|
@@ -58,6 +58,8 @@ endif
|
|
|
GPU_COMPILER_CUFLAGS = \
|
|
|
$(GPU_COMPILER_FPIC) \
|
|
|
$(addprefix -m,$(GPU_RUNNER_CPU_FLAGS)) \
|
|
|
+ -mf16c \
|
|
|
+ -mfma \
|
|
|
-parallel-jobs=2 \
|
|
|
-c \
|
|
|
-O3 \
|
|
@@ -77,6 +79,9 @@ GPU_COMPILER_CUFLAGS = \
|
|
|
-D_CRT_SECURE_NO_WARNINGS \
|
|
|
-D_GNU_SOURCE \
|
|
|
-D_XOPEN_SOURCE=600 \
|
|
|
+ -DUSE_PROF_API=1 \
|
|
|
+ -std=gnu++14 \
|
|
|
+ -x hip \
|
|
|
-mllvm=-amdgpu-early-inline-all=true \
|
|
|
-mllvm=-amdgpu-function-calls=false \
|
|
|
-Wno-expansion-to-defined \
|