Upstream Contribution Candidates

Fork changes worth contributing upstream

31 candidate areas · 3 open-work entries · Evidence and caveats · Stable/main

Fix / benefit names the observed repair or measured benefit. Named-game fixes apply to the tested path; proposed fixes remain HOLD.

First batch: #2 + #201 + #121 + #138 + #305.

CPU, GPU and frame counters

Before→after. CPU: process usage, 100%=one core. GPU: whole-device usage. FPS: source-reported FPS; host presentation is not independently verified. Guest/s: guest event rate, not FPS. — = unreported, never zero.

Each row is a separate retained campaign. † historical/diagnostic; ‡ mixed or adverse; § different GPUs. Scope and regressions.

PR	Fix / benefit	Workload	CPU core-%	GPU device-%	Reported FPS	Guest/s	p99 ms
#335	Optimization • CPU usage reduction • PGR2 — frame cadence increase	Deck / PGR2	380.947→ 363.672	51.43→ 57.67	—	7.388→ 9.001	346.647→ 221.125
#336	Optimization / risk • CPU usage reduction • PGR2 — frame cadence increase • Seeded-history failure remains	Deck / PGR2	364.131→ 276.695	58.34→ 59.53	—	9.505→ 14.974	199.448→ 102.395
#296	Fix + optimization • FPU state correctness • FNSTSW work-time reduction • PGR2 — cadence/tails worse	Deck / PGR2	251.5653→ 251.4226	—	—	18.1705→ 18.0346	77.2801→ 78.2095
#191	Small efficiency gain • PGR2 — CPU usage reduction	Deck / PGR2	252.845→ 250.937	—	—	17.623→ 17.739	79.532→ 77.919
#283	Mixed results • Morrowind — cadence increase • Morrowind — worse p99	Deck / Morrowind †	222.318→ 221.828	70.34→ 75.76	—	12.765→ 13.045	110.005→ 114.591
#187	Mixed experiment • No repeatable game gain • PGR2 — worse p99/max	Deck / PGR2 ‡	324.401→ 322.967	—	—	18.585→ 18.62	—
#187	Mixed experiment • No repeatable game gain	Windows / Conker ‡	247.732→ 247.661	—	—	30→ 30	—
#190	Historical efficiency • Conker — CPU usage reduction	Windows / Conker †	250.915→ 241.573	—	—	30→ 30	—
#190	Mixed results • Morrowind — near-flat cadence • Morrowind — worse tails	Deck / Morrowind †	—	—	—	13.192→ 13.207	—
#190	Mixed results • Conker — lower CPU reported • Conker — slightly lower cadence	Deck / Conker †	—	—	—	8.348→ 8.325	—
#275	Mixed results • Voice processing — less work • Morrowind — cadence/p99 regression	Deck / Morrowind ‡	221.962→ 219.785	—	—	13.01→ 12.676	100.145→ 116.647
#158	Fix • PGR2 — menu/atlas visual corruption	Windows / PGR2 †	—	—	—	30.00017→ 30.00059	34.28→ 35.223
#158	Fix • PGR2 — menu/atlas visual corruption	Windows / Morrowind †	—	—	—	29.213→ 29.135	50.01→ 50.03
#76	Fix • Dual GPU — black screen	Windows / PGR2 shared †	—	—	29.408→ 29.536	—	43.144→ 43.917
#76	Fix • Dual GPU — black screen	Windows / PGR2 cross-GPU §	—	—	29.536→ 25.39	—	43.917→ 56.18
#104	Fix • Vulkan — invalid depth/texture copies • Device loss — unresolved	Windows / PGR2 †	—	—	29.412→ 29.665	—	43.558→ 42.184
#104	Fix • Vulkan — invalid depth/texture copies • Device loss — unresolved	Windows / Morrowind †	—	—	28.312→ 28.963	—	50.046→ 50.034

#182, earlier head / DSP JIT off: CPU reduction +3.530%; guest-cadence change +1.968%. Absolute counters unreported in the retained summary. Scope.

#200: Auto-four, all six cohorts

Separate 48-run clean-upstream study, Vulkan, ABBA/BAAB; medians of four per-run summaries per arm. Scene equality remains unresolved. Deck Conker p99 worsens 9.30%.

Fix / benefit	Workload	CPU core-%	CPU improvement %	GPU device-%	Guest/s	p99 ms
Efficiency • CPU usage reduction	Deck / Morrowind	180.33→163.48	+9.35%	70.6→68.7	6.79→7.19	226.03→217
Efficiency / mixed • CPU usage reduction • Conker — worse p99	Deck / Conker	217.85→199.26	+8.53%	25.56→26.13	7.86→7.9	158.12→172.83
Efficiency • CPU usage reduction	Deck / PGR2	293.12→268.97	+8.24%	38.17→34.26	8.32→8.39	210.24→208.08
Efficiency • CPU usage reduction	Windows / Morrowind	250.49→244.66	+2.33%	38.45→38.84	24.98→25.62	50.41→50.23
Efficiency • CPU usage reduction	Windows / Conker	254.38→247.21	+2.82%	53.09→53.74	29.17→29.56	50.09→50.04
Efficiency • CPU usage reduction	Windows / PGR2	292.12→256.65	+12.14%	35.2→35.83	24.17→24.65	52.05→51.87

Correctness contributions

PR	Summary	Fix / benefit	Result / counter	Next
#2	Release owned SDL resources	Fix • SDL — leaked GL context/window	64/64 cycles; 0 added windows	Port + lifecycle tests
#201	Unlock on worker exit	Fix • Shutdown — voice-worker deadlock	3/3 timeouts → 3/3 joins	Production teardown tests
#121	Bound complete ZPASS reports	Fix • ZPASS reports — out-of-bounds writes	16/16 + 2/2; GL/VK PASS	Check upstream writer
#132	Fix renderer-switch lock order	Fix • Renderer switching — deadlock	3/3 concurrency cases	Lock/lifetime review
#124 + #298	Revalidate texture DMA sources	Fix • Texture DMA remap — stale colors	Stale red → correct green	Separate GL/VK ports
#127 + #256	Fix cubemap stride and GL crop	Fix • Cubemaps — wrong face/mip pixels • Converted mips — buffer overread	1/12 → 12/12; 96/96 repetitions	Separate layout/crop patches
#76	Cross-GPU presentation fallback	Fix • Dual GPU — black screen	Black screen → visible output	Interop + transfer cost
#10 + #123	Reject oversized packets atomically	Fix • Oversized packets — partial state writes	24 accepted + 12 oversized modes	Coordinate upstream #2930
#138	Initialize bump uniforms	Fix • Shader uniforms — uninitialized values	7/7; deterministic bytes	Small producer patch
#305	Recheck completion predicates	Fix • Voice workers — spurious-wake abort	Spurious-wake abort → PASS	Port wake regressions
#279	Guard missing input provider	Fix • Input tab — controller disconnect crash	SIGSEGV → tab opens	Check UI lifetime
#125	Read GPU surfaces before GL vertices	Fix • GL surface→vertex — stale attributes	5/5 + 4/4; guest PASS	Coordinate upstream #2984
#119 + #158	Preserve newer guest writes	Fix • PGR2 — menu/atlas visual corruption	15/15; 6/6 short reloads	Ownership review; p99 +2.75%
#50	Balance paused snapshot cleanup	Fix • Paused snapshots — ownership/cleanup	Ownership repaired; FPS —	Save/failure regressions
#53	Quiesce APU before restore	Fix • Morrowind — snapshot-load DSP assertion	20 s GL/VK smoke PASS	Lock/IRQ/lifecycle checks
#104	Fix Vulkan copy usage/alignment	Fix • Vulkan — invalid depth/texture copies • Device loss — unresolved	4/4; copy VUIDs removed	Device loss still unresolved
#264	Implement DOT_ZW depth	Fix • Ghoulies — depth/rendering corruption	17/17 GL + 7/7 VK	Hardware/oracle coverage
#271	Decouple timer/control waits	Fix • Def Jam — freeze, tested path	2.406 ms lateness; 2.826 ms BQL wait	GL/reset/save-load coverage

Performance contributions

PR	Summary	Fix / benefit	Other counter / improvement	Next
#101	Reuse completed host-copy upload	Work reduction • Host copy — duplicate uploads removed	2,991 skips; 45 uploads	Requires host-copy path
#283	Direct Linux surface uploads	Optimization / mixed • Surface upload — lower 1x latency • Morrowind — worse p99	1x: +9.13…14.92%; RGBA8 2x: −12.41%	Retain historical p99 cost
#182	Resample mono once	Optimization • CPU usage reduction, earlier head • PGR2 — frame cadence increase, earlier head	CPU +3.530%; cadence +1.968%	Earlier head; lifecycle checks
#200	Auto workers = 4	Efficiency / mixed • CPU usage reduction • Deck Conker — worse p99	CPU +2.33…12.14%	Resolve scene/tail attribution
#147	Persist driver pipeline cache	Optimization • Pipeline construction — lower time	44.941→4.740 ms; +89.5%	One pair; coordinate #2993
#296	Generate FNSTSW AX	Fix + optimization • FPU — checkpoint/FIP correctness • FNSTSW — lower work time	GL +77.837%; VK +79.144%	Game cadence −0.748%; context gaps
#335	Wait for compiler completion	Optimization • CPU usage reduction • PGR2 — frame cadence increase	Cadence +21.83%; CPU +4.53%	Hybrid dependency; control tail +2.40%
#336	Defer optional live compilation	Optimization / risk • CPU usage reduction • PGR2 — frame cadence increase • Seeded history — black/stopped output	Cadence +57.53%; CPU +24.01%	HOLD port: seeded-history failure

Percentages above use positive = improvement; component timings are not game FPS. GPU utilization has no preferred direction by itself.

Follow-ups

PR	Summary	Fix / benefit	Result / counter	Next
#47	Skip saturated LRU free scan	Historical optimization • LRU — wasted scan work removed	4,194,304→0 failed visits	Historical; absent from current fork
#11	Preserve failed-prepare retry state	Fix • Vulkan — lost failure/retry state	17/17; isolated cost −0.392%	Upstream ownership audit
#191	Read callback format once	Small efficiency gain • PGR2 — CPU usage reduction	9→1 reads; CPU +0.755%	GL CPU control −2.475%
#187	Reuse decoded ADPCM blocks	Mixed experiment • No repeatable game gain • PGR2 — worse p99/max	CPU +0.442% / +0.029%	No stable gain; Deck p99/max worse
#190 + #275	Reuse live RAM mappings	Efficiency / mixed • Conker — historical CPU reduction • Morrowind stack — cadence/tail regression	Deck cadence −2.57%; p99 cost +16.48%	Resolve #310

Open work — HOLD

PR	Summary	Fix / benefit	Result / counter	Blocker
#362 + #363	Format guards + Vulkan mip crop	Proposed fix — HOLD • R6G5B5 — wrong colors • Vulkan cubemaps — wrong border/mip pixels	11/11 + 19/19; 12 historical failures repaired	Basic conversion already upstream; qualify missing delta
#377	Export JIT state without DMA reads	Proposed fix — HOLD • Pause/save — consumes DSP DMA polls	Regression PASS; 2 pause/resume cycles	PCM/lifecycle/runtime qualification
#367	Finish GPU work before eviction retry	Proposed fix — HOLD • Texture-cache pressure — assertion/crash	3/3; historical 1,088-identity crash repaired	Current-tree GPU/game qualification

Full review, dependencies, failed checks and contribution requirements. Upstream availability and fork merge status are separate; no row is an approved upstream port.