Merge branch 'patch_merge' into extract_outliers

1 files changed, 27 insertions, 10 deletions

diff --git a/Makefile b/Makefile
index b58e233..10f267a 100644
--- a/Makefile
+++ b/Makefile
@@ -16,7 +16,7 @@ FILES_CUDA := $(CSRC)/ops.cu $(CSRC)/kernels.cu
 FILES_CPP := $(CSRC)/common.cpp $(CSRC)/cpu_ops.cpp $(CSRC)/pythonInterface.c
 
 INCLUDE :=  -I $(CUDA_HOME)/include -I $(ROOT_DIR)/csrc -I $(CONDA_PREFIX)/include -I $(ROOT_DIR)/dependencies/cub -I $(ROOT_DIR)/include
-LIB := -L $(CUDA_HOME)/lib64 -lcudart -lcuda -lcublas -lcurand -lcusparse -L $(CONDA_PREFIX)/lib
+LIB := -L $(CUDA_HOME)/lib64 -lcudart -lcublas -lcublasLt -lcurand -lcusparse -L $(CONDA_PREFIX)/lib
 
 # NVIDIA NVCC compilation flags
 COMPUTE_CAPABILITY := -gencode arch=compute_35,code=sm_35 # Kepler 
@@ -27,7 +27,6 @@ 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
@@ -43,31 +42,49 @@ 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
 
+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
-	$(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)
 
-cuda10x: $(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)
+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) -D NO_CUBLASLT
 	$(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)
 
-cuda110: $(BUILD_DIR) env
-	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
+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) -D NO_CUBLASLT
 	$(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)
 
-cuda11x: $(BUILD_DIR) env
-	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
+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) -D NO_CUBLASLT
 	$(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)
 
+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
 	$(GPP) -std=c++14 -shared -fPIC -I $(ROOT_DIR)/csrc -I $(ROOT_DIR)/include $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so