[NicoPV94]

What and why

Symptom: on main (b2d46b8, M4 Max), everything Ocarina of Time (0400486024DEC000 v131072) draws in real time has dark, dotted vertical bands across the whole picture: a few pixels wide, about 64 output pixels apart, on Link, the grass and the scenery, in the opening's forest scene, Link's house and Kokiri Forest. The movies, the title, the menus' 2D and the HUD are clean. DM_UPSCALER=metalfx does not remove them.

Where: a colour dump of one frame (DM_CTL texdump) shows the scene's own targets at 1536x864 clean and only the upscaled 2048x1152 picture banded. That picture is written by agl_ndsr_vr.pck's compute network (f8965bffbdf51d78, then its HMMA shaders), which MetalFX does not replace (it takes agl_ndsr.pck's place only).

Cause: commit bd9f422, "HMMA chains keep their accumulators and A in the SIMD-group's matrices". With that commit reverted on top of main (the translator and dmsh_build.py's hmma_msl together; reverting only the translator leaves HMMA rewrite failed on 62 shaders), the same save, moment and screenshot draw without bands. Which of its flags gets this network wrong is not known yet. The commit's own reference entry says the NDSR upscaler loads none of its A where its LDSM read it, so the suspects are the accumulator slots, A reuse and C = 0.

Ruled out on the way:

• LDSM ordered by subgroup barriers (LDSM is fenced by subgroup barriers: Ocarina of Time's movie filter read other lanes' shared-memory stores before they landed #1376's change, ported to main's ALdsm): no difference.
• The network's image atomics (d17eefaff4f0bc23's dispatches dropped): no difference.
• Its first shader (f8965bffbdf51d78): the same MSL with and without the change.

Change: bd9f422 reverted: TensorPass.cs, InstEmitAmpereTensor.cs, HelperFunctionManager.cs, Instruction.cs, TransformPasses.cs, dmsh_build.py's hmma_msl and test_dmsh_hmma.py; the translator rebuilt, tool_lock.py --update-lock, dmsh_chain.py --header (chain 1681473c4e47). The commit's reference entry is replaced by one for the bands (Dotted bands in the variable-resolution NDSR upscaler's picture), and the game note says what is lost: the movie networks go back from 11.4 to 14.8 ms of compute a frame on an M6, as measured in bd9f422.

This is a revert, not the fix of the flag itself: it would be better to land bd9f422 again without whatever breaks agl_ndsr_vr, and the reference entry leaves that open. Until then, every Mac player of this game on main sees the bands.

Checked

• Tests: test_docs.py, test_nsp_chain.py, test_dmsh_hmma.py (the version from before bd9f422) and every dm/gfx/tests/test_dmsh_*.py pass, except test_dmsh_pairs.py: 2 errors (NameError: name 'output' is not defined in dmsh_build.py vs_pad), the same on main b2d46b8 without this change.
• Game: Ocarina of Time on an M4 Max (macOS 27.0.1), from a save in Kokiri Forest, screenshots at the same moment:
  • main b2d46b8: banded;
  • main + LDSM is fenced by subgroup barriers: Ocarina of Time's movie filter read other lanes' shared-memory stores before they landed #1376's barriers: banded;
  • main with d17eefaff4f0bc23 dropped: banded;
  • main with bd9f422 reverted (this PR): clean, 60 fps, no red lines beyond those of a partial library.
• A playtest of a Mac app built from this branch (all 42,848 shaders, --platform mac-app; plus Ocarina of Time's first start no longer shows a black picture with the title music for minutes: pso_manifest.tsv seeds the prepare step #1493's prewarm list on a local branch, not in this PR) on a new file through the opening and Kokiri Forest, with a controller: no bands; nearly smooth, with some dips. Seen but not from this change: some floor textures missing and some sparks on the screen (not looked into), and on the first start a black picture with the music for minutes while pipelines compiled (Ocarina of Time (0400486024DEC000 v131072, Mac M3 Pro): black picture with sound from the first frame; Game's upscaler = MetalFX (DM_UPSCALER=metalfx) brings the picture back #1477, which Ocarina of Time's first start no longer shows a black picture with the title music for minutes: pso_manifest.tsv seeds the prepare step #1493 fixes).
• Not run: scripts/regress.py check. regress.py impact origin/main..HEAD names no game that uses what the change touches; Breath of the Wild, Pokémon Scarlet and Tears of the Kingdom only through the broad rule (the chain version: every NVN game), and 146 profiles have no usage manifest. HMMA is an Ampere (Switch 2) instruction, so the games it changes are the Switch 2 ones with agl's upscaler or NDVE movies, among them Donkey Kong Bananza, whose upscaler bd9f422 measured unchanged. Ocarina of Time is the only one on this Mac.

🤖 Generated with Claude Code

[lunks]

Thanks for the bisect. #1544 fixes the cause: a plain HMMA call overwrote the chain waiting in slot 0, which is agl_ndsr_vr's, so the revert is not needed. In game, with the fix and #1456 kept: the dotted columns are gone (esdras-kyo, M5 Max; aksmith-git, M4 Pro, library rebuilt) and so are the horizontal rows (PucaVaz, M4 Air); see #1544. Could you try it? Chain 5c852d7989ac.

[NicoPV94]

Closing: 13b432b on main fixes the cause (a plain HMMA call overwrote the chain waiting in slot 0, chain d866f9700d78), so the revert isn't needed and the speed-up stays. Thanks @lunks and everyone who tested it.