Oxide Room 104 (PPSA07653)

[JJenkx]

Existing thread

• I searched this category for PPSA07653 and "Oxide Room"; no thread exists.

Title

Oxide Room 104 (PPSA07653)

Content version tested

The extraction's eboot.bin. The dump has no PlayGo metadata files.

AnyPS5 commit

main 474030e584, with no extra PRs.

System tested

Linux 7.2.11 (Arch), Ryzen 9 7950X, Radeon RX 6650 XT, RADV Mesa 26.2.6, GCC, Release.

Furthest point

Boot screen at 25 s.

What is shown

The title starts at about 28 FPS, then aborts at 36 s, exit 134, when it first uses compute shader 0x20084d0000. Run with APS5_PLAYGO_UNLISTED_CHUNKS=installed.

Current blocker

The shader recompiler cannot resolve an image descriptor read through an address chosen between two constant-buffer reads:

compute shader 0x20084d0000: shader 0x20084d0000 stage 0: ShaderRecompiler::Recompile: GetImageResource dword 0 is not a valid runtime value; chain: LoadAddressU32@pc=0x00004244(GetAddressResource(Phi(Phi(ReadConstBuffer@pc=0x00000418(GetBufferResource(ReadConst(GetSrtResource, Immediate), ReadConst(GetSrtResource, Immediate), ReadConst(GetSrtResource, Immediate), ReadConst(GetSrtResource, Immediate)), ShiftLeftLogical32(ReadFirstLane(SelectU32(LogicalAnd, IAdd32, LoadBufferU32), Phi(LogicalAnd, LogicalAnd)), Immediate)), Phi(ReadConstBuffer@pc=0x00000418(GetBufferResource(ReadConst(GetSrtResource, Immediate), ReadConst(GetSrtResource, Immediate), ReadConst(GetSrtResource, Immediate), ReadConst(GetSrtResource, Immediate)), ShiftLeftLogical32(ReadFirstLane(SelectU32, Phi), Immediate)), Phi(ReadConstBuffer@pc=0x00000418(GetBufferResource(ReadConst, ReadConst, ReadConst, ReadConst), ShiftLeftLogical32(ReadFirstLane, Immediate)), ReadConstBuffer@pc=0x00003b70(GetBufferResource(LoadAddressU32@pc=0x00003b64, LoadAddressU32@pc=0x00003b64, LoadAddressU32@pc=0x00003b64, LoadAddressU32@pc=0x00003b64), Immediate)))), ReadConstBuffer@pc=0x00000418(GetBufferResource(ReadConst(GetSrtResource, Im …

I also tried two open descriptor-tracking PRs, each merged onto the same main. Neither changes the stop:

• fix(shader): fold descriptor web paths that hold the lane index or texels #12132 (fold descriptor web paths that hold the lane index or texels): the same error at pc=0x00003f28.
• feat(shader): resolve nested image tables through runtime pointers #7223 (resolve nested image tables through runtime pointers): the same error at pc=0x00004290.

How it was run

relinker --registry --rpath <build>/core/libs/libs eboot.elf app.elf
APS5_PLAYGO_UNLISTED_CHUNKS=installed ./app.elf

AI-assisted testing and report: yes.

[JJenkx]

Update on the descriptor refusal from my first post. On current main 4569fe654a, the title still stops at about 36 s:

GetImageResource dword 0 is not a valid runtime value; chain: LoadAddressU32(GetAddressResource(Phi(Phi(ReadConstBuffer ...

I merged each open PR below onto that main locally and ran 300 s with the same scripted Cross presses, APS5_PLAYGO_UNLISTED_CHUNKS=absent, on an RX 6650 XT with RADV:

build	result
+ #8939 df80d9b636	same refusal, 36 s
+ #14636 79d205afdb	refusal gone; stops at 38 s on sampled image/sampler pair limit exceeded
+ #9633 984e14dc49	refusal gone; stops at 38 s on sampler resource limit exceeded
+ #14636 + #10000 a803318e65	no refusal; runs the full 300 s at about 27 presents/s

The last build is not gameplay yet: every screenshot from 60 s to 285 s is pure black, while the boot screen at 25 s matches main. #14636 and #10000 conflict only on ShaderDiskCache.hpp's FormatVersion, which I set to a new value for the test. #9633 + #10000 also conflicts in docs/dev/debt/shader.md and was not run.

AI-assisted: yes (Claude Code).

[JJenkx]

Follow-up: adding #6959 (head f0bf17912a, the Unreal Engine 4 lighting fix) to the #14636 + #10000 build does not change the black screen. It runs the full 300 s at about 28 presents/s, and every screenshot from 60 s to 285 s is still pure black. The black frames are not the lighting issue; finding their cause needs a frame-level capture.

AI-assisted: yes (Claude Code).

[JJenkx]

More on the black screen. The build is main + #14636 + #10000 + #6959, plus #4449 88b99b28b6.

• Rendering is happening: the draw profile shows about 18,000 draws per 10 s, and every present comes from the scanout buffer's resident GPU image. So the frames are rendered, but the presented image is black.
• Import failures are not the cause: feat(libSceAgcDriver): import scattered direct memory as a dma-buf run list and probe its coherence #4449 removes the roughly 1,770 host-import failures (to 0) and raises the rate to about 32 presents/s, but the frames stay black.
• Pending storage results are not the cause: APS5_STORE_AT_FLIP=1 changes nothing.
• Input is not the cause: Options, Cross, Circle, Triangle and Square change nothing either. The boot image is still shown at 45 s, and the screen is black from about 60 s.

The next step is to find which pass writes the scanout buffer, for example with APS5_DUMP_DRAW_SHADERS on the flip address.

AI-assisted: yes (Claude Code).

[JJenkx]

Narrowed further, using the dumped draws (APS5_DUMP_DRAW_SHADERS=8fc0000000, display buffers at 0x8fc0000000 and 0x8fc2000000) and #15001, which writes a draw's fragment program for offline disassembly. Same build as in my previous comment.

Of the 19 draws into the display buffer, four are full-screen triangles. The fragment program of one of them (0x200e100000, behind vertex program 0x200e270000) is a typical Unreal Engine 4 tonemap pass:

• it samples the scene colour and two other 2D images;
• it multiplies them by one scalar loaded from a constant buffer at offset 0x878, likely an exposure value;
• it grades the result through a 3D lookup table (image_sample ... dim:SQ_RSRC_IMG_3D);
• it exports to mrt0.

So pure-black frames most likely mean that scalar is 0 or the 3D table is all black.

APS5_DUMP_STORAGE=32x32x32 dumped nothing, so if the table is 32x32x32 it is not a storage image here. It may be drawn as a layered render target. Next steps: read the fragment program's user data and constant buffer at 0x878, and find the pass that writes the 3D table.

AI-assisted: yes (Claude Code).