devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix cuda device warnings

Browse Source
This commit is contained in:
Antonio Sanchez
2025-02-28 15:21:23 -08:00
parent c53002f5fb
commit 23b1682723
53 changed files with 655 additions and 600 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
test/CMakeLists.txt
View File

@@ -362,14 +362,35 @@ if(EIGEN_TEST_CUDA)
find_package(CUDA 5.0)
if(CUDA_FOUND)
set(CUDA_PROPAGATE_HOST_FLAGS OFF)
if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
set(CUDA_NVCC_FLAGS "-ccbin ${CMAKE_C_COMPILER}" CACHE STRING "nvcc flags" FORCE)
if( (NOT EIGEN_TEST_CXX11) OR (CMAKE_VERSION VERSION_LESS 3.3))
string(APPEND EIGEN_CUDA_CXX11_FLAGS " -std=c++11")
endif()
if(EIGEN_TEST_CUDA_CLANG)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 --cuda-gpu-arch=sm_30")
string(APPEND CMAKE_CXX_FLAGS " --cuda-path=${CUDA_TOOLKIT_ROOT_DIR}")
foreach(GPU IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
string(APPEND CMAKE_CXX_FLAGS " --cuda-gpu-arch=sm_${GPU}")
endforeach()
string(APPEND CMAKE_CXX_FLAGS " ${EIGEN_CUDA_CXX_FLAGS}")
else()
set(CUDA_PROPAGATE_HOST_FLAGS OFF)
set(NVCC_ARCH_FLAGS)
# Define an -arch=sm_<arch>, otherwise if GPU does not exactly match one of
# those in the arch list for -gencode, the kernels will fail to run with
# cudaErrorNoKernelImageForDevice
# This can happen with newer cards (e.g. sm_75) and compiling with older
# versions of nvcc (e.g. 9.2) that do not support their specific arch.
list(LENGTH EIGEN_CUDA_COMPUTE_ARCH EIGEN_CUDA_COMPUTE_ARCH_SIZE)
if(EIGEN_CUDA_COMPUTE_ARCH_SIZE)
list(GET EIGEN_CUDA_COMPUTE_ARCH 0 EIGEN_CUDA_COMPUTE_DEFAULT)
set(NVCC_ARCH_FLAGS " -arch=sm_${EIGEN_CUDA_COMPUTE_DEFAULT}")
endif()
foreach(ARCH IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
string(APPEND NVCC_ARCH_FLAGS " -gencode arch=compute_${ARCH},code=sm_${ARCH}")
endforeach()
set(CUDA_NVCC_FLAGS "--expt-relaxed-constexpr -Xcudafe \"--display_error_number\" ${NVCC_ARCH_FLAGS} ${CUDA_NVCC_FLAGS} ${EIGEN_CUDA_CXX_FLAGS}")
cuda_include_directories("${CMAKE_CURRENT_BINARY_DIR}" "${CUDA_TOOLKIT_ROOT_DIR}/include")
endif()
cuda_include_directories(${CMAKE_CURRENT_BINARY_DIR})
set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu")
ei_add_test(cuda_basic)

14
test/cuda_common.h
View File

@@ -37,26 +37,26 @@ void run_on_cuda(const Kernel& ker, int n, const Input& in, Output& out)
typename Output::Scalar* d_out;
std::ptrdiff_t in_bytes = in.size() * sizeof(typename Input::Scalar);
std::ptrdiff_t out_bytes = out.size() * sizeof(typename Output::Scalar);
cudaMalloc((void**)(&d_in), in_bytes);
cudaMalloc((void**)(&d_out), out_bytes);
cudaMemcpy(d_in, in.data(), in_bytes, cudaMemcpyHostToDevice);
cudaMemcpy(d_out, out.data(), out_bytes, cudaMemcpyHostToDevice);
// Simple and non-optimal 1D mapping assuming n is not too large
// That's only for unit testing!
dim3 Blocks(128);
dim3 Grids( (n+int(Blocks.x)-1)/int(Blocks.x) );
cudaThreadSynchronize();
cudaDeviceSynchronize();
run_on_cuda_meta_kernel<<<Grids,Blocks>>>(ker, n, d_in, d_out);
cudaThreadSynchronize();
cudaDeviceSynchronize();
// check inputs have not been modified
cudaMemcpy(const_cast<typename Input::Scalar*>(in.data()), d_in, in_bytes, cudaMemcpyDeviceToHost);
cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost);
cudaFree(d_in);
cudaFree(d_out);
}