trunk/137a3dbeb0e2632f8655c2452a1d3ebfef5218a6: [Native DSL] Register-resident topk for small K, small N (#184927)
🔒
https://github.com
«Adds a second cutedsl topk kernel covering K in {16, 32}, where the
radix kernel can't run (K>=64) and one-CTA-per-row's launch overhead
dominates anyway. One warp per row, register-resident keys, no smem
or global st...»
Automatische Weiterleitung...
1.5s