1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

1. CMAKE_MINIMUM_REQUIRED(VERSION 3.24.1)

2. SET(PROJECT_NAME demo)

3. PROJECT(${PROJECT_NAME} LANGUAGES CXX CUDA)
4. set(CMAKE_CUDA_ARCHITECTURES 75)

5. include(GNUInstallDirs)

6. find_package(CUDAToolkit)

7. add_executable(${PROJECT_NAME} main.cu)
8. target_link_libraries(${PROJECT_NAME} PRIVATE CUDA::cublas)

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101

#include <cstdio>
#include <cstdlib>
#include <vector>

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

#include "cbblas_utils.h"

using data_type = double;

int main(int argc, char *argv[]) {
    cublasHandle_t cublasH = NULL;
    cudaStream_t stream = NULL;

    const int m = 2;
    const int n = 2;
    const int k = 2;
    const int lda = 2;
    const int ldb = 2;
    const int ldc = 2;
    /*
     *   A = | 1.0 | 2.0 |
     *       | 3.0 | 4.0 |
     *
     *   B = | 5.0 | 6.0 |
     *       | 7.0 | 8.0 |
     */

    const std::vector<data_type> A = {1.0, 8.0, 3.0, 4.0};
    const std::vector<data_type> B = {5.0, 6.0, 7.0, 8.0};
    std::vector<data_type> C(m * n);
    const data_type alpha = 1.0;
    const data_type beta = 0.0;

    data_type *d_A = nullptr;
    data_type *d_B = nullptr;
    data_type *d_C = nullptr;

    cublasOperation_t transa = CUBLAS_OP_N;
    cublasOperation_t transb = CUBLAS_OP_N;

    printf("A\n");
    print_matrix(m, k, A.data(), lda);
    printf("=====\n");

    printf("B\n");
    print_matrix(k, n, B.data(), ldb);
    printf("=====\n");

    /* step 1: create cublas handle, bind a stream */
    CUBLAS_CHECK(cublasCreate(&cublasH));

    CUDA_CHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
    CUBLAS_CHECK(cublasSetStream(cublasH, stream));

    /* step 2: copy data to device */
    CUDA_CHECK(cudaMalloc(reinterpret_cast<void **>(&d_A), sizeof(data_type) * A.size()));
    CUDA_CHECK(cudaMalloc(reinterpret_cast<void **>(&d_B), sizeof(data_type) * B.size()));
    CUDA_CHECK(cudaMalloc(reinterpret_cast<void **>(&d_C), sizeof(data_type) * C.size()));

    CUDA_CHECK(cudaMemcpyAsync(d_A, A.data(), sizeof(data_type) * A.size(), cudaMemcpyHostToDevice,
                               stream));
    CUDA_CHECK(cudaMemcpyAsync(d_B, B.data(), sizeof(data_type) * B.size(), cudaMemcpyHostToDevice,
                               stream));

    while(1) {
        /* step 3: compute */
    CUBLAS_CHECK(
        cublasDgemm(cublasH, transa, transb, m, n, k, &alpha, d_A, lda, d_B, ldb, &beta, d_C, ldc));
    }
    

    /* step 4: copy data to host */
    CUDA_CHECK(cudaMemcpyAsync(C.data(), d_C, sizeof(data_type) * C.size(), cudaMemcpyDeviceToHost,
                               stream));

    CUDA_CHECK(cudaStreamSynchronize(stream));

    /*
     *   C = | 23.0 | 31.0 |
     *       | 34.0 | 46.0 |
     */

    printf("C\n");
    print_matrix(m, n, C.data(), ldc);
    printf("=====\n");

    /* free resources */
    CUDA_CHECK(cudaFree(d_A));
    CUDA_CHECK(cudaFree(d_B));
    CUDA_CHECK(cudaFree(d_C));

    CUBLAS_CHECK(cublasDestroy(cublasH));

    CUDA_CHECK(cudaStreamDestroy(stream));

    CUDA_CHECK(cudaDeviceReset());

    return EXIT_SUCCESS;
}