MilkDrop: preset thumbnails in a grid (#575), with engine buffer and light sampler fixes

[CaYatur]

#575: preset thumbnails. Two engine fixes found along the way go beyond it and are listed separately below.

Thumbnails

The MilkDrop panel's list gets a Layout choice, List or Grid. In the grid, every preset shows a thumbnail. Each thumbnail is drawn once in a hidden window and kept in userData/milkdrop-thumbs/.

• Recipe. Measured on a seeded 200-preset sample of the corpus.
  • Drawing starts from black buffers and runs 60 frames of the demo sound at 1/30 s. It is drawn at 640x360 and scaled to a 256x144 WebP of about 3.6 KB.
  • I tried a fixed seed image instead: a builtin preset for 45 frames, then a hard cut. It cut black thumbnails only from 20 to 18, made each job 45% longer, and painted its own colours into presets that draw nothing like it.
  • More frames did not help: 90 frames gave 20 black and 150 gave 22.
  • Drawing at 640x360 instead of 320x180 barely changed the time (250 ms vs 259 ms median).
• Same bytes every time.
  • Each thumbnail uses a new engine instance, whose context is released on dispose.
  • Math.random and the preset seed are derived from the key.
  • Requested textures are waited for. A thumbnail whose texture does not arrive, or whose frames were not all drawn on a live context, is not written.
  • A crashed drawing process is replaced.
  • Checked in the real page:
    • 7 of 7 presets gave the same bytes after two different busy predecessors: the five busiest of 40, a textured one, and one using sampler_randNN.
    • Under load, one preset drawn 48 times gave the same bytes every time.
• Key. A hash of the recipe (the bytes of every file the page loads, plus the sizes and frame count), the preset source, and a signature of the textures it uses.
  • An engine change retires every thumbnail by itself.
  • An edited preset or a changed texture gets a new thumbnail.
  • Files that no current preset produces are deleted once per session.
• Panel.
  • The panel requests the visible cells. Cached thumbnails come back at once; the rest arrive as they are drawn, newest request first.
  • A waiting cell is striped and a failed one has a dashed frame, so neither looks like a thumbnail that is really black.
  • Images come through an sv-thumb: protocol that accepts only a key.
  • The hidden window closes when idle and when the panel closes.
• Measured.
  • First thumbnail after 1.0 s; the 20 visible cells were full in 10–17 s.
  • With a visualizer window drawing MilkDrop at 75 Hz: frame-time p99 was 13.5 ms idle vs 13.6 ms while drawing thumbnails. The longest frame was 17.2 ms vs 26.7 ms, and no frame took over 33 ms.
  • Asking for 40 cached thumbnails took 2.9–4.2 ms.
• Smoke. The smoke test draws a builtin's thumbnail through the whole chain into a temporary folder: the recipe files read from the package, the hidden window, the protocol and the panel's CSP.

Found on the way

• A picture from another window in a new buffer (engine).
  • Render targets were allocated empty and cleared with gl.clear.
  • With another MilkDrop context drawing on the same GPU and the CPU busy, a freshly allocated feedback buffer could still hold that context's picture, and feedback grew it. A preset whose thumbnail is black came out as another window's starburst in 13 of 24 draws.
  • Neither condition alone caused it: 18 of 18 were clean with each one.
  • Targets are now allocated from a shared zero buffer that is never written. The same conditions gave 0 of 48.
  • Drawing alone is unchanged: 200 presets gave identical results before and after.
  • Cost: only when buffers are allocated. A full rebuild at 1080p takes 8.6 ms instead of 3.3 ms.
  • Live windows (a new window, a resize, a restored context) and the video export were exposed too.
• Light colours from MilkDrop (MilkDrop: drive the lights from the colours MilkDrop draws #589).
  • The sampler read its canvas with getImageData about 30 times a second, and Chromium warned about it. The smoke test caught the warning because lights in the current settings take MilkDrop's colours.
  • The downscale now stays on the GPU; only the 4 KB result is read back, from a canvas made for reading.
• Automation never drives physical lights.
  • The smoke test and the screenshot tool run on the user's own profile. With Dynamic Lighting on, they drove the lights at start-up. The smoke test also turns OpenRGB on in the panel.
  • As with the camera, in these runs:
    • every Dynamic Lighting setting is sent switched off;
    • the device scan is skipped;
    • OpenRGB and Art-Net are never started.
  • --diag still drives the lights.
• Smaller fixes.
  • sampler_randNN picked from an unsorted texture list; the list is now sorted.
  • The importer (MilkDrop: import from ZIP archives, and find libraries already on the machine #574) missed textures requested with filter/wrap prefixes (sampler_fw_worms).
  • Builtin preset sources are now readable in Node.

Not done

• Thumbnails in the web remote and in the list view.
• Drawing every thumbnail ahead of time.
• Thumbnails show the first two seconds and use the engine defaults.

Checks

• npm test: 2108/2108.
• New tests pass with process.platform set to Linux.
• npm run smoke: PASS. User-data checksums were unchanged.
• 57 of 58 mutations are caught. The survivor is an equivalent check.

Refs #575, #589, #560