j2k-transcode-cuda
CUDA acceleration adapter for supported stages in J2K's JPEG-to-HTJ2K coefficient-domain transcode pipeline.
This crate accelerates supported transform and code-block preparation stages.
It does not replace the transcode API in j2k-transcode.
Enable the cuda-runtime feature for CUDA Driver API dispatch. Default builds
expose the API surface but do not require a CUDA runtime.
cuda-runtime is not proof that every CUDA Oxide kernel was built on the local
host. Product PTX is generated only on supported Linux cuda-oxide build hosts;
other builds may embed placeholder PTX. Set J2K_REQUIRE_CUDA_OXIDE_BUILD=1
on CUDA validation and benchmark hosts to fail the build when PTX is missing.
Runtime errors for placeholder kernels state that CUDA Oxide PTX was not built.
Auto routing is conservative. Single transform jobs are offered only after the
shared 224 * 224 component-sample floor, while same-shape reversible 5/3 and
9/7 batches use the same 32-job / 224 * 224 * 32 floors as the Metal adapter.
Callers with local evidence can lower the batch floors with
with_auto_reversible_batch_thresholds or with_auto_dwt97_batch_thresholds;
the defaults are routing policy, not a speedup promise.
Links
- API docs: https://docs.rs/j2k-transcode-cuda
- Repository: https://github.com/frames-sg/j2k
- Support policy: https://github.com/frames-sg/j2k/blob/main/docs/public-support.md