From 8400b58cbbc06e0a434cfa71f76c2efd713473fc Mon Sep 17 00:00:00 2001 From: Tim Dettmers Date: Sun, 17 Oct 2021 21:21:39 -0700 Subject: Added Kepler and fixed V100+CUDA101 support. #4 #5 --- Makefile | 37 ++++++++++++++++++++++++++++--------- 1 file changed, 28 insertions(+), 9 deletions(-) (limited to 'Makefile') diff --git a/Makefile b/Makefile index 753ee3d..1a5f17f 100644 --- a/Makefile +++ b/Makefile @@ -15,11 +15,30 @@ INCLUDE := -I $(CUDA_HOME)/include -I $(ROOT_DIR)/csrc -I $(CONDA_PREFIX)/inclu LIB := -L $(CUDA_HOME)/lib64 -lcudart -lcuda -lcublas -lcurand -lcusparse -L $(CONDA_PREFIX)/lib # NVIDIA NVCC compilation flags -COMPUTE_CAPABILITY := -gencode arch=compute_50,code=sm_50 # Maxwell +COMPUTE_CAPABILITY := -gencode arch=compute_35,code=sm_35 # Kepler +COMPUTE_CAPABILITY += -gencode arch=compute_37,code=sm_37 # Kepler +COMPUTE_CAPABILITY += -gencode arch=compute_50,code=sm_50 # Maxwell COMPUTE_CAPABILITY += -gencode arch=compute_52,code=sm_52 # Maxwell +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_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 + +# CUDA 9.2 supports CC 3.0, but CUDA >= 11.0 does not +CC_CUDA92 := -gencode arch=compute_30,code=sm_30 + +# Later versions of CUDA support the new architectures +CC_CUDA10x := -gencode arch=compute_30,code=sm_30 +CC_CUDA10x += -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_CUDA11x := -gencode arch=compute_75,code=sm_75 +CC_CUDA11x += -gencode arch=compute_80,code=sm_80 +CC_CUDA11x += -gencode arch=compute_86,code=sm_86 + all: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) @@ -27,23 +46,23 @@ all: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) $(GPP) -std=c++11 -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB) cuda92: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) - $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) - $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o + $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) + $(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(GPP) -std=c++11 -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) - $(NVCC) $(COMPUTE_CAPABILITY) -gencode arch=compute_75,code=sm_75 -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) - $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o + $(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' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(GPP) -std=c++11 -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) - $(NVCC) $(COMPUTE_CAPABILITY) -gencode arch=compute_80,code=sm_80 -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) - $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o + $(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' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(GPP) -std=c++11 -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) - $(NVCC) $(COMPUTE_CAPABILITY) -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) - $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o + $(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' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o $(GPP) -std=c++11 -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB) $(BUILD_DIR): -- cgit v1.2.3