Skip to main content

PyAV extension with hardware encoding/decoding support on Nvidia GPUs.

Project description

PyAV-CUDA

PyPI version

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.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)

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

avcuda-0.2.0.tar.gz (9.8 kB view details)

Uploaded Source

File details

Details for the file avcuda-0.2.0.tar.gz.

File metadata

  • Download URL: avcuda-0.2.0.tar.gz
  • Upload date:
  • Size: 9.8 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/5.1.1 CPython/3.12.7

File hashes

Hashes for avcuda-0.2.0.tar.gz
Algorithm Hash digest
SHA256 e5094d0607abb16ce624a055a88d88694883d3d93385d51b4356bae7c54e5f75
MD5 6da04bd3d5758c343d4185d61b6a12d9
BLAKE2b-256 4057869826dbfec69cd0e4cbfcff1b2c369a2dc2640f0e709142ee906987d6f6

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page