c - CUDA での非正方行列の乗算

Question

私の GPU プログラミング クラスでは、非正方行列の乗算プログラムの特定の部分を完了するように割り当てられました。具体的には、カーネル関数と、スレッド ブロックとカーネル グリッドの次元の初期化です。

I've based my code on the CUDA C Programming Guide's matrix multiplication code, but instead of using structs as they do, I have modified mine to use only the parameters given (since we're not allowed to change parameters). We are provided with the 3 matrices A, B, and C, as well as the dimensions of them- m x k, k x n, and m x n, respectively. Where the struct used A.height, I've used dimension m, where it used B.width, I've used dimension n, etc.

I've run into several problems, the first of which is that my program doesn't pass the included test, which verifies the correctness of the product matrix C. I assume that there is something wrong in my matrix multiplication code, then, and that the issue probably arises from me adapting the struct code.

#include <stdio.h>
__global__ void mysgemm(int m, int n, int k, const float *A, const float *B,
        float* C) {

    /********************************************************************
     *
     * Compute C = A x B
     *   where A is a (m x k) matrix
     *   where B is a (k x n) matrix
     *   where C is a (m x n) matrix
     *
     ********************************************************************/

    // INSERT KERNEL CODE HERE
    // Each thread computes one element of C
    // by accumulating results into Cvalue
    float Cvalue = 0;
    int row = blockIdx.y * blockDim.y + threadIdx.y;
    int col = blockIdx.x * blockDim.x + threadIdx.x;

    for (int e = 0; e < k; ++e){
        Cvalue += (A[row * k + e]) * (B[e * n + col]);
    }
    C[row * n + col] = Cvalue;
}

My other problem, which I'm even less sure about, involves the code to initialize the thread block and kernel grid dimensions.

// Initialize thread block and kernel grid dimensions ---------------------
    const unsigned int BLOCK_SIZE = 16; // Use 16x16 thread blocks
//INSERT CODE HERE
    dim3 dimBlock(BLOCK_SIZE, BLOCK_SIZE);
    dim3 dimGrid(n / dimBlock.x, m / dimBlock.y);
// Invoke CUDA kernel -----------------------------------------------------
//INSERT CODE HERE
    mysgemm<<<dimGrid, dimBlock>>>(m, n, k, A, B, C);

I understand dimBlock, but I don't understand dimGrid, and don't have a proper idea of what to use as parameters for it. When I run the code as is, the kernel won't even launch if the matrix I pass in doesn't have a dimension that is a power of 2. And if I do use a power of 2, the test still fails.

I apologize if I've been too wordy. This is my first post and I wanted to give as many details as possible. Hopefully someone can help walk me through these issues.