From 9268dc9d887a3d54cd1f008dcb628aaa5b5bd90a Mon Sep 17 00:00:00 2001 From: Tim Dettmers Date: Mon, 25 Jul 2022 19:30:37 -0700 Subject: Some progress on build script; added multi-cuda install script. --- Makefile | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) (limited to 'Makefile') diff --git a/Makefile b/Makefile index 2e1d265..328faa5 100644 --- a/Makefile +++ b/Makefile @@ -27,13 +27,14 @@ 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_75,code=sm_75 # 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_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 @@ -43,12 +44,12 @@ 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) env - $(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) + $(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 $(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) cuda92: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env - $(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' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT $(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) -- cgit v1.2.3