Is it possible not to use cuGraphicsEGLRegisterImage for every frame?

[user43281]

I am using this code to map from nvsufracebuffer to opencv GPU mat

    for (int i = 0; i < surface->numFilled; i++) {
        // GST_DEBUG("NvBufSurfaceMapEglImage: %d", i);
        auto registerStart = std::chrono::high_resolution_clock::now();  
       status = cuGraphicsEGLRegisterImage(&resources[i], 
                surface->surfaceList[i].mappedAddr.eglImage, CU_GRAPHICS_MAP_RESOURCE_FLAGS_NONE);
       
        if (status != CUDA_SUCCESS) {
            GST_ERROR("cuGraphicsEGLRegisterImage failed: %d", status); 
            break;  
        }
        registered[i] = true; 

        // map egl image for every call 
        // GST_DEBUG("cuGraphicsResourceGetMappedEglFrame: %d", i);
        status = cuGraphicsResourceGetMappedEglFrame(&frames[i], resources[i], 0, 0);
        if (status != CUDA_SUCCESS) {
            GST_ERROR("cuGraphicsResourceGetMappedEglFrame failed: %d", status); 
            break; 
        }
        auto registerEnd = std::chrono::high_resolution_clock::now();
        auto registerDuration = std::chrono::duration_cast<std::chrono::microseconds>(registerEnd - registerStart);
        GST_INFO("cuGraphicsResourceGetMappedEglFrame took %ld microseconds", registerDuration.count());
 
        int stream_i = stream_idx.empty() ? i : stream_idx[i]; 
        assert(stream_i < surface->batchSize);
        mats[stream_i] = cv::cuda::GpuMat(frames[i].height, frames[i].width, CV_8UC4, 
                                   frames[i].frame.pPitch[0], frames[i].pitch);   
    }

I have 3 4k 30 fps streams and cuGraphicsEGLRegisterImage takes a pretty long time, is there a way to avoid calling cuGraphicsEGLRegisterImage for every frame and doing it only once at the start ?

[AastaLLL]

Hi,

Depends on your use case.

If the frame data uses the same location, you can do that only in the initialization and termination time.
Please give it a try to see if any issues.

Thanks.

[user43281]

Hi,
Thank you for your response.

I have a custom deepstream element, that gets the frames from the nvarguscamera element, I checked there it uses a nv buffer pool, I assume this could be the reason I was seeing a bit of weird behavior, frames updating slower, and sometimes with weird data.

[AastaLLL]

Hi,

Could you share the frame size and the frame updating elapsed time with us?
Does the perf improve without the custom Deepstream element?

Thanks.

[user43281]

Hi

This worked by some modifications of the nvarguscamera element to send ids with buffers from the buffer pool and registering each buffer only once.

My image size is 4032x3040, I had 3 streams at 30 fps,
On average registering each resource of this size was about 3.5ms, saving me about 10 milliseconds for each frame.

Thanks for the help.