OpenCL入门

初入OpenCL,做个记录。

在Windows下开发OpenCL程序，必须先下载OpenCL的SDK，现在AMD，NVIDIA，Intel均提供各自的OpenCL库，基本是大同小异。安装好SDK后新建Win32控制台项目，然后需要配置下包含文件路径和库路径，具体见下图（我安装的Intel的SDK ）。

1.其中那个包含Intel的路径就是包含cl.h文件的目录。

2.如图中那个Intel的lib目录

3.添加需要连接的静态库OpenCL.lib

配置完成后就可以开始写代码调试了，OpenCL的初始化还是很复杂的，和CUDA几行代码搞定完全没可比性，刚开始可能对流程不太熟悉，慢慢熟悉就好，当然也可以自己写个框架来做这些复杂的初始化工作。OpenCL的内核代码是即时编译的，代码中我为了方便没有从cl文件中读入Kernel代码，直接以字符串的形式定义了。

#include "stdafx.h"

#include <iostream>
#include <fstream>
#include <string.h>
#include <vector>
using namespace std;

#if defined(__APPLE__) || defined(__MACOSX)
#include <OpenCL/cl.hpp>
#else
#include <CL/cl.h>
#endif

#define KERNEL(...) #__VA_ARGS__

#define ARRAY_X_LEN 16
#define ARRAY_Y_LEN 16

const char *kernelSourceCode = KERNEL(
    __kernel void VecAdd(__global int *buffer1, __global int *buffer2, __global int *buffer3)
{
        size_t idx = get_global_id(0);
        size_t idy = get_global_id(1);
        int dimX = get_global_size(0);
        int dimY = get_global_size(1);
        int id = idx + idy*dimX;
        buffer3[id] = buffer1[id] + buffer2[id];
    });

int main()
{
    cl_int status = 0;
    size_t deviceListSize;
    cl_uint numPlatforms;
    cl_platform_id platform = NULL;
    status = clGetPlatformIDs(0, NULL, &numPlatforms);
    if (status != CL_SUCCESS)
    {
        printf("获取平台数目失败");
        return EXIT_FAILURE;
    }
    if (numPlatforms >0)
    {
        cl_platform_id* platforms = (cl_platform_id*)malloc(numPlatforms*sizeof(cl_platform_id));
        status = clGetPlatformIDs(numPlatforms, platforms, NULL);
        if (status != CL_SUCCESS)
        {
            printf("初始化平台失败");
            return -1;
        }
        for (unsigned int i = 0; i<numPlatforms; ++i)
        {
            char *vendor = (char*)malloc(100);
            status = clGetPlatformInfo(platforms[i], CL_PLATFORM_VENDOR, sizeof(vendor), vendor, NULL);
            platform = platforms[i];
            if (!strcmp(vendor, "NVIDIA Corporation"))
            {
                break;
            }
        }
        delete platforms;
    }
    cl_context_properties cps[3] = { CL_CONTEXT_PLATFORM, (cl_context_properties)platform, 0 };
    cl_context_properties* cprops = (NULL == platform) ? NULL : cps;
    cl_context context = clCreateContextFromType(cprops, CL_DEVICE_TYPE_GPU, NULL, NULL, &status);
    if (status != CL_SUCCESS)
    {
        printf("创建上下文失败");
        return EXIT_FAILURE;
    }
    status = clGetContextInfo(context, CL_CONTEXT_DEVICES, 0, NULL, &deviceListSize);
    if (status != CL_SUCCESS)
    {
        printf("获取设备数目失败");
        return EXIT_FAILURE;
    }
    cl_device_id *devices = (cl_device_id *)malloc(deviceListSize);
    if (devices == 0)
    {
        printf("为设备分配空间失败");
        return EXIT_FAILURE;
    }
    status = clGetContextInfo(context, CL_CONTEXT_DEVICES, deviceListSize, devices, NULL);
    if (status != CL_SUCCESS)
    {
        printf("初始化设备失败");
        return EXIT_FAILURE;
    }

    size_t sourceSize[] = { strlen(kernelSourceCode) };
    cl_program program = clCreateProgramWithSource(context, 1, &kernelSourceCode, sourceSize, &status);
    if (status != CL_SUCCESS)
    {
        printf("创建程序失败");
        return EXIT_FAILURE;
    }
    status = clBuildProgram(program, 1, devices, NULL, NULL, NULL);
    if (status != CL_SUCCESS)
    {
        printf("编译程序失败");
        return EXIT_FAILURE;
    }
    cl_kernel kernel = clCreateKernel(program, "VecAdd", &status);
    if (status != CL_SUCCESS)
    {
        printf("创建内核失败");
        return EXIT_FAILURE;
    }
    cl_command_queue commandQueue = clCreateCommandQueue(context, devices[0], 0, &status);
    if (status != CL_SUCCESS)
    {
        printf("创建命令队列失败");
        return EXIT_FAILURE;
    }
    int arrayLenght = ARRAY_X_LEN*ARRAY_Y_LEN;
    int arraySize = arrayLenght*sizeof(int);

    int *hA = new int[arrayLenght];
    int *hB = new int[arrayLenght];
    int *hC = new int[arrayLenght];

    memset(hA, 0, arraySize);
    memset(hB, 0, arraySize);
    memset(hC, 0, arraySize);

    for (int i = 0; i<arrayLenght; i++)
    {
        hA[i] = i;
        hB[i] = i;
    }

    cl_mem dA = clCreateBuffer(context, CL_MEM_ALLOC_HOST_PTR, arraySize, NULL, &status);
    if (status != CL_SUCCESS)
    {
        printf("创建内存对象失败");
        return EXIT_FAILURE;
    }
    cl_mem dB = clCreateBuffer(context, CL_MEM_ALLOC_HOST_PTR, arraySize, NULL, &status);
    if (status != CL_SUCCESS)
    {
        printf("创建内存对象失败");
        return EXIT_FAILURE;
    }
    cl_mem dC = clCreateBuffer(context, CL_MEM_ALLOC_HOST_PTR, arraySize, NULL, &status);
    if (status != CL_SUCCESS)
    {
        printf("创建内存对象失败");
        return EXIT_FAILURE;
    }
    status = clEnqueueWriteBuffer(commandQueue, dA, CL_TRUE, 0, arraySize, hA, 0, NULL, NULL);
    if (status != CL_SUCCESS)
    {
        printf("输入值写入内存对象失败");
        return EXIT_FAILURE;
    }
    status = clEnqueueWriteBuffer(commandQueue, dB, CL_TRUE, 0, arraySize, hB, 0, NULL, NULL);
    if (status != CL_SUCCESS)
    {
        printf("输入值写入内存对象失败");
        return EXIT_FAILURE;
    }
    status = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void*)&dA);
    if (status != CL_SUCCESS)
    {
        printf("设置内核参数失败");
        return EXIT_FAILURE;
    }
    status = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void*)&dB);
    if (status != CL_SUCCESS)
    {
        printf("设置内核参数失败");
        return EXIT_FAILURE;
    }
    status = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void*)&dC);
    if (status != CL_SUCCESS)
    {
        printf("设置内核参数失败");
        return EXIT_FAILURE;
    }
    size_t globalThreads[] = { ARRAY_X_LEN, ARRAY_Y_LEN };
    size_t localThreads[] = { 4, 4 };
    status = clEnqueueNDRangeKernel(commandQueue, kernel, 2, NULL, globalThreads, localThreads, 0, NULL, NULL);
    if (status != CL_SUCCESS)
    {
        printf("将内核放入命令队列失败");
        return EXIT_FAILURE;
    }
    status = clFinish(commandQueue);
    if (status != CL_SUCCESS)
    {
        printf("队列还没有完成");
        return EXIT_FAILURE;
    }
    status = clEnqueueReadBuffer(commandQueue, dC, CL_TRUE, 0, arraySize, hC, 0, NULL, NULL);
    if (status != CL_SUCCESS)
    {
        printf("读内存对象失败");
        return EXIT_FAILURE;
    }
    printf("结果:\n");
    for (int i = 0; i<arrayLenght; i++)
    {
        printf("%d ", hC[i]);
        if ((i + 1) % ARRAY_X_LEN == 0)
            printf("\n");
    }
    status = clReleaseKernel(kernel);
    status = clReleaseProgram(program);
    status = clReleaseMemObject(dA);
    status = clReleaseMemObject(dB);
    status = clReleaseMemObject(dC);
    status = clReleaseCommandQueue(commandQueue);
    status = clReleaseContext(context);
    free(devices);
    delete [] hA;
    delete [] hB;
    delete [] hC;
    return 0;
}

运行结果：