Hi @DaneLLL, I think it is better to explain my problem:
I grab a frame from the camera, appsink, do some processing etc; then I want to send it to udpsink in H264 through appsrc and using nvv4l2enc.
Right now I’m using OpenCV to do the processing between appsink and appsrc, and using the CVVideoWriter with Gstreamer API, the main problem is latency and I identified it in the time it takes for the frame to being copy from memory to NVMM. I’m looking for a way to speed up this process.
I’m using Unified Memory so I have both a GpuMat, a Mat, and a pointer to the memory that is wrapped by OpenCV mats. None of my tries works after the first frame: copy operations (both Opencv copy that call cudaMemcpy2D or cudaMemcpy using the pointer) or color conversion (BGR->RGBA) directly on the GpuMat that wraps the EGLImage.
I think the problem is not in OpenCV functions but something about the feeding function and appsrc. I also tried without forcing a cuda context with cudaFree(0) because I’m not familiar with the Cuda Driver API but seems like the context “protects” the data.
Any way to solve this problem?
P.S.For my understanding 2 alternative are NVEnc from the jetson multimedia api or creating a eglstream with cuda Producer, but I’d like to use this code (for compatibility reasons).