Repository navigation
Lies of P (PPSA10621 / PPSA10261) #22009
Replies: 1 comment
Update: Workaround found to get past the BaselineMiss::UploadCopyDiffers crash (progressed past first boss into Hotel Krat)Baseline Crash: Every baseline run aborted after ~215–554 s (2,604–4,115 frames) with: With local [agc-unsaved-debug] diagnostics, the failing surface was consistently a 512x512 512 KiB (0x80000, 8 × 64 KiB units, fmt=7, tile=4, blockUnits=1, dcc=0x0) storage image backed by a 512 KiB host import. Because 512 KiB is below settleFloorBytes (1 MiB / 0x100000 in GuestBufferMemory.cpp), those imports leave the write watch on Windows and rewatchSettledImports keeps them compared while a live StorageTexture overlaps them. Once unwatched, CPU/GPU reuse over those units later triggered BaselineMiss::UploadCopyDiffers (or UnwatchUnproven during re-sync) when markLayersPending / writeBackLayers ran. Temporary Workaround Tested:Locally lowering settleFloorBytes in core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp from 0x100000 (1 MiB) to 0x80000 (512 KiB) keeps Lies of P's 512 KiB imports synchronously settled and write-watched so they never enter the unwatched generationBaseline path. Result : 2 consecutive crash-free runs, including a 1331.2 s (~22 min) clean run (8,458 frames presented, exit code 0, 0 BaselineMiss errors) from the beginning area through defeating the first boss and loading into Hotel Krat.Why this is only a temporary workaround:It only bypasses the precondition for >= 512 KiB imports — it does not fix the underlying generationBaseline / CpuWroteSince / unrefreshed markLayersPending bug itself, so any < 512 KiB storage surface would still be exposed.
|



Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Existing thread
Title
Lies of P (PPSA10621)
Content version tested
01.012.000
AnyPS5 commit
main d31916e + local revert of #16053 + local sceSaveDataUmount2 backup/event fix (branch fix/savedata-umount2-backup on fork; PR will be opened and linked once my current first open PR merges)
System tested
Windows 11 (build 26300), Intel Core i9-9900K, 32 GiB RAM, AMD Radeon RX 6800 (AMD proprietary driver 26.10.1, Vulkan 1.4.315), MinGW-w64 GCC 15.3.0 (Release)
Furthest point
In game
What is shown
Reaches the main menu, loads into a New Game, plays through the starting train, build selection, Krat Central Station (including combat and defeating the first miniboss), opens the door into the second area, and reaches up to the second checkpoint cleanly.
#16053is reverted.~2–5 FPSin heavier scenes with occasional >10 s shader/storage-image stalls).Account1/Account2/OptionSlotsave creation, backups, and checkpoint autosaves all work cleanly. (Audio: [worked normally / had issues — update before posting].)Current blocker
Blockers on plain
main(d31916ed37)#16053(tracked in The Pathless (PPSA01825): upside-down picture and garbage-descriptor crashes since #16053 #20835 , fix open in fix(agc): start the viewport control and color masks at the AGC register defaults again #20827 ); bypassed locally by reverting#16053.Account1,OptionSlot) and New Game save creation (Account2),TaskGraphThreadHPcallssceSaveDataUmount2(0x10000, "/savedata0")while holding the game's save mutex and pollssceSaveDataGetEventResultforSAVE_DATA_EVENT_TYPE_UMOUNT_BACKUPwhileGameThreadwaits on that mutex. PreviouslysceSaveDataUmount2ignoredmodeand never backed up or queued the event (similar symptom to#5983onPPSA10261). Fixed locally infix/savedata-umount2-backup; PR will be opened and linked here once my first open PR is merged.Current in-game blocker (with the two fixes above applied)
Intermittent
libSceAgcDriverabort after ~8–10 minutes of play / VRAM pressure (not deterministic on a specific action; restarting from a checkpoint passes the same fight cleanly, though longer runs still eventually hit it):(Also observed on
unit 4 of 8near the second checkpoint, followed by secondary teardown exceptionVideoOut: presentation thread cannot stop itself.)Diagnostics on the crashing surface point to a
512x512fmt=7tile=4(kR64KBX,0x80000) storage image where low units left the Windows write watch (overlapping< 1 MiBhost imports) whilebaselineDuestayed set across GPU write-backs andgenerationBaselinestill held the initial0xffffffffupload copy (UploadCopyDiffers). No existing open PR was found targeting this exactBaselineMiss::UploadCopyDifferspath.How it was run
relinker --windows ---to-intel eboot.bin app.exe
launched
app.exe -vwith no special runtime environment flags required once the two local fixes above are applied (APS5_OUTPUT_RESOLUTION=1920x1080andAPS5_REPORT=1/APS5_TRACE_BENCH=1were used only to capture metrics).All reactions