Changes from 3.6.2 to 3.7.0
- Faster shuffle and bitshuffle. Speed up shuffle and unshuffle for 2-, 4-,
8-, and 16-byte elements on NEON, SSE2, and AVX2. Accelerate bitshuffle and
bitunshuffle with SIMD, including a bitshuffle fast path for constant
byte-plane prefixes. Enable fused SHUFFLE+BYTEDELTA by default and extend
it to SSSE3 and 2-byte groups. - BloscLZ 2.6.0. Improve match finding on periodic filtered data, reuse
worker-local hash tables, and enable decoder fast paths on ARM64 and AMD64. - Less worker setup and synchronization. Reuse compression state for
independent blocks with bundled LZ4, and claim the next block while reserving
compressed output to reduce scheduler lock acquisitions. - Full-write prefilters. The new
blosc2_set_prefilter_full_write()lets a
compression context skip zero-initializing prefilter output when the callback
writes every byte it returns, including padding. Existing prefilters still
receive zeroed output by default. B2ND slice gathering uses the opt-in for
fully overwritten blocks. - Safer chunk decoding. Validate the stored size of special-value chunks
against the supplied buffer even when the lazy flag is set (#881). Size
worker scratch buffers for both the block size and element width, and handle
allocation failures so later jobs can retry (#882).