[itsyunusacar]

What and why

Fixes #1477. A first start of Ocarina of Time (0400486024DEC000 v131072) on a Mac with no pso_user.tsv yet and a cold Metal shader cache shows a black picture with the title music for minutes: the first title frame builds 11 pipelines and loads 12 shader libraries inside one submission, 252.6 s on an M3 Pro (macOS 27.2) with MTLCompilerService at 100 %, while the window keeps presenting the last frame. The profile's pso_manifest.tsv was header-only, so the prepare step had nothing to build before the first frame.

This seeds dm/games/0400486024DEC000/pso_manifest.tsv with the list this Mac recorded on v131072 (chain adec7d81d970) with no lists present, so everything was built on the spot: 9,620 lines (8,376 shaders, 1,112 render pipelines, 132 compute pipelines) over boot, the title and its movie, a new game and the opening, Link's house, Kokiri Forest and a save load. The game note has the finding.

Checked

• Tests: python3 -m unittest dm/tools/tests/test_nsp_incremental.py (the manifest's derivation into the build).
• Games run: Ocarina of Time 0400486024DEC000 v131072, Mac app on 671c68e / adec7d81d970. Cold start (lists and Metal cache set aside): frame 108 whole frame 252618.4 ms … new pipelines 11 + shader libraries 12, 252581.4 ms in all, then the title at 287 s. Same start with the list in place: the title in about 20 s, screenshots at frames 300/800/1500 full range, GPU 25 ms a frame in a window. Not checked: a rebuilt app with this manifest bundled (the list was tested through pso_user.tsv, which the prepare step reads the same way), and the first-start time on an M4 Max (the other report in 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).
• Devices and OS: MacBook Pro M3 Pro (18-core GPU, applegpu_g15s), 18 GB, macOS 27.2, Xcode 27.2.
• Not a chain, host or tool change: no regress.py run.

Docs

• The finding is in its game note (dm/games/0400486024DEC000/README.md); no docs/reference entry, so docs_symptoms.py was not needed
• No chain tool changed
• No game files, keys or data extracted from a game in the change or the attachments (the list holds shader content keys only)

🤖 Generated with Claude Code

[jonpage0]

Same comparison as on #1488, since the two lists conflict on one file:

Both lists compared, since they conflict on the same file: #1488 (9,784 rows: S 8,355, R 1,227, C 144, M 58) and #1493 (9,620 rows: S 8,376, R 1,112, C 132) share 9,558 rows; #1493 adds 21 shaders (its S set is a superset of #1488's) and 40 render pipelines #1488 lacks, #1488 adds 155 render pipelines, 12 compute pipelines and the 58 M rows. A third list, this Mac's own pso_user.tsv after a few hours of Kokiri Forest routes (9,102 rows), lies entirely within #1488. Their union is 9,846 rows (S 8,376, R 1,267, C 145, M 58); we build with that.

Cost of the prepare step with the union list, Ocarina of Time v131072 Mac app on main bd9f422 + #1368 + #1307, M3 Max (36 GB, macOS 27.0.1), the device archive empty each time:

• first start after a fresh pack: pipeline prepare: 9846 lines (8376 shaders, 1267 render pipelines, 145 compute pipelines) in 84.56 s on 14 workers before the first frame, progress screen shown; bases 1017 compiled, 44 from the archive, 205 reused; 1266 specializations;
• second start, another empty device archive (Metal's own cache warm): the same line in 36.67 s, first frame at 39 s, then the title movie at 60 fps; after a repack of the same shaders, 65.29 s; with a device archive from an earlier run in place, 0.66 s (progress screen not shown).
  With Ocarina of Time: prewarm native mesh pipelines before attacks and enemies appear #1368 the 58 M rows are prepared too (pipeline prewarm: 58 native task / mesh pipelines). The bundled offline archive (4.3 GB, Compute pipelines in the pipeline archives: compiled offline for every compute shader of the map, and captured into the device's archive when the bundle lacks one #1307's compute pipelines in it) held the bases of every probed pair but only 44 of the 1,266 bases the R rows need, so the first start is still a compile; a prepare that read R-row bases from the offline archive would make it seconds.

[PucaVaz]

Compared both lists on a true first start, M4 MacBook Air (24 GB, macOS 27.2, 10 cores), Mac app from main b2d46b8 (+ #1544, #1543, #1555), --seen-shaders build (about 6,930 of each list's ~8,370 shaders are in its library). Each list was given with DM_PSO_MANIFEST=<file> (main's bundled list is header-only). Each run started with no pso_user.tsv, an empty pipeline archive, and the app's Metal caches (com.apple.metal, metalfe, gpuarchiver under the user cache folder) set aside. One scripted run per list from launch through the title, a save load into Link's house, out of the door and Saria's greeting.

	#1493	#1488
prepare step	9,620 lines in 134.7 s (887 bases compiled, 29 from the bundled archive)	9,726 lines in 150.4 s (1,001 compiled, 44 from the archive; the 58 M rows skipped)
host start → first frame	135.9 s	151.7 s
first frame → save loaded: frames over 0.25 s	0	0
stalls over 1 s	3.7 s + 2.9 s at Saria's greeting	3.4 s + 2.7 s at Saria's greeting
built on the spot	7 render, 5 compute pipelines	5 compute pipelines

Both lists miss the same five compute pipelines: C rows for 2de44613e731e867, e28d865e2cc82d36, 983d63a70f1d7fcc (Enemy_Scrub_Deku), 50bb5cb052951045 (Enemy_Scrub_Business) and fef9f883f289b61b (Enemy_Dekubaba). Both have them as S rows (stage 5), so the shaders load early, but the pipelines are built when Saria's greeting starts: about 6 s of frozen picture on a first start with either list. Adding those five C rows to whichever list lands should remove it.

The 7 render pipelines #1488 has and #1493 lacks on this route never made a frame over 100 ms (hybrid pipelines cover them), and #1488's longer list costs about 16 s more before the first frame here. On this route, then, #1493's list (or the union, plus the five C rows) gives the shorter wait. I haven't covered later areas, where #1488's iPad rows may matter, or a no-list cold start on this Mac.

[seedds]

Compute pipelines are now in the offline archive (30fb213), so the title's kernels behind #1477's black screen come from the bundle on a first start. This list (the first-scenes route, as docs/development.md asks) plus the five C rows PucaVaz found at Saria's greeting is the planned seed list. It's waiting on a true first-start measurement on main now that compute is archived, and the result decides whether it lands.

[seedds]

Landed in 540f797 with credit: this list unchanged, plus the five C rows PucaVaz found at Saria's greeting (on #1488), 9,625 rows. Thank you for the recording and for tracking #1477 down to the title's pipeline builds. The first start on main (compute pipelines are in the offline archive since 30fb213) is not measured yet; the game note lists that as open.