Files
foxhunt/crates
jgrusewski f0b8044b4a refactor: replace segment tree with prefix sum buffers, rewrite update_priorities_gpu
- Replace segment tree fields (seg_tree, capacity_pow2) with prefix sum
  buffers (priorities_pa, prefix_sum, scan_tile_state/aggregate/prefix)
- Replace 5 seg_tree kernel fields with 4 PER prefix sum kernels
  (per_update_pa, per_insert_pa, per_prefix_scan, per_sample)
- Delete CastKernels infrastructure (struct, OnceLock, get_cast_kernels)
  — per_update_pa reads bf16 td_errors directly, no cast needed
- Delete rebuild_tree method (no tree to rebuild)
- Delete update_td_f32 scratch buffer (bf16→f32 conversion eliminated)
- Remove all PER_DIAG diagnostic eprintln calls

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-06 14:32:09 +02:00
..