Repository navigation
feat(agc): render into mipmapped 3D color targets - #7310
oneandonlydean merged 25 commits into
Conversation
|
Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.
Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.
|
|
Thanks, the mip addressing and the per-texel The PR now decides "3D target" by Repro: replace the A linear depth-1 volume without an existing host import reaches the same refusal. It used to render, and linear 3D at depth 1 has the same layout as 2D. Could you:
|
ce9264c to
a96dad4
Compare
|
@oneandonlydean 2026-10-20 CI follow-up: I traced the failed Linux job to exact #7310 head The Windows job succeeds: 457 executed tests pass and 645 skip. Main has since reverted the Xvfb workflow setup. Could you start a fresh Build workflow from current main for #7310, rather than rerunning the old job with its old workflow? The PR currently merges cleanly and its attached Conventions check passes. No patch code change is indicated by this setup failure. AI-assisted: yes (Codex); author responsible. Earlier depth-one review response and validationThanks for catching the depth-one regression. Addressed in a96dad4ef, after rebasing onto mainf663 and removing the now-merged #7294 dependency. All three requested changes are covered: volumes bypass the2D byte comparison; single-level depth-one linear volumes keep their valid existing copy path; AI-assisted: Codex; author responsible. |
a96dad4 to
bf98421
Compare
github-actions
Bot
added
unblocks title
|
Apple GPU / MoltenVK run of the new Machine: Apple M2 Max, macOS 27.2, x86-64 build under Rosetta. Head bf98421. Method: main 81cb863 + @mugurc's
Not done: the full suite with the PR. GPUs tested: Apple M2 Max (MoltenVK 1.4.2; MoltenVK main + SPIRV-Cross #2704) Written with AI assistance (Claude Code). I ran the builds and tests above on my machine. |
bf98421 to
f09a55b
Compare
github-actions
Bot
added
needs rebase
f09a55b to
9b1d53c
Compare
|
@oneandonlydean thanks, addressed both metadata write origins in Fresh main:11 focused passes. Corrected PR:20 passes (16 physical-GPU,4CPU), including all14 mip cases3/3 on each RTX5080/Radeon. Zero VUID/errors/skips/timeouts; max7.55s. Known layer-count warnings match main in shared groups. Pins, full synthetic logs and advisory triage ( |
|
Refreshed and qualified at Current report,immutable raw logs and manifest:36positives(30actualGPU/sixCPUheader),including14volume-mip selections on each GPU. Same-machine main/change controls,actualRTX5080/RadeonGraphics and loadedKhronos1.4.363 verified;no errors/VUIDs/skips/timeouts. Known compatibility warnings remain80main/100candidate;shared groups match exactly. Builds/public_headers/conventions/whitespace pass.18static advisories at unchanged main locations retained,no parser failures;no whole-program claim. The independent NVIDIA Linux run remains olderd917/main73f4:2430selected passes,no new failures/validation IDs,50skips. BDA55.9s/maps31.3s/graphics30.1s warnings remain open;this Windows selection does not clear them. The bounded serial main/change request remains;no timeout increase. Its two capability modes were validated without a main baseline because absent there. Previous37-pass qualification/controls,d917 hosted compile/link job,andJJenkx Linux6959 composition retain their historical pins;not new-head Linux/title evidence. No new game/FPS/full-suite claim. AI-assisted:yes(Codex);author responsible. |
github-actions
Bot
added
needs rebase
|
Linux compatibility evidence for exact #7310 head Private base RX 6650 XT / RADV Mesa 26.2.4, Linux 7.2.9, Release: 6/6 focused checks execute/pass, zero skips/VUIDs, four confirmed Khronos layer insertions, with This is candidate-only private composition smoke; it does not replace the paired plain-main qualification or establish title/FPS results. Known non-VUID layer-compatibility warnings remain. The first attempt stopped during module staging before any test; corrected raw-module paths were used in the fresh successful receipt. #6959's declared #7310-first merge order remains unchanged; no public source branch was modified. Source pins, fixture resolution, binary hashes, loader/test discovery and full synthetic CTest output, SHA-256 AI-assisted: yes (Codex). |
✅ NVIDIA GPU test: passedPR head
Unstable on main (2, these also failed in recent main runs on this machine; not counted against this PR)Failing on main too (3, not counted against this PR)Passed on retry (1, failed in the parallel run, passed when re-run with a 360 s timeout; not counted against this PR)Validated but not compared (2, not in the main baseline's validation pass)Validation IDs already present on main (5 IDs in 426 tests)Skipped (50)ReproduceAutomated build and test on a contributor's NVIDIA machine; it does not replace upstream CI. A new comment is posted for each new push. Full logs. |
|
@JAB-dev the refreshed overlap check flags #11557 against this PR. I checked #11557 at Would #7310 first, then #11557 rebased while retaining both additions fit your plan? This is a merge-order proposal, not an agreed dependency or a combined build/runtime result. I have not modified your branch or changed our Depends on line. |
What
Render into one mip of an uncompressed3D color surface, preserving the base dimensions and volume identity, including depth1. Use its3D mip layout, bound its slice view by the selected mip's depth, validate the full allocation, and create the resident attachment view at that mip. Unverified mipmapped3D DCC/CMASK metadata paths still throw.
Depends on: none. #7294 is merged. Earlier refreshed head 9a3ae675277b incorporates frozen main
6e9b9134f6ed14e659d870a33118d164096551dfthrough a normal merge; publishedd82f817df54416cc8ad9792382f6018bb57b32a1remains an ancestor and is archived. The eight production/test topic files retain exactly their previous added/removed lines. The #9050 buffer-range helper uses the volume's full backing-allocation address, including cached/moved binding checks; its two focused regressions are retained. No force push.Sonic Unleashed Recompiled PPSA99902, #7260 / #4443, requests a two-level128x128x8 RGBA8_UINT
SW_4KB_S_Xtarget after the opening movie. The ordinary bring-up composition stops at the guard still present on main:The diagnostic captured
CB_COLOR_INFO=0x00070428,ATTRIB2=0x101fc07f,ATTRIB3=0x0ea54007,VIEW=0x0000e000(mip0, slices0..7). #7274 merged the diagnostic; this capability change replaces the general refusal. Its64x4x4/mip1 descriptor test verifies the selected32x2 one-slice view. No title-specific branch or game modification is added.Review correction: non-mipmapped depth-one volumes now bypass the2D resident-byte comparison as all other volumes do; their complete3D allocation was already validated against
DescribeSurface. A single-level linear volume of depth1 retains the existing linear copy path, whose2D/3D layout is identical. Other nonresident volume cases continue to throw. Four new--mipscases cover96x32x1/one level in Linear, Standard4KB, RenderTarget and S4KBX modes, with exact per-texel readback.Overlap review: #7199 and #7254 are merged; their 2D CMASK accounting/refusals and layered-array attachment/dirty-range handling are retained. #7256 array attachment windows remain separate. #6959 implements mesh layer windows. The author confirms #7310-first landing order; its shared fixture must retain both optional Run arguments and enough vertices for the 33-layer mesh case and eight-layer mip cases. JJenkx's attributed Linux composition results below cover that fixture union at their stated pins; our own Windows matrix does not qualify #6959.
Tested
Current refreshed head
2c1f59472aeb, tested maind2d2a866d9a9ed65e9d3871f83dcf427195aa362. Normal merge after#12862 preserves published history/authorship,identical production/regression topic deltas,and all selected test-registration properties. No force push.Exact pins,immutable raw logs,artifact inventories,controls and analysis: 36 positives(30 verified GPU/six CPU or header). Main14GPU/twoCPU;candidate16GPU/threeCPU plus public_headers. Shared layered draw/submit/vertex/no-GPL,XOR volume,depth bounds and new heap-mirror controls run on both GPUs. Candidate additionally passes all14volume-mip selections/vendor(28total) and CPU linear-nonresident refusal. Graphics/ShaderMemory pass on both sources,including intervening descriptor force-sync changes. Full native/relinker/real-fixture builds,conventions/whitespace/public_headers pass;binary inventories stable per source.
Required Vulkan,app cache/journal off,inserted Khronos1.4.363 and actual device identity verified;verbose serial30s. Windows11Enterprise26200/WinLibsGCC15.2,RTX5080/NVIDIA621.01 and RadeonGraphics1002:13c0/AMD26.10.1. Zero validation errors/VUIDs/skips/timeouts;known layer-compatibility warnings main80/candidate100,identical IDs/counts for shared groups. Two extra volume cases account for20warnings. Max wholeGPUprocess1.231s;max anyCTest17.794s. Not warning-free or a performance comparison.
Cppcheck/clang-tidy:115.463 analyzer-seconds,18advisories at unchanged main locations,no parser/configuration failures/timeouts or automatic fixes;no static-clean/whole-program claim. Historical isolated address/metadata mutation controls and previous fixture analysis retain their original pins;no new current-main failing control claimed. Candidate-only registrations are absent on literal main,not passing main controls. Latest inspected6190 has unrelated exports/relinker changes and is not the tested base.
Independent NVIDIA Linux retest of olderd917/main73f4:2430selected passes,no new failures/validation IDs,50skips;BDA/maps/graphics timing warnings remain source-pinned and unresolved by this focused Windows qualification. Prior hosted d917 compile/link evidence is historical;new-head hosted checks pending. No new title/FPS/full-suite/Linux/macOS/console or dependent-stack execution claim.
Additional Linux composition evidence: JJenkx tested this exact head with #6959 head327addf79eae on RX6650XT/RADV Mesa26.2.4. The report has six passing checks, zero skips/VUIDs and four confirmed Khronos layer insertions, covering all14 volume-mip selections plus the33-layer mesh case. Its private base also contains capture/diagnostic changes; this is candidate-only composed evidence, not a paired plain-main qualification or title/FPS result. The shared fixture retains both optional Run arguments and enough vertices for both mip and mesh cases. The agreed #7310-first order remains; our source is unchanged.
Checklist
main; no other open PR implements the same functionsNotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file.mdfiles or images added to the repository (attach them to this PR)