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/Compiler/Clang-CUDA.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/Compiler/Clang-CUDA.cmake')
| -rw-r--r-- | Modules/Compiler/Clang-CUDA.cmake | 1 |
1 files changed, 1 insertions, 0 deletions
diff --git a/Modules/Compiler/Clang-CUDA.cmake b/Modules/Compiler/Clang-CUDA.cmake index 336827b..fd8c2b7 100644 --- a/Modules/Compiler/Clang-CUDA.cmake +++ b/Modules/Compiler/Clang-CUDA.cmake @@ -13,6 +13,7 @@ __compiler_clang_cxx_standards(CUDA) set(CMAKE_CUDA_COMPILER_HAS_DEVICE_LINK_PHASE TRUE) set(_CMAKE_COMPILE_AS_CUDA_FLAG "-x cuda") set(_CMAKE_CUDA_PTX_FLAG "--cuda-device-only -S") +set(_CMAKE_CUDA_DEVICE_CODE "-fgpu-rdc -c") # RulePlaceholderExpander expands crosscompile variables like sysroot and target only for CMAKE_<LANG>_COMPILER. Override the default. set(CMAKE_CUDA_LINK_EXECUTABLE "<CMAKE_CUDA_COMPILER> <LINK_FLAGS> <OBJECTS> -o <TARGET> <LINK_LIBRARIES>${__IMPLICT_LINKS}") |