Silt ps5 (PPSA03306) on windows: boots and plays with four local changes”

[billythemenace2]

Status: boots, reaches the main menu, starts a new game and plays through several areas on Windows with four local changes, with sound. Open: no antialiasing (workaround) and one rendering artefact.

Setup

• Title: Silt, PPSA03306 (EP2520-PPSA03306_00-SIEESILTFULLGAME), version 01.003.000, master 01.00. Unity (IL2CPP), FMOD.
• AnyPS5: main at ba2a0a31 plus the attached patch (12 files), built with WinLibs GCC 15.2.0 r7 (x86_64-ucrt-posix-seh).
• Host: Windows 10.0.26200, AMD Radeon RX 5700 XT. The window title shows 44 to 52 FPS in game.
• Inputs: eboot.bin and all modules unwrapped from SELF to plain ELF. Media/Modules/*.prx copied into sce_module/, Media/Plugins/*.prx copied into prx/.
• Command: relinker --windows --windows-diagnostics --registry eboot.elf app.exe, run with APS5_IGNORE_MSAA=1 (added by the patch).
• Sound: audible in the menu and in game.

What works with the patch

• All 15 bundled modules and every system library load; no absent imports.
• FMOD opens its four banks and audio plays.
• Save data mounts and GameSave.bin is written.
• Menu, new game, gameplay, and area transitions.

1. Multisampling

Unpatched, the first scene draw is rejected:

[videoout] presentation failed: multisampling or coverage conversion is unsupported (register 0x2f8 = 0x0020c002)

The scene target is 4x (2x once the saved settings load: 0x00108001), uncompressed: CB_COLOR_INFO 0x00040718, CB_COLOR_ATTRIB 0x00012000. Each frame ends the scene with a hardware resolve, CB_COLOR_CONTROL 0x00cc0030 (mode 3), from color 0 to a single-sample color 1 that the bloom chain then samples.

Local workaround, behind APS5_IGNORE_MSAA: multisampled color and depth targets are decoded as single-sample surfaces at the same address, and the mode 3 draw becomes StorageTexture::CopyFrom from the source image to the destination. No antialiasing, but the frame is otherwise complete.

2. Modules the title loads by path

Unity loads /app0/Media/Modules/PS5Util.prx and Il2CppUserAssemblies.prx (both needed by the executable) and its plugins under /app0/Media/Plugins/ with sceKernelLoadStartModule. The relinker writes them under app0/sce_module and app0/prx, so dlopen finds no <path>.guest.prx:

unable to load /app0/Media/Modules/PS5Util.prx, error:0x80020002
unable to load /app0/Media/Modules/Il2CppUserAssemblies.prx, error:0x80020002
...
FATAL: unhandled exception 0xc0000005
  execute of 0x0000000000000000

On Windows a second image loaded from that path would also have no imports bound, since the entry stub binds them. Local change: Windows dlopen first looks for an already loaded <file name>.guest.prx and returns that module.

3. Plugin module_start arguments (Windows)

The entry stub calls DT_INIT of every guest module with (0, NULL). Unity's PSN plugin needs the arguments Unity passes with its load request:

ERROR...
 PSNCore is an old version that cannot be used by the current player runtime.
FATAL: unhandled exception 0xc0000005
  rip PSNCore.prx.guest.prx+0x113b4
  read of 0x0000000000000004

Local change: for a module under prx/ that neither the executable nor another module depends on, the stub skips DT_INIT and DT_FINI, the guest image exports its DT_INIT, and dlopen calls it on the first load request with the pending sceKernelLoadStartModule arguments.

4. Stall while loading: deadlock between signal delivery and the guest mapping lock (Windows)

Before the change the title stalled while loading the first level or the next area (three runs). All file, texture and shader activity stops and Unity's watchdog prints, until the window is closed:

Wanted to force a call to sce::Agc::suspendPoint but not safe...(66 seconds)

Thread dump taken with APS5_HANG_DUMP_SECS (log attached; offsets are from the local build, names resolved with nm):

• 31 threads are inside the title's signal 30 handler (Il2CppUserAssemblies+0xe2405a, the garbage collector stopping the world), waiting in sceKernelWaitEventFlag: 29 delivered through WaitingEntry, 2 through RedirectedEntry.
• One redirected thread was interrupted inside GuestArenaHandleWrite:

thread 3524 'Job.Worker 12'
  libkernel.prx+0x3af6e   sceKernelWaitEventFlag
  Il2CppUserAssemblies.prx.guest.prx+0xe2405a
  libkernel.prx+0x4338c   RedirectedEntry
  libkernel.prx+0x43df9   Aps5RedirectedEntry
  libc.prx+0xab69         ReportCrash, return address of its call to GuestArenaHandleWrite
  ntdll.dll+0xcc196

• Three threads, the process's first thread among them, wait in the winpthreads mutex lock path (libwinpthread-1.dll+0x453c).

Reading: the worker was stopped while it held WindowsMappings::mutex, so every later write fault waits for a thread that only the end of the collection resumes. This is the case TechnicalDebt already lists for sceKernelRaiseException on Windows (a handler run where the thread was interrupted can deadlock on a lock of the emulator). That the three waiting threads are on this mutex is inferred; the dump has no deeper frames for them.

Local change: WindowsMappings uses a lock that records its holder and waits by yielding, and RaiseOn resumes the target and retries while it holds that lock (30 s limit). After it: no stall in a run crossing several area loads. Other host locks have the same exposure and are not covered.

5. Open: rendering

• No antialiasing, from the workaround in 1.
• When a pillar breaks, the dust cloud ends at a straight vertical line, at the same x as the edge of the textured background layer, and the loose pieces look translucent. Nothing is logged and no draw is rejected. Not compared against the console and not investigated; it is unknown whether the workaround in 1 is involved.

Smaller points

• sceSysmoduleLoadModule logs unknown ids 272 and 295.
• The patch is attached for reference, not as proposed fixes: it bundles four topics, is commented throughout, and 1 is a workaround. The relinker part passes the existing guest module tests; the Windows parts were only run with this title.

Attached: anyps5-silt-bringup.patch, silt-run7-trimmed.log (the stall with the thread dump; the repeated watchdog lines are cut). Investigated and patched with AI assistance (Claude).
anyps5-silt-bringup.patch
silt-run7-trimmed.log

[atikrh-beep]

Could you share the relevant log output and specify which windows version you're using? It would also help to know whether the issue occurs consistently or only with a particular module. With those details, it may be easier to identify whether the problem is related to module loading, rendering, or a possible deadlock.

[billythemenace2]

Windows version: Windows 11, build 10.0.26200. GPU is an AMD Radeon RX 5700 XT.

Logs: the stall log with the full thread dump is attached as silt-run7-trimmed.log at to bottom of the post.

[atikrh-beep]

Since you're using an AMD Radeon RX 5700 XT, it would be helpful to determine whether the stall is related to GPU rendering or a thread waiting on another operation.

Could you check the thread dump for the thread that appears to be stalled and identify what it is waiting on? If the log shows a thread blocked on a mutex, condition variable, or graphics-related call, that could help narrow down the cause. Comparing the same workload with any available debug logs or a different rendering configuration may also help isolate the issue.

[billythemenace2]

Windows version: Windows 11, build 10.0.26200. GPU is an AMD Radeon RX 5700 XT.

Logs: the stall log with the full thread dump is attached as silt-run7-trimmed.log at to bottom of the post.