Upstream Work

Upstream Work

For fork changes worth offering upstream, use the dedicated contribution shortlist: PR, summary, improvement metrics/correctness outcome, upstream overlap and remaining work. This hub tracks existing upstream PRs and overlap. On November 1 the 26-page September review set contains 23 open PRs, two closed without merge, and one merged PR. That count describes this review set, not all upstream work.

Reviewed: November 1, 2026. Fork source: b7140848 (full SHA b71408480a3a979a9c584663e1d1f400904c8bb7). Latest stable checked: v0.8.136-0-g458730bf5373, September 30, source 458730bf5373f1f0d027450fef7443e3af7a3ae1. GitHub state, branch recommendation, source presence and release availability are separate facts. See current status and release versus main.

Current feature register · Actionable research · Research and retired work

Current upstream disposition

All 26 current metadata records were retrieved from GitHub. The dedicated pages record full current/original heads and metadata timestamps. Of 24 pages with recorded original heads, 21 heads are identical and three differ (#2965, #3014, #3058). #2977/#3028 omitted original heads, so their source-change status cannot be reconstructed from the old pages alone. A newer metadata timestamp does not prove changed code.

Upstream page	GitHub state / draft	Current head	Change since original review	Current action
#819	Open; not draft	3e633bbbc049	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2929	Open; not draft	4316955c55ad	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2930	Open; not draft	e53fd6f4857d	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2956	Closed, unmerged; draft	966416c9931d	Same head; metadata refreshed	Retired parallel candidate; follow #2988.
#2965	Open; not draft	f0db53c49360	Changed head; review update required	Review current timestamp order and retained hardware gap.
#2977	Open; not draft	cc9546771924	Original head omitted	Open proposal; refresh integration and focused tests before handoff.
#2984	Open; not draft	cc576956ee5b	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2988	Open; not draft	8a6a4b87001a	Same head; metadata refreshed	Active upstream guard-set proposal; #2956 retired.
#2993	Open; not draft	98ab977bb75b	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2994	Open; not draft	781f5a560933	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2995	Open; not draft	42cf316e59b0	Original head omitted	Preserve opportunity; reopen only for an affected workload.
#2996	Open; not draft	4ab056ef5bfa	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2997	Open; not draft	9e9773a53e31	Same head; metadata refreshed	Contribute missing conversion/lifetime tests; avoid duplicate packing.
#2998	Open; not draft	651079ce4de1	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#2999	Open; not draft	bf3f610d0518	Same head; metadata refreshed	Review feature query/enablement separately from #3000.
#3000	Merged; not draft	c53e813ec7a2	Same head; metadata refreshed	Upstream history; fork equivalent present in stable and main.
#3001	Open; not draft	5b42f9c466a8	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#3002	Open; not draft	873181abd5db	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#3004	Open; not draft	d9dc31924b55	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#3006	Open; draft	a0e2d3e7ee0e	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#3010	Open; draft	2e39fed36955	Same head; metadata refreshed	Follow #319 current-head border repair and stale B3/B4 diagnosis.
#3014	Closed, unmerged; draft	d07fd5f8d75b	Changed head; review update required	Closed research; no verified rationale or replacement.
#3028	Open; not draft	792e00bd42bd	Original head omitted	Recover original self-page/SmartXX reproducer before qualification.
#3047	Open; draft	5b1c1404c55a	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#3051	Open; not draft	3b4c7f97f0f3	Same head; metadata refreshed	Open proposal; refresh integration and focused tests before handoff.
#3058	Open; not draft	885319c94a4a	Changed head; review update required	Re-review new locking, callback context and unload sequence.

Dependencies and delivery boundaries

• #2956 closed for #2988: use the author’s closure explanation, not two active parallel vertex-guard candidates. #2988 remains open; unchanged-byte comparison is only a possible follow-up.
• #3000 merged; #2999 still open: supported SPIR-V target selection and required Vulkan feature enablement are different contracts. Independently checked fork target selection exists in both main and September 30 stable; the upstream merge commit itself is not a fork ancestor. See #3000’s source checks.
• #2965 / #3058 changed source: re-review texture timestamps after submit-count advancement, and plugin locking/unload/asynchronous callback lifetime. The September reviews do not cover those later changes. #2965 retains the hardware no-semaphore agreement limitation.
• #3014 closed without merge: no verified closing reason or replacement is asserted. Preserve the input-provider lifetime and freshness questions as research.
• Fork source is independent: #317/#319/#320/#321/#351 remain open/HOLD or SKIPPED; #318 is closed/SKIPPED. None is a delivered main/stable feature. Their bounded next actions supersede the old port queue.
• Retired product work: Shader Browser was removed by #268 before the checked stable release. Workbench #241/#259/#260 closed unmerged; #242 was superseded by #259. They are preserved branch research rather than current UI or upstream port promises. Neural, depth/color consumers, linear resampling and GP/HLE work also remain separately held.

No upstream comments, code submissions or native campaigns were made by this documentation refresh. Existing fork correctness fixes may still provide useful tests; each port needs an exact current upstream base, focused review and its own applicable qualification.

Historical upstream and code-priority reviews — September 21–26, 2026

The complete earlier review follows for provenance. Its old open-PR counts, source identities, active-candidate labels and Shader Browser port advice are historical, not today’s recommendation. Use the current table and delivery boundaries above.

This hub tracks open xemu-project/xemu pull requests that overlap or interact with work already implemented or tested in this fork. It exists to turn downstream experience into accurate, bounded upstream contributions.

Audit snapshot: 2026-09-21. Upstream had 104 open pull requests. Fork main was 82d04d84bdb3de956f20c3537e1da2cece0e8da2; reviewed upstream base was f9b14039e5bb56ae2d8f028e31e7cc19f13f7e12. This is source review and test-planning material, not proof that any branch merges cleanly or passes native qualification. Check each linked PR before posting.

Code-priority refresh: 2026-09-26. Fork main is 9a58e261; upstream master is still f9b14039. The upstream open-PR listing returned 112 PRs. The overlapping PRs named in the first-submission list below were rechecked as open. The older index and dedicated comparison pages retain their September 21 review scope; this refresh is a prioritized port plan, not a new full source audit or a claim that patches apply cleanly.

Best fork code to take upstream

Start with correctness repairs that can be reviewed as small changes and have a focused regression. Rank below weighs user-visible failure, applicability to upstream, and how much of the fork's implementation can be separated from fork-specific infrastructure. A merged fork PR is a source candidate, not an upstream-qualified patch. Where an upstream PR already exists, contribute a focused patch or test to that work after coordinating with its author; do not open a competing copy of the feature.

Order	Fork code	Why it matters	First upstream action and acceptance gate
1 — prepare first	#10 + #123: atomic PGRAPH array-packet preflight	An oversized packet must not partially mutate draw state, including when bulk processing is off or tracing chooses the scalar path. Narrow correctness rule with a production-linked test.	Coordinate a focused validation patch with upstream #2930's author, keeping upstream's bulk implementation. Test INLINE_ARRAY, ARRAY_ELEMENT16/32, exact capacity and one word over capacity in bulk, scalar and traced modes.
2 — prepare first	#125: OpenGL GPU-surface reconciliation before vertex fetch	A draw can consume stale vertex bytes when a GPU-authored surface overlaps the actual fetch range and the quick-upload path skips the changed data.	Offer the checked fetch-span/readback and mirror-invalidation repair to upstream #2984. Test zero stride, element size versus stride, exact overlap and the production-linked surface-to-vertex case. Upstream's native-scale policy remains separate.
3 — high-impact correctness	#119 + #158: Vulkan surface authority	A stale cached-surface readback can overwrite newer guest writes and corrupt PGR2's menu texture. This is a user-visible data-authority failure.	Extract the smallest guest-write notification and stale-readback veto series with an overlap/replacement regression. Compare with upstream #2965, which protects a different reuse boundary; do not call it a duplicate.
4 — reliability, split by subsystem	#50: paused NV2A snapshot cleanup; #53: DSP quiescence before load	Failed or already-paused snapshot operations must release ownership correctly; restored DSP state must not race the running audio thread.	Compare the actual save/load hook ordering with upstream #819. Prepare separate graphics and audio patches, with paused save/failure cleanup and active-DSP load tests respectively. Do not treat #819 as equivalent without hook-level review.
5 — small correctness patch	#138: initialize unused bump-uniform entries	Reflected bytes that both renderers compare or copy must be deterministic even when the guest leaves stage zero unused.	Check current upstream source, then submit only the shared-producer initialization and sentinel regression. This complements upstream #2998; its uniform-copy change did not create the pre-existing producer defect.
6 — focused Vulkan validation	#104: transfer-source usage and BC staging alignment	The depth scratch buffer lacked transfer-source usage, and a compressed-texture upload could start between BC blocks. The fork's validation removed those specific copy errors but did not resolve the investigated device loss.	Upstream #2999 already addresses the transfer-source usage. Check current upstream source, then offer the distinct BC block-alignment fix and its focused copy test if still missing; run Vulkan validation on the upstream-based head.

Important code that needs a narrower upstream route

Fork work	Recommended route
#70, #139, #147: persistent SPIR-V and pipeline caches	High-value loading work, but upstream #2993 already proposes the feature. Lead with cache-key, bounded-load, worker-quiescence and Vulkan-recreation tests; port only a missing lifecycle repair after reviewing upstream ownership. No broad FPS claim follows from cache hits alone.
#228: learned Vulkan fallback readiness	Important shader-stall work now merged in the fork. Separate the missing-stage request/retention and publication rules before proposing upstream patches. Its reported synchronous-work reduction had mixed frame tails, so qualify an upstream-based head before claiming game-performance benefit.
#75 + #76: GPU identity and cross-adapter presentation	Potentially valuable multi-GPU compatibility work. First diff upstream's current device selection and presenter contracts, then test requested versus initialized GPU and incompatible-adapter host-copy fallback on suitable hosts.
Shader Browser #237, #238, #239	Merged opt-in debugging features, but the UI, persistence, recipe and override contracts form a larger series. Seek upstream design interest and propose stages before porting. Preview #241 and profiling #242 remain open drafts, so exclude them from a main-based port.

An exact-match overlap such as upstream #2994 should receive tests or review evidence instead of a duplicate fork patch. Keep unmerged APU experiments in the dated development snapshot until their own merge and qualification decisions are made.

How to use these pages

Each upstream PR has a dedicated page containing:

• the refreshed upstream state and exact reviewed head;
• the corresponding fork work;
• which mechanism is equivalent, partial, alternative, or merely adjacent;
• transferable tests and implementation lessons;
• evidence limits and claims to avoid;
• a suggested handoff that can be adapted into an upstream review comment.

No comment should be copied blindly. Refresh the upstream head, read new discussion, verify the referenced fork code still exists, and post only evidence that can be shared publicly.

Relationship vocabulary

Label	Meaning
Exact mechanism	The same core mechanism exists downstream; offer validation and edge cases rather than duplicate code.
Substantial overlap	Most responsibilities exist downstream, but lifecycle, API, or ownership differs.
Partial overlap	The fork covers only part of the upstream PR. State the uncovered portion explicitly.
Alternative strategy	Both changes address the same hazard with different ownership or data structures; tests transfer more safely than code.
Adjacent hazard	The changes concern the same subsystem but protect different directions or lifetimes.
Complementary	Downstream work supplies integration tests or constraints but not the proposed feature.

Upstream PR index

Upstream PR	State at audit	Fork relationship	Fork references	Priority
#2993 — Persist pipeline and SPIR-V caches to disk	Open	Substantial downstream implementation overlap	#70, #139, #147	Highest
#2998 — Derive uniform dirtiness during copying	Open	Core mechanism already present downstream; companion fixes available	#138, #148 and current vk/glsl.h	Highest
#2994 — Visit active LRU nodes without scanning every bin	Open	Exact mechanism already present downstream	Current include/qemu/lru.h; #47 is separate	High
#2930 — Consume INLINE_ARRAY vertex data in bulk	Open	Bulk path present downstream with stronger packet validation	#10, #123	Highest
#2988 — Guard vertex RAM writes on what draws read	Open	Same lifetime problem; alternative downstream strategy	#85, #89	Highest
#2956 — Finish only when vertex RAM updates affect pending draws	Open	Same lifetime problem; more precise upstream range strategy	#85, #89	High
#2984 — OpenGL render-to-vertex-buffer support	Open	Partial overlap	#125	Highest
#2999 — Enable required Vulkan features and buffer usage	Open	One direct overlap plus separate feature-negotiation work	#104	High
#3000 — Compile SPIR-V for the supported device version	Open	Related target-selection and persistent-identity work	#70, #75	High
#2965 — Flush when reusing Vulkan surface memory	Open	Adjacent coherence hazard, not duplicate	#119, #158	High
#3010 — Native S3TC uploads and GL texture recycling	Open draft	Complementary compressed-layout tests	#127 and existing OpenGL texture work	Medium
#2997 — Speed up remapped vertex copies	Open	Complementary to bounded vertex lifetime work	#85, #89	Medium
#3002 — Convert vertex attributes the device cannot consume	Open	Complementary feature; no equivalent downstream conversion claimed	Vertex lifetime work and prewarm format admission	Medium
#3006 — Convert unsupported inline vertex formats	Open draft	Third step in upstream vertex-conversion series	#10, #123 plus vertex-range work	Medium
#2996 — Size Vulkan storage buffers from device limits	Open	Complementary allocation work	Related transient-buffer growth/failure-handling work	Medium
#3004 — Retry a smaller Vulkan buffer after allocation failure	Open	Complementary allocation fallback	Related transient-buffer growth/failure-handling work	Medium
#3001 — Prefer standard Vulkan border colors	Open	Focused complementary sampler optimization	No direct downstream equivalent claimed	Medium
#3014 — Cached host input outside the BQL	Open draft	Complementary to virtual-controller presence repair	#160	Medium
#2929 — Defer PGRAPH IRQ delivery from PFIFO	Open	Related lock-order work, distinct path	#132	Medium
#3047 — DSP core and DMA performance/accuracy overhaul	Open draft	Complementary snapshot lifecycle requirement	#53	Medium
#3058 — TCG plugin and monitor interaction	Open	Supports downstream profiling workflow	Diagnostic counters and bounded performance protocol	Medium
#3051 — Guest-visible xemu identification and version	Open	Useful for test identity; not a replacement for artifact hashes	Release/version identity workflow	Medium
#819 — Move snapshot surface download into savevm hook	Open	Older lifecycle overlap requiring hook-level comparison	#50, #53, #119, #158	High

Best first review handoffs on existing upstream PRs

1. #2993: cache lifecycle, identity, bounded loading, worker quiescence, and renderer-recreation coverage from #70/#139/#147.
2. #2998: deterministic fragment-uniform construction from #138, while acknowledging upstream binding-switch handling; #148 remains a separate dirty-row optimization.
3. #2984: exact vertex fetch spans, quick-upload invalidation after surface download, and the #125 production-linked test.
4. #2930: whole-packet validation before bulk/scalar/traced path selection from #10/#123.
5. #2988/#2956: three-generation vertex lifetime and rollover coverage from #85/#89.

These are focused review and test handoffs for existing PRs. The ranked code-submission plan above identifies what to extract from the fork first. Exact matches such as #2994 should receive downstream validation rather than duplicate code. Partial overlaps such as #2984 and #2965 must name what the fork does not cover.

Downstream reference commits

These are the merged fork PR identities observed during this audit. Refresh them if a PR is reverted or superseded.

Fork PR	Merged commit	Responsibility used by this audit
#10	c5a598a0ea42	Atomic oversized inline-packet rejection
#47	bac2cae991a2	Separate full-cache free-node scan optimization
#50	5151c3887a3f	Paused snapshot save/failure cleanup
#53	bd1fecb93353	DSP quiescence before snapshot load
#70	fc8c5dec9c1a	Persistent exact-source SPIR-V identity/reuse
#75	7c6a1a51fe84	Explicit selected-GPU identity/reporting
#85	ade1c546e338	Surface/vertex source freshness
#89	003b88475584	Bounded immutable vertex versions
#104	ef09f061d766	Transfer usage and depth/BC copy repair
#119	a179d9842046	Preserve guest writes over cached surfaces
#123	1fcf4434e933	Validate packets before bulk/scalar selection
#125	dcd0fa349b95	OpenGL surface reconciliation before vertex fetch
#127	08844c897bca	Checked compressed cubemap face stride
#132	3cb55dffdd46	Renderer-switch lock/progress phase repair
#138	43f84505c295	Deterministic unused bump uniforms
#139	f929fe5a8bb1	Persistent SPIR-V in fallback preparation
#147	1b582a35b04e	Persistent Vulkan driver pipeline cache
#148	ef1a7fc4b62b	Aggregate VSH dirty-row state
#158	9c529ab612dd	Stale surface-readback veto
#160	9f756c9ba036	Virtual controller without host provider

Upstream handoff workflow

1. Refresh identity. Record upstream PR state/head, upstream base, fork main, and referenced fork PR merge commits.
2. Recheck discussion. A new comment may already address the concern or invalidate the suggested path.
3. Inspect the exact boundary. Compare ownership, lifecycle, cache key, state transition, or packet admission—not just similar filenames.
4. Run the smallest transferable regression. Prefer production-linked unit/guest tests over unrelated gameplay volume.
5. Separate claims. Label source review, controlled reproduction, native result, and unmeasured risk distinctly.
6. Prepare a concise comment. Lead with the exact condition and test; link the fork PR or commit; state what the evidence does not prove.
7. Do not post without maintainer authorization. This wiki prepares collaboration material; it does not authorize contacting upstream.
8. Record the result. Update the dedicated page with posted-comment URL, upstream response, revised head, and disposition.

Comment template

I compared this head with downstream work at <fork commit>.

Relevant overlap:
- <exact mechanism or boundary>

Focused condition worth testing:
- <minimal sequence>
- expected: <invariant>

Downstream reference:
- <PR/test/commit link>

Scope limit:
- This establishes <specific result>.
- It does not establish <uncovered behavior or performance claim>.

Evidence rules

• Never call a partial overlap “already implemented.”
• Do not present one fork’s hardware result as a prediction for upstream users.
• A source-level concern is not a reproduced native failure.
• Keep historical performance results attached to the exact head they tested.
• Alternative implementations transfer regression cases more reliably than patches.
• Persistent cache identity must describe the complete compilation/driver compatibility request.
• “Uploaded this batch” and “referenced by this batch” are different vertex-memory facts.
• Surface coherence is directional: protecting pending GPU consumers differs from preventing stale GPU data from overwriting newer guest writes.
• Guest-visible version strings supplement rather than replace host-side executable hashes.

Maintenance checklist

Update this hub when an upstream PR changes state, is rebased materially, splits, merges, or closes. On each dedicated page, update:

last checked date
upstream state and exact head
fork main and referenced merge commits
posted-comment/review URL, if any
upstream response and required follow-up
native or focused validation status
final disposition

Related pages

• Development
• Upstream Work In Need — candidate mechanisms that need a fork-specific rework before implementation
• Pull Request Lessons
• Performance Methodology
• Debugging Tools and Commands