glFinish在clEnqueueReleaseGLObjects后永远挂起

Question

以下代码在Windows 7上正常运行（使用Nvidia GTX 750 Ti），并且与Nvidia驱动程序361.91（以及早期版本）兼容，但是与新版本如364.72和368.69卡住。现在，glFinish仅在调用clEnqueueReleaseGLObjects之后阻止程序的执行。在指责驱动程序之前，我会怀疑我如何进行OpenCL / OpenGL交互存在问题，因此这里是一个小型完整程序的代码，可以重现问题，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;
}

在我更大的程序中（存在完全相同的问题），直到驱动程序更新之前，所有东西都运行得很完美。现在，即使是在更新之前编译的二进制文件也显示出上述相同的冻结。为了可读性，我删除了返回码检查，但无论是在这个小程序中还是在更大的程序中，都没有报告任何问题。我看不到我可能在做错的明显错误...

- Michel Rouzic

---

顺便提一下，从技术上讲，glFinish不保证在调用之后返回，除非先调用glFlush，以确保在“有限时间”内执行。只有少数供应商够恶心，不为每个finish做隐式刷新。而有缺陷的代码可能会一直运行得很好... - Andreas

我最终再次尝试了最新的驱动程序并添加了glFlush()，但遗憾的是它仍然挂起。 - Michel Rouzic

Answer 1

显然，我的问题来自于倒过来的操作，因为我在将要执行的任务入队之前尝试从上一帧获取结果。

如果在第一帧中跳过该部分并直接进入排队第一个任务，则不再阻塞。

- Michel Rouzic