gpu: introduce render passes

Modern GPU API such as Metal and Vulkan use explicit render passes
and command buffers for recording rendering commands. They don't have
global state; each render pass starts with a clean set of bound
textures, pipeline etc.

Change our GPU abstraction to better match newer API and modify our two
renderers to explicitly describe their render passes.

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

gpu/compute.go

@@ -623,17 +623,18 @@ func (g *compute) frame(target RenderTarget) error {
 		return err
 	}
 	t.render.end()
-	var d driver.LoadDesc
+	d := driver.LoadDesc{
+		ClearColor: g.collector.clearColor,
+	}
 	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)
+	g.ctx.BeginRenderPass(defFBO, d)
 	t.blit.begin()
 	g.blitLayers(viewport)
 	t.blit.end()
+	g.ctx.EndRenderPass()
 	t.compact.begin()
 	if err := g.compactAllocs(); err != nil {
 		return err
@@ -993,18 +994,18 @@ func (g *compute) renderMaterials() error {
 		m.vert.uniforms.scale = [2]float32{2, -2}
 		m.vert.uniforms.pos = [2]float32{-1, +1}
 		m.vert.buf.Upload(byteslice.Struct(m.vert.uniforms))
-		g.ctx.BindVertexUniforms(m.vert.buf)
-		g.ctx.BindFragmentUniforms(m.frag.buf)
 		vertexData := byteslice.Slice(m.quads)
 		n := pow2Ceil(len(vertexData))
 		m.buffer.ensureCapacity(false, g.ctx, driver.BufferBindingVertices, n)
 		m.buffer.buffer.Upload(vertexData)
-		g.ctx.BindTexture(0, imgAtlas.image)
 		var d driver.LoadDesc
 		if !realized {
 			d.Action = driver.LoadActionClear
 		}
-		g.ctx.BindFramebuffer(atlas.fbo, d)
+		g.ctx.BeginRenderPass(atlas.fbo, d)
+		g.ctx.BindVertexUniforms(m.vert.buf)
+		g.ctx.BindFragmentUniforms(m.frag.buf)
+		g.ctx.BindTexture(0, imgAtlas.image)
 		g.ctx.BindPipeline(m.pipeline)
 		g.ctx.BindVertexBuffer(m.buffer.buffer, 0)
 		newAllocs := atlas.allocs[allocStart:]
@@ -1013,6 +1014,7 @@ func (g *compute) renderMaterials() error {
 			g.ctx.Viewport(a.rect.Min.X, a.rect.Min.Y, sz.X, sz.Y)
 			g.ctx.DrawArrays(driver.DrawModeTriangles, i*6, 6)
 		}
+		g.ctx.EndRenderPass()
 		if !g.useCPU {
 			continue
 		}
@@ -1263,18 +1265,6 @@ func (g *compute) render(images *textureAtlas, dst driver.Texture, cpuDst cpu.Im
 		}
 	}
 
-	if !g.useCPU {
-		g.ctx.BindImageTexture(kernel4OutputUnit, dst, driver.AccessWrite, driver.TextureFormatRGBA8)
-		if images != nil {
-			g.ctx.BindImageTexture(kernel4AtlasUnit, images.image, driver.AccessRead, driver.TextureFormatRGBA8)
-		}
-	} else {
-		*g.output.descriptors.Binding2() = cpuDst
-		if images != nil {
-			*g.output.descriptors.Binding3() = images.cpuImage
-		}
-	}
-
 	for {
 		*g.memHeader = memoryHeader{
 			mem_offset: alloc,
@@ -1282,6 +1272,19 @@ func (g *compute) render(images *textureAtlas, dst driver.Texture, cpuDst cpu.Im
 		g.buffers.memory.upload(byteslice.Struct(g.memHeader))
 		g.buffers.state.upload(g.zeros(clearSize))
 
+		if !g.useCPU {
+			g.ctx.BeginCompute()
+			g.ctx.BindImageTexture(kernel4OutputUnit, dst, driver.AccessWrite, driver.TextureFormatRGBA8)
+			if images != nil {
+				g.ctx.BindImageTexture(kernel4AtlasUnit, images.image, driver.AccessRead, driver.TextureFormatRGBA8)
+			}
+		} else {
+			*g.output.descriptors.Binding2() = cpuDst
+			if images != nil {
+				*g.output.descriptors.Binding3() = images.cpuImage
+			}
+		}
+
 		g.bindBuffers()
 		g.memoryBarrier()
 		g.dispatch(g.programs.elements, numPartitions, 1, 1)
@@ -1298,7 +1301,9 @@ func (g *compute) render(images *textureAtlas, dst driver.Texture, cpuDst cpu.Im
 		g.memoryBarrier()
 		g.dispatch(g.programs.kernel4, tileDims.X, tileDims.Y, 1)
 		g.memoryBarrier()
-		if g.useCPU {
+		if !g.useCPU {
+			g.ctx.EndCompute()
+		} else {
 			g.dispatcher.Sync()
 		}
 

gpu/gpu.go

@@ -424,7 +424,6 @@ func (g *gpu) frame(target RenderTarget) error {
 	for _, img := range g.drawOps.imageOps {
 		expandPathOp(img.path, img.clip)
 	}
-	g.ctx.Viewport(0, 0, viewport.X, viewport.Y)
 	g.stencilTimer.begin()
 	g.renderer.packStencils(&g.drawOps.pathOps)
 	g.renderer.stencilClips(g.drawOps.pathCache, g.drawOps.pathOps)
@@ -432,17 +431,19 @@ func (g *gpu) frame(target RenderTarget) error {
 	g.renderer.intersect(g.drawOps.imageOps)
 	g.stencilTimer.end()
 	g.coverTimer.begin()
-	var d driver.LoadDesc
+	g.renderer.uploadImages(g.cache, g.drawOps.imageOps)
+	d := driver.LoadDesc{
+		ClearColor: g.drawOps.clearColor,
+	}
 	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.BeginRenderPass(defFBO, d)
 	g.ctx.Viewport(0, 0, viewport.X, viewport.Y)
 	g.renderer.drawOps(g.cache, g.drawOps.imageOps)
 	g.coverTimer.end()
+	g.ctx.EndRenderPass()
 	g.cleanupTimer.begin()
 	g.cache.frame()
 	g.drawOps.pathCache.frame()
@@ -669,13 +670,21 @@ func (r *renderer) stencilClips(pathCache *opCache, ops []*pathOp) {
 	r.pather.begin(r.packer.sizes)
 	for _, p := range ops {
 		if fbo != p.place.Idx {
+			if fbo != -1 {
+				r.ctx.EndRenderPass()
+			}
 			fbo = p.place.Idx
 			f := r.pather.stenciler.cover(fbo)
-			r.ctx.BindFramebuffer(f.fbo, driver.LoadDesc{Action: driver.LoadActionClear})
+			r.ctx.BeginRenderPass(f.fbo, driver.LoadDesc{Action: driver.LoadActionClear})
+			r.ctx.BindPipeline(r.pather.stenciler.pipeline.pipeline.pipeline)
+			r.ctx.BindIndexBuffer(r.pather.stenciler.indexBuf)
 		}
 		v, _ := pathCache.get(p.pathKey)
 		r.pather.stencilPath(p.clip, p.off, p.place.Pos, v.data)
 	}
+	if fbo != -1 {
+		r.ctx.EndRenderPass()
+	}
 }
 
 func (r *renderer) intersect(ops []imageOp) {
@@ -684,21 +693,28 @@ func (r *renderer) intersect(ops []imageOp) {
 	}
 	fbo := -1
 	r.pather.stenciler.beginIntersect(r.intersections.sizes)
-	r.ctx.BindVertexBuffer(r.blitter.quadVerts, 0)
 	for _, img := range ops {
 		if img.clipType != clipTypeIntersection {
 			continue
 		}
 		if fbo != img.place.Idx {
+			if fbo != -1 {
+				r.ctx.EndRenderPass()
+			}
 			fbo = img.place.Idx
 			f := r.pather.stenciler.intersections.fbos[fbo]
 			d := driver.LoadDesc{Action: driver.LoadActionClear}
 			d.ClearColor.R = 1.0
-			r.ctx.BindFramebuffer(f.fbo, d)
+			r.ctx.BeginRenderPass(f.fbo, d)
+			r.ctx.BindPipeline(r.pather.stenciler.ipipeline.pipeline.pipeline)
+			r.ctx.BindVertexBuffer(r.blitter.quadVerts, 0)
 		}
 		r.ctx.Viewport(img.place.Pos.X, img.place.Pos.Y, img.clip.Dx(), img.clip.Dy())
 		r.intersectPath(img.path, img.clip)
 	}
+	if fbo != -1 {
+		r.ctx.EndRenderPass()
+	}
 }
 
 func (r *renderer) intersectPath(p *pathOp, clip image.Rectangle) {
@@ -1077,8 +1093,16 @@ func (d *drawState) materialFor(rect f32.Rectangle, off f32.Point, partTrans f32
 	return m
 }
 
+func (r *renderer) uploadImages(cache *resourceCache, ops []imageOp) {
+	for _, img := range ops {
+		m := img.material
+		if m.material == materialTexture {
+			r.texHandle(cache, m.data)
+		}
+	}
+}
+
 func (r *renderer) drawOps(cache *resourceCache, ops []imageOp) {
-	r.ctx.BindVertexBuffer(r.blitter.quadVerts, 0)
 	var coverTex driver.Texture
 	for _, img := range ops {
 		m := img.material
@@ -1092,6 +1116,9 @@ func (r *renderer) drawOps(cache *resourceCache, ops []imageOp) {
 		var fbo stencilFBO
 		switch img.clipType {
 		case clipTypeNone:
+			p := r.blitter.pipelines[m.material]
+			r.ctx.BindPipeline(p.pipeline)
+			r.ctx.BindVertexBuffer(r.blitter.quadVerts, 0)
 			r.blitter.blit(m.material, m.color, m.color1, m.color2, scale, off, m.uvTrans)
 			continue
 		case clipTypePath:
@@ -1108,6 +1135,9 @@ func (r *renderer) drawOps(cache *resourceCache, ops []imageOp) {
 			Max: img.place.Pos.Add(drc.Size()),
 		}
 		coverScale, coverOff := texSpaceTransform(layout.FRect(uv), fbo.size)
+		p := r.pather.coverer.pipelines[m.material]
+		r.ctx.BindPipeline(p.pipeline)
+		r.ctx.BindVertexBuffer(r.blitter.quadVerts, 0)
 		r.pather.cover(m.material, m.color, m.color1, m.color2, scale, off, m.uvTrans, coverScale, coverOff)
 	}
 }
@@ -1165,12 +1195,12 @@ func (u *uniformBuffer) Release() {
 
 func (p *pipeline) UploadUniforms(ctx driver.Device) {
 	if p.vertUniforms != nil {
-		ctx.BindVertexUniforms(p.vertUniforms.buf)
 		p.vertUniforms.Upload()
+		ctx.BindVertexUniforms(p.vertUniforms.buf)
 	}
 	if p.fragUniforms != nil {
-		ctx.BindFragmentUniforms(p.fragUniforms.buf)
 		p.fragUniforms.Upload()
+		ctx.BindFragmentUniforms(p.fragUniforms.buf)
 	}
 }
 

gpu/headless/driver_test.go

@@ -27,7 +27,15 @@ var clearColExpect = f32color.NRGBAToRGBA(clearCol)
 func TestFramebufferClear(t *testing.T) {
 	b := newDriver(t)
 	sz := image.Point{X: 800, Y: 600}
-	fbo := setupFBO(t, b, sz)
+	fbo := newFBO(t, b, sz)
+	d := driver.LoadDesc{
+		// ClearColor accepts linear RGBA colors, while 8-bit colors
+		// are in the sRGB color space.
+		ClearColor: f32color.LinearFromSRGB(clearCol),
+		Action:     driver.LoadActionClear,
+	}
+	b.BeginRenderPass(fbo, d)
+	b.EndRenderPass()
 	img := screenshot(t, b, fbo, sz)
 	if got := img.RGBAAt(0, 0); got != clearColExpect {
 		t.Errorf("got color %v, expected %v", got, clearColExpect)
@@ -37,13 +45,13 @@ func TestFramebufferClear(t *testing.T) {
 func TestSimpleShader(t *testing.T) {
 	b := newDriver(t)
 	sz := image.Point{X: 800, Y: 600}
-	fbo := setupFBO(t, b, sz)
 	vsh, fsh, err := newShaders(b, gio.Shader_simple_vert, gio.Shader_simple_frag)
 	if err != nil {
 		t.Fatal(err)
 	}
 	defer vsh.Release()
 	defer fsh.Release()
+	fbo := newFBO(t, b, sz)
 	p, err := b.NewPipeline(driver.PipelineDesc{
 		VertexShader:   vsh,
 		FragmentShader: fsh,
@@ -53,8 +61,15 @@ func TestSimpleShader(t *testing.T) {
 		t.Fatal(err)
 	}
 	defer p.Release()
+	d := driver.LoadDesc{
+		ClearColor: f32color.LinearFromSRGB(clearCol),
+		Action:     driver.LoadActionClear,
+	}
+	b.BeginRenderPass(fbo, d)
+	b.Viewport(0, 0, sz.X, sz.Y)
 	b.BindPipeline(p)
 	b.DrawArrays(driver.DrawModeTriangles, 0, 3)
+	b.EndRenderPass()
 	img := screenshot(t, b, fbo, sz)
 	if got := img.RGBAAt(0, 0); got != clearColExpect {
 		t.Errorf("got color %v, expected %v", got, clearColExpect)
@@ -70,7 +85,6 @@ func TestSimpleShader(t *testing.T) {
 func TestInputShader(t *testing.T) {
 	b := newDriver(t)
 	sz := image.Point{X: 800, Y: 600}
-	fbo := setupFBO(t, b, sz)
 	vsh, fsh, err := newShaders(b, gio.Shader_input_vert, gio.Shader_simple_frag)
 	if err != nil {
 		t.Fatal(err)
@@ -87,6 +101,7 @@ func TestInputShader(t *testing.T) {
 		},
 		Stride: 4 * 4,
 	}
+	fbo := newFBO(t, b, sz)
 	pipe, err := b.NewPipeline(driver.PipelineDesc{
 		VertexShader:   vsh,
 		FragmentShader: fsh,
@@ -97,7 +112,6 @@ func TestInputShader(t *testing.T) {
 		t.Fatal(err)
 	}
 	defer pipe.Release()
-	b.BindPipeline(pipe)
 	buf, err := b.NewImmutableBuffer(driver.BufferBindingVertices,
 		byteslice.Slice([]float32{
 			0, .5, .5, 1,
@@ -109,8 +123,16 @@ func TestInputShader(t *testing.T) {
 		t.Fatal(err)
 	}
 	defer buf.Release()
+	d := driver.LoadDesc{
+		ClearColor: f32color.LinearFromSRGB(clearCol),
+		Action:     driver.LoadActionClear,
+	}
+	b.BeginRenderPass(fbo, d)
+	b.Viewport(0, 0, sz.X, sz.Y)
+	b.BindPipeline(pipe)
 	b.BindVertexBuffer(buf, 0)
 	b.DrawArrays(driver.DrawModeTriangles, 0, 3)
+	b.EndRenderPass()
 	img := screenshot(t, b, fbo, sz)
 	if got := img.RGBAAt(0, 0); got != clearColExpect {
 		t.Errorf("got color %v, expected %v", got, clearColExpect)
@@ -137,19 +159,20 @@ func newShaders(ctx driver.Device, vsrc, fsrc shader.Sources) (vert driver.Verte
 func TestFramebuffers(t *testing.T) {
 	b := newDriver(t)
 	sz := image.Point{X: 800, Y: 600}
-	fbo1 := newFBO(t, b, sz)
-	fbo2 := newFBO(t, b, sz)
 	var (
 		col1 = color.NRGBA{R: 0xac, G: 0xbd, B: 0xef, A: 0xde}
 		col2 = color.NRGBA{R: 0xfe, G: 0xba, B: 0xbe, A: 0xca}
 	)
+	fbo1 := newFBO(t, b, sz)
+	fbo2 := newFBO(t, b, sz)
 	fcol1, fcol2 := f32color.LinearFromSRGB(col1), f32color.LinearFromSRGB(col2)
 	d := driver.LoadDesc{Action: driver.LoadActionClear}
-	c := &d.ClearColor
-	c.R, c.G, c.B, c.A = fcol1.Float32()
-	b.BindFramebuffer(fbo1, d)
-	c.R, c.G, c.B, c.A = fcol2.Float32()
-	b.BindFramebuffer(fbo2, d)
+	d.ClearColor = fcol1
+	b.BeginRenderPass(fbo1, d)
+	b.EndRenderPass()
+	d.ClearColor = fcol2
+	b.BeginRenderPass(fbo2, d)
+	b.EndRenderPass()
 	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))
@@ -160,19 +183,6 @@ func TestFramebuffers(t *testing.T) {
 	}
 }
 
-func setupFBO(t *testing.T, b driver.Device, size image.Point) driver.Framebuffer {
-	fbo := newFBO(t, b, size)
-	// ClearColor accepts linear RGBA colors, while 8-bit colors
-	// are in the sRGB color space.
-	col := f32color.LinearFromSRGB(clearCol)
-	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
-}
-
 func newFBO(t *testing.T, b driver.Device, size image.Point) driver.Framebuffer {
 	fboTex, err := b.NewTexture(
 		driver.TextureFormatSRGBA,

gpu/headless/headless.go

@@ -47,7 +47,6 @@ func NewWindow(width, height int) (*Window, error) {
 		if err != nil {
 			return err
 		}
-		dev.Viewport(0, 0, width, height)
 		fboTex, err := dev.NewTexture(
 			driver.TextureFormatSRGBA,
 			width, height,

gpu/internal/d3d11/d3d11_windows.go

@@ -789,7 +789,13 @@ func (f *Framebuffer) ReadPixels(src image.Rectangle, pixels []byte, stride int)
 	return nil
 }
 
-func (b *Backend) BindFramebuffer(fbo driver.Framebuffer, d driver.LoadDesc) {
+func (b *Backend) BeginCompute() {
+}
+
+func (b *Backend) EndCompute() {
+}
+
+func (b *Backend) BeginRenderPass(fbo driver.Framebuffer, d driver.LoadDesc) {
 	b.fbo = fbo.(*Framebuffer)
 	b.ctx.OMSetRenderTargets(b.fbo.renderTarget, nil)
 	if d.Action == driver.LoadActionClear {
@@ -799,6 +805,9 @@ func (b *Backend) BindFramebuffer(fbo driver.Framebuffer, d driver.LoadDesc) {
 	}
 }
 
+func (b *Backend) EndRenderPass() {
+}
+
 func (f *Framebuffer) Release() {
 	if f.foreign {
 		panic("framebuffer not created by NewFramebuffer")

gpu/internal/driver/driver.go

@@ -7,6 +7,7 @@ import (
 	"image"
 	"time"
 
+	"gioui.org/internal/f32color"
 	"gioui.org/shader"
 )
 
@@ -34,9 +35,10 @@ type Device interface {
 	DrawArrays(mode DrawMode, off, count int)
 	DrawElements(mode DrawMode, off, count int)
 
+	BeginRenderPass(f Framebuffer, desc LoadDesc)
+	EndRenderPass()
 	BindProgram(p Program)
 	BindPipeline(p Pipeline)
-	BindFramebuffer(f Framebuffer, a LoadDesc)
 	BindTexture(unit int, t Texture)
 	BindVertexBuffer(b Buffer, offset int)
 	BindIndexBuffer(b Buffer)
@@ -45,6 +47,8 @@ type Device interface {
 	BindFragmentUniforms(buf Buffer)
 	BindStorageBuffer(binding int, buf Buffer)
 
+	BeginCompute()
+	EndCompute()
 	CopyTexture(dst Texture, dstOrigin image.Point, src Framebuffer, srcRect image.Rectangle)
 	MemoryBarrier()
 	DispatchCompute(x, y, z int)
@@ -54,12 +58,7 @@ type Device interface {
 
 type LoadDesc struct {
 	Action     LoadAction
-	ClearColor struct {
-		R float32
-		G float32
-		B float32
-		A float32
-	}
+	ClearColor f32color.RGBA
 }
 
 type Pipeline interface {

gpu/internal/metal/metal_darwin.go

@@ -387,12 +387,12 @@ type Backend struct {
 	computeEnc    C.CFTypeRef
 	blitEnc       C.CFTypeRef
 
+	prog *Program
+
 	stagingBuf C.CFTypeRef
 	stagingOff int
 
 	indexBuf *Buffer
-	state    state
-	newState state
 
 	// bufSizes is scratch space for filling out the spvBufferSizeConstants
 	// that spirv-cross generates for emulating buffer.length expressions in
@@ -400,25 +400,6 @@ type Backend struct {
 	bufSizes []uint32
 }
 
-type state struct {
-	renderPass struct {
-		framebuffer *Framebuffer
-		loadAction  driver.LoadAction
-		clearColor  [4]float32
-	}
-	pipeline *Pipeline
-	program  *Program
-	vertices struct {
-		buffer *Buffer
-		offset int
-	}
-	buffers      [bufferUnits]*Buffer
-	vertUniforms *Buffer
-	fragUniforms *Buffer
-	textures     [texUnits]*Texture
-	viewport     C.MTLViewport
-}
-
 type Texture struct {
 	backend *Backend
 	texture C.CFTypeRef
@@ -517,7 +498,6 @@ func (b *Backend) startBlit() C.CFTypeRef {
 	if b.blitEnc == 0 {
 		panic("metal: [MTLCommandBuffer blitCommandEncoder:] failed")
 	}
-	b.state = state{}
 	return b.blitEnc
 }
 
@@ -805,24 +785,34 @@ func (b *Backend) newShader(src shader.Sources) (*Shader, error) {
 }
 
 func (b *Backend) Viewport(x, y, width, height int) {
-	b.newState.viewport = C.MTLViewport{
+	enc := b.renderEnc
+	if enc == 0 {
+		panic("no active render pass")
+	}
+	C.renderEncViewport(enc, C.MTLViewport{
 		originX: C.double(x),
 		originY: C.double(y),
 		width:   C.double(width),
 		height:  C.double(height),
 		znear:   0.0,
 		zfar:    1.0,
-	}
+	})
 }
 
 func (b *Backend) DrawArrays(mode driver.DrawMode, off, count int) {
-	enc := b.encodeState()
+	enc := b.renderEnc
+	if enc == 0 {
+		panic("no active render pass")
+	}
 	t := primitiveFor(mode)
 	C.renderEncDrawPrimitives(enc, t, C.NSUInteger(off), C.NSUInteger(count))
 }
 
 func (b *Backend) DrawElements(mode driver.DrawMode, off, count int) {
-	enc := b.encodeState()
+	enc := b.renderEnc
+	if enc == 0 {
+		panic("no active render pass")
+	}
 	t := primitiveFor(mode)
 	C.renderEncDrawIndexedPrimitives(enc, t, b.indexBuf.buffer, C.NSUInteger(off), C.NSUInteger(count))
 }
@@ -839,59 +829,46 @@ func primitiveFor(mode driver.DrawMode) C.MTLPrimitiveType {
 }
 
 func (b *Backend) BindImageTexture(unit int, tex driver.Texture, access driver.AccessBits, f driver.TextureFormat) {
-	b.newState.textures[unit] = tex.(*Texture)
+	b.BindTexture(unit, tex)
 }
 
 func (b *Backend) MemoryBarrier() {}
 
-func (b *Backend) startCompute() C.CFTypeRef {
-	if b.computeEnc != 0 {
-		return b.computeEnc
-	}
+func (b *Backend) BeginCompute() {
 	b.endEncoder()
 	b.ensureCmdBuffer()
+	for i := range b.bufSizes {
+		b.bufSizes[i] = 0
+	}
 	b.computeEnc = C.cmdBufferComputeEncoder(b.cmdBuffer)
 	if b.computeEnc == 0 {
 		panic("metal: [MTLCommandBuffer computeCommandEncoder:] failed")
 	}
-	b.state = state{}
-	return b.computeEnc
+}
+
+func (b *Backend) EndCompute() {
+	if b.computeEnc == 0 {
+		panic("no active compute pass")
+	}
+	C.computeEncEnd(b.computeEnc)
+	C.CFRelease(b.computeEnc)
+	b.computeEnc = 0
 }
 
 func (b *Backend) DispatchCompute(x, y, z int) {
-	enc := b.startCompute()
-	p := b.newState.program
-	if p != b.state.program {
-		C.computeEncSetPipeline(enc, p.pipeline)
-	}
-	for i, t := range b.newState.textures {
-		current := b.state.textures[i]
-		if t != current {
-			C.computeEncSetTexture(enc, C.NSUInteger(i), t.texture)
-		}
-	}
-	for i, buf := range b.newState.buffers {
-		b.bufSizes[i] = uint32(buf.size)
-		current := b.state.buffers[i]
-		if buf.buffer != 0 {
-			if buf != current {
-				C.computeEncSetBuffer(enc, C.NSUInteger(i), buf.buffer)
-			}
-		} else if buf.size > 0 {
-			C.computeEncSetBytes(enc, unsafe.Pointer(&buf.store[0]), C.NSUInteger(buf.size), C.NSUInteger(i))
-		}
-	}
-	if n := len(b.newState.buffers); n > 0 {
-		C.computeEncSetBytes(enc, unsafe.Pointer(&b.bufSizes[0]), C.NSUInteger(n*4), spvBufferSizeConstantsBinding)
+	enc := b.computeEnc
+	if enc == 0 {
+		panic("no active compute pass")
 	}
+	C.computeEncSetBytes(enc, unsafe.Pointer(&b.bufSizes[0]), C.NSUInteger(len(b.bufSizes)*4), spvBufferSizeConstantsBinding)
 	threadgroupsPerGrid := C.MTLSize{
 		width: C.NSUInteger(x), height: C.NSUInteger(y), depth: C.NSUInteger(z),
 	}
+	sz := b.prog.groupSize
 	threadsPerThreadgroup := C.MTLSize{
-		width: C.NSUInteger(p.groupSize[0]), height: C.NSUInteger(p.groupSize[1]), depth: C.NSUInteger(p.groupSize[2]),
+		width: C.NSUInteger(sz[0]), height: C.NSUInteger(sz[1]), depth: C.NSUInteger(sz[2]),
 	}
 	C.computeEncDispatch(enc, threadgroupsPerGrid, threadsPerThreadgroup)
-	b.state = b.newState
 }
 
 func (b *Backend) stagingBuffer(size int) (C.CFTypeRef, int) {
@@ -956,34 +933,14 @@ func (p *Pipeline) Release() {
 
 func (b *Backend) BindTexture(unit int, tex driver.Texture) {
 	t := tex.(*Texture)
-	b.newState.textures[unit] = t
-}
-
-func (b *Backend) beginPass() C.CFTypeRef {
-	r := b.newState.renderPass
-	if r == b.state.renderPass {
-		return b.renderEnc
+	if enc := b.renderEnc; enc != 0 {
+		C.renderEncSetFragmentTexture(enc, C.NSUInteger(unit), t.texture)
+		C.renderEncSetFragmentSamplerState(enc, C.NSUInteger(unit), t.sampler)
+	} else if enc := b.computeEnc; enc != 0 {
+		C.computeEncSetTexture(enc, C.NSUInteger(unit), t.texture)
+	} else {
+		panic("no active render nor compute pass")
 	}
-	b.endEncoder()
-	var act C.MTLLoadAction
-	switch r.loadAction {
-	case driver.LoadActionKeep:
-		act = C.MTLLoadActionLoad
-	case driver.LoadActionClear:
-		act = C.MTLLoadActionClear
-	case driver.LoadActionInvalidate:
-		act = C.MTLLoadActionDontCare
-	}
-	b.ensureCmdBuffer()
-	c := r.clearColor
-	b.renderEnc = C.cmdBufferRenderEncoder(b.cmdBuffer, r.framebuffer.texture, act, C.float(c[0]), C.float(c[1]), C.float(c[2]), C.float(c[3]))
-	if b.renderEnc == 0 {
-		panic("metal: [MTLCommandBuffer renderCommandEncoderWithDescriptor:] failed")
-	}
-	r.loadAction = driver.LoadActionKeep
-	b.newState.renderPass = r
-	b.state.renderPass = r
-	return b.renderEnc
 }
 
 func (b *Backend) ensureCmdBuffer() {
@@ -996,59 +953,23 @@ func (b *Backend) ensureCmdBuffer() {
 	}
 }
 
-func (b *Backend) encodeState() C.CFTypeRef {
-	enc := b.beginPass()
-	for i, t := range b.newState.textures {
-		current := b.state.textures[i]
-		if t != current {
-			C.renderEncSetFragmentTexture(enc, C.NSUInteger(i), t.texture)
-			C.renderEncSetFragmentSamplerState(enc, C.NSUInteger(i), t.sampler)
-		}
-	}
-	if p := b.newState.pipeline; p != b.state.pipeline {
-		C.renderEncSetRenderPipelineState(enc, p.pipeline)
-	}
-	if bf := b.newState.fragUniforms; bf != nil {
-		if bf.buffer != 0 {
-			if bf != b.state.fragUniforms {
-				C.renderEncSetFragmentBuffer(enc, bf.buffer, uniformBufferIndex, 0)
-			}
-		} else if bf.size > 0 {
-			C.renderEncSetFragmentBytes(enc, unsafe.Pointer(&bf.store[0]), C.NSUInteger(bf.size), uniformBufferIndex)
-		}
-	}
-	if bf := b.newState.vertUniforms; bf != nil {
-		if bf.buffer != 0 {
-			if bf != b.state.vertUniforms {
-				C.renderEncSetVertexBuffer(enc, bf.buffer, uniformBufferIndex, 0)
-			}
-		} else if bf.size > 0 {
-			C.renderEncSetVertexBytes(enc, unsafe.Pointer(&bf.store[0]), C.NSUInteger(bf.size), uniformBufferIndex)
-		}
-	}
-	if v := b.newState.vertices; v.buffer != nil {
-		if v.buffer.buffer != 0 {
-			if v != b.state.vertices {
-				C.renderEncSetVertexBuffer(enc, v.buffer.buffer, attributeBufferIndex, C.NSUInteger(v.offset))
-			}
-		} else if n := v.buffer.size - v.offset; n > 0 {
-			C.renderEncSetVertexBytes(enc, unsafe.Pointer(&v.buffer.store[v.offset]), C.NSUInteger(n), attributeBufferIndex)
-		}
-	}
-	if vp := b.newState.viewport; vp != b.state.viewport {
-		C.renderEncViewport(enc, vp)
-	}
-	b.state = b.newState
-	return enc
-}
-
 func (b *Backend) BindPipeline(pipe driver.Pipeline) {
 	p := pipe.(*Pipeline)
-	b.newState.pipeline = p
+	enc := b.renderEnc
+	if enc == 0 {
+		panic("no active render pass")
+	}
+	C.renderEncSetRenderPipelineState(enc, p.pipeline)
 }
 
 func (b *Backend) BindProgram(prog driver.Program) {
-	b.newState.program = prog.(*Program)
+	enc := b.computeEnc
+	if enc == 0 {
+		panic("no active compute pass")
+	}
+	p := prog.(*Program)
+	C.computeEncSetPipeline(enc, p.pipeline)
+	b.prog = p
 }
 
 func (s *Shader) Release() {
@@ -1062,23 +983,56 @@ func (p *Program) Release() {
 }
 
 func (b *Backend) BindStorageBuffer(binding int, buffer driver.Buffer) {
-	b.newState.buffers[binding] = buffer.(*Buffer)
+	buf := buffer.(*Buffer)
+	b.bufSizes[binding] = uint32(buf.size)
+	enc := b.computeEnc
+	if enc == 0 {
+		panic("no active compute pass")
+	}
+	if buf.buffer != 0 {
+		C.computeEncSetBuffer(enc, C.NSUInteger(binding), buf.buffer)
+	} else if buf.size > 0 {
+		C.computeEncSetBytes(enc, unsafe.Pointer(&buf.store[0]), C.NSUInteger(buf.size), C.NSUInteger(binding))
+	}
 }
 
 func (b *Backend) BindVertexUniforms(buf driver.Buffer) {
 	bf := buf.(*Buffer)
-	b.newState.vertUniforms = bf
+	enc := b.renderEnc
+	if enc == 0 {
+		panic("no active render pass")
+	}
+	if bf.buffer != 0 {
+		C.renderEncSetVertexBuffer(enc, bf.buffer, uniformBufferIndex, 0)
+	} else if bf.size > 0 {
+		C.renderEncSetVertexBytes(enc, unsafe.Pointer(&bf.store[0]), C.NSUInteger(bf.size), uniformBufferIndex)
+	}
 }
 
 func (b *Backend) BindFragmentUniforms(buf driver.Buffer) {
 	bf := buf.(*Buffer)
-	b.newState.fragUniforms = bf
+	enc := b.renderEnc
+	if enc == 0 {
+		panic("no active render pass")
+	}
+	if bf.buffer != 0 {
+		C.renderEncSetFragmentBuffer(enc, bf.buffer, uniformBufferIndex, 0)
+	} else if bf.size > 0 {
+		C.renderEncSetFragmentBytes(enc, unsafe.Pointer(&bf.store[0]), C.NSUInteger(bf.size), uniformBufferIndex)
+	}
 }
 
 func (b *Backend) BindVertexBuffer(buf driver.Buffer, offset int) {
 	bf := buf.(*Buffer)
-	b.newState.vertices.buffer = bf
-	b.newState.vertices.offset = offset
+	enc := b.renderEnc
+	if enc == 0 {
+		panic("no active render pass")
+	}
+	if bf.buffer != 0 {
+		C.renderEncSetVertexBuffer(enc, bf.buffer, attributeBufferIndex, C.NSUInteger(offset))
+	} else if n := bf.size - offset; n > 0 {
+		C.renderEncSetVertexBytes(enc, unsafe.Pointer(&bf.store[offset]), C.NSUInteger(n), attributeBufferIndex)
+	}
 }
 
 func (b *Backend) BindIndexBuffer(buf driver.Buffer) {
@@ -1162,32 +1116,46 @@ func (f *Framebuffer) ReadPixels(src image.Rectangle, pixels []byte, stride int)
 	return nil
 }
 
-func (b *Backend) BindFramebuffer(fbo driver.Framebuffer, desc driver.LoadDesc) {
-	b.newState.renderPass.framebuffer = fbo.(*Framebuffer)
-	b.newState.renderPass.loadAction = desc.Action
-	c := desc.ClearColor
-	b.newState.renderPass.clearColor = [4]float32{c.R, c.G, c.B, c.A}
-	b.beginPass()
+func (b *Backend) BeginRenderPass(fbo driver.Framebuffer, d driver.LoadDesc) {
+	b.endEncoder()
+	b.ensureCmdBuffer()
+	f := fbo.(*Framebuffer)
+	col := d.ClearColor
+	var act C.MTLLoadAction
+	switch d.Action {
+	case driver.LoadActionKeep:
+		act = C.MTLLoadActionLoad
+	case driver.LoadActionClear:
+		act = C.MTLLoadActionClear
+	case driver.LoadActionInvalidate:
+		act = C.MTLLoadActionDontCare
+	}
+	b.renderEnc = C.cmdBufferRenderEncoder(b.cmdBuffer, f.texture, act, C.float(col.R), C.float(col.G), C.float(col.B), C.float(col.A))
+	if b.renderEnc == 0 {
+		panic("metal: [MTLCommandBuffer renderCommandEncoderWithDescriptor:] failed")
+	}
+}
+
+func (b *Backend) EndRenderPass() {
+	if b.renderEnc == 0 {
+		panic("no active render pass")
+	}
+	C.renderEncEnd(b.renderEnc)
+	C.CFRelease(b.renderEnc)
+	b.renderEnc = 0
 }
 
 func (b *Backend) endEncoder() {
 	if b.renderEnc != 0 {
-		C.renderEncEnd(b.renderEnc)
-		C.CFRelease(b.renderEnc)
-		b.renderEnc = 0
-		b.state = state{}
+		panic("active render pass")
 	}
 	if b.computeEnc != 0 {
-		C.computeEncEnd(b.computeEnc)
-		C.CFRelease(b.computeEnc)
-		b.computeEnc = 0
-		b.state = state{}
+		panic("active compute pass")
 	}
 	if b.blitEnc != 0 {
 		C.blitEncEnd(b.blitEnc)
 		C.CFRelease(b.blitEnc)
 		b.blitEnc = 0
-		b.state = state{}
 	}
 }
 

gpu/internal/opengl/opengl.go

@@ -660,7 +660,7 @@ 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, driver.LoadDesc{})
+	b.glstate.bindFramebuffer(b.funcs, gl.FRAMEBUFFER, fbo.obj)
 	if err := glErr(b.funcs); err != nil {
 		fbo.Release()
 		return nil, err
@@ -1194,7 +1194,7 @@ func (b *Backend) CopyTexture(dst driver.Texture, dstOrigin image.Point, src dri
 
 func (f *framebuffer) ReadPixels(src image.Rectangle, pixels []byte, stride int) error {
 	glErr(f.backend.funcs)
-	f.backend.BindFramebuffer(f, driver.LoadDesc{})
+	f.backend.glstate.bindFramebuffer(f.backend.funcs, gl.FRAMEBUFFER, f.obj)
 	if len(pixels) < src.Dx()*src.Dy()*4 {
 		return errors.New("unexpected RGBA size")
 	}
@@ -1217,7 +1217,13 @@ func (b *Backend) BindPipeline(pl driver.Pipeline) {
 	b.BlendFunc(p.blend.SrcFactor, p.blend.DstFactor)
 }
 
-func (b *Backend) BindFramebuffer(fbo driver.Framebuffer, desc driver.LoadDesc) {
+func (b *Backend) BeginCompute() {
+}
+
+func (b *Backend) EndCompute() {
+}
+
+func (b *Backend) BeginRenderPass(fbo driver.Framebuffer, desc driver.LoadDesc) {
 	b.glstate.bindFramebuffer(b.funcs, gl.FRAMEBUFFER, fbo.(*framebuffer).obj)
 	switch desc.Action {
 	case driver.LoadActionClear:
@@ -1228,6 +1234,9 @@ func (b *Backend) BindFramebuffer(fbo driver.Framebuffer, desc driver.LoadDesc)
 	}
 }
 
+func (b *Backend) EndRenderPass() {
+}
+
 func (f *framebuffer) Release() {
 	if f.foreign {
 		panic("framebuffer not created by NewFramebuffer")

gpu/path.go

@@ -346,7 +346,6 @@ func (s *stenciler) beginIntersect(sizes []image.Point) {
 	// floating point formats. Replace with GL_RGB+GL_UNSIGNED_BYTE if
 	// no floating point support is available.
 	s.intersections.resize(s.ctx, sizes)
-	s.ctx.BindPipeline(s.ipipeline.pipeline.pipeline)
 }
 
 func (s *stenciler) cover(idx int) stencilFBO {
@@ -355,8 +354,6 @@ func (s *stenciler) cover(idx int) stencilFBO {
 
 func (s *stenciler) begin(sizes []image.Point) {
 	s.fbos.resize(s.ctx, sizes)
-	s.ctx.BindPipeline(s.pipeline.pipeline.pipeline)
-	s.ctx.BindIndexBuffer(s.indexBuf)
 }
 
 func (s *stenciler) stencilPath(bounds image.Rectangle, offset f32.Point, uv image.Point, data pathData) {
@@ -388,8 +385,6 @@ func (p *pather) cover(mat materialType, col f32color.RGBA, col1, col2 f32color.
 }
 
 func (c *coverer) cover(mat materialType, col f32color.RGBA, col1, col2 f32color.RGBA, scale, off f32.Point, uvTrans f32.Affine2D, coverScale, coverOff f32.Point) {
-	p := c.pipelines[mat]
-	c.ctx.BindPipeline(p.pipeline)
 	var uniforms *coverUniforms
 	switch mat {
 	case materialColor:
@@ -411,7 +406,7 @@ func (c *coverer) cover(mat materialType, col f32color.RGBA, col1, col2 f32color
 	}
 	uniforms.transform = [4]float32{scale.X, scale.Y, off.X, off.Y}
 	uniforms.uvCoverTransform = [4]float32{coverScale.X, coverScale.Y, coverOff.X, coverOff.Y}
-	p.UploadUniforms(c.ctx)
+	c.pipelines[mat].UploadUniforms(c.ctx)
 	c.ctx.DrawArrays(driver.DrawModeTriangleStrip, 0, 4)
 }