diff --git a/gpu/caches.go b/gpu/caches.go
index 3ad5ba1b..b4ab29db 100644
--- a/gpu/caches.go
+++ b/gpu/caches.go
@@ -27,9 +27,10 @@ type opCache struct {
 
 type opCacheValue struct {
 	data pathData
-	// cpuData is the retained data, if kept by the compute renderer.
-	cpuData []byte
-	bounds  f32.Rectangle
+	// computePath is the encoded path for compute.
+	computePath encoder
+
+	bounds f32.Rectangle
 	// the fields below are handled by opCache
 	key  ops.Key
 	keep bool
diff --git a/gpu/compute.go b/gpu/compute.go
index 2df1dfa2..811a7ab2 100644
--- a/gpu/compute.go
+++ b/gpu/compute.go
@@ -409,7 +409,9 @@ func (g *compute) encodeClipStack(clip, bounds f32.Rectangle, p *pathOp) int {
 	}
 	if p != nil && p.path {
 		pathData, _ := g.drawOps.pathCache.get(p.pathKey)
-		g.encodePath(p.off, pathData.cpuData)
+		g.enc.transform(f32.Affine2D{}.Offset(p.off))
+		g.enc.append(pathData.computePath)
+		g.enc.transform(f32.Affine2D{}.Offset(p.off.Mul(-1)))
 	} else {
 		g.enc.rect(bounds, false)
 	}
@@ -419,7 +421,8 @@ func (g *compute) encodeClipStack(clip, bounds f32.Rectangle, p *pathOp) int {
 // encodePath takes a Path encoded with quadSplitter and encode it for elements.comp.
 // This is certainly wasteful, but minimizes implementation differences to the old
 // renderer.
-func (g *compute) encodePath(off f32.Point, p []byte) {
+func encodePath(p []byte) encoder {
+	var enc encoder
 	for len(p) > 0 {
 		// p contains quadratic curves encoded in vertex structs.
 		vertex := p[:vertStride]
@@ -436,12 +439,13 @@ func (g *compute) encodePath(off f32.Point, p []byte) {
 			math.Float32frombits(bo.Uint32(vertex[24:])),
 			math.Float32frombits(bo.Uint32(vertex[28:])),
 		)
-		g.enc.quad(from.Add(off), ctrl.Add(off), to.Add(off), false)
+		enc.quad(from, ctrl, to, false)
 
 		// The vertex is duplicated 4 times, one for each corner of quads drawn
 		// by the old renderer.
 		p = p[vertStride*4:]
 	}
+	return enc
 }
 
 func (g *compute) render(tileDims image.Point) error {
@@ -696,6 +700,12 @@ func (e *encoder) numElements() int {
 	return len(e.scene) / sceneElemSize
 }
 
+func (e *encoder) append(e2 encoder) {
+	e.scene = append(e.scene, e2.scene...)
+	e.npath += e2.npath
+	e.npathseg += e2.npathseg
+}
+
 func (e *encoder) transform(m f32.Affine2D) {
 	sx, hx, ox, hy, sy, oy := m.Elems()
 	cmd := make([]byte, sceneElemSize)
diff --git a/gpu/gpu.go b/gpu/gpu.go
index eaa000f5..0079dfd1 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -824,15 +824,14 @@ func (d *drawOps) collect(ctx backend.Device, cache *resourceCache, root *op.Ops
 	for _, p := range d.pathOps {
 		if v, exists := d.pathCache.get(p.pathKey); !exists || v.data.data == nil {
 			data := buildPath(ctx, p.pathVerts)
-			var pathVerts []byte
+			var computePath encoder
 			if d.retainPathData {
-				pathVerts = make([]byte, len(p.pathVerts))
-				copy(pathVerts, p.pathVerts)
+				computePath = encodePath(p.pathVerts)
 			}
 			d.pathCache.put(p.pathKey, opCacheValue{
-				data:    data,
-				bounds:  p.bounds,
-				cpuData: pathVerts,
+				data:        data,
+				bounds:      p.bounds,
+				computePath: computePath,
 			})
 		}
 		p.pathVerts = nil