tracking: AA rasterization improvements roadmap

[kolkov]

Summary

Tracking issue for AA (Anti-Aliasing) rasterization improvements in gg. Created after deep investigation of #95 and comprehensive AA architecture audit.

Current State (v0.31.0)

Production CPU path works correctly after two bug fixes (#95):

• AdvanceX() drift removed — edges no longer expand toward bottom of shapes
• coverageToRuns() maxValue bug fixed — vertex pixels get correct partial coverage
• 4 regression tests added

Vello 9-stage GPU compute pipeline (v0.30.0–v0.31.0):

• CPU reference implementation in internal/gpu/tilecompute/
• 9 WGSL shaders written and dispatched via GPU compute
• Fixed: fine.wgsl y_edge sign, coarse.wgsl Z-order
• Full demo scene (6 paths, transparency, EvenOdd) — 0% pixel diff GPU=CPU
• v0.31.0: Tier 5 scene accumulation wired into rendering cascade (GG-COMPUTE-008)
• v0.31.0: PipelineMode wiring — Context.SetPipelineMode() propagates to GPU accelerator (GG-COMPUTE-006)
• v0.31.0: Removed 2 naga WGSL workarounds from path_tiling.wgsl (see tracking: remaining naga SPIR-V workarounds in compute shaders #139 for remaining 3)

Remaining pixel-level differences vs fogleman/gg (max 7-8/255) are inherent algorithmic differences, not bugs — confirmed by mathematical proof (trapezoidal area vs 4x4 supersampling).

Architecture (5 rendering systems)

System	Status	Description
AnalyticFiller (Vello) + EdgeBuilder (tiny-skia)	Production	Scanline AA with trapezoidal coverage, FDot16 fixed-point
CPU SDF	Production	Smoothstep SDF for circles/rects/rrects
GPU 5-Tier Pipeline	Production	SDF + convex + stencil-then-cover + MSDF text + Vello compute
TileRasterizer (Vello 16×16)	Validation	CPU reference for GPU compute golden tests
GPU Compute Pipeline (9-stage)	Production (v0.31.0)	9 WGSL shaders, VelloAccelerator dispatches via hal.Queue

Current rendering cascade (automatic per-draw fallback via ErrFallbackToCPU):

GPU 5-Tier → CPU SDF → AnalyticFiller

Improvement Tasks

Quality (v0.32.0+):

• GG-AA-018: Native curve evaluation — eliminate Bezier flattening for CPU path (P1)
• GG-AA-020: Float32 pipeline — remove FDot16 quantization noise (P2)
• GG-AA-019: Configurable AA quality — hint-based API (P3)

Integration (completed):

• GG-AA-016: TileRasterizer validation — CPU reference + golden test suite for GPU compute ✅ v0.30.0
• GG-AA-017: GPU Compute Pipeline — 9 WGSL shaders written, CPU reference complete ✅ v0.30.0
• GG-COMPUTE-008: Tier 5 scene accumulation — VelloAccelerator accumulates paths during FillPath/StrokePath, dispatches via compute on Flush ✅ v0.31.0
• GG-COMPUTE-006: PipelineMode wiring — Context.SetPipelineMode() propagates to GPU accelerator, SDFAccelerator routes to compute when active ✅ v0.31.0

Shader cleanup:

• Removed span() inlining workaround from path_tiling.wgsl ✅ v0.31.0
• Removed var reassignment workaround from path_tiling.wgsl ✅ v0.31.0
• 3 naga SPIR-V workarounds remain — tracked in tracking: remaining naga SPIR-V workarounds in compute shaders #139

Testing:

• GG-AA-015: Visual AA comparison test suite — automated quality metrics (P2)

Related Issues

• Pattern result is strange #95 — Pattern rendering AA quality (fixed in v0.29.4)
• tracking: remaining naga SPIR-V workarounds in compute shaders #139 — naga SPIR-V workarounds tracking (3 remaining)
• help wanted: Bottom circle artifact in Vello tile rasterizer #72 — Circle bottom artifact (closed — resolved by GPU compute pipeline, v0.30.0)
• feat: Vello-style tile-based analytic anti-aliasing #50 — Vello tile-based analytic AA (closed — implemented in v0.30.0)