Compress fatbin to fit into 32bit indexing (#43074)

zasdfgbnm · facebook-github-bot · commit ee74c2e5bec8 · 2020-08-18T09:48:54.000-07:00
Summary: Fixes #39968 tested with `TORCH_CUDA_ARCH_LIST='3.5 5.2 6.0 6.1 7.0 7.5 8.0+PTX'`, before this PR, it was failing, and with this PR, the build succeed. With `TORCH_CUDA_ARCH_LIST='7.0 7.5 8.0+PTX'`, `libtorch_cuda.so` with symbols changes from 2.9GB -> 2.2GB cc: ptrblck mcarilli jjsjann123 Pull Request resolved: #43074 Reviewed By: mrshenli Differential Revision: D23176095 Pulled By: malfet fbshipit-source-id: 7b3e6d049fc080e519f21e80df05ef68e7bea57e
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -357,6 +357,10 @@ if(MSVC)
   list(APPEND CUDA_NVCC_FLAGS "-Xcompiler /w -w")
 endif(MSVC)
 
+list(APPEND CUDA_NVCC_FLAGS "-Xfatbin" "-compress-all")
+list(APPEND CUDA_NVCC_FLAGS_DEBUG "-Xfatbin" "-compress-all")
+list(APPEND CUDA_NVCC_FLAGS_RELWITHDEBINFO "-Xfatbin" "-compress-all")
+
 if(NOT MSVC)
   list(APPEND CUDA_NVCC_FLAGS_DEBUG "-g" "-lineinfo" "--source-in-ptx")
   list(APPEND CUDA_NVCC_FLAGS_RELWITHDEBINFO "-g" "-lineinfo" "--source-in-ptx")
