Repository navigation
Tags
Tags
fix(renderer): reset() zeroes a slot's geometry, so a grown SpriteLay… …er cannot draw ghosts (#86) resize(n) documents "new slots are zeroed, white, visible", but reset() left x/y/w/h alone. Slots past the count keep the geometry of whatever sprite last used them, so after clear(), removeAt() or a shrinking resize() the next grow redraws removed sprites at their old position and size. Zero the geometry in reset() itself (add() writes it right after, so the only other caller is unaffected) rather than at the resize() loop. Regression tests: clear() then resize(1), and removeAt()/shrink() then grow. Both fail on the previous code with the stale x/y/w/h. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
perf(renderer): radix sort for the draw-order full sorts (25k keys 4.… …6 -> 0.6 ms) (#93) * perf(renderer): radix sort for the SpriteLayer and TextLayer full draw-order sorts (25k keys 4.6 -> 0.6 ms) The full sort that SpriteLayer.drawOrder falls back to, and the TextLayer rebuild sort, compared with a JavaScript callback. They now run an 8-pass LSD radix sort over the sortable 64-bit pattern of the key (a second key first when sortKey2 is set, so equal keys keep the slot tie-break). NaN keys keep the comparator. TextLayer.sortKey is Float64Array now, as SpriteLayer's is, so both use the same helper. Steady-state insertion sorts are unchanged. Measured (Node, bundled radixSort.ts vs the comparator each layer used, same process, 40 runs each, random keys, median / p95): - SpriteLayer 1 key, 25,000: comparator 4.59 / 5.25 ms, radix 0.63 / 0.79 ms - SpriteLayer 2 keys, 25,000: comparator 4.62 / 7.27 ms, radix 1.39 / 2.52 ms - TextLayer 1 key, 25,000: comparator 4.67 / 7.56 ms, radix 0.62 / 1.32 ms - 3,000 keys: 0.40 -> 0.08 ms (1 key), 0.37 -> 0.17 ms (2 keys) The radix timings include a scratch allocation per call, so the real gain is slightly larger. Tests: radixSort.test.ts checks the order against the comparator for sizes 0..4000, with and without a second key, -0/+0, infinities, subnormals and NaN fallback. Layer tests cover a 2,000-sprite rebuild with hidden sprites and a 1,500-label TextLayer rebuild followed by a steady-state shuffle. The renderer typecheck and the full suite (2,439 tests) pass; bench:ci is within budget. Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com> * fix(renderer): radix sort leaves order untouched when keys2 has a NaN The NaN check ran after the second key had already permuted order. Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Haiku 5.5 <noreply@anthropic.com>
fix(xip): reset virtual joystick input after teardown (#73) * fix(xip): reset virtual joystick input on teardown Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(xip): guard virtual input pointer ownership Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(xip): clear only joystick-owned input Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
perf(renderer): retained packed instance buffer for SpriteLayer (touc… …hRange, dirty-range packing and uploads) (#55) * perf(renderer): retained packed instance buffer for SpriteLayer (dirty ranges, touchRange, partial uploads) * test(gpu-smoke): retained SpriteLayer partial uploads on real GL
feat: engine.captureFrame() and useCaptureFrame() (#28) * feat: engine.captureFrame() and useCaptureFrame() Renders a frame and reads it back in the same task, so a WebGL canvas that is not preserving its drawing buffer no longer yields a black image. Result as PNG/JPEG/WebP blob, ImageBitmap or detached canvas, at a chosen size, scaled or re-rendered at that resolution. Loaded on first use (bundle unchanged). Co-Authored-By: Claude <noreply@anthropic.com> * Remove an accidentally committed node_modules symlink * Fix the merged gpu-smoke runner: main's checks plus the capture block and Chrome flag --------- Co-authored-by: Claude <noreply@anthropic.com>
perf(bundle): load tile/sprite layer renderers only when a layer exists Layer data classes register their GL renderers; RenderSystem creates them on first use. The 11-symbol import drops to 30.5 kB gzip (from 33.1); using layers adds ~8 kB. Size budgets tightened to 32.5 kB. Co-Authored-By: Claude <noreply@anthropic.com>
PreviousNext