Oskiewar raster benchmark #

Replay the same recorded triangle stream through Canvas2D and WebGL 2. This is an isolated rasterization experiment, not a replacement game shell.

node xbox/live/bench/render.mjs --headed --out /tmp/oskiewar-render-bench
node xbox/live/bench/render.mjs --headed --debug --out /tmp/oskiewar-render-bench-debug

Use --frames 30..600 to set measured callbacks per case (default 120). --chrome /path/to/browser selects a browser; otherwise the runner searches common Chrome locations and then uses Playwright's installed Chromium. Without --headed, it launches headless Chrome. The runner closes its browser and loopback server when finished. It neither contacts production nor deploys.

--capture-only writes the trace and its metadata without opening a browser. Two captures with the same source and debug flag should produce the same hash.

Comparison #

The fixture warms up the current game, then captures 60 frames from three seconds of fixed-step skateboard play. Both rasterizers consume those exact frames in the same order, at 1920×1080 and 1280×720 backing resolution. Each resolution runs the cases forward and backward to expose order effects.

• canvas: same-color triangle paths with the browser shell's winding rule.
• webgl-stream: pack the recorded projected vertices into one reusable interleaved buffer each callback; upload and draw through the existing WebGLOskiewarScene3D renderer. This bypasses the allocation-heavy per-triangle scene adapter, so its overhead is not mistaken for a GPU limitation.
• webgl-prepacked: copy an already packed trace frame, upload and draw. This is a lower bound on CPU packing cost, not retained world meshes: vertices still upload each frame, and a live game cannot precompute arbitrary frames.

Depth testing is disabled to preserve Canvas painter order. Enabling depth, retaining world-space terrain, or moving projection onto the GPU are separate experiments. The present experiment changes neither camera nor clipping.

Evidence and limits #

report.json includes source/trace hashes, triangle counts, omitted command counts, CPU submission p50/p95/p99, callback intervals, empty callback baselines, WebGL GPU elapsed queries when available, renderer identity, and image differences. The output also contains trace.json, trace-meta.json, and matching-frame PNGs.

Boxes, host lines and text are counted but omitted. Triangle-based world geometry, fighters, shadows, and debug capsules remain. Debug results therefore do not measure the complete debug meter or HUD. Simulation and command-generation CPU timings are reported separately from browser replay and come from Node.

CPU submission excludes later GPU/compositor work. GPU elapsed queries are invalidated when the driver reports a disjoint timer. Canvas has no equivalent GPU timer here; null means unavailable, not zero. The separate completion probes use getImageData or gl.finish: they stall and have different overheads, so they are diagnostic values, not comparable end-to-end presentation measurements.

Callback gaps over 25 ms are reported as counts, not labeled dropped display frames: a 30 Hz scheduling policy produces those gaps even without drawing. Screenshots are inspected at the common CSS display size; antialiasing may differ. Image differences are evidence of similarity, not a proof of full visual parity.

A production decision still needs a complete shell trial preserving text, translucency and UI/world ordering, followed by played-session measurements on browser, Xbox and Apple targets. See RESULTS.md for the first run.