ccleste (PPSA59322)

[rooklz]

Existing thread

• I searched this category for the title ID and no thread exists for this title on any platform.

Title

ccleste (PPSA59322)

Content version tested

The homebrew catalog build (release zip PPSA59322), the Celeste Classic port with its own assets in app0/assets.

AnyPS5 commit

main 8b3e398 plus the libSceAgcDriver change for #9783 on macOS; main a6dcea9 on Linux. Last updated 2026-10-22.

System tested

macOS 26.2, Apple M5, Apple clang 21, x86-64 under Rosetta 2, MoltenVK (MBridge build); and an Ubuntu 24.04 x86-64 container under Rosetta with llvmpipe (Mesa, LLVM 20.1.2).

Furthest point

• Does not convert
• Stops before the first frame
• Logos or loading screen
• Menu
• In game

What is shown

Nothing is presented yet. The title starts, opens the pad, loads its sounds and music (ui kit: 0 songs, 67 sounds (0 rejected), output playing, celeste: 23 effects, 5 songs) and stops at its first draw.

Current blocker

On macOS with MoltenVK, plain main stops at the first draw because the pixel shader exports raw 32-bit values into a COLOR_8 UINT target and Metal refuses the pipeline (output of type float4 is not compatible with a MTLPixelFormatR8Uint color attachment): #9783, pull request linked there. With that change the next stop on this MoltenVK is

AGC driver: a pixel shader on the mesh path reads barycentrics, and this device has neither VK_KHR_fragment_shader_barycentric nor mesh shaders

which is a capability of the MoltenVK build (the extension is listed, its feature reports false), noted on #1.

On Linux with llvmpipe the stop is llvmpipe's mesh output limit: ShaderRecompiler::Recompile: barycentric emulation on the mesh path needs 36864 bytes of mesh output memory, the device allows 32768. A real GPU is the next step.

The macOS conversion needs main's relinker; the ci/macos-series-2 tree's refuses the eboot with Guest TLS access without a non-empty PT_TLS.

How it was run

macOS: relinker --macos --to-rosetta --registry src/eboot.elf app (main's relinker) with the package as app0/, then ./app with VK_DRIVER_FILES pointing at MoltenVK's ICD and SDL_AUDIODRIVER=dummy. Linux: --to-rosetta, run through /lib64/ld-linux-x86-64.so.2 ./app.elf under xvfb-run.

[ryanc2112]

Steam Deck (Linux, RADV): in game, 59.94 FPS.

Furthest point

• In game

Converts, reaches the first level and plays with the Deck's controls; a steady 59.94 FPS (3510 frames in 59.4 s, median 59.94), music and sound effects play, and it exits cleanly. No local changes. On this device the mesh-path barycentric stop seen on MoltenVK and the llvmpipe mesh output limit do not apply: the Deck has fragmentShaderBarycentric and mesh shaders.

System tested

Steam Deck LCD (Jupiter), AMD Custom GPU 0405 (RADV VANGOGH), SteamOS 3.8.28 (build 20260922.1, kernel 6.18.50-valve2), Mesa 26.2.5 in an Arch distrobox, Desktop Mode (KDE Wayland).

AnyPS5 commit

main 4777caca3, GCC 16.2.1 Release.

How it was run

Catalog release PPSA59322.zip (darkxex/PS5_Celeste 1.00, SHA-256 checked). tools/unwrap_self.py on eboot.bin and sce_module/libc.prx (both SELF), then relinker eboot.elf app.elf (1 converted module, 352 external prx references), the package without sce_module/ as app0/, the converted libc.prx.guest.prx in app0/sce_module/, and build/core/libs/libs/*.prx in libs/. Started with ANYPS5_GPU=VANGOGH ./app.elf. /dev/udmabuf is accessible to the deck user on SteamOS through its ACL, so no group change was needed.

The only non-fatal log lines are three optional files the title probes and does not ship (/app0/hui/theme.txt and two assets/hui/audio/sfx/*/index.txt).

Screenshot (F12/Enter, 3840x2160): in the reply below.

AI-assisted: yes. I ran the title on my Deck.

[ryanc2112]

[valenciana-nc]

Windows / NVIDIA: first level reached on current main; #9789 removes the validation warning.

Tested release 1.00 / PPSA59322 01.000.000, published ZIP SHA-256 verified (96861b81c1fbaac818080d1bf958b575e6b44d2c258684e1931f2940cee2fb43). Built actual libs and relinker on clean main c383e3772be19e6f9139f1a620d37569981614ac, converted with --windows --to-intel --registry and staged the assets plus converted GPL platform libc.prx. No absent imports; five throwing import entries remain but were not runtime stops in these runs.

Windows 11 26300.9711, Core i7-13650HX, RTX 4060 Laptop GPU / NVIDIA 610.62, WinLibs GCC 15.2.0 r7 Release. AC power, internal display 1920×1200 / 165 Hz. Required Vulkan and Khronos validation enabled, application shader cache disabled, driver cache uncontrolled.

Two comparable runs of main and two with #9789 head f9b349a056 merged onto the same main (A/B/B/A). The same executable, guest module, assets and other libraries were used; only libSceAgcDriver.prx differs. All four show the first level in display readbacks, respond to scripted movement/jump input, and exit naturally with code 0 through an Options hold, about 25.5–25.8 s after launch, before the 40 s cap. The logs report roughly 1,200 presents per 10 s; this is not a performance comparison.

Main renders here but reports six Undefined-Value-ShaderFragmentOutputMismatch-DynamicRendering warnings per run (FLOAT fragment output into R8_UINT). With #9789 there are no validation warnings/errors or VUID messages. Neither build has a fatal/driver stop in these runs. This confirms Windows/NVIDIA operation, not reproduction of the MoltenVK pipeline failure or a newly playable-title claim.

Exact input schedule, all environment settings, build hashes, focused GPU test results, and the separately disclosed preliminary capped runs: #9789 (comment)

SDL_AUDIODRIVER=dummy: audible audio, physical controller, motion/rumble, longer gameplay, console and other OSes not tested by me. AI-assisted execution, log/frame inspection and report.