Reland: graphics: Reuse vertices backend and reduce allocating

This is a reland of 74f1e5519f.

The vertex slice is now copied before saving as the drawing
history items.

internal/restorable/image.go

@@ -414,12 +414,13 @@ func (i *Image) appendDrawTrianglesHistory(image *Image, vertices []float32, ind
 	// All images must be resolved and not stale each after frame.
 	// So we don't have to care if image is stale or not here.
 
-	// vertices is generated at ebiten package and doesn't have to be copied so far.
+	vs := make([]float32, len(vertices))
+	copy(vs, vertices)
 	is := make([]uint16, len(indices))
 	copy(is, indices)
 	item := &drawTrianglesHistoryItem{
 		image:    image,
-		vertices: vertices,
+		vertices: vs,
 		indices:  is,
 		colorm:   colorm,
 		mode:     mode,

mipmap.go

@@ -121,9 +121,14 @@ func (m *mipmap) drawImage(src *mipmap, bounds image.Rectangle, geom *GeoM, colo
 		colorm = nil
 	}
 
+	screen := filter == driver.FilterScreen
+	if screen && level != 0 {
+		panic("ebiten: Mipmap must not be used when the filter is FilterScreen")
+	}
+
 	a, b, c, d, tx, ty := geom.elements()
 	if level == 0 {
-		vs := quadVertices(bounds.Min.X, bounds.Min.Y, bounds.Max.X, bounds.Max.Y, a, b, c, d, tx, ty, cr, cg, cb, ca)
+		vs := quadVertices(bounds.Min.X, bounds.Min.Y, bounds.Max.X, bounds.Max.Y, a, b, c, d, tx, ty, cr, cg, cb, ca, screen)
 		is := graphics.QuadIndices()
 		m.orig.DrawTriangles(src.orig, vs, is, colorm, mode, filter, driver.AddressClampToZero)
 	} else if buf := src.level(bounds, level); buf != nil {
@@ -133,7 +138,7 @@ func (m *mipmap) drawImage(src *mipmap, bounds image.Rectangle, geom *GeoM, colo
 		b *= s
 		c *= s
 		d *= s
-		vs := quadVertices(0, 0, w, h, a, b, c, d, tx, ty, cr, cg, cb, ca)
+		vs := quadVertices(0, 0, w, h, a, b, c, d, tx, ty, cr, cg, cb, ca, false)
 		is := graphics.QuadIndices()
 		m.orig.DrawTriangles(buf, vs, is, colorm, mode, filter, driver.AddressClampToZero)
 	}
@@ -149,7 +154,7 @@ func (m *mipmap) drawTriangles(src *mipmap, bounds image.Rectangle, vertices []V
 	// TODO: Needs boundary check optimization?
 	// See https://go101.org/article/bounds-check-elimination.html
 
-	vs := vertexSlice(len(vertices))
+	vs := vertexSlice(len(vertices), false)
 	for i, v := range vertices {
 		vs[i*graphics.VertexFloatNum] = v.DstX
 		vs[i*graphics.VertexFloatNum+1] = v.DstY
@@ -194,7 +199,7 @@ func (m *mipmap) level(r image.Rectangle, level int) *buffered.Image {
 	switch {
 	case level == 1:
 		src = m.orig
-		vs = quadVertices(r.Min.X, r.Min.Y, r.Max.X, r.Max.Y, 0.5, 0, 0, 0.5, 0, 0, 1, 1, 1, 1)
+		vs = quadVertices(r.Min.X, r.Min.Y, r.Max.X, r.Max.Y, 0.5, 0, 0, 0.5, 0, 0, 1, 1, 1, 1, false)
 		filter = driver.FilterLinear
 	case level > 1:
 		src = m.level(r, level-1)
@@ -203,11 +208,11 @@ func (m *mipmap) level(r image.Rectangle, level int) *buffered.Image {
 			return nil
 		}
 		w, h := sizeForLevel(r.Dx(), r.Dy(), level-1)
-		vs = quadVertices(0, 0, w, h, 0.5, 0, 0, 0.5, 0, 0, 1, 1, 1, 1)
+		vs = quadVertices(0, 0, w, h, 0.5, 0, 0, 0.5, 0, 0, 1, 1, 1, 1, false)
 		filter = driver.FilterLinear
 	case level == -1:
 		src = m.orig
-		vs = quadVertices(r.Min.X, r.Min.Y, r.Max.X, r.Max.Y, 2, 0, 0, 2, 0, 0, 1, 1, 1, 1)
+		vs = quadVertices(r.Min.X, r.Min.Y, r.Max.X, r.Max.Y, 2, 0, 0, 2, 0, 0, 1, 1, 1, 1, false)
 		filter = driver.FilterNearest
 	case level < -1:
 		src = m.level(r, level+1)
@@ -216,7 +221,7 @@ func (m *mipmap) level(r image.Rectangle, level int) *buffered.Image {
 			return nil
 		}
 		w, h := sizeForLevel(r.Dx(), r.Dy(), level+1)
-		vs = quadVertices(0, 0, w, h, 2, 0, 0, 2, 0, 0, 1, 1, 1, 1)
+		vs = quadVertices(0, 0, w, h, 2, 0, 0, 2, 0, 0, 1, 1, 1, 1, false)
 		filter = driver.FilterNearest
 	default:
 		panic(fmt.Sprintf("ebiten: invalid level: %d", level))
@@ -292,6 +297,10 @@ func (m *mipmap) mipmapLevel(geom *GeoM, width, height int, filter driver.Filter
 		panic("ebiten: dst must be non zero at mipmapLevel")
 	}
 
+	if filter == driver.FilterScreen {
+		return 0
+	}
+
 	// Use 'negative' mipmap to render edges correctly (#611, #907).
 	// It looks like 128 is the enlargement factor that causes edge missings to pass the test TestImageStretch.
 	const tooBigScale = 128

vertex.go

@@ -21,7 +21,9 @@ import (
 )
 
 var (
-	theVerticesBackend = &verticesBackend{}
+	theVerticesBackend = &verticesBackend{
+		backend: make([]float32, graphics.VertexFloatNum*1024),
+	}
 )
 
 type verticesBackend struct {
@@ -30,42 +32,42 @@ type verticesBackend struct {
 	m       sync.Mutex
 }
 
-func (v *verticesBackend) slice(n int) []float32 {
+func (v *verticesBackend) slice(n int, last bool) []float32 {
 	v.m.Lock()
 
 	need := n * graphics.VertexFloatNum
-	if v.head+need > len(v.backend) {
+	if l := len(v.backend); v.head+need > l {
-		v.backend = nil
+		for v.head+need > l {
+			l *= 2
+		}
+		v.backend = make([]float32, l)
 		v.head = 0
 	}
 
-	if v.backend == nil {
-		l := 1024
-		if n > l {
-			l = n
-		}
-		v.backend = make([]float32, graphics.VertexFloatNum*l)
-	}
-
 	s := v.backend[v.head : v.head+need]
+	if last {
+		// If last is true, the vertices backend is sent to GPU and it is fine to reuse the slice.
+		v.head = 0
+	} else {
 		v.head += need
+	}
 
 	v.m.Unlock()
 	return s
 }
 
-func vertexSlice(n int) []float32 {
+func vertexSlice(n int, last bool) []float32 {
-	return theVerticesBackend.slice(n)
+	return theVerticesBackend.slice(n, last)
 }
 
-func quadVertices(sx0, sy0, sx1, sy1 int, a, b, c, d, tx, ty float32, cr, cg, cb, ca float32) []float32 {
+func quadVertices(sx0, sy0, sx1, sy1 int, a, b, c, d, tx, ty float32, cr, cg, cb, ca float32, last bool) []float32 {
 	x := float32(sx1 - sx0)
 	y := float32(sy1 - sy0)
 	ax, by, cx, dy := a*x, b*y, c*x, d*y
 	u0, v0, u1, v1 := float32(sx0), float32(sy0), float32(sx1), float32(sy1)
 
 	// This function is very performance-sensitive and implement in a very dumb way.
-	vs := vertexSlice(4)
+	vs := vertexSlice(4, last)
 	_ = vs[:48]
 
 	// For each values, see the comment at shareable.(*Image).DrawTriangles.