Skip to content

J2K 0.13.0

Choose a tag to compare

@jcwal1516 jcwal1516 released this 26 Oct 18:08
· 444 commits to main since this release
5012426

J2K 0.13.0 publishes all 25 workspace crates and requires Rust 1.99.0. The shared kernel crates j2k-types and j2k-codec-math retain Rust 1.96 support.

CUDA and Metal device decodes now follow CPU sample conversion for signed and deeper-than-8-bit sources, fixing #135. JPEG decode verifies pixel integrity by default, j2k_jpeg::JpegBatchDecoder decodes complete JPEGs into dense RGB8 batches, and j2k-ml and the Python extension accept JPEG and JPEG 2000 inputs in one batch. The Python extension can return PyTorch mps tensors on Apple Silicon. Classic CPU encode is faster, and CUDA HTJ2K encode and region decode batch and bound their GPU work more tightly. Encoded bytes and decoded pixels are unchanged. See the changelog and migration notes for details.

Security: j2k-tilecodec now requires zstd 0.14.3, which fixes RUSTSEC-2026-0391, RUSTSEC-2026-0398, RUSTSEC-2026-0403 and RUSTSEC-2026-0410. ZstdCodec only uses the streaming read decoder, which the first three do not affect. RUSTSEC-2026-0410 affects wasm32 builds that decode untrusted Zstandard tiles, so wasm32 users of j2k-tilecodec should upgrade. zstd 0.14 is licensed BSD-3-Clause.

Migration: J2kDirectGrayscalePlan adds a required signed field. BatchDecodeOptions, j2k_jpeg::ColorTransform and the Burn tensor and error enums are #[non_exhaustive]. BatchLayout moves to j2k-core and is still re-exported by j2k. JPEG decode returns JpegError::RecoveryRequired for damaged input unless you opt into recovery with DecodeOptions::with_strict(false). The cuda-oxide-j2k-ml feature of j2k-cuda-j2k-engine is removed.

Hosted validation and Metal hardware validation passed for the tagged source. The release verifier checked the four attached T.803 reports: CPU on Linux x86-64, macOS arm64, and Windows x86-64, plus Metal on macOS.

The CUDA full validation job and CUDA T.803 report are waived for this release only, because the self-hosted CUDA runner is offline. The waiver records the CUDA suites, Python CUDA tests and byte-identity checks that passed on the same crate sources. The CUDA T.803 report is pending and will be attached here once the runner is back.

The public API comparison and review describe compatibility with 0.12.0.