Question

我试图分析简单卷积的不同实现。我已经在不同的CPU（i5，xeon等等）上获得了几个结果，现在我正在通过intel beignet尝试使用intel HD4000。

我在主机端使用clock_gettime，在设备端使用CL_QUEUE_PROFILING_QUEUE和事件。代码的精简版本将是：

clock_gettime(CLOCK_REALTIME, &start);

err = clEnqueueNDRangeKernel(queue, img_conv_kernel, 2, NULL,
        &global_ws[0], &local_ws[0], 0, NULL, &event_clock);

if (err)
    die("can not launch kernel %d\n", err);

/* profiling */
clWaitForEvents(1, &event_clock);
clGetEventProfilingInfo(event_clock, CL_PROFILING_COMMAND_START, 
        sizeof(cl_ulong), &cl_start, NULL);
clGetEventProfilingInfo(event_clock, CL_PROFILING_COMMAND_END, 
        sizeof(cl_ulong), &cl_stop, NULL);

clock_gettime(CLOCK_REALTIME, &end);
printf("%f %f ", double) (cl_stop - cl_start) * 1e-6, 
            time_elapsed(start, end));

/* read data */
clock_gettime(CLOCK_REALTIME, &start);
err = clEnqueueReadBuffer(queue, res_d, CL_TRUE, 0, N*sizeof(float),
        res_h, 0, NULL, NULL);
clock_gettime(CLOCK_REALTIME, &end);

printf("%f ", time_elapsed(start, end));

/* C implementation */
clock_gettime(CLOCK_REALTIME, &start);
conv(img_data, res_h, &sobel_gx[0][0], k, k);
clock_gettime(CLOCK_REALTIME, &end);
printf("%f\n", time_elapsed(start, end));

结果是：

231.592960 16.701613 3.995006 151.874017
/* (device / host / reading-data / basic-c implementation )*/

我不明白内核执行时间实际上大于通过来确定 cpu-time clock_gettime ，但根据[0]我使用clWaitForEvents（）来确保内核完全执行。

[0]：https://software.intel.com/en-us/articles/intel-sdk-for-opencl-applications-performance-debugging-intro

Answer 1

请运行此代码＆amp;显示结果。

static long Time_Elapsed(
    long start,
    long end)
{
    return end - start;
}

static long Get_CL_Time(
    cl_event event)
{
    cl_ulong start, end;

    clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &start, NULL);     
    clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, NULL);

    return Time_Elapsed(start, end);
}

timespec start, end;
cl_event event_clock;

clock_gettime(CLOCK_REALTIME, &start);

/* run kernel */
err = clEnqueueNDRangeKernel(queue, img_conv_kernel, 2, NULL,
        &global_ws[0], &local_ws[0], 0, NULL, &event_clock);

clWaitForEvents(1, &event_clock);
long kernel_time = Get_CL_Time(event_clock);

/* read data */
err = clEnqueueReadBuffer(queue, res_d, CL_TRUE, 0, N*sizeof(float),
        res_h, 0, NULL, &event_clock);

clWaitForEvents(1, &event_clock);
long io_time = Get_CL_Time(event_clock);    

clock_gettime(CLOCK_REALTIME, &end);
long host_time = Time_Elapsed(start.tv_nsec, end.tv_nsec);

printf( "Kernel time: %l nanoseconds \n"
        "IO time:     %l nanoseconds \n"
        "Host time:   %l nanoseconds \n", 

        kernel_time,
        io_time,
        host_time);

/* C implementation */
clock_gettime(CLOCK_REALTIME, &start);
conv(img_data, res_h, &sobel_gx[0][0], k, k);
clock_gettime(CLOCK_REALTIME, &end);
host_time = Time_Elapsed(start.tv_nsec, end.tv_nsec);

printf("C implementation time: %l nanoseconds\n", host_time);

OpenCL clock_gettime vs内核分析：奇怪的结果

1 个答案: