345ba07f1d
test cuda: import fmt in CUDA source code Current test is only for Windows(cl.exe). Need to test more with the other host compilers... * Activate the test when `find_package(CUDA)` worked * The test runs with C++14 Detailed comments in 'test/cuda-test' test cuda: add more comment / macro check * checks both `__NVCC__` and `__CUDACC__` More comments for CMake and CUDA source file. test cuda: <fmt/core.h> checks NVCC and CUDA The header file checks 2 things. * __NVCC__: if the compiler is from NVIDIA * __CUDACC__: if the source code is CUDA(.cu) file Since we can't sure all users prefer latest, Version for `find_pacakge(CUDA)` is downgraded to 9.0. This is the minimum version for C++14 in CUDA
56 lines
2.3 KiB
CMake
56 lines
2.3 KiB
CMake
# We can find some usecases which follow the guide of CMake which uses
|
|
# `enable_language(CUDA)` instead of `find_package(CUDA)` and let the CMake
|
|
# built-in functions use NVCC.
|
|
|
|
# See: https://cmake.org/cmake/help/latest/module/FindCUDA.html#replacement
|
|
#
|
|
# However, this requires CMake version 3.10 or higher and we can't be sure most
|
|
# of the CUDA projects are using those.
|
|
#
|
|
# This test relies on `find_package(CUDA)` in the parent CMake config.
|
|
|
|
# These can be updated when NVCC becomes ready for C++ 17 features
|
|
# https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#cpp14-language-features
|
|
set(CMAKE_CUDA_STANDARD 14)
|
|
set(CMAKE_CUDA_STANDARD_REQUIRED 14)
|
|
|
|
# https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html
|
|
list(APPEND CUDA_NVCC_FLAGS "-std=c++14")
|
|
if (MSVC)
|
|
# This is the solution of pytorch:
|
|
# https://github.com/pytorch/pytorch/pull/7118
|
|
list(APPEND CUDA_NVCC_FLAGS "-Xcompiler" "/std:c++14")
|
|
list(APPEND CUDA_NVCC_FLAGS "-Xcompiler" "/Zc:__cplusplus")
|
|
# for the reason of this -Xcompiler options, see below.
|
|
endif ()
|
|
|
|
# In this test, we assume that the user is going to compile CUDA source code
|
|
# with some libraries (fmt in this case).
|
|
#
|
|
# In addition to that, this test invokes both the C++ host compiler and NVCC
|
|
# by providing another (non-CUDA) C++ source code.
|
|
cuda_add_executable(fmt-in-cuda-test cuda-cpp14.cu cpp14.cc)
|
|
target_compile_features(fmt-in-cuda-test PRIVATE cxx_std_14)
|
|
|
|
get_target_property(IN_USE_CUDA_STANDARD fmt-in-cuda-test CUDA_STANDARD)
|
|
message(STATUS "cuda_standard: ${IN_USE_CUDA_STANDARD}")
|
|
|
|
get_target_property(IN_USE_CUDA_STANDARD_REQUIRED
|
|
fmt-in-cuda-test CUDA_STANDARD_REQUIRED)
|
|
message(STATUS "cuda_standard_required: ${IN_USE_CUDA_STANDARD_REQUIRED}")
|
|
|
|
# We don't use PUBLIC or other keyword for reasons explained in the
|
|
# CUDA_LINK_LIBRARIES_KEYWORD section in
|
|
# https://cmake.org/cmake/help/latest/module/FindCUDA.html
|
|
target_link_libraries(fmt-in-cuda-test fmt::fmt)
|
|
|
|
if (MSVC)
|
|
# This part is for (non-CUDA) C++ code. MSVC can define incorrect
|
|
# `__cplusplus` macro. Fix for the issue is to use additional compiler flag.
|
|
#
|
|
# See Also:
|
|
# https://devblogs.microsoft.com/cppblog/msvc-now-correctly-reports-__cplusplus/
|
|
# https://github.com/Microsoft/vscode-cpptools/issues/2595
|
|
target_compile_options(fmt-in-cuda-test PRIVATE /Zc:__cplusplus /permissive-)
|
|
endif ()
|