op/clip,gpu: move approximation of complex strokes to op/clip.Op.Add

Before this change, the two renderers both had special case code for
approximating strokes they don't support natively. This change moves
that conversion to clip.Op.Add, for several reasons:

- The compute renderer no longer need fallback logic and caches for
  strokes it doesn't support.
- The approximation logic is slow. Moving it to clip.Op.Add will not
  speed it up, but will make the cost easier to spot in profiles. Until all
  strokes are supported natively, users can use macros to cache
  expensive strokes.
- Reduced garbage: Op.Add takes an op.Ops anyway, and can use that for
  storing the approximated stroke outline.

Signed-off-by: Elias Naur <mail@eliasnaur.com>

gpu/compute.go

@@ -18,10 +18,8 @@ import (
 	"gioui.org/internal/f32color"
 	"gioui.org/internal/ops"
 	"gioui.org/internal/scene"
-	"gioui.org/internal/stroke"
 	"gioui.org/layout"
 	"gioui.org/op"
-	"gioui.org/op/clip"
 )
 
 type compute struct {
@@ -671,9 +669,9 @@ func (g *compute) encodeClipStack(clip, bounds f32.Rectangle, p *pathOp, begin b
 		nclips += g.encodeClipStack(clip, bounds, p.parent, true)
 		nclips += 1
 	}
-	s := isStroke(p)
+	isStroke := p.stroke.Width > 0
 	if p != nil && p.path {
-		if s {
+		if isStroke {
 			g.enc.fillMode(scene.FillModeStroke)
 			g.enc.lineWidth(p.stroke.Width)
 		}
@@ -685,7 +683,7 @@ func (g *compute) encodeClipStack(clip, bounds f32.Rectangle, p *pathOp, begin b
 	}
 	if begin {
 		g.enc.beginClip(clip)
-		if s {
+		if isStroke {
 			g.enc.fillMode(scene.FillModeNonzero)
 		}
 		if p != nil && p.path {
@@ -695,31 +693,8 @@ func (g *compute) encodeClipStack(clip, bounds f32.Rectangle, p *pathOp, begin b
 	return nclips
 }
 
-func supportsStroke(p *pathOp) bool {
+func encodePath(verts []byte) encoder {
-	return stroke.IsSolidLine(p.dashes) && p.stroke.Miter == 0 && p.stroke.Join == clip.RoundJoin && p.stroke.Cap == clip.RoundCap
-}
-
-func isStroke(p *pathOp) bool {
-	return p.stroke.Width > 0 && supportsStroke(p)
-}
-
-func encodePath(p *pathOp) encoder {
 	var enc encoder
-	verts := p.pathVerts
-	if p.stroke.Width > 0 && !supportsStroke(p) {
-		ss := stroke.StrokeStyle{
-			Width: p.stroke.Width,
-			Miter: p.stroke.Miter,
-			Cap:   stroke.StrokeCap(p.stroke.Cap),
-			Join:  stroke.StrokeJoin(p.stroke.Join),
-		}
-		quads := stroke.StrokePathCommands(ss, p.dashes, verts)
-		for _, quad := range quads {
-			q := quad.Quad
-			enc.quad(q.From, q.Ctrl, q.To)
-		}
-		return enc
-	}
 	for len(verts) >= scene.CommandSize+4 {
 		cmd := ops.DecodeCommand(verts[4:])
 		enc.scene = append(enc.scene, cmd)

gpu/gpu.go

@@ -131,7 +131,6 @@ type pathOp struct {
 	// For compute
 	trans  f32.Affine2D
 	stroke clip.StrokeStyle
-	dashes stroke.DashOp
 }
 
 type imageOp struct {
@@ -143,29 +142,14 @@ type imageOp struct {
 	place    placement
 }
 
-func decodeDashOp(data []byte) stroke.DashOp {
-	_ = data[5]
-	if opconst.OpType(data[0]) != opconst.TypeDash {
-		panic("invalid op")
-	}
-	bo := binary.LittleEndian
-	return stroke.DashOp{
-		Phase:  math.Float32frombits(bo.Uint32(data[1:])),
-		Dashes: make([]float32, data[5]),
-	}
-}
-
 func decodeStrokeOp(data []byte) clip.StrokeStyle {
-	_ = data[10]
+	_ = data[4]
 	if opconst.OpType(data[0]) != opconst.TypeStroke {
 		panic("invalid op")
 	}
 	bo := binary.LittleEndian
 	return clip.StrokeStyle{
 		Width: math.Float32frombits(bo.Uint32(data[1:])),
-		Miter: math.Float32frombits(bo.Uint32(data[5:])),
-		Cap:   clip.StrokeCap(data[9]),
-		Join:  clip.StrokeJoin(data[10]),
 	}
 }
 
@@ -827,7 +811,7 @@ func (d *drawOps) collect(ctx driver.Device, cache *resourceCache, root *op.Ops,
 			data := buildPath(ctx, p.pathVerts)
 			var computePath encoder
 			if d.compute {
-				computePath = encodePath(p)
+				computePath = encodePath(p.pathVerts)
 			}
 			d.pathCache.put(p.pathKey, opCacheValue{
 				data:        data,
@@ -844,7 +828,7 @@ func (d *drawOps) newPathOp() *pathOp {
 	return &d.pathOpCache[len(d.pathOpCache)-1]
 }
 
-func (d *drawOps) addClipPath(state *drawState, aux []byte, auxKey ops.Key, bounds f32.Rectangle, off f32.Point, tr f32.Affine2D, stroke clip.StrokeStyle, dashes stroke.DashOp) {
+func (d *drawOps) addClipPath(state *drawState, aux []byte, auxKey ops.Key, bounds f32.Rectangle, off f32.Point, tr f32.Affine2D, stroke clip.StrokeStyle) {
 	npath := d.newPathOp()
 	*npath = pathOp{
 		parent: state.cpath,
@@ -852,7 +836,6 @@ func (d *drawOps) addClipPath(state *drawState, aux []byte, auxKey ops.Key, boun
 		off:    off,
 		trans:  tr,
 		stroke: stroke,
-		dashes: dashes,
 	}
 	state.cpath = npath
 	if len(aux) > 0 {
@@ -882,10 +865,9 @@ func (d *drawOps) save(id int, state drawState) {
 
 func (d *drawOps) collectOps(r *ops.Reader, state drawState) {
 	var (
-		quads  quadsOp
+		quads quadsOp
-		str    clip.StrokeStyle
+		str   clip.StrokeStyle
-		dashes stroke.DashOp
+		z     int
-		z      int
 	)
 	d.save(opconst.InitialStateID, state)
 loop:
@@ -897,22 +879,6 @@ loop:
 			dop := ops.DecodeTransform(encOp.Data)
 			state.t = state.t.Mul(dop)
 
-		case opconst.TypeDash:
-			dashes = decodeDashOp(encOp.Data)
-			if len(dashes.Dashes) > 0 {
-				encOp, ok = r.Decode()
-				if !ok {
-					panic("gpu: could not decode dashes pattern")
-				}
-				data := encOp.Data[1:]
-				bo := binary.LittleEndian
-				for i := range dashes.Dashes {
-					dashes.Dashes[i] = math.Float32frombits(bo.Uint32(
-						data[i*4:],
-					))
-				}
-			}
-
 		case opconst.TypeStroke:
 			str = decodeStrokeOp(encOp.Data)
 
@@ -940,7 +906,7 @@ loop:
 					op.bounds = v.bounds
 				} else {
 					pathData, bounds := d.buildVerts(
-						quads.aux, trans, op.outline, str, dashes,
+						quads.aux, trans, op.outline, str,
 					)
 					op.bounds = bounds
 					if !d.compute {
@@ -956,10 +922,9 @@ loop:
 				quads.key.SetTransform(trans)
 			}
 			state.clip = state.clip.Intersect(op.bounds.Add(off))
-			d.addClipPath(&state, quads.aux, quads.key, op.bounds, off, state.t, str, dashes)
+			d.addClipPath(&state, quads.aux, quads.key, op.bounds, off, state.t, str)
 			quads = quadsOp{}
 			str = clip.StrokeStyle{}
-			dashes = stroke.DashOp{}
 
 		case opconst.TypeColor:
 			state.matType = materialColor
@@ -997,7 +962,7 @@ loop:
 				// The paint operation is sheared or rotated, add a clip path representing
 				// this transformed rectangle.
 				encOp.Key.SetTransform(trans)
-				d.addClipPath(&state, clipData, encOp.Key, bnd, off, state.t, clip.StrokeStyle{}, stroke.DashOp{})
+				d.addClipPath(&state, clipData, encOp.Key, bnd, off, state.t, clip.StrokeStyle{})
 			}
 
 			bounds := boundRectF(cl)
@@ -1349,7 +1314,7 @@ func (d *drawOps) writeVertCache(n int) []byte {
 }
 
 // transform, split paths as needed, calculate maxY, bounds and create GPU vertices.
-func (d *drawOps) buildVerts(pathData []byte, tr f32.Affine2D, outline bool, str clip.StrokeStyle, dashes stroke.DashOp) (verts []byte, bounds f32.Rectangle) {
+func (d *drawOps) buildVerts(pathData []byte, tr f32.Affine2D, outline bool, str clip.StrokeStyle) (verts []byte, bounds f32.Rectangle) {
 	inf := float32(math.Inf(+1))
 	d.qs.bounds = f32.Rectangle{
 		Min: f32.Point{X: inf, Y: inf},
@@ -1367,7 +1332,7 @@ func (d *drawOps) buildVerts(pathData []byte, tr f32.Affine2D, outline bool, str
 			Cap:   stroke.StrokeCap(str.Cap),
 			Join:  stroke.StrokeJoin(str.Join),
 		}
-		quads := stroke.StrokePathCommands(ss, dashes, pathData)
+		quads := stroke.StrokePathCommands(ss, stroke.DashOp{}, pathData)
 		for _, quad := range quads {
 			d.qs.contour = quad.Contour
 			quad.Quad = quad.Quad.Transform(tr)

internal/opconst/ops.go

@@ -33,7 +33,6 @@ const (
 	TypeCursor
 	TypePath
 	TypeStroke
-	TypeDash
 )
 
 const (
@@ -61,8 +60,7 @@ const (
 	TypeProfileLen         = 1
 	TypeCursorLen          = 1 + 1
 	TypePathLen            = 1
-	TypeStrokeLen          = 1 + 4 + 4 + 1 + 1
+	TypeStrokeLen          = 1 + 4
-	TypeDashLen            = 1 + 4 + 1
 )
 
 // StateMask is a bitmask of state types a load operation
@@ -106,7 +104,6 @@ func (t OpType) Size() int {
 		TypeCursorLen,
 		TypePathLen,
 		TypeStrokeLen,
-		TypeDashLen,
 	}[t-firstOpIndex]
 }
 

op/clip/clip.go

@@ -27,29 +27,31 @@ type Op struct {
 }
 
 func (p Op) Add(o *op.Ops) {
-	if p.path.hasSegments {
+	str := p.stroke
-		data := o.Write(opconst.TypePathLen)
+	dashes := p.dashes
-		data[0] = byte(opconst.TypePath)
+	path := p.path
-		p.path.spec.Add(o)
+	outline := p.outline
+	approx := str.Width > 0 && !(dashes == DashSpec{} && str.Miter == 0 && str.Join == RoundJoin && str.Cap == RoundCap)
+	if approx {
+		// If the stroke is not natively supported by the compute renderer, construct a filled path
+		// that approximates it.
+		path = p.approximateStroke(o)
+		dashes = DashSpec{}
+		str = StrokeStyle{}
+		outline = true
 	}
 
-	if p.stroke.Width > 0 {
+	if path.hasSegments {
+		data := o.Write(opconst.TypePathLen)
+		data[0] = byte(opconst.TypePath)
+		path.spec.Add(o)
+	}
+
+	if str.Width > 0 {
 		data := o.Write(opconst.TypeStrokeLen)
 		data[0] = byte(opconst.TypeStroke)
 		bo := binary.LittleEndian
-		bo.PutUint32(data[1:], math.Float32bits(p.stroke.Width))
+		bo.PutUint32(data[1:], math.Float32bits(str.Width))
-		bo.PutUint32(data[5:], math.Float32bits(p.stroke.Miter))
-		data[9] = uint8(p.stroke.Cap)
-		data[10] = uint8(p.stroke.Join)
-	}
-
-	if p.dashes.phase != 0 || p.dashes.size > 0 {
-		data := o.Write(opconst.TypeDashLen)
-		data[0] = byte(opconst.TypeDash)
-		bo := binary.LittleEndian
-		bo.PutUint32(data[1:], math.Float32bits(p.dashes.phase))
-		data[5] = p.dashes.size // FIXME(sbinet) uint16? uint32?
-		p.dashes.spec.Add(o)
 	}
 
 	data := o.Write(opconst.TypeClipLen)
@@ -59,17 +61,77 @@ func (p Op) Add(o *op.Ops) {
 	bo.PutUint32(data[5:], uint32(p.bounds.Min.Y))
 	bo.PutUint32(data[9:], uint32(p.bounds.Max.X))
 	bo.PutUint32(data[13:], uint32(p.bounds.Max.Y))
-	if p.outline {
+	if outline {
 		data[17] = byte(1)
 	}
 }
 
+func (p Op) approximateStroke(o *op.Ops) PathSpec {
+	if !p.path.hasSegments {
+		return PathSpec{}
+	}
+
+	var r ops.Reader
+	// Add path op for us to decode. Use a macro to omit it from later decodes.
+	ignore := op.Record(o)
+	r.ResetAt(o, ops.NewPC(o))
+	p.path.spec.Add(o)
+	ignore.Stop()
+	encOp, ok := r.Decode()
+	if !ok || opconst.OpType(encOp.Data[0]) != opconst.TypeAux {
+		panic("corrupt path data")
+	}
+	pathData := encOp.Data[opconst.TypeAuxLen:]
+
+	// Decode dashes in a similar way.
+	var dashes stroke.DashOp
+	if p.dashes.phase != 0 || p.dashes.size > 0 {
+		ignore := op.Record(o)
+		r.ResetAt(o, ops.NewPC(o))
+		p.dashes.spec.Add(o)
+		ignore.Stop()
+		encOp, ok := r.Decode()
+		if !ok || opconst.OpType(encOp.Data[0]) != opconst.TypeAux {
+			panic("corrupt dash data")
+		}
+		dashes.Dashes = make([]float32, p.dashes.size)
+		dashData := encOp.Data[opconst.TypeAuxLen:]
+		bo := binary.LittleEndian
+		for i := range dashes.Dashes {
+			dashes.Dashes[i] = math.Float32frombits(bo.Uint32(dashData[i*4:]))
+		}
+		dashes.Phase = p.dashes.phase
+	}
+
+	// Approximate and output path data.
+	var outline Path
+	outline.Begin(o)
+	ss := stroke.StrokeStyle{
+		Width: p.stroke.Width,
+		Miter: p.stroke.Miter,
+		Cap:   stroke.StrokeCap(p.stroke.Cap),
+		Join:  stroke.StrokeJoin(p.stroke.Join),
+	}
+	quads := stroke.StrokePathCommands(ss, dashes, pathData)
+	pen := f32.Pt(0, 0)
+	for _, quad := range quads {
+		q := quad.Quad
+		if q.From != pen {
+			pen = q.From
+			outline.MoveTo(pen)
+		}
+		outline.contour = int(quad.Contour)
+		outline.QuadTo(q.Ctrl, q.To)
+	}
+	return outline.End()
+}
+
 type PathSpec struct {
 	spec op.CallOp
 	// open is true if any path contour is not closed. A closed contour starts
 	// and ends in the same point.
 	open bool
-	// hasSegments tracks whether there is more than one path segment in the path.
+	// hasSegments tracks whether there are any segments in the path.
 	hasSegments bool
 }
 

op/clip/stroke.go

@@ -93,7 +93,7 @@ func (d *Dash) Phase(v float32) {
 
 func (d *Dash) Dash(length float32) {
 	if d.size == math.MaxUint8 {
-		panic("clip: too large dash pattern")
+		panic("clip: dash pattern too large")
 	}
 	data := d.ops.Write(4)
 	bo := binary.LittleEndian