[besmpl]

Source-aware GPU font fallback (#451)

MultiFace now exposes its fallback chain and contiguous source-font runs, and shaping retains the owning face on every ShapedGlyph. The glyph-mask and MSDF GPU paths rasterize those runs into ordered batches instead of returning a nil-source error; shaped CPU fallback also resolves each glyph's source. Mixed-font tests cover deterministic run offsets, source identity, glyph-mask batching, and MSDF batching.

Fixes #521.

Verification: go test ./..., go test -tags nogpu ./..., go vet ./....

[codecov]

Codecov Report

❌ Patch coverage is 86.98347% with 63 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/gpu/glyph_mask_engine.go	64.40%	23 Missing and 19 partials ⚠️
internal/gpu/gpu_text.go	63.79%	11 Missing and 10 partials ⚠️

📢 Thoughts on this report? Let us know!

[kolkov]

@besmpl, same apology here for the long silence, for the same reason.

Nine days after you opened this, I described the fix in #521 without linking your PR, which already implements it. That was a miss on my side. The direction matches what we planned: split by source face before shaping, keep the owning face on each glyph, and send single-font runs to the glyph-mask and MSDF paths. That replaces our own plan to build it.

I've read the diff, but I haven't run it yet. Its CI ran against the August main, and that run can't be restarted.

Needed before merge:

1. A rebase onto current main, so CI runs on the code this would merge into. text/ and internal/gpu/ changed since August, and more lands with the next release. If you'd prefer to rebase once after that release, say so and I'll ping you when it's on main.
2. This adds public API: FontRun, MultiFace.Faces/FaceForRune/FontRuns/ShapeRuns, FilteredFace.FontRuns, text.ShapeRuns and ShapedGlyph.Face. We're freezing the API for v1.0, so each exported name has to be one we'd keep forever. Two questions:
   • Does FontRun.IsCJK need to be exported? It's a policy flag for our atlas and hinting choice. An unexported field, or a script value rather than a bool, would age better.
   • Do we need both MultiFace.ShapeRuns and text.ShapeRuns? The function form already dispatches to FontRuns when the face has no single source, so it covers the method's case.

Optional: ShapedGlyph gains an interface field, 16 bytes per glyph. If you have numbers for a long paragraph through scene before and after, they'd help. I'll measure it myself otherwise.

[kolkov]

@besmpl, same here: I've done the changes on your branch, your authorship kept. Your three commits are rebased onto main without conflicts.

On the two API questions:

• text.ShapeRuns is the only form now. MultiFace.ShapeRuns is gone. It had no callers outside its tests, and the function already sends a face without a single source to FontRuns. MultiFace.FontRuns now accepts a nil receiver, which the method used to guard.
• FontRun.IsCJK stays exported. I asked about it before noticing that ShapedGlyph.IsCJK has been public since v0.46.6, and that the GPU code reads the run's flag from another package. Whether both should become a script value instead of a bool is a separate decision we'll make for both types before v1.0.

Cost of ShapedGlyph.Face: the struct goes from 56 to 72 bytes on amd64, the price of a glyph knowing its font.

The CHANGELOG credits you. Thanks for building this.

[kolkov]

Rebased onto current main, CI green, merging for v0.52.7. Your commits stay yours; the one on top drops the duplicate ShapeRuns and the helper only it used.