gpu: fold buffer clearing into framebuffer bind

In Metal, clearing a framebuffer is most efficiently done during bind.
Modify our driver accordingly.

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

gpu/compute.go

@@ -582,12 +582,6 @@ func (g *compute) Frame(target RenderTarget) error {
 		t.blit = t.t.newTimer()
 	}
 
-	g.ctx.BindFramebuffer(defFBO)
-	if g.collector.clear {
-		g.collector.clear = false
-		g.ctx.Clear(g.collector.clearColor.Float32())
-	}
-
 	t.compact.begin()
 	if err := g.compactLayers(); err != nil {
 		return err
@@ -598,7 +592,14 @@ func (g *compute) Frame(target RenderTarget) error {
 		return err
 	}
 	t.render.end()
-	g.ctx.BindFramebuffer(defFBO)
+	var d driver.LoadDesc
+	if g.collector.clear {
+		g.collector.clear = false
+		d.Action = driver.LoadActionClear
+		c := &d.ClearColor
+		c.R, c.G, c.B, c.A = g.collector.clearColor.Float32()
+	}
+	g.ctx.BindFramebuffer(defFBO, d)
 	t.blit.begin()
 	g.blitLayers(viewport)
 	t.blit.end()
@@ -691,9 +692,6 @@ func (g *compute) compactLayers() error {
 			dr := image.Rectangle{Min: l.newPlace.pos, Max: l.newPlace.pos.Add(sz)}
 			g.ctx.BlitFramebuffer(dst, src, sr, dr)
 			l.place.atlas.layers--
-			if l.place.atlas.layers == 0 {
-				l.place.atlas.fbo.Invalidate()
-			}
 			layers[i].place = l.newPlace
 		}
 		layers = layers[end:]
@@ -926,11 +924,12 @@ restart:
 	m.buffer.ensureCapacity(false, g.ctx, driver.BufferBindingVertices, n)
 	m.buffer.buffer.Upload(vertexData)
 	g.ctx.BindTexture(0, g.images.tex)
-	g.ctx.BindFramebuffer(m.fbo)
+	var d driver.LoadDesc
-	g.ctx.Viewport(0, 0, texSize, texSize)
 	if reclaimed {
-		g.ctx.Clear(0, 0, 0, 0)
+		d.Action = driver.LoadActionClear
 	}
+	g.ctx.BindFramebuffer(m.fbo, d)
+	g.ctx.Viewport(0, 0, texSize, texSize)
 	g.ctx.BindPipeline(m.pipeline)
 	g.ctx.BindVertexBuffer(m.buffer.buffer, int(unsafe.Sizeof(m.quads[0])), 0)
 	g.ctx.DrawArrays(driver.DrawModeTriangles, 0, len(m.quads))

gpu/gpu.go

@@ -445,11 +445,6 @@ func (g *gpu) Frame(target RenderTarget) error {
 	for _, img := range g.drawOps.imageOps {
 		expandPathOp(img.path, img.clip)
 	}
-	g.ctx.BindFramebuffer(defFBO)
-	if g.drawOps.clear {
-		g.drawOps.clear = false
-		g.ctx.Clear(g.drawOps.clearColor.Float32())
-	}
 	g.ctx.Viewport(0, 0, viewport.X, viewport.Y)
 	g.stencilTimer.begin()
 	g.renderer.packStencils(&g.drawOps.pathOps)
@@ -458,12 +453,17 @@ func (g *gpu) Frame(target RenderTarget) error {
 	g.renderer.intersect(g.drawOps.imageOps)
 	g.stencilTimer.end()
 	g.coverTimer.begin()
-	g.ctx.BindFramebuffer(defFBO)
+	var d driver.LoadDesc
+	if g.drawOps.clear {
+		g.drawOps.clear = false
+		d.Action = driver.LoadActionClear
+		c := &d.ClearColor
+		c.R, c.G, c.B, c.A = g.drawOps.clearColor.Float32()
+	}
+	g.ctx.BindFramebuffer(defFBO, d)
 	g.ctx.Viewport(0, 0, viewport.X, viewport.Y)
 	g.renderer.drawOps(g.cache, g.drawOps.imageOps)
-	g.renderer.pather.stenciler.invalidateFBO()
 	g.coverTimer.end()
-	g.ctx.BindFramebuffer(defFBO)
 	g.cleanupTimer.begin()
 	g.cache.frame()
 	g.drawOps.pathCache.frame()
@@ -689,8 +689,7 @@ func (r *renderer) stencilClips(pathCache *opCache, ops []*pathOp) {
 		if fbo != p.place.Idx {
 			fbo = p.place.Idx
 			f := r.pather.stenciler.cover(fbo)
-			r.ctx.BindFramebuffer(f.fbo)
+			r.ctx.BindFramebuffer(f.fbo, driver.LoadDesc{Action: driver.LoadActionClear})
-			r.ctx.Clear(0.0, 0.0, 0.0, 0.0)
 		}
 		v, _ := pathCache.get(p.pathKey)
 		r.pather.stencilPath(p.clip, p.off, p.place.Pos, v.data)
@@ -711,8 +710,9 @@ func (r *renderer) intersect(ops []imageOp) {
 		if fbo != img.place.Idx {
 			fbo = img.place.Idx
 			f := r.pather.stenciler.intersections.fbos[fbo]
-			r.ctx.BindFramebuffer(f.fbo)
+			d := driver.LoadDesc{Action: driver.LoadActionClear}
-			r.ctx.Clear(1.0, 0.0, 0.0, 0.0)
+			d.ClearColor.R = 1.0
+			r.ctx.BindFramebuffer(f.fbo, d)
 		}
 		r.ctx.Viewport(img.place.Pos.X, img.place.Pos.Y, img.clip.Dx(), img.clip.Dy())
 		r.intersectPath(img.path, img.clip)

gpu/headless/driver_test.go

@@ -141,10 +141,12 @@ func TestFramebuffers(t *testing.T) {
 		col2 = color.NRGBA{R: 0xfe, G: 0xba, B: 0xbe, A: 0xca}
 	)
 	fcol1, fcol2 := f32color.LinearFromSRGB(col1), f32color.LinearFromSRGB(col2)
-	b.BindFramebuffer(fbo1)
+	d := driver.LoadDesc{Action: driver.LoadActionClear}
-	b.Clear(fcol1.Float32())
+	c := &d.ClearColor
-	b.BindFramebuffer(fbo2)
+	c.R, c.G, c.B, c.A = fcol1.Float32()
-	b.Clear(fcol2.Float32())
+	b.BindFramebuffer(fbo1, d)
+	c.R, c.G, c.B, c.A = fcol2.Float32()
+	b.BindFramebuffer(fbo2, d)
 	img := screenshot(t, b, fbo1, sz)
 	if got := img.RGBAAt(0, 0); got != f32color.NRGBAToRGBA(col1) {
 		t.Errorf("got color %v, expected %v", got, f32color.NRGBAToRGBA(col1))
@@ -157,11 +159,13 @@ func TestFramebuffers(t *testing.T) {
 
 func setupFBO(t *testing.T, b driver.Device, size image.Point) driver.Framebuffer {
 	fbo := newFBO(t, b, size)
-	b.BindFramebuffer(fbo)
 	// ClearColor accepts linear RGBA colors, while 8-bit colors
 	// are in the sRGB color space.
 	col := f32color.LinearFromSRGB(clearCol)
-	b.Clear(col.Float32())
+	d := driver.LoadDesc{Action: driver.LoadActionClear}
+	c := &d.ClearColor
+	c.R, c.G, c.B, c.A = col.Float32()
+	b.BindFramebuffer(fbo, d)
 	b.Viewport(0, 0, size.X, size.Y)
 	return fbo
 }

gpu/internal/d3d11/d3d11_windows.go

@@ -432,11 +432,6 @@ func (b *Backend) NewFragmentShader(src shader.Sources) (driver.FragmentShader,
 	return &FragmentShader{b, fs}, nil
 }
 
-func (b *Backend) Clear(colr, colg, colb, cola float32) {
-	b.clearColor = [4]float32{colr, colg, colb, cola}
-	b.ctx.ClearRenderTargetView(b.fbo.renderTarget, &b.clearColor)
-}
-
 func (b *Backend) Viewport(x, y, width, height int) {
 	b.viewport = d3d11.VIEWPORT{
 		TopLeftX: float32(x),
@@ -643,9 +638,14 @@ func (f *Framebuffer) ReadPixels(src image.Rectangle, pixels []byte) error {
 	return nil
 }
 
-func (b *Backend) BindFramebuffer(fbo driver.Framebuffer) {
+func (b *Backend) BindFramebuffer(fbo driver.Framebuffer, d driver.LoadDesc) {
 	b.fbo = fbo.(*Framebuffer)
 	b.ctx.OMSetRenderTargets(b.fbo.renderTarget, nil)
+	if d.Action == driver.LoadActionClear {
+		c := d.ClearColor
+		b.clearColor = [4]float32{c.R, c.G, c.B, c.A}
+		b.ctx.ClearRenderTargetView(b.fbo.renderTarget, &b.clearColor)
+	}
 }
 
 func (f *Framebuffer) Invalidate() {

gpu/internal/driver/driver.go

@@ -30,14 +30,13 @@ type Device interface {
 	NewFragmentShader(src shader.Sources) (FragmentShader, error)
 	NewPipeline(desc PipelineDesc) (Pipeline, error)
 
-	Clear(r, g, b, a float32)
 	Viewport(x, y, width, height int)
 	DrawArrays(mode DrawMode, off, count int)
 	DrawElements(mode DrawMode, off, count int)
 
 	BindProgram(p Program)
 	BindPipeline(p Pipeline)
-	BindFramebuffer(f Framebuffer)
+	BindFramebuffer(f Framebuffer, a LoadDesc)
 	BindTexture(unit int, t Texture)
 	BindVertexBuffer(b Buffer, stride, offset int)
 	BindIndexBuffer(b Buffer)
@@ -53,6 +52,16 @@ type Device interface {
 	Release()
 }
 
+type LoadDesc struct {
+	Action     LoadAction
+	ClearColor struct {
+		R float32
+		G float32
+		B float32
+		A float32
+	}
+}
+
 type Pipeline interface {
 	Release()
 }
@@ -89,6 +98,8 @@ type TextureFormat uint8
 
 type BufferBinding uint8
 
+type LoadAction uint8
+
 type Features uint
 
 type Caps struct {
@@ -119,7 +130,6 @@ type Buffer interface {
 
 type Framebuffer interface {
 	RenderTarget
-	Invalidate()
 	Release()
 	ReadPixels(src image.Rectangle, pixels []byte) error
 }
@@ -182,6 +192,12 @@ const (
 	BlendFactorDstColor
 )
 
+const (
+	LoadActionKeep LoadAction = iota
+	LoadActionClear
+	LoadActionInvalidate
+)
+
 var ErrContentLost = errors.New("buffer content lost")
 
 func (f Features) Has(feats Features) bool {

gpu/internal/opengl/opengl.go

@@ -268,7 +268,7 @@ func (b *Backend) BeginFrame(target driver.RenderTarget, clear bool, viewport im
 	}
 	b.glstate.bindFramebuffer(b.funcs, gl.FRAMEBUFFER, renderFBO)
 	if b.sRGBFBO != nil && !clear {
-		b.Clear(0, 0, 0, 0)
+		b.clearOutput(0, 0, 0, 0)
 	}
 	return &framebuffer{backend: b, obj: renderFBO, foreign: true}
 }
@@ -652,7 +652,7 @@ func (b *Backend) NewFramebuffer(tex driver.Texture) (driver.Framebuffer, error)
 	gltex := tex.(*texture)
 	fb := b.funcs.CreateFramebuffer()
 	fbo := &framebuffer{backend: b, obj: fb}
-	b.BindFramebuffer(fbo)
+	b.BindFramebuffer(fbo, driver.LoadDesc{})
 	if err := glErr(b.funcs); err != nil {
 		fbo.Release()
 		return nil, err
@@ -849,7 +849,7 @@ func (b *Backend) Viewport(x, y, width, height int) {
 	b.glstate.setViewport(b.funcs, x, y, width, height)
 }
 
-func (b *Backend) Clear(colR, colG, colB, colA float32) {
+func (b *Backend) clearOutput(colR, colG, colB, colA float32) {
 	b.glstate.setClearColor(b.funcs, colR, colG, colB, colA)
 	b.funcs.Clear(gl.COLOR_BUFFER_BIT)
 }
@@ -1179,7 +1179,7 @@ func (b *Backend) BlitFramebuffer(dst, src driver.Framebuffer, srect, drect imag
 
 func (f *framebuffer) ReadPixels(src image.Rectangle, pixels []byte) error {
 	glErr(f.backend.funcs)
-	f.backend.BindFramebuffer(f)
+	f.backend.BindFramebuffer(f, driver.LoadDesc{})
 	if len(pixels) < src.Dx()*src.Dy()*4 {
 		return errors.New("unexpected RGBA size")
 	}
@@ -1195,13 +1195,15 @@ func (b *Backend) BindPipeline(pl driver.Pipeline) {
 	b.BlendFunc(p.blend.SrcFactor, p.blend.DstFactor)
 }
 
-func (b *Backend) BindFramebuffer(fbo driver.Framebuffer) {
+func (b *Backend) BindFramebuffer(fbo driver.Framebuffer, desc driver.LoadDesc) {
 	b.glstate.bindFramebuffer(b.funcs, gl.FRAMEBUFFER, fbo.(*framebuffer).obj)
+	switch desc.Action {
+	case driver.LoadActionClear:
+		c := desc.ClearColor
+		b.clearOutput(c.R, c.G, c.B, c.A)
+	case driver.LoadActionInvalidate:
+		b.funcs.InvalidateFramebuffer(gl.FRAMEBUFFER, gl.COLOR_ATTACHMENT0)
 	}
-
-func (f *framebuffer) Invalidate() {
-	f.backend.BindFramebuffer(f)
-	f.backend.funcs.InvalidateFramebuffer(gl.FRAMEBUFFER, gl.COLOR_ATTACHMENT0)
 }
 
 func (f *framebuffer) Release() {

gpu/path.go

@@ -284,12 +284,6 @@ func (s *fboSet) resize(ctx driver.Device, sizes []image.Point) {
 	s.delete(ctx, len(sizes))
 }
 
-func (s *fboSet) invalidate(ctx driver.Device) {
-	for _, f := range s.fbos {
-		f.fbo.Invalidate()
-	}
-}
-
 func (s *fboSet) delete(ctx driver.Device, idx int) {
 	for i := idx; i < len(s.fbos); i++ {
 		f := s.fbos[i]
@@ -348,11 +342,6 @@ func (s *stenciler) beginIntersect(sizes []image.Point) {
 	s.ctx.BindPipeline(s.ipipeline.pipeline.pipeline)
 }
 
-func (s *stenciler) invalidateFBO() {
-	s.intersections.invalidate(s.ctx)
-	s.fbos.invalidate(s.ctx)
-}
-
 func (s *stenciler) cover(idx int) stencilFBO {
 	return s.fbos.fbos[idx]
 }