi__get32be(s));
+
+    // Skip the reserved data.
+    stbi__skip(s, stbi__get32be(s));
+
+    // Find out if the data is compressed.
+    // Known values:
+    //   0: no compression
+    //   1: RLE compressed
+    compression = stbi__get16be(s);
+    if (compression > 1)
+        return stbi__errpuc("bad compression", "PSD has an unknown compression format");
+
+    // Check size
+    if (!stbi__mad3sizes_valid(4, w, h, 0))
+        return stbi__errpuc("too large", "Corrupt PSD");
+
+    // Create the destination image.
+
+    if (!compression && bitdepth == 16 && bpc == 16) {
+        out = (stbi_uc *)stbi__malloc_mad3(8, w, h, 0);
+        ri->bits_per_channel = 16;
+    } else
+        out = (stbi_uc *)stbi__malloc(4 * w * h);
+
+    if (!out)
+        return stbi__errpuc("outofmem", "Out of memory");
+    pixelCount = w * h;
+
+    // Initialize the data to zero.
+    // memset( out, 0, pixelCount * 4 );
+
+    // Finally, the image data.
+    if (compression) {
+        // RLE as used by .PSD and .TIFF
+        // Loop until you get the number of unpacked bytes you are expecting:
+        //     Read the next source byte into n.
+        //     If n is between 0 and 127 inclusive, copy the next n+1 bytes literally.
+        //     Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times.
+        //     Else if n is 128, noop.
+        // Endloop
+
+        // The RLE-compressed data is preceded by a 2-byte data count for each row in the data,
+        // which we're going to just skip.
+        stbi__skip(s, h * channelCount * 2);
+
+        // Read the RLE data by channel.
+        for (channel = 0; channel < 4; channel++) {
+            stbi_uc * p;
+
+            p = out + channel;
+            if (channel >= channelCount) {
+                // Fill this channel with default data.
+                for (i = 0; i < pixelCount; i++, p += 4)
+                    *p = (channel == 3 ? 255 : 0);
+            } else {
+                // Read the RLE data.
+                if (!stbi__psd_decode_rle(s, p, pixelCount)) {
+                    STBI_FREE(out);
+                    return stbi__errpuc("corrupt", "bad RLE data");
+                }
+            }
+        }
+    } else {
+        // We're at the raw image data.  It's each channel in order (Red, Green, Blue, Alpha, ...)
+        // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image.
+
+        // Read the data by channel.
+        for (channel = 0; channel < 4; channel++) {
+            if (channel >= channelCount) {
+                // Fill this channel with default data.
+                if (bitdepth == 16 && bpc == 16) {
+                    stbi__uint16 * q = ((stbi__uint16 *)out) + channel;
+                    stbi__uint16 val = channel == 3 ? 65535 : 0;
+                    for (i = 0; i < pixelCount; i++, q += 4)
+                        *q = val;
+                } else {
+                    stbi_uc * p = out + channel;
+                    stbi_uc val = channel == 3 ? 255 : 0;
+                    for (i = 0; i < pixelCount; i++, p += 4)
+                        *p = val;
+                }
+            } else {
+                if (ri->bits_per_channel == 16) { // output bpc
+                    stbi__uint16 * q = ((stbi__uint16 *)out) + channel;
+                    for (i = 0; i < pixelCount; i++, q += 4)
+                        *q = (stbi__uint16)stbi__get16be(s);
+                } else {
+                    stbi_uc * p = out + channel;
+                    if (bitdepth == 16) { // input bpc
+                        for (i = 0; i < pixelCount; i++, p += 4)
+                            *p = (stbi_uc)(stbi__get16be(s) >> 8);
+                    } else {
+                        for (i = 0; i < pixelCount; i++, p += 4)
+                            *p = stbi__get8(s);
+                    }
+                }
+            }
+        }
+    }
+
+    // remove weird white matte from PSD
+    if (channelCount >= 4) {
+        if (ri->bits_per_channel == 16) {
+            for (i = 0; i < w * h; ++i) {
+                stbi__uint16 * pixel = (stbi__uint16 *)out + 4 * i;
+                if (pixel[3] != 0 && pixel[3] != 65535) {
+                    float a = pixel[3] / 65535.0f;
+                    float ra = 1.0f / a;
+                    float inv_a = 65535.0f * (1 - ra);
+                    pixel[0] = (stbi__uint16)(pixel[0] * ra + inv_a);
+                    pixel[1] = (stbi__uint16)(pixel[1] * ra + inv_a);
+                    pixel[2] = (stbi__uint16)(pixel[2] * ra + inv_a);
+                }
+            }
+        } else {
+            for (i = 0; i < w * h; ++i) {
+                unsigned char * pixel = out + 4 * i;
+                if (pixel[3] != 0 && pixel[3] != 255) {
+                    float a = pixel[3] / 255.0f;
+                    float ra = 1.0f / a;
+                    float inv_a = 255.0f * (1 - ra);
+                    pixel[0] = (unsigned char)(pixel[0] * ra + inv_a);
+                    pixel[1] = (unsigned char)(pixel[1] * ra + inv_a);
+                    pixel[2] = (unsigned char)(pixel[2] * ra + inv_a);
+                }
+            }
+        }
+    }
+
+    // convert to desired output format
+    if (req_comp && req_comp != 4) {
+        if (ri->bits_per_channel == 16)
+            out = (stbi_uc *)stbi__convert_format16((stbi__uint16 *)out, 4, req_comp, w, h);
+        else
+            out = stbi__convert_format(out, 4, req_comp, w, h);
+        if (out == NULL)
+            return out; // stbi__convert_format frees input on failure
+    }
+
+    if (comp)
+        *comp = 4;
+    *y = h;
+    *x = w;
+
+    return out;
+}
+#endif
+
+// *************************************************************************************************
+// Softimage PIC loader
+// by Tom Seddon
+//
+// See http://softimage.wiki.softimage.com/index.php/INFO:_PIC_file_format
+// See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_is4(stbi__context * s, const char * str) {
+    int i;
+    for (i = 0; i < 4; ++i)
+        if (stbi__get8(s) != (stbi_uc)str[i])
+            return 0;
+
+    return 1;
+}
+
+static int stbi__pic_test_core(stbi__context * s) {
+    int i;
+
+    if (!stbi__pic_is4(s, "\x53\x80\xF6\x34"))
+        return 0;
+
+    for (i = 0; i < 84; ++i)
+        stbi__get8(s);
+
+    if (!stbi__pic_is4(s, "PICT"))
+        return 0;
+
+    return 1;
+}
+
+typedef struct {
+    stbi_uc size, type, channel;
+} stbi__pic_packet;
+
+static stbi_uc * stbi__readval(stbi__context * s, int channel, stbi_uc * dest) {
+    int mask = 0x80, i;
+
+    for (i = 0; i < 4; ++i, mask >>= 1) {
+        if (channel & mask) {
+            if (stbi__at_eof(s))
+                return stbi__errpuc("bad file", "PIC file too short");
+            dest[i] = stbi__get8(s);
+        }
+    }
+
+    return dest;
+}
+
+static void stbi__copyval(int channel, stbi_uc * dest, const stbi_uc * src) {
+    int mask = 0x80, i;
+
+    for (i = 0; i < 4; ++i, mask >>= 1)
+        if (channel & mask)
+            dest[i] = src[i];
+}
+
+static stbi_uc * stbi__pic_load_core(stbi__context * s, int width, int height, int * comp, stbi_uc * result) {
+    int act_comp = 0, num_packets = 0, y, chained;
+    stbi__pic_packet packets[10];
+
+    // this will (should...) cater for even some bizarre stuff like having data
+    // for the same channel in multiple packets.
+    do {
+        stbi__pic_packet * packet;
+
+        if (num_packets == sizeof(packets) / sizeof(packets[0]))
+            return stbi__errpuc("bad format", "too many packets");
+
+        packet = &packets[num_packets++];
+
+        chained = stbi__get8(s);
+        packet->size = stbi__get8(s);
+        packet->type = stbi__get8(s);
+        packet->channel = stbi__get8(s);
+
+        act_comp |= packet->channel;
+
+        if (stbi__at_eof(s))
+            return stbi__errpuc("bad file", "file too short (reading packets)");
+        if (packet->size != 8)
+            return stbi__errpuc("bad format", "packet isn't 8bpp");
+    } while (chained);
+
+    *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel?
+
+    for (y = 0; y < height; ++y) {
+        int packet_idx;
+
+        for (packet_idx = 0; packet_idx < num_packets; ++packet_idx) {
+            stbi__pic_packet * packet = &packets[packet_idx];
+            stbi_uc * dest = result + y * width * 4;
+
+            switch (packet->type) {
+            default:
+                return stbi__errpuc("bad format", "packet has bad compression type");
+
+            case 0: { // uncompressed
+                int x;
+
+                for (x = 0; x < width; ++x, dest += 4)
+                    if (!stbi__readval(s, packet->channel, dest))
+                        return 0;
+                break;
+            }
+
+            case 1: // Pure RLE
+            {
+                int left = width, i;
+
+                while (left > 0) {
+                    stbi_uc count, value[4];
+
+                    count = stbi__get8(s);
+                    if (stbi__at_eof(s))
+                        return stbi__errpuc("bad file", "file too short (pure read count)");
+
+                    if (count > left)
+                        count = (stbi_uc)left;
+
+                    if (!stbi__readval(s, packet->channel, value))
+                        return 0;
+
+                    for (i = 0; i < count; ++i, dest += 4)
+                        stbi__copyval(packet->channel, dest, value);
+                    left -= count;
+                }
+            } break;
+
+            case 2: { // Mixed RLE
+                int left = width;
+                while (left > 0) {
+                    int count = stbi__get8(s), i;
+                    if (stbi__at_eof(s))
+                        return stbi__errpuc("bad file", "file too short (mixed read count)");
+
+                    if (count >= 128) { // Repeated
+                        stbi_uc value[4];
+
+                        if (count == 128)
+                            count = stbi__get16be(s);
+                        else
+                            count -= 127;
+                        if (count > left)
+                            return stbi__errpuc("bad file", "scanline overrun");
+
+                        if (!stbi__readval(s, packet->channel, value))
+                            return 0;
+
+                        for (i = 0; i < count; ++i, dest += 4)
+                            stbi__copyval(packet->channel, dest, value);
+                    } else { // Raw
+                        ++count;
+                        if (count > left)
+                            return stbi__errpuc("bad file", "scanline overrun");
+
+                        for (i = 0; i < count; ++i, dest += 4)
+                            if (!stbi__readval(s, packet->channel, dest))
+                                return 0;
+                    }
+                    left -= count;
+                }
+                break;
+            }
+            }
+        }
+    }
+
+    return result;
+}
+
+static void * stbi__pic_load(stbi__context * s, int * px, int * py, int * comp, int req_comp, stbi__result_info * ri) {
+    stbi_uc * result;
+    int i, x, y, internal_comp;
+    STBI_NOTUSED(ri);
+
+    if (!comp)
+        comp = &internal_comp;
+
+    for (i = 0; i < 92; ++i)
+        stbi__get8(s);
+
+    x = stbi__get16be(s);
+    y = stbi__get16be(s);
+
+    if (y > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+    if (x > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+
+    if (stbi__at_eof(s))
+        return stbi__errpuc("bad file", "file too short (pic header)");
+    if (!stbi__mad3sizes_valid(x, y, 4, 0))
+        return stbi__errpuc("too large", "PIC image too large to decode");
+
+    stbi__get32be(s); // skip `ratio'
+    stbi__get16be(s); // skip `fields'
+    stbi__get16be(s); // skip `pad'
+
+    // intermediate buffer is RGBA
+    result = (stbi_uc *)stbi__malloc_mad3(x, y, 4, 0);
+    if (!result)
+        return stbi__errpuc("outofmem", "Out of memory");
+    memset(result, 0xff, x * y * 4);
+
+    if (!stbi__pic_load_core(s, x, y, comp, result)) {
+        STBI_FREE(result);
+        result = 0;
+    }
+    *px = x;
+    *py = y;
+    if (req_comp == 0)
+        req_comp = *comp;
+    result = stbi__convert_format(result, 4, req_comp, x, y);
+
+    return result;
+}
+
+static int stbi__pic_test(stbi__context * s) {
+    int r = stbi__pic_test_core(s);
+    stbi__rewind(s);
+    return r;
+}
+#endif
+
+// *************************************************************************************************
+// GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb
+
+#ifndef STBI_NO_GIF
+typedef struct {
+    stbi__int16 prefix;
+    stbi_uc first;
+    stbi_uc suffix;
+} stbi__gif_lzw;
+
+typedef struct {
+    int w, h;
+    stbi_uc * out;        // output buffer (always 4 components)
+    stbi_uc * background; // The current "background" as far as a gif is concerned
+    stbi_uc * history;
+    int flags, bgindex, ratio, transparent, eflags;
+    stbi_uc pal[256][4];
+    stbi_uc lpal[256][4];
+    stbi__gif_lzw codes[8192];
+    stbi_uc * color_table;
+    int parse, step;
+    int lflags;
+    int start_x, start_y;
+    int max_x, max_y;
+    int cur_x, cur_y;
+    int line_size;
+    int delay;
+} stbi__gif;
+
+static int stbi__gif_test_raw(stbi__context * s) {
+    int sz;
+    if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8')
+        return 0;
+    sz = stbi__get8(s);
+    if (sz != '9' && sz != '7')
+        return 0;
+    if (stbi__get8(s) != 'a')
+        return 0;
+    return 1;
+}
+
+static int stbi__gif_test(stbi__context * s) {
+    int r = stbi__gif_test_raw(s);
+    stbi__rewind(s);
+    return r;
+}
+
+static void stbi__gif_parse_colortable(stbi__context * s, stbi_uc pal[256][4], int num_entries, int transp) {
+    int i;
+    for (i = 0; i < num_entries; ++i) {
+        pal[i][2] = stbi__get8(s);
+        pal[i][1] = stbi__get8(s);
+        pal[i][0] = stbi__get8(s);
+        pal[i][3] = transp == i ? 0 : 255;
+    }
+}
+
+static int stbi__gif_header(stbi__context * s, stbi__gif * g, int * comp, int is_info) {
+    stbi_uc version;
+    if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8')
+        return stbi__err("not GIF", "Corrupt GIF");
+
+    version = stbi__get8(s);
+    if (version != '7' && version != '9')
+        return stbi__err("not GIF", "Corrupt GIF");
+    if (stbi__get8(s) != 'a')
+        return stbi__err("not GIF", "Corrupt GIF");
+
+    stbi__g_failure_reason = "";
+    g->w = stbi__get16le(s);
+    g->h = stbi__get16le(s);
+    g->flags = stbi__get8(s);
+    g->bgindex = stbi__get8(s);
+    g->ratio = stbi__get8(s);
+    g->transparent = -1;
+
+    if (g->w > STBI_MAX_DIMENSIONS)
+        return stbi__err("too large", "Very large image (corrupt?)");
+    if (g->h > STBI_MAX_DIMENSIONS)
+        return stbi__err("too large", "Very large image (corrupt?)");
+
+    if (comp != 0)
+        *comp = 4; // can't actually tell whether it's 3 or 4 until we parse the comments
+
+    if (is_info)
+        return 1;
+
+    if (g->flags & 0x80)
+        stbi__gif_parse_colortable(s, g->pal, 2 << (g->flags & 7), -1);
+
+    return 1;
+}
+
+static int stbi__gif_info_raw(stbi__context * s, int * x, int * y, int * comp) {
+    stbi__gif * g = (stbi__gif *)stbi__malloc(sizeof(stbi__gif));
+    if (!g)
+        return stbi__err("outofmem", "Out of memory");
+    if (!stbi__gif_header(s, g, comp, 1)) {
+        STBI_FREE(g);
+        stbi__rewind(s);
+        return 0;
+    }
+    if (x)
+        *x = g->w;
+    if (y)
+        *y = g->h;
+    STBI_FREE(g);
+    return 1;
+}
+
+static void stbi__out_gif_code(stbi__gif * g, stbi__uint16 code) {
+    stbi_uc *p, *c;
+    int idx;
+
+    // recurse to decode the prefixes, since the linked-list is backwards,
+    // and working backwards through an interleaved image would be nasty
+    if (g->codes[code].prefix >= 0)
+        stbi__out_gif_code(g, g->codes[code].prefix);
+
+    if (g->cur_y >= g->max_y)
+        return;
+
+    idx = g->cur_x + g->cur_y;
+    p = &g->out[idx];
+    g->history[idx / 4] = 1;
+
+    c = &g->color_table[g->codes[code].suffix * 4];
+    if (c[3] > 128) { // don't render transparent pixels;
+        p[0] = c[2];
+        p[1] = c[1];
+        p[2] = c[0];
+        p[3] = c[3];
+    }
+    g->cur_x += 4;
+
+    if (g->cur_x >= g->max_x) {
+        g->cur_x = g->start_x;
+        g->cur_y += g->step;
+
+        while (g->cur_y >= g->max_y && g->parse > 0) {
+            g->step = (1 << g->parse) * g->line_size;
+            g->cur_y = g->start_y + (g->step >> 1);
+            --g->parse;
+        }
+    }
+}
+
+static stbi_uc * stbi__process_gif_raster(stbi__context * s, stbi__gif * g) {
+    stbi_uc lzw_cs;
+    stbi__int32 len, init_code;
+    stbi__uint32 first;
+    stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear;
+    stbi__gif_lzw * p;
+
+    lzw_cs = stbi__get8(s);
+    if (lzw_cs > 12)
+        return NULL;
+    clear = 1 << lzw_cs;
+    first = 1;
+    codesize = lzw_cs + 1;
+    codemask = (1 << codesize) - 1;
+    bits = 0;
+    valid_bits = 0;
+    for (init_code = 0; init_code < clear; init_code++) {
+        g->codes[init_code].prefix = -1;
+        g->codes[init_code].first = (stbi_uc)init_code;
+        g->codes[init_code].suffix = (stbi_uc)init_code;
+    }
+
+    // support no starting clear code
+    avail = clear + 2;
+    oldcode = -1;
+
+    len = 0;
+    for (;;) {
+        if (valid_bits < codesize) {
+            if (len == 0) {
+                len = stbi__get8(s); // start new block
+                if (len == 0)
+                    return g->out;
+            }
+            --len;
+            bits |= (stbi__int32)stbi__get8(s) << valid_bits;
+            valid_bits += 8;
+        } else {
+            stbi__int32 code = bits & codemask;
+            bits >>= codesize;
+            valid_bits -= codesize;
+            // @OPTIMIZE: is there some way we can accelerate the non-clear path?
+            if (code == clear) { // clear code
+                codesize = lzw_cs + 1;
+                codemask = (1 << codesize) - 1;
+                avail = clear + 2;
+                oldcode = -1;
+                first = 0;
+            } else if (code == clear + 1) { // end of stream code
+                stbi__skip(s, len);
+                while ((len = stbi__get8(s)) > 0)
+                    stbi__skip(s, len);
+                return g->out;
+            } else if (code <= avail) {
+                if (first) {
+                    return stbi__errpuc("no clear code", "Corrupt GIF");
+                }
+
+                if (oldcode >= 0) {
+                    p = &g->codes[avail++];
+                    if (avail > 8192) {
+                        return stbi__errpuc("too many codes", "Corrupt GIF");
+                    }
+
+                    p->prefix = (stbi__int16)oldcode;
+                    p->first = g->codes[oldcode].first;
+                    p->suffix = (code == avail) ? p->first : g->codes[code].first;
+                } else if (code == avail)
+                    return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+
+                stbi__out_gif_code(g, (stbi__uint16)code);
+
+                if ((avail & codemask) == 0 && avail <= 0x0FFF) {
+                    codesize++;
+                    codemask = (1 << codesize) - 1;
+                }
+
+                oldcode = code;
+            } else {
+                return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+            }
+        }
+    }
+}
+
+// this function is designed to support animated gifs, although stb_image doesn't support it
+// two back is the image from two frames ago, used for a very specific disposal format
+static stbi_uc * stbi__gif_load_next(stbi__context * s, stbi__gif * g, int * comp, int req_comp, stbi_uc * two_back) {
+    int dispose;
+    int first_frame;
+    int pi;
+    int pcount;
+    STBI_NOTUSED(req_comp);
+
+    // on first frame, any non-written pixels get the background colour (non-transparent)
+    first_frame = 0;
+    if (g->out == 0) {
+        if (!stbi__gif_header(s, g, comp, 0))
+            return 0; // stbi__g_failure_reason set by stbi__gif_header
+        if (!stbi__mad3sizes_valid(4, g->w, g->h, 0))
+            return stbi__errpuc("too large", "GIF image is too large");
+        pcount = g->w * g->h;
+        g->out = (stbi_uc *)stbi__malloc(4 * pcount);
+        g->background = (stbi_uc *)stbi__malloc(4 * pcount);
+        g->history = (stbi_uc *)stbi__malloc(pcount);
+        if (!g->out || !g->background || !g->history)
+            return stbi__errpuc("outofmem", "Out of memory");
+
+        // image is treated as "transparent" at the start - ie, nothing overwrites the current background;
+        // background colour is only used for pixels that are not rendered first frame, after that "background"
+        // color refers to the color that was there the previous frame.
+        memset(g->out, 0x00, 4 * pcount);
+        memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent)
+        memset(g->history, 0x00, pcount);        // pixels that were affected previous frame
+        first_frame = 1;
+    } else {
+        // second frame - how do we dispose of the previous one?
+        dispose = (g->eflags & 0x1C) >> 2;
+        pcount = g->w * g->h;
+
+        if ((dispose == 3) && (two_back == 0)) {
+            dispose = 2; // if I don't have an image to revert back to, default to the old background
+        }
+
+        if (dispose == 3) { // use previous graphic
+            for (pi = 0; pi < pcount; ++pi) {
+                if (g->history[pi]) {
+                    memcpy(&g->out[pi * 4], &two_back[pi * 4], 4);
+                }
+            }
+        } else if (dispose == 2) {
+            // restore what was changed last frame to background before that frame;
+            for (pi = 0; pi < pcount; ++pi) {
+                if (g->history[pi]) {
+                    memcpy(&g->out[pi * 4], &g->background[pi * 4], 4);
+                }
+            }
+        } else {
+            // This is a non-disposal case eithe way, so just
+            // leave the pixels as is, and they will become the new background
+            // 1: do not dispose
+            // 0:  not specified.
+        }
+
+        // background is what out is after the undoing of the previou frame;
+        memcpy(g->background, g->out, 4 * g->w * g->h);
+    }
+
+    // clear my history;
+    memset(g->history, 0x00, g->w * g->h); // pixels that were affected previous frame
+
+    for (;;) {
+        int tag = stbi__get8(s);
+        switch (tag) {
+        case 0x2C: /* Image Descriptor */
+        {
+            stbi__int32 x, y, w, h;
+            stbi_uc * o;
+
+            x = stbi__get16le(s);
+            y = stbi__get16le(s);
+            w = stbi__get16le(s);
+            h = stbi__get16le(s);
+            if (((x + w) > (g->w)) || ((y + h) > (g->h)))
+                return stbi__errpuc("bad Image Descriptor", "Corrupt GIF");
+
+            g->line_size = g->w * 4;
+            g->start_x = x * 4;
+            g->start_y = y * g->line_size;
+            g->max_x = g->start_x + w * 4;
+            g->max_y = g->start_y + h * g->line_size;
+            g->cur_x = g->start_x;
+            g->cur_y = g->start_y;
+
+            // if the width of the specified rectangle is 0, that means
+            // we may not see *any* pixels or the image is malformed;
+            // to make sure this is caught, move the current y down to
+            // max_y (which is what out_gif_code checks).
+            if (w == 0)
+                g->cur_y = g->max_y;
+
+            g->lflags = stbi__get8(s);
+
+            if (g->lflags & 0x40) {
+                g->step = 8 * g->line_size; // first interlaced spacing
+                g->parse = 3;
+            } else {
+                g->step = g->line_size;
+                g->parse = 0;
+            }
+
+            if (g->lflags & 0x80) {
+                stbi__gif_parse_colortable(s, g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1);
+                g->color_table = (stbi_uc *)g->lpal;
+            } else if (g->flags & 0x80) {
+                g->color_table = (stbi_uc *)g->pal;
+            } else
+                return stbi__errpuc("missing color table", "Corrupt GIF");
+
+            o = stbi__process_gif_raster(s, g);
+            if (!o)
+                return NULL;
+
+            // if this was the first frame,
+            pcount = g->w * g->h;
+            if (first_frame && (g->bgindex > 0)) {
+                // if first frame, any pixel not drawn to gets the background color
+                for (pi = 0; pi < pcount; ++pi) {
+                    if (g->history[pi] == 0) {
+                        g->pal[g->bgindex][3] =
+                            255; // just in case it was made transparent, undo that; It will be reset next frame if need be;
+                        memcpy(&g->out[pi * 4], &g->pal[g->bgindex], 4);
+                    }
+                }
+            }
+
+            return o;
+        }
+
+        case 0x21: // Comment Extension.
+        {
+            int len;
+            int ext = stbi__get8(s);
+            if (ext == 0xF9) { // Graphic Control Extension.
+                len = stbi__get8(s);
+                if (len == 4) {
+                    g->eflags = stbi__get8(s);
+                    g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths.
+
+                    // unset old transparent
+                    if (g->transparent >= 0) {
+                        g->pal[g->transparent][3] = 255;
+                    }
+                    if (g->eflags & 0x01) {
+                        g->transparent = stbi__get8(s);
+                        if (g->transparent >= 0) {
+                            g->pal[g->transparent][3] = 0;
+                        }
+                    } else {
+                        // don't need transparent
+                        stbi__skip(s, 1);
+                        g->transparent = -1;
+                    }
+                } else {
+                    stbi__skip(s, len);
+                    break;
+                }
+            }
+            while ((len = stbi__get8(s)) != 0) {
+                stbi__skip(s, len);
+            }
+            break;
+        }
+
+        case 0x3B:               // gif stream termination code
+            return (stbi_uc *)s; // using '1' causes warning on some compilers
+
+        default:
+            return stbi__errpuc("unknown code", "Corrupt GIF");
+        }
+    }
+}
+
+static void * stbi__load_gif_main_outofmem(stbi__gif * g, stbi_uc * out, int ** delays) {
+    STBI_FREE(g->out);
+    STBI_FREE(g->history);
+    STBI_FREE(g->background);
+
+    if (out)
+        STBI_FREE(out);
+    if (delays && *delays)
+        STBI_FREE(*delays);
+    return stbi__errpuc("outofmem", "Out of memory");
+}
+
+static void * stbi__load_gif_main(stbi__context * s, int ** delays, int * x, int * y, int * z, int * comp, int req_comp) {
+    if (stbi__gif_test(s)) {
+        int layers = 0;
+        stbi_uc * u = 0;
+        stbi_uc * out = 0;
+        stbi_uc * two_back = 0;
+        stbi__gif g;
+        int stride;
+        int out_size = 0;
+        int delays_size = 0;
+
+        STBI_NOTUSED(out_size);
+        STBI_NOTUSED(delays_size);
+
+        memset(&g, 0, sizeof(g));
+        if (delays) {
+            *delays = 0;
+        }
+
+        do {
+            u = stbi__gif_load_next(s, &g, comp, req_comp, two_back);
+            if (u == (stbi_uc *)s)
+                u = 0; // end of animated gif marker
+
+            if (u) {
+                *x = g.w;
+                *y = g.h;
+                ++layers;
+                stride = g.w * g.h * 4;
+
+                if (out) {
+                    void * tmp = (stbi_uc *)STBI_REALLOC_SIZED(out, out_size, layers * stride);
+                    if (!tmp)
+                        return stbi__load_gif_main_outofmem(&g, out, delays);
+                    else {
+                        out = (stbi_uc *)tmp;
+                        out_size = layers * stride;
+                    }
+
+                    if (delays) {
+                        int * new_delays = (int *)STBI_REALLOC_SIZED(*delays, delays_size, sizeof(int) * layers);
+                        if (!new_delays)
+                            return stbi__load_gif_main_outofmem(&g, out, delays);
+                        *delays = new_delays;
+                        delays_size = layers * sizeof(int);
+                    }
+                } else {
+                    out = (stbi_uc *)stbi__malloc(layers * stride);
+                    if (!out)
+                        return stbi__load_gif_main_outofmem(&g, out, delays);
+                    out_size = layers * stride;
+                    if (delays) {
+                        *delays = (int *)stbi__malloc(layers * sizeof(int));
+                        if (!*delays)
+                            return stbi__load_gif_main_outofmem(&g, out, delays);
+                        delays_size = layers * sizeof(int);
+                    }
+                }
+                memcpy(out + ((layers - 1) * stride), u, stride);
+                if (layers >= 2) {
+                    two_back = out - 2 * stride;
+                }
+
+                if (delays) {
+                    (*delays)[layers - 1U] = g.delay;
+                }
+            }
+        } while (u != 0);
+
+        // free temp buffer;
+        STBI_FREE(g.out);
+        STBI_FREE(g.history);
+        STBI_FREE(g.background);
+
+        // do the final conversion after loading everything;
+        if (req_comp && req_comp != 4)
+            out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h);
+
+        *z = layers;
+        return out;
+    } else {
+        return stbi__errpuc("not GIF", "Image was not as a gif type.");
+    }
+}
+
+static void * stbi__gif_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) {
+    stbi_uc * u = 0;
+    stbi__gif g;
+    memset(&g, 0, sizeof(g));
+    STBI_NOTUSED(ri);
+
+    u = stbi__gif_load_next(s, &g, comp, req_comp, 0);
+    if (u == (stbi_uc *)s)
+        u = 0; // end of animated gif marker
+    if (u) {
+        *x = g.w;
+        *y = g.h;
+
+        // moved conversion to after successful load so that the same
+        // can be done for multiple frames.
+        if (req_comp && req_comp != 4)
+            u = stbi__convert_format(u, 4, req_comp, g.w, g.h);
+    } else if (g.out) {
+        // if there was an error and we allocated an image buffer, free it!
+        STBI_FREE(g.out);
+    }
+
+    // free buffers needed for multiple frame loading;
+    STBI_FREE(g.history);
+    STBI_FREE(g.background);
+
+    return u;
+}
+
+static int stbi__gif_info(stbi__context * s, int * x, int * y, int * comp) { return stbi__gif_info_raw(s, x, y, comp); }
+#endif
+
+// *************************************************************************************************
+// Radiance RGBE HDR loader
+// originally by Nicolas Schulz
+#ifndef STBI_NO_HDR
+static int stbi__hdr_test_core(stbi__context * s, const char * signature) {
+    int i;
+    for (i = 0; signature[i]; ++i)
+        if (stbi__get8(s) != signature[i])
+            return 0;
+    stbi__rewind(s);
+    return 1;
+}
+
+static int stbi__hdr_test(stbi__context * s) {
+    int r = stbi__hdr_test_core(s, "#?RADIANCE\n");
+    stbi__rewind(s);
+    if (!r) {
+        r = stbi__hdr_test_core(s, "#?RGBE\n");
+        stbi__rewind(s);
+    }
+    return r;
+}
+
+#define STBI__HDR_BUFLEN 1024
+static char * stbi__hdr_gettoken(stbi__context * z, char * buffer) {
+    int len = 0;
+    char c = '\0';
+
+    c = (char)stbi__get8(z);
+
+    while (!stbi__at_eof(z) && c != '\n') {
+        buffer[len++] = c;
+        if (len == STBI__HDR_BUFLEN - 1) {
+            // flush to end of line
+            while (!stbi__at_eof(z) && stbi__get8(z) != '\n')
+                ;
+            break;
+        }
+        c = (char)stbi__get8(z);
+    }
+
+    buffer[len] = 0;
+    return buffer;
+}
+
+static void stbi__hdr_convert(float * output, stbi_uc * input, int req_comp) {
+    if (input[3] != 0) {
+        float f1;
+        // Exponent
+        f1 = (float)ldexp(1.0f, input[3] - (int)(128 + 8));
+        if (req_comp <= 2)
+            output[0] = (input[0] + input[1] + input[2]) * f1 / 3;
+        else {
+            output[0] = input[0] * f1;
+            output[1] = input[1] * f1;
+            output[2] = input[2] * f1;
+        }
+        if (req_comp == 2)
+            output[1] = 1;
+        if (req_comp == 4)
+            output[3] = 1;
+    } else {
+        switch (req_comp) {
+        case 4:
+            output[3] = 1; /* fallthrough */
+        case 3:
+            output[0] = output[1] = output[2] = 0;
+            break;
+        case 2:
+            output[1] = 1; /* fallthrough */
+        case 1:
+            output[0] = 0;
+            break;
+        }
+    }
+}
+
+static float * stbi__hdr_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) {
+    char buffer[STBI__HDR_BUFLEN];
+    char * token;
+    int valid = 0;
+    int width, height;
+    stbi_uc * scanline;
+    float * hdr_data;
+    int len;
+    unsigned char count, value;
+    int i, j, k, c1, c2, z;
+    const char * headerToken;
+    STBI_NOTUSED(ri);
+
+    // Check identifier
+    headerToken = stbi__hdr_gettoken(s, buffer);
+    if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0)
+        return stbi__errpf("not HDR", "Corrupt HDR image");
+
+    // Parse header
+    for (;;) {
+        token = stbi__hdr_gettoken(s, buffer);
+        if (token[0] == 0)
+            break;
+        if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0)
+            valid = 1;
+    }
+
+    if (!valid)
+        return stbi__errpf("unsupported format", "Unsupported HDR format");
+
+    // Parse width and height
+    // can't use sscanf() if we're not using stdio!
+    token = stbi__hdr_gettoken(s, buffer);
+    if (strncmp(token, "-Y ", 3))
+        return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+    token += 3;
+    height = (int)strtol(token, &token, 10);
+    while (*token == ' ')
+        ++token;
+    if (strncmp(token, "+X ", 3))
+        return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+    token += 3;
+    width = (int)strtol(token, NULL, 10);
+
+    if (height > STBI_MAX_DIMENSIONS)
+        return stbi__errpf("too large", "Very large image (corrupt?)");
+    if (width > STBI_MAX_DIMENSIONS)
+        return stbi__errpf("too large", "Very large image (corrupt?)");
+
+    *x = width;
+    *y = height;
+
+    if (comp)
+        *comp = 3;
+    if (req_comp == 0)
+        req_comp = 3;
+
+    if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0))
+        return stbi__errpf("too large", "HDR image is too large");
+
+    // Read data
+    hdr_data = (float *)stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0);
+    if (!hdr_data)
+        return stbi__errpf("outofmem", "Out of memory");
+
+    // Load image data
+    // image data is stored as some number of sca
+    if (width < 8 || width >= 32768) {
+        // Read flat data
+        for (j = 0; j < height; ++j) {
+            for (i = 0; i < width; ++i) {
+                stbi_uc rgbe[4];
+            main_decode_loop:
+                stbi__getn(s, rgbe, 4);
+                stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp);
+            }
+        }
+    } else {
+        // Read RLE-encoded data
+        scanline = NULL;
+
+        for (j = 0; j < height; ++j) {
+            c1 = stbi__get8(s);
+            c2 = stbi__get8(s);
+            len = stbi__get8(s);
+            if (c1 != 2 || c2 != 2 || (len & 0x80)) {
+                // not run-length encoded, so we have to actually use THIS data as a decoded
+                // pixel (note this can't be a valid pixel--one of RGB must be >= 128)
+                stbi_uc rgbe[4];
+                rgbe[0] = (stbi_uc)c1;
+                rgbe[1] = (stbi_uc)c2;
+                rgbe[2] = (stbi_uc)len;
+                rgbe[3] = (stbi_uc)stbi__get8(s);
+                stbi__hdr_convert(hdr_data, rgbe, req_comp);
+                i = 1;
+                j = 0;
+                STBI_FREE(scanline);
+                goto main_decode_loop; // yes, this makes no sense
+            }
+            len <<= 8;
+            len |= stbi__get8(s);
+            if (len != width) {
+                STBI_FREE(hdr_data);
+                STBI_FREE(scanline);
+                return stbi__errpf("invalid decoded scanline length", "corrupt HDR");
+            }
+            if (scanline == NULL) {
+                scanline = (stbi_uc *)stbi__malloc_mad2(width, 4, 0);
+                if (!scanline) {
+                    STBI_FREE(hdr_data);
+                    return stbi__errpf("outofmem", "Out of memory");
+                }
+            }
+
+            for (k = 0; k < 4; ++k) {
+                int nleft;
+                i = 0;
+                while ((nleft = width - i) > 0) {
+                    count = stbi__get8(s);
+                    if (count > 128) {
+                        // Run
+                        value = stbi__get8(s);
+                        count -= 128;
+                        if ((count == 0) || (count > nleft)) {
+                            STBI_FREE(hdr_data);
+                            STBI_FREE(scanline);
+                            return stbi__errpf("corrupt", "bad RLE data in HDR");
+                        }
+                        for (z = 0; z < count; ++z)
+                            scanline[i++ * 4 + k] = value;
+                    } else {
+                        // Dump
+                        if ((count == 0) || (count > nleft)) {
+                            STBI_FREE(hdr_data);
+                            STBI_FREE(scanline);
+                            return stbi__errpf("corrupt", "bad RLE data in HDR");
+                        }
+                        for (z = 0; z < count; ++z)
+                            scanline[i++ * 4 + k] = stbi__get8(s);
+                    }
+                }
+            }
+            for (i = 0; i < width; ++i)
+                stbi__hdr_convert(hdr_data + (j * width + i) * req_comp, scanline + i * 4, req_comp);
+        }
+        if (scanline)
+            STBI_FREE(scanline);
+    }
+
+    return hdr_data;
+}
+
+static int stbi__hdr_info(stbi__context * s, int * x, int * y, int * comp) {
+    char buffer[STBI__HDR_BUFLEN];
+    char * token;
+    int valid = 0;
+    int dummy;
+
+    if (!x)
+        x = &dummy;
+    if (!y)
+        y = &dummy;
+    if (!comp)
+        comp = &dummy;
+
+    if (stbi__hdr_test(s) == 0) {
+        stbi__rewind(s);
+        return 0;
+    }
+
+    for (;;) {
+        token = stbi__hdr_gettoken(s, buffer);
+        if (token[0] == 0)
+            break;
+        if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0)
+            valid = 1;
+    }
+
+    if (!valid) {
+        stbi__rewind(s);
+        return 0;
+    }
+    token = stbi__hdr_gettoken(s, buffer);
+    if (strncmp(token, "-Y ", 3)) {
+        stbi__rewind(s);
+        return 0;
+    }
+    token += 3;
+    *y = (int)strtol(token, &token, 10);
+    while (*token == ' ')
+        ++token;
+    if (strncmp(token, "+X ", 3)) {
+        stbi__rewind(s);
+        return 0;
+    }
+    token += 3;
+    *x = (int)strtol(token, NULL, 10);
+    *comp = 3;
+    return 1;
+}
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_info(stbi__context * s, int * x, int * y, int * comp) {
+    void * p;
+    stbi__bmp_data info;
+
+    info.all_a = 255;
+    p = stbi__bmp_parse_header(s, &info);
+    if (p == NULL) {
+        stbi__rewind(s);
+        return 0;
+    }
+    if (x)
+        *x = s->img_x;
+    if (y)
+        *y = s->img_y;
+    if (comp) {
+        if (info.bpp == 24 && info.ma == 0xff000000)
+            *comp = 3;
+        else
+            *comp = info.ma ? 4 : 3;
+    }
+    return 1;
+}
+#endif
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_info(stbi__context * s, int * x, int * y, int * comp) {
+    int channelCount, dummy, depth;
+    if (!x)
+        x = &dummy;
+    if (!y)
+        y = &dummy;
+    if (!comp)
+        comp = &dummy;
+    if (stbi__get32be(s) != 0x38425053) {
+        stbi__rewind(s);
+        return 0;
+    }
+    if (stbi__get16be(s) != 1) {
+        stbi__rewind(s);
+        return 0;
+    }
+    stbi__skip(s, 6);
+    channelCount = stbi__get16be(s);
+    if (channelCount < 0 || channelCount > 16) {
+        stbi__rewind(s);
+        return 0;
+    }
+    *y = stbi__get32be(s);
+    *x = stbi__get32be(s);
+    depth = stbi__get16be(s);
+    if (depth != 8 && depth != 16) {
+        stbi__rewind(s);
+        return 0;
+    }
+    if (stbi__get16be(s) != 3) {
+        stbi__rewind(s);
+        return 0;
+    }
+    *comp = 4;
+    return 1;
+}
+
+static int stbi__psd_is16(stbi__context * s) {
+    int channelCount, depth;
+    if (stbi__get32be(s) != 0x38425053) {
+        stbi__rewind(s);
+        return 0;
+    }
+    if (stbi__get16be(s) != 1) {
+        stbi__rewind(s);
+        return 0;
+    }
+    stbi__skip(s, 6);
+    channelCount = stbi__get16be(s);
+    if (channelCount < 0 || channelCount > 16) {
+        stbi__rewind(s);
+        return 0;
+    }
+    STBI_NOTUSED(stbi__get32be(s));
+    STBI_NOTUSED(stbi__get32be(s));
+    depth = stbi__get16be(s);
+    if (depth != 16) {
+        stbi__rewind(s);
+        return 0;
+    }
+    return 1;
+}
+#endif
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp) {
+    int act_comp = 0, num_packets = 0, chained, dummy;
+    stbi__pic_packet packets[10];
+
+    if (!x)
+        x = &dummy;
+    if (!y)
+        y = &dummy;
+    if (!comp)
+        comp = &dummy;
+
+    if (!stbi__pic_is4(s, "\x53\x80\xF6\x34")) {
+        stbi__rewind(s);
+        return 0;
+    }
+
+    stbi__skip(s, 88);
+
+    *x = stbi__get16be(s);
+    *y = stbi__get16be(s);
+    if (stbi__at_eof(s)) {
+        stbi__rewind(s);
+        return 0;
+    }
+    if ((*x) != 0 && (1 << 28) / (*x) < (*y)) {
+        stbi__rewind(s);
+        return 0;
+    }
+
+    stbi__skip(s, 8);
+
+    do {
+        stbi__pic_packet * packet;
+
+        if (num_packets == sizeof(packets) / sizeof(packets[0]))
+            return 0;
+
+        packet = &packets[num_packets++];
+        chained = stbi__get8(s);
+        packet->size = stbi__get8(s);
+        packet->type = stbi__get8(s);
+        packet->channel = stbi__get8(s);
+        act_comp |= packet->channel;
+
+        if (stbi__at_eof(s)) {
+            stbi__rewind(s);
+            return 0;
+        }
+        if (packet->size != 8) {
+            stbi__rewind(s);
+            return 0;
+        }
+    } while (chained);
+
+    *comp = (act_comp & 0x10 ? 4 : 3);
+
+    return 1;
+}
+#endif
+
+// *************************************************************************************************
+// Portable Gray Map and Portable Pixel Map loader
+// by Ken Miller
+//
+// PGM: http://netpbm.sourceforge.net/doc/pgm.html
+// PPM: http://netpbm.sourceforge.net/doc/ppm.html
+//
+// Known limitations:
+//    Does not support comments in the header section
+//    Does not support ASCII image data (formats P2 and P3)
+
+#ifndef STBI_NO_PNM
+
+static int stbi__pnm_test(stbi__context * s) {
+    char p, t;
+    p = (char)stbi__get8(s);
+    t = (char)stbi__get8(s);
+    if (p != 'P' || (t != '5' && t != '6')) {
+        stbi__rewind(s);
+        return 0;
+    }
+    return 1;
+}
+
+static void * stbi__pnm_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) {
+    stbi_uc * out;
+    STBI_NOTUSED(ri);
+
+    ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n);
+    if (ri->bits_per_channel == 0)
+        return 0;
+
+    if (s->img_y > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+    if (s->img_x > STBI_MAX_DIMENSIONS)
+        return stbi__errpuc("too large", "Very large image (corrupt?)");
+
+    *x = s->img_x;
+    *y = s->img_y;
+    if (comp)
+        *comp = s->img_n;
+
+    if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0))
+        return stbi__errpuc("too large", "PNM too large");
+
+    out = (stbi_uc *)stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0);
+    if (!out)
+        return stbi__errpuc("outofmem", "Out of memory");
+    if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) {
+        STBI_FREE(out);
+        return stbi__errpuc("bad PNM", "PNM file truncated");
+    }
+
+    if (req_comp && req_comp != s->img_n) {
+        if (ri->bits_per_channel == 16) {
+            out = (stbi_uc *)stbi__convert_format16((stbi__uint16 *)out, s->img_n, req_comp, s->img_x, s->img_y);
+        } else {
+            out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y);
+        }
+        if (out == NULL)
+            return out; // stbi__convert_format frees input on failure
+    }
+    return out;
+}
+
+static int stbi__pnm_isspace(char c) { return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r'; }
+
+static void stbi__pnm_skip_whitespace(stbi__context * s, char * c) {
+    for (;;) {
+        while (!stbi__at_eof(s) && stbi__pnm_isspace(*c))
+            *c = (char)stbi__get8(s);
+
+        if (stbi__at_eof(s) || *c != '#')
+            break;
+
+        while (!stbi__at_eof(s) && *c != '\n' && *c != '\r')
+            *c = (char)stbi__get8(s);
+    }
+}
+
+static int stbi__pnm_isdigit(char c) { return c >= '0' && c <= '9'; }
+
+static int stbi__pnm_getinteger(stbi__context * s, char * c) {
+    int value = 0;
+
+    while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) {
+        value = value * 10 + (*c - '0');
+        *c = (char)stbi__get8(s);
+        if ((value > 214748364) || (value == 214748364 && *c > '7'))
+            return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int");
+    }
+
+    return value;
+}
+
+static int stbi__pnm_info(stbi__context * s, int * x, int * y, int * comp) {
+    int maxv, dummy;
+    char c, p, t;
+
+    if (!x)
+        x = &dummy;
+    if (!y)
+        y = &dummy;
+    if (!comp)
+        comp = &dummy;
+
+    stbi__rewind(s);
+
+    // Get identifier
+    p = (char)stbi__get8(s);
+    t = (char)stbi__get8(s);
+    if (p != 'P' || (t != '5' && t != '6')) {
+        stbi__rewind(s);
+        return 0;
+    }
+
+    *comp = (t == '6') ? 3 : 1; // '5' is 1-component .pgm; '6' is 3-component .ppm
+
+    c = (char)stbi__get8(s);
+    stbi__pnm_skip_whitespace(s, &c);
+
+    *x = stbi__pnm_getinteger(s, &c); // read width
+    if (*x == 0)
+        return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+    stbi__pnm_skip_whitespace(s, &c);
+
+    *y = stbi__pnm_getinteger(s, &c); // read height
+    if (*y == 0)
+        return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+    stbi__pnm_skip_whitespace(s, &c);
+
+    maxv = stbi__pnm_getinteger(s, &c); // read max value
+    if (maxv > 65535)
+        return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images");
+    else if (maxv > 255)
+        return 16;
+    else
+        return 8;
+}
+
+static int stbi__pnm_is16(stbi__context * s) {
+    if (stbi__pnm_info(s, NULL, NULL, NULL) == 16)
+        return 1;
+    return 0;
+}
+#endif
+
+static int stbi__info_main(stbi__context * s, int * x, int * y, int * comp) {
+#ifndef STBI_NO_JPEG
+    if (stbi__jpeg_info(s, x, y, comp))
+        return 1;
+#endif
+
+#ifndef STBI_NO_PNG
+    if (stbi__png_info(s, x, y, comp))
+        return 1;
+#endif
+
+#ifndef STBI_NO_GIF
+    if (stbi__gif_info(s, x, y, comp))
+        return 1;
+#endif
+
+#ifndef STBI_NO_BMP
+    if (stbi__bmp_info(s, x, y, comp))
+        return 1;
+#endif
+
+#ifndef STBI_NO_PSD
+    if (stbi__psd_info(s, x, y, comp))
+        return 1;
+#endif
+
+#ifndef STBI_NO_PIC
+    if (stbi__pic_info(s, x, y, comp))
+        return 1;
+#endif
+
+#ifndef STBI_NO_PNM
+    if (stbi__pnm_info(s, x, y, comp))
+        return 1;
+#endif
+
+#ifndef STBI_NO_HDR
+    if (stbi__hdr_info(s, x, y, comp))
+        return 1;
+#endif
+
+// test tga last because it's a crappy test!
+#ifndef STBI_NO_TGA
+    if (stbi__tga_info(s, x, y, comp))
+        return 1;
+#endif
+    return stbi__err("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static int stbi__is_16_main(stbi__context * s) {
+#ifndef STBI_NO_PNG
+    if (stbi__png_is16(s))
+        return 1;
+#endif
+
+#ifndef STBI_NO_PSD
+    if (stbi__psd_is16(s))
+        return 1;
+#endif
+
+#ifndef STBI_NO_PNM
+    if (stbi__pnm_is16(s))
+        return 1;
+#endif
+    return 0;
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int stbi_info(char const * filename, int * x, int * y, int * comp) {
+    FILE * f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f)
+        return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_info_from_file(f, x, y, comp);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_info_from_file(FILE * f, int * x, int * y, int * comp) {
+    int r;
+    stbi__context s;
+    long pos = ftell(f);
+    stbi__start_file(&s, f);
+    r = stbi__info_main(&s, x, y, comp);
+    fseek(f, pos, SEEK_SET);
+    return r;
+}
+
+STBIDEF int stbi_is_16_bit(char const * filename) {
+    FILE * f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f)
+        return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_is_16_bit_from_file(f);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_is_16_bit_from_file(FILE * f) {
+    int r;
+    stbi__context s;
+    long pos = ftell(f);
+    stbi__start_file(&s, f);
+    r = stbi__is_16_main(&s);
+    fseek(f, pos, SEEK_SET);
+    return r;
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int stbi_info_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp) {
+    stbi__context s;
+    stbi__start_mem(&s, buffer, len);
+    return stbi__info_main(&s, x, y, comp);
+}
+
+STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const * c, void * user, int * x, int * y, int * comp) {
+    stbi__context s;
+    stbi__start_callbacks(&s, (stbi_io_callbacks *)c, user);
+    return stbi__info_main(&s, x, y, comp);
+}
+
+STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const * buffer, int len) {
+    stbi__context s;
+    stbi__start_mem(&s, buffer, len);
+    return stbi__is_16_main(&s);
+}
+
+STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * c, void * user) {
+    stbi__context s;
+    stbi__start_callbacks(&s, (stbi_io_callbacks *)c, user);
+    return stbi__is_16_main(&s);
+}
+
+#endif // STB_IMAGE_IMPLEMENTATION
+
+/*
+   revision history:
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) change sbti__shiftsigned to avoid clang -O2 bug
+                         1-bit BMP
+                         *_is_16_bit api
+                         avoid warnings
+      2.16  (2017-07-23) all functions have 16-bit variants;
+                         STBI_NO_STDIO works again;
+                         compilation fixes;
+                         fix rounding in unpremultiply;
+                         optimize vertical flip;
+                         disable raw_len validation;
+                         documentation fixes
+      2.15  (2017-03-18) fix png-1,2,4 bug; now all Imagenet JPGs decode;
+                         warning fixes; disable run-time SSE detection on gcc;
+                         uniform handling of optional "return" values;
+                         thread-safe initialization of zlib tables
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-11-29) add 16-bit API, only supported for PNG right now
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) allocate large structures on the stack
+                         remove white matting for transparent PSD
+                         fix reported channel count for PNG & BMP
+                         re-enable SSE2 in non-gcc 64-bit
+                         support RGB-formatted JPEG
+                         read 16-bit PNGs (only as 8-bit)
+      2.10  (2016-01-22) avoid warning introduced in 2.09 by STBI_REALLOC_SIZED
+      2.09  (2016-01-16) allow comments in PNM files
+                         16-bit-per-pixel TGA (not bit-per-component)
+                         info() for TGA could break due to .hdr handling
+                         info() for BMP to shares code instead of sloppy parse
+                         can use STBI_REALLOC_SIZED if allocator doesn't support realloc
+                         code cleanup
+      2.08  (2015-09-13) fix to 2.07 cleanup, reading RGB PSD as RGBA
+      2.07  (2015-09-13) fix compiler warnings
+                         partial animated GIF support
+                         limited 16-bpc PSD support
+                         #ifdef unused functions
+                         bug with < 92 byte PIC,PNM,HDR,TGA
+      2.06  (2015-04-19) fix bug where PSD returns wrong '*comp' value
+      2.05  (2015-04-19) fix bug in progressive JPEG handling, fix warning
+      2.04  (2015-04-15) try to re-enable SIMD on MinGW 64-bit
+      2.03  (2015-04-12) extra corruption checking (mmozeiko)
+                         stbi_set_flip_vertically_on_load (nguillemot)
+                         fix NEON support; fix mingw support
+      2.02  (2015-01-19) fix incorrect assert, fix warning
+      2.01  (2015-01-17) fix various warnings; suppress SIMD on gcc 32-bit without -msse2
+      2.00b (2014-12-25) fix STBI_MALLOC in progressive JPEG
+      2.00  (2014-12-25) optimize JPG, including x86 SSE2 & NEON SIMD (ryg)
+                         progressive JPEG (stb)
+                         PGM/PPM support (Ken Miller)
+                         STBI_MALLOC,STBI_REALLOC,STBI_FREE
+                         GIF bugfix -- seemingly never worked
+                         STBI_NO_*, STBI_ONLY_*
+      1.48  (2014-12-14) fix incorrectly-named assert()
+      1.47  (2014-12-14) 1/2/4-bit PNG support, both direct and paletted (Omar Cornut & stb)
+                         optimize PNG (ryg)
+                         fix bug in interlaced PNG with user-specified channel count (stb)
+      1.46  (2014-08-26)
+              fix broken tRNS chunk (colorkey-style transparency) in non-paletted PNG
+      1.45  (2014-08-16)
+              fix MSVC-ARM internal compiler error by wrapping malloc
+      1.44  (2014-08-07)
+              various warning fixes from Ronny Chevalier
+      1.43  (2014-07-15)
+              fix MSVC-only compiler problem in code changed in 1.42
+      1.42  (2014-07-09)
+              don't define _CRT_SECURE_NO_WARNINGS (affects user code)
+              fixes to stbi__cleanup_jpeg path
+              added STBI_ASSERT to avoid requiring assert.h
+      1.41  (2014-06-25)
+              fix search&replace from 1.36 that messed up comments/error messages
+      1.40  (2014-06-22)
+              fix gcc struct-initialization warning
+      1.39  (2014-06-15)
+              fix to TGA optimization when req_comp != number of components in TGA;
+              fix to GIF loading because BMP wasn't rewinding (whoops, no GIFs in my test suite)
+              add support for BMP version 5 (more ignored fields)
+      1.38  (2014-06-06)
+              suppress MSVC warnings on integer casts truncating values
+              fix accidental rename of 'skip' field of I/O
+      1.37  (2014-06-04)
+              remove duplicate typedef
+      1.36  (2014-06-03)
+              convert to header file single-file library
+              if de-iphone isn't set, load iphone images color-swapped instead of returning NULL
+      1.35  (2014-05-27)
+              various warnings
+              fix broken STBI_SIMD path
+              fix bug where stbi_load_from_file no longer left file pointer in correct place
+              fix broken non-easy path for 32-bit BMP (possibly never used)
+              TGA optimization by Arseny Kapoulkine
+      1.34  (unknown)
+              use STBI_NOTUSED in stbi__resample_row_generic(), fix one more leak in tga failure case
+      1.33  (2011-07-14)
+              make stbi_is_hdr work in STBI_NO_HDR (as specified), minor compiler-friendly improvements
+      1.32  (2011-07-13)
+              support for "info" function for all supported filetypes (SpartanJ)
+      1.31  (2011-06-20)
+              a few more leak fixes, bug in PNG handling (SpartanJ)
+      1.30  (2011-06-11)
+              added ability to load files via callbacks to accomidate custom input streams (Ben Wenger)
+              removed deprecated format-specific test/load functions
+              removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks
+   anyway error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha) fix inefficiency in
+   decoding 32-bit BMP (David Woo) 1.29  (2010-08-16) various warning fixes from Aurelien Pocheville 1.28  (2010-08-01)
+              fix bug in GIF palette transparency (SpartanJ)
+      1.27  (2010-08-01)
+              cast-to-stbi_uc to fix warnings
+      1.26  (2010-07-24)
+              fix bug in file buffering for PNG reported by SpartanJ
+      1.25  (2010-07-17)
+              refix trans_data warning (Won Chun)
+      1.24  (2010-07-12)
+              perf improvements reading from files on platforms with lock-heavy fgetc()
+              minor perf improvements for jpeg
+              deprecated type-specific functions so we'll get feedback if they're needed
+              attempt to fix trans_data warning (Won Chun)
+      1.23    fixed bug in iPhone support
+      1.22  (2010-07-10)
+              removed image *writing* support
+              stbi_info support from Jetro Lauha
+              GIF support from Jean-Marc Lienher
+              iPhone PNG-extensions from James Brown
+              warning-fixes from Nicolas Schulz and Janez Zemva (i.stbi__err. Janez (U+017D)emva)
+      1.21    fix use of 'stbi_uc' in header (reported by jon blow)
+      1.20    added support for Softimage PIC, by Tom Seddon
+      1.19    bug in interlaced PNG corruption check (found by ryg)
+      1.18  (2008-08-02)
+              fix a threading bug (local mutable static)
+      1.17    support interlaced PNG
+      1.16    major bugfix - stbi__convert_format converted one too many pixels
+      1.15    initialize some fields for thread safety
+      1.14    fix threadsafe conversion bug
+              header-file-only version (#define STBI_HEADER_FILE_ONLY before including)
+      1.13    threadsafe
+      1.12    const qualifiers in the API
+      1.11    Support installable IDCT, colorspace conversion routines
+      1.10    Fixes for 64-bit (don't use "unsigned long")
+              optimized upsampling by Fabian "ryg" Giesen
+      1.09    Fix format-conversion for PSD code (bad global variables!)
+      1.08    Thatcher Ulrich's PSD code integrated by Nicolas Schulz
+      1.07    attempt to fix C++ warning/errors again
+      1.06    attempt to fix C++ warning/errors again
+      1.05    fix TGA loading to return correct *comp and use good luminance calc
+      1.04    default float alpha is 1, not 255; use 'void *' for stbi_image_free
+      1.03    bugfixes to STBI_NO_STDIO, STBI_NO_HDR
+      1.02    support for (subset of) HDR files, float interface for preferred access to them
+      1.01    fix bug: possible bug in handling right-side up bmps... not sure
+              fix bug: the stbi__bmp_load() and stbi__tga_load() functions didn't work at all
+      1.00    interface to zlib that skips zlib header
+      0.99    correct handling of alpha in palette
+      0.98    TGA loader by lonesock; dynamically add loaders (untested)
+      0.97    jpeg errors on too large a file; also catch another malloc failure
+      0.96    fix detection of invalid v value - particleman@mollyrocket forum
+      0.95    during header scan, seek to markers in case of padding
+      0.94    STBI_NO_STDIO to disable stdio usage; rename all #defines the same
+      0.93    handle jpegtran output; verbose errors
+      0.92    read 4,8,16,24,32-bit BMP files of several formats
+      0.91    output 24-bit Windows 3.0 BMP files
+      0.90    fix a few more warnings; bump version number to approach 1.0
+      0.61    bugfixes due to Marc LeBlanc, Christopher Lloyd
+      0.60    fix compiling as c++
+      0.59    fix warnings: merge Dave Moore's -Wall fixes
+      0.58    fix bug: zlib uncompressed mode len/nlen was wrong endian
+      0.57    fix bug: jpg last huffman symbol before marker was >9 bits but less than 16 available
+      0.56    fix bug: zlib uncompressed mode len vs. nlen
+      0.55    fix bug: restart_interval not initialized to 0
+      0.54    allow NULL for 'int *comp'
+      0.53    fix bug in png 3->4; speedup png decoding
+      0.52    png handles req_comp=3,4 directly; minor cleanup; jpeg comments
+      0.51    obey req_comp requests, 1-component jpegs return as 1-component,
+              on 'test' only check type, not whether we support this variant
+      0.50  (2006-11-19)
+              first released version
+*/
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/

scripts/sync_llama.sh

@@ -44,6 +44,12 @@ 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/
 
+# clip
+cp $src_dir/examples/llava/clip.cpp $dst_dir/clip.cpp
+cp $src_dir/examples/llava/clip.h $dst_dir/clip.h
+cp $src_dir/common/log.h $dst_dir/log.h
+cp $src_dir/common/stb_image.h $dst_dir/stb_image.h
+
 # apply patches
 for patch in $dst_dir/patches/*.patch; do
   git apply "$patch"
@@ -77,7 +83,6 @@ for IN in $dst_dir/*.{c,h,cpp,m,metal,cu}; do
 done
 
 # ggml-metal
-sed -i '' '1s;^;// go:build darwin && arm64\n\n;' $dst_dir/ggml-metal.m
 sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < $dst_dir/ggml-metal.metal > temp.metal
 TEMP_ASSEMBLY=$(mktemp)
 echo ".section __DATA, __ggml_metallib"   >  $TEMP_ASSEMBLY