Support for gpu-accelerated decoding?

[morphogencc]

Does avcpp support gpu-accelerated decoding/encoding via nvenc?

[teodly]

I've tested hardware decoding with h264_cuvid. It works. You need to create codec context using something like this:

av::VideoDecoderContext vdec(stream, av::findDecodingCodec("h264_cuvid"))

[tehkillerbee]

It works with Nvidia GPU decoding, if you build ffmpeg with support for it. I have not tested with encoding, however.
https://docs.nvidia.com/video-technologies/video-codec-sdk/ffmpeg-with-nvidia-gpu/
https://trac.ffmpeg.org/wiki/HWAccelIntro

[teodly]

Hardware-accelerated encoding also works (tested with nvenc). You may need to set low-level libavcodec's fields: hw_device_ctx and hw_frames_ctx using av::VideoDecoderContext::raw() accessor.

Transcoding without raw frames leaving GPU's internal RAM (VRAM) is also possible if hw_frames_ctx->format is set to appropriate hardware "pixel format", e.g. AV_PIX_FMT_CUDA. It makes transcoding less PCIe-bandwidth-intensive. I can provide more details if anyone is interested in it.

[h4tr3d]

@teowoz , thanks for answer! I am so happy if you send a PR with minimal sample for nvenc usage. And will thinking about proper wrapping with high level code...

[teodly]

Apologies for the delay, now the backend app we are using avcpp in is open source, so I can provide a well-tested example, working on both CUDA and VAAPI:

https://github.com/amagimedia/avplumber/blob/fbaf51b8387c9a58efa84f691ac365c2cfc586e9/src/nodes/decoders.cpp#L81

(note the #if 0 // not really needed, the actual code is simpler than it looks)