OpenCL 初实践（1）矩阵相乘

来源：互联网发布：我是歌手4网络踢馆赛编辑：程序博客网时间：2024/05/19 15:24

对2000*2000的两个矩阵做矩阵相乘，

执行内核设定的globalsize 是二维的，200*200，实验环境是centos OpenCL 2.0 AMD-APP (1800.5) AMD Accelerated Parallel Processing，

MaxItemSize=256,256,256,

故选取的globalsize global item size 是200,200.

size_t globalSize[2] = {200,200};
size_t localSize[1] = {200};
err = clEnqueueNDRangeKernel(commands, kernel, 2, NULL, globalSize, NULL, 0, NULL, NULL);

核函数：

__kernel void mmult(__global int* a, __global int* b, __global int* output)

{
int width = get_global_id(1);
int height = get_global_id(0);
int num_size = get_global_size(0);
int rank = 2000;
for (int total = 1;total<=10;total++)
{
int running = 0;
for(int num =0;num<2000;num++)
{
int aIndex = width*rank+num;
int bIndex = num*rank + height*10+total-1;
running += a[aIndex]*b[bIndex];
}
output[width*rank+height*10+total-1] = running;
}

return;

}

0 0