[humanperzeus]

What and why

Ocarina of Time (0400486024DEC000, v131072) on an M1 Max: the flower bushes in Hyrule Castle's courtyard showed their dark inner body through holes in about 40 % of frames, standing still, at random.

The foliage is GPU-driven: a chain of compute dispatches culls the clusters (83df46a0b1b379b4, 60dd0e6eac91f712, cee08c634ef1f343, 58f47abfa0a84773, e855420a39e60874) and builds the list the leaf and flower mesh shaders draw (d98076e4148db34f, eebdb9347d09b344). A later dispatch of that chain is indirect, its group counts written by an earlier one. Dispatches share the copy encoder, each after a barrier inside it (dispatch_enc), and that barrier did not order the indirect dispatch's counts after the dispatch that wrote them: in the frames that went wrong the cull kept 260, 268 or 250 of its 276 clusters, and the leaf list held 1,584 / 1,680 / 1,488 leaves instead of 1,749.

• dispatch (ind_split): an indirect dispatch does not join a copy encoder that already holds work. The encoder ends, and the dispatch goes into the next one, which begins with a queue barrier (blit_open) that waits for everything before it. Direct dispatches keep sharing the encoder. DM_NVN_IND_SPLIT=0 brings the old order back, to measure.
• Which of Metal's reads misses the barrier inside the encoder, the indirect-argument fetch or the copy of the counts into the driver segment, is not known.

How it was found, in short: the leaf list's count read once each submission completed gave one number a frame; it dropped in 39 % of frames, the same share as the holes in screenshots, and each frame independently of the one before. The cull's own list dropped with it. Every uniform and small storage input of the five cull shaders, dumped each frame for 662 frames, followed none of the drops (|r| ≤ 0.12), so the inputs were the same and the result was not: an ordering problem, not the shaders. Lab switches flipped live in one session (not part of this change) then gave: an encoder of its own for every dispatch 0 %, for the indirect ones only 0 %, for the direct ones only 41 %.

Checked

• Games run: The Legend of Zelda: Ocarina of Time 0400486024DEC000 v131072, standing still in Hyrule Castle's courtyard, on our build (base 641636e2 plus our changes, the same dispatch code as main), M1 Max, macOS 27.2:
  • the leaf list's count, 240-300 frames a setting, one session: old order 39 %, 41 %, 47 % of frames short; this change 0 of 600 (two runs of 300);
  • 24 shots each: 0 with holes, 9 with DM_NVN_IND_SPLIT=0.
• Cost: GPU time per frame in the courtyard (DM_NVN_GPUPROF=every:30, frame-rate target 60, M1 Max): 21.08 ms mean / 22.36 ms median over 43 frames with this change, 20.50 / 21.95 ms over 44 frames with DM_NVN_IND_SPLIT=0; 141 against 131 encoders a frame. One run each, so about +0.4 to +0.6 ms (2 to 3 %); how much of that is run-to-run spread is not measured.
• Tests: make -C dm/gfx/tests GAME=0400486024DEC000 run on this branch: exit 0, 75,079 checks in 158 result lines, 0 failed, the Python groups OK. Left out: t_gl_imports, which does not link on main itself (f3ec8fc8: _g_key18, referenced from hle.c, defined in loader.c, which the test does not link); built on a clean checkout of main it fails the same way. No test reproduces the race: a synthetic pair (a dispatch writes the counts, an indirect one reads them, in a small pool and through two chunk buffers of a large one) passed with the old order too (0 of 100 runs each), so it is not part of this PR; what in the game's chain lets the old order fail is not pinned down.
• python3 dm/tools/tests/test_docs.py: OK. Host builds (make -C dm/host GAME=0400486024DEC000 dmhost).
• scripts/regress.py: impact reaches every NVN game (with manifests: Breath of the Wild, Pokémon Scarlet, Tears of the Kingdom; Hades not affected). check --games was not run: only Ocarina of Time is set up here, and its library is of an older chain than main's.
• Not checked: a game other than Ocarina of Time; the iPhone; a full Ocarina build on main's chain (the change is host-only; the run above is on our build with the same dispatch code).

Docs

• docs/reference/rendering.md: "An indirect dispatch read its group counts before the dispatch that wrote them" (symptom, cause, fix, check, with the measurements and what is not known).
• The game note's "Found on this game"; docs/agents/symptoms.md regenerated.

🤖 Generated with Claude Code