diff options
| author | Raul Tambre <raul@tambre.ee> | 2020-09-05 16:40:02 (GMT) |
|---|---|---|
| committer | Brad King <brad.king@kitware.com> | 2020-09-24 19:19:54 (GMT) |
| commit | c63fe018353cf6afb30980c4cac7493be7cd0a82 (patch) | |
| tree | 68d2daf0cd8ab91a9feaa49392607c6cfecd2ac4 /Modules/CMakeCUDAInformation.cmake | |
| parent | c98ec731f90eb0180c89108b7d2e42263b66d1ed (diff) | |
| download | CMake-c63fe018353cf6afb30980c4cac7493be7cd0a82.zip CMake-c63fe018353cf6afb30980c4cac7493be7cd0a82.tar.gz CMake-c63fe018353cf6afb30980c4cac7493be7cd0a82.tar.bz2 | |
CUDA: Clang separable compilation
For NVCC the compiler takes care of device linking when passed the "-dlink"
flag.
Clang doesn't support such magic and requires the buildsystem to do the work
that NVCC does behind the scenes.
The implementation is based on Bazel's device linking documentation:
https://github.com/tensorflow/tensorflow/blob/7cabcdf073abad8c46e9dda62bb8fa4682d2061e/third_party/nccl/build_defs.bzl.tpl#L259
Closes: #20726
Diffstat (limited to 'Modules/CMakeCUDAInformation.cmake')
| -rw-r--r-- | Modules/CMakeCUDAInformation.cmake | 7 |
1 files changed, 6 insertions, 1 deletions
diff --git a/Modules/CMakeCUDAInformation.cmake b/Modules/CMakeCUDAInformation.cmake index f9f7574..58e6e29 100644 --- a/Modules/CMakeCUDAInformation.cmake +++ b/Modules/CMakeCUDAInformation.cmake @@ -145,7 +145,7 @@ endif() #Specify how to compile when separable compilation has been requested if(NOT CMAKE_CUDA_COMPILE_SEPARABLE_COMPILATION) set(CMAKE_CUDA_COMPILE_SEPARABLE_COMPILATION - "<CMAKE_CUDA_COMPILER> ${_CMAKE_CUDA_EXTRA_FLAGS} <DEFINES> <INCLUDES> <FLAGS> ${_CMAKE_COMPILE_AS_CUDA_FLAG} -dc <SOURCE> -o <OBJECT>") + "<CMAKE_CUDA_COMPILER> ${_CMAKE_CUDA_EXTRA_FLAGS} <DEFINES> <INCLUDES> <FLAGS> ${_CMAKE_COMPILE_AS_CUDA_FLAG} ${_CMAKE_CUDA_DEVICE_CODE} <SOURCE> -o <OBJECT>") endif() #Specify how to compile when whole compilation has been requested @@ -200,6 +200,11 @@ if(NOT CMAKE_CUDA_DEVICE_LINK_EXECUTABLE) "<CMAKE_CUDA_COMPILER> ${_CMAKE_CUDA_EXTRA_FLAGS} <LANGUAGE_COMPILE_FLAGS> <LINK_FLAGS> ${CMAKE_CUDA_COMPILE_OPTIONS_PIC} ${_CMAKE_CUDA_EXTRA_DEVICE_LINK_FLAGS} -shared -dlink <OBJECTS> -o <TARGET> <LINK_LIBRARIES>${__IMPLICT_DLINK_FLAGS}") endif() +# Used when device linking is handled by CMake. +if(NOT CMAKE_CUDA_DEVICE_LINK_COMPILE) + set(CMAKE_CUDA_DEVICE_LINK_COMPILE "<CMAKE_CUDA_COMPILER> ${_CMAKE_CUDA_EXTRA_FLAGS} <FLAGS> -D__CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__ -D__NV_EXTRA_INITIALIZATION=\"\" -D__NV_EXTRA_FINALIZATION=\"\" -DREGISTERLINKBINARYFILE=\\\"<REGISTER_FILE>\\\" -DFATBINFILE=\\\"<FATBINARY>\\\" ${_CMAKE_COMPILE_AS_CUDA_FLAG} -c \"${CMAKE_CUDA_COMPILER_TOOLKIT_LIBRARY_ROOT}/bin/crt/link.stub\" -o <OBJECT>") +endif() + unset(__IMPLICT_DLINK_FLAGS) set(CMAKE_CUDA_INFORMATION_LOADED 1) |