cutlass-matmul.h

parasailteam · Nov 10, 2022 · 9930431 · 9930431
1 parent d3a7d22
commit 9930431
Show file tree

Hide file tree

Showing 2 changed files with 635 additions and 1 deletion.
diff --git a/examples/model-parallel/Makefile b/examples/model-parallel/Makefile
@@ -21,7 +21,7 @@ model-parallel-mm-rs-c-ag: model-parallel-mm-rs-c-ag.cu
 	nvcc $< -std=c++11 -Xcompiler -fopenmp -lcudadevrt -lcudart -I/usr/include/x86_64-linux-gnu/openmpi -I.. -I/usr/local/cuda/include/ -I../../ -I$(NCCL_BUILD_PATH)/include -L$(NCCL_BUILD_PATH)/lib/ -L/usr/local/cuda/lib64/ $(GENCODE) -lnccl -lcublas -lcurand -c -o $@.o && $(MPI_CXX) $@.o -L$(NCCL_BUILD_PATH)/lib -L/usr/local/cuda/lib64/ -lcudart -lnccl -lcublas -Wall -lcurand -lcudadevrt -std=c++11 -fopenmp -o $@
 
 model-parallel-ol-mm-fuse-rs-c-ag: model-parallel-ol-mm-fuse-rs-c-ag.cu
-	nvcc $< -std=c++11 -Xcompiler -fopenmp -lcudadevrt -lcudart -I/usr/include/x86_64-linux-gnu/openmpi -I.. -I/usr/local/cuda/include/ -I../../ -I$(NCCL_BUILD_PATH)/include -L$(NCCL_BUILD_PATH)/lib/ -L/usr/local/cuda/lib64/ -lnccl -lcublas -lcurand -c -o $@.o && $(MPI_CXX) $@.o -L$(NCCL_BUILD_PATH)/lib -L/usr/local/cuda/lib64/ -lcudart -lnccl -lcublas -Wall -lcurand -lcudadevrt -std=c++11 -fopenmp -o $@
+	nvcc $< -std=c++11 -Xcompiler -fopenmp -lcudadevrt -lcudart -I ../../cutlass/include -I ../../cutlass/tools/util/include -I ../../cutlass/examples/common -I/usr/include/x86_64-linux-gnu/openmpi -I.. -I/usr/local/cuda/include/ -I../../ -I$(NCCL_BUILD_PATH)/include -L$(NCCL_BUILD_PATH)/lib/ -L/usr/local/cuda/lib64/ -lnccl -lcublas -lcurand -c -o $@.o && $(MPI_CXX) $@.o -L$(NCCL_BUILD_PATH)/lib -L/usr/local/cuda/lib64/ -lcudart -lnccl -lcublas -Wall -lcurand -lcudadevrt -std=c++11 -fopenmp -o $@
 
 clean:
 	rm -rf model-parallel model-parallel-ar-c.cu model-parallel-ar-c model-parallel-rs-c-ag.cu model-parallel-rs-c-ag model-parallel-ol-mm-fuse-rs-c-ag model-parallel-ol-mm-fuse-rs-c-ag.cu