Hollow Knight: Silksong (PPSA12544) on Windows with NVIDIA

[lubaxe]

Bring-up notes for Hollow Knight: Silksong (PPSA12544) on Windows with an NVIDIA GPU. #726 covers Windows on AMD (RX 9070 XT). On NVIDIA with current main, the title needs a few open PRs and one opt-in switch to reach playable gameplay. With them it runs at a steady 60 fps.

Investigated with Claude Code (AI); the runs are on my machine and I checked the result in game.

Status (updated 2026-10-09, main d70b899)

• ✅ Playable at 60 fps. It boots, plays the intro, and gets through the title menu, profile select and the opening sequence into gameplay (tut_01, then tut_02). Backgrounds, HUD, prompts and menu text render correctly.
• 📈 Frame rate in gameplay: 600 presents per 10 s.
• ⚠️ It needs one switch, APS5_STORE_AT_FLIP=1 (below). Without it, the backgrounds are garbled or the title crashes while loading tut_01.

Setup: Windows 11 Pro, Ryzen 7 9800X3D, RTX 5090 (driver 617.42), WinLibs GCC 15.2.0 r7, Release, keyboard input.

Patches on top of main d70b899

	For
#846	--module-dir: the Unity plugins in Media/Plugins (PSN, SaveData, CommonDialog, ...) are relinked and started with their start arguments
#425	depth surfaces sampled as textures
#1836	a type 4 shader header without a program address: missing static shader ABI register 201
#1599	boxed menu glyphs (#865)
#1968	the primitive restart index counted as a vertex: vertex fetch exceeds descriptor record count
#1969	a null V# vertex attribute: unsupported vertex format 0
#1995	APS5_STORE_AT_FLIP=1: garbled backgrounds and the tut_01 crash (#1970)

What the switch fixes

Garbled backgrounds and the tut_01 crash (APS5_STORE_AT_FLIP=1, #1970). Unity frees render targets and its allocator reuses the memory while AnyPS5 still holds the GPU results. A write trap caught +0x1a578fe and +0x1a57931 writing allocator block headers into a pending 1920x1080 target. The late store then either drops the results (noise) or overwrites the headers (the crash at +0x1a5771d). Storing pending results at each flip fixes both. APS5_EAGER_WRITEBACK=1 also renders correctly, but costs most of the frame time.

Frame rate. Gameplay used to run at about 5 fps on Windows, because since 838d02a every imported render target was stored and uploaded again at each use (about 2.1 GiB per frame). #2096 on main fixes that, so #1994 (an opt-in switch for the same problem) is closed.

Still open

• agc-driver: write-back drops GPU results of full 64 KiB blocks of 1920x1080 render targets (PPSA12544 backgrounds garbled) #1970: store pending results at the title's GPU waits instead of at the flip, so the switch is no longer needed.
• Seen once, not reproduced in two later runs: after profile select and New Game, the title stopped in the opening sequence with sampling the depth plane of depth surface 0x265f10000 (1920x1080, vk format 130) as a 1920x1080 texture of guest format 56 (vk 37) ... is not implemented. It looks like a depth surface's memory reused as a color target.
• agc_driver_pixel_lds fails on unmodified main on this GPU (VK_ERROR_DEVICE_LOST), and agc_driver_recorder_tests stops in storageRefreshTests on Windows. Neither affects the title.

If you have another NVIDIA card, please post how far it gets with these patches.