LEGO® Builder's Journey (PPSA05496)

[whamre]

Tested
Title	LEGO® Builder's Journey (PPSA05496)
Version	content 03.002.583, master 01.00, SDK 0x04000000
Engine	Unity 2021.2.13f1, IL2CPP, FMOD
Host	Windows 11 (build 26300), i5-10500, RTX 2080 SUPER 8 GB (NVIDIA 616.92)
Build	MinGW GCC 15.2.0, main f6eeca4 plus the PRs below
Relink	--windows --module-dir Media/Modules --module-dir Media/Plugins. eboot, sce_module and the Unity plugins unwrapped from SELF with tools/unwrap_self.py
Updated	2026-10-19

Status

With the PRs below applied, it boots, plays the LEGO logo and reaches the first playable scene at about 7 FPS (frame p50 ≈ 145 ms). It hasn't been played beyond that scene yet.

Stops on main, in the order they're hit

#	Stop	Covered by
1	Black screen after 2 presents. IL2CPP's GC raises SIGUSR1 to stop the world, and on Windows the main thread is redirected into the suspend handler inside host code, then waits for a resume that never comes (resumeEvent; the render thread then waits on the main thread).	#7652 (same root cause as Sonic Superstars PPSA06888); details in this comment
2	depth mips (DB_DEPTH_VIEW MIP_LEVEL) is unsupported (register 0x2 = 0x04000000) right after the logo: depth is rendered into mip 1 of a 64 KiB Z_X chain, without HTILE.	#7927

Performance notes

• Storage-image cache thrash on 8 GB cards. The first scene's working set is between 2.0 and 2.5 GiB, and the automatic budget here is 1804 MiB, so every frame re-uploads about 4.7 GB and writes back about 2.3 GB. Frame p50 is about 630 ms at the old budget and about 145 ms at about 3.6 GiB. fix(agc): reserve only the sampled textures alive from the storage image budget #7862 plus the cap change proposed there. Workaround: APS5_TEXTURE_CACHE_MIB=3600.
• First run only: about 70 wave64 compute pipelines of 0.25–0.93 M SPIR-V words, compiled as two lanes per invocation on NVIDIA's 32-wide subgroups. NVIDIA needs about 22 µs per word (up to 20 s per module, measured with agc_compute_pipeline_replay), so the first visit after the logo stalls for 2–3.5 minutes until the driver cache holds them. APS5_SPANNED_WAVE=1 didn't help: the largest modules stay two-lane, and steady state is unchanged.

Imports that can still throw (from the relinker registry)

sceNpTrophy2GetRewardIcon (PSNCore.prx), sceAgcAcbSetFlip, sceAgcUpdateInterpolantMapping_0100, and the AgcDriver capture Start/Stop/Trigger calls (via libSceJobManager). Partial: sceAgcAcbSetWorkload*, sceAjmInstanceCreate, sceAjmBatchStart/sceAjmBatchJob*, sceSystemServiceLoadExec. None of these were hit before the first scene.

Stack tried

main f6eeca4 + #7652 (an equivalent local change) + #7927 + #7873 (hang-dump stack candidates, used to find stop 1) + the storage budget change above.

AI-assisted: yes. The runs and numbers are from my machine.