Add swizzle_dyn_precise, mirroring swizzle_dyn from `std::simd` - #276

Merged

Shnatsel merged 26 commits into

linebender:mainfrom

Shnatsel:swizzle-dyn-precise

Aug 4, 2026

Commits on Jul 23, 2026

Initial pass at swizzle_dyn, ported from std::simd
Shnatsel
committed
b24a2f2
Add a 256-bit SSE4.2 implementation as an experiment
Shnatsel
committed
54d6760
Split out recursive swizzle formulation into a helper and apply it to AVX2
Shnatsel
committed
29262e1
Implement a dedicated AVX2 512-bit shuffle path
Shnatsel
committed
51113e8
Apply the or trick to the dedicated AVX2 512-bit implementation to see what happens
Shnatsel
committed
e037904
Revert "Apply the or trick to the dedicated AVX2 512-bit implementation to see what happens"
Shnatsel
committed
6fd5d86
More detailed comment on the AVX2-optimized shuffle
Shnatsel
committed
c79bf96
Reuse the recursive decomposition for WASM 256-bit case
Shnatsel
committed
e1e605d
Add random testing for swizzle_dyn_precise correctness, since it's precise and we can assert it has to always match the scalar fallback
Shnatsel
committed
1ff3d16
Revert "Add random testing for swizzle_dyn_precise correctness, since it's precise and we can assert it has to always match the scalar fallback" to avoid adding extra dependencies
Shnatsel
committed
1d95366
Use an unwrap instead of a cast to appease clippy
Shnatsel
committed
e0a3225
Fix typo
Shnatsel
committed
710825b
Remove 512-bit shuffle specialized for AVX2 and route through the generic split/combine since there's no difference on benchmarks on real Zen2 hardware, which was the best-case scenario for this im…
Shnatsel
committed
79de6c0
Implement a custom AVX2 shuffle instead of reusing the std::simd one. llvm-mca results: Haswell, Broadwell and Skylake effectively unchanged for the 256-bit case; Zen1, Tiger Lake show 23% improvem…
Shnatsel
committed
8aee113
Empty commit to re-run stuck CI
Shnatsel
committed
ab3fd43

Commits on Jul 24, 2026

Merge branch 'main' into swizzle-dyn-precise
Shnatsel
committed
19a1ed9

Commits on Jul 25, 2026

Optimize SSE4.2 as suggested by axnsan12. Changes the control sequence from vpcmpgtb + vpblendvb to vpaddb + vpor
Shnatsel
committed
3e3a00c

Commits on Jul 28, 2026

Commits on Jul 30, 2026

Use a faster formulation of swizzle_dyn_precise on AVX-512 VBMI.
Shnatsel
committed
af454fe

Commits on Aug 4, 2026