参考资料

CUDA编程模型系列三(矩阵乘)_哔哩哔哩_bilibili

极力推荐的良心课程~

代码片段

#include <stdio.h>const int matrix_size = 32;
const int BLOCK_SIZE = 16;__global__ void matrix_multi(const int* a1, const int *b1, int *c1, const int size)
{int y = blockDim.y * blockIdx.y + threadIdx.y;int x = blockDim.x * blockIdx.x + threadIdx.x;int t_sum = 0;if(x < size && y < size) {for(int step = 0; step < size; step++) {t_sum += a1[y * size + step] * b1[step * size + x];}c1[y * size + x] = t_sum;printf("x: %d  y: %d  sum: %d \n", x, y, t_sum);}
}int main()
{int* a1, *b1, *c1;int msize = matrix_size * matrix_size * sizeof(int);// 分配内存空间cudaMallocHost((void**)&a1, msize);cudaMallocHost((void**)&b1, msize);cudaMallocHost((void**)&c1, msize);// 初始化for(int i = 0; i < matrix_size; i++) {for(int j = 0; j < matrix_size; j++) {a1[i * matrix_size + j] = 1;b1[i * matrix_size + j] = 1;}}printf("Check result a1: %d \n", a1[1]);printf("Check result b1: %d \n", b1[1]);// 分配设备空间int *x1, *y1, *z1;cudaMalloc((void**)&x1, msize);cudaMalloc((void**)&y1, msize);cudaMalloc((void**)&z1, msize); // Host => DevicecudaMemcpy(x1, a1, msize, cudaMemcpyHostToDevice);cudaMemcpy(y1, b1, msize, cudaMemcpyHostToDevice);// printf("Check result: %d \n", x1[1]);// printf("Check result: %d \n", y1[1]);int grid_rows = (matrix_size + BLOCK_SIZE - 1) / BLOCK_SIZE;int grid_cols = (matrix_size + BLOCK_SIZE - 1) / BLOCK_SIZE;dim3 gridDim(grid_cols, grid_rows);dim3 blockDim(BLOCK_SIZE, BLOCK_SIZE);// 计算matrix_multi<<<gridDim, blockDim>>>(x1, y1, z1, matrix_size);// Device => HostcudaMemcpy(c1, z1, msize, cudaMemcpyDeviceToHost);printf("Check result c1: %d \n", c1[1]);// 释放空间cudaFreeHost(a1);cudaFreeHost(b1);cudaFreeHost(c1);cudaFree(x1);cudaFree(y1);cudaFree(z1);return 0;
}