Disney Epic Mickey: Rebrushed (PPSA15657) on Linux

[johnmaia]

Tested

• Title: Disney Epic Mickey: Rebrushed, PPSA15657 (EU), v01.000.000 (master 01.02). Unreal Engine 4, FMOD, Bink 2. Relinked with its own sce_module and prx (libfmod, libfmodstudio).
• Host: Ubuntu 24.04.5 (kernels 6.8 and 7.0), Intel Core i3-12100, AMD Radeon RX 7600 (RADV, Mesa 25.2.8), 32 GB, GCC 13.3 Release build. Headless Xwayland at 1920x1080. Input is injected with XTest; audio is captured with SDL's disk driver.
• Last update: 2026-10-09

Status

Reaches the first level and gameplay on Linux. Warnings, logos and the opening cinematic play. "Press X to start" opens the main menu, and Continue loads Dark Beauty Castle. The in-engine opening scene in the Mad Doctor's lab plays with sound, and then gameplay starts with Mickey in the third-person camera. One run stayed in the level for 44 minutes without an error, until the test script's time limit stopped it.

Two things are needed for that today:

• APS5_NO_HOST_IMPORT=1. With host import on, pressing Cross at the title ends in an Unreal fatal error in most runs (UE4 title (PPSA15657): LowLevelFatalError "RemoveFromOuterMap NumRemoved = 0" when the title screen opens the menu #2365, open issue E).
• The testing-only stand-ins listed under How to run it.

The image fills only the middle 1152x648 of the 1920x1080 window (open issue G).

Clips with sound (2.5 min each, about 5 fps): opening scene, gameplay.

Earlier screenshots:

How to run it

The mickey/testing branch is main with the pull requests below merged, plus these testing-only commits. It is a testing branch, not for merging.

• Layered color views render into their first slice (Layered rendering: a draw into all slices of a 3D color target aborts with "color views of several array slices are unsupported" #2102).
• A CB metadata pass over mixed DCC keys keeps the resident texels (B).
• A T# over a depth surface's memory in another layout is read as a plain texture (Sampling a D16 depth surface's memory as a half-width 32-bit texture (PPSA15657): "sampling the depth plane ... is not implemented" #2366).
• On top of feat(agc): vertex clip and cull distances #423: vertex programs are prepared with the position-export bits of PA_CL_VS_OUT_CNTL, and prepared at draw when a layout was not prepared ahead (details on feat(agc): vertex clip and cull distances #423).

Run with APS5_NO_HOST_IMPORT=1. The first launch compiles about 4,100 shaders. The title screen comes up after about 15 minutes, with or without the shader cache, because loading dominates.

Performance

First level

Measured in the opening scene, 5-15 minutes after Continue, with host import off:

• The GPU is about 1.4% busy: 140 ms of GPU work per 10 s.
• One thread is the limit: the GPU submission thread runs at about 95% of a core and holds the queue's GPU mutex 9.8 s of every 10.
• That thread is mostly writing GPU results back into guest memory. The level's compute shaders rewrite mirrored buffers, about 1 MB per write-back at about 500 write-backs per second. Writing back only the bytes that changed (GuestMemory::WriteChanged) took 43% of the thread, at about 1 ms per MiB. Together with the memcmp and memmove of refreshing the mirrors, memory sync was about 59% of the process's CPU time.

#2481 stores those changes 16 bytes at a time (SSE2 MASKMOVDQU for partly changed chunks):

Opening scene, 5-15 min after Continue	Before	With #2481
Average fps (300 samples)	4.60	6.92
Write-back cost	~1.0 ms per MiB	~0.26 ms per MiB
Write-back share of the submission thread	43%	17%

The rest is the round trip itself: each heavy dispatch copies a buffer in from guest memory, runs, and writes it back. Host import avoids that by sharing the memory, so fixing #2365 should matter for speed as well as stability.

Boot and title screen

On Linux kernels before 6.11, GuestMemory::describePages has no PROCMAP_QUERY and re-reads /proc/self/maps on every query. This title has about 5,400 mappings (518 KB of text), and about 85% of the busiest thread went into parsing it. #1810 (with #861) caches the results, as the Windows path does:

Same build, 25 min, kernel 6.8	Without #1810	With #1810 + #861
Logos and intro movies	5-6 fps	57-60 fps
Title screen	not reached (3 fps in the cinematic)	reached at ~14 min, 11.4 fps
Frames in 25 min	4,323	23,801

On kernel 7.0, which has PROCMAP_QUERY, the build without #1810 runs about the same as with it on 6.8. #1810 matters on kernels before 6.11, which includes Ubuntu 24.04's default kernel.

Open issues

	Issue	Status
A	Layered rendering into a 3D color target: a 48³ RGBA16F volume filled in one rect-list draw (#2102)	Testing-only stand-in renders the first slice; volume textures are wrong until it's implemented
B	CB metadata pass over mixed DCC keys during the first level load	Per-block DCC metadata is not modeled. Issue not filed yet; testing-only stand-in on the branch
C	Compute shaders that call fetch routines through s_swappc_b64 (#1817)	Deferred by #1836; about 380 of them, none dispatched so far
D	About 5-7 fps in the first level	GPU submission thread writing back mirrored buffers; #2481, then host import (E)
E	RemoveFromOuterMap NumRemoved = 0 fatal error when Cross opens the menu (#2365)	Avoided with APS5_NO_HOST_IMPORT=1: 0 of 5 runs, against 5 of 8 with import on. Cause not found yet
F	A D16 depth surface's memory sampled as a half-width 32-bit texture (#2366)	Testing-only stand-in reads it as a plain texture
G	The image fills only the middle 1152x648 of a 1920x1080 window	Not investigated

Pull requests

Everything the title needs on top of main:

Shader preparation:

• fix(agc): defer shaders that cannot be prepared at registration to their first use #1836 defer shaders that cannot be prepared at registration (the Ghost of Yōtei: compute shader calls a user-data function pointer with s_swappc_b64 (not statically resolvable) #1817 family)
• fix(agc): skip registration-time preparation for unsupported shader types #1893 skip registration-time preparation for unsupported shader types (Registering a shader of type 8 aborts the title: "AGC driver: unsupported registered shader type" #1892)
• fix(agc): prepare the null pixel program at draw when its variant was not prepared #1902 prepare the null pixel program at draw (Draw with no pixel program fails: "prepared shader artifact is missing for the requested static ABI" (null pixel program) #1901); fix(agc): prepare a rect-list draw's missing rectangle at draw from its prepared stages #1928 its rect-list rectangles (Rect-list draw aborts with "prepared rectangle artifacts are missing" when its pixel variant was prepared at draw #1927)
• fix(agc): accept bound copies of a registered shader header that differ only in user_data #1837 bound copies of a registered shader header
• fix(agc): read an unwritten pixel program address as unset #2022 read an unwritten pixel program address as unset

Shaders:

• fix(shader): decode the scalar aperture sources 235-238 instead of failing the shader #1940 decode the scalar aperture sources 235-238
• fix(shader): lower the packed pixel ancillary when its extract operands fold after it #1904 and fix(shader): carry a packed pixel ancillary through phis when every reader extracts a modelled field #1934 the packed pixel ancillary (Recompiler: "packed pixel ancillary input has an unsupported live use" for a plain v_bfe_u32 of the render target index #1903, Recompiler: packed pixel ancillary rejected when the shader reuses its VGPR ("unsupported live use") #1933)

Rendering:

• feat(agc): depth surfaces read as textures and storage images #425 depth surfaces read as textures and storage images (AGC: storage image access to depth/stencil surface 0x201a800000 is not implemented #1409)
• feat(agc): vertex clip and cull distances #423 vertex clip and cull distances, with the two changes above
• feat(agc): accept UINT16_ABGR color exports into unsigned integer targets #1073 UINT16_ABGR color exports (merged)

Memory and performance:

• fix(agc): copy an imported range once a merge extends it past its import #1489 copy an imported range once a merge extends it past its import (merged; replaces fix(agc): reuse a host import only while it covers its region #2101)
• fix(agc): cache Linux describePages results #1810 cache Linux describePages results, with fix(guest-arena): place guest address space in the PS5 window on Linux #861 (guest arena in the PS5 window on Linux)
• perf(agc): store write-back changes sixteen bytes at a time #2481 store write-back changes sixteen bytes at a time

Relinker and libraries:

• fix(relinker): end the guest frame-pointer chain at the entry stub #2339 end the guest frame-pointer chain at the entry stub, so Unreal's call-stack capture stops there instead of crashing
• feat(libSceNpEntitlementAccess): report signed out for entitlement info lists #2077 NpEntitlementAccess entitlement info lists report signed out (merged)

Tested on a contributor's machine with Claude (AI assistant).

[johnmaia]

Update: the game now reaches the first level and gameplay. Continue loads Dark Beauty Castle, the opening scene in the Mad Doctor's lab plays with sound, and then Mickey is playable. One run stayed in the level for 44 minutes without an error.

Clips with sound: opening scene, gameplay.

What changed since the last update (details in the post above):

• Rendering: feat(agc): vertex clip and cull distances #423 (vertex clip and cull distances) plus two changes described on that PR, and feat(agc): accept UINT16_ABGR color exports into unsigned integer targets #1073 (UINT16_ABGR color exports, merged).
• Title crash: UE4 title (PPSA15657): LowLevelFatalError "RemoveFromOuterMap NumRemoved = 0" when the title screen opens the menu #2365, the RemoveFromOuterMap fatal error when Cross opens the menu, is avoided with APS5_NO_HOST_IMPORT=1: 0 of 5 runs, against 5 of 8 with import on.
• Performance: in the level the GPU is about 1.4% busy, and the GPU submission thread is the limit, writing GPU results back into mirrored guest buffers. perf(agc): store write-back changes sixteen bytes at a time #2481 makes that write-back about 4x cheaper: 4.6 to 6.9 fps over the same stretch of the level.
• New open issues: Sampling a D16 depth surface's memory as a half-width 32-bit texture (PPSA15657): "sampling the depth plane ... is not implemented" #2366 (a depth surface's memory sampled as a 32-bit texture), and the image filling only the middle of the window.