Merge pull request #89 from guzba/master

image minifyBy2 faster, mask minifyBy2, path fills take blendMode

src/pixie/blends.nim

@@ -1,9 +1,6 @@
 ## Blending modes.
 import chroma, math, common
 
-when defined(amd64) and not defined(pixieNoSimd):
-  import nimsimd/sse2
-
 # See https://www.w3.org/TR/compositing-1/
 # See https://www.khronos.org/registry/OpenGL/extensions/KHR/KHR_blend_equation_advanced.txt
 
@@ -34,12 +31,20 @@ type
     bmIntersectMask
     bmExcludeMask
 
-  Blender* = proc(a, b: ColorRGBA): ColorRGBA
+  Blender* = proc(backdrop, source: ColorRGBA): ColorRGBA
+
+when defined(amd64) and not defined(pixieNoSimd):
+  import nimsimd/sse2
+
+  type BlenderSimd* = proc(blackdrop, source: M128i): M128i
 
 when defined(release):
   {.push checks: off.}
 
-proc blendNormalPremultiplied*(backdrop, source: ColorRGBA): ColorRGBA {.inline.} =
+proc blendAlpha(backdrop, source: uint8): uint8 {.inline.} =
+  source + ((backdrop.uint32 * (255 - source)) div 255).uint8
+
+proc blendNormalPremultiplied*(backdrop, source: ColorRGBA): ColorRGBA =
   if backdrop.a == 0:
     return source
   if source.a == 255:
@@ -51,17 +56,21 @@ proc blendNormalPremultiplied*(backdrop, source: ColorRGBA): ColorRGBA {.inline.
   result.r = source.r + ((backdrop.r.uint32 * k) div 255).uint8
   result.g = source.g + ((backdrop.g.uint32 * k) div 255).uint8
   result.b = source.b + ((backdrop.b.uint32 * k) div 255).uint8
-  result.a = source.a + ((backdrop.a.uint32 * k) div 255).uint8
+  result.a = blendAlpha(backdrop.a, source.a)
+
+proc blenderPremultiplied*(blendMode: BlendMode): Blender =
+  case blendMode:
+  of bmNormal: blendNormalPremultiplied
+  else:
+    raise newException(PixieError, "No premultiplied blender for " & $blendMode)
 
 when defined(amd64) and not defined(pixieNoSimd):
-  proc blendNormalPremultiplied*(backdrop, source: M128i): M128i {.inline.} =
+  proc blendNormalPremultipliedSimd*(backdrop, source: M128i): M128i =
     let
       alphaMask = mm_set1_epi32(cast[int32](0xff000000))
       oddMask = mm_set1_epi16(cast[int16](0xff00))
       div255 = mm_set1_epi16(cast[int16](0x8081))
 
-    # Shortcuts didn't help (backdrop.a == 0, source.a == 0, source.a == 255)
-
     var
       sourceAlpha = mm_and_si128(source, alphaMask)
       backdropEven = mm_slli_epi16(backdrop, 8)
@@ -85,6 +94,13 @@ when defined(amd64) and not defined(pixieNoSimd):
       mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
     )
 
+  proc blenderSimd*(blendMode: BlendMode): BlenderSimd =
+    case blendMode:
+    of bmNormal: blendNormalPremultipliedSimd
+    else:
+      raise newException(PixieError, "No SIMD blender for " & $blendMode)
+
+
 when defined(release):
   {.pop.}
 
@@ -534,7 +550,7 @@ proc blendOverwrite(backdrop, source: ColorRGBA): ColorRGBA =
   source
 
 proc blender*(blendMode: BlendMode): Blender =
-  case blendMode
+  case blendMode:
   of bmNormal: blendNormal
   of bmDarken: blendDarken
   of bmMultiply: blendMultiply

src/pixie/images.nim

@@ -176,16 +176,28 @@ proc minifyBy2*(image: Image, power = 1): Image =
   if power == 0:
     return image.copy()
 
+  var src = image
   for i in 1 .. power:
-    result = newImage(image.width div 2, image.height div 2)
+    result = newImage(src.width div 2, src.height div 2)
     for y in 0 ..< result.height:
       for x in 0 ..< result.width:
-        var color =
+        let
-          image.getRgbaUnsafe(x * 2 + 0, y * 2 + 0).color / 4.0 +
+          a = src.getRgbaUnsafe(x * 2 + 0, y * 2 + 0)
-          image.getRgbaUnsafe(x * 2 + 1, y * 2 + 0).color / 4.0 +
+          b = src.getRgbaUnsafe(x * 2 + 1, y * 2 + 0)
-          image.getRgbaUnsafe(x * 2 + 1, y * 2 + 1).color / 4.0 +
+          c = src.getRgbaUnsafe(x * 2 + 1, y * 2 + 1)
-          image.getRgbaUnsafe(x * 2 + 0, y * 2 + 1).color / 4.0
+          d = src.getRgbaUnsafe(x * 2 + 0, y * 2 + 1)
-        result.setRgbaUnsafe(x, y, color.rgba)
+
+        let color = rgba(
+          ((a.r.uint32 + b.r + c.r + d.r) div 4).uint8,
+          ((a.g.uint32 + b.g + c.g + d.g) div 4).uint8,
+          ((a.b.uint32 + b.b + c.b + d.b) div 4).uint8,
+          ((a.a.uint32 + b.a + c.a + d.a) div 4).uint8
+        )
+
+        result.setRgbaUnsafe(x, y, color)
+
+    # Set src as this result for if we do another power
+    src = result
 
 proc magnifyBy2*(image: Image, power = 1): Image =
   ## Scales image image up by 2 ^ power.

src/pixie/masks.nim

@@ -63,5 +63,23 @@ proc `[]=`*(mask: Mask, x, y: int, value: uint8) {.inline.} =
   if mask.inside(x, y):
     mask.setValueUnsafe(x, y, value)
 
+proc minifyBy2*(mask: Mask, power = 1): Mask =
+  ## Scales the mask down by an integer scale.
+  if power < 0:
+    raise newException(PixieError, "Cannot minifyBy2 with negative power")
+  if power == 0:
+    return mask.copy()
+
+  for i in 1 .. power:
+    result = newMask(mask.width div 2, mask.height div 2)
+    for y in 0 ..< result.height:
+      for x in 0 ..< result.width:
+        let value =
+          mask.getValueUnsafe(x * 2 + 0, y * 2 + 0).uint32 +
+          mask.getValueUnsafe(x * 2 + 1, y * 2 + 0) +
+          mask.getValueUnsafe(x * 2 + 1, y * 2 + 1) +
+          mask.getValueUnsafe(x * 2 + 0, y * 2 + 1)
+        result.setValueUnsafe(x, y, (value div 4).uint8)
+
 when defined(release):
   {.pop.}

src/pixie/paths.nim

@@ -360,6 +360,9 @@ proc rect*(path: var Path, x, y, w, h: float32) =
   path.lineTo(x, y + h)
   path.closePath()
 
+proc rect*(path: var Path, pos: Vec2, wh: Vec2) {.inline.} =
+  path.rect(pos.x, pos.y, wh.x, wh.y)
+
 proc polygon*(path: var Path, x, y, size: float32, sides: int) =
   ## Draws a n sided regular polygon at (x, y) with size.
   path.moveTo(x + size * cos(0.0), y + size * sin(0.0))
@@ -886,7 +889,8 @@ proc fillShapes(
   image: Image,
   shapes: seq[seq[Vec2]],
   color: ColorRGBA,
-  windingRule: WindingRule
+  windingRule: WindingRule,
+  blendMode: BlendMode
 ) =
   let (topHalf, bottomHalf, fullHeight) =
     partitionSegments(shapes, image.height div 2)
@@ -898,6 +902,10 @@ proc fillShapes(
     startX = max(0, bounds.x.int)
     startY = max(0, bounds.y.int)
     stopY = min(image.height, (bounds.y + bounds.h).int)
+    blender = blendMode.blenderPremultiplied()
+
+  when defined(amd64) and not defined(pixieNoSimd):
+    let blenderSimd = blendMode.blenderSimd()
 
   var
     coverages = newSeq[uint8](image.width)
@@ -979,10 +987,7 @@ proc fillShapes(
           let
             index = image.dataIndex(x, y)
             backdrop = mm_loadu_si128(image.data[index].addr)
-          mm_storeu_si128(
+          mm_storeu_si128(image.data[index].addr, blenderSimd(backdrop, source))
-            image.data[index].addr,
-            blendNormalPremultiplied(backdrop, source)
-          )
 
         x += 4
 
@@ -1003,7 +1008,7 @@ proc fillShapes(
           source.a = ((color.a.uint16 * coverage) div 255).uint8
 
         let backdrop = image.getRgbaUnsafe(x, y)
-        image.setRgbaUnsafe(x, y, blendNormalPremultiplied(backdrop, source))
+        image.setRgbaUnsafe(x, y, blender(backdrop, source))
       inc x
 
 proc fillShapes(
@@ -1160,35 +1165,38 @@ proc fillPath*(
   image: Image,
   path: SomePath,
   color: ColorRGBA,
-  windingRule = wrNonZero
+  windingRule = wrNonZero,
+  blendMode = bmNormal
 ) {.inline.} =
-  image.fillShapes(parseSomePath(path), color, windingRule)
+  image.fillShapes(parseSomePath(path), color, windingRule, blendMode)
 
 proc fillPath*(
   image: Image,
   path: SomePath,
   color: ColorRGBA,
   pos: Vec2,
-  windingRule = wrNonZero
+  windingRule = wrNonZero,
+  blendMode = bmNormal
 ) =
   var shapes = parseSomePath(path)
   for shape in shapes.mitems:
     for segment in shape.mitems:
       segment += pos
-  image.fillShapes(shapes, color, windingRule)
+  image.fillShapes(shapes, color, windingRule, blendMode)
 
 proc fillPath*(
   image: Image,
   path: SomePath,
   color: ColorRGBA,
   mat: Mat3,
-  windingRule = wrNonZero
+  windingRule = wrNonZero,
+  blendMode = bmNormal
 ) =
   var shapes = parseSomePath(path)
   for shape in shapes.mitems:
     for segment in shape.mitems:
       segment = mat * segment
-  image.fillShapes(shapes, color, windingRule)
+  image.fillShapes(shapes, color, windingRule, blendMode)
 
 proc fillPath*(
   mask: Mask,
@@ -1226,14 +1234,15 @@ proc strokePath*(
   path: SomePath,
   color: ColorRGBA,
   strokeWidth = 1.0,
-  windingRule = wrNonZero
+  windingRule = wrNonZero,
+  blendMode = bmNormal
 ) =
   let strokeShapes = strokeShapes(
     parseSomePath(path),
     strokeWidth,
     windingRule
   )
-  image.fillShapes(strokeShapes, color, windingRule)
+  image.fillShapes(strokeShapes, color, windingRule, blendMode)
 
 proc strokePath*(
   image: Image,
@@ -1241,7 +1250,8 @@ proc strokePath*(
   color: ColorRGBA,
   strokeWidth = 1.0,
   pos: Vec2,
-  windingRule = wrNonZero
+  windingRule = wrNonZero,
+  blendMode = bmNormal
 ) =
   var strokeShapes = strokeShapes(
     parseSomePath(path),
@@ -1251,7 +1261,7 @@ proc strokePath*(
   for shape in strokeShapes.mitems:
     for segment in shape.mitems:
       segment += pos
-  image.fillShapes(strokeShapes, color, windingRule)
+  image.fillShapes(strokeShapes, color, windingRule, blendMode)
 
 proc strokePath*(
   image: Image,
@@ -1259,7 +1269,8 @@ proc strokePath*(
   color: ColorRGBA,
   strokeWidth = 1.0,
   mat: Mat3,
-  windingRule = wrNonZero
+  windingRule = wrNonZero,
+  blendMode = bmNormal
 ) =
   var strokeShapes = strokeShapes(
     parseSomePath(path),
@@ -1269,7 +1280,7 @@ proc strokePath*(
   for shape in strokeShapes.mitems:
     for segment in shape.mitems:
       segment = mat * segment
-  image.fillShapes(strokeShapes, color, windingRule)
+  image.fillShapes(strokeShapes, color, windingRule, blendMode)
 
 proc strokePath*(
   mask: Mask,

tests/benchmark_images.nim

@@ -13,6 +13,13 @@ timeIt "fill_rgba":
 timeIt "subImage":
   keep a.subImage(0, 0, 256, 256)
 
+# timeIt "superImage":
+#   discard
+
+timeIt "minifyBy2":
+  let minified = a.minifyBy2()
+  doAssert minified[0, 0] == rgba(63, 127, 191, 191)
+
 timeIt "invert":
   a.invert()
   keep(a)

tests/test_images.nim

@@ -79,3 +79,15 @@ block:
     a = readImage("tests/images/flipped1.png")
     b = a.superImage(45, 45, 20, 20)
   b.writeFile("tests/images/superimage6.png")
+
+block:
+  let
+    a = readImage("tests/images/flipped1.png")
+    b = a.minifyBy2()
+  b.writeFile("tests/images/minifiedBy2.png")
+
+block:
+  let
+    a = readImage("tests/images/flipped1.png")
+    b = a.minifyBy2(2)
+  b.writeFile("tests/images/minifiedBy4.png")

tests/test_masks.nim

@@ -0,0 +1,19 @@
+import pixie, pixie/fileformats/png
+
+block:
+  let
+    mask = newMask(100, 100)
+    r = 10.0
+    x = 10.0
+    y = 10.0
+    h = 80.0
+    w = 80.0
+  var path: Path
+  path.moveTo(x + r, y)
+  path.arcTo(x + w, y, x + w, y + h, r)
+  path.arcTo(x + w, y + h, x, y + h, r)
+  path.arcTo(x, y + h, x, y, r)
+  path.arcTo(x, y, x + w, y, r)
+  mask.fillPath(path)
+
+  writeFile("tests/images/masks/maskMinified.png", mask.minifyBy2().encodePng())