CUDA矩阵乘法

1 、乘法

#include <cstdio>
#include <iostream>
#include <chrono>

#define TILE_WIDTH 32

__global__ void MatrixMulKernel(float* md, float* nd, float* pd, int width) {
    __shared__ float mds[TILE_WIDTH][TILE_WIDTH];
    __shared__ float nds[TILE_WIDTH][TILE_WIDTH];

    int bx = blockIdx.x;
    int by = blockIdx.y;
    int tx = threadIdx.x;
    int ty = threadIdx.y;

    int row = by * TILE_WIDTH + ty;
    int col = bx * TILE_WIDTH + tx;

    float p_value = 0;
    for (int i = 0; i < width/TILE_WIDTH; i++) {
        mds[ty][tx] = md[row * width + (i*TILE_WIDTH + tx)];
        nds[ty][tx] = nd[(i*TILE_WIDTH + ty)*width + col];
        __syncthreads();

        for (int j = 0; j < TILE_WIDTH; j++) {
            p_value += mds[ty][j] * nds[j][tx];
        }
        __syncthreads();
    }

    pd[row*width + col] = p_value;
}

int main() {
	......
    MatrixMulKernel<<<dim3(width/TILE_WIDTH, width/TILE_WIDTH), dim3(TILE_WIDTH, TILE_WIDTH)>>>(md, nd, pd, width);
    ......
}