OpenCL多次循环执行内核的一个简单样例

#include 
#include <string.h>
#include 

#ifdef __APPLE__
#include 
#else
#include 
#endif


int main(void)
{
    cl_int ret;
    
    cl_platform_id platform_id = NULL;
    cl_device_id device_id = NULL;
    cl_context context = NULL;
    cl_command_queue command_queue = NULL;
    cl_mem memObj1 = NULL;
    cl_mem memObj2 = NULL;
    char *kernelSource = NULL;
    cl_program program = NULL;
    cl_kernel kernel = NULL;
    int *pInputBuffer1 = NULL;
    int *pInputBuffer2 = NULL;
    int *pOutputBuffer = NULL;
    
    clGetPlatformIDs(1, &platform_id, NULL);
    if(platform_id == NULL)
    {
        puts("Get OpenCL platform failed!");
        goto FINISH;
    }
    
    clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_GPU, 1, &device_id, NULL);
    if(device_id == NULL)
    {
        puts("No GPU available as a compute device!");
        goto FINISH;
    }
    
    context = clCreateContext(NULL, 1, &device_id, NULL, NULL, &ret);
    if(cOntext== NULL)
    {
        puts("Context not established!");
        goto FINISH;
    }
    
    command_queue = clCreateCommandQueue(context, device_id, 0, &ret);
    if(command_queue == NULL)
    {
        puts("Command queue cannot be created!");
        goto FINISH;
    }
    
    // Specify the path of the kernel source
    const char *pFileName = "/Users/zennychen/Downloads/test.cl";
    
    FILE *fp = fopen(pFileName, "r");
    if (fp == NULL)
    {
        puts("The specified kernel source file cannot be opened!");
    goto FINISH;
    }
    fseek(fp, 0, SEEK_END);
    const long kernelLength = ftell(fp);
    fseek(fp, 0, SEEK_SET);
    
    kernelSource = malloc(kernelLength);
    
    fread(kernelSource, 1, kernelLength, fp);
    fclose(fp);
    
    program = clCreateProgramWithSource(context, 1, (const char**)&kernelSource, (const size_t*)&kernelLength, &ret);
    ret = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL);
    if (ret != CL_SUCCESS)
    {
        size_t len;
        char buffer[8 * 1024];
        
        printf("Error: Failed to build program executable!\n");
        clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, sizeof(buffer), buffer, &len);
        printf("%s\n", buffer);
        goto FINISH;
    }
    
    kernel = clCreateKernel(program, "test", &ret);
    if(kernel == NULL)
    {
        puts("Kernel failed to create!");
        goto FINISH;
    }
    
    const size_t cOntentLength= sizeof(*pInputBuffer1) * 1024 * 1024;
    
    // 这里预分配的缓存大小为4MB，第一个参数是读写的
    memObj1 = clCreateBuffer(context, CL_MEM_READ_WRITE, contentLength, NULL, &ret);
    if(memObj1 == NULL)
    {
        puts("Memory object1 failed to create!");
        goto FINISH;
    }
    
    // 这里预分配的缓存大小为4MB，第一个参数是只读的
    memObj2 = clCreateBuffer(context, CL_MEM_READ_ONLY, contentLength, NULL, &ret);
    if(memObj1 == NULL)
    {
        puts("Memory object2 failed to create!");
        goto FINISH;
    }
    
    ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&memObj1);
    ret |= clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&memObj2);
    
    if(ret != CL_SUCCESS)
    {
        puts("Set arguments error!");
        goto FINISH;
    }
    
    // 以下为在主机端分配输入缓存
    pInputBuffer1 = malloc(contentLength);
    pInputBuffer2 = malloc(contentLength);
    
    // 然后对此工作缓存进行初始化
    for(int i = 0; i <1024 * 1024; i++)
        pInputBuffer1[i] = i + 1;
    
    memset(pInputBuffer2, 0, contentLength);
    
    // 然后分配输出缓存
    pOutputBuffer = malloc(contentLength);
    
    // 先将第一个参数的数据传入GPU端，以后就不去改动了
    ret = clEnqueueWriteBuffer(command_queue, memObj1, CL_TRUE, 0, contentLength, pInputBuffer1, 0, NULL, NULL);
    if(ret != CL_SUCCESS)
    {
        puts("Data transfer failed");
        goto FINISH;
    }
    
    int count = 5;  // 执行5次循环
    
    do
    {
        // 先将第二个参数传给GPU
        ret = clEnqueueWriteBuffer(command_queue, memObj2, CL_TRUE, 0, contentLength, pInputBuffer2, 0, NULL, NULL);
        if(ret != CL_SUCCESS)
        {
            puts("Data transfer failed");
            goto FINISH;
        }
        
        // 这里指定将总共有1024 * 1024个work-item
        ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, (const size_t[]){1024 * 1024}, NULL, 0, NULL, NULL);
        
        // 将结果拷贝给主机端
        ret = clEnqueueReadBuffer(command_queue, memObj1, CL_TRUE, 0, contentLength, pOutputBuffer, 0, NULL, NULL);
        
        // 做次同步，这里偷懒，不用wait event机制了～
        clFinish(command_queue);
        
        // 做校验
        const int newValue = 5 - count + 1;
        const int addition = (5 - count) * newValue / 2;
        for(int i = 0; i <1024 * 1024; i++)
        {
            if(pOutputBuffer[i] != i + 1 + addition)
            {
                puts("Result error!");
                break;
            }
        }
        
        // 最后，给第二个缓存初始化新数据
        for(int i = 0; i <1024 * 1024; i++)
            pInputBuffer2[i] = newValue;
    }
    while(--count > 0);
    
    
FINISH:
    
    /* Finalization */
    if(pInputBuffer1 != NULL)
        free(pInputBuffer1);
    if(pInputBuffer2 != NULL)
        free(pInputBuffer2);
    if(pOutputBuffer != NULL)
        free(pOutputBuffer);
    
    if(kernelSource != NULL)
        free(kernelSource);
    
    if(memObj1 != NULL)
        clReleaseMemObject(memObj1);
    if(memObj2 != NULL)
        clReleaseMemObject(memObj2);
    
    if(kernel != NULL)
        clReleaseKernel(kernel);
    
    if(program != NULL)
        clReleaseProgram(program);
    
    if(command_queue != NULL)
        clReleaseCommandQueue(command_queue);
    
    if(context != NULL)
        clReleaseContext(context);
    
    return 0;
}

__kernel void test(__global int *pInOut, __global int *pIn)
{
    int index = get_global_id(0);
    
    pInOut[index] += pIn[index];
}