From 32c3072e0a92c04977aba2198447cc61b75bdc2f Mon Sep 17 00:00:00 2001
From: Ryan Oldenburg <ryan@guzba.com>
Date: Sun, 27 Jun 2021 15:55:13 -0500
Subject: [PATCH] get smooth a bit smarter (faster in solid color case)

---
 src/pixie/images.nim            | 18 ++++++++++++---
 src/pixie/masks.nim             | 10 ++++++++-
 tests/benchmark_images_draw.nim | 39 ++++++++++++++++++++++++++++++---
 3 files changed, 60 insertions(+), 7 deletions(-)

diff --git a/src/pixie/images.nim b/src/pixie/images.nim
index b1bfe7e..7c5fd88 100644
--- a/src/pixie/images.nim
+++ b/src/pixie/images.nim
@@ -535,11 +535,18 @@ proc getRgbaSmooth*(image: Image, x, y: float32, wrapped = false): ColorRGBX =
     x0y1 = image[x0, y1]
     x1y1 = image[x1, y1]
 
-  let
+  var topMix = x0y0
+  if xFractional > 0 and x0y0 != x1y0:
     topMix = lerp(x0y0, x1y0, xFractional)
+
+  var bottomMix = x0y1
+  if xFractional > 0 and x0y1 != x1y1:
     bottomMix = lerp(x0y1, x1y1, xFractional)
 
-  lerp(topMix, bottomMix, yFractional)
+  if yFractional != 0 and topMix != bottomMix:
+    lerp(topMix, bottomMix, yFractional)
+  else:
+    topMix
 
 proc drawCorrect(
   a, b: Image | Mask, mat = mat3(), tiled = false, blendMode = bmNormal
@@ -678,9 +685,11 @@ proc drawUber(a, b: Image | Mask, mat = mat3(), blendMode = bmNormal) =
         zeroMem(a.data[a.dataIndex(0, y)].addr, 4 * xMin)
 
     if smooth:
+      var srcPos = p + dx * xMin.float32 + dy * y.float32
+      srcPos = vec2(max(0, srcPos.x), max(0, srcPos.y))
+
       for x in xMin ..< xMax:
         let
-          srcPos = p + dx * x.float32 + dy * y.float32
           xFloat = srcPos.x - h
           yFloat = srcPos.y - h
         when type(a) is Image:
@@ -701,6 +710,9 @@ proc drawUber(a, b: Image | Mask, mat = mat3(), blendMode = bmNormal) =
           else: # b is a Mask
             let sample = b.getValueSmooth(xFloat, yFloat)
           a.setValueUnsafe(x, y, masker(backdrop, sample))
+
+        srcPos += dx
+
     else:
       var x = xMin
       when defined(amd64) and not defined(pixieNoSimd):
diff --git a/src/pixie/masks.nim b/src/pixie/masks.nim
index e81d16a..fdcaa01 100644
--- a/src/pixie/masks.nim
+++ b/src/pixie/masks.nim
@@ -173,10 +173,18 @@ proc getValueSmooth*(mask: Mask, x, y: float32): uint8 =
     x0y1 = mask[x0, y1]
     x1y1 = mask[x1, y1]
 
+  var topMix = x0y0
+  if xFractional > 0 and x0y0 != x1y0:
     topMix = lerp(x0y0, x1y0, xFractional)
+
+  var bottomMix = x0y1
+  if xFractional > 0 and x0y1 != x1y1:
     bottomMix = lerp(x0y1, x1y1, xFractional)
 
-  lerp(topMix, bottomMix, yFractional)
+  if yFractional != 0 and topMix != bottomMix:
+    lerp(topMix, bottomMix, yFractional)
+  else:
+    topMix
 
 proc spread*(mask: Mask, spread: float32) =
   ## Grows the mask by spread.
diff --git a/tests/benchmark_images_draw.nim b/tests/benchmark_images_draw.nim
index bf4a925..3ae2443 100644
--- a/tests/benchmark_images_draw.nim
+++ b/tests/benchmark_images_draw.nim
@@ -40,7 +40,40 @@ block:
   a.fill(rgba(255, 0, 0, 255))
   b.fill(rgba(0, 255, 0, 255))
 
-  timeIt "draw big-on-bigger Smooth bmNormal":
+  timeIt "draw [scale 0.5]":
+    a.draw(b, translate(vec2(25, 25)) * scale(vec2(0.5, 0.5)), bmNormal)
+    keep(b)
+
+block:
+  let
+    a = newImage(1000, 1000)
+    b = newImage(500, 500)
+  a.fill(rgba(255, 0, 0, 255))
+  b.fill(rgba(0, 255, 0, 255))
+
+  timeIt "draw Smooth [x translate]":
+    a.draw(b, translate(vec2(25.2, 0)), bmNormal)
+    keep(b)
+
+block:
+  let
+    a = newImage(1000, 1000)
+    b = newImage(500, 500)
+  a.fill(rgba(255, 0, 0, 255))
+  b.fill(rgba(0, 255, 0, 255))
+
+  timeIt "draw Smooth [y translate]":
+    a.draw(b, translate(vec2(0, 25.2)), bmNormal)
+    keep(b)
+
+block:
+  let
+    a = newImage(1000, 1000)
+    b = newImage(500, 500)
+  a.fill(rgba(255, 0, 0, 255))
+  b.fill(rgba(0, 255, 0, 255))
+
+  timeIt "draw Smooth [x + y translate]":
     a.draw(b, translate(vec2(25.2, 25.2)), bmNormal)
     keep(b)
 
@@ -51,8 +84,8 @@ block:
   a.fill(rgba(255, 0, 0, 255))
   b.fill(rgba(0, 255, 0, 255))
 
-  timeIt "draw big-on-bigger bmNormal scale(0.5)":
-    a.draw(b, translate(vec2(25, 25)) * scale(vec2(0.5, 0.5)), bmNormal)
+  timeIt "draw Smooth [rotate 45 deg]":
+    a.draw(b, translate(vec2(0, 500)) * rotate(toRadians(45)), bmNormal)
     keep(b)
 
 block: