Measured Results

Measured results and limitations

Reviewed: 2026-11-01 · Source: b7140848 · Availability: current main; stable/main boundaries are stated below.

Results are bound to the source, executable, workload and settings actually tested. This page publishes existing evidence; no new emulator campaign was run for the wiki repair. Guest cadence is labelled explicitly and does not establish host-presented FPS or identical simulation progress.

October 2 cumulative comparison

The immutable report, campaign manifest and fixed inputs retain complete records and all>1% changes.

Reference A is upstream v0.8.136 source fc24584c plus matched buffered flip telemetry at 6571e4db; emulation/locking stays upstream. Candidate B is fork ee5ce48b48784f999af374c1452003f8b2b1230f. A is instrumented, rather than an untouched official binary. B is not November main; the cumulative difference cannot be attributed to any one PR.

Each host/title has ABBA then BAAB, eight attempts/four per arm: 48 final completed/passed/eligible attempts. Earlier Windows Conker disk-exhaustion schedule was retained and excluded in full before repeating the entire schedule; successful cells were not cherry-picked into the final cohort.

Host / bounded scene	Guest flips/s A → B	Cadence change	Process CPU core-% A → B	p99 interval ms A → B
Windows / PGR2 parked start	14.66 →30.00	+104.67%	316.7 →321.2	83.18 →34.00
Windows / Conker bar menu	29.18 →30.00	+2.80%	252.6 →248.2	50.06 →33.75
Windows / DOAXBV island menu	60.00 →60.00	0.00%	280.6 →247.4	17.49 →17.04
Deck / PGR2 parked start	7.38 →7.15	−3.18%	299.8 →403.5	242.72 →335.92
Deck / Conker bar menu	9.12 →9.59	+5.17%	210.2 →185.3	140.70 →161.21
Deck / DOAXBV island menu	42.74 →47.95	+12.19%	189.6 →196.3	37.13 →26.51

Windows PGR2 is the strongest scene gain. Windows DOAXBV retains60 guest flips/s while mean CPU falls11.82%. Deck DOAXBV gains cadence and p99 with higher CPU. Deck PGR2 regresses: CPU+34.60%, p95+42.77%, p99+38.40%, alongside lower cadence. Deck Conker gains mean cadence but p99 worsens14.57%. Neither loss is hidden by the favorable rows.

Conditions and units

• Windows: Ryzen9 6900HX, RTX3070Ti Laptop GPU, NVIDIA581.95, Windows10 build19045. Deck: AMD Custom APU0405, RADV VANGOGH, Mesa25.3.0 git59b552c765.
• Vulkan guest1×, host1280×720, VSync Off, shader cache Off, isolated state/private HDD/EEPROM. Fork DSP JIT and tested Prewarm/fast defaults; upstream original interpreter.
• CPU 100%=one logical core. Whole-device GPU utilization is secondary context, not xemu-exclusive cost. Concurrent CPU/GPU spans are not additive frame time.
• Table values are medians of four per-attempt summaries, not percentiles of pooled raw samples. CPU/GPU cover authored30 s plus screenshot overhead; guest statistics cover a bounded25 s tail/flip-group calculation.
• PGR2 ends parked at0mph rather than active racing; cross-host starts differ. Menus are not phase-locked. There is no qualified external frame observer.
• Windows driver cache was uncontrolled and accepted; Deck used a verified private Mesa namespace. OS/page caches, initial thermals and power state were not fully controlled. Sampling itself has matched, nonzero overhead.

November 1 generated status path #296

Candidate0bfa34ff versus parent 47dc9059; merged mainb7140848 has the reviewed candidate tree. PR #296, final public review, open context research #236.

Fixed-work test	Backend	Before µs	After µs	Time difference µs	Improvement %	Correctness boundary
x87_status_ax	OpenGL	53785.25	11920.25	−41865.00	+77.837%	Returned-state checks; captured-context limitation below
x87_status_ax	Vulkan	54446.25	11355.25	−43091.00	+79.144%	Same boundary
x87_compare_status_ax	OpenGL	289622.00	228587.75	−61034.25	+21.074%	Same boundary
x87_compare_status_ax	Vulkan	297903.25	230984.00	−66919.25	+22.463%	Same boundary

These are guest-QPC focused work times, medians of four run medians per arm; they are not game-FPS percentages. Deck Vulkan PGR2 whole-window READ3D cadence18.1705→18.0346/s(−0.748%); CPU essentially flat; p99 77.2801→78.2095 ms; median per-run maximum182.3605→211.5880 ms. Isolated275.443/238.581 ms stalls remain. A favorable last-five-group calculation does not replace the adverse whole-window result.

Forty attempts generated140 leaf records:120 PASS plus20 retained captured-context FAIL/ineligible records. The shared fault-context leaf fails both arms; direct pre-resume FP state was unobserved. Status timing retains generic inconclusive classification; checkpoint comparisons remain ineligible. Source/returned-state checks supported the bounded architectural repair, rather than making the missing contract pass. Current raw records are private workspace evidence; the public compact table has less independent reproducibility than October's linked package.

Pipeline cache #147: component result

In one matched released-candidate PGR2 intro pair, restored cache bytes reduce summed worker pipeline-construction44.941→4.740 ms(about89.5%). Both runs retain roughly3.4 s maximum intervals. This is a component-time reduction, not89.5% FPS, loading-time or repeated all-title proof. PR #147, recorded prerelease.

One157/157-record XISO run has152 leaves plus5 groups,1789 measured iterations and80.616 s summed guest leaf time. There is no matched pre-patch XISO control. The separate80.965 s result at 1b582a35 is already on the merged #147 tree; those totals cannot establish the patch's speedup. Operationally, compatible bytes persist/restore, invalid identity falls back and hot-draw disk I/O is avoided.

Reliability and progress outcomes

Repair	Before / after observation	Qualified boundary
Def Jam PGRAPH/PTIMER #271	Owner reaches active Windows Vulkan match beyond former freeze;10266 callbacks/10267 alarm writes; max lateness2.406 ms, max BQL-held MMIO wait2.826 ms	Diagnostic exits0 after about 13.8 min; planned1 hour not completed; OpenGL/save-load unqualified. Release evidence.
APU snapshot #53	Baseline asserts before Morrowind readiness; repaired candidate completes20 s OpenGL/Vulkan snapshot cells	No comparable baseline timed completion, so no causal speedup percentage. Evidence.
Input #279 / worker #305	Avoid absent-provider crash / recheck completion predicates after condition wakes	Correctness and operational improvements; no invented performance percentage.

Withdrawn, adverse and historical results

The September19 original XISO percentages were withdrawn because Advanced settings were not explicitly matched. Preserve raw functional observations, but never revive withdrawn speedups. Original correction.

The matched rerun has four157/157-pass runs and150/152 matching leaf hashes. Of150 comparable leaves:54 favor candidate,95 main,1 tie; median latency-ratio speedup−0.59%. Focused sampler−4.62% under that report's ratio formula equals4.84% longer latency. These are diagnostic guest work times, not pooled game FPS.

Historical September PGR2+110.8% guest-write evidence is its own two-run study and cannot be pooled with October+104.67%, renamed loading time or extended to a snapshot comparison that failed to run. Historical upstream report. Issue310 retains mapping-stack Morrowind cadence/tail losses; #335's2.40% texture-tail tradeoff and #336 seeded black/stopped-flip observation remain disclosed in specialist pages.

How to interpret a percentage

Time/cost improvement=100*(before-after)/before; throughput improvement=100*(after-before)/before; positive is favorable. Latency-ratio speedup uses a different denominator and must retain its formula. Zero baseline has no meaningful percentage. Distinguish attempts, captures, eligible comparisons, leaf tests, groups and iterations. Current methodology.