glFinish hanging forever after clEnqueueReleaseGLObjects - opengl

Question

The following code runs fine (Windows 7, Nvidia GTX 750 Ti) with Nvidia drivers 361.91 (and earlier), but hangs with newer versions like 364.72 and 368.69. Now glFinish blocks the execution of the program only after clEnqueueReleaseGLObjects is invoked. Before blaming the drivers I'd suspect there is something wrong with how I do my OpenCL/OpenGL interop, so here is the code for a small entire program that reproduces the problem, the problem being at the very end:
#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;
}
In my much larger program (which had the exact same problem) everything ran perfectly until the driver update, and now even binaries compiled before the update show the same freeze demonstrated above. I removed the checks of return codes for the sake of readability but both in this small program and in the larger one it's from there were no reported problems at all. And I can't see anything obviously wrong that I might be doing...

Answer 1

Apparently my problem came from doing things backwards, as I try to get results from the previous frame before I enqueue things to do.
If on the first frame I skip that part and move on straight to enqueuing the first tasks then it doesn't block anymore.