[eljackjoestar-cmd]

What

Collapsed color array views move the surface address to the selected slice but lose that slice's contribution to the XOR swizzle equation. A draw can look correct in its separate 2D render image while publication and a subsequent array read place its texels at the wrong addresses. Lost Soul Aside's reflection cube then reads green clear values from packed mip tails, producing flashing green lighting.

Fold the selected slice's equation bits into pipeBankXor when CB_COLOR_ATTRIB3 represents the view as a shifted 2D surface. Full-array and volume views retain their existing addressing. The regression extends the existing R32 R_X native test to collapsed slices 1, 2 and 3, checking the array read, untouched neighbouring slices, publication and re-upload.

Addresses #4387. Main now includes #4318 (D_X color support) and #4328 (nonresident packed tails); neither implements the collapsed-view slice-XOR correction. The rebased fixture retains both R_X/D_X full-array transport cases and adds R_X collapsed slices 1/2/3. This change uses the existing swizzle equations and adds no shader instruction semantics.

Tested

Windows 11 x64, WinLibs GCC 15.2, Release, NVIDIA RTX 5070; isolated branch rebased onto main fc1f063f5, emulator SPIRV-Tools disabled.

• Full Windows Release build passes on fc1f063f5. agc_driver_array_layer_target and agc_driver_graphics pass without skips in 5.40 s, retaining full-array and collapsed-slice 1/2/3 coverage. Conventions and whitespace checks pass. The earlier negative control failed on slice 1 without this fix; it was not repeated for this rebase.
• Earlier integration validation of Lost Soul Aside PPSA26358: before the fix, separate render images were blue while the reconstructed cube contained green values up to 10,000. After the fix, all eight copied cube mips contain finite blue values, with maximum green below 0.3. A bounded 256-draw HDR readback found no large green values or nonfinite values.
• Two title-screen runs were clean, including a run on newer main with the temporary capture hooks removed. The user independently confirmed that the green flashes are gone. The integration build also includes the title's other bring-up changes; the native regression above tests this patch independently on main.
• The title scene still differs from a reference image, and gameplay and startup audio problems remain. This PR claims the slice-addressing/green-flash fix, not complete visual correctness or playability. Linux, other GPUs and console hardware remain unverified.

No game run was repeated for this rebase. The debt diff contains only the collapsed-color-array entry; all other entries match this main. The debt entry follows the new layout in docs/dev/debt/libSceAgcDriver.md. No game shaders, requests, assets, capture hooks or trace files are committed.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw; no silent stubs added
• No notes, investigation files or images added to the repository
• No new third-party code
• Generic behaviour, not specific to one title
• Depends on: none; feat(agc): support 64 KiB display XOR color targets #4318 is now merged, and its R_X/D_X fixture checks are retained
• Unblocks: PPSA26358 Lost Soul Aside (PPSA26358): bring-up and remaining blockers #4401
• AI-assisted: yes

[Amoenus]

I checked the overlap with my #4318. The runtime fixes are independent, but both extend ArrayLayerTarget.cpp by changing DecodeTarget, DrawSlice and Run: yours adds collapsed slices, mine adds the swizzle mode. The current guide asks us to agree an order for this test conflict.

I propose #4551 first, #4318 afterward. I will mark #4318 as depending on #4551 for fixture sequencing and rebase its regression to keep your R_X collapsed-slice 1/2/3 checks plus our R_X/D_X full-array transport checks. I have already rebased #4318's unchanged three commits onto d592d54e; Windows full-build/retests are running, and no new validation result is claimed yet. Does this order work for you?

[eljackjoestar-cmd]

Yes, that order works: #4551 first, then #4318. The runtime changes are independent; the dependency is for sequencing the shared ArrayLayerTarget.cpp fixture changes.

Please keep the R_X collapsed-slice 1/2/3 checks alongside your R_X/D_X full-array transport checks when rebasing. #4551 has passed the native array-target and graphics tests on Windows/RTX 5070. Its Depends on remains none, and I will record that #4318 follows this PR for the fixture update.

[blawesom]

Small conventions catch — full suite is green locally (957/957, Linux/RADV), but tools/check_conventions.py fails with [binary] on docs/dev/TechnicalDebt.md: the file contains one invalid UTF-8 byte.

At byte 191725, slices 1\x963 (the Lost Soul Aside entry) uses a Windows-1252 en-dash (0x96) — it needs the UTF-8 encoding (– = 0xE2 0x80 0x93), or a plain ASCII hyphen. One-character fix:

python3 -c "d=open('docs/dev/TechnicalDebt.md','rb').read(); open('docs/dev/TechnicalDebt.md','wb').write(d.replace(b'1\x963', b'1\xe2\x80\x933'))"

[johnmaia]

Tested on Linux, RX 7600 (RADV NAVI33, Mesa 25.2.8) with Disney Epic Mickey: Rebrushed (PPSA15657). It's the same bug as Lost Soul Aside's, and this PR fixes it there too.

Root cause, found independently in Mickey before I saw this PR:

• Unreal's sky capture renders each cube face with CB_COLOR0_VIEW slice start = max = face, while CB_COLOR0_ATTRIB3 = 0x4dc6c000 declares no array (depth 0).
• The surface is RGBA16F 128x128 with 8 mips in SW_64KB_R_X, so mips 1 to 7 sit in each face's 64 KiB mip tail.
• AnyPS5 rendered faces 1 to 5 into shifted 2D surfaces without the slice's XOR bits.
• The cube then read back through the array had faces 1, 2, 3 and 5 scrambled: green about 233 at mip 0 and 10,000 in the mip tails. Face 0 was clean, and face 4 was nearly clean (a 16-texel shift). The sky light's specular term turned the whole level green.

With this PR's State.cpp change applied alone (my own local fix removed) on my Mickey test bed:

• The final sky cube, read back right after the game's compute copy, holds finite values in all 6 faces × 8 mips. No texel has green far above red and blue, against about 16,000 per face before.
• The level renders with its intended colors. Before: green; after: fixed.

My local fix took a different route: it gave such a view arrayLastSlice = slice, so the draw used the resident array image. It worked in the game too. Folding the slice into pipeBankXor keeps the existing 2D path and is the smaller change.

One problem in the diff: docs/dev/debt/libSceAgcDriver.md re-encodes characters on lines this PR doesn't otherwise change: ⌊ becomes ⌊, × becomes × and ± becomes ±. It looks like a UTF-8 file was saved as Windows-1252. Those lines should probably keep main's text.

I haven't run agc_driver_array_layer_target on Linux yet: the newer main needs libxrandr-dev, which this machine is missing. I'll follow up with that result.

[johnmaia]

Follow-up: the regression tests on Linux, RX 7600 (RADV NAVI33, Mesa 25.2.8), at head ff4b7d901.

• agc_driver_array_layer_target and agc_driver_graphics pass.
• Negative control: with State.cpp from main, agc_driver_array_layer_target fails at the collapsed slice 1. The message is slice 1 read through the array after the draw: texel (0, 0) reads 0x5a5a5a5a, expected 0x3e800000, the same mismatch you saw on NVIDIA.
• All agc ctests pass: 449 run, the usual 2 skipped (ngg_vertex_wave_reduction, oversized_dispatch).

[oneandonlydean]

docs/dev/debt/libSceAgcDriver.md re-encodes three entries this PR doesn't otherwise change (partial rect-list primitives, the nullDescriptor placeholder image and the shader float mode): ⌊ becomes ⌊, × becomes × and ± becomes ±, as johnmaia noted. Please restore main's text for those lines and keep only the new entry.

[eljackjoestar-cmd]

Restored the partial rect-list, nullDescriptor placeholder image and shader float-mode entries byte-for-byte from main. The documentation diff now contains only the new collapsed color array views entry. Conventions and whitespace checks pass; runtime code is unchanged.

[eljackjoestar-cmd]

Rebased onto main fc1f063f5, preserving main’s updated depth-storage documentation and keeping only the collapsed-color-array debt entry in this PR’s documentation diff. Full Windows Release build passes; agc_driver_array_layer_target and agc_driver_graphics both pass without skips in 5.40 s. Conventions and whitespace checks pass. No new game-runtime claim is made for this rebase. The agreed #4551-before-#4318 fixture order is retained.

[adriadam10]

Checked the slice-XOR fold against the swizzle equations used by the detile shader (slice << 24), and the new R_X collapsed slice 1/2/3 checks. Without the State.cpp change agc_driver_array_layer_target fails (negative control); with it, it passes together with the rest of the libSceAgcDriver tests on the rebase (only agc_driver_tessellation crashes, also on main). The only conflict with main was the adjacent line in docs/dev/debt/libSceAgcDriver.md; I rebased and kept your entry. Merging once the full build is green.

[github-actions]

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.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
#4937	core/libs/prx/libSceAgcDriver/tests/execution/ArrayLayerTarget.cpp L168, L175, L179, L183, L210, L249, L256-261, L282-283
#5003	docs/dev/debt/libSceAgcDriver.md L30