pub fn topk_tile() -> usizeExpand description
Tile width dsa_topk_pools walks the pool axis in.
The block holds two tiles — the running best list and the candidate tile — each a
[f32, i32] pair, so a tile costs 16 × T bytes. The largest power of two under the
49,152 B ceiling is 3,072 → 2,048. Bigger is better (the walk is
O(P/T · log²T)), so this is the ceiling, not a taste.
🪤 Mirrored by dsa_write_geom’s tile argument, which this module passes explicitly
rather than duplicating as a #define, so the two cannot drift.