Repository navigation

Add byte compress and expand operations - #395

Merged

Shnatsel merged 10 commits into

linebender:mainfrom

Novum:compress-expand

Oct 3, 2026

Commits on Sep 30, 2026

Add portable SIMD compress and expand operations

Novum
authored and
Shnatsel
committed
681b1cd
Route wider compress/expand ops through byte-level ones, so that all SimdBase vectors can implement compress/expand
Shnatsel
committed
261407e
Simplify mask control for a 13-28% measured speedup on zen5 in AVX2 mode
Shnatsel
committed
0b4639a
Simplify mask control on NEON and WASM
Shnatsel
committed
92fe110
Implement a dedicated 128-bit compress/expand path for wider elements for 1.3x to 1.8x speedup over the byte variant
Shnatsel
committed
6b27eee
Implement a specialized AVX2 codepath for expand/compress for 32 bit values and larger that is 3x faster than the generic byte codepath
Shnatsel
committed
12e56ed
Extend the AVX2 optimization for 32 bits and up to 512-bit wide vectors, same 3x gains seen there as for the 256-bit wide vectors
Shnatsel
committed
e187a6b
Drop redundant test
Shnatsel
committed
bf7d734

Commits on Oct 3, 2026