[bymaxmarc]

What and why

Pixel shaders of a Switch 2 game read special register 0x84 with S2UR and take its bytes 1 and 2 with PRMT, multiply them and divide by the product. Every one of Nintendo Switch Sports Resort's (04008E001F098000) 612 reads, in 611 pixel shaders, does this. The bytes are the fragment's width and height under variable-rate shading; that meaning is inferred from this use. The register is not in Ryujinx's SReg list, and EmitSReg read it as 0, so 1 / (0 x 0) gave +Inf. The sea shaders (agl AnnaSea: 01a6b85e2f6ea3d4, b3203175c76cda20, 2271b2e0293b3c55) made NaN at the horizon and the auto-exposure spread it: the main menu's sea was black with a band of coloured noise above it. The avatar material c30f0edd6c669a61 takes the register into a shading factor, and the avatar's hair and clothes drew flat black.

The fix, in the translator (InstEmitAmpereAlu.cs AS2r): register 0x84 answers a 1 x 1 fragment (0x00010100), since the host shades every pixel. Any other special register the SReg list does not know now leaves a translator note (S2R of special register 0x.., which the translator does not know: read as 0) instead of a silent 0.

This is a chain change. dmsh_xlate was rebuilt with dm/tools/native_tools.sh dmsh_xlate from main's sources plus this change (Microsoft's .NET SDK 10.0.403, universal arm64 + x86_64, no weak imports; --source-sha fcb37fccd2f53012 on both, as main, since that digest covers only Program.cs and the csproj). The lock was updated with tool_lock.py --update-lock (dmsh_xlate 90c1264b7831364c -> 09dfad6902250b49), and dmsh_chain.py --header gives d24f05cf9d78 -> 3df0ed1c9c0a. The build is not byte-reproducible (native_tools.sh: each link gets its own UUID). If you would rather build the binary yourself, take the source and test commit and rebuild with native_tools.sh, tool_lock.py --update-lock and dmsh_chain.py --header.

Reference entry: The fragment size register (special register 0x84) reads 1 x 1, symptom index regenerated.

This PR and the uniform-tests PR (#2372) each change dm/gfx/nvn/dmsh_chain.h, so whichever lands second needs dmsh_chain.py --header run again on top of the first.

Checked

• Tests on this branch (main ccd834e + 1 commit), Apple-silicon Mac, macOS 27.2:
  • DMSH_TEST_XLATE=dm/gfx/tools/dmsh_xlate_bin/dmsh_xlate python3 dm/gfx/tests/test_dmsh_ampere_alu.py: 47 tests pass, FragmentSize among them (the game's own S2UR, UPRMT and UIMAD encodings; the product folds to 1, and the register is 65792). With main's translator, the two FragmentSize tests fail and the other 45 pass.
  • python3 dm/gfx/tools/tool_lock.py --check-tools: passes. dm/tools/tests/test_nsp_chain.py: 17 tests OK.
  • python -m unittest discover -s dm/tools/tests: 1,469 tests OK, 27 skipped, as on main.
  • python -m unittest discover -s dm/gfx/tests -p 'test_*.py' with DMSH_TEST_XLATE set: 491 tests, the 489 that main has plus the 2 new ones. The same 10 modules fail to import on this branch and on main, because they load a game profile at import and DM_GAME was not set (test_gfpak, test_kpka, test_monolith, test_mt_framework, test_needle, test_pg_shader, test_pkz, test_shader_split_progress, test_snowdrop, test_sumo). No other failure.
  • make -k -C dm/gfx/tests GAME=010015100B514000 -j8 run (M2 Max): exit 0, every check passed.
• Translator output against main's, as native_tools.sh asks: I ran dmsh_build.py on 304 of Nintendo Switch Sports Resort's shaders once in a main tree and once in this branch, each with its own translator. The 304 were the 4 named above and 300 vertex and pixel shader keys drawn at random from its 10,718. The MSL of 412 of the 486 outputs is identical. The other 74 shaders all read 0x84. That is more than the 611 above suggests. The 612 reads were counted in the code stored in the game's shader packages, so the likely reason is that many shader keys share one code body; this was not checked. Their MSL differs in one line only, the folded value of that read: float _N = float(0); becomes float(1). The reflection JSON differs only in its chain field, the metallibs only for those 74, and no new translator note appears.
• Games: Nintendo Switch Sports Resort 04008E001F098000 v65536 on the user's Mac (Apple silicon, macOS 27.2). On the app rebuilt on the fork's work branch (an earlier main with the same source change), the main menu's sea is blue, the noise band and washed-out exposure are gone, and two default avatars show their colours in the avatar editor. That app was not rebuilt from this branch: the game has no profile on main.
• regress.py: not run. Only this game has a build on this Mac. The change reaches only shaders that read special register 0x84 (Ampere, NVN2), so it most likely reaches the other Switch 2 games (Mario Kart World, Donkey Kong Bananza, PRAGMATA). Their shaders were not rebuilt here.

Docs

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

[seedds]

Thanks for this: the same fix landed from #2103 as b581efc, which reads 0x10100 for system register 0x84 in EmitSReg, the path both Maxwell's S2R and Ampere's S2R / S2UR go through, rather than in AS2r only. Your finding about the avatars' material shader (c30f0edd6c669a61) is in the game's note.