Entry 175

"Skeith's stall and the drain's dither: stream files read ahead, vertex alpha as the GS's integer"

2026-09-26render

Filescrates/piney-gs/src/lib.rscrates/piney-gs/src/assets.rscrates/piney-stream/tests/drain_render.rscrates/piney-stream/examples/stream_time.rscrates/piney-game/src/session.rsdocs/engine/render.mddocs/engine/stream.mdBUGS.md

Two reports from playing event 4's streams in the window. The game lagged badly as Skeith's animation began. The drain on Orca looked like its textures z-fought.

The stall. event_4_stream_times (ignored) plays event 4's streams 4-6 through the session as the window does. Each step, its sound (the events handed to a headless Audio, a frame of the driver, a 30th of a second rendered) and its GPU draw are timed. Stream steps take about 1 ms and draws about 5 ms. The exceptions:

• Each stream's load at its first frame: 78 ms for stream 4, 199 for 5, 48 for 6. Stream 5's str0120 is 31 MB, 83 MB inflated. Inflating it takes 107 ms and reading its scene 68 ms. It falls at the cut between two streams. flate2's zlib-rs backend saved only 8%, so it was not taken.
• One draw of 138-151 ms at stream 5's frame 1834. It made no new pipeline and only one texture. The draw commands of frames 1833 and 1834 (stream_time --dump) differ by one new model, the book's bubble, and it is the first draw from str0120 rather than str0120e. The renderer reads a file the first time a draw names it, which here means inflating and parsing all 83 MB mid-scene.

piney_gs::Assets now starts a thread when it is given a new overlay (a stream's archive). The thread reads every member in turn and sends each file as it is read. A lookup takes whatever has arrived, and reads the file itself only if the thread has not got there yet. Files are Arc now, not Rc. Stream 5's worst draw after the first frame is now 6.7 ms.

The dither. The pictures of stream 5's frames 2690-2697 showed the figure in the drain's cylinder with its hat broken into a fine dither of itself and what lay behind. It was not the texture: TEX_ckitbodh2 reads clean.

Leaving commands out of one frame's picture (stream_time --drop) narrowed it down:

• Only draw 34 matters: a gold ring, str0120e object 0x20c, on layer 3.
• The figure is on layer 4 and draws after it.
• 8x the depth precision changed nothing, so it was not the depth.

The ring is fading in at transparency 0.02. Its vertex alpha, trunc(0x80 * t), is 2, and its AREF, trunc(aref * t) (ccSetMaterialPacket, docs/engine/desktop.md), is 2 as well. So on the GS every pixel passes and writes Z. The ring draws nearly nothing and erases the figure behind it, as a band sweeping down before the figure turns to particles.

The shader interpolated that constant 2 as a float. At some pixels it landed a hair under 2, floored to 1, and failed. The shader now snaps the interpolated vertex alpha to the GS's integer (floor(a + 1/32)) before the modulate and the test. The erasure is clean now. The cylinder's dotted floor, the same threshold, is smooth.

tests/drain_render.rs (piney-stream) renders frame 2694 and measures the hat's place: 76 with the old shader, 24 with the new, and it asserts under 40.

stream_time gained --shots DIR FROM TO EVERY, --dump FRAME (the model draws with their edits), --drop I,J (commands left out of the dumped frame), the textures and pipelines each draw made (Gs::counts), and the load split into the disc, the stream and its effects.