mrq/bitsandbytes-rocm
1
0
Fork 1
Code Issues 1 Pull Requests Packages Projects Releases Wiki Activity

Added matmul build and flags.

Browse Source
This commit is contained in:
Tim Dettmers 2022-07-25 22:34:14 -07:00
parent 9268dc9d88
commit f2dd703251
2 changed files with 171 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
Makefile
View File

@ -27,14 +27,13 @@ COMPUTE_CAPABILITY += -gencode arch=compute_60,code=sm_60 # Pascal
COMPUTE_CAPABILITY += -gencode arch=compute_61,code=sm_61 # Pascal COMPUTE_CAPABILITY += -gencode arch=compute_61,code=sm_61 # Pascal
COMPUTE_CAPABILITY += -gencode arch=compute_70,code=sm_70 # Volta COMPUTE_CAPABILITY += -gencode arch=compute_70,code=sm_70 # Volta
COMPUTE_CAPABILITY += -gencode arch=compute_72,code=sm_72 # Volta COMPUTE_CAPABILITY += -gencode arch=compute_72,code=sm_72 # Volta
COMPUTE_CAPABILITY := -gencode arch=compute_75,code=sm_75 # Volta
# CUDA 9.2 supports CC 3.0, but CUDA >= 11.0 does not # CUDA 9.2 supports CC 3.0, but CUDA >= 11.0 does not
CC_CUDA92 := -gencode arch=compute_30,code=sm_30 CC_CUDA92 := -gencode arch=compute_30,code=sm_30
# Later versions of CUDA support the new architectures # Later versions of CUDA support the new architectures
CC_CUDA10x := -gencode arch=compute_30,code=sm_30 CC_CUDA10x := -gencode arch=compute_30,code=sm_30
CC_CUDA10x := -gencode arch=compute_75,code=sm_75 CC_CUDA10x += -gencode arch=compute_75,code=sm_75
CC_CUDA110 := -gencode arch=compute_75,code=sm_75 CC_CUDA110 := -gencode arch=compute_75,code=sm_75
CC_CUDA110 += -gencode arch=compute_80,code=sm_80 CC_CUDA110 += -gencode arch=compute_80,code=sm_80
@ -43,6 +42,14 @@ CC_CUDA11x := -gencode arch=compute_75,code=sm_75
CC_CUDA11x += -gencode arch=compute_80,code=sm_80 CC_CUDA11x += -gencode arch=compute_80,code=sm_80
CC_CUDA11x += -gencode arch=compute_86,code=sm_86 CC_CUDA11x += -gencode arch=compute_86,code=sm_86
CC_cublasLt110 := -gencode arch=compute_75,code=sm_75
CC_cublasLt110 += -gencode arch=compute_80,code=sm_80
CC_cublasLt111 := -gencode arch=compute_75,code=sm_75
CC_cublasLt111 += -gencode arch=compute_80,code=sm_80
CC_cublasLt111 += -gencode arch=compute_86,code=sm_86
all: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env all: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env
$(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
$(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
@ -53,21 +60,31 @@ cuda92: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env
$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB) $(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
cuda10x: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env cuda10x_nomatmul: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env
$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB) $(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
cuda110: $(BUILD_DIR) env cuda110_nomatmul: $(BUILD_DIR) env
$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB) $(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
cuda11x: $(BUILD_DIR) env cuda11x_nomatmul: $(BUILD_DIR) env
$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB) $(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
cuda110: $(BUILD_DIR) env
$(NVCC) $(CC_cublasLt110) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
$(NVCC) $(CC_cublasLt110) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
cuda11x: $(BUILD_DIR) env
$(NVCC) $(CC_cublasLt111) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
$(NVCC) $(CC_cublasLt111) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
cpuonly: $(BUILD_DIR) env cpuonly: $(BUILD_DIR) env
$(GPP) -std=c++14 -shared -fPIC -I $(ROOT_DIR)/csrc -I $(ROOT_DIR)/include $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(GPP) -std=c++14 -shared -fPIC -I $(ROOT_DIR)/csrc -I $(ROOT_DIR)/include $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so

163
deploy_from_slurm.sh
View File

@ -4,88 +4,217 @@ BASE_PATH=$1
module unload cuda module unload cuda
module unload gcc module unload gcc
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.0
#make cuda110
#
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=110 python -m build
##python -m twine upload dist/* --verbose
#
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.1
#make cuda11x
#
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=111 python -m build
##python -m twine upload dist/* --verbose
#
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.2
#make cuda11x
#
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=112 python -m build
##python -m twine upload dist/* --verbose
#
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.3
#make cuda11x
#
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=113 python -m build
##python -m twine upload dist/* --verbose
#
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.4
#make cuda11x
#
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=114 python -m build
###python -m twine upload dist/* --verbose
#
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.5
#make cuda11x
#
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=115 python -m build
##python -m twine upload dist/* --verbose
#
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.6
#
#make cuda11x
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=116 python -m build
##python -m twine upload dist/* --verbose
#
#rm -rf dist build
#make clean
#make cleaneggs
#export CUDA_HOME=$BASE_PATH/cuda-11.7
#make cuda11x
#
#if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# # Control will enter here if $DIRECTORY doesn't exist.
# echo "Compilation unsuccessul!" 1>&2
# exit 64
#fi
##CUDA_VERSION=117 python -m build
##python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
make clean make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.0 export CUDA_HOME=$BASE_PATH/cuda-10.2
make cuda110 make cuda10x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=110 python -m build #CUDA_VERSION=102-nomatmul python -m build
#python -m twine upload dist/* --verbose #python -m twine upload dist/* --verbose
rm -rf dist build
make clean
make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.0
make cuda110_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2
exit 64
fi
#CUDA_VERSION=110-nomatmul python -m build
#python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
make clean make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.1 export CUDA_HOME=$BASE_PATH/cuda-11.1
make cuda11x make cuda11x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=111 python -m build #CUDA_VERSION=111-nomatmul python -m build
#python -m twine upload dist/* --verbose #python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
make clean make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.2 export CUDA_HOME=$BASE_PATH/cuda-11.2
make cuda11x make cuda11x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=112 python -m build #CUDA_VERSION=112-nomatmul python -m build
#python -m twine upload dist/* --verbose #python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
make clean make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.3 export CUDA_HOME=$BASE_PATH/cuda-11.3
make cuda11x make cuda11x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=113 python -m build #CUDA_VERSION=113-nomatmul python -m build
#python -m twine upload dist/* --verbose #python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
make clean make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.4 export CUDA_HOME=$BASE_PATH/cuda-11.4
make cuda11x make cuda11x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=114 python -m build #CUDA_VERSION=114-nomatmul python -m build
##python -m twine upload dist/* --verbose ##python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
make clean make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.5 export CUDA_HOME=$BASE_PATH/cuda-11.5
make cuda11x make cuda11x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=115 python -m build #CUDA_VERSION=115-nomatmul python -m build
#python -m twine upload dist/* --verbose #python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
@ -93,25 +222,25 @@ make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.6 export CUDA_HOME=$BASE_PATH/cuda-11.6
make cuda11x make cuda11x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=116 python -m build #CUDA_VERSION=116-nomatmul python -m build
#python -m twine upload dist/* --verbose #python -m twine upload dist/* --verbose
rm -rf dist build rm -rf dist build
make clean make clean
make cleaneggs make cleaneggs
export CUDA_HOME=$BASE_PATH/cuda-11.7 export CUDA_HOME=$BASE_PATH/cuda-11.7
make cuda11x make cuda11x_nomatmul
if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then if [ ! -f "./bitsandbytes/libbitsandbytes.so" ]; then
# Control will enter here if $DIRECTORY doesn't exist. # Control will enter here if $DIRECTORY doesn't exist.
echo "Compilation unsuccessul!" 1>&2 echo "Compilation unsuccessul!" 1>&2
exit 64 exit 64
fi fi
#CUDA_VERSION=117 python -m build #CUDA_VERSION=117-nomatmul python -m build
#python -m twine upload dist/* --verbose #python -m twine upload dist/* --verbose