openresearch
|
3cbf5a2560
|
feat(gpu): cross-enriched chiral Sidon filter — 4^8=65K configs on GPU
Cross-enrichment: each strand can have MULTIPLE chiral types
contributing simultaneously. With 4 ChiralLabel types × 8 strands
= 4^8 = 65,536 configurations (vs 2^8=256 from binary swaps).
This is why GPU is needed: 65K configs × pairwise quaternion product
checks = millions of operations.
WGSL shader uses:
- ChiralLabel types from BraidStateN.lean (achiral/scarred/left/right)
- Rossby drift weights (left=+65536, right=-65536, scarred=+32768, achiral=0)
- Quaternion basis from HopfFibration.lean (1,i,j,k)
- Golden angle winding mod 28 (HopfFibration.lean)
- COUCH gate as cheap pre-filter (scarred count → self-loop → threshold)
Pipeline: COUCH filter (cheap, O(1)) → Quaternion Sidon filter
(expensive, O(n²)) — only COUCH-passing configs reach the Sidon check.
256 threads/workgroup × 256 workgroups = 65,536 threads = all configs
checked in ONE dispatch. Zero-copy via existing dna_radix_gpu.py
infrastructure.
|
2026-07-04 21:04:21 +00:00 |
|