Current Change Register

Current change register

Reviewed: 2026-11-01 · Source: b7140848 · Availability: current main; release differences are explicit below.

This register covers all 43 merged PRs after the September 26 snapshot through reviewed main. It includes tests, documentation, diagnostics, removals and runtime work:43 merges do not mean43 features or speedups.

Current source is139 commits beyond the September snapshot; the latest stable is103 commits and35 of these merges behind. Those are version-distance counts, not performance measurements. Full comparison.

Included in latest stable means its source contains the merge; the first earlier stable release was not independently determined. Main only means no latest-stable availability claim. Release guide. Earlier delivered mechanisms are in Implemented. The machine-readable owner record is .wiki/current-changes.json in the wiki Git repository.

PR / change	Availability	Mechanism and user problem	Default / apply / platform	Evidence / limits	Read next
#182 — apu: Resample mono voices as one channel	included in latest stable	Mono processing duplicated input and resampler work. One resampling channel for mono, duplicate output afterwards; sinc retained.	Automatic; no new toggle Apply: Automatic at the next applicable operation; existing guards remain Platform: Audio platforms	Historical workload evidence only; no November game speedup attributed. Limits: No new November-main native performance measurement.	Technical page
#183 — audio: Enable the DSP JIT by default	included in latest stable	Users otherwise selected the existing JIT manually. DSP JIT is the fresh selection default; saved false remains false.	audio.use_dsp_jit=true; saved false preserved Apply: Engine preference applied at an audio frame boundary Platform: Audio platforms	Default selection is delivered; all-title PCM/hardware parity remains bounded. Limits: Default selection does not establish all-title PCM or hardware parity.	Technical page
#268 — xui: Remove experimental Shader Browser stages from main	included in latest stable	Experimental Shader Browser stages were unsuitable for current product. Remove menu, product files and SQLite/override integration.	Feature absent Apply: Removal already present in reviewed stable and main Platform: All	Removal; historical prereleases are not current availability. Limits: Old user databases/prereleases do not restore current-main UI.	Technical page
#269 — ci: Use tracked xemu version for untagged release builds	included in latest stable	Tagless builds lacked a reliable product identity. Use tracked xemu version in release builds.	Tracked version used by eligible builds Apply: At build/package generation Platform: All supported builds	Build traceability; no runtime speedup. Limits: Build identity does not improve emulator runtime.	Technical page
#267 — nv2a: reconcile PTIMER semantics and queue scheduling	included in latest stable	PTIMER semantics and scheduling needed reconciliation. Reconcile clock/deadline/queue behavior without claiming a faster guest clock.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: NV2A, both renderers	Timer correctness, not a clock-speed optimization. Limits: Timer semantics are not guest-clock speedup evidence.	Technical page
#265 — nv2a: Remove the prior issue #149 suppression workaround	included in latest stable	Old Ghoulies suppression workaround masked rendering. Remove the earlier issue149 suppression workaround.	Automatic removal Apply: Automatic at the next applicable operation; existing guards remain Platform: Both renderers	Paired with semantic DOT_ZW correction; not independent speedup. Limits: Workaround removal is separate from frame-exact transient validation.	Technical page
#271 — nv2a: decouple PGRAPH interrupt controls from renderer work	included in latest stable	PGRAPH control accesses could wait behind renderer work under BQL. Decouple interrupt/control work and preserve timely PTIMER responsiveness.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: Shared NV2A interrupt/control implementation, both backends	Active Def Jam match; about13.8min, exit0; planned1h/OpenGL/save-load not completed. Limits: Only the recorded Windows Vulkan path; planned hour, GL and save/load incomplete.	Technical page
#264 — nv2a: Implement DOT_ZW depth replacement for issue #149	included in latest stable	Ghoulies DOT_ZW depth output was incorrect. Validate stages and implement depth replacement with normalized fragment modules.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: Both renderers	Owner observed correction; issue149 closure does not supply missing frame-exact transient proof. Limits: Owner correction observed; frame-exact transient comparison remains unproved by issue closure.	Technical page
#279 — xui: avoid Input-tab crash without a bound controller	main only at this review	Input tab could dereference an absent physical provider. Keep configuration usable without a bound controller.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: Input UI	Null-provider crash repair; no snapshot topology guarantee. Limits: Does not reconstruct all controller topology on snapshot restore.	Technical page
#295 — docs: update benchmark README with external evidence links	main only at this review	Product benchmark summary and evidence were hard to find. Publish October cumulative comparison in README with evidence links.	Documentation only Apply: Read the current record; no process setting Platform: Windows and Deck recorded cohorts	48 attempts; includes Deck PGR2 regression; not November-main attribution. Limits: October tested builds differ from November main; Deck regressions retained.	Technical page
#256 — nv2a: Crop bordered OpenGL cubemap mips by level	main only at this review	Bordered OpenGL cubemap mips used the wrong crop dimensions. Crop applicable levels by their actual border and mip size.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: OpenGL	Applicable cubemap correction, not every compressed-border case. Limits: Does not cover every compressed-border format.	Technical page
#298 — nv2a: revalidate OpenGL texture DMA sources	main only at this review	Unchanged handles hid texture/palette DMA source changes. Revalidate source identity before OpenGL binding reuse.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: OpenGL	Correctness under remapping; no standalone game gain inferred. Limits: No standalone game throughput attribution.	Technical page
#187 — apu: Reuse unchanged decoded ADPCM blocks	main only at this review	Identical ADPCM blocks were repeatedly decoded across callbacks. Owned per-voice decoded-block cache compares fresh encoded bytes/configuration.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: APU voices	No repeatable final native game gain; retained PGR2 tail costs. Limits: No repeatable final native gain; PGR2 tail costs remain.	Technical page
#190 — apu: Cache live SGE descriptor RAM mappings	main only at this review	Descriptor reads repeated RAM mapping work. Retain bounded live SGE descriptor mappings; read current values.	Callback-local guarded reuse Apply: Automatic at the next applicable operation; existing guards remain Platform: APU RAM; fallback for other mappings	Component savings; accepted stack has adverse Morrowind evidence in issue310. Limits: Issue310 retains Morrowind cadence/tail losses for the mapping stack.	Technical page
#275 — apu: bound ADPCM sample-memory reads to RAM cache spans	main only at this review	ADPCM payload reads repeated translations. Reuse bounded payload RAM spans while reading current bytes.	Callback-local guarded reuse Apply: Automatic at the next applicable operation; existing guards remain Platform: APU RAM; non-RAM fallback	Partial span work only; issue310 native regression retained. Limits: Full span engine/SIMD conversion is not implemented; issue310 remains.	Technical page
#303 — tests/apu: qualify production resampler state fixtures	main only at this review	Resampler-state coverage did not exercise production paths sufficiently. Production state/throughput fixtures characterize completion and source behavior.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Host fixtures	Test coverage; known finite-source behavior is not repaired by a passing fixture. Limits: Fixtures characterize inherited finite-source behavior rather than repairing all of it.	Technical page
#305 — apu: Recheck voice worker completion after condition wakes	main only at this review	Condition wakes could be mistaken for completed voice jobs. Recheck worker completion predicates after wakes.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: APU worker threads	Reliability repair; no invented FPS percentage. Limits: No independent FPS percentage inferred.	Technical page
#302 — nv2a: validate offline Vulkan finish attribution	main only at this review	Vulkan finish reasons lacked qualified offline attribution. Validate read-only finish analysis and its measurement contract.	Off unless explicitly invoked Apply: Offline or paused-process workflow with declared inputs Platform: Retained trace inputs	Attribution tooling; not a runtime speedup. Limits: Offline trace attribution does not measure exclusive GPU/frame time.	Technical page
#274 — tests/nv2a: retain deterministic PGRAPH/PTIMER contention coverage	main only at this review	PGRAPH/PTIMER lock interaction lacked deterministic coverage. Production contention controls exercise completion/locking boundaries.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Xbox test workload	Detects regressions; not game performance evidence. Limits: Test coverage is not evidence of game acceleration.	Technical page
#292 — audio: qualify read-only ownership inspector	main only at this review	Audio ownership diagnosis lacked bounded source-matched tooling. Read-only Linux GDB inspector for attached paused processes and matching DWARF.	Off unless explicitly invoked Apply: Offline or paused-process workflow with declared inputs Platform: Linux GDB	Ownership snapshot; no PCM/audibility proof. Limits: Requires already-attached paused process/matching DWARF; no audibility proof.	Technical page
#257 — ui: Add typed Advanced tweak policy resolver	main only at this review	Requested/effective settings and restart state could diverge silently. Typed policy resolver and owned status publication; existing toggles stay Boolean.	Preserves saved Boolean requests Apply: Live profile publication for live controls; process-start choices remain latched until restart Platform: Applicable controls/backends	Infrastructure landed; #258/#314 CLI/evidence consumers remain unmerged. Limits: Boolean UI remains Boolean; branch-only#314/#258 consumers absent.	Technical page
#200 — audio: Default Auto voice workers to four	main only at this review	Auto could create too many voice workers. Resolve Auto from logical CPUs with minimum1, cap4; explicit1–16 remains.	Auto(0) resolves1–4; explicit1–16 Apply: Quit and restart xemu Platform: Host voice workers; normal sinc	Independent of held linear converter#189; not multiple Xbox CPUs. Limits: Restart required; not guest multi-CPU emulation or linear-resampler delivery.	Technical page
#282 — nv2a/vk: preserve read-only depth and stencil ownership	main only at this review	Read-only depth/stencil access could lose authoritative GPU ownership. Preserve read-only ownership before consumer work.	Automatic foundation Apply: Automatic at the next applicable operation; existing guards remain Platform: Vulkan	Foundation delivered; runtime consumer#290 held. Limits: Does not activate held end-to-end consumer#290.	Technical page
#283 — nv2a/vk: upload unscaled surfaces directly on Linux	main only at this review	Applicable Linux unscaled uploads used avoidable blits. Upload eligible unscaled surfaces directly.	Automatic eligible route Apply: Automatic at the next applicable operation; existing guards remain Platform: Linux Vulkan	Bounded path, not general zero-copy or Windows qualification. Limits: Linux unscaled eligibility only; not general zero-copy.	Technical page
#284 — nv2a/vk: retain byte-identical depth images across alias views	main only at this review	Byte-identical depth aliases could discard reusable storage. Retain compatible images with ownership/version checks.	Automatic foundation Apply: Automatic at the next applicable operation; existing guards remain Platform: Vulkan	Retention foundation; no claim that full#290 consumer shipped. Limits: Retention foundation does not ship the entire alias proposal.	Technical page
#286 — nv2a/vk: map packed depth aliases using guest Morton layout	main only at this review	Packed depth aliases require guest addressing, not host linear assumptions. Map guest Morton layout.	Foundation present; held consumer not active Apply: No consumer activation by a user switch Platform: Vulkan	No runtime caller for held conversion consumer. Limits: Mapping helper lacks runtime activation for held consumer.	Technical page
#287 — nv2a/vk: bound packed depth alias compute dispatch	main only at this review	Packed depth work needs bounded dispatch. Bound depth-alias compute dispatch.	Foundation present; held consumer not active Apply: No consumer activation by a user switch Platform: Vulkan	Dispatch bounds do not prove game benefit. Limits: Bounded dispatch is not a game speedup measurement.	Technical page
#288 — nv2a/vk: qualify read-only depth aliases and buffer extents	main only at this review	Depth alias buffers and extents need independent qualification. Add read-only alias/extent correctness coverage.	Tests and foundation checks Apply: Automatic at the next applicable operation; existing guards remain Platform: Vulkan	Qualified boundaries; consumer#290 remains held. Limits: Qualification is limited to declared extents/oracles.	Technical page
#289 — nv2a/vk: record bounded GPU depth alias conversion	main only at this review	Depth conversion needs an ordered recorder. Record bounded GPU depth-alias conversion.	Foundation present; held consumer not active Apply: No consumer activation by a user switch Platform: Vulkan	Recorder present; end-to-end GPU-resident aliases not activated. Limits: No runtime caller for the held conversion consumer path.	Technical page
#335 — nv2a/vk: stop polling pending pipeline jobs	main only at this review	Pending shader jobs caused zero-deadline polling. Wake on compiler completion; retain retries where required.	Automatic Apply: Automatic at the next applicable operation; existing guards remain Platform: Vulkan compiler	Accepted2.40% texture-tail tradeoff in recorded cohort. Limits: Accepted2.40% texture-tail tradeoff; recorded comparison/oracle limits retained.	Technical page
#337 — docs: judge performance PRs by overall game outcomes	main only at this review	Leaf thresholds obscured overall useful game outcomes. Document cadence/tail/CPU/correctness acceptance policy.	Documentation only Apply: Read the current record; no process setting Platform: All evaluations	>1% disclosure, not arbitrary per-leaf veto; runtime performance N/A. Limits: Policy is documentation, not new runtime evidence.	Technical page
#336 — nv2a/vk: defer live fallback-family compilation	main only at this review	Speculative live family work could compete with active play. Learn demand for future launches; optional admission uses launch-loaded history.	Current hybrid policy Apply: Automatic at the next applicable operation; existing guards remain Platform: Vulkan	Seeded-history candidate black image/stopped flips retained; later pass does not erase it. Limits: Seeded candidate black image/stopped flips remains unresolved; later pass does not erase it.	Technical page
#338 — tests: wait for async compiler result publication	main only at this review	Async compiler test checked results before publication. Wait for actual compiler-result publication.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Host shader tests	Synchronization coverage, no runtime speedup. Limits: Test synchronization repair only.	Technical page
#360 — tests/nv2a: keep GL lifecycle fixtures within stack limits	main only at this review	Large GL fixture objects exceeded stack limits. Move bounded fixture state to owned heap storage.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: GL lifecycle fixtures	Fixture robustness; no product behavior change. Limits: Fixture heap allocation has no product runtime delta.	Technical page
#361 — tests/nv2a: scope Xbox build flags to production fixtures	main only at this review	Production Vulkan fixtures lacked scoped Xbox build flags. Apply Xbox flags only to intended fixture dependencies.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Host build	Build coverage; no new emulation functionality. Limits: Scoped fixture flags do not alter product defaults.	Technical page
#364 — tests/apu: link KVM stubs into memory fixtures	main only at this review	APU memory fixtures lacked KVM stub linkage. Link required KVM stubs into relevant fixtures.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Host build	Build coverage only. Limits: Linkage repair does not validate native KVM emulation.	Technical page
#369 — tests: Observe APU worker wakeups across LTO	main only at this review	LTO bypassed an ineffective worker-wake observation hook. Observe actual pthread wake boundaries.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: APU host fixture	Measurement integrity; no scheduler change. Limits: Observation integrity does not change the scheduler.	Technical page
#368 — tests: isolate Vulkan ownership GPU hooks under LTO	main only at this review	LTO made GPU-boundary fixture hooks unreliable. Isolate production ownership hooks under LTO.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Vulkan host fixture	Retains real ownership assertions; not native GPU qualification. Limits: Host LTO fixture repair is not native GPU qualification.	Technical page
#371 — tests: iterate the attached AIO source context	main only at this review	AIO tests iterated the wrong GLib context. Iterate the context attached to the AIO source.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Xbox host test driver	Repairs a fixture hang; generic suite limits remain explicit. Limits: Does not repair all generic character/WebSocket coverage.	Technical page
#374 — tests: give file-monitor worker its Xbox loop context	main only at this review	File-monitor worker lacked the Xbox context/lock it polled. Transfer context ownership and loop-lock lifecycle to the worker.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: Xbox host test driver	Normal handoff covered; abnormal-stop/join limitations remain. Limits: Existing abnormal-stop/join limitations remain.	Technical page
#375 — tests: drain retired views before SGE reference baseline	main only at this review	Deferred RCU retirements changed owner reference baseline. Drain earlier retired views before exact leak-check baseline.	No runtime setting; fixture only Apply: When the test is invoked; no product behavior change Platform: APU SGE fixture	Deterministic false-failure repair; leak negative control retained. Limits: Repairs a test baseline race, not emulated APU behavior.	Technical page
#191 — apu: Snapshot callback-local voice sample format	main only at this review	CFG_FMT fields repeated the same read within one callback. Decode an immutable callback-local voice format once.	Automatic callback-local preparation Apply: Fresh object at every voice_get_samples callback Platform: APU voice callback	Narrow landing; no persistent envelope/control cache or universal game gain. Limits: No persistent cache of envelopes, routing or guest control writes.	Technical page
#296 — tcg: Generate hard-FPU FNSTSW AX with preserved checkpoints	main only at this review	Hard-FPU FNSTSW AX used avoidable helper work and needed correct checkpoints. Generate status extraction within existing hard-FPU mode, retaining IP/checkpoint/rounding behavior.	perf.hard_fpu=true on supported x86_64 Apply: Select existing hard-FPU policy, then restart xemu Platform: Supported x86_64; ARM branch not landed	Focused77.837–79.144% status-time reductions; PGR2 whole-window cadence−0.748%; context failures retained. Limits: 20 retained context FAIL/ineligible records; direct pre-resume state missing; modest game cost retained.	Technical page

Review contract

Presence, release availability, default/apply behavior, validation and measured benefit are separate fields. Unknown gains remain unknown. The record retains source/PR, owner, reviewed date, source paths and limitations; publication checks compare this table with the records. Use Measured results, Controls and Maintenance for interpretation.