ssimulacra2-cuda
| Crates.io | ssimulacra2-cuda |
| lib.rs | ssimulacra2-cuda |
| version | 0.1.0 |
| source | src |
| created_at | 2024-10-12 13:07:34.647204 |
| updated_at | 2024-10-12 13:07:34.647204 |
| description | Ssimulacra2 implementation running on CUDA |
| homepage | https://github.com/Gui-Yom/turbo-metrics |
| repository | https://github.com/Gui-Yom/turbo-metrics |
| max_upload_size | |
| id | 1406487 |
| size | 89,658 |
(Gui-Yom)
documentation
README
ssimulacra2-cuda
An implementation of ssimulacra2 using CUDA.
Features
- Close to the original implementation, and with close results.
- Leverages many custom kernels written in Rust and a few CUDA NPP primitives.
- Uses CUDA graphs to alleviate the cost of launching the 200+ kernels per image pair.
TODO
- Investigate if it is possible to change some computations to accelerate processing without deviating from the original implementation too much. Maybe making it configurable.
- More custom kernels, is it possible to run the whole computation in a single fused kernel launch ?
- Use less memory (currently 500MB for 1080p), might be possible by using a single fused kernel.
Credits
Original reference implementation : https://github.com/cloudinary/ssimulacra2
With inspiration from : https://github.com/rust-av/ssimulacra2