由C主机代码调用cublasSgemm的结果不正确

从C主机代码调用cuBLAS库函数cublasSgemm时，我得到了一些奇怪的数字。它确实编译并运行，但结果矩阵中的数字不正确

C宿主代码调用这些函数的问题在于，C语言按行主顺序读取矩阵，而cuBLAS函数是用FORTRAN编写的，按列主顺序读取矩阵

我尝试了许多CubLassegm的参数组合，但似乎没有一个能正常工作

我需要在m1和m2之间执行矩阵求积，所以我首先传递了m2，然后传递了m1，所以cublas函数应该读取（m2）T和（m1）T，其中T是转换形式；通过这样做，我应该得到结果（r）T=（m2.m1）T。 我的C代码最终应该将（r）T读作r，但我无法得到正确的数字。。。 代码如下：

cudaError_t vector_matrix_molt(float *m1, float *m2, float *r, int row1, int col1, int row2, int col2) {

    //Device Memory allocation
    float *d_m1;
    float *d_m2;
    float *d_r;
    float a = 1.0f;
    float b = 0.0f;
    int stride = 1;
    //CUDA stuff
    cublasHandle_t handle;
    cudaError_t cudaStatus;


    cudaStatus = cudaMalloc((void**)&d_m1, col1*row1*sizeof(float));
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaMalloc failed!");
        goto Error;
    }

    cudaStatus = cudaMalloc((void**)&d_m2, row2*col2*sizeof(float));
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaMalloc failed!");
        goto Error;
    }

    cudaStatus = cudaMalloc((void**)&d_r, row1*col2*sizeof(float));
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaMalloc failed!");
        goto Error;
    }

    cublasCreate(&handle);

    // Copy Data to Device Memory
    cudaStatus = cudaMemcpy(d_m1, m1, row1*col1*sizeof(float), cudaMemcpyHostToDevice);
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaMemcpy 1 failed!");
        goto Error;
    }

    cudaStatus = cudaMemcpy(d_m2, m2, row2*col2*sizeof(float), cudaMemcpyHostToDevice);
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaMemcpy 2 failed!");
        goto Error;
    }

    /*cublasStatus_t cublasSgemm(cublasHandle_t handle, cublasOperation_t transa, cublasOperation_t transb,
    int m, int n, int k, const float *alpha, const float *A, int lda, const float *B, int ldb, const float *beta, float *C, int ldc
    */
    //Calling cuBLAS library function... 
    cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, col2, row1, col1, &a, d_m2, col2, d_m1, col1, &b, d_r, row1);

    // Check for any errors launching the kernel
    cudaStatus = cudaGetLastError();
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "moltKernel launch failed: %s\n", cudaGetErrorString(cudaStatus));
        goto Error;
    }

    // cudaDeviceSynchronize waits for the kernel to finish, and returns
    // any errors encountered during the launch.
    cudaStatus = cudaDeviceSynchronize();
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaDeviceSynchronize returned error code %d after launching cublasSgemv!\n", cudaStatus);
        //printf("Cuda Error: %s\n", cudaGetErrorString(cudaStatus));
        goto Error;
    }

    // Copy output vector from GPU buffer to host memory.
    cudaStatus = cudaMemcpy(r, d_r, row1*col2* sizeof(float), cudaMemcpyDeviceToHost);
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaMemcpy 3 failed!");
        goto Error;
    }

Error:
    cudaFree(d_m1);
    cudaFree(d_m2);
    cudaFree(d_r);

    return cudaStatus;
}

---

唯一需要更改的是

r

的前导dim

cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, col2, row1, col1, &a, d_m2, col2, d_m1, col1, &b, d_r, col2);

你可以参考这个答案来获得更详细的解释

---

非常感谢！我刚刚在两个矩阵中存储了一些值，它是有效的。由于某些原因，我不能用随机生成的浮点矩阵得到相同的结果，但我可以说我的目标已经实现了。再次感谢；）