Repository navigation
Conversation
This was referenced Oct 8, 2026
9 tasks done
9 tasks done
Contributor
Author
|
Rebased on Fedora 44 Release: |
TibixDev
force-pushed
the
pr/depth-pass-recognition
branch
from
October 8, 2026 07:21
bfd1d9e to
5b34fef
Compare
9 tasks done
github-actions
Bot
added
needs rebase
Assemble 2D depth/stencil array views from the resident render-target slices, copying only slices changed by attachment use or explicit clears. Preserve BASE_ARRAY subviews and first-layer views, and reject missing resident layers. Record the remaining layout and guest-memory coherence limitations. Validation: agc_driver_graphics passed. GPU depth-array and depth-clear tests passed with Khronos core and synchronization validation enabled, covering D16, D32/S8, cached binding refresh, render-pass writes, stencil isolation, subviews, and missing layers. Match the test command pool to the runtime's reset flags. The Skyrim menu was observed through GPU captures; the desktop was locked before a new loading attempt, so gameplay remains unverified. AI-assisted implementation.
Detile nonresident layers using their absolute guest slice index while preserving resident depth and stencil images. Refresh guest-backed layers after writes and switch to rendered contents when a resident layer appears. Reject unbacked and compressed guest layers. Validation: depth-array and depth-clear GPU regressions pass with Khronos core and synchronization validation. Tests cover mixed D16 and D32/S8 arrays, guest writes, slice swizzles, subviews and newly rendered layers. Graphics CTest passes. AI-assisted implementation.
Version depth and stencil independently after draws and clears. Keep attachment lookup read-only so a gather during resource binding cannot hide a later attachment write. Validated GPU sampling, aspect isolation and copy counts with Vulkan synchronization validation; graphics tests pass. Written with AI assistance.
Cache gathered depth storage and guest snapshots separately from component views. Keep gathered images alive through surviving views and preserve independent aspect, layer and metadata identities. Add allocation/copy/upload checks, view lifetime sampling and pending-copy exception cleanup coverage. Correct test fixture destruction order after an old-code failing assertion caused an RX6600 reset. Remaining validation used CPU-only lavapipe; depth and graphics checks pass. Written with AI assistance.
Avoid assuming low virtual addresses remain unmapped after driver activity. Reserve guard storage for the unbacked-slice checks and release it after GPU cleanup. AI-assisted: yes.
Exclude COLOR_INVALID attachments from the effective write mask while preserving later MRT export indices and depth-only rendering. Validated with the graphics tests and Skyrim menu draws. AI-assisted implementation and testing by OpenAI Codex.
Extract the held draw decoders and resident-image copy path for review. Register bounds do not prove rasterized coverage; preserve this as an explicit draft blocker in TechnicalDebt. CPU decoding and Vulkan image-operation tests pass on RX 6600 and llvmpipe, without establishing arbitrary geometry or console equivalence. AI-assisted: yes.
TibixDev
force-pushed
the
pr/depth-pass-recognition
branch
from
October 8, 2026 08:28
5b34fef to
cedec6e
Compare
This was referenced Oct 15, 2026
Closed
12 tasks done
11 tasks done
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Status after review
Draft: rasterized coverage still unproven.
Rechecked the implementation in
a27e8e19f0e4: bothDecodeDepthClearPass(const QueueState&)andDecodeDepthCopyPass(const QueueState&)decide from register state. Their full viewport/scissor checks do not examine the draw's vertex/index geometry. Two draws with the same registers but different coverage can therefore select the same full-image operation.The decoder and Vulkan image-operation fixtures cannot close that gap. This remains a draft until there is a reference-backed coverage proof with partial/cull/clip negative cases, or the original rasterized operation is implemented. Main's explicit rejection must remain the baseline. #1346 and #1371 have merged; #1368 is the remaining open feature dependency, and its newly rebased implementation must replace the old embedded copy before any eventual integration. No unsafe recognition path was promoted or marked complete.
What
Draft for coverage review. Recognize depth/stencil clear and override-copy draws before shader compilation. Record the selected image operations on the GPU and refresh sampled views after copies.
Tested
Earlier branch-specific tests and measurements remain in the discussion. They have not been rerun on a redesigned current-main version of this PR. The review audit and shared NGS-assisted title exploration do not establish that this head is merge-ready. No full Linux/Windows platform CI was dispatched in this pass.
Checklist