我想编译CUDALibrarySamples。 cuFFT 使用 cmake,我想用静态版本的 cufft lib (
1d_c2c
) 编译和链接-lcufft_static
应用程序。使用 Makefiles 很简单我在中添加了
-lcufft_static
nvcc -x cu $(FLAGS) $(INC) 1d_c2c_example.cpp -o 1d_c2c_example $(LIBS)
但是,我不确定如何使用 cmake 做同样的事情。我注意到 cmake 有静态标志:
CUDA_cublasLt_static_LIBRARY
、CUDA_cufft_static_LIBRARY
等。所以我的问题是如何启用它们?
提前谢谢你!
我试过了
target_link_libraries(${ROUTINE}_example PRIVATE ${CUDA_cufft_static_LIBRARY})
但是好像不行。
根据@paleonix 的建议,我做了以下事情:
target_link_libraries(${ROUTINE}_example PRIVATE CUDA::cufft_static CUDA::cudart).
但是我收到以下错误:
/usr/bin/ld: /opt/cuda/lib64/libcufft_static.a(cbdouble_32bit_prime_callback_RT_SM35_plus.o): in function __sti____cudaRegisterAll()': cbdouble_32bit_prime_callback_RT_SM35_plus.compute_86.cudafe1.cpp:(.text.startup+0x1d): undefined reference to __cudaRegisterLinkedBinary_61_cbdouble_32bit_prime_callback_RT_SM35_plus_compute_86_cpp1_ii_dc5d5345
我正在尝试为 CUDA 库示例构建以下示例:
#include <complex>
#include <iostream>
#include <random>
#include <vector>
#include <cuda_runtime.h>
#include <cufftXt.h>
#include "cufft_utils.h"
int main(int argc, char *argv[]) {
cufftHandle plan;
cudaStream_t stream = NULL;
int n = 8;
int batch_size = 2;
int fft_size = batch_size * n;
using scalar_type = float;
using data_type = std::complex<scalar_type>;
std::vector<data_type> data(fft_size);
for (int i = 0; i < fft_size; i++) {
data[i] = data_type(i, -i);
}
std::printf("Input array:\n");
for (auto &i : data) {
std::printf("%f + %fj\n", i.real(), i.imag());
}
std::printf("=====\n");
cufftComplex *d_data = nullptr;
CUFFT_CALL(cufftCreate(&plan));
CUFFT_CALL(cufftPlan1d(&plan, data.size(), CUFFT_C2C, batch_size));
CUDA_RT_CALL(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
CUFFT_CALL(cufftSetStream(plan, stream));
// Create device data arrays
CUDA_RT_CALL(cudaMalloc(reinterpret_cast<void **>(&d_data), sizeof(data_type) * data.size()));
CUDA_RT_CALL(cudaMemcpyAsync(d_data, data.data(), sizeof(data_type) * data.size(), cudaMemcpyHostToDevice, stream));
CUFFT_CALL(cufftExecC2C(plan, d_data, d_data, CUFFT_FORWARD));
CUFFT_CALL(cufftExecC2C(plan, d_data, d_data, CUFFT_INVERSE));
CUDA_RT_CALL(cudaMemcpyAsync(data.data(), d_data, sizeof(data_type) * data.size(), cudaMemcpyDeviceToHost, stream));
CUDA_RT_CALL(cudaStreamSynchronize(stream));
/* free resources */
CUDA_RT_CALL(cudaFree(d_data))
CUFFT_CALL(cufftDestroy(plan));
CUDA_RT_CALL(cudaStreamDestroy(stream));
CUDA_RT_CALL(cudaDeviceReset());
return EXIT_SUCCESS;
}
cmake_minimum_required(VERSION 3.18)
set(ROUTINE 1d_c2c)
project(
"${ROUTINE}_example"
DESCRIPTION "GPU-Accelerated Fast Fourier Transforms"
HOMEPAGE_URL "https://docs.nvidia.com/cuda/cufft/index.html"
LANGUAGES CXX CUDA)
set(CMAKE_CUDA_ARCHITECTURES 80)
find_package(CUDAToolkit REQUIRED)
set(CMAKE_CXX_STANDARD 11)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
if("${CMAKE_BUILD_TYPE}" STREQUAL "")
set(CMAKE_BUILD_TYPE Release)
endif()
set(CMAKE_CUDA_ARCHITECTURES 80)
#if(CMAKE_CUDA_ARCHITECTURES LESS 60)
#set(CMAKE_CUDA_ARCHITECTURES 60 70 75 80 86)
#endif()
set(BUILD_SHARED_LIBS OFF)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CUFFT_LIBRARIES ${CUDA_cufft_LIBRARY} ${CUDA_culibos_LIBRARY} ${CUDA_cudart_LIBRARY})
add_executable(${ROUTINE}_example)
target_include_directories(${ROUTINE}_example
PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
${CMAKE_SOURCE_DIR}/../utils)
target_sources(${ROUTINE}_example
PRIVATE ${PROJECT_SOURCE_DIR}/${ROUTINE}_example.cpp)
set(CMAKE_CUDA_ARCHITECTURES 80)
#target_link_libraries(${ROUTINE}_example PRIVATE ${CUDA_cufft_static_LIBRARY} CUDA::cufft CUDA::cudart)
target_link_libraries(${ROUTINE}_example PRIVATE CUDA::cufft_static CUDA::cudart)
当我删除
find_package(CUDAToolkit REQUIRED)
cmake 显示以下错误:
CMake Error at CMakeLists.txt:82 (target_link_libraries):
Target "1d_c2c_example" links to:
CUDA::cufft_static
but the target was not found.
使用
CUDA::cufft_static
后仍然存在链接器问题的主要原因是静态 cuFFT 需要启用可重定位设备代码。这是通过 CUDA_SEPARABLE_COMPILATION
属性在 CMake 中完成的。
我会收回我的声明,即一个人不应该同时使用
CUDA
语言和find_package(CUDAToolkit REQUIRED)
。虽然 cufft_static
目标在仅使用语言时可用,但它不会自动链接 culibos
。所以更优雅的解决方案似乎是使用包中的CUDA::cufft_static
。
我从链接的存储库中获取了
cufft_utils.h
并将其放入${PROJECT_SOURCE_DIR}/utils
。如果您的项目结构不同,您将不得不调整 target_include_directories
命令。
Setting
CMAKE_
variables is bad style.要使其中一些起作用,您必须在 project
命令之前设置它们。但是例如CUDA 体系结构和构建类型应在第一次配置期间通过命令行参数设置为 cmake
或使用 ccmake
获得漂亮的控制台 ui.
cmake_minimum_required(VERSION 3.18)
set(ROUTINE 1d_c2c)
project(
"${ROUTINE}_example"
DESCRIPTION "GPU-Accelerated Fast Fourier Transforms"
HOMEPAGE_URL "https://docs.nvidia.com/cuda/cufft/index.html"
LANGUAGES CXX CUDA)
find_package(CUDAToolkit REQUIRED)
add_executable(${ROUTINE}_example)
set_target_properties(${ROUTINE}_example
PROPERTIES
CUDA_SEPARABLE_COMPILATION ON
RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
target_compile_features(${ROUTINE}_example
PRIVATE cuda_std_11)
target_sources(${ROUTINE}_example
PRIVATE ${PROJECT_SOURCE_DIR}/${ROUTINE}_example.cu)
target_include_directories(${ROUTINE}_example
PRIVATE ${PROJECT_SOURCE_DIR}/utils)
target_link_libraries(${ROUTINE}_example PRIVATE
PRIVATE CUDA::cufft_static)
如您所见,我给源文件打了
.cu
。让它为 .cpp
文件工作似乎有点复杂。阅读该问题的一个很好的起点是 cuFFT 文档的本章。虽然可以通过使用 find_package(Threads REQUIRED)
然后将 ${CMAKE_DL_LIBS}
和 Threads::Threads
添加到 target_link_libraries
命令来消除大多数链接器错误,但我无法使可重定位设备代码工作,因为我无法使用 CMake nvcc
用于链接。我尝试在 LINKER_LANGUAGE CUDA
中设置 set_target_properties
这在理论上应该有效,但实际上 CMake 一直使用 g++
进行链接(用 make VERBOSE=1
可见)。这可能是 CMake 中的错误。