Ready or Not (PPSA18605)

[Nesyth]

Tested

• Title: Ready or Not, PPSA18605, v01.001.008 (Unreal Engine 5)

• Host: Windows 11 Pro 10.0.22631, AMD Ryzen 7 7800X3D, 32 GB RAM, AMD Radeon RX 7900 XTX.

• Build: main at bab7b32, WinLibs GCC 15.2.0, Release, plus:

  • fix(libkernel): restore Windows select build for Demons Souls testing #4021 and fix(recompiler): include implicit inputs in relative vector windows #4057, and commit 1dc342d (all in main since);
  • open pull requests feat(agc): render layered draws into 3D color targets #1077 (with feat(agc): vertex clip and cull distances #423), feat(agc): depth planes sampled as 2D arrays, cubes and integer views #425, fix(libSceAgcDriver): retire a depth surface once another writer covers its memory #454, feat(shader): give reads the masked-select pass keeps the whole pixel ancillary word #4031, perf(libc): walk shared views by iterator and remember clean pages per run in the write collect #3864, perf(agc): clear integer-format storage images with DCC code 0000 on the GPU #3865, perf(agc): handle the zero dword fill kernel as a buffer fill #4012, perf(agc): sort guest buffer regions through indices before merging them #4248, fix(agc): order direct reads after copied writes #1359, perf(agc): disambiguate completion label overlaps in CopyBuffer #4496, perf(agc): keep adjacent ranges one import serves as one region of the cached address space #4428 and perf(agc): maintain draw fingerprints on register writes #1312, and fix(libc): alias non-contiguous shared guest views on Windows #2254 (closed; its code is now draft fix(libc): alias non-contiguous shared guest views on Windows #4367);
  • my draft pull requests perf(libc): share one snapshot of the registered guest ranges between leases until the registry changes #4698 and fix(agc): read the DCC keys of sampled textures after the GPU writes recorded before them #4699, and an earlier version of fix(agc): read comp-to-single DCC keys of sampled textures and storage images #4700;
  • local changes:
    • view imports (the cause in AGC: a buffer that crosses registered ranges outside a whole import window is copied on the CPU at every use (Palworld: 15–56 GB per 10 s) #4260);
    • two changes to the waits in fix(agc): order direct reads after copied writes #1359's CompletePendingCpuWrites (see the notes on blocker B);
    • a relinker fix for modules whose names differ from their files (fix(relinker): load needed providers for unscoped Windows imports #3365 should replace it);
    • NpAuth answering signed out (feat(libSceNpAuth): implement offline auth requests #3604 implements offline auth);
    • game mouse scaling (as feat(libSceVideoOut): capture the host cursor for guest mice and scale their deltas #4217);
    • held pad sticks (now in main as feat(pad): hold the sticks from the controller command file #4032).

  The eboot and modules are fSELFs, unwrapped to plain ELF before relinking with --windows --registry.

• Last update: 2026-10-12.

Status

In game.

• Path: the title boots, shows the PSN sign-in dialog and the language selection, and reaches the main menu. Continue loads the station. The briefing room desk opens Mission Select. The first mission ("Thank You, Come Again", 4U Gas Station) loads and can be played through slowly, with the squad following orders.
• Speed: with APS5_PROFILE_DRAW=1, about 6 FPS in the menu and about 4 FPS at the station once it settles (39–43 presents per 10 s), or 3.0–3.5 in a session with blocker F. The mission ran at 2.4–2.7 FPS on an earlier build.
• Loading: the first load of the mission level took about 14 minutes while its shaders were translated and pipelines compiled. With warm caches, boot to the station takes about 3 minutes.

Test-only settings:

• APS5_HOST_IMPORT_MIB=10240: while the level loads, frames come about 2.4 times faster than with the default 6 GiB (1.1 vs 0.45 FPS).
• For the first level load, download0/readyornot/saved/ps5/engine.ini containing [SystemSettings] and g.TimeoutForBlockOnRenderFence=1800000 (see the stops below).

The station lobby in a session without blocker F.

The same spot in a session with blocker F.

The first mission (earlier build): blocker F on the canopy and the wet ground.

First post (2026-10-12): black silhouettes, coloured patches, and a fine grid over the image.

Open blockers

	Blocker	Seen
B	Performance: CPU-bound on the GPU packet thread (see the notes under this table).	Every launch
D	saved/ps5/gameusersettings.ini is sometimes left as a single [, and the next launch aborts with APR: read of …gameusersettings.ini at offset 0 returned 1 of 5480 bytes. Restoring the file and marking it read-only avoids it. Cause in the notes below.	7 of the first 14 launches
E	A fine grid over the image (part of A in the first post).	Mostly the menu now
F	A colour cast that lasts the whole session: green, or purple and blue, glow in the air around lights, and tinted characters. In the mission, also rainbow blotches and white splashes on reflective surfaces. These are the coloured patches under A in the first post.	8 of 22 station loads on 2026-10-12; the mission's one load

Notes on blocker B, per 10 s at the station once it settles:

• About 56,000 draws at about 95 µs each (7 s), 1.1 s in dispatches and 0.1 s in DMA_DATA packets, all on the queue 0 packet thread. The host GPU is busy for about 1.9 s of those 10.
• perf(agc): disambiguate completion label overlaps in CopyBuffer #4496 cut DMA_DATA from about 1.2 s to 0.1 s. Together with fix(agc): order direct reads after copied writes #1359, it added about 1.1 s of GPU waits per 10 s in CompletePendingCpuWrites, because fix(agc): order direct reads after copied writes #1359's label check there also matched labels the GPU stores in place. A local change waits only for labels that a completion stores on the CPU, and that site no longer shows among the top sync sites. I'll describe it on fix(agc): order direct reads after copied writes #1359.
• Turning Lumen GI off (r.DynamicGlobalIlluminationMethod=0 in engine.ini) makes the station about 30 % faster.

Notes on blocker D:

• Ready or Not keeps a copy of gameusersettings.ini in its "GameUserSettings" save slot. On every launch it writes that copy over the local file, then reloads the file.
• UE's PS5 file layer resolves each path once per process through sceKernelAprResolveFilepathsToIdsAndFileSizes and keeps the size. The reload therefore asks for the size of the first read, 5480 bytes.
• AnyPS5's APR read throws on the short read, which aborts the title. UE expects such a failure in the APR result buffer and logs Apr read failure there instead; ampr_emu reports SCE_AMPR_ERROR_APR_INVALIDFILEOFFSET (0x8191203A).
• In the failing launches the saved copy itself was a single [, so every launch wrote it back. How the first 1-byte copy appeared is not known.
• Not tested yet: whether Ready or Not carries on when the error is reported that way.

Notes on blocker E:

• The grid follows UE's temporal upscaler jitter (Halton 2,3 over 17 frames), so it is a pattern TSR leaves in the image.
• It became 2–4 times weaker after the DCC fixes below. It is now mostly in the menu, faint at the station, and not seen in the mission.

Notes on blocker F:

• It starts when the level loads and stays for the whole session. The colour differs between sessions.
• It is not tied to one rendering feature. It still showed with each of these turned off in engine.ini:
  • Lumen GI (r.DynamicGlobalIlluminationMethod=0): 1 of 3 runs;
  • Lumen reflections (r.ReflectionMethod=0): 1 of 1;
  • volumetric fog (r.VolumetricFog=0): 1 of 1, faint.
• It is not saved state: the save data and the shader cache were byte-identical in clean and glowing sessions.
• The two 2612x1472 RGBA8 surfaces with mixed DCC keys (the first 36864 key bytes 0x00, the rest 0xff; texels read as stored) are present in clean and glowing sessions alike, so they don't explain it on their own.
• Sessions with F run 10–25 % slower.

Fixed since the first post:

• A, lighting (black silhouettes, missing city lights, black logo and menu text): gone after merging perf(libc): walk shared views by iterator and remember clean pages per run in the write collect #3864, perf(agc): clear integer-format storage images with DCC code 0000 on the GPU #3865, perf(agc): handle the zero dword fill kernel as a buffer fill #4012 and perf(agc): sort guest buffer regions through indices before merging them #4248 locally; the same build without them stayed black. Not bisected.
  • perf(agc): handle the zero dword fill kernel as a buffer fill #4012 (the zero dword fill kernel handled as a buffer fill) is the likely fix, so this looks like a DCC key race.
  • That build also had fix(agc): read the DCC keys of sampled textures after the GPU writes recorded before them #4699, which did not fix it on its own. fix(agc): read the DCC keys of sampled textures after the GPU writes recorded before them #4699 fixes a related race, where a sampled texture read its keys before the GPU writes recorded ahead of it.
• C, comp-to-single DCC keys read as stored: fix(agc): read comp-to-single DCC keys of sampled textures and storage images #4700.
• Performance, from under 2 FPS to about 4 at the station:
  • perf(libc): share one snapshot of the registered guest ranges between leases until the registry changes #4698: about 2.0 to 3.0–3.3 FPS, measured with and without it on the same build.
  • A local change imports buffer views that cross adjacent registered ranges instead of copying them on every draw (about +40 %). The same cause is AGC: a buffer that crosses registered ranges outside a whole import window is copied on the CPU at every use (Palworld: 15–56 GB per 10 s) #4260, so it is not a pull request. It needs fix(agc): order direct reads after copied writes #1359: without it, copied writes race with in-place reads and give a colour haze.
  • perf(agc): disambiguate completion label overlaps in CopyBuffer #4496, perf(agc): keep adjacent ranges one import serves as one region of the cached address space #4428, perf(agc): maintain draw fingerprints on register writes #1312, 1dc342d and the local CompletePendingCpuWrites changes: about 3.4 to 4.1 FPS.

Stops on main, in the order they are hit

Stop	Covered by
Unresolved imports of modules whose names differ from their files, such as libSceAppContentUtil, provided by libSceAppContent.prx	#3365 (merged) should cover it: with unscoped imports, the startup stub now loads every needed library. Not yet confirmed by a startup run on main; my build carries a local relinker fix for it
packed pixel ancillary input has an unsupported live use (used by SelectU32 ? ?)	#4031
color views of several array slices are unsupported (layered draws into 3D color targets)	#1077 (includes #423)
sceNpAuthGetIdTokenV3 not implemented	Local only: answers signed out, so the title shows the PSN dialog and carries on (#3604 implements offline auth)
storage image access to depth/stencil surface … is not implemented	#425
With #425: storage image access to depth surface … as a 2612x1472 image of vk format 97, dimension 1, 6 mips is not implemented	#454
Level load: the default 6 GiB host import budget fills, and imports past it are refused and copied by the CPU instead. The load still works as far as the briefing, at about 0.45 FPS.	Test-only APS5_HOST_IMPORT_MIB=10240. #290 (open, not tested here) adds least-recently-used eviction under the budget
First level load: GameThread timed out waiting for RenderThread after 120.00 secs, then an access violation, while the level's shaders are translated and their pipelines compiled	Test-only engine.ini setting above; later loads hit the caches
Level load, with the raised import budget: read-write alias of shared guest memory 0x328fb20000+0x8080000: the range is not one contiguous run of views of a section	#4367 (draft), which carries closed #2254's code unchanged; my build uses #2254

The stops before the level load were found on main 7028457 and 09bd4aa (2026-10-11). Apart from the relinker one, they were not re-checked on bab7b32 without the pull requests that cover them. Three others hit then are fixed on main since: select on descriptors that are not sockets, _umtx_op, and an incomplete rect-list primitive.

[Nesyth]

Update (2026-10-12):

• Lighting (A) is fixed by four open PRs merged locally (perf(agc): handle the zero dword fill kernel as a buffer fill #4012 most likely; not bisected).
• The first mission ("Thank You, Come Again") loads and can be played through slowly. The station now runs at about 4 FPS, up from under 2.
• New draft pull requests from this bring-up: perf(libc): share one snapshot of the registered guest ranges between leases until the registry changes #4698 (performance), fix(agc): read the DCC keys of sampled textures after the GPU writes recorded before them #4699 and fix(agc): read comp-to-single DCC keys of sampled textures and storage images #4700 (DCC keys).
• The cause of blocker D is found: UE's PS5 file layer keeps each file's size, and AnyPS5's APR read throws on the short read that follows when the title rewrites its settings file.
• Still open: a colour cast that lasts the whole session in about 1 station load in 3 (blocker F, screenshots in the post), and a faint TSR grid, now mostly in the menu (E).