diff --git a/script/cmake-cuda.sh b/script/cmake-cuda.sh index 0110075c0d..0e7d00c469 100755 --- a/script/cmake-cuda.sh +++ b/script/cmake-cuda.sh @@ -15,11 +15,12 @@ cmake -D DEVICE_BACKEND=CUDA \ -D BOOST_ROOT="/package/install/boost_1.67.0" \ -D CUDA_COMMON_INCLUDE_DIR="/home/chao/code/test_feature/cuda_common/cuda_10.0_common/inc" \ --D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61" \ +-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -Xptxas -v -maxrregcount=128" \ ${MY_PROJECT_SOURCE} #-D CMAKE_CUDA_COMPILER="/package/install/cuda_10.0/bin/nvcc" \ -#-D CMAKE_CUDA_FLAGS="-G -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61" \ -#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61" \ -#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61 -Xptxas -v -maxrregcount=128" \ +#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61" \ +#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -Xptxas -v -maxrregcount=128" \ +#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -gencode=arch=compute_70,code=sm_70" \ +#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -gencode=arch=compute_70,code=sm_70 -Xptxas -v -maxrregcount=128" \ diff --git a/script/extract_asm-cuda.sh b/script/extract_asm-cuda.sh new file mode 100755 index 0000000000..84a0af76dd --- /dev/null +++ b/script/extract_asm-cuda.sh @@ -0,0 +1 @@ +cuobjdump -xelf all ./driver/driver && nvdisasm --print-code -g driver.sm_61.cubin > driver.sm_61.asm && nvdisasm --print-code -g driver.sm_70.cubin > driver.sm_70.asm