diff options
| author | Brad King <brad.king@kitware.com> | 2022-03-04 13:51:49 -0500 |
|---|---|---|
| committer | Brad King <brad.king@kitware.com> | 2022-03-10 09:27:29 -0500 |
| commit | d1b48bfabd6157309b3056967e6e30cc0ce07983 (patch) | |
| tree | 3ce109b078bf5da4d866f6e2994bbce8246e4df1 /Modules/CMakeDetermineCUDACompiler.cmake | |
| parent | 632752d62e8de2730796f509dbb10551351309c2 (diff) | |
| download | cmake-d1b48bfabd6157309b3056967e6e30cc0ce07983.tar.gz | |
CUDA: Add support for CUDA_ARCHITECTURES=native
CUDA 11.6 added the `nvcc -arch=native` flag to automatically compile
for the host GPUs' architectures. Add support for specifying this
special `native` value in `CMAKE_CUDA_ARCHITECTURES` and
`CUDA_ARCHITECTURES`. During the compiler ABI detection step,
detect the native architectures so we can pass them explicitly
when using Clang or older versions of nvcc.
Fixes: #22375
Diffstat (limited to 'Modules/CMakeDetermineCUDACompiler.cmake')
| -rw-r--r-- | Modules/CMakeDetermineCUDACompiler.cmake | 15 |
1 files changed, 13 insertions, 2 deletions
diff --git a/Modules/CMakeDetermineCUDACompiler.cmake b/Modules/CMakeDetermineCUDACompiler.cmake index 4263bed392..f68c4b235f 100644 --- a/Modules/CMakeDetermineCUDACompiler.cmake +++ b/Modules/CMakeDetermineCUDACompiler.cmake @@ -249,7 +249,7 @@ if(NOT CMAKE_CUDA_COMPILER_ID_RUN) set(CMAKE_CUDA_COMPILER_TOOLKIT_VERSION "${CMAKE_MATCH_1}") endif() - # Make the all and all-major architecture information available. + # Make the all, all-major, and native architecture information available. # FIXME(#23161): Defer architecture detection until compiler testing. include(${CMAKE_ROOT}/Modules/CUDA/architectures.cmake) endif() @@ -291,6 +291,17 @@ if(NOT CMAKE_CUDA_COMPILER_ID_RUN) set(architectures_test ${CMAKE_CUDA_ARCHITECTURES_ALL_MAJOR}) endif() endif() + elseif(CMAKE_CUDA_ARCHITECTURES STREQUAL "native") + # For sufficiently new NVCC we can just use the 'native' value directly. + # For VS we don't test since we can't find nvcc this early (see #23161). + if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA" AND CMAKE_CUDA_COMPILER_TOOLKIT_VERSION VERSION_GREATER_EQUAL 11.6) + string(APPEND nvcc_test_flags " -arch=${CMAKE_CUDA_ARCHITECTURES}") + set(architectures_tested "${CMAKE_CUDA_ARCHITECTURES}") + elseif(CMAKE_GENERATOR MATCHES "Visual Studio") + set(architectures_tested "${CMAKE_CUDA_ARCHITECTURES}") + else() + set(architectures_test ${_CUDA_ARCHITECTURES_NATIVE}) + endif() elseif(CMAKE_CUDA_ARCHITECTURES OR "${CMAKE_CUDA_ARCHITECTURES}" STREQUAL "") # Explicit architectures. Test them during detection. set(architectures_explicit TRUE) @@ -636,7 +647,7 @@ if("${CMAKE_CUDA_ARCHITECTURES}" STREQUAL "") message(FATAL_ERROR "Failed to detect a default CUDA architecture.\n\nCompiler output:\n${CMAKE_CUDA_COMPILER_PRODUCED_OUTPUT}") endif() endif() -elseif(CMAKE_CUDA_ARCHITECTURES AND NOT "${architectures_tested}" MATCHES "^(all|all-major)$") +elseif(CMAKE_CUDA_ARCHITECTURES AND NOT "${architectures_tested}" MATCHES "^(all|all-major|native)$") # Sort since order mustn't matter. list(SORT architectures_detected) list(SORT architectures_tested) |