diff --git a/gpu/gpu.go b/gpu/gpu.go
index 334f03a4..0cb78a4a 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -69,7 +69,7 @@ type drawOps struct {
 
 type drawState struct {
 	clip  f32.Rectangle
-	t     op.TransformOp
+	t     f32.Affine2D
 	cpath *pathOp
 	rect  bool
 	z     int
@@ -691,8 +691,8 @@ loop:
 		case opconst.TypeProfile:
 			d.profile = true
 		case opconst.TypeTransform:
-			dop := ops.DecodeTransformOp(encOp.Data)
-			state.t = state.t.Multiply(op.TransformOp(dop))
+			dop := ops.DecodeTransform(encOp.Data)
+			state.t = state.t.Mul(dop)
 		case opconst.TypeAux:
 			aux = encOp.Data[opconst.TypeAuxLen:]
 			auxKey = encOp.Key
diff --git a/internal/opconst/ops.go b/internal/opconst/ops.go
index aca341f9..62365b8f 100644
--- a/internal/opconst/ops.go
+++ b/internal/opconst/ops.go
@@ -31,7 +31,7 @@ const (
 const (
 	TypeMacroLen        = 1 + 4 + 4
 	TypeCallLen         = 1 + 4 + 4
-	TypeTransformLen    = 1 + 4*2
+	TypeTransformLen    = 1 + 4*6
 	TypeLayerLen        = 1
 	TypeRedrawLen       = 1 + 8
 	TypeImageLen        = 1 + 4*4
diff --git a/internal/ops/ops.go b/internal/ops/ops.go
index 80994e17..721f4628 100644
--- a/internal/ops/ops.go
+++ b/internal/ops/ops.go
@@ -8,7 +8,6 @@ import (
 
 	"gioui.org/f32"
 	"gioui.org/internal/opconst"
-	"gioui.org/op"
 )
 
 const QuadSize = 4 * 2 * 3
@@ -38,13 +37,23 @@ func DecodeQuad(d []byte) (q Quad) {
 	return
 }
 
-func DecodeTransformOp(d []byte) op.TransformOp {
-	bo := binary.LittleEndian
+func DecodeTransform(d []byte) (t f32.Affine2D) {
 	if opconst.OpType(d[0]) != opconst.TypeTransform {
 		panic("invalid op")
 	}
-	return op.TransformOp{}.Offset(f32.Point{
-		X: math.Float32frombits(bo.Uint32(d[1:])),
-		Y: math.Float32frombits(bo.Uint32(d[5:])),
-	})
+	if len(d) < 1+6*4 {
+		panic("too short buffer")
+	}
+	return decodeAffine2D(d[1:])
+}
+
+func decodeAffine2D(data []byte) f32.Affine2D {
+	bo := binary.LittleEndian
+	a := math.Float32frombits(bo.Uint32(data))
+	b := math.Float32frombits(bo.Uint32(data[4*1:]))
+	c := math.Float32frombits(bo.Uint32(data[4*2:]))
+	d := math.Float32frombits(bo.Uint32(data[4*3:]))
+	e := math.Float32frombits(bo.Uint32(data[4*4:]))
+	f := math.Float32frombits(bo.Uint32(data[4*5:]))
+	return f32.NewAffine2D(a, b, c, d, e, f)
 }
diff --git a/io/router/pointer.go b/io/router/pointer.go
index ec3bb397..0d0977f6 100644
--- a/io/router/pointer.go
+++ b/io/router/pointer.go
@@ -56,7 +56,7 @@ type areaOp struct {
 }
 
 type areaNode struct {
-	trans op.TransformOp
+	trans f32.Affine2D
 	next  int
 	area  areaOp
 }
@@ -68,7 +68,7 @@ const (
 	areaEllipse
 )
 
-func (q *pointerQueue) collectHandlers(r *ops.Reader, events *handlerEvents, t op.TransformOp, area, node int, pass bool) {
+func (q *pointerQueue) collectHandlers(r *ops.Reader, events *handlerEvents, t f32.Affine2D, area, node int, pass bool) {
 	for encOp, ok := r.Decode(); ok; encOp, ok = r.Decode() {
 		switch opconst.OpType(encOp.Data[0]) {
 		case opconst.TypePush:
@@ -90,8 +90,8 @@ func (q *pointerQueue) collectHandlers(r *ops.Reader, events *handlerEvents, t o
 			})
 			node = len(q.hitTree) - 1
 		case opconst.TypeTransform:
-			dop := ops.DecodeTransformOp(encOp.Data)
-			t = t.Multiply(op.TransformOp(dop))
+			dop := ops.DecodeTransform(encOp.Data)
+			t = t.Mul(dop)
 		case opconst.TypePointerInput:
 			op := decodePointerInputOp(encOp.Data, encOp.Refs)
 			q.hitTree = append(q.hitTree, hitNode{
@@ -175,7 +175,7 @@ func (q *pointerQueue) Frame(root *op.Ops, events *handlerEvents) {
 	q.hitTree = q.hitTree[:0]
 	q.areas = q.areas[:0]
 	q.reader.Reset(root)
-	q.collectHandlers(&q.reader, events, op.TransformOp{}, -1, -1, false)
+	q.collectHandlers(&q.reader, events, f32.Affine2D{}, -1, -1, false)
 	for k, h := range q.handlers {
 		if !h.active {
 			q.dropHandlers(events, k)
diff --git a/op/op.go b/op/op.go
index 82f1457a..3aafbc33 100644
--- a/op/op.go
+++ b/op/op.go
@@ -290,8 +290,13 @@ func (t TransformOp) Add(o *Ops) {
 	data := o.Write(opconst.TypeTransformLen)
 	data[0] = byte(opconst.TypeTransform)
 	bo := binary.LittleEndian
-	bo.PutUint32(data[1:], math.Float32bits(t.offset.X))
-	bo.PutUint32(data[5:], math.Float32bits(t.offset.Y))
+	// write it out as an affine matrix although we only support offset yet
+	bo.PutUint32(data[1:], math.Float32bits(1.0))
+	bo.PutUint32(data[1+4*1:], math.Float32bits(0))
+	bo.PutUint32(data[1+4*2:], math.Float32bits(t.offset.X))
+	bo.PutUint32(data[1+4*3:], math.Float32bits(0))
+	bo.PutUint32(data[1+4*4:], math.Float32bits(1))
+	bo.PutUint32(data[1+4*5:], math.Float32bits(t.offset.Y))
 }
 
 func (s *stack) push() stackID {