Skip to content

GPU acceleration for images without restart markers #111

Open

Description

@simonedeiana

nvJPEG is able to encode JPEGs without restart markers almost completely on the GPU. Decoding is much more difficult to parallelize due to Huffman decoding being mostly serial. However, a research paper and software project JPEGGPU has discovered a way of making Huffman decoding more parallel on GPUs.

Since GPUJPEG falls back to CPU-only processing when no restart markers are present, it would be great to create a GPU path for this case too. Sure, it will not be as fast as using restart markers, but it would be a big improvement in interoperability with other pipelines and software not making use of this obscure feature.

I have forked this repo and used Codex to implement nvJPEG's and JPEGGPU's strategy for encoding and decoding, leading to massively better performance. Of course, the thing is extremely experimental and mostly AI slop, but I just wanted to show it was something possible to achieve in the future.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions