SAROS - PPSA07631

[rhya119]

I will update as I go

[rhya119]

posting to menu for now

[rhya119]

Progress update (2026-10-10): I implemented separate depth and stencil storage-image seeding and writeback. I added a GPU regression test covering both planes, read/write persistence, and pending fast clears. I ran agc_driver_depth_fast_clear, agc_driver_depth_fast_clear_submit, and agc_driver_depth_storage on Windows; all three passed.

I committed this locally as 1697da56 (fix(agc): keep depth and stencil storage writes independent) on saros-sparse-memory-20261009. I have not pushed it to boykopovar/AnyPS5 because my account currently has read-only access there. These tests verify driver behavior only; I have not yet reverified SAROS reaching the main menu.

[boykopovar]

Do you know what stage the game has reached on the current main branch?

[rhya119]

I haven’t verified SAROS on an unmodified current main yet. My local branch is based on d70b899 with additional upstream PRs and local fixes.

That build plays the intro and reaches the language/subtitle setup screen at approximately 60 FPS, with keyboard input working. After Continue, it enters further rendering but has hit a GPU hang/device-loss error. I haven’t verified the main menu or gameplay.

I’m testing the coherent BDA load fix from PR #2259 now. Correct textures also remain unverified because the current diagnostic run uses the bindless overflow fallback.

[JJenkx]

Linux/RADV update for SAROS PPSA07631 v01.000.000:

Unmodified main a3aeaf884b converts and loads, but aborts in presentation after about seven seconds with [videoout] presentation failed: YNw9Xk5iNWo not implemented. This is a host symbol collision: internal MouseIsOpen_nid_postfix hashes to the same NID as the VideoOut stub added by merged #5608. Patched VideoOut imports and defines that same name, so its mouse-capture helper call resolves to the throwing stub. This does not establish a title call to the unknown VideoOut API.

With only the three-line internal helper rename in fork PR55, the same converted title passes that stop and reaches the language/subtitle setup screen in a verified90-second snapshot. The patched binding check and existing mouse_input test pass; full Linux/Windows CI has passed (1,068 Linux /735 Windows tests; expected skips remain unvalidated). RX6650XT/RADV, 3840x2160 guest output.

No input yet, so Continue, subsequent menus and gameplay are not verified. No game files or screenshots attached. AI-assisted: yes.

Follow-up: Cross/Continue advances to the cycle-saving information screen and writes settings. Continuing from there reaches a separate stop: AGC driver: DISPATCH_DIRECT at DWORD 4169: dispatch modifiers 0x49 are not implemented (SIGABRT). The frozen window counter was stale during core collection, not steady rendering. Upstream draft5716 contains only the mouse-helper fix; existing PR5176 implements ordered-append dispatch support and is the next local integration to validate. No title-menu/gameplay claim.

Ordered-append integration result: main a3aeaf884b plus only the mouse rename and #5176 (bb17e61105fe) passes the earlier0x49 modifier rejection. The three focused PM4/wave-decode/GDS execution checks pass on AMD/RADV, zero skips. The same Continue sequence then aborts with bindless image table has198656 entries (0 materials), limit16; a deferred BDA access error preceded it. Title-menu and gameplay remain unverified. Existing stacked #757/#758 are being built locally for descriptor-table validation; no result is claimed for them yet.

Descriptor-table integration result: stacked #757/#758 (5595a163c0d0) builds with the prior local integration; eight focused table/cache/shader-memory/BDA checks pass on AMD/RADV, zero skips. The same Continue sequence no longer reports the198656-entry capacity rejection. It aborts later on computed/data-dependent s_swappc_b64 call target at program counter0x130 is not statically resolvable; the deferred unaligned BDA access fault also persists. Existing #4706 is the next local integration under validation. Still no verified title menu or gameplay.

Latest-main follow-up (31bfc4e869 plus the recorded local peer stack): #4706 passes scalar-call, compute-table-call, shader-memory serialization and BDA tests on RADV after local conflict resolution, but its native benefit is not established: an earlier run stopped on the nonzero AJM ratio ramp before the shader-call point. With #5647 added, eight focused tests pass and that ratio rejection does not recur. The new native stop is sceAudioPropagationSourceGetAudioPath: index0 out of range (0 audio paths). A behavior-preserving diagnostic records source creation immediately followed by path0 lookup, before source attributes, path calculation or path-list updates. System-option header is0x010107d5, size56. This identifies source-owned configurable path state as the next implementation requirement. No title menu or gameplay claim; no game files or raw logs attached.

New local library implementation: fork draft56 adds source-owned configurable path handles and copied path data. The observed source-create/path0 lookup and setter now pass. Saros configures18 empty Invalid slots and calls render with mono float PCM (4096 bytes) and format0 output (16384 bytes, 1024 four-channel frames). Inactive-path rendering now preserves the caller's existing mix for the documented layouts; active nonempty rendering still throws. Host/project tests pass, including stale-handle and untouched-output cases; full Linux/Windows CI is running. The fixed36-slot capacity follows the public maximum and remains an allocation/count-model inference, not a console measurement.

The updated native integration passes those audio stops and reaches the computed s_swappc_b64 target at pc0x130 again; #4706 does not resolve this shader's shape. The unaligned BDA access fault persists. No title menu, gameplay or active acoustic rendering is verified. No game assets, screenshots or raw logs attached.