PyAV-CUDA

PyAV-CUDA is an extension of PyAV that adds support for hardware-accelerated video decoding using Nvidia GPUs. It integrates with FFmpeg and PyTorch, providing CUDA-accelerated kernels for efficient color space conversion.

Installation

1. Build and install FFmpeg with hardware acceleration support.

2. To enable hardware acceleration in PyAV, it needs to be reinstalled from source. Assuming FFmpeg is installed in /opt/ffmpeg, run:

   pip uninstall av
   PKG_CONFIG_LIBDIR="/opt/ffmpeg/lib/pkgconfig" pip install av --no-binary av --no-cache
   

   If the installation was successful, h264_cuvid should appear between the available codecs:

   import av
   print(av.codecs_available)
   

3. Install PyAV-CUDA:

   PKG_CONFIG_LIBDIR="/opt/ffmpeg/lib/pkgconfig" CUDA_HOME="/usr/local/cuda" pip install avcuda
   

4. Test the installation by running python examples/benchmark_decode.py. The output should show something like:

   Running CPU decoding... took 34.99s
   Running GPU decoding... took 8.30s
   

Usage

Decoding

import av
import avcuda

CUDA_DEVICE = 0

with av.open("video.mp4") as container:
    stream = container.streams.video[0]
    avcuda.init_hwcontext(stream.codec_context, CUDA_DEVICE)

    for avframe in container.decode(stream):
        frame_tensor = avcuda.to_tensor(avframe, CUDA_DEVICE)

Encoding

import av
import avcuda

CUDA_DEVICE = 0

NUM_FRAMES = 100
FPS = 30
WIDTH = 640
HEIGHT = 480

with av.open("video.mp4", "w") as container:
    stream = container.add_stream("h264_nvenc", rate=FPS)
    stream.pix_fmt, stream.width, stream.height = "yuv420p", WIDTH, HEIGHT

    avcuda.init_hwcontext(stream.codec_context, CUDA_DEVICE)

    for _ in range(NUM_FRAMES):
        frame_tensor = torch.randint(0, 255, (HEIGHT, WIDTH, 3), dtype=torch.uint8, device=CUDA_DEVICE)
        avframe = avcuda.from_tensor(frame_tensor, stream.codec_context) 

        for packet in stream.encode(avframe):
            container.mux(packet)