fix and re-enable fill optimization

src/pixie/images.nim

@@ -792,16 +792,15 @@ proc drawUber(
           sy = srcPos.y.int
         var sx = srcPos.x.int
 
-        # TODO Fix
+        when type(a) is Image and type(b) is Image:
-        # when type(a) is Image and type(b) is Image:
+          if blendMode in {bmNormal, bmOverwrite} and
-        #   if blendMode in {bmNormal, bmOverwrite} and
+            isOpaque(b.data, b.dataIndex(sx, sy), xStop - xStart):
-        #     isOpaque(b.data, b.dataIndex(sx, sy), xStop - sx):
+              copyMem(
-        #       copyMem(
+                a.data[a.dataIndex(x, y)].addr,
-        #         a.data[a.dataIndex(x, y)].addr,
+                b.data[b.dataIndex(sx, sy)].addr,
-        #         b.data[b.dataIndex(sx, sy)].addr,
+                (xStop - xStart) * 4
-        #         (xStop - xStart) * 4
+              )
-        #       )
+              continue
-        #       continue
 
         when defined(amd64) and not defined(pixieNoSimd):
           case blendMode:

src/pixie/internal.nim

@@ -139,24 +139,23 @@ proc toPremultipliedAlpha*(data: var seq[ColorRGBA | ColorRGBX]) {.raises: [].}
 proc isOpaque*(data: var seq[ColorRGBX], start, len: int): bool =
   result = true
 
-  var i: int = start
+  var i = start
-  # TODO FIX:
+  when defined(amd64) and not defined(pixieNoSimd):
-  # when defined(amd64) and not defined(pixieNoSimd):
+    let
-  #   let
+      vec255 = mm_set1_epi32(cast[int32](uint32.high))
-  #     vec255 = mm_set1_epi32(cast[int32](uint32.high))
+      colorMask = mm_set1_epi32(cast[int32]([255.uint8, 255, 255, 0]))
-  #     colorMask = mm_set1_epi32(cast[int32]([255.uint8, 255, 255, 0]))
+    for _ in start ..< (start + len) div 16:
-  #   for _ in start div 16 ..< (start + len) div 16:
+      let
-  #     let
+        values0 = mm_loadu_si128(data[i + 0].addr)
-  #       values0 = mm_loadu_si128(data[i + 0].addr)
+        values1 = mm_loadu_si128(data[i + 4].addr)
-  #       values1 = mm_loadu_si128(data[i + 4].addr)
+        values2 = mm_loadu_si128(data[i + 8].addr)
-  #       values2 = mm_loadu_si128(data[i + 8].addr)
+        values3 = mm_loadu_si128(data[i + 12].addr)
-  #       values3 = mm_loadu_si128(data[i + 12].addr)
+        values01 = mm_and_si128(values0, values1)
-  #       values01 = mm_and_si128(values0, values1)
+        values23 = mm_and_si128(values2, values3)
-  #       values23 = mm_and_si128(values2, values3)
+        values = mm_or_si128(mm_and_si128(values01, values23), colorMask)
-  #       values = mm_or_si128(mm_and_si128(values01, values23), colorMask)
+      if mm_movemask_epi8(mm_cmpeq_epi8(values, vec255)) != 0xffff:
-  #     if mm_movemask_epi8(mm_cmpeq_epi8(values, vec255)) != 0xffff:
+        return false
-  #       return false
+      i += 16
-  #     i += 16
 
   for j in i ..< start + len:
     if data[j].a != 255:

tests/test_images_draw.nim

@@ -270,3 +270,49 @@ block:
     rock = readImage("tests/images/rock.png")
     minified = rock.minifyBy2(2)
   doDiff(minified, "rock_minified2")
+
+block:
+  let pathStr = """
+  M 0 0
+  L 20 0
+  L 20 20
+  L 0 20
+  z
+  """
+  let
+    image = newImage(20, 20)
+    strokeImage = newImage(20, 20)
+  image.fillPath(pathStr, color(1.0, 0.5, 0.25, 1.0))
+  strokeImage.strokePath(pathStr, color(1, 1, 1, 1), strokeWidth = 4)
+  image.draw(strokeImage)
+
+  image.writeFile("tests/images/fillOptimization.png")
+  doAssert image[10, 10] == rgbx(255, 127, 63, 255)
+
+block:
+  let a = newImage(100, 100)
+  a.fill(color(1, 1, 1, 1))
+
+
+  let draws = [
+    # Overlaps in bounds
+    (vec2(-50, -50), color(1, 0, 0, 1)),
+    (vec2(50, -50), color(1, 0, 0, 1)),
+    (vec2(50, 50), color(1, 0, 0, 1)),
+    (vec2(-50, 50), color(1, 0, 0, 1)),
+    # Out of bounds
+    (vec2(-100, 00), color(1, 0, 0, 1)),
+    (vec2(0, -100), color(1, 0, 0, 1)),
+    (vec2(100, 00), color(1, 0, 0, 1)),
+    (vec2(0, 100), color(1, 0, 0, 1))
+  ]
+  for (translation, color) in draws:
+    let
+      b = newImage(100, 100)
+      path = newPath()
+    path.rect(0, 0, b.width.float32, b.height.float32)
+    b.strokePath(path, color, strokeWidth = 20)
+
+    a.draw(b, translate(translation))
+
+  a.writeFile("tests/images/fillOptimization2.png")

tests/test_paths.nim

@@ -1,24 +1,5 @@
 import chroma, pixie, pixie/fileformats/png, strformat
 
-
-block:
-  let pathStr = """
-  M 0 0
-  L 20 0
-  L 20 20
-  L 0 20
-  z
-  """
-  let
-    image = newImage(20, 20)
-    strokeImage = newImage(20, 20)
-  image.fillPath(pathStr, color(1.0, 0.5, 0.25, 1.0))
-  strokeImage.strokePath(pathStr, color(1, 1, 1, 1), strokeWidth = 4)
-  image.draw(strokeImage)
-
-  image.writeFile("tests/paths/fillOptimization.png")
-  doAssert image[10, 10] == rgbx(255, 127, 63, 255)
-
 block:
   let pathStr = """
   m 1 2 3 4 5 6