Add regression-tracking benchmarks for common colour/interop paths

[mattleibow]

What

Expand benchmarks/SkiaSharp.Benchmarks.Tracking with a small, curated set of permanent regression benchmarks, derived from the areas the recent perf/agentic PRs touch. Each new benchmark is one .cs file in Tracking/Benchmarks/, so it's automatically picked up by both the package-mode project and the source-mode …Tracking.Source project (the ⭐ pr column).

Following the project's philosophy (and how dotnet/performance / aspnetcore structure their suites), this is a scaled pyramid — a few small common-path trackers + one large "touch many things" indicator — not a micro-benchmark per PR.

Tier 1 — small common-path trackers

• ColorMathBenchmark — managed SKColor → SKColorF ([performance] Port SKColor -> SKColorF conversion to managed C# #4370) + SKPMColor.PreMultiply / UnPreMultiply ([performance] Port single-value SKPMColor premultiply/unpremultiply to managed C# #4385). Pure managed, zero-alloc; the ToColorF/ToColor ratio (the reverse operator stays native) is a clean managed-regression signal even on noisy shared runners.
• RasterImageLifecycleBenchmark — create+dispose churn of SKImage.Create raster ([memory-leak] Fix CoTaskMem pixel-buffer leak in SKImage.Create raster failure path #4455) and SKData.Create with a managed release proc ([memory-leak] Fix GCHandle leak on release-proc factory failure path #4453). MemoryDiagnoser's Allocated bytes are the tracked signal for the leak-fix area.
• RuntimeEffectShaderBenchmark — a per-frame animated SkSL shader: build uniforms (an SKColor uniform hits the [performance] Port SKColor -> SKColorF conversion to managed C# #4370 convert), ToShader, draw. Exercises the uniforms lifecycle whose SKData leak [memory-leak] Fix SKData leak in SKRuntimeEffectUniforms.Reset() #4442 fixed; per-frame allocation is the headline.

Tier 2 — large composite

• SceneRenderBenchmark — one broad frame (transforms, linear + radial gradients, filled/stroked primitives, a built path, a scaled image draw = the [All, critical on Windows] Severe Software (CPU) Rasterization Performance Regression / Flickering starting in Milestone 146` #4421 CPU sampling path, a colour-filter layer, clipping, managed colour maths). Deliberately broad so many unrelated future optimizations each nudge it — a merge indicator.

Not added: #4372 (niche SKNWayCanvas/SKOverdrawCanvas debug wrappers) and #4468 (HarfBuzz — a separate assembly the tracking project doesn't reference).

Constraints honoured

• Public API only (project references the nightly NuGet, no internal SkiaApi).
• Deterministic + machine-independent: fixed seeds, no fonts/external content.
• Never-rename names chosen for permanent history keys.
• Compile against both the 3.119 baseline and 4.x nightly (only SKPath/SKPathBuilder needs the existing #if).

Validation

Local (short job)

Built against SkiaSharp 4.151.0-preview.2.1 and 3.119.4; a --job short run emits time + allocation data for every case:

Benchmark	Mean	Allocated
ColorMathBenchmark.ToColorF (managed, #4370)	9,986 ns	0 B
ColorMathBenchmark.ToColor (native reverse, sentinel)	15,930 ns	0 B
ColorMathBenchmark.PreMultiply (#4385)	10,239 ns	0 B
ColorMathBenchmark.UnPreMultiply (#4385)	6,986 ns	0 B
RasterImageLifecycleBenchmark.CreateRasterImage (#4455)	235 µs	55,296 B
RasterImageLifecycleBenchmark.CreateDataWithReleaseProc (#4453)	57.7 µs	34,816 B
RuntimeEffectShaderBenchmark.DrawFrame (#4442)	328 µs	680 B
SceneRenderBenchmark.RenderFrame (Complexity 1 / 4)	1.05 ms / 4.53 ms	0 B

(The managed ToColorF already measuring faster than the still-native ToColor is the #4370 win showing through.)

CI baseline — pr column, run 29608575980 (2026-07-17)

Full source build of this branch (the ⭐ pr role) on the GitHub-hosted runners — one point per OS, the baseline each future run diffs against. Managed allocation is OS-independent (single column). On pr/HEAD ToColorF is still native, so expect its line to drop once #4370 merges.

Benchmark (params)	Linux	Windows	macOS	Allocated
ColorMathBenchmark.ToColorF (4096)	21.5 µs	32.6 µs	14.5 µs	0 B
ColorMathBenchmark.ToColor (4096)	14.1 µs	15.8 µs	21.4 µs	0 B
ColorMathBenchmark.PreMultiply (4096)	15.9 µs	23.1 µs	11.1 µs	0 B
ColorMathBenchmark.UnPreMultiply (4096)	14.6 µs	19.2 µs	9.0 µs	0 B
RasterImageLifecycleBenchmark.CreateRasterImage (256)	263.5 µs	283.7 µs	166.3 µs	54.0 KB
RasterImageLifecycleBenchmark.CreateDataWithReleaseProc (256)	117.2 µs	131.1 µs	84.6 µs	34.0 KB
RuntimeEffectShaderBenchmark.DrawFrame	2.83 ms	1.12 ms	558 µs	~0.7 KB
SceneRenderBenchmark.RenderFrame (Complexity 1)	4.90 ms	2.56 ms	1.46 ms	0 B
SceneRenderBenchmark.RenderFrame (Complexity 4)	19.18 ms	10.76 ms	6.04 ms	0 B

All 21 legs green (3 OSes × source pr + 5 package roles incl. prev-major 3.119.4); the perf-dashboard artifact renders all four benchmarks.

Absolute ns are environment-dependent; the tracked signals are per-OS version-role ratios + allocations.

[github-actions]

📦 Try the packages from this PR

Warning

Do not run these scripts without first reviewing the code in this PR.

Step 1 — Download the packages

bash / macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/mono/SkiaSharp/main/scripts/get-skiasharp-pr.sh | bash -s -- 4476

PowerShell / Windows:

iex "& { $(irm https://raw.githubusercontent.com/mono/SkiaSharp/main/scripts/get-skiasharp-pr.ps1) } 4476"

Step 2 — Add the local NuGet source

dotnet nuget add source ~/.skiasharp/hives/pr-4476/packages --name skiasharp-pr-4476

More options

Option	Description
--successful-only / -SuccessfulOnly	Only use successful builds
--force / -Force	Overwrite previously downloaded packages
--list / -List	List available artifacts without downloading
--build-id ID / -BuildId ID	Download from a specific build

Or download manually from Azure Pipelines — look for the nuget artifact on the build for this PR.

Remove the source when you're done:

dotnet nuget remove source skiasharp-pr-4476