Merge pull request #103 from guzba/master

sharpen mask, some mask blends, draw y optimization

src/pixie/blends.nim

@@ -466,12 +466,13 @@ proc blender*(blendMode: BlendMode): Blender =
   of bmSubtractMask: blendSubtractMask
   of bmIntersectMask: blendIntersectMask
   of bmExcludeMask: blendExcludeMask
-  else:
-    # blendWhite
-    # blendNormal
-    raise newException(PixieError, "No blender for " & $blendMode)
+
+proc maskNormal(backdrop, source: uint8): uint8 =
+  ## Blending masks
+  blendAlpha(backdrop, source)
 
 proc maskMask(backdrop, source: uint8): uint8 =
+  ## Masking masks
   ((backdrop.uint32 * source) div 255).uint8
 
 proc maskSubtract(backdrop, source: uint8): uint8 =
@@ -488,6 +489,7 @@ proc maskOverwrite(backdrop, source: uint8): uint8 =
 
 proc masker*(blendMode: BlendMode): Masker =
   case blendMode:
+  of bmNormal: maskNormal
   of bmMask: maskMask
   of bmOverwrite: maskOverwrite
   of bmSubtractMask: maskSubtract
@@ -499,7 +501,9 @@ proc masker*(blendMode: BlendMode): Masker =
 when defined(amd64) and not defined(pixieNoSimd):
   import nimsimd/sse2
 
-  type BlenderSimd* = proc(blackdrop, source: M128i): M128i
+  type
+    BlenderSimd* = proc(blackdrop, source: M128i): M128i
+    MaskerSimd* = proc(blackdrop, source: M128i): M128i
 
   proc blendNormalSimd*(backdrop, source: M128i): M128i =
     let
@@ -540,5 +544,51 @@ when defined(amd64) and not defined(pixieNoSimd):
     else:
       raise newException(PixieError, "No SIMD blender for " & $blendMode)
 
+  proc maskNormalSimd*(backdrop, source: M128i): M128i =
+    ## Blending masks
+    let
+      oddMask = mm_set1_epi16(cast[int16](0xff00))
+      v255high = mm_set1_epi16(cast[int16](255.uint16 shl 8))
+      div255 = mm_set1_epi16(cast[int16](0x8081))
+
+    var
+      sourceEven = mm_slli_epi16(mm_andnot_si128(oddMask, source), 8)
+      sourceOdd = mm_and_si128(source, oddMask)
+
+    let
+      evenK = mm_sub_epi16(v255high, sourceEven)
+      oddK = mm_sub_epi16(v255high, sourceOdd)
+
+    var
+      backdropEven = mm_slli_epi16(mm_andnot_si128(oddMask, backdrop), 8)
+      backdropOdd = mm_and_si128(backdrop, oddMask)
+
+    # backdrop * k
+    backdropEven = mm_mulhi_epu16(backdropEven, evenK)
+    backdropOdd = mm_mulhi_epu16(backdropOdd, oddK)
+
+    # div 255
+    backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
+    backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
+
+    # Shift from high to low bits
+    sourceEven = mm_srli_epi16(sourceEven, 8)
+    sourceOdd = mm_srli_epi16(sourceOdd, 8)
+
+    var
+      blendedEven = mm_add_epi16(sourceEven, backdropEven)
+      blendedOdd = mm_add_epi16(sourceOdd, backdropOdd)
+
+    blendedOdd = mm_slli_epi16(blendedOdd, 8)
+
+    mm_or_si128(blendedEven, blendedOdd)
+
+  proc maskerSimd*(blendMode: BlendMode): MaskerSimd =
+    case blendMode:
+    of bmNormal: maskNormalSimd
+    of bmOverwrite: blendOverwriteSimd
+    else:
+      raise newException(PixieError, "No SIMD masker for " & $blendMode)
+
 when defined(release):
   {.pop.}

src/pixie/images.nim

@@ -612,15 +612,6 @@ proc blur*(target: Image | Mask, radius: float32) =
 when defined(release):
   {.pop.}
 
-proc sharpOpacity*(image: Image) =
-  ## Sharpens the opacity to extreme.
-  ## A = 0 stays 0. Anything else turns into 255.
-  for rgba in image.data.mitems:
-    if rgba.a == 0:
-      rgba = rgba(0, 0, 0, 0)
-    else:
-      rgba = rgba(255, 255, 255, 255)
-
 proc drawUber(
   a, b: Image,
   p, dx, dy: Vec2,
@@ -629,12 +620,29 @@ proc drawUber(
   smooth: bool
 ) =
   let blender = blendMode.blender()
-  for y in 0 ..< a.height:
+
+  # Determine where we should start and stop drawing in the y dimension
+  var yMin, yMax: int
+  if blendMode == bmIntersectMask:
+    yMin = 0
+    yMax = a.height
+  else:
+    yMin = a.height
+    yMax = 0
+    for segment in perimeter:
+        yMin = min(yMin, segment.at.y.floor.int)
+        yMax = max(yMax, segment.at.y.ceil.int)
+
+  yMin = yMin.clamp(0, a.height)
+  yMax = yMax.clamp(0, a.height)
+
+  for y in yMin ..< yMax:
+    # Determine where we should start and stop drawing in the x dimension
     var
       xMin = a.width
       xMax = 0
     for yOffset in [0.float32, 1]:
-      var scanLine = Line(
+      let scanLine = Line(
         a: vec2(-1000, y.float32 + yOffset),
         b: vec2(1000, y.float32 + yOffset)
       )

src/pixie/masks.nim

@@ -1,5 +1,8 @@
 import common, vmath, system/memory
 
+when defined(amd64) and not defined(pixieNoSimd):
+  import nimsimd/sse2
+
 type
   Mask* = ref object
     ## Mask object that holds mask opacity data.
@@ -132,5 +135,23 @@ proc spread*(mask: Mask, spread: float32) =
               break blurBox
       mask.setValueUnsafe(x, y, maxValue)
 
+proc ceil*(mask: Mask) =
+  ## A value of 0 stays 0. Anything else turns into 255.
+  var i: int
+  when defined(amd64) and not defined(pixieNoSimd):
+    let
+      vZero = mm_setzero_si128()
+      vMax = mm_set1_epi32(cast[int32](uint32.high))
+    for _ in countup(0, mask.data.len - 16, 16):
+      var values = mm_loadu_si128(mask.data[i].addr)
+      values = mm_cmpeq_epi8(values, vZero)
+      values = mm_andnot_si128(values, vMax)
+      mm_storeu_si128(mask.data[i].addr, values)
+      i += 16
+
+  for j in i ..< mask.data.len:
+    if mask.data[j] != 0:
+      mask.data[j] = 255
+
 when defined(release):
   {.pop.}

src/pixie/paths.nim

@@ -373,11 +373,18 @@ proc roundedRect*(
     se = min(se, maxRadius)
     sw = min(sw, maxRadius)
 
+  if clockwise:
     path.moveTo(pos.x + nw, pos.y)
     path.arcTo(pos.x + wh.x, pos.y, pos.x + wh.x, pos.y + wh.y, ne)
     path.arcTo(pos.x + wh.x, pos.y + wh.y, pos.x, pos.y + wh.y, se)
     path.arcTo(pos.x, pos.y + wh.y, pos.x, pos.y, sw)
     path.arcTo(pos.x, pos.y, pos.x + wh.x, pos.y, nw)
+  else:
+    path.moveTo(pos.x + wh.x + ne, pos.y)
+    path.arcTo(pos.x, pos.y, pos.x, pos.y + wh.y, nw)
+    path.arcTo(pos.x, pos.y + wh.y, pos.x + wh.x, pos.y + wh.y, sw)
+    path.arcTo(pos.x + wh.x, pos.y + wh.y, pos.x + wh.x, pos.y, se)
+    path.arcTo(pos.x + wh.x, pos.y, pos.x, pos.y, ne)
   path.closePath()
 
 proc ellipse*(path: var Path, cx, cy, rx, ry: float32) =
@@ -1077,51 +1084,17 @@ proc fillShapes(
     var x = startX
     when defined(amd64) and not defined(pixieNoSimd):
       # When supported, SIMD blend as much as possible
-      let
-        oddMask = mm_set1_epi16(cast[int16](0xff00))
-        v255high = mm_set1_epi16(cast[int16](255.uint16 shl 8))
-        div255 = mm_set1_epi16(cast[int16](0x8081))
-
       for _ in countup(x, coverages.len - 16, 16):
         var coverage = mm_loadu_si128(coverages[x].addr)
 
         let eqZero = mm_cmpeq_epi16(coverage, mm_setzero_si128())
         if mm_movemask_epi8(eqZero) != 0xffff:
           # If the coverages are not all zero
-          var
-            coverageEven = mm_slli_epi16(mm_andnot_si128(oddMask, coverage), 8)
-            coverageOdd = mm_and_si128(coverage, oddMask)
-
-          let
-            evenK = mm_sub_epi16(v255high, coverageEven)
-            oddK = mm_sub_epi16(v255high, coverageOdd)
-
-          var
-            backdrop = mm_loadu_si128(mask.data[mask.dataIndex(x, y)].addr)
-            backdropEven = mm_slli_epi16(mm_andnot_si128(oddMask, backdrop), 8)
-            backdropOdd = mm_and_si128(backdrop, oddMask)
-
-          # backdrop * k
-          backdropEven = mm_mulhi_epu16(backdropEven, evenK)
-          backdropOdd = mm_mulhi_epu16(backdropOdd, oddK)
-
-          # div 255
-          backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
-          backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
-
-          # Shift from high to low bits
-          coverageEven = mm_srli_epi16(coverageEven, 8)
-          coverageOdd = mm_srli_epi16(coverageOdd, 8)
-
-          var
-            blendedEven = mm_add_epi16(coverageEven, backdropEven)
-            blendedOdd = mm_add_epi16(coverageOdd, backdropOdd)
-
-          blendedOdd = mm_slli_epi16(blendedOdd, 8)
+          let backdrop = mm_loadu_si128(mask.data[mask.dataIndex(x, y)].addr)
 
           mm_storeu_si128(
             mask.data[mask.dataIndex(x, y)].addr,
-            mm_or_si128(blendedEven, blendedOdd)
+            maskNormalSimd(backdrop, coverage)
           )
 
         x += 16

tests/benchmark_images.nim

@@ -45,11 +45,6 @@ timeIt "applyOpacity":
 
 reset()
 
-timeIt "sharpOpacity":
-  image.sharpOpacity()
-
-reset()
-
 timeIt "toPremultipliedAlpha":
   image.toPremultipliedAlpha()
 

tests/benchmark_masks.nim

@@ -25,3 +25,8 @@ reset()
 
 timeIt "blur":
   mask.blur(40)
+
+reset()
+
+timeIt "ceil":
+  mask.ceil()

tests/test_masks.nim

@@ -91,3 +91,14 @@ block:
   a.spread(10)
 
   writeFile("tests/images/masks/spread.png", a.encodePng())
+
+block:
+  let mask = newMask(100, 100)
+
+  var path: Path
+  path.ellipse(mask.width / 2, mask.height / 2, 25, 25)
+
+  mask.fillPath(path)
+  mask.ceil()
+
+  writeFile("tests/images/masks/circleMaskSharpened.png", mask.encodePng())