Repository navigation

Prevent temporary SIMD allocations in Panama vector kernels - #755

Open

tlwillke wants to merge 29 commits into

mainfrom

fix-pq-vector-allocations

Open

Prevent temporary SIMD allocations in Panama vector kernels#755
tlwillke wants to merge 29 commits into
mainfrom
fix-pq-vector-allocations

Commits on Oct 2, 2026
- Keep concrete vector types in SIMD PQ diversity scoring
  tlwillke
  committed
  79dc2a2
- Specialize NVQ vector conversions to avoid temporary arrays
  tlwillke
  committed
  31b411e
Commits on Oct 3, 2026
- Keep concrete vector types in PQ ADC scoring
  tlwillke
  committed
  51adb5b
- Keep concrete vector types through NVQ decoding and dot products
  tlwillke
  committed
  e0af68f
- Keep concrete vector types in eight-coordinate squared distance
  tlwillke
  committed
  0161b64
- Avoid temporary vector allocations in PQ cosine gather chains
  tlwillke
  committed
  2e74b8f
- Keep NVQ encoding and scoring vector chains concrete
  tlwillke
  committed
  02b9107
Commits on Oct 13, 2026
Commits on Oct 14, 2026
- Fix older JDK gathers and portable SIMD regression checks
  tlwillke
  committed
  0dc6a17
- Widen binary compatibility self-query candidate budget
  tlwillke
  committed
  c2c7bd0
- Compare compatibility search scores numerically
  tlwillke
  committed
  3270244
- Preserve SIMD species and native storage compatibility
  tlwillke
  committed
  2aae57a
- Verify NVQ byte conversions across fixed and MAX species
  tlwillke
  committed
  8623080
- Merge remote-tracking branch 'origin/main' into fix-pq-vector-allocations
  tlwillke
  committed
  0c1042b
Commits on Oct 15, 2026
Commits on Oct 16, 2026
- Hoist backing-array access out of ADC SIMD loops
  tlwillke
  committed
  1234e77
- Separate ADC array access from SIMD arithmetic
  tlwillke
  committed
  12d30cb
- Initialize native library in inherited-kernel tests
  tlwillke
  committed
  0b4fd47
- Keep PQ cosine SIMD kernels small enough to inline
  tlwillke
  committed
  b47eef1
- Keep concrete cosine accumulators in a compact 512-bit loop
  tlwillke
  committed
  b68abca
- Preserve native-backed gathers in the compact cosine kernel
  tlwillke
  committed
  0a44033
- Preserve compact typed 512-bit cosine gathers across providers
  tlwillke
  committed
  0f92ea6

Navigation Menu

Repository navigation

Prevent temporary SIMD allocations in Panama vector kernels - #755

Prevent temporary SIMD allocations in Panama vector kernels#755
tlwillke wants to merge 29 commits into
mainfrom
fix-pq-vector-allocations

Commits on Oct 2, 2026

Keep concrete vector types in SIMD PQ diversity scoring

Specialize NVQ vector conversions to avoid temporary arrays

Commits on Oct 3, 2026

Keep concrete vector types in PQ ADC scoring

Keep concrete vector types through NVQ decoding and dot products

Keep concrete vector types in eight-coordinate squared distance

Avoid temporary vector allocations in PQ cosine gather chains

Keep NVQ encoding and scoring vector chains concrete

Commits on Oct 13, 2026

Merge remote-tracking branch 'origin/main' into fix-pq-vector-allocations

Keep eight-coordinate dot-product SIMD chains concrete

Keep shared vector arithmetic concrete and resolve arrays once

Preserve concrete SIMD types across distance and reduction kernels

Cover SIMD kernel computations and native inherited arithmetic

Add release note for Panama allocation fixes

Commits on Oct 14, 2026

Fix older JDK gathers and portable SIMD regression checks

Widen binary compatibility self-query candidate budget

Compare compatibility search scores numerically

Preserve SIMD species and native storage compatibility

Verify NVQ byte conversions across fixed and MAX species

Merge remote-tracking branch 'origin/main' into fix-pq-vector-allocations

Commits on Oct 15, 2026

Keep Panama gather and reduction kernels within HotSpot inlining budgets

Keep Panama gather and NVQ intermediates inside compiled kernels

Merge remote-tracking branch 'origin/main' into HEAD

Commits on Oct 16, 2026

Hoist backing-array access out of ADC SIMD loops

Separate ADC array access from SIMD arithmetic

Initialize native library in inherited-kernel tests

Keep PQ cosine SIMD kernels small enough to inline

Keep concrete cosine accumulators in a compact 512-bit loop

Preserve native-backed gathers in the compact cosine kernel

Preserve compact typed 512-bit cosine gathers across providers