Why the cvtColor use a lot of CPU when process UMat in opencl gpu mode

edit

OpenCV version is 3.3. Windows 10.

My GPU is GTX 750, I'm sure the cvtColor is call in opencl mode, because I set

_putenv("OPENCV_OPENCL_DEVICE=:GPU:750");

and print the device name use follow code

{
        char        *value;
        size_t      valueSize;
        //print the device name
        clGetDeviceInfo((cl_device_id)cv::ocl::Device::getDefault().ptr(),
            CL_DEVICE_NAME, 0, NULL, &valueSize);
        value = (char*)malloc(valueSize);
        clGetDeviceInfo((cl_device_id)cv::ocl::Device::getDefault().ptr(),
            CL_DEVICE_NAME, valueSize, value, NULL);
        OutputDebugStringA("Device Name: ");
        OutputDebugStringA(value);
        OutputDebugStringA("\n");
        free(value);
}

the result is :

Device Name: GeForce GTX 750

and I'm sure cvtColor is call ocl_cvtColor to do the task.

when I call

cv::cvtColor(*nv12_frame_, *src_frame_, cv::COLOR_YUV420p2BGR);

30 times per second, the app CPU usage in task manager is 8. BUT, if use OpenCL API and kernel function directly, the CPU usage in task manger is almost 0.

I'm curious about why the cvtColor use a lot of CPU in opencl mode, but use opencl directly, the cpu usage is almost 0.