From 27097612559b61f36a88f87647848bf2014fa85c Mon Sep 17 00:00:00 2001
From: "Sascha L. Teichmann" <sascha.teichmann@intevation.de>
Date: Sat, 20 Sep 2014 23:53:09 +0200
Subject: [PATCH] mtwebmapper: Added fast path for rendering RGBA sub base
 level tiles. Up to 5-10x faster.

---
 cmd/mtwebmapper/subbaseline.go | 79 ++++++++++++++++++++++++++++------
 1 file changed, 67 insertions(+), 12 deletions(-)

diff --git a/cmd/mtwebmapper/subbaseline.go b/cmd/mtwebmapper/subbaseline.go
index f28e033..4260d65 100644
--- a/cmd/mtwebmapper/subbaseline.go
+++ b/cmd/mtwebmapper/subbaseline.go
@@ -7,7 +7,6 @@ package main
 import (
 	"fmt"
 	"image"
-	"image/color"
 	"image/png"
 	"log"
 	"net/http"
@@ -105,7 +104,56 @@ func (sb *subBaseLine) ServeHTTP(rw http.ResponseWriter, r *http.Request) {
 
 func blowUp(src image.Image) *image.RGBA {
 
-	// TODO: Fast path for image.RGBA
+	// Fast path for RGBA -> RGBA
+	if rgba, ok := src.(*image.RGBA); ok {
+		return blowUpRGBA(rgba)
+	}
+
+	// Fallback
+	dst := image.NewRGBA(image.Rect(0, 0, 256, 256))
+
+	// fix point numbers x:8
+	dx, dy := src.Bounds().Dx(), src.Bounds().Dy()
+
+	bx, by := src.Bounds().Min.X<<8, src.Bounds().Min.Y<<8
+
+	//start := time.Now()
+
+	pix := dst.Pix
+
+	lineOfs := dst.PixOffset(0, 0) // Should be 0.
+
+	py := by
+	var r, g, b, a uint8
+	for y := 0; y < 256; y++ {
+		sy := (py >> 8) & 0xff
+		ox := -1
+		px := bx
+		ofs := lineOfs // Should not really b needed
+		lineOfs += dst.Stride
+		for x := 0; x < 256; x++ {
+			sx := (px >> 8) & 0xff
+			if sx != ox { // Minimize interface indirection access.
+				ox = sx
+				xr, xg, xb, xa := src.At(sx, sy).RGBA()
+				r, g, b, a = uint8(xr), uint8(xg), uint8(xb), uint8(xa)
+			}
+			pix[ofs] = r
+			pix[ofs+1] = g
+			pix[ofs+2] = b
+			pix[ofs+3] = a
+			ofs += 4
+			px += dx
+		}
+		py += dy
+	}
+
+	//log.Printf("Rendering took: %s", time.Since(start))
+
+	return dst
+}
+
+func blowUpRGBA(src *image.RGBA) *image.RGBA {
 
 	dst := image.NewRGBA(image.Rect(0, 0, 256, 256))
 
@@ -114,24 +162,31 @@ func blowUp(src image.Image) *image.RGBA {
 
 	bx, by := src.Bounds().Min.X<<8, src.Bounds().Min.Y<<8
 
+	//start := time.Now()
+
+	sPix := src.Pix
+	dPix := dst.Pix
+
 	py := by
-	for y := 0; y < 256; y++ {
+	// Assuming memory layout is packed 256*256*4 with stride of 4*256.
+	// for dLineOfs, dEnd := dst.PixOffset(0, 0), dst.PixOffset(0, 256); dLineOfs < dEnd; dLineOfs += dst.Stride {
+	for ofs := 0; ofs < 256*256*4; {
 		sy := (py >> 8) & 0xff
-		ox := -1
+		sLineOfs := src.PixOffset(0, sy)
 		px := bx
-		var col color.Color
-		for x := 0; x < 256; x++ {
-			sx := (px >> 8) & 0xff
-			if sx != ox { // Minimize interface indirection access.
-				ox = sx
-				col = src.At(sx, sy)
-			}
-			dst.Set(x, y, col)
+		// ofs := dLineOfs
+		for end := ofs + 4*256; ofs < end; ofs += 4 {
+			sOfs := sLineOfs + ((px >> 6) & 0x3fc)
 			px += dx
+			dPix[ofs] = sPix[sOfs]
+			dPix[ofs+1] = sPix[sOfs+1]
+			dPix[ofs+2] = sPix[sOfs+2]
+			dPix[ofs+3] = sPix[sOfs+3]
 		}
 		py += dy
 	}
 
+	//log.Printf("Rendering took: %s", time.Since(start))
 	return dst
 }