以一个博文《CUDA例子》来说明。
#include "cuda_runtime.h"#include "device_launch_parameters.h"#include <stdio.h>#define N 10__global__ void add(int *a, int *b, int *c){int tid =blockIdx.x;c[tid] = a[tid] + b[tid];}int main(){int a[N], b[N], c[N];int *deva, *devb, *devc;//在GPU上分配内存cudaMalloc((void **)&deva, N*sizeof(int));cudaMalloc((void **)&devb, N*sizeof(int));cudaMalloc((void **)&devc, N*sizeof(int));//在CPU上为数组赋值for (int i = 0; i < N; i++){a[i] = -i;b[i] = i*i;}//将数组a和b传到GPUcudaMemcpy(deva, a, N*sizeof(int), cudaMemcpyHostToDevice);cudaMemcpy(devb, b, N*sizeof(int), cudaMemcpyHostToDevice);cudaMemcpy(devc, c, N*sizeof(int), cudaMemcpyHostToDevice);add <<<N, 1 >> >(deva, devb, devc);//将数组c从GPU传到CPUcudaMemcpy(c, devc, N*sizeof(int), cudaMemcpyDeviceToHost);for (int i = 0; i < N; i++){printf("%d+%d=%d\n", a[i], b[i], c[i]);}cudaFree(deva);cudaFree(devb);cudaFree(devc);return 0;}
cpu
//#include "cuda_runtime.h"//#include "device_launch_parameters.h"#include <stdio.h>#define N 10//加一个全局并行变量int blockIdx_x; //__global__ void add(int *a, int *b, int *c){int tid =blockIdx_x;//blockIdx.x;c[tid] = a[tid] + b[tid];}int main(){int a[N], b[N], c[N];int *deva, *devb, *devc;// //在GPU上分配内存// cudaMalloc((void **)&deva, N*sizeof(int));// cudaMalloc((void **)&devb, N*sizeof(int));// cudaMalloc((void **)&devc, N*sizeof(int));//在CPU上为数组赋值for (int i = 0; i < N; i++){a[i] = -i;b[i] = i*i;}// //将数组a和b传到GPU// cudaMemcpy(deva, a, N*sizeof(int), cudaMemcpyHostToDevice);// cudaMemcpy(devb, b, N*sizeof(int), cudaMemcpyHostToDevice);// cudaMemcpy(devc, c, N*sizeof(int), cudaMemcpyHostToDevice);deva = a; devb = b; devc = c;for (int blockIdx_x = 0; blockIdx_x < N; blockIdx_x++){//<<<N, 1 >> >add (deva, devb, devc);//add (a, b, c);//也可以直接这句}// //将数组c从GPU传到CPU// cudaMemcpy(c, devc, N*sizeof(int), cudaMemcpyDeviceToHost);for (int i = 0; i < N; i++){printf("%d+%d=%d\n", a[i], b[i], c[i]);}// cudaFree(deva);// cudaFree(devb);// cudaFree(devc);return 0;}
对于OpenMP 加上 “#pragma omp parallel for”就是CPU并行了
总结：
1。去掉数据从cpu-->gpu 和 gpu-->cpu 的相关代码
2。核函数作为cpu并行对象放到for循环中
3。把双角括号<<>>的数据变为for的循环变量
------------------------------------------------------分割线----------------------------------------------------
glsl
先来对比一下glsl 和 cuda 计算单位结构图
可见，除了名称叫法不同，组成方式是一样的。
//#include "cuda_runtime.h"//#include "device_launch_parameters.h"#include <stdio.h>#define N 10//__global__ //void add(int *a, int *b, int *c)//{// int tid =blockIdx_x;//blockIdx.x;// c[tid] = a[tid] + b[tid];// //}//核函数放入Source中，作为glsl源码char *Source={"#version 430 core\n""layout (local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = 1) in;\n""layout(std430, binding = 0) buffer buffer0\n""{\n""int a[];\n""};\n""layout(std430, binding = 1) buffer buffer1\n""{\n""int b[];\n""};\n""layout(std430, binding = 2) buffer buffer2\n""{\n""int c[];\n""};\n""void main(void)\n""{\n""int tid =int(gl_WorkGroupID.x);\n""c[tid] = a[tid] + b[tid];\n""}\n"}int main(int argc,char **argv){//glsl 初始化init_glut_glew(argc,argv);int a[N], b[N], c[N];int *deva, *devb, *devc;// //在GPU上分配内存// cudaMalloc((void **)&deva, N*sizeof(int));// cudaMalloc((void **)&devb, N*sizeof(int));// cudaMalloc((void **)&devc, N*sizeof(int));分配glsl缓存ID();//glGenBuffers()//分配缓存 和传送数据一起//在CPU上为数组赋值for (int i = 0; i < N; i++){a[i] = -i;b[i] = i*i;}// //将数组a和b传到GPU// cudaMemcpy(deva, a, N*sizeof(int), cudaMemcpyHostToDevice);// cudaMemcpy(devb, b, N*sizeof(int), cudaMemcpyHostToDevice);// cudaMemcpy(devc, c, N*sizeof(int), cudaMemcpyHostToDevice);传送数组(a,b,c,N*sizeof(int));//glBindBuffer();glBufferData();glBindBufferBase();glUnmapBuffer();设置glsl运行参数(N);//local_size_x = N, local_size_y = 1, local_size_z = 1;//代入Source中//NumGroupsX = 1, NumGroupsY = 1, NumGroupsZ = 1;编译链接源码(Source);//glCompileShader();glLinkProgram()运行计算着色器();//performCompute();//add (deva, devb, devc);// //将数组c从GPU传到CPU// cudaMemcpy(c, devc, N*sizeof(int), cudaMemcpyDeviceToHost);取回结果(c);//glBindBuffer();glMapBuffer();glUnmapBuffer();for (int i = 0; i < N; i++){printf("%d+%d=%d\n", a[i], b[i], c[i]);}// cudaFree(deva);// cudaFree(devb);// cudaFree(devc);return 0;}
总结：
1。核函数作为源码放入一个字符串中
2。计算单位的设置都有对应的名称代入
3。分配存储传送取回数据也有对应的函数
