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 /Tests/Cuda/CMakeLists.txt | |
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 'Tests/Cuda/CMakeLists.txt')
-rw-r--r-- | Tests/Cuda/CMakeLists.txt | 7 |
1 files changed, 3 insertions, 4 deletions
diff --git a/Tests/Cuda/CMakeLists.txt b/Tests/Cuda/CMakeLists.txt index 35ceb33..be5ccac 100644 --- a/Tests/Cuda/CMakeLists.txt +++ b/Tests/Cuda/CMakeLists.txt @@ -17,13 +17,12 @@ add_cuda_test_macro(Cuda.SeparableCompCXXOnly SeparableCompCXXOnly) add_cuda_test_macro(Cuda.Toolkit Toolkit) add_cuda_test_macro(Cuda.IncludePathNoToolkit IncludePathNoToolkit) add_cuda_test_macro(Cuda.SharedRuntimePlusToolkit SharedRuntimePlusToolkit) +add_cuda_test_macro(Cuda.Complex CudaComplex) +add_cuda_test_macro(Cuda.ProperLinkFlags ProperLinkFlags) -# Separable compilation is currently only supported on NVCC. Disable tests -# using it for other compilers. if(CMake_TEST_CUDA AND NOT CMake_TEST_CUDA STREQUAL "Clang") - add_cuda_test_macro(Cuda.Complex CudaComplex) + # Clang lacks __CUDACC_VER*__ defines. add_cuda_test_macro(Cuda.ProperDeviceLibraries ProperDeviceLibraries) - add_cuda_test_macro(Cuda.ProperLinkFlags ProperLinkFlags) endif() # The CUDA only ships the shared version of the toolkit libraries |