Skip to content

j2k 0.10.0

Choose a tag to compare

View all tags
@jcwal1516 jcwal1516 released this 06 Sep 18:39
· 47 commits to main since this release
v0.10.0
e19fccf

  • Adds the experimental j2k-mpsgraph crate for Apple Silicon macOS 11+.
    Completed resident batches and pipelined direct decode feed static rank-four
    MPSGraph programs without decoded-pixel readback/re-upload; blocking,
    nonblocking, reference-oracle, benchmark, package, and release validation
    paths are included. No zero-copy or speed claim is made.

  • Breaking: narrows j2k-cuda-runtime to codec-neutral CUDA allocation,
    launch, completion, and diagnostics. CUDA JPEG 2000, JPEG, and transcode
    ownership now lives in the published implementation crates
    j2k-cuda-j2k-engine, j2k-cuda-jpeg-engine, and
    j2k-cuda-transcode-engine; callers of
    j2k_cuda_runtime::transcode_kernels_built must use
    j2k_cuda_transcode_engine::transcode_kernels_built.

  • Breaking: generalizes
    j2k_transcode_metal::resident_codestream_buffer_from_metal_encoded_j2k
    from &j2k_metal::MetalEncodedJ2k to any &impl DeviceCodestream. Existing
    MetalEncodedJ2k calls continue to type-check, while custom device
    codestreams implement j2k_core::DeviceCodestream; this removes the
    transcode adapter's dependency on the full public Metal adapter.

  • Moves stable codec contracts, prepared-plan geometry, capability rejection,
    host allocation phases, packing, and encode geometry into single typed
    owners. Native produces typed plans consumed by CPU, CUDA, and Metal; the
    adapters no longer downcast erased plans or duplicate resource policy.

  • Decomposes the native color, Metal JPEG/JPEG 2000, CUDA, and transcode
    orchestration roots by pipeline ownership, with repository policies enforcing
    dependency direction, source boundaries, unsafe inventory, clone ceilings,
    benchmark registration, and generated routing evidence.

  • Promotes measured staged CUDA JPEG encoding and adaptive CUDA JPEG checkpoint
    launch geometry. The staged encoder improves the representative 512 x 512
    batch-8 cells by about 95%, while the checkpoint policy uses one-thread
    launches below 128 checkpoints and 128-thread blocks at or above that
    threshold.

  • Records and removes rejected Metal and CUDA candidates whose product
    intervals did not qualify, including terminal IDWT/store specialization,
    wide-axis and FDWT staging, RGB input fusion, packetization, terminal
    column/quantization fusion, and JPEG coefficient/IDCT defusion. No rejected
    selector, fallback, or benchmark-only production branch remains.

  • Repairs ties-to-even irreversible color rounding in Metal and CUDA, the JPEG
    restart-marker bit-reader boundary, CUDA restart-coded capability routing,
    clean packaged CUDA-Oxide sibling resolution, and lane-specific changed-line
    coverage orchestration.

  • Refactors the CPU JPEG AVX2 and NEON paths around unforgeable capability
    tokens and safe kernel entry points. SIMD dispatch, IDCT/color arithmetic,
    tests, and benchmark wrappers no longer require unsafe calls; the remaining
    raw loads/stores and exact-AVX2 bridge are isolated behind fixed-size safe
    interfaces. Ten production SIMD unsafe blocks remain, with no unsafe fn,
    under an AST-enforced 24-block cap.

  • Uses fearless_simd 0.7.0 for AArch64 NEON kernels while preserving the
    existing AVX2-only x86 acceleration envelope with a private exact-AVX2 token;
    fearless_simd::Avx2 is intentionally not used because its 0.7 contract is
    the broader x86-64-v3 feature set.

  • Adds deterministic CPU JPEG decode and expanded SIMD microbenchmarks covering
    grayscale, 4:4:4, 4:2:2, full/odd/cropped 4:2:0, row streaming, specialized
    IDCT, tail widths, and unaligned source rows.

All 24 crates in the tagged release manifest are available on crates.io. This GitHub release records the existing v0.10.0 distribution; later main-branch and PR changes are not included.

Assets 2
Loading