[Tracking] Re-enable the GPU backends currently opted out of CI (no OpenGL or Metal coverage anywhere)

[mattleibow]

Description

Umbrella issue. #4575 made GPU backends required rather than silently skipped — a backend that cannot run must now be declared via SKIASHARP_TEST_SKIP_GPU, not inferred from a caught exception. That change did its job: it turned a set of invisible skips into a visible, enumerable list of gaps.

This issue tracks closing them. Every opt-out below is a coverage hole, each with a known cause. None of them existed as "we chose not to test this" — they were all previously skipping without anything in the build saying so.

Today CI has no OpenGL coverage anywhere, and no Metal coverage anywhere.

Current opt-outs

Leg	Backends skipped	Why	Tracking
Linux	ganesh-gl	Nine test classes create GLX contexts concurrently → host crashes, exit 7, no results	#4590
Windows (.NET Framework)	ganesh-gl	Agent ships only the generic OpenGL 1.1 implementation — no WGL_ARB_* extensions, so WglContext cannot come up	this issue
Windows (.NET Core)	ganesh-gl	same	this issue
Linux (Azure Linux)	ganesh-gl	container image has no X server, Mesa or ICD	this issue
Linux (Azure Linux, No Deps)	ganesh-gl	same	this issue
Linux (Alpine)	ganesh-gl	same	this issue
Linux (Alpine, No Deps)	ganesh-gl	same	this issue
Windows (Nano Server)	ganesh-gl	same	this issue
macOS	ganesh-metal, graphite-metal	x64 pool virtualizes Metal; creating a device + queue leaves state that hangs the host on shutdown (job killed at the 180 min timeout, no TRX)	#4555
iOS	ganesh-metal, graphite-metal	simulator's virtualized Metal; also cannot compile Graphite's gradient pipeline	#4555
Mac Catalyst	ganesh-metal, graphite-metal	a Catalyst app never sees the Mac2 GPU family Skia requires	#4555
WASM	graphite-dawn	navigator.gpu.requestAdapter() returns null — the DeviceRunners CLI's wasm test launches a headless browser with no WebGPU and exposes no browser-argument passthrough	upstream (DeviceRunners)
WASM (.NET Preview)	graphite-dawn	same	upstream (DeviceRunners)

What still works

Not everything is opted out — these are genuinely covered and must stay green:

• raster on all 13 legs
• ganesh-vulkan + graphite-vulkan on Windows, Linux, Android (including software lavapipe on Linux and SwiftShader on Windows)
• ganesh-direct3d and ganesh-vulkan-sharpvk on Windows

Work items

1. Linux OpenGL — #4590. Put every class calling SKTest.CreateGlContext() into one xUnit collection so GLX contexts are never created concurrently, mirroring GpuRenderingCollection (visual) and VulkanGpuRenderingCollection (Vulkan). Then drop the Linux opt-out. This is the most tractable item and the one most likely to restore real coverage.

2. Windows OpenGL. The agents have no usable GL. Investigated during #4575:

• Mesa (mesa-dist-win) works — verified correct per-arch PE headers and the required WGL extensions present. opengl32.dll + libgallium_wgl.dll would go into System32/SysWOW64 exactly as the Vulkan loader already does in scripts/infra/native/windows/install-vulkan-icd.ps1, and needs no test-code change. Blocker is that it has no trustworthy package feed — it would need mirroring.
• Silk.NET.OpenGLES.ANGLE.Native does not work — every published version ships 32-bit binaries in runtimes/win-x64/. Verified across all 4 versions. Do not reach for it.

3. Container images. Same Mesa question as (2) but for Linux/Alpine images. Lower value — these images exist to prove SkiaSharp works without extra dependencies, so software GL there may be out of scope by design. Worth an explicit decision so the opt-out becomes intentional rather than pending.

4. Metal — #4555. All three Apple legs are virtualized and each fails differently. We already run device tests, so an Apple Silicon pool (or a real device for iOS) would likely resolve all three at once and is probably cheaper than chasing each symptom.

5. WASM / Dawn — upstream. Needs a browser-argument passthrough in the DeviceRunners CLI. Measured on Chrome 150 (the version scripts/infra/managed/install-chrome.sh now pins):

Flags	requestAdapter()
--disable-gpu (CI-like)	null
(none)	ok — real GPU
--enable-unsafe-webgpu --use-webgpu-adapter=swiftshader	ok — SwiftShader
the above plus --disable-gpu	ok — SwiftShader

The last row matters: WebGPU works with no GPU at all, so this needs no special agent hardware — only the ability to pass the flags. forceFallbackAdapter: true does not help; the software adapter must be enabled at launch.

Golden coverage gaps

Independent of the opt-outs, these have no committed goldens, so re-enabling a backend also means harvesting and reviewing them:

• ganesh-gl.linux — nothing; blocked on (1)
• graphite-metal.maccatalyst — nothing; blocked on (4)
• graphite-dawn.* — nothing on any platform; blocked on (5)
• graphite-metal.ios — 4 of 5 scenes (GradientBlend missing, see [BUG] Metal is unusable on every CI host: macOS agent hangs on shutdown, Catalyst reports no GPU family, simulator cannot compile Graphite's gradient pipeline #4555)

Definition of done

scripts/azure-templates-stages-test.yml contains no SKIASHARP_TEST_SKIP_GPU entries, and GpuPolicy.RequiredOn is satisfied on every leg it names.

Partial progress is fine and expected — each row removed is a real gain. The opt-out mechanism is deliberately per-leg so items can be closed one at a time.

Related

• Make GPU tests required; skip only when declared #4575 — made GPU backends required; introduced these declarations
• [BUG] Concurrent GLX context creation crashes the Linux test host (exit 7, no results published) #4590 — Linux GLX concurrency crash
• [BUG] Metal is unusable on every CI host: macOS agent hangs on shutdown, Catalyst reports no GPU family, simulator cannot compile Graphite's gradient pipeline #4555 — Metal unusable on every CI host

Version of SkiaSharp

Nightly / CI.

IDE / Editor

N/A — CI only (dnceng-public → mono-SkiaSharp).

Platform / Operating System

Windows, macOS, Linux, iOS, Mac Catalyst, WASM

Platform / Operating System Version

Azure DevOps hosted agents and container images