C vs OpenCL，如何比较时间测量结果？

Question

因此，在另一篇文章中，我对 C 时间测量提出了质疑。现在，我想知道如何比较 C“函数”与 OpenCL“函数”的结果

这是主机OpenCL和C的代码

#define PROGRAM_FILE "sum.cl"
#define KERNEL_FUNC "float_sum"
#define ARRAY_SIZE 1000000


#include <stdio.h>
#include <stdlib.h>
#include <time.h>

#include <CL/cl.h>

int main()
{
    /* OpenCL Data structures */

    cl_platform_id platform;
    cl_device_id device;
    cl_context context;
    cl_program program;
    cl_kernel kernel;    
    cl_command_queue queue;
    cl_mem vec_buffer, result_buffer;

    cl_event prof_event;;

    /* ********************* */

    /* C Data Structures / Data types */
    FILE *program_handle; //Kernel file handle
    char *program_buffer; //Kernel buffer

    float *vec, *non_parallel;
    float result[ARRAY_SIZE];

    size_t program_size; //Kernel file size

    cl_ulong time_start, time_end, total_time;

    int i;
    /* ****************************** */

    /* Errors */
    cl_int err;
    /* ****** */

    non_parallel = (float*)malloc(ARRAY_SIZE * sizeof(float));
    vec          = (float*)malloc(ARRAY_SIZE * sizeof(float));

    //Initialize the vector of floats
    for(i = 0; i < ARRAY_SIZE; i++)
    vec[i] = i + 1;

    /************************* C Function **************************************/
    clock_t start, end;

    start = clock();

    for( i = 0; i < ARRAY_SIZE; i++) 
    {
    non_parallel[i] = vec[i] * vec[i];
    }
    end = clock();
    printf( "Number of seconds: %f\n", (clock()-start)/(double)CLOCKS_PER_SEC );

    free(non_parallel);
    /***************************************************************************/




    clGetPlatformIDs(1, &platform, NULL);//Just want NVIDIA platform
    clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &device, NULL);
    context = clCreateContext(NULL, 1, &device, NULL, NULL, &err);

    // Context error?
    if(err)
    {
    perror("Cannot create context");
    return 1;
    }

    //Read the kernel file
    program_handle = fopen(PROGRAM_FILE,"r");
    fseek(program_handle, 0, SEEK_END);
    program_size = ftell(program_handle);
    rewind(program_handle);

    program_buffer = (char*)malloc(program_size + 1);
    program_buffer[program_size] = '\0';
    fread(program_buffer, sizeof(char), program_size, program_handle);
    fclose(program_handle);

    //Create the program
    program = clCreateProgramWithSource(context, 1, (const char**)&program_buffer, 
                    &program_size, &err);

    if(err)
    {
    perror("Cannot create program");
    return 1;
    }

    free(program_buffer);

    clBuildProgram(program, 0, NULL, NULL, NULL, NULL);

    kernel = clCreateKernel(program, KERNEL_FUNC, &err);

    if(err)
    {
    perror("Cannot create kernel");
    return 1;
    }

    queue = clCreateCommandQueue(context, device, CL_QUEU_PROFILING_ENABLE, &err);

    if(err)
    {
    perror("Cannot create command queue");
    return 1;
    }

    vec_buffer = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                sizeof(float) * ARRAY_SIZE, vec, &err);
    result_buffer = clCreateBuffer(context, CL_MEM_WRITE_ONLY, sizeof(float)*ARRAY_SIZE, NULL, &err);

    if(err)
    {
    perror("Cannot create the vector buffer");
    return 1;
    }

    clSetKernelArg(kernel, 0, sizeof(cl_mem), &vec_buffer);
    clSetKernelArg(kernel, 1, sizeof(cl_mem), &result_buffer);

    size_t global_size = ARRAY_SIZE;
    size_t local_size = 0;

    clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global_size, NULL, 0, NULL, &prof_event);

    clEnqueueReadBuffer(queue, result_buffer, CL_TRUE, 0, sizeof(float)*ARRAY_SIZE, &result, 0, NULL, NULL);
    clFinish(queue);



     clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_START,
           sizeof(time_start), &time_start, NULL);
     clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_END,
           sizeof(time_end), &time_end, NULL);
     total_time += time_end - time_start;

    printf("\nAverage time in nanoseconds = %lu\n", total_time/ARRAY_SIZE);



    clReleaseMemObject(vec_buffer);
    clReleaseMemObject(result_buffer);
    clReleaseKernel(kernel);
    clReleaseCommandQueue(queue);
    clReleaseProgram(program);
    clReleaseContext(context);

    free(vec);

    return 0;
}

内核是:

__kernel void float_sum(__global float* vec,__global float* result){
    int gid = get_global_id(0);
    result[gid] = vec[gid] * vec[gid];
}

现在，结果是:

Number of seconds: 0.010000 <- This is the for the C code

Average time in nanoseconds = 140737284 <- OpenCL function

0,1407 秒是OpenCL 时间内核执行的时间，它比C 函数多，对吗？因为我认为 OpenCL 应该比 C 非并行算法最快...

Answer 1

在 GPU 上执行并行代码不一定比在 CPU 上执行更快。考虑到除了计算之外，您还必须将数据传入和传出 GPU 内存。

在您的示例中，您正在传输 2 * N 项并并行执行 O(N) 操作，这是对 GPU 的非常低效的使用。因此，很可能 CPU 对于这个特定的计算确实更快。