Anti-Aliasing Cost at Sub-Pixel Scale

Date: 2026-04-07 Status: Investigation findings

Summary

At fit-zoom on large documents (0.02x on 135K nodes), anti-aliased sub-pixel geometry is the dominant GPU cost. Disabling AA on sub-pixel nodes is a viable path to 2x+ frame time reduction during settle frames.

Discovery

Benchmark setup

Isolated GPU benchmark (skia_bench_subpixel): pre-recorded SkPictures of simple rects drawn at 1.0x vs 0.02x zoom, with AA on vs off. GPU synced via flush_submit_and_sync_cpu() for accurate timing.

Hardware: Apple M2 Pro, Metal 4.1, 1000x1000 viewport.

Results

Nodes	full (AA on)	0.02x (AA on)	0.02x (AA off)	skip (0 draws)
1,000	621 µs	809 µs	442 µs	286 µs
5,000	1,923 µs	3,426 µs	1,535 µs	271 µs
10,000	2,986 µs	6,221 µs	2,180 µs	458 µs
40,000	9,628 µs	21,878 µs	6,804 µs	324 µs

Key findings

1. Sub-pixel with AA is 2.3x MORE expensive than full-size with AA. At 0.02x zoom, 40K rects: 21,878 µs vs 9,628 µs. Counter-intuitive — smaller geometry costs more.

2. AA is the dominant cost at sub-pixel scale. AA on vs AA off at 0.02x: 21,878 vs 6,804 µs = 3.2x overhead. Skia's AA rasterizer computes edge coverage for each sub-pixel edge, and this work is proportionally more expensive when the geometry is smaller than a pixel.

3. Without AA, sub-pixel draws are near-free. Per-node cost at 0.02x: AA off = 0.16 µs/node vs AA on = 0.54 µs/node. The AA-off cost approaches the skip baseline (0 draws).

4. Text is not the bottleneck. A/B test skipping all text nodes (22% of layers) on the 135K fixture showed 0% frame time difference. At 0.02x zoom, text and shapes have identical per-node cost — both dominated by AA overhead.

Implications for optimization

Adaptive AA by screen size

When a node's screen-space area falls below a threshold (e.g. 4 px²), disable AA for that node. The visual difference is invisible (the node is sub-pixel) but the GPU cost drops 3x.

This is similar to Chromium's approach: content below a certain screen size gets rasterized with reduced quality during pinch-zoom.

Where AA is set

All set_anti_alias(true) calls go through a few central functions in crates/grida-canvas/src/painter/:

• paint.rs — sk_solid_paint(), sk_paint_stack(), sk_paint_stack_without_images()
• gradient.rs — gradient paint creation
• painter.rs — shadow, inner shadow, outline paints
• shadow.rs — drop/inner shadow paints
• effects_noise.rs — noise effect paints

A force_no_aa field on RenderPolicy controls AA globally. The bench CLI exposes this as --no-aa. For production, the approach should be per-node based on screen-space size, computed during the frame plan.

Benchmark measurement fix

During this investigation, we discovered that gpu_flush() was using flush_and_submit() (async, non-blocking) instead of flush_submit_and_sync_cpu() (blocking). This meant:

• mid_flush_us measured command buffer submission time, not GPU execution
• Per-stage breakdowns in FrameFlushStats were unreliable
• A/B comparisons that changed GPU workload showed false-negative results

Fixed: added sync_gpu config flag on RuntimeRendererConfig. Benchmarks enable this, making per-stage timing accurate. Note: synced benchmarks serialize CPU/GPU and understate pipelined throughput — they measure isolated GPU cost, not real-world frame rate.

Real-scene results (135K nodes, 01-135k.perf.grida)

Scenario	AA on	AA off	Delta
baseline_nocache_zoom_slow_fit (0.02x)	62,038 µs	60,103 µs	-3%
mid_flush at fit	50,599 µs	48,773 µs	-4%
baseline_nocache_zoom_slow_high (zoomed in)	21,190 µs	19,491 µs	-8%
mid_flush at high zoom	16,542 µs	15,009 µs	-9%

The improvement is smaller than the isolated bench predicted (3-9% vs 3.2x) because the real scene has complex Path nodes where picture cache replay and path tessellation overhead dominate over AA cost.

Interpretation: AA is a contributor but not the primary bottleneck on real scenes with complex geometry. The dominant cost is the per-node draw_picture dispatch + replay + GPU pipeline overhead for 41K nodes, regardless of AA state.

Related

• crates/grida-canvas/examples/skia_bench/skia_bench_subpixel.rs — isolated benchmark
• docs/wg/feat-2d/optimization.md — master optimization catalog
• Chromium pinch-zoom: reduced rasterization quality during interaction