gpu: switch from uniform buffers to combined push buffers

Vulkan doesn't support changing uniforms during a render pass. However,
push constants *can* be changed. The gio-shader repository was changed
to use push constants instead of uniforms, this change implements the
corresponding driver and renderer change.

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

go.mod

@@ -10,5 +10,5 @@ require (
 
 require (
 	gioui.org/cpu v0.0.0-20210817075930-8d6a761490d2
-	gioui.org/shader v1.0.2
+	gioui.org/shader v1.0.3
 )

go.sum

@@ -4,8 +4,8 @@ dmitri.shuralyov.com/gpu/mtl v0.0.0-20201218220906-28db891af037/go.mod h1:H6x//7
 gioui.org/cpu v0.0.0-20210808092351-bfe733dd3334/go.mod h1:A8M0Cn5o+vY5LTMlnRoK3O5kG+rH0kWfJjeKd9QpBmQ=
 gioui.org/cpu v0.0.0-20210817075930-8d6a761490d2 h1:AGDDxsJE1RpcXTAxPG2B4jrwVUJGFDjINIPi1jtO6pc=
 gioui.org/cpu v0.0.0-20210817075930-8d6a761490d2/go.mod h1:A8M0Cn5o+vY5LTMlnRoK3O5kG+rH0kWfJjeKd9QpBmQ=
-gioui.org/shader v1.0.2 h1:5JrtNdOqmY+tQzBKKVFqf0YIziUK38OEw8mxxEbXdc8=
-gioui.org/shader v1.0.2/go.mod h1:mWdiME581d/kV7/iEhLmUgUK5iZ09XR5XpduXzbePVM=
+gioui.org/shader v1.0.3 h1:xllhh4w2XWej9lfkHMtOtI3ES05T8xZzxrSSWbMZgTw=
+gioui.org/shader v1.0.3/go.mod h1:mWdiME581d/kV7/iEhLmUgUK5iZ09XR5XpduXzbePVM=
 github.com/BurntSushi/toml v0.3.1/go.mod h1:xHWCNGjB5oqiDr8zfno3MHue2Ht5sIBksp03qcyfWMU=
 github.com/BurntSushi/xgb v0.0.0-20160522181843-27f122750802/go.mod h1:IVnqGOEym/WlBOVXweHU+Q+/VP0lqqI8lqeDx9IjBqo=
 github.com/Knetic/govaluate v3.0.1-0.20171022003610-9aa49832a739+incompatible/go.mod h1:r7JcOSlj0wfOMncg0iLm8Leh48TZaKVeNIfJntJ2wa0=

gpu/compute.go

@@ -88,11 +88,8 @@ type compute struct {
 		pipeline driver.Pipeline
 		buffer   sizedBuffer
 		quads    []materialVertex
-		vert     struct {
-			uniforms *materialVertUniforms
-			buf      driver.Buffer
-		}
-		frag struct {
+		uniforms struct {
+			u   *materialUniforms
 			buf driver.Buffer
 		}
 	}
@@ -172,13 +169,10 @@ type copyUniforms struct {
 	_       [8]byte // Pad to 16 bytes.
 }
 
-type materialVertUniforms struct {
-	scale [2]float32
-	pos   [2]float32
-}
-
-type materialFragUniforms struct {
-	emulateSRGB float32
+type materialUniforms struct {
+	scale       [2]float32
+	pos         [2]float32
+	emulatesRGB float32
 	_           [12]byte // Pad to 16 bytes
 }
 
@@ -504,25 +498,14 @@ func newCompute(ctx driver.Device) (*compute, error) {
 		return nil, err
 	}
 	g.materials.pipeline = pipe
-	g.materials.vert.uniforms = new(materialVertUniforms)
+	g.materials.uniforms.u = new(materialUniforms)
 
-	buf, err = ctx.NewBuffer(driver.BufferBindingUniforms, int(unsafe.Sizeof(*g.materials.vert.uniforms)))
+	buf, err = ctx.NewBuffer(driver.BufferBindingUniforms, int(unsafe.Sizeof(*g.materials.uniforms.u)))
 	if err != nil {
 		g.Release()
 		return nil, err
 	}
-	g.materials.vert.buf = buf
-	var emulateSRGB materialFragUniforms
-	if !g.srgb {
-		emulateSRGB.emulateSRGB = 1.0
-	}
-	buf, err = ctx.NewBuffer(driver.BufferBindingUniforms, int(unsafe.Sizeof(emulateSRGB)))
-	if err != nil {
-		g.Release()
-		return nil, err
-	}
-	buf.Upload(byteslice.Struct(&emulateSRGB))
-	g.materials.frag.buf = buf
+	g.materials.uniforms.buf = buf
 
 	for _, shader := range shaders {
 		if !g.useCPU {
@@ -874,7 +857,6 @@ func (g *compute) blitLayers(viewport image.Point) {
 	layers := g.collector.frame.layers
 	g.ctx.Viewport(0, 0, viewport.X, viewport.Y)
 	g.ctx.BindPipeline(g.output.blitPipeline)
-	g.ctx.BindVertexUniforms(g.output.uniBuf)
 	for len(layers) > 0 {
 		g.output.layerVertices = g.output.layerVertices[:0]
 		atlas := layers[0].alloc.atlas
@@ -906,6 +888,7 @@ func (g *compute) blitLayers(viewport image.Point) {
 		g.output.uniforms.pos = [2]float32{ox, oy}
 		g.output.uniforms.uvScale = [2]float32{1 / float32(atlas.size.X), 1 / float32(atlas.size.Y)}
 		g.output.uniBuf.Upload(byteslice.Struct(g.output.uniforms))
+		g.ctx.BindUniforms(g.output.uniBuf)
 		vertexData := byteslice.Slice(g.output.layerVertices)
 		g.output.buffer.ensureCapacity(false, g.ctx, driver.BufferBindingVertices, len(vertexData))
 		g.output.buffer.buffer.Upload(vertexData)
@@ -1001,9 +984,14 @@ func (g *compute) renderMaterials() error {
 		}
 		// Transform to clip space: [-1, -1] - [1, 1] and flip Y-axis to cancel the implied transformation
 		// between framebuffer and texture space.
-		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))
+		*m.uniforms.u = materialUniforms{
+			scale: [2]float32{2, -2},
+			pos:   [2]float32{-1, +1},
+		}
+		if !g.srgb {
+			m.uniforms.u.emulatesRGB = 1.0
+		}
+		m.uniforms.buf.Upload(byteslice.Struct(m.uniforms.u))
 		vertexData := byteslice.Slice(m.quads)
 		n := pow2Ceil(len(vertexData))
 		m.buffer.ensureCapacity(false, g.ctx, driver.BufferBindingVertices, n)
@@ -1013,10 +1001,9 @@ func (g *compute) renderMaterials() error {
 			d.Action = driver.LoadActionClear
 		}
 		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.BindUniforms(m.uniforms.buf)
 		g.ctx.BindVertexBuffer(m.buffer.buffer, 0)
 		newAllocs := atlas.allocs[allocStart:]
 		for i, a := range newAllocs {
@@ -1507,8 +1494,7 @@ func (g *compute) Release() {
 		&g.buffers.config,
 		g.materials.pipeline,
 		&g.materials.buffer,
-		g.materials.vert.buf,
-		g.materials.frag.buf,
+		g.materials.uniforms.buf,
 		g.timers.t,
 	}
 	for _, r := range res {

gpu/gpu.go

@@ -282,30 +282,20 @@ type blitter struct {
 }
 
 type blitColUniforms struct {
-	vert struct {
-		blitUniforms
-		_ [12]byte // Padding to a multiple of 16.
-	}
-	frag struct {
-		colorUniforms
-	}
+	blitUniforms
+	_ [128 - unsafe.Sizeof(blitUniforms{}) - unsafe.Sizeof(colorUniforms{})]byte // Padding to 128 bytes.
+	colorUniforms
 }
 
 type blitTexUniforms struct {
-	vert struct {
-		blitUniforms
-		_ [12]byte // Padding to a multiple of 16.
-	}
+	blitUniforms
+	_ [12]byte // Padding to 16 bytes.
 }
 
 type blitLinearGradientUniforms struct {
-	vert struct {
-		blitUniforms
-		_ [12]byte // Padding to a multiple of 16.
-	}
-	frag struct {
-		gradientUniforms
-	}
+	blitUniforms
+	_ [128 - unsafe.Sizeof(blitUniforms{}) - unsafe.Sizeof(gradientUniforms{})]byte // Padding to 128 bytes.
+	gradientUniforms
 }
 
 type uniformBuffer struct {
@@ -314,9 +304,8 @@ type uniformBuffer struct {
 }
 
 type pipeline struct {
-	pipeline     driver.Pipeline
-	vertUniforms *uniformBuffer
-	fragUniforms *uniformBuffer
+	pipeline driver.Pipeline
+	uniforms *uniformBuffer
 }
 
 type blitUniforms struct {
@@ -536,8 +525,7 @@ func newBlitter(ctx driver.Device) *blitter {
 	b.texUniforms = new(blitTexUniforms)
 	b.linearGradientUniforms = new(blitLinearGradientUniforms)
 	pipelines, err := createColorPrograms(ctx, gio.Shader_blit_vert, gio.Shader_blit_frag,
-		[3]interface{}{&b.colUniforms.vert, &b.linearGradientUniforms.vert, &b.texUniforms.vert},
-		[3]interface{}{&b.colUniforms.frag, &b.linearGradientUniforms.frag, nil},
+		[3]interface{}{b.colUniforms, b.linearGradientUniforms, b.texUniforms},
 	)
 	if err != nil {
 		panic(err)
@@ -553,7 +541,7 @@ func (b *blitter) release() {
 	}
 }
 
-func createColorPrograms(b driver.Device, vsSrc shader.Sources, fsSrc [3]shader.Sources, vertUniforms, fragUniforms [3]interface{}) ([3]*pipeline, error) {
+func createColorPrograms(b driver.Device, vsSrc shader.Sources, fsSrc [3]shader.Sources, uniforms [3]interface{}) ([3]*pipeline, error) {
 	var pipelines [3]*pipeline
 	blend := driver.BlendDesc{
 		Enable:    true,
@@ -589,17 +577,14 @@ func createColorPrograms(b driver.Device, vsSrc shader.Sources, fsSrc [3]shader.
 		if err != nil {
 			return pipelines, err
 		}
-		var vertBuffer, fragBuffer *uniformBuffer
-		if u := vertUniforms[materialTexture]; u != nil {
+		var vertBuffer *uniformBuffer
+		if u := uniforms[materialTexture]; u != nil {
 			vertBuffer = newUniformBuffer(b, u)
 		}
-		if u := fragUniforms[materialTexture]; u != nil {
-			fragBuffer = newUniformBuffer(b, u)
-		}
-		pipelines[materialTexture] = &pipeline{pipe, vertBuffer, fragBuffer}
+		pipelines[materialTexture] = &pipeline{pipe, vertBuffer}
 	}
 	{
-		var vertBuffer, fragBuffer *uniformBuffer
+		var vertBuffer *uniformBuffer
 		fsh, err := b.NewFragmentShader(fsSrc[materialColor])
 		if err != nil {
 			pipelines[materialTexture].Release()
@@ -618,16 +603,13 @@ func createColorPrograms(b driver.Device, vsSrc shader.Sources, fsSrc [3]shader.
 			pipelines[materialTexture].Release()
 			return pipelines, err
 		}
-		if u := vertUniforms[materialColor]; u != nil {
+		if u := uniforms[materialColor]; u != nil {
 			vertBuffer = newUniformBuffer(b, u)
 		}
-		if u := fragUniforms[materialColor]; u != nil {
-			fragBuffer = newUniformBuffer(b, u)
-		}
-		pipelines[materialColor] = &pipeline{pipe, vertBuffer, fragBuffer}
+		pipelines[materialColor] = &pipeline{pipe, vertBuffer}
 	}
 	{
-		var vertBuffer, fragBuffer *uniformBuffer
+		var vertBuffer *uniformBuffer
 		fsh, err := b.NewFragmentShader(fsSrc[materialLinearGradient])
 		if err != nil {
 			pipelines[materialTexture].Release()
@@ -648,13 +630,10 @@ func createColorPrograms(b driver.Device, vsSrc shader.Sources, fsSrc [3]shader.
 			pipelines[materialColor].Release()
 			return pipelines, err
 		}
-		if u := vertUniforms[materialLinearGradient]; u != nil {
+		if u := uniforms[materialLinearGradient]; u != nil {
 			vertBuffer = newUniformBuffer(b, u)
 		}
-		if u := fragUniforms[materialLinearGradient]; u != nil {
-			fragBuffer = newUniformBuffer(b, u)
-		}
-		pipelines[materialLinearGradient] = &pipeline{pipe, vertBuffer, fragBuffer}
+		pipelines[materialLinearGradient] = &pipeline{pipe, vertBuffer}
 	}
 	if err != nil {
 		for _, p := range pipelines {
@@ -1151,21 +1130,21 @@ func (b *blitter) blit(mat materialType, col f32color.RGBA, col1, col2 f32color.
 	var uniforms *blitUniforms
 	switch mat {
 	case materialColor:
-		b.colUniforms.frag.color = col
-		uniforms = &b.colUniforms.vert.blitUniforms
+		b.colUniforms.color = col
+		uniforms = &b.colUniforms.blitUniforms
 	case materialTexture:
 		t1, t2, t3, t4, t5, t6 := uvTrans.Elems()
-		b.texUniforms.vert.blitUniforms.uvTransformR1 = [4]float32{t1, t2, t3, 0}
-		b.texUniforms.vert.blitUniforms.uvTransformR2 = [4]float32{t4, t5, t6, 0}
-		uniforms = &b.texUniforms.vert.blitUniforms
+		b.texUniforms.blitUniforms.uvTransformR1 = [4]float32{t1, t2, t3, 0}
+		b.texUniforms.blitUniforms.uvTransformR2 = [4]float32{t4, t5, t6, 0}
+		uniforms = &b.texUniforms.blitUniforms
 	case materialLinearGradient:
-		b.linearGradientUniforms.frag.color1 = col1
-		b.linearGradientUniforms.frag.color2 = col2
+		b.linearGradientUniforms.color1 = col1
+		b.linearGradientUniforms.color2 = col2
 
 		t1, t2, t3, t4, t5, t6 := uvTrans.Elems()
-		b.linearGradientUniforms.vert.blitUniforms.uvTransformR1 = [4]float32{t1, t2, t3, 0}
-		b.linearGradientUniforms.vert.blitUniforms.uvTransformR2 = [4]float32{t4, t5, t6, 0}
-		uniforms = &b.linearGradientUniforms.vert.blitUniforms
+		b.linearGradientUniforms.blitUniforms.uvTransformR1 = [4]float32{t1, t2, t3, 0}
+		b.linearGradientUniforms.blitUniforms.uvTransformR2 = [4]float32{t4, t5, t6, 0}
+		uniforms = &b.linearGradientUniforms.blitUniforms
 	}
 	uniforms.transform = [4]float32{scale.X, scale.Y, off.X, off.Y}
 	p.UploadUniforms(b.ctx)
@@ -1197,23 +1176,16 @@ func (u *uniformBuffer) Release() {
 }
 
 func (p *pipeline) UploadUniforms(ctx driver.Device) {
-	if p.vertUniforms != nil {
-		p.vertUniforms.Upload()
-		ctx.BindVertexUniforms(p.vertUniforms.buf)
-	}
-	if p.fragUniforms != nil {
-		p.fragUniforms.Upload()
-		ctx.BindFragmentUniforms(p.fragUniforms.buf)
+	if p.uniforms != nil {
+		p.uniforms.Upload()
+		ctx.BindUniforms(p.uniforms.buf)
 	}
 }
 
 func (p *pipeline) Release() {
 	p.pipeline.Release()
-	if p.vertUniforms != nil {
-		p.vertUniforms.Release()
-	}
-	if p.fragUniforms != nil {
-		p.fragUniforms.Release()
+	if p.uniforms != nil {
+		p.uniforms.Release()
 	}
 	*p = pipeline{}
 }

gpu/internal/d3d11/d3d11_windows.go

@@ -681,13 +681,9 @@ func (b *Backend) BindStorageBuffer(binding int, buffer driver.Buffer) {
 	}
 }
 
-func (b *Backend) BindVertexUniforms(buffer driver.Buffer) {
+func (b *Backend) BindUniforms(buffer driver.Buffer) {
 	buf := buffer.(*Buffer)
 	b.ctx.VSSetConstantBuffers(buf.buf)
-}
-
-func (b *Backend) BindFragmentUniforms(buffer driver.Buffer) {
-	buf := buffer.(*Buffer)
 	b.ctx.PSSetConstantBuffers(buf.buf)
 }
 

gpu/internal/driver/driver.go

@@ -43,8 +43,7 @@ type Device interface {
 	BindVertexBuffer(b Buffer, offset int)
 	BindIndexBuffer(b Buffer)
 	BindImageTexture(unit int, texture Texture, access AccessBits, format TextureFormat)
-	BindVertexUniforms(buf Buffer)
-	BindFragmentUniforms(buf Buffer)
+	BindUniforms(buf Buffer)
 	BindStorageBuffer(binding int, buf Buffer)
 
 	BeginCompute()

gpu/internal/metal/metal_darwin.go

@@ -1005,20 +1005,9 @@ func (b *Backend) BindVertexUniforms(buf driver.Buffer) {
 	}
 	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)
-	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.renderEncSetVertexBytes(enc, unsafe.Pointer(&bf.store[0]), C.NSUInteger(bf.size), uniformBufferIndex)
 		C.renderEncSetFragmentBytes(enc, unsafe.Pointer(&bf.store[0]), C.NSUInteger(bf.size), uniformBufferIndex)
 	}
 }

gpu/internal/opengl/opengl.go

@@ -26,17 +26,14 @@ type Backend struct {
 
 	glver [2]int
 	gles  bool
-	ubo   bool
 	feats driver.Caps
 	// floatTriple holds the settings for floating point
 	// textures.
 	floatTriple textureTriple
 	// Single channel alpha textures.
-	alphaTriple  textureTriple
-	srgbaTriple  textureTriple
-	vertUniforms *buffer
-	fragUniforms *buffer
-	storage      [storageBindings]*buffer
+	alphaTriple textureTriple
+	srgbaTriple textureTriple
+	storage     [storageBindings]*buffer
 
 	sRGBFBO *SRGBFBO
 
@@ -190,13 +187,10 @@ func newOpenGLDevice(api driver.OpenGL) (driver.Device, error) {
 	}
 	floatTriple, ffboErr := floatTripleFor(f, ver, exts)
 	srgbaTriple, srgbErr := srgbaTripleFor(ver, exts)
-	gles30 := gles && ver[0] >= 3
 	gles31 := gles && (ver[0] > 3 || (ver[0] == 3 && ver[1] >= 1))
-	gl40 := !gles && ver[0] >= 4
 	b := &Backend{
 		glver:       ver,
 		gles:        gles,
-		ubo:         gles30 || gl40,
 		funcs:       f,
 		floatTriple: floatTriple,
 		alphaTriple: alphaTripleFor(ver),
@@ -712,12 +706,9 @@ func (b *Backend) NewBuffer(typ driver.BufferBinding, size int) (driver.Buffer,
 		if typ != driver.BufferBindingUniforms {
 			return nil, errors.New("uniforms buffers cannot be bound as anything else")
 		}
-		if !b.ubo {
-			// GLES 2 doesn't support uniform buffers.
-			buf.data = make([]byte, size)
-		}
+		buf.data = make([]byte, size)
 	}
-	if typ&^driver.BufferBindingUniforms != 0 || b.ubo {
+	if typ&^driver.BufferBindingUniforms != 0 {
 		buf.hasBuffer = true
 		buf.obj = b.funcs.CreateBuffer()
 		if err := glErr(b.funcs); err != nil {
@@ -841,7 +832,6 @@ func (b *Backend) prepareDraw() {
 		return
 	}
 	b.setupVertexArrays()
-	p.prog.updateUniforms()
 }
 
 func toGLDrawMode(mode driver.Topology) gl.Enum {
@@ -954,27 +944,8 @@ func (b *Backend) newProgram(desc driver.PipelineDesc) (*program, error) {
 			b.funcs.Uniform1i(u, tex.Binding)
 		}
 	}
-	if b.ubo {
-		for _, block := range vsh.src.Uniforms.Blocks {
-			blockIdx := b.funcs.GetUniformBlockIndex(p, block.Name)
-			if blockIdx != gl.INVALID_INDEX {
-				b.funcs.UniformBlockBinding(p, blockIdx, uint(block.Binding))
-			}
-		}
-		// To match Direct3D 11 with separate vertex and fragment
-		// shader uniform buffers, offset all fragment blocks to be
-		// located after the vertex blocks.
-		off := len(vsh.src.Uniforms.Blocks)
-		for _, block := range fsh.src.Uniforms.Blocks {
-			blockIdx := b.funcs.GetUniformBlockIndex(p, block.Name)
-			if blockIdx != gl.INVALID_INDEX {
-				b.funcs.UniformBlockBinding(p, blockIdx, uint(block.Binding+off))
-			}
-		}
-	} else {
-		prog.vertUniforms.setup(b.funcs, p, vsh.src.Uniforms.Size, vsh.src.Uniforms.Locations)
-		prog.fragUniforms.setup(b.funcs, p, fsh.src.Uniforms.Size, fsh.src.Uniforms.Locations)
-	}
+	prog.vertUniforms.setup(b.funcs, p, vsh.src.Uniforms.Size, vsh.src.Uniforms.Locations)
+	prog.fragUniforms.setup(b.funcs, p, fsh.src.Uniforms.Size, fsh.src.Uniforms.Locations)
 	return prog, nil
 }
 
@@ -994,38 +965,13 @@ func (b *Backend) BindStorageBuffer(binding int, buf driver.Buffer) {
 	b.storage[binding] = bf
 }
 
-func (b *Backend) BindVertexUniforms(buf driver.Buffer) {
+func (b *Backend) BindUniforms(buf driver.Buffer) {
 	bf := buf.(*buffer)
 	if bf.typ&driver.BufferBindingUniforms == 0 {
 		panic("not a uniform buffer")
 	}
-	b.vertUniforms = bf
-}
-
-func (b *Backend) BindFragmentUniforms(buf driver.Buffer) {
-	bf := buf.(*buffer)
-	if bf.typ&driver.BufferBindingUniforms == 0 {
-		panic("not a uniform buffer")
-	}
-	b.fragUniforms = bf
-}
-
-func (p *program) updateUniforms() {
-	f := p.backend.funcs
-	if b := p.backend.vertUniforms; b != nil {
-		if p.backend.ubo {
-			p.backend.glstate.bindBufferBase(f, gl.UNIFORM_BUFFER, 0, b.obj)
-		} else {
-			p.vertUniforms.update(f, b)
-		}
-	}
-	if b := p.backend.fragUniforms; b != nil {
-		if p.backend.ubo {
-			p.backend.glstate.bindBufferBase(f, gl.UNIFORM_BUFFER, 1, b.obj)
-		} else {
-			p.fragUniforms.update(f, b)
-		}
-	}
+	b.state.pipeline.prog.vertUniforms.update(b.funcs, bf)
+	b.state.pipeline.prog.fragUniforms.update(b.funcs, bf)
 }
 
 func (b *Backend) BindProgram(prog driver.Program) {

gpu/path.go

@@ -37,30 +37,20 @@ type coverer struct {
 }
 
 type coverTexUniforms struct {
-	vert struct {
-		coverUniforms
-		_ [12]byte // Padding to multiple of 16.
-	}
+	coverUniforms
+	_ [12]byte // Padding to multiple of 16.
 }
 
 type coverColUniforms struct {
-	vert struct {
-		coverUniforms
-		_ [12]byte // Padding to multiple of 16.
-	}
-	frag struct {
-		colorUniforms
-	}
+	coverUniforms
+	_ [128 - unsafe.Sizeof(coverUniforms{}) - unsafe.Sizeof(colorUniforms{})]byte // Padding to 128 bytes.
+	colorUniforms
 }
 
 type coverLinearGradientUniforms struct {
-	vert struct {
-		coverUniforms
-		_ [12]byte // Padding to multiple of 16.
-	}
-	frag struct {
-		gradientUniforms
-	}
+	coverUniforms
+	_ [128 - unsafe.Sizeof(coverUniforms{}) - unsafe.Sizeof(gradientUniforms{})]byte // Padding to 128.
+	gradientUniforms
 }
 
 type coverUniforms struct {
@@ -87,11 +77,9 @@ type stenciler struct {
 }
 
 type stencilUniforms struct {
-	vert struct {
-		transform  [4]float32
-		pathOffset [2]float32
-		_          [8]byte // Padding to multiple of 16.
-	}
+	transform  [4]float32
+	pathOffset [2]float32
+	_          [8]byte // Padding to multiple of 16.
 }
 
 type intersectUniforms struct {
@@ -161,8 +149,7 @@ func newCoverer(ctx driver.Device) *coverer {
 	c.texUniforms = new(coverTexUniforms)
 	c.linearGradientUniforms = new(coverLinearGradientUniforms)
 	pipelines, err := createColorPrograms(ctx, gio.Shader_cover_vert, gio.Shader_cover_frag,
-		[3]interface{}{&c.colUniforms.vert, &c.linearGradientUniforms.vert, &c.texUniforms.vert},
-		[3]interface{}{&c.colUniforms.frag, &c.linearGradientUniforms.frag, nil},
+		[3]interface{}{c.colUniforms, c.linearGradientUniforms, c.texUniforms},
 	)
 	if err != nil {
 		panic(err)
@@ -215,7 +202,7 @@ func newStenciler(ctx driver.Device) *stenciler {
 	defer vsh.Release()
 	defer fsh.Release()
 	st.pipeline.uniforms = new(stencilUniforms)
-	vertUniforms := newUniformBuffer(ctx, &st.pipeline.uniforms.vert)
+	vertUniforms := newUniformBuffer(ctx, st.pipeline.uniforms)
 	pipe, err := st.ctx.NewPipeline(driver.PipelineDesc{
 		VertexShader:   vsh,
 		FragmentShader: fsh,
@@ -228,7 +215,7 @@ func newStenciler(ctx driver.Device) *stenciler {
 		PixelFormat: driver.TextureFormatFloat,
 		Topology:    driver.TopologyTriangles,
 	})
-	st.pipeline.pipeline = &pipeline{pipe, vertUniforms, nil}
+	st.pipeline.pipeline = &pipeline{pipe, vertUniforms}
 	if err != nil {
 		panic(err)
 	}
@@ -252,7 +239,7 @@ func newStenciler(ctx driver.Device) *stenciler {
 		PixelFormat: driver.TextureFormatFloat,
 		Topology:    driver.TopologyTriangleStrip,
 	})
-	st.ipipeline.pipeline = &pipeline{ipipe, vertUniforms, nil}
+	st.ipipeline.pipeline = &pipeline{ipipe, vertUniforms}
 	return st
 }
 
@@ -364,8 +351,8 @@ func (s *stenciler) stencilPath(bounds image.Rectangle, offset f32.Point, uv ima
 	texSize := f32.Point{X: float32(bounds.Dx()), Y: float32(bounds.Dy())}
 	scale := f32.Point{X: 2 / texSize.X, Y: 2 / texSize.Y}
 	orig := f32.Point{X: -1 - float32(bounds.Min.X)*2/texSize.X, Y: -1 - float32(bounds.Min.Y)*2/texSize.Y}
-	s.pipeline.uniforms.vert.transform = [4]float32{scale.X, scale.Y, orig.X, orig.Y}
-	s.pipeline.uniforms.vert.pathOffset = [2]float32{offset.X, offset.Y}
+	s.pipeline.uniforms.transform = [4]float32{scale.X, scale.Y, orig.X, orig.Y}
+	s.pipeline.uniforms.pathOffset = [2]float32{offset.X, offset.Y}
 	s.pipeline.pipeline.UploadUniforms(s.ctx)
 	// Draw in batches that fit in uint16 indices.
 	start := 0
@@ -390,21 +377,21 @@ func (c *coverer) cover(mat materialType, col f32color.RGBA, col1, col2 f32color
 	var uniforms *coverUniforms
 	switch mat {
 	case materialColor:
-		c.colUniforms.frag.color = col
-		uniforms = &c.colUniforms.vert.coverUniforms
+		c.colUniforms.color = col
+		uniforms = &c.colUniforms.coverUniforms
 	case materialLinearGradient:
-		c.linearGradientUniforms.frag.color1 = col1
-		c.linearGradientUniforms.frag.color2 = col2
+		c.linearGradientUniforms.color1 = col1
+		c.linearGradientUniforms.color2 = col2
 
 		t1, t2, t3, t4, t5, t6 := uvTrans.Elems()
-		c.linearGradientUniforms.vert.uvTransformR1 = [4]float32{t1, t2, t3, 0}
-		c.linearGradientUniforms.vert.uvTransformR2 = [4]float32{t4, t5, t6, 0}
-		uniforms = &c.linearGradientUniforms.vert.coverUniforms
+		c.linearGradientUniforms.uvTransformR1 = [4]float32{t1, t2, t3, 0}
+		c.linearGradientUniforms.uvTransformR2 = [4]float32{t4, t5, t6, 0}
+		uniforms = &c.linearGradientUniforms.coverUniforms
 	case materialTexture:
 		t1, t2, t3, t4, t5, t6 := uvTrans.Elems()
-		c.texUniforms.vert.uvTransformR1 = [4]float32{t1, t2, t3, 0}
-		c.texUniforms.vert.uvTransformR2 = [4]float32{t4, t5, t6, 0}
-		uniforms = &c.texUniforms.vert.coverUniforms
+		c.texUniforms.uvTransformR1 = [4]float32{t1, t2, t3, 0}
+		c.texUniforms.uvTransformR2 = [4]float32{t4, t5, t6, 0}
+		uniforms = &c.texUniforms.coverUniforms
 	}
 	uniforms.transform = [4]float32{scale.X, scale.Y, off.X, off.Y}
 	uniforms.uvCoverTransform = [4]float32{coverScale.X, coverScale.Y, coverOff.X, coverOff.Y}