Files
foxhunt/crates
jgrusewski 8c7e907fcf perf(cuda): eliminate GPU→CPU roundtrips from backtest evaluate loop
- gather_states: replace memcpy_dtoh + Tensor::from_vec with DtoD copy
  (cuMemcpyDtoDAsync) — state tensor stays on device, zero CPU touch
- actions: replace to_vec1 + memcpy_htod with DtoD copy from argmax
  tensor directly into actions_buf — eliminates per-step PCIe upload
- batch_q_values (RegimeConditionalDQN): replace CPU-side regime
  classification (to_vec2 + serial loop + sub-batch re-upload) with
  on-device classify_regime_masks_gpu + all-heads forward + masked blend

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-11 13:05:02 +01:00
..