Skip to content

STATUS.md

Latest commit

 

History

History
48 lines (39 loc) · 2.42 KB

STATUS.md

File metadata and controls

48 lines (39 loc) · 2.42 KB
Raw

Runtime — status

Reconciled 2026-08-10 (S50). Code: crates/mapal-rt/src/lib.rs. Gate: 21 tests in-file, plus every workspace suite that runs a compiled program.

Built

  • Task DAG execution: register → seal → launch → wait → finish
  • Deps, pinning, host-spine pinned execution, trap flag, watermarks
  • Compile-time slice sizing honoured (slice_elems, oversub, width)
  • Work stealing, and the heterogeneity tolerance it buys (S33)
  • Per-lane wake — a dispatch wakes only the lanes it filled (S50 P0b)
  • Lock-free completion — execute takes no lock; every slice but the last is one atomic (S50 P2)
  • Host spin before parking, budget zero when oversubscribed (S50 P1)
  • One executor per lane, host included (S50 P1)
  • One-slice tasks run inline, never queued (S50 P3)

Not built / refuted

  • Spinning workers — REFUTED, do not retry as specified. 0.80–0.93× on real shapes; handoff_floor over-predicted because it gave every thread work every wave. See performance/s50c-p1-spinning-workers-refuted.md.
  • Persistent kernel / static per-lane schedule with a counter barrier — the structural end state; not designed.
  • Work still carries an Arc<Run> (one atomic pair per slice); take_any's miss path still locks every lane.
  • path_plan_is_acyclic in the backend is still load-bearing — path_plan does not guarantee a DAG for every shape, only for the sibling-merge class S49 fixed.

Needs work

Item Where Note
raptorlake dispatch economics backend-llvm/profile.rs owed; box unreachable at S50 close, rule inert there
dispatch_ns tracks the RUNTIME, not the machine same re-derive whenever the handoff changes; recipe in performance/s50e-*.md §2
ARCHITECTURE §4 rules have no automated pin here rules 1–6 are prose; rule 1 was violated and only found by a spinning host

Where to dig

Question Doc
what a dispatch costs, and why performance/s49b-the-dispatch-price.md
the nested-run deletion performance/s50-p0-the-nested-runs.md
the shared lock performance/s50b-p2-the-shared-lock.md
spinning, refuted performance/s50c-p1-spinning-workers-refuted.md
the dispatch threshold performance/s50d-*.md, s50e-*.md
the design that produced all of it plans/plan-s50-scheduler-handoff.md