Skip to content

README.md

Latest commit

 

History

History
121 lines (100 loc) · 6.15 KB

README.md

File metadata and controls

121 lines (100 loc) · 6.15 KB

Spec 047 §14 Phase 3 close-out, x64 advisory

This is an advisory x64 capture, NOT authoritative. Cloud PC (CPC-ander-YTZ3O, AMD EPYC 7763 64-Core Processor, x64), not on AC/dedicated hardware. Do not cite these numbers in §13 or §14 spec text. A stable-AC ARM64 re-capture on LAPTOP-4MEP83VI should ratify the matrix before §14 Phase 3 is closed.

Why this capture exists

This run extends the 2026-05-27-phase3-final-3x5/ matrix with the four close-out commits landed on spec/047-phase3-close-out (off PR #436 HEAD):

  • Engine (1) — Panel<>.PerChildAttachedAfterAll two-pass shape.
  • Engine (2) — TemplatedItems<> strategy + Reconciler.BindKeyedItemsSource, plus the T-erased TemplatedItemsErased<> + BindErasedKeyedItemsSource variant.
  • Port (4) — RelativePanel descriptor uses the new after-all callback.
  • Port (5) G2 — TemplatedListView<T> / TemplatedGridView<T> via base-derived registration. Adds a base-walk to V1HandlerRegistry.TryGet (exact-type entries always win; base-derived entries cached per concrete type for O(1) steady state).

DescriptorVariantFactory now registers 52 ported controls (50 from prior + 2 new via RegisterHandlerForDerivedTypes for the typed templated lists). The bench matrix detects (a) dispatch-table shape change from the +2 entries and the new strategy pattern-match arms in V1HandlerAdapter, (b) the registry's base-walk fallback cost on non-derived lookups, and (c) the DescriptorHandler.Children switch gaining the two new templated-items cases.

Capture environment

CPC-ander-YTZ3O, x64 (AMD EPYC 7763 64-Core Processor), Release, .NET 10.0.8, Windows 11 26200. Cloud PC — not on AC/dedicated hardware. 3 process launches × 5 reps × 13 benches × 4 variants = 780 measurements across launch-1.jsonl + launch-2.jsonl + launch-3.jsonl.

Headline — V1 ON (descriptors, post close-out) vs V1 OFF (today)

Median of n=15 (3 launches × 5 reps) per cell. Compared against the prior 2026-05-27-phase3-final-3x5/ headline (50 controls) to surface the close-out delta.

Bench This capture (V1 ON vs V1 OFF) Prior phase3-final-3x5 Delta vs prior Notes
M1 Mount_Leaf_NoCallback +21.2% +14.9% +6.3pp regression New strategy pattern-match arms in V1HandlerAdapter.DispatchChildrenMount add two upfront is-checks (ITemplatedItemsStrategy, IErasedTemplatedItemsStrategy).
M2 Mount_Leaf_OneCallback -0.1% -1.7% within noise
M3 Mount_Leaf_ThreeCallbacks -1.0% +3.3% improvement
M4 Dispatch_Switch_Cold -20.8% -21.2% held Dispatch wins persist — the +2 entries don't push past the inflection.
M5 Dispatch_Switch_Warm -23.9% -24.3% held Same.
M6 Dispatch_ExternalType -0.5% +0.2% within noise The new base-walk fallback in V1HandlerRegistry.TryGet is gated on _baseEntries.Count == 0 so external-type lookups skip it.
M7 Update_NoChange +6.3% +7.4% minor improvement
M8 Update_OneLeafChanged +18.9% +25.5% improvement (-6.6pp) Largest movement. DescriptorHandler.Children switch refactor (added ITemplatedItemsStrategy / IErasedTemplatedItemsStrategy arms returning null so dispatch happens inline) shortens the non-ItemsHost Update path.
M9 Update_AllChanged +4.5% +3.6% within noise
M10 EventHandlerState_Alloc -1.7% +8.7% improvement (-10.4pp) Volatile run-to-run; not load-bearing.
M11 ModifierEHS_Frequency +9.7% +8.5% within judgment band
M12 Pool_Rent_HotPath +18.5% +20.9% held Descriptor-interpreter pool-rent overhead persists; known regression.
M13 Setters_Suppression_Scope -2.1% -0.9% within noise

Net signal:

  • M1 regressed +6.3pp from the prior advisory — directly attributable to the strategy pattern-match arms added in V1HandlerAdapter.DispatchChildrenMount. Two is-checks fire before the pattern switch on every Mount, even for leaves that don't use templated items. Worth folding into the prior case switch in a Phase 4 perf-tuning pass; not load-bearing for correctness.
  • M8 improved -6.6pp — the DescriptorHandler.Children switch short-circuit for ItemsHost / templated-items strategies is a structural win.
  • M4 / M5 dispatch wins held.
  • M12 Pool_Rent_HotPath +18.5% carry-over — same descriptor-rent overhead the prior capture already documented; nothing in this branch intersects.

Q1 decision matrix — for completeness

Per §13 Q1's pre-committed decision matrix applied to ReactorDescriptors vs ReactorV2:

Bench vs ReactorV2 ns Q1 band
M1 +20.2% exceeds 15% — judgment call vs LOC/readability
M2 -0.5% ship descriptors
M5 -19.5% ship descriptors (improvement)
M7 +7.3% judgment-call band
M10 +5.2% judgment-call band

Verdict: No reopen condition for Q1 — Q1's reopen is gated on source-gen (§7) landing, not advisory perf noise. The close-out scope's M1 +21.2% and M12 +18.5% should be confirmed on stable-AC ARM64 before any spec-text change.

Caveats

  • Cloud PC noise. Per the prior README: "noise-prone, advisory. Do not cite in §13/§14 spec text."
  • ARM64 stable-AC re-capture on LAPTOP-4MEP83VI is deferred for the §14 ratification gate.

Reproduce

cd C:\Users\andersonch\Code\reactor2
dotnet build tests/perf_bench/PerfBench.ControlModel -c Release -p:Platform=x64
$exe = "tests\perf_bench\PerfBench.ControlModel\bin\x64\Release\net10.0-windows10.0.22621.0\PerfBench.ControlModel.exe"
$out = "docs\specs\047\phase3-results\CPC-ander-YTZ3O-x64-advisory\2026-05-27-phase3-closeout-3x5"
$results = "tests\perf_bench\PerfBench.ControlModel\bin\x64\Release\net10.0-windows10.0.22621.0\results.jsonl"
for ($i = 1; $i -le 3; $i++) {
    Remove-Item $results -ErrorAction SilentlyContinue
    Start-Process -FilePath $exe -Wait -NoNewWindow   # Start-Process -Wait is required;
                                                      # `& $exe` does not block on this WinUI app.
    Copy-Item $results "$out\launch-$i.jsonl"
}
python "$out\aggregate.py" > "$out\summary.md"

See summary.md for the full per-bench table.