Repository navigation

Planar NEON kernels, >= 2x faster on Apple Silicon & Raspberry Pi, bit-identical to a2_fast - #313

Open

rikkus wants to merge 9 commits into

sdatkinson:mainfrom

rikkus:apple-silicon-a2-planar

Open

Planar NEON kernels, >= 2x faster on Apple Silicon & Raspberry Pi, bit-identical to a2_fast#313
rikkus wants to merge 9 commits into
sdatkinson:mainfrom
rikkus:apple-silicon-a2-planar

Commits on Sep 11, 2026
- Planar NEON kernels for the A2 fast path (AArch64), bit-identical
  rikkus
  committed
  62ab274
- Gate the planar kernels to Apple Silicon, and fix the tool's build off it
  rikkus
  committed
  cbacb76
- Widen the planar gate from Apple Silicon to AArch64
  rikkus
  committed
  6c2a821
- Widen the planar gate to 32-bit ARM, and re-tune the tiles for it
  
  rikkus
  and
  claude
  committed
  a2b71b7
- Give ARMv7 its own C=8 conv loop shape
  
  rikkus
  and
  claude
  committed
  99f8fff
- Summarise the four configurations at the top of the header
  
  rikkus
  and
  claude
  committed
  d3814ad
- Give the planar models the cached-prewarm path A2FastModel gained upstream
  
  rikkus
  and
  claude
  committed
  552c5ab
Commits on Sep 12, 2026
- Build the A2 kernels optimised for run_tests, so the parity test can hold
  
  rikkus
  and
  claude
  committed
  44412fa
Commits on Sep 25, 2026
- Merge branch 'main' of https://github.com/sdatkinson/NeuralAmpModelerCore into apple-silicon-a2-planar
  rikkus
  committed
  d01960c