torchcodec-xpu 0.14.0

Intel extension for torchcodec

Intel Plugin for TorchCodec

Overview

TorchCodec is a high-performance Python library designed for media processing (decoding and encoding) using PyTorch* tensors. Intel® XPU plugin for TorchCodec enables hardware acceleration for video operations (only decoding at the moment) on Linux. Both TorchCodec and Intel® plugin rely on the FFmpeg libraries for their operations which must be pre-installed on the system. Intel® plugin further assumes that FFmpeg is built with the VAAPI support.

To use Intel® XPU plugin for TorchCodec, load it in the Python script and pass XPU device to initialize TorchCodec decoder or encoder:

import torchcodec
import torchcodec_xpu

decoder = torchcodec.decoders.VideoDecoder(
    "input.mp4", device="xpu:0")

Supported hardware

All the Intel GPU hardware enabled for XPU PyTorch backend with hardware media decoding capabilities is supported.

Installation

Pre-built release wheels are available at PyPI. Installation requires PyTorch with enabled XPU support which can be fetched from https://download.pytorch.org/whl/xpu:

pip install torchcodec-xpu \
  --extra-index-url https://download.pytorch.org/whl/xpu

Environment variables

The following environment variables can be used to customize the behavior of Intel Plugin for TorchCodec:

• USE_SYCL_KERNELS = on|off (default: off) - use SYCL kernels for augmentation such as color space conversion instead of VAAPI interface. If SYCL kernels are requested but can not be used due to hardware limitations, then fallback to VAAPI will be attempted.

• FORCE_CPU_FALLBACL=on|off (default: off) - force CPU fallback.

Known limitations

• Intel® Data Center GPU Max Series (Ponte Vecchio, PVC) GPUs are not supported due to missing hardware media engines. Execution will fallback to CPU.
• SYCL color space conversion kernel is not supported on Intel® Arc™ Pro A-Series Graphics (Alchemist, DG2) and Intel® Data Center GPU Flex Series (Archtic Sound, ATS-M) GPUs as 64-bit floating point operations used in the kernel are not available on these GPUs. Execution will fallback to less performant color space conversion on media engines via FFmpeg VAAPI.