◐ Shell
clean mode source ↗

GitHub - doterax/guetzli-cuda-opencl: Perceptual JPEG encoder, optimized with CUDA&OpenCL, full JPEG format support.

Guetzli

CI

GPU-accelerated JPEG encoder based on Google Guetzli. Guetzli generates JPEG images that are typically 20-30% smaller than libjpeg at equivalent visual quality.

This fork adds CUDA and OpenCL support to dramatically speed up encoding:

Method Example (750×400)
Original CPU 14.7 s
Procedure Optimised (--c) 8.2 s
OpenCL (--opencl) 1.5 s
CUDA (--cuda) 0.8 s

The binary auto-detects the best available backend at runtime. On systems without an NVIDIA GPU, CUDA is simply skipped — no driver required to start the program.

Quick Start

# Build dependencies (one-time)
make external

# Build
make CONFIG=release

# Run
guetzli [--quality Q] [--verbose] input.png output.jpg

See BUILD.md for full build instructions, prerequisites, and CI/CD documentation.

Usage

guetzli [--quality Q] [--verbose] original.png output.jpg
guetzli [--quality Q] [--verbose] original.jpg output.jpg   # requires FULL_JPEG
guetzli [--quality Q] [--verbose] original.tiff output.jpg

You can also force a specific backend:

guetzli --cuda   input.png output.jpg   # Force CUDA
guetzli --opencl input.png output.jpg   # Force OpenCL
guetzli --c      input.png output.jpg   # CPU with procedure optimisation
guetzli --auto   input.png output.jpg   # Auto-detect best backend (default)

Notes:

  • Guetzli uses ~300 MB of memory per 1 MPix of input.
  • Guetzli assumes sRGB input with a gamma of 2.2.
  • Prefer uncompressed (PNG/TIFF) input for best results.
  • JPEG alpha channels are not supported; PNG alpha is composited on black.

Supported Platforms

Platform CUDA OpenCL Tested
Windows x64 (MinGW-w64 / MSYS2) ✅ ✅ CI
Linux x64 (GCC) ✅ ✅ CI
macOS arm64 (Apple Silicon) — ✅ CI

Building

Step Command
Build external deps make external
Release build make CONFIG=release
Debug build make CONFIG=debug
OpenCL-only make FEATURES="OPENCL FULL_JPEG"
Run tests make test
Show help make help

Full details (prerequisites, GitHub Actions setup, troubleshooting) are in BUILD.md.

GitHub Actions CI/CD

Every pull request is built and tested on Linux, Windows (MSYS2), and macOS. Pushing a v* tag (e.g. v1.3.0) creates a GitHub Release with pre-built binaries for all platforms.

See BUILD.md § GitHub Actions CI/CD for setup instructions.

Project Structure

guetzli/          Core JPEG optimiser (C++)
clguetzli/        GPU acceleration — CUDA & OpenCL kernels + host code
third_party/      Bundled: butteraugli, minilzo, OpenCL headers
external/         CMake superbuild for zlib, libpng, libtiff, libjpeg-turbo,
                  OpenCL ICD loader
tests/            Smoke-test suite + sample images
.github/          CI/CD workflow

Contributing

  1. Fork and create a feature branch.
  2. Ensure make test passes on at least one platform.
  3. Open a pull request — CI will build on all three platforms.
  4. See CONTRIBUTING.md for details.

License

Apache 2.0 — see LICENSE.

Acknowledgements

  • Google Guetzli — original JPEG encoder.
  • Original CUDA/OpenCL acceleration by strongtu, ianhuang, chriskzhou, stephendeng @ Tencent.