[lunks]

What and why

Ampere's MUFU.SIN / COS take the angle in turns and use only its fraction, which is exact for any float. The translator multiplied the turns by 2π in float and called Metal's sin / cos. The product rounds the fraction away (a step of 0.4 rad at 2^20 turns), and Metal's fast sin / cos return 0 from about 2^20 turns on.

Found on Donkey Kong Bananza's SSAO (a6f04605183d9915). It rotates its kernel per pixel by an angle that grows with x·y. Past x·y ≈ 659,000 the rotation is (0, 0), every sample lands on the pixel itself, and the lower right of the picture has no ambient occlusion, bounded by an arc centred off screen. The SSAO runs at half the scene's size, which dynamic resolution moves:

SSAO target	When	Pixels without occlusion
838x471	scene 1676x943	0 (angles up to 0.23 rad off)
1152x648	docked scene at its top, 2304x1296	5,425, a wedge in the lower-right corner
1676x943	effects at full size (a local patch of the game's effects-resolution levels)	345,255, a fifth of the picture

Any Ampere shader that passes sin / cos an angle that large, such as a per-pixel rotation built from pixel coordinates, has the same fault.

Fix: AMufu lowers both as sin / cos((t − rint(t)) · 2π) (ATurnsToRadians in InstEmitAmpereAlu.cs). t − rint(t) is exact and in [−0.5, 0.5], where the fast functions are accurate. Maxwell code is unchanged; its MUFU follows RRO and takes radians.

Reference entry: MUFU.SIN and MUFU.COS of a large number of turns.

Checked

• Tests:
  • t_mufu_turns (new, GPU) compares the lowering with a double-precision reference over the SSAO's angles at 838x471, 1152x648, 1676x943 and 2560x1440. Largest error 2e-7, no collapsed rotations. The old lowering collapses 0, 5,425, 345,255 and 1,893,254.
  • test_dmsh_ampere_mem.py MufuTurns checks the lowered MSL.
  • make -C dm/gfx/tests GAME=040074A01BF12000 run passes.
  • lint-python: no new findings. lint-c on the changed files: none (LLVM 23.1.1).
• Games run: Donkey Kong Bananza (040074A01BF12000, v458752), docked on the Mac, with the effects at full size: the arc without ambient occlusion is gone. Other Switch 2 games were not checked.
• Devices and OS: M5 Pro, macOS 27.0, Xcode 27.2.
• scripts/regress.py check --games 040074A01BF12000, on main at 4447be1 (the branch was then rebased over Mario Kart World 1.8.0 (v917504) on main: plays well, one sky artifact #1186's c0 offsets, which changed only the chain header here): CHANGED.
  • 23,638 of 23,638 shaders build on both sides; 0 newly failing.
  • 8,258 with changed MSL (8,161 through variants); 0 interface changes.
  • No boot run in the report.
• regress.py impact lists Breath of the Wild, Pokémon Scarlet and Tears of the Kingdom only through the chain header. They are Maxwell games, whose code this does not change.

Docs

• The finding is in docs/reference/shaders.md and the game note, and python3 dm/tools/docs_symptoms.py was run
• A chain tool changed: dmsh_chain.py --header was run and dm/gfx/nvn/dmsh_chain.h is part of the change
• No game files, keys or data extracted from a game in the change or the attachments

[seedds]

Landed as b594834 with you as co-author (translator rebuilt here; t_mufu_turns and MufuTurns pass, MufuTurns fails before). Thank you.