编译CUDA示例程序

Question

我有一个非常简单的 CUDA 程序，但无法编译

这是

main.cpp

#include <iostream>
#include <cstdlib>
#include "/opt/cuda/targets/x86_64-linux/include/cuda_runtime.h"
#include "/opt/cuda/targets/x86_64-linux/include/cublas_v2.h"

// Function to initialize a matrix with random values
void initMatrix(float* matrix, int rows, int cols) {
    for (int i = 0; i < rows * cols; ++i) {
        matrix[i] = rand() / (float)RAND_MAX;  // Initialize with random values
    }
}

// Function to print a matrix
void printMatrix(float* matrix, int rows, int cols) {
    for (int i = 0; i < rows; ++i) {
        for (int j = 0; j < cols; ++j) {
            std::cout << matrix[i * cols + j] << " ";
        }
        std::cout << std::endl;
    }
}

int main() {
    // Matrix dimensions
    int m = 3;  // Number of rows
    int n = 3;  // Number of columns

    // Allocate host memory for the matrix and vector
    float* h_matrix = new float[m * n];
    float* h_vector = new float[n];

    // Initialize matrix and vector with random values
    initMatrix(h_matrix, m, n);
    initMatrix(h_vector, n, 1);

    // Print the matrix and vector
    std::cout << "Matrix:" << std::endl;
    printMatrix(h_matrix, m, n);

    std::cout << "\nVector:" << std::endl;
    printMatrix(h_vector, n, 1);

    // Allocate device memory for the matrix and vector
    float* d_matrix;
    float* d_vector;
    cudaMalloc((void**)&d_matrix, m * n * sizeof(float));
    cudaMalloc((void**)&d_vector, n * sizeof(float));

    // Transfer the matrix and vector from host to device
    cudaMemcpy(d_matrix, h_matrix, m * n * sizeof(float), cudaMemcpyHostToDevice);
    cudaMemcpy(d_vector, h_vector, n * sizeof(float), cudaMemcpyHostToDevice);

    // cuBLAS handle
    cublasHandle_t handle;
    cublasCreate(&handle);

    // Perform matrix-vector multiplication
    float alpha = 1.0;  // Scaling factor
    float beta = 0.0;   // Scaling factor
    cublasSgemv(handle, CUBLAS_OP_N, m, n, &alpha, d_matrix, m, d_vector, 1, &beta, d_vector, 1);

    // Transfer the result back to the host
    cudaMemcpy(h_vector, d_vector, m * sizeof(float), cudaMemcpyDeviceToHost);

    // Print the result
    std::cout << "\nResult:" << std::endl;
    printMatrix(h_vector, m, 1);

    // Clean up
    delete[] h_matrix;
    delete[] h_vector;
    cudaFree(d_matrix);
    cudaFree(d_vector);
    cublasDestroy(handle);

    return 0;
}

这是

CMakeLists.txt

:

cmake_minimum_required(VERSION 3.17)
project(CUBLAS_Matrix_Vector_Multiplication LANGUAGES CXX CUDA)

find_package(CUDAToolkit REQUIRED)
find_package(CUDA REQUIRED)

list(APPEND CUDA_NVCC_FLAGS "-std=c++11")
set(CUDA_PROPAGATE_HOST_FLAGS OFF)
add_definitions(-std=c++11)

set(CUDA_SEPARABLE_COMPILATION ON)

add_executable(matrix_vector_multiplication main.cpp)

target_include_directories(matrix_vector_multiplication PRIVATE ${CUDAToolkit_INCLUDE_DIRS})
set_target_properties(matrix_vector_multiplication PROPERTIES CUDA_SEPARABLE_COMPILATION ON CUDA_STANDARD 11)
target_link_libraries(matrix_vector_multiplication PRIVATE ${CUDAToolkit_LIBRARIES})

尝试通过

cmake

创建配置文件来编译它。但随后

make

会导致很多错误：

/usr/bin/ld: CMakeFiles/matrix_vector_multiplication.dir/main.cpp.o: in function `main':
main.cpp:(.text+0x8056): undefined reference to `cudaMalloc'
/usr/bin/ld: main.cpp:(.text+0x8072): undefined reference to `cudaMalloc'
/usr/bin/ld: main.cpp:(.text+0x8098): undefined reference to `cudaMemcpy'
/usr/bin/ld: main.cpp:(.text+0x80ba): undefined reference to `cudaMemcpy'
/usr/bin/ld: main.cpp:(.text+0x80c6): undefined reference to `cublasCreate_v2'
/usr/bin/ld: main.cpp:(.text+0x8112): undefined reference to `cublasSgemv_v2'
/usr/bin/ld: main.cpp:(.text+0x8138): undefined reference to `cudaMemcpy'
/usr/bin/ld: main.cpp:(.text+0x81c6): undefined reference to `cudaFree'
/usr/bin/ld: main.cpp:(.text+0x81d2): undefined reference to `cudaFree'
/usr/bin/ld: main.cpp:(.text+0x81de): undefined reference to `cublasDestroy_v2'
collect2: Fehler: ld gab 1 als Ende-Status zurück
make[2]: *** [CMakeFiles/matrix_vector_multiplication.dir/build.make:97: matrix_vector_multiplication] Fehler 1
make[1]: *** [CMakeFiles/Makefile2:83: CMakeFiles/matrix_vector_multiplication.dir/all] Fehler 2
make: *** [Makefile:91: all] Fehler 2

尝试解决

我尝试遵循这个问题，并使用：

 g++ main.cpp -o main -L/opt/cuda/lib64/ -lcudart -lcuda -L/opt/cuda

这会引发错误：

In file included from /opt/cuda/targets/x86_64-linux/include/cublas_v2.h:69,
                 from main.cpp:4:
/opt/cuda/targets/x86_64-linux/include/cublas_api.h:77:10: fatal error: cuda_fp16.h: No such file or directory
   77 | #include <cuda_fp16.h>
      |          ^~~~~~~~~~~~~
compilation terminated.

如何解决这个问题？

 nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Fri_Sep__8_19:17:24_PDT_2023
Cuda compilation tools, release 12.3, V12.3.52
Build cuda_12.3.r12.3/compiler.33281558_0

Answer 1

对我来说这很好用：

cmake_minimum_required(VERSION 3.17)
project(CUBLAS_Matrix_Vector_Multiplication LANGUAGES CXX)

find_package(CUDAToolkit REQUIRED)

add_executable(matrix_vector_multiplication main.cpp)

target_link_libraries(matrix_vector_multiplication PRIVATE
        CUDA::cublas
        CUDA::cudart
)
target_compile_features(matrix_vector_multiplication PRIVATE cxx_std_11)

当您将包含更改为

#include <cuda_runtime.h>
#include <cublas_v2.h>