Hi Skia team,
I don't have permission to create on issues.skia.org, so I'm posting here in the hope someone can file it.
SUMMARY
When a paint has a stroke style and a path effect that overrides the stroke rec -- SkLine2DPathEffect is the clear case -- the CPU backend applies its thin-stroke-to-hairline alpha modulation, and the path effect then restores the geometric width. The width change is undone; the alpha change is not. Rendered coverage comes out proportional to width^2 instead of width. The Metal backend renders the same drawing correctly, so the two backends disagree.
REPRO
A line grid on an 8px period. Expected mean alpha over the rect is 255 * w / 8, for any w.
SkPaint paint;
paint.setAntiAlias(true);
paint.setColor(SK_ColorRED);
paint.setStyle(SkPaint::kStroke_Style);
SkMatrix lattice;
lattice.setScale(8, 8);
const SkScalar w = 0.5f; // any width below 1 device pixel
paint.setPathEffect(SkLine2DPathEffect::Make(w, lattice));
paint.setStrokeWidth(w);
canvas->drawRect(SkRect::MakeXYWH(10, 10, 40, 40), paint);
Measured mean alpha, sampled over a whole number of periods:
width expected CPU Metal
2.0 63.75 63.75 63.75
1.0 31.88 32.00 32.00
0.5 15.94 8.00 16.00
0.25 7.97 1.88 8.00
0.125 3.98 0.00 0.00
CPU/expected is 0.50 at w=0.5 and 0.24 at w=0.25, i.e. the measured coverage tracks w^2. Metal matches expected to within rounding.
CAUSE
DrawTreatAsHairline() (src/core/SkDrawProcs.h:25) checks style, stroke width and anti-alias, but does not check whether the paint carries a path effect, so it assumes paint.getStrokeWidth() is what will actually be stroked.
modifyPaintForHairlines() (src/core/SkDraw.cpp:979) then acts on that:
paint.setStrokeWidth(0); // 1px hairline
paint.setAlpha(origAlpha * coverage); // dimmed by the real width
which is correct in isolation.
SkLine2DPathEffectImpl::onFilterPath() (src/effects/Sk2DPathEffect.cpp:127) subsequently calls rec->setStrokeStyle(fWidth), restoring the geometric width while the alpha reduction stays applied. Net: geometry at width w, alpha scaled by w, so coverage w^2.
modifyPaintForHairlines is CPU-only, which is why the GPU path is unaffected.
CONFIRMING TEST
Raising only the paint's stroke width to 10 -- which the effect overrides, so it cannot change the stroking -- suppresses the hairline substitution and restores exact coverage:
effect width expected setStrokeWidth(w) setStrokeWidth(10)
0.5 15.94 8.00 16.00
0.25 7.97 1.88 8.00
Separately, SkPathUtils::FillPathWithPaint() with the same paint produces correct geometry at every width (bounds exactly centre +/- w/2, same verb count) and filling that path renders correctly. So the effect's output is fine; only the direct drawPath route loses it. Two public entry points that should agree, disagree.
SUGGESTED FIX
An early return in DrawTreatAsHairline when a path effect is present:
if (paint.getPathEffect()) {
return false; // the effect may rewrite the stroke rec
}
or re-evaluating the hairline decision after filterPath has returned its final SkStrokeRec.
VERSION
Still present on main as of today -- DrawTreatAsHairline has no path-effect check. Observed on SK_MILESTONE 150 (via skia-safe 0.99.0), macOS arm64, CPU raster vs Metal.
LOWER-PRIORITY ASIDE
At w = 0.125 both backends render nothing, including when the geometry is materialised first via FillPathWithPaint -- so this is separate from the above. A plain drawLine at the same width renders correctly (coverage within 6% of ideal), so strokes and fills differ here. Possibly intended, but worth confirming.
Thanks,