Performance optimization for the path without a shader

Fixes hajimehoshi#1355
momodezhou · Sep 20, 2020 · f3ef2e2 · f3ef2e2
1 parent cb73230
commit f3ef2e2
Show file tree

Hide file tree

Showing 4 changed files with 59 additions and 48 deletions.
diff --git a/internal/buffered/image.go b/internal/buffered/image.go
@@ -265,26 +265,24 @@ func (i *Image) DrawTriangles(srcs [graphics.ShaderImageNum]*Image, vertices []f
 		}
 	}
 
-	for _, src := range srcs {
-		if src == nil {
-			continue
-		}
-		src.resolvePendingPixels(true)
-	}
-	i.resolvePendingPixels(false)
-
 	var s *shareable.Shader
-	if shader != nil {
-		s = shader.shader
-	}
-
 	var imgs [graphics.ShaderImageNum]*shareable.Image
-	for i, img := range srcs {
-		if img == nil {
-			continue
+	if shader == nil {
+		// Fast path for rendering without a shader (#1355).
+		img := srcs[0]
+		img.resolvePendingPixels(true)
+		imgs[0] = img.img
+	} else {
+		for i, img := range srcs {
+			if img == nil {
+				continue
+			}
+			img.resolvePendingPixels(true)
+			imgs[i] = img.img
 		}
-		imgs[i] = img.img
+		s = shader.shader
 	}
+	i.resolvePendingPixels(false)
 
 	i.img.DrawTriangles(imgs, vertices, indices, colorm, mode, filter, address, sourceRegion, subimageOffsets, s, uniforms)
 	i.invalidatePendingPixels()

diff --git a/internal/graphicscommand/image.go b/internal/graphicscommand/image.go
@@ -157,14 +157,23 @@ func (i *Image) DrawTriangles(srcs [graphics.ShaderImageNum]*Image, offsets [gra
 		}
 	}
 
-	for _, src := range srcs {
-		if src == nil {
-			continue
-		}
-		if src.screen {
+	if shader == nil {
+		// Fast path for rendering without a shader (#1355).
+		img := srcs[0]
+		if img.screen {
 			panic("graphicscommand: the screen image cannot be the rendering source")
 		}
-		src.resolveBufferedReplacePixels()
+		img.resolveBufferedReplacePixels()
+	} else {
+		for _, src := range srcs {
+			if src == nil {
+				continue
+			}
+			if src.screen {
+				panic("graphicscommand: the screen image cannot be the rendering source")
+			}
+			src.resolveBufferedReplacePixels()
+		}
 	}
 	i.resolveBufferedReplacePixels()
 

diff --git a/internal/restorable/image.go b/internal/restorable/image.go
@@ -389,17 +389,20 @@ func (i *Image) DrawTriangles(srcs [graphics.ShaderImageNum]*Image, offsets [gra
 	} else {
 		i.appendDrawTrianglesHistory(srcs, offsets, vertices, indices, colorm, mode, filter, address, sourceRegion, shader, uniforms)
 	}
-	var s *graphicscommand.Shader
-	if shader != nil {
-		s = shader.shader
-	}
 
+	var s *graphicscommand.Shader
 	var imgs [graphics.ShaderImageNum]*graphicscommand.Image
-	for i, src := range srcs {
-		if src == nil {
-			continue
+	if shader == nil {
+		// Fast path for rendering without a shader (#1355).
+		imgs[0] = srcs[0].image
+	} else {
+		for i, src := range srcs {
+			if src == nil {
+				continue
+			}
+			imgs[i] = src.image
 		}
-		imgs[i] = src.image
+		s = shader.shader
 	}
 	i.image.DrawTriangles(imgs, offsets, vertices, indices, colorm, mode, filter, address, sourceRegion, s, uniforms)
 }

diff --git a/internal/shareable/image.go b/internal/shareable/image.go
@@ -354,27 +354,28 @@ func (i *Image) DrawTriangles(srcs [graphics.ShaderImageNum]*Image, vertices []f
 	}
 
 	var offsets [graphics.ShaderImageNum - 1][2]float32
-	for i, subimageOffset := range subimageOffsets {
-		src := srcs[i+1]
-		if src == nil {
-			continue
-		}
-		ox, oy, _, _ := src.regionWithPadding()
-		offsets[i][0] = float32(ox) + paddingSize - oxf + subimageOffset[0]
-		offsets[i][1] = float32(oy) + paddingSize - oyf + subimageOffset[1]
-	}
-
 	var s *restorable.Shader
-	if shader != nil {
-		s = shader.shader
-	}
-
 	var imgs [graphics.ShaderImageNum]*restorable.Image
-	for i, src := range srcs {
-		if src == nil {
-			continue
+	if shader == nil {
+		// Fast path for rendering without a shader (#1355).
+		imgs[0] = srcs[0].backend.restorable
+	} else {
+		for i, subimageOffset := range subimageOffsets {
+			src := srcs[i+1]
+			if src == nil {
+				continue
+			}
+			ox, oy, _, _ := src.regionWithPadding()
+			offsets[i][0] = float32(ox) + paddingSize - oxf + subimageOffset[0]
+			offsets[i][1] = float32(oy) + paddingSize - oyf + subimageOffset[1]
+		}
+		s = shader.shader
+		for i, src := range srcs {
+			if src == nil {
+				continue
+			}
+			imgs[i] = src.backend.restorable
 		}
-		imgs[i] = src.backend.restorable
 	}
 
 	i.backend.restorable.DrawTriangles(imgs, offsets, vertices, indices, colorm, mode, filter, address, sourceRegion, s, uniforms)