CMake编译CUDA程序时如何静态链接cuFFT库？

我需要编译NVIDIA CUDALibrarySamples中的1d_c2c示例，目标是通过静态版本的cuFFT库（-lcufft_static）完成编译链接。在Makefile中只需在编译命令添加-lcufft_static即可实现，但在CMake中尝试多种写法都遇到问题：

尝试过的方法及错误 #

1. 直接引用静态库变量：

   target_link_libraries(${ROUTINE}_example PRIVATE ${CUDA_cufft_static_LIBRARY})
   

   无效果，链接时仍使用动态库。

2. 使用CMake CUDA目标：
   根据建议改用：

   target_link_libraries(${ROUTINE}_example PRIVATE CUDA::cufft_static CUDA::cudart)
   

   出现链接错误：

   /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'
   

3. 移除find_package(CUDAToolkit REQUIRED)：
   CMake直接报错找不到目标：

   CMake Error at CMakeLists.txt:82 (target_link_libraries):
      Target "1d_c2c_example" links to:
        CUDA::cufft_static
      but the target was not found.
   

示例代码与当前CMake配置 #

1d_c2c_example.cpp #

#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;
}

当前CMakeLists.txt #

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)

解决方案 #

静态链接cuFFT时，需要补充依赖库并调整链接配置，具体修改如下：

修改后的CMakeLists.txt关键部分 #

# 确保静态链接全局设置
set(BUILD_SHARED_LIBS OFF)

# 替换原target_link_libraries为以下内容
target_link_libraries(${ROUTINE}_example PRIVATE
    CUDA::cufft_static
    CUDA::cudart_static
    CUDA::culibos
)

关键说明 #

1. 使用静态CUDA运行时库：CUDA::cudart_static替代CUDA::cudart，确保运行时库也静态链接，避免动态依赖。
2. 添加culibos依赖：静态cuFFT库依赖CUDA的culibos库，提供基础设备函数支持，缺失会导致链接错误。
3. 确保BUILD_SHARED_LIBS为OFF：全局禁用动态库构建，强制CMake优先选择静态库目标。
4. 确认CUDA架构匹配：CMAKE_CUDA_ARCHITECTURES设置需与你的GPU架构一致（示例中为80，对应Ampere架构），避免设备代码不兼容。

完成上述修改后，重新执行CMake配置与编译即可成功静态链接cuFFT库。

内容的提问来源于stack exchange，提问作者MANOS