Intel Plugin for TorchCodec

This repo contains a prototype of the Intel Plugin for TorchCodec.

Prerequisites

To run:

• System with Intel Xe GPU capable to:
  • Run PyTorch* XPU backend
  • Run HW accelerated video decoding
• Linux operating system
• FFmpeg version 6 or later with enabled VAAPI hardware acceleration
• PyTorch* with enabled XPU backend
• TorchCodec version of v0.10.x series

Additionally, to build:

• CMake 3.18 or later
• Intel oneAPI matching the version used to build PyTorch

How to build

• Install oneAPI 2025.3

• Install FFmpeg with enabled VAAPI hardware acceleration. For example:

git clone https://git.ffmpeg.org/ffmpeg.git && cd ffmpeg
./configure \
  --prefix=$HOME/_install \
  --libdir=$HOME/_install/lib \
  --disable-static \
  --disable-stripping \
  --disable-doc \
  --enable-shared \
  --enable-vaapi
make -j$(nproc) && make install

• Build and install Intel plugin for TorchCodec:

git clone https://github.com/dvrogozh/torchcodec-xpu.git
cd torchcodec-xpu

# If ffmpeg was installed per above example
export PKG_CONFIG_PATH=$HOME/_install/lib/pkgconfig
export LD_LIBRARY_PATH=$HOME/_install/lib:$LD_LIBRARY_PATH

uv venv && uv pip install torch~=2.10.0 -e . \
  --index https://download.pytorch.org/whl/xpu -vv

How to use

Import the project in your Python script to register Intel device interface for the TorchCodec:

import torchcodec
import torchcodec_xpu

Before executing, make sure to export environment variables to point to FFmpeg installation:

export PATH=$HOME/_install/bin:$PATH
export LD_LIBRARY_PATH=$HOME/_install/lib:$LD_LIBRARY_PATH