opengl - glFinish 在 clEnqueueReleaseGLObjects 之后永远挂起

Question

以下代码在 Nvidia 驱动程序 361.91(及更早版本)中运行良好(Windows 7、Nvidia GTX 750 Ti)，但在 364.72 和 368.69 等较新版本中挂起。现在，只有在调用 clEnqueueReleaseGLObjects 后，glFinish 才会阻止程序的执行。在指责驱动程序之前，我怀疑我的 OpenCL/OpenGL 互操作方式有问题，因此这里是重现该问题的整个小程序的代码，问题出现在最后:

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

#include <SDL.h>
#include <gl/glew.h>
#include <SDL_opengl.h>
#include <gl/glut.h>

#pragma comment (lib, "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\lib\\x64\\OpenCL.lib")
#include <CL/cl.h>
#include <CL/cl_gl.h>

cl_int init_cl_context(cl_context *context, cl_command_queue *command_queue)
{
    cl_int i, ret, pf_index=-1;
    cl_platform_id  platform_id[16];
    cl_device_id    device_id[16];
    cl_uint     ret_num_platforms;
    cl_uint     ret_num_devices;

    ret = clGetPlatformIDs(sizeof(platform_id)/sizeof(*platform_id), platform_id, &ret_num_platforms);  // get all the platforms

    for (i=0; i<ret_num_platforms; i++)     // go through all the platforms
    {
        ret = clGetDeviceIDs(platform_id[i], CL_DEVICE_TYPE_GPU, sizeof(device_id)/sizeof(*device_id), device_id, &ret_num_devices);    // get all the suitable GPU devices

        if (ret_num_devices > 0)        // stop trying platforms when a suitable device is found
        {
            pf_index = i;
            break;
        }
    }

    cl_context_properties properties[] = { CL_GL_CONTEXT_KHR, (cl_context_properties)wglGetCurrentContext(), CL_WGL_HDC_KHR, (cl_context_properties)wglGetCurrentDC(), CL_CONTEXT_PLATFORM, (cl_context_properties)platform_id[pf_index], 0 };
    *context = clCreateContextFromType(properties, CL_DEVICE_TYPE_GPU, NULL, NULL, &ret);
    *command_queue = clCreateCommandQueue(*context, device_id[0], 0*CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE | 0*CL_QUEUE_PROFILING_ENABLE, &ret);

    return ret;
}

int main(int argc, char *argv[])
{
    cl_int ret=0;
    int w = 800, h = 600;
    SDL_Window *window;
    SDL_Renderer *renderer;
    cl_context context;
    cl_command_queue command_queue;
    cl_mem cltex;           // CL buffer of type image_2d_t pointing to the GL texture
    uint32_t gltex;         // ID of the GL texture for cltex

    //**** Init SDL, OpenGL/glew ****
    SDL_Init(SDL_INIT_TIMER | SDL_INIT_VIDEO | SDL_INIT_AUDIO);

    window = SDL_CreateWindow ("Title", SDL_WINDOWPOS_UNDEFINED, SDL_WINDOWPOS_UNDEFINED, w, h, SDL_WINDOW_OPENGL);
    SDL_GetWindowSize(window, &w, &h);

    SDL_GL_CreateContext(window);
    glewExperimental = 1; 
    glewInit();

    renderer = SDL_CreateRenderer(window, -1, 0*SDL_RENDERER_PRESENTVSYNC);
    //-------------------------------

    ret = init_cl_context(&context, &command_queue);    // initialise the CL context to match GL as to make the interop possible

    // create an OpenGL 2D texture normally
    glEnable(GL_TEXTURE_2D);
    glGenTextures(1, &gltex);                                   // generate the texture ID
    glBindTexture(GL_TEXTURE_2D, gltex);                                // binding the texture
    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, w, h, 0, GL_RGBA, GL_UNSIGNED_BYTE, 0);        // specify texture dimensions, format etc

    cltex = clCreateFromGLTexture(context, CL_MEM_WRITE_ONLY, GL_TEXTURE_2D, 0, gltex, &ret);   // Creating the OpenCL image corresponding to the texture (once)

    ret = clFinish(command_queue);
    //glFinish();                  // this works fine
    ret = clEnqueueReleaseGLObjects(command_queue, 1, &cltex, 0, 0, NULL);          // release the ownership from CL back to GL
    clFinish(command_queue);
    glFlush();

    printf("This blocks the execution forever:\n");
    glFinish();                   // this blocks everything
    printf("This never gets printed\n");

    return 0;
}

在我的更大的程序中(有完全相同的问题)，在驱动程序更新之前一切都运行得很好，现在甚至在更新之前编译的二进制文件也显示出与上面演示的相同的卡住。为了可读性，我删除了返回码的检查，但是在这个小程序和它所在的较大程序中，根本没有报告任何问题。我看不出我可能做的任何明显错误的事情......

Answer 1

显然我的问题来自于向后做事，因为我尝试在将要做的事情排队之前从前一帧中获取结果。

如果在第一帧上我跳过该部分并直接将第一个任务排入队列，那么它就不会再阻塞。