feat(sp22-vnext): Phase A4 — aux_trade_outcome loss reduce kernel

K=3 sparse cross-entropy reduce over the trade-outcome softmax tile
produced by `aux_trade_outcome_forward` (Phase A3). Mirrors the K=2
sibling `aux_next_bar_loss_reduce` structurally: single-block shmem-tree
reduce, two parallel partial strips (loss_numer + valid_count) reduced
lockstep, fmaxf(p_tgt, 1e-30) numerical floor, fmaxf(valid, 1.0) all-
skip-batch guard, valid_count_out[1] save-for-backward.

Kept as SEPARATE kernel from the K=2 sibling:
- Diagnostic isolation (distinct HEALTH_DIAG slot, distinct cubin in
  profiles for clean per-loss-source attribution)
- Sparse-label semantic clarity (~95-99% mask=-1 vs ~50-100% valid for
  the K=2 next-bar head)
- Future per-class weighting headroom (Profit/Stop/Timeout 3:1-10:1
  imbalance will likely need class-weighted CE — surgical mod here
  without touching the K=2 head's contract)

Phase A4 (this commit) is dead code — no Rust launcher yet. Phase A5
lands backward; Phase B wires the full forward→loss→backward chain.

Discipline: feedback_no_atomicadd (single-block tree-reduce), feedback_
cpu_is_read_only (pure GPU), pearl_first_observation_bootstrap (sentinel
0 valid_count produces zero gradients gracefully on cold start).

Audit: docs/dqn-wire-up-audit.md Phase A4 section.
Cubin: aux_trade_outcome_loss_reduce_kernel.cubin (9.9 KB) compiles clean.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-05-13 23:41:59 +02:00
parent 07728f9efc
commit 3ddcfb8868
3 changed files with 188 additions and 0 deletions

View File

@@ -812,6 +812,15 @@ fn main() {
// `docs/plans/2026-05-14-sp22-h6-vNext-trade-outcome-aux.md` for
// the full vNext architecture spec.
"aux_trade_outcome_forward_kernel.cu",
// SP22 H6 vNext Phase A4 (2026-05-13): sparse cross-entropy reduce
// over the K=3 softmax tile from `aux_trade_outcome_forward`.
// Mirrors `aux_next_bar_loss_reduce`'s single-block shmem-tree
// pattern: mean over B_valid (skip label=-1 rows), fmaxf(p, 1e-30)
// numerical floor, saved valid_count for backward. Dead code until
// Phase A5 (backward) + Phase B (Rust launcher wireup) land. Kept
// separate from the K=2 sibling for diagnostic isolation, sparse-
// label semantic clarity, and future per-class weighting headroom.
"aux_trade_outcome_loss_reduce_kernel.cu",
// SP22 H6 (2026-05-12): per-env p_up extractor that copies
// `aux_softmax[env, 1]` → `prev_aux_dir_prob[env]` after each aux
// forward. The cache buffer is read by `experience_state_gather`