Merge pull request #485 from treeform/guzba

switch to split draw, either pixel-aligned fast or smooth slower

experiments/bench_cairo_draw.nim

@@ -1,39 +1,4 @@
-import benchy, cairo, pixie, pixie/blends, pixie/internal
+import benchy, cairo, pixie
-
-when defined(amd64) and not defined(pixieNoSimd):
-  import nimsimd/sse2
-
-when defined(release):
-  {.push checks: off.}
-
-proc drawBasic(backdrop, source: Image) =
-  for y in 0 ..< min(backdrop.height, source.height):
-    if isOpaque(source.data, source.dataIndex(0, y), source.width):
-      copyMem(
-        backdrop.data[backdrop.dataIndex(0, y)].addr,
-        source.data[source.dataIndex(0, y)].addr,
-        min(backdrop.width, source.width) * 4
-      )
-    else:
-      var x: int
-      when defined(amd64) and not defined(pixieNoSimd):
-        let vec255 = mm_set1_epi32(cast[int32](uint32.high))
-        for _ in 0 ..< min(backdrop.width, source.width) div 4:
-          let sourceVec = mm_loadu_si128(source.data[source.dataIndex(x, y)].addr)
-          if mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, mm_setzero_si128())) != 0xffff:
-            if (mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, vec255)) and 0x8888) == 0x8888:
-              mm_storeu_si128(backdrop.data[backdrop.dataIndex(x, y)].addr, sourceVec)
-            else:
-              let backdropVec = mm_loadu_si128(backdrop.data[backdrop.dataIndex(x, y)].addr)
-              mm_storeu_si128(
-                backdrop.data[backdrop.dataIndex(x, y)].addr,
-                blendNormalInlineSimd(backdropVec, sourceVec)
-              )
-          x += 4
-      # No scalar for now
-
-when defined(release):
-  {.pop.}
 
 block:
   let
@@ -43,12 +8,6 @@ block:
     ctx = tmp.create()
 
   timeIt "cairo draw normal":
-    # ctx.setSourceRgba(0.5, 0.5, 0.5, 1)
-    # let operator = ctx.getOperator()
-    # ctx.setOperator(OperatorSource)
-    # ctx.paint()
-    # ctx.setOperator(operator)
-
     ctx.setSource(backdrop, 0, 0)
     ctx.paint()
     ctx.setSource(source, 0, 0)
@@ -64,7 +23,6 @@ block:
     tmp = newImage(1568, 940)
 
   timeIt "pixie draw normal":
-    # tmp.fill(rgbx(127, 127, 127, 255))
     tmp.draw(backdrop)
     tmp.draw(source)
 
@@ -77,25 +35,11 @@ block:
     tmp = newImage(1568, 940)
 
   timeIt "pixie draw overwrite":
-    # tmp.fill(rgbx(127, 127, 127, 255))
     tmp.draw(backdrop, blendMode = OverwriteBlend)
     tmp.draw(source)
 
   # tmp.writeFile("tmp2.png")
 
-block:
-  let
-    backdrop = readImage("tests/fileformats/svg/masters/dragon2.png")
-    source = readImage("tests/fileformats/svg/masters/Ghostscript_Tiger.png")
-    tmp = newImage(1568, 940)
-
-  timeIt "pixie draw basic":
-    # tmp.fill(rgbx(127, 127, 127, 255))
-    tmp.drawBasic(backdrop)
-    tmp.drawBasic(source)
-
-  # tmp.writeFile("tmp2.png")
-
 block:
   let
     backdrop = imageSurfaceCreateFromPng("tests/fileformats/svg/masters/dragon2.png")
@@ -104,17 +48,11 @@ block:
     ctx = tmp.create()
 
   timeIt "cairo draw mask":
-    ctx.setSourceRgba(1, 1, 1, 1)
-    let operator = ctx.getOperator()
-    ctx.setOperator(OperatorSource)
-    ctx.paint()
-    ctx.setOperator(operator)
-
     ctx.setSource(backdrop, 0, 0)
     ctx.mask(source, 0, 0)
     tmp.flush()
 
-  # echo tmp.writeToPng("tmp_masked.png")
+  # tmp.writeToPng("tmp_masked.png")
 
 block:
   let

src/pixie/blends.nim

@@ -1,6 +1,6 @@
 ## Blending modes.
 
-import chroma, common, simd, std/math
+import chroma, common, std/math
 
 # See https://www.w3.org/TR/compositing-1/
 # See https://www.khronos.org/registry/OpenGL/extensions/KHR/KHR_blend_equation_advanced.txt
@@ -412,82 +412,5 @@ proc blender*(blendMode: BlendMode): Blender {.raises: [].} =
   of SubtractMaskBlend: subtractMaskBlender
   of ExcludeMaskBlend: excludeMaskBlender
 
-when defined(amd64) and allowSimd:
-  type
-    BlenderSimd* = proc(blackdrop, source: M128i): M128i {.gcsafe, raises: [].}
-      ## Function signature returned by blenderSimd.
-
-  proc blendNormalSimd*(backdrop, source: M128i): M128i {.inline.} =
-    let
-      alphaMask = mm_set1_epi32(cast[int32](0xff000000))
-      oddMask = mm_set1_epi16(cast[int16](0xff00))
-      div255 = mm_set1_epi16(cast[int16](0x8081))
-
-    var
-      sourceAlpha = mm_and_si128(source, alphaMask)
-      backdropEven = mm_slli_epi16(backdrop, 8)
-      backdropOdd = mm_and_si128(backdrop, oddMask)
-
-    sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
-
-    let k = mm_sub_epi32(
-      mm_set1_epi32(cast[int32]([0.uint8, 255, 0, 255])),
-      sourceAlpha
-    )
-
-    backdropEven = mm_mulhi_epu16(backdropEven, k)
-    backdropOdd = mm_mulhi_epu16(backdropOdd, k)
-
-    backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
-    backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
-
-    mm_add_epi8(
-      source,
-      mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
-    )
-
-  proc blendMaskSimd*(backdrop, source: M128i): M128i {.inline.} =
-    let
-      alphaMask = mm_set1_epi32(cast[int32](0xff000000))
-      oddMask = mm_set1_epi16(cast[int16](0xff00))
-      div255 = mm_set1_epi16(cast[int16](0x8081))
-
-    var
-      sourceAlpha = mm_and_si128(source, alphaMask)
-      backdropEven = mm_slli_epi16(backdrop, 8)
-      backdropOdd = mm_and_si128(backdrop, oddMask)
-
-    sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
-
-    backdropEven = mm_mulhi_epu16(backdropEven, sourceAlpha)
-    backdropOdd = mm_mulhi_epu16(backdropOdd, sourceAlpha)
-
-    backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
-    backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
-
-    mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
-
-  proc normalSimdBlender(backdrop, source: M128i): M128i =
-    blendNormalSimd(backdrop, source)
-
-  proc maskSimdBlender(backdrop, source: M128i): M128i =
-    blendMaskSimd(backdrop, source)
-
-  proc overwriteSimdBlender(backdrop, source: M128i): M128i =
-    source
-
-  proc blenderSimd*(blendMode: BlendMode): BlenderSimd {.raises: [PixieError].} =
-    ## Returns a blend function for a given blend mode with SIMD support.
-    case blendMode:
-    of NormalBlend: normalSimdBlender
-    of MaskBlend: maskSimdBlender
-    of OverwriteBlend: overwriteSimdBlender
-    else:
-      raise newException(PixieError, "No SIMD blender for " & $blendMode)
-
-  proc hasSimdBlender*(blendMode: BlendMode): bool {.inline, raises: [].} =
-    ## Is there a blend function for a given blend mode with SIMD support?
-    blendMode in {NormalBlend, MaskBlend, OverwriteBlend}
-
 when defined(release):
   {.pop.}

src/pixie/images.nim

@@ -394,63 +394,11 @@ proc getRgbaSmooth*(
   else:
     topMix
 
-proc blitLine(a, b: ptr UncheckedArray[ColorRGBX], len: int, blender: Blender) =
-  for i in 0 ..< len:
-    a[i] = blender(a[i], b[i])
-
-proc blitLineNormal(a, b: ptr UncheckedArray[ColorRGBX], len: int) {.hasSimd.} =
-  for i in 0 ..< len:
-    a[i] = blendNormal(a[i], b[i])
-
-proc blitLineOverwrite(a, b: ptr UncheckedArray[ColorRGBX], len: int) =
-  copyMem(a[0].addr, b[0].addr, len * 4)
-
-template getUncheckedArray(a: Image, x, y: int): ptr UncheckedArray[ColorRGBX] =
-  cast[ptr UncheckedArray[ColorRGBX]](a.data[a.dataIndex(x, y)].addr)
-
-proc blitRect(
-  a, b: Image, pos = ivec2(0, 0), blendMode = NormalBlend
-) =
-  ## Blits one image onto another using integer position with color blending.
-  let
-    px = pos.x.int
-    py = pos.y.int
-    xStart = max(-px, 0)
-    yStart = max(-py, 0)
-    xEnd = min(b.width, a.width - px)
-    yEnd = min(b.height, a.height - py)
-
-  case blendMode:
-  of NormalBlend:
-    for y in yStart ..< yEnd:
-      blitLineNormal(
-        a.getUncheckedArray(xStart + px, y + py),
-        b.getUncheckedArray(xStart, y),
-        xEnd - xStart
-      )
-  of OverwriteBlend:
-    {.linearScanEnd.}
-    for y in yStart ..< yEnd:
-      blitLineOverwrite(
-        a.getUncheckedArray(xStart + px, y + py),
-        b.getUncheckedArray(xStart, y),
-        xEnd - xStart
-      )
-  else:
-    let blender = blendMode.blender()
-    for y in yStart ..< yEnd:
-      blitLine(
-        a.getUncheckedArray(xStart + px, y + py),
-        b.getUncheckedArray(xStart, y),
-        xEnd - xStart,
-        blender
-      )
-
 proc drawCorrect(
-  a, b: Image, transform = mat3(), blendMode = NormalBlend, tiled = false
+  a, b: Image, transform = mat3(), blendMode: BlendMode, tiled: bool
-) {.raises: [PixieError].} =
+) =
-  ## Draws one image onto another using matrix with color blending.
+  ## Draws one image onto another using a matrix transform and color blending.
-
+  ## This proc is not about performance, it should be as simple as possible.
   var
     inverseTransform = transform.inverse()
     # Compute movement vectors
@@ -466,7 +414,131 @@ proc drawCorrect(
     dx /= 2
     dy /= 2
     filterBy2 /= 2
-    inverseTransform = scale(vec2(1/2, 1/2)) * inverseTransform
+    inverseTransform = scale(vec2(0.5, 0.5)) * inverseTransform
+
+  while filterBy2 <= 0.5:
+    b = b.magnifyBy2()
+    p *= 2
+    dx *= 2
+    dy *= 2
+    filterBy2 *= 2
+    inverseTransform = scale(vec2(2, 2)) * inverseTransform
+
+  let blender = blendMode.blender()
+  for y in 0 ..< a.height:
+    for x in 0 ..< a.width:
+      let
+        samplePos = inverseTransform * vec2(x.float32 + h, y.float32 + h)
+        xFloat = samplePos.x - h
+        yFloat = samplePos.y - h
+        backdrop = a.unsafe[x, y]
+        sample = b.getRgbaSmooth(xFloat, yFloat, tiled)
+        blended = blender(backdrop, sample)
+      a.unsafe[x, y] = blended
+
+template getUncheckedArray(
+  image: Image, x, y: int
+): ptr UncheckedArray[ColorRGBX] =
+  cast[ptr UncheckedArray[ColorRGBX]](image.data[image.dataIndex(x, y)].addr)
+
+proc blitLine(a, b: ptr UncheckedArray[ColorRGBX], len: int, blender: Blender) {.inline.} =
+  for i in 0 ..< len:
+    a[i] = blender(a[i], b[i])
+
+proc blitLineOverwrite(a, b: ptr UncheckedArray[ColorRGBX], len: int) {.inline.} =
+  copyMem(a[0].addr, b[0].addr, len * 4)
+
+proc blitLineNormal(a, b: ptr UncheckedArray[ColorRGBX], len: int) {.hasSimd.} =
+  for i in 0 ..< len:
+    a[i] = blendNormal(a[i], b[i])
+
+proc blitLineMask(a, b: ptr UncheckedArray[ColorRGBX], len: int) {.hasSimd.} =
+  for i in 0 ..< len:
+    a[i] = blendMask(a[i], b[i])
+
+proc blitRect(a, b: Image, pos: Ivec2, blendMode: BlendMode) =
+  let
+    px = pos.x.int
+    py = pos.y.int
+
+  if px >= a.width or px + b.width <= 0 or py >= a.height or py + b.height <= 0:
+    if blendMode == MaskBlend:
+      a.fill(rgbx(0, 0, 0, 0))
+    return
+
+  let
+    xStart = max(-px, 0)
+    yStart = max(-py, 0)
+    xEnd = min(b.width, a.width - px)
+    yEnd = min(b.height, a.height - py)
+
+  case blendMode:
+  of NormalBlend:
+    for y in yStart ..< yEnd:
+      blitLineNormal(
+        a.getUncheckedArray(xStart + px, y + py),
+        b.getUncheckedArray(xStart, y),
+        xEnd - xStart
+      )
+  of OverwriteBlend:
+    for y in yStart ..< yEnd:
+      blitLineOverwrite(
+        a.getUncheckedArray(xStart + px, y + py),
+        b.getUncheckedArray(xStart, y),
+        xEnd - xStart
+      )
+  of MaskBlend:
+    {.linearScanEnd.}
+    if yStart + py > 0:
+      zeroMem(a.data[0].addr, (yStart + py) * a.width * 4)
+    for y in yStart ..< yEnd:
+      if xStart + px > 0:
+        zeroMem(a.data[a.dataIndex(0, y + py)].addr, (xStart + px) * 4)
+      blitLineMask(
+        a.getUncheckedArray(xStart + px, y + py),
+        b.getUncheckedArray(xStart, y),
+        xEnd - xStart
+      )
+      if xEnd + px < a.width:
+        zeroMem(
+          a.data[a.dataIndex(xEnd + px, y + py)].addr,
+          (a.width - (xEnd + px)) * 4
+        )
+    if yEnd + py < a.height:
+      zeroMem(
+        a.data[a.dataIndex(0, yEnd + py)].addr,
+        (a.height - (yEnd + py)) * a.width * 4
+      )
+  else:
+    let blender = blendMode.blender()
+    for y in yStart ..< yEnd:
+      blitLine(
+        a.getUncheckedArray(xStart + px, y + py),
+        b.getUncheckedArray(xStart, y),
+        xEnd - xStart,
+        blender
+      )
+
+proc draw*(
+  a, b: Image, transform = mat3(), blendMode = NormalBlend
+) {.raises: [PixieError].} =
+  ## Draws one image onto another using a matrix transform and color blending.
+  var
+    inverseTransform = transform.inverse()
+    # Compute movement vectors
+    p = inverseTransform * vec2(0 + h, 0 + h)
+    dx = inverseTransform * vec2(1 + h, 0 + h) - p
+    dy = inverseTransform * vec2(0 + h, 1 + h) - p
+    filterBy2 = max(dx.length, dy.length)
+    b = b
+
+  while filterBy2 >= 2.0:
+    b = b.minifyBy2()
+    p /= 2
+    dx /= 2
+    dy /= 2
+    filterBy2 /= 2
+    inverseTransform = scale(vec2(0.5, 0.5)) * inverseTransform
 
   while filterBy2 <= 0.5:
     b = b.magnifyBy2()
@@ -485,277 +557,10 @@ proc drawCorrect(
       transform[2, 1].fractional == 0.0
     )
 
-  if not hasRotationOrScaling and not smooth and not tiled:
-    blitRect(a, b, ivec2(transform[2, 0].int32, transform[2, 1].int32), blendMode)
-    return
-
-  let blender = blendMode.blender()
-  for y in 0 ..< a.height:
-    for x in 0 ..< a.width:
-      let
-        samplePos = inverseTransform * vec2(x.float32 + h, y.float32 + h)
-        xFloat = samplePos.x - h
-        yFloat = samplePos.y - h
-        backdrop = a.unsafe[x, y]
-        sample = b.getRgbaSmooth(xFloat, yFloat, tiled)
-        blended = blender(backdrop, sample)
-      a.unsafe[x, y] = blended
-
-proc drawUber(
-  a, b: Image, transform = mat3(), blendMode: BlendMode
-) {.raises: [PixieError].} =
-  let
-    corners = [
-      transform * vec2(0, 0),
-      transform * vec2(b.width.float32, 0),
-      transform * vec2(b.width.float32, b.height.float32),
-      transform * vec2(0, b.height.float32)
-    ]
-    perimeter = [
-      segment(corners[0], corners[1]),
-      segment(corners[1], corners[2]),
-      segment(corners[2], corners[3]),
-      segment(corners[3], corners[0])
-    ]
-
-  var
-    inverseTransform = transform.inverse()
-    # Compute movement vectors
-    p = inverseTransform * vec2(0 + h, 0 + h)
-    dx = inverseTransform * vec2(1 + h, 0 + h) - p
-    dy = inverseTransform * vec2(0 + h, 1 + h) - p
-    filterBy2 = max(dx.length, dy.length)
-    b = b
-
-  while filterBy2 >= 2.0:
-    b = b.minifyBy2()
-    p /= 2
-    dx /= 2
-    dy /= 2
-    filterBy2 /= 2
-
-  while filterBy2 <= 0.5:
-    b = b.magnifyBy2()
-    p *= 2
-    dx *= 2
-    dy *= 2
-    filterBy2 *= 2
-
-  let
-    hasRotationOrScaling = not(dx == vec2(1, 0) and dy == vec2(0, 1))
-    smooth = not(
-      dx.length == 1.0 and
-      dy.length == 1.0 and
-      transform[2, 0].fractional == 0.0 and
-      transform[2, 1].fractional == 0.0
-    )
-
-  # Determine where we should start and stop drawing in the y dimension
-  var
-    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)
-
-  let blender = blendMode.blender()
-
-  if blendMode == MaskBlend:
-    if yMin > 0:
-      zeroMem(a.data[0].addr, 4 * yMin * a.width)
-
-  for y in yMin ..< yMax:
-    # Determine where we should start and stop drawing in the x dimension
-    var
-      xMin = a.width.float32
-      xMax = 0.float32
-    for yOffset in [0.float32, 1]:
-      let scanLine = Line(
-        a: vec2(-1000, y.float32 + yOffset),
-        b: vec2(1000, y.float32 + yOffset)
-      )
-      for segment in perimeter:
-        var at: Vec2
-        if scanline.intersects(segment, at) and segment.to != at:
-          xMin = min(xMin, at.x)
-          xMax = max(xMax, at.x)
-
-    var xStart, xStop: int
   if hasRotationOrScaling or smooth:
-      xStart = xMin.floor.int
+    a.drawCorrect(b, inverseTransform.inverse(), blendMode, false)
-      xStop = xMax.ceil.int
   else:
-      # Rotation of 360 degrees can cause knife-edge issues with floor and ceil
+    a.blitRect(b, ivec2(transform[2, 0].int32, transform[2, 1].int32), blendMode)
-      xStart = xMin.round().int
-      xStop = xMax.round().int
-    xStart = xStart.clamp(0, a.width)
-    xStop = xStop.clamp(0, a.width)
-
-    # Skip this row if there is nothing in-bounds to draw
-    if xStart == a.width or xStop == 0:
-      continue
-
-    if blendMode == MaskBlend:
-      if xStart > 0:
-        zeroMem(a.data[a.dataIndex(0, y)].addr, 4 * xStart)
-
-    if smooth:
-      var srcPos = p + dx * xStart.float32 + dy * y.float32
-      srcPos = vec2(srcPos.x - h, srcPos.y - h)
-      for x in xStart ..< xStop:
-        let
-          backdrop = a.unsafe[x, y]
-          sample = b.getRgbaSmooth(srcPos.x, srcPos.y)
-          blended = blender(backdrop, sample)
-        a.unsafe[x, y] = blended
-        srcPos += dx
-
-    else:
-      var x = xStart
-      if not hasRotationOrScaling:
-        let
-          srcPos = p + dx * x.float32 + dy * y.float32
-          sy = srcPos.y.int
-        var sx = srcPos.x.int
-
-        if blendMode in {NormalBlend, OverwriteBlend} and
-          isOpaque(b.data, b.dataIndex(sx, sy), xStop - xStart):
-          copyMem(
-            a.data[a.dataIndex(x, y)].addr,
-            b.data[b.dataIndex(sx, sy)].addr,
-            (xStop - xStart) * 4
-          )
-          continue
-
-        when defined(amd64) and allowSimd:
-          case blendMode:
-          of OverwriteBlend:
-            for _ in 0 ..< (xStop - xStart) div 16:
-              for q in [0, 4, 8, 12]:
-                let sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx + q, sy)].addr)
-                mm_storeu_si128(a.data[a.dataIndex(x + q, y)].addr, sourceVec)
-              x += 16
-              sx += 16
-          of NormalBlend:
-            let vec255 = mm_set1_epi32(cast[int32](uint32.high))
-            for _ in 0 ..< (xStop - xStart) div 16:
-              for q in [0, 4, 8, 12]:
-                let
-                  sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx + q, sy)].addr)
-                  eqZer0 = mm_cmpeq_epi8(sourceVec, mm_setzero_si128())
-                if mm_movemask_epi8(eqZer0) != 0xffff:
-                  let eq255 = mm_cmpeq_epi8(sourceVec, vec255)
-                  if (mm_movemask_epi8(eq255) and 0x8888) == 0x8888:
-                    mm_storeu_si128(a.data[a.dataIndex(x + q, y)].addr, sourceVec)
-                  else:
-                    let
-                      backdropIdx = a.dataIndex(x + q, y)
-                      backdropVec = mm_loadu_si128(a.data[backdropIdx].addr)
-                    mm_storeu_si128(
-                      a.data[backdropIdx].addr,
-                      blendNormalSimd(backdropVec, sourceVec)
-                    )
-              x += 16
-              sx += 16
-          of MaskBlend:
-            let vec255 = mm_set1_epi32(cast[int32](uint32.high))
-            for _ in 0 ..< (xStop - xStart) div 16:
-              for q in [0, 4, 8, 12]:
-                let
-                  sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx + q, sy)].addr)
-                  eqZer0 = mm_cmpeq_epi8(sourceVec, mm_setzero_si128())
-                if mm_movemask_epi8(eqZer0) == 0xffff:
-                  mm_storeu_si128(
-                    a.data[a.dataIndex(x + q, y)].addr,
-                    mm_setzero_si128()
-                  )
-                elif mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, vec255)) != 0xffff:
-                  let backdropVec = mm_loadu_si128(a.data[a.dataIndex(x + q, y)].addr)
-                  mm_storeu_si128(
-                    a.data[a.dataIndex(x + q, y)].addr,
-                    blendMaskSimd(backdropVec, sourceVec)
-                  )
-              x += 16
-              sx += 16
-          else:
-            when type(a) is Image:
-              if blendMode.hasSimdBlender():
-                let blenderSimd = blendMode.blenderSimd()
-                for _ in 0 ..< (xStop - xStart) div 16:
-                  for q in [0, 4, 8, 12]:
-                    let
-                      backdrop = mm_loadu_si128(a.data[a.dataIndex(x + q, y)].addr)
-                      source = mm_loadu_si128(b.data[b.dataIndex(sx + q, sy)].addr)
-                    mm_storeu_si128(
-                      a.data[a.dataIndex(x + q, y)].addr,
-                      blenderSimd(backdrop, source)
-                    )
-                  x += 16
-                  sx += 16
-
-      var srcPos = p + dx * x.float32 + dy * y.float32
-      srcPos = vec2(
-        clamp(srcPos.x, 0, b.width.float32),
-        clamp(srcPos.y, 0, b.height.float32)
-      )
-
-      case blendMode:
-      of OverwriteBlend:
-        for x in x ..< xStop:
-          let
-            samplePos = ivec2((srcPos.x - h).int32, (srcPos.y - h).int32)
-            source = b.unsafe[samplePos.x, samplePos.y]
-          if source.a > 0:
-            a.unsafe[x, y] = source
-          srcPos += dx
-      of NormalBlend:
-        for x in x ..< xStop:
-          let
-            samplePos = ivec2((srcPos.x - h).int32, (srcPos.y - h).int32)
-            source = b.unsafe[samplePos.x, samplePos.y]
-          if source.a > 0:
-            if source.a == 255:
-              a.unsafe[x, y] = source
-            else:
-              let backdrop = a.unsafe[x, y]
-              a.unsafe[x, y] = blendNormal(backdrop, source)
-          srcPos += dx
-      of MaskBlend:
-        for x in x ..< xStop:
-          let
-            samplePos = ivec2((srcPos.x - h).int32, (srcPos.y - h).int32)
-            source = b.unsafe[samplePos.x, samplePos.y]
-          if source.a == 0:
-            a.unsafe[x, y] = rgbx(0, 0, 0, 0)
-          elif source.a != 255:
-            let backdrop = a.unsafe[x, y]
-            a.unsafe[x, y] = blendMask(backdrop, source)
-          srcPos += dx
-      else:
-        for x in x ..< xStop:
-          let
-            samplePos = ivec2((srcPos.x - h).int32, (srcPos.y - h).int32)
-            backdrop = a.unsafe[x, y]
-            sample = b.unsafe[samplePos.x, samplePos.y]
-            blended = blender(backdrop, sample)
-          a.unsafe[x, y] = blended
-          srcPos += dx
-
-    if blendMode == MaskBlend:
-      if a.width - xStop > 0:
-        zeroMem(a.data[a.dataIndex(xStop, y)].addr, 4 * (a.width - xStop))
-
-  if blendMode == MaskBlend:
-    if a.height - yMax > 0:
-      zeroMem(a.data[a.dataIndex(0, yMax)].addr, 4 * a.width * (a.height - yMax))
-
-proc draw*(
-  a, b: Image, transform = mat3(), blendMode = NormalBlend
-) {.inline, raises: [PixieError].} =
-  ## Draws one image onto another using matrix with color blending.
-  a.drawUber(b, transform, blendMode)
 
 proc drawTiled*(
   dst, src: Image, mat: Mat3, blendMode = NormalBlend

src/pixie/simd/avx2.nim

@@ -1,4 +1,4 @@
-import avx, chroma, internal, nimsimd/avx2, pixie/common, vmath
+import avx, chroma, internal, nimsimd/avx2, pixie/blends, pixie/common, vmath
 
 when defined(gcc) or defined(clang):
   {.localPassc: "-mavx2".}
@@ -107,6 +107,10 @@ proc toPremultipliedAlphaAvx2*(data: var seq[ColorRGBA | ColorRGBX]) {.simd.} =
 
   let
     alphaMask = mm256_set1_epi32(cast[int32](0xff000000))
+    shuffleControl = mm256_set_epi8(
+      15, -1, 15, -1, 11, -1, 11, -1, 7, -1, 7, -1, 3, -1, 3, -1,
+      15, -1, 15, -1, 11, -1, 11, -1, 7, -1, 7, -1, 3, -1, 3, -1
+    )
     oddMask = mm256_set1_epi16(0xff00)
     vec128 = mm256_set1_epi16(128)
     hiMask = mm256_set1_epi16(255 shl 8)
@@ -118,7 +122,7 @@ proc toPremultipliedAlphaAvx2*(data: var seq[ColorRGBA | ColorRGBX]) {.simd.} =
       eq = mm256_cmpeq_epi8(values, alphaMask)
     if (mm256_movemask_epi8(eq) and 0x88888888) != 0x88888888:
       let
-        evenMultiplier = mm256_or_si256(alpha, mm256_srli_epi32(alpha, 16))
+        evenMultiplier = mm256_shuffle_epi8(alpha, shuffleControl)
         oddMultiplier = mm256_or_si256(evenMultiplier, alphaMask)
       var
         colorsEven = mm256_slli_epi16(values, 8)
@@ -376,5 +380,100 @@ proc minifyBy2Avx2*(image: Image, power = 1): Image {.simd.} =
     # Set src as this result for if we do another power
     src = result
 
+proc blitLineNormalAvx2*(
+  a, b: ptr UncheckedArray[ColorRGBX], len: int
+) {.simd.} =
+  let
+    alphaMask = mm256_set1_epi32(cast[int32](0xff000000))
+    oddMask = mm256_set1_epi16(cast[int16](0xff00))
+    div255 = mm256_set1_epi16(cast[int16](0x8081))
+    vec255 = mm256_set1_epi8(255)
+    vecAlpha255 = mm256_set1_epi32(cast[int32]([0.uint8, 255, 0, 255]))
+    shuffleControl = mm256_set_epi8(
+      15, -1, 15, -1, 11, -1, 11, -1, 7, -1, 7, -1, 3, -1, 3, -1,
+      15, -1, 15, -1, 11, -1, 11, -1, 7, -1, 7, -1, 3, -1, 3, -1
+    )
+
+  var i: int
+  while i < len - 8:
+    let
+      source = mm256_loadu_si256(b[i].addr)
+      eq255 = mm256_cmpeq_epi8(source, vec255)
+    if (mm256_movemask_epi8(eq255) and 0x88888888) == 0x88888888: # Opaque source
+      mm256_storeu_si256(a[i].addr, source)
+    else:
+      let backdrop = mm256_loadu_si256(a[i].addr)
+
+      var
+        sourceAlpha = mm256_and_si256(source, alphaMask)
+        backdropEven = mm256_slli_epi16(backdrop, 8)
+        backdropOdd = mm256_and_si256(backdrop, oddMask)
+
+      sourceAlpha = mm256_shuffle_epi8(sourceAlpha, shuffleControl)
+
+      let multiplier = mm256_sub_epi32(vecAlpha255, sourceAlpha)
+
+      backdropEven = mm256_mulhi_epu16(backdropEven, multiplier)
+      backdropOdd = mm256_mulhi_epu16(backdropOdd, multiplier)
+      backdropEven = mm256_srli_epi16(mm256_mulhi_epu16(backdropEven, div255), 7)
+      backdropOdd = mm256_srli_epi16(mm256_mulhi_epu16(backdropOdd, div255), 7)
+
+      let added = mm256_add_epi8(
+        source,
+        mm256_or_si256(backdropEven, mm256_slli_epi16(backdropOdd, 8))
+      )
+
+      mm256_storeu_si256(a[i].addr, added)
+
+    i += 8
+
+  for i in i ..< len:
+    a[i] = blendNormal(a[i], b[i])
+
+proc blitLineMaskAvx2*(
+  a, b: ptr UncheckedArray[ColorRGBX], len: int
+) {.simd.} =
+  let
+    alphaMask = mm256_set1_epi32(cast[int32](0xff000000))
+    oddMask = mm256_set1_epi16(cast[int16](0xff00))
+    div255 = mm256_set1_epi16(cast[int16](0x8081))
+    vec255 = mm256_set1_epi8(255)
+    shuffleControl = mm256_set_epi8(
+      15, -1, 15, -1, 11, -1, 11, -1, 7, -1, 7, -1, 3, -1, 3, -1,
+      15, -1, 15, -1, 11, -1, 11, -1, 7, -1, 7, -1, 3, -1, 3, -1
+    )
+
+  var i: int
+  while i < len - 8:
+    let
+      source = mm256_loadu_si256(b[i].addr)
+      eq255 = mm256_cmpeq_epi8(source, vec255)
+    if (mm256_movemask_epi8(eq255) and 0x88888888) == 0x88888888: # Opaque source
+      discard
+    else:
+      let backdrop = mm256_loadu_si256(a[i].addr)
+
+      var
+        sourceAlpha = mm256_and_si256(source, alphaMask)
+        backdropEven = mm256_slli_epi16(backdrop, 8)
+        backdropOdd = mm256_and_si256(backdrop, oddMask)
+
+      sourceAlpha = mm256_shuffle_epi8(sourceAlpha, shuffleControl)
+
+      backdropEven = mm256_mulhi_epu16(backdropEven, sourceAlpha)
+      backdropOdd = mm256_mulhi_epu16(backdropOdd, sourceAlpha)
+      backdropEven = mm256_srli_epi16(mm256_mulhi_epu16(backdropEven, div255), 7)
+      backdropOdd = mm256_srli_epi16(mm256_mulhi_epu16(backdropOdd, div255), 7)
+
+      mm256_storeu_si256(
+        a[i].addr,
+        mm256_or_si256(backdropEven, mm256_slli_epi16(backdropOdd, 8))
+      )
+
+    i += 8
+
+  for i in i ..< len:
+    a[i] = blendMask(a[i], b[i])
+
 when defined(release):
   {.pop.}

src/pixie/simd/sse2.nim

@@ -1,4 +1,4 @@
-import chroma, internal, nimsimd/sse2, pixie/common, vmath
+import chroma, internal, nimsimd/sse2, pixie/blends, pixie/common, vmath
 
 when defined(release):
   {.push checks: off.}
@@ -15,6 +15,56 @@ proc unpackAlphaValues*(v: M128i): M128i {.inline, raises: [].} =
   result = mm_unpacklo_epi8(mm_setzero_si128(), v)
   result = mm_unpacklo_epi8(mm_setzero_si128(), result)
 
+proc blendNormalSimd*(backdrop, source: M128i): M128i {.inline.} =
+  let
+    alphaMask = mm_set1_epi32(cast[int32](0xff000000))
+    oddMask = mm_set1_epi16(cast[int16](0xff00))
+    div255 = mm_set1_epi16(cast[int16](0x8081))
+
+  var
+    sourceAlpha = mm_and_si128(source, alphaMask)
+    backdropEven = mm_slli_epi16(backdrop, 8)
+    backdropOdd = mm_and_si128(backdrop, oddMask)
+
+  sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
+
+  let k = mm_sub_epi32(
+    mm_set1_epi32(cast[int32]([0.uint8, 255, 0, 255])),
+    sourceAlpha
+  )
+
+  backdropEven = mm_mulhi_epu16(backdropEven, k)
+  backdropOdd = mm_mulhi_epu16(backdropOdd, k)
+
+  backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
+  backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
+
+  mm_add_epi8(
+    source,
+    mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
+  )
+
+proc blendMaskSimd*(backdrop, source: M128i): M128i {.inline.} =
+  let
+    alphaMask = mm_set1_epi32(cast[int32](0xff000000))
+    oddMask = mm_set1_epi16(cast[int16](0xff00))
+    div255 = mm_set1_epi16(cast[int16](0x8081))
+
+  var
+    sourceAlpha = mm_and_si128(source, alphaMask)
+    backdropEven = mm_slli_epi16(backdrop, 8)
+    backdropOdd = mm_and_si128(backdrop, oddMask)
+
+  sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
+
+  backdropEven = mm_mulhi_epu16(backdropEven, sourceAlpha)
+  backdropOdd = mm_mulhi_epu16(backdropOdd, sourceAlpha)
+
+  backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
+  backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
+
+  mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
+
 proc fillUnsafeSse2*(
   data: var seq[ColorRGBX],
   color: SomeColor,
@@ -480,18 +530,22 @@ proc magnifyBy2Sse2*(image: Image, power = 1): Image {.simd.} =
 proc blitLineNormalSse2*(
   a, b: ptr UncheckedArray[ColorRGBX], len: int
 ) {.simd.} =
-
-  # TODO align to 16
-
-  var i = 0
-  while i < len - 4:
-
   let
-      source = mm_loadu_si128(b[i].addr)
-      backdrop = mm_loadu_si128(a[i].addr)
     alphaMask = mm_set1_epi32(cast[int32](0xff000000))
     oddMask = mm_set1_epi16(cast[int16](0xff00))
     div255 = mm_set1_epi16(cast[int16](0x8081))
+    vec255 = mm_set1_epi8(255)
+    vecAlpha255 = mm_set1_epi32(cast[int32]([0.uint8, 255, 0, 255]))
+
+  var i: int
+  while i < len - 4:
+    let
+      source = mm_loadu_si128(b[i].addr)
+      eq255 = mm_cmpeq_epi8(source, vec255)
+    if (mm_movemask_epi8(eq255) and 0x00008888) == 0x00008888: # Opaque source
+      mm_storeu_si128(a[i].addr, source)
+    else:
+      let backdrop = mm_loadu_si128(a[i].addr)
 
       var
         sourceAlpha = mm_and_si128(source, alphaMask)
@@ -500,29 +554,65 @@ proc blitLineNormalSse2*(
 
       sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
 
-    let k = mm_sub_epi32(
+      let multiplier = mm_sub_epi32(vecAlpha255, sourceAlpha)
-      mm_set1_epi32(cast[int32]([0.uint8, 255, 0, 255])),
-      sourceAlpha
-    )
-
-    backdropEven = mm_mulhi_epu16(backdropEven, k)
-    backdropOdd = mm_mulhi_epu16(backdropOdd, k)
 
+      backdropEven = mm_mulhi_epu16(backdropEven, multiplier)
+      backdropOdd = mm_mulhi_epu16(backdropOdd, multiplier)
       backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
       backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
 
-    let done = mm_add_epi8(
+      let added = mm_add_epi8(
         source,
         mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
       )
 
-    mm_storeu_si128(a[i].addr, done)
+      mm_storeu_si128(a[i].addr, added)
 
     i += 4
 
-  # TODO last 1-3 pixels
+  for i in i ..< len:
-  # for i in i ..< len:
+    a[i] = blendNormal(a[i], b[i])
-  #   a[i] = blendNormal(a[i], b[i])
+
+proc blitLineMaskSse2*(
+  a, b: ptr UncheckedArray[ColorRGBX], len: int
+) {.simd.} =
+  let
+    alphaMask = mm_set1_epi32(cast[int32](0xff000000))
+    oddMask = mm_set1_epi16(cast[int16](0xff00))
+    div255 = mm_set1_epi16(cast[int16](0x8081))
+    vec255 = mm_set1_epi8(255)
+
+  var i: int
+  while i < len - 4:
+    let
+      source = mm_loadu_si128(b[i].addr)
+      eq255 = mm_cmpeq_epi8(source, vec255)
+    if (mm_movemask_epi8(eq255) and 0x00008888) == 0x00008888: # Opaque source
+      discard
+    else:
+      let backdrop = mm_loadu_si128(a[i].addr)
+
+      var
+        sourceAlpha = mm_and_si128(source, alphaMask)
+        backdropEven = mm_slli_epi16(backdrop, 8)
+        backdropOdd = mm_and_si128(backdrop, oddMask)
+
+      sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
+
+      backdropEven = mm_mulhi_epu16(backdropEven, sourceAlpha)
+      backdropOdd = mm_mulhi_epu16(backdropOdd, sourceAlpha)
+      backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
+      backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
+
+      mm_storeu_si128(
+        a[i].addr,
+        mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
+      )
+
+    i += 4
+
+  for i in i ..< len:
+    a[i] = blendMask(a[i], b[i])
 
 when defined(release):
   {.pop.}

tests/bench_blends.nim

@@ -133,17 +133,3 @@ reset()
 timeIt "blendExcludeMask":
   for i in 0 ..< backdrop.data.len:
     backdrop.data[i] = blendExcludeMask(backdrop.data[i], source.data[i])
-
-when defined(amd64) and not defined(pixieNoSimd):
-  import nimsimd/sse2
-
-  reset()
-
-  timeIt "blendNormal [simd]":
-    var i: int
-    while i < backdrop.data.len - 4:
-      let
-        b = mm_loadu_si128(backdrop.data[i].addr)
-        s = mm_loadu_si128(source.data[i].addr)
-      mm_storeu_si128(backdrop.data[i].addr, blendNormalSimd(b, s))
-      i += 4

tests/bench_images.nim

@@ -32,12 +32,12 @@ timeIt "isOpaque":
 reset()
 
 timeIt "subImage":
-  keep image.subImage(0, 0, 256, 256)
+  discard image.subImage(0, 0, 256, 256)
 
 reset()
 
 timeIt "superImage":
-  keep image.superImage(-10, -10, 2580, 1460)
+  discard image.superImage(-10, -10, 2580, 1460)
 
 reset()
 

tests/bench_images_draw.nim

@@ -9,7 +9,26 @@ block:
 
   timeIt "big-on-bigger NormalBlend":
     a.draw(b, translate(vec2(25, 25)), NormalBlend)
-    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 "big-on-bigger MaskBlend":
+    a.draw(b, translate(vec2(25, 25)), MaskBlend)
+
+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 "big-on-bigger OverwriteBlend":
+    a.draw(b, translate(vec2(25, 25)), OverwriteBlend)
 
 block:
   let
@@ -20,7 +39,6 @@ block:
 
   timeIt "scale x0.5":
     a.draw(b, translate(vec2(25, 25)) * scale(vec2(0.5, 0.5)), NormalBlend)
-    keep(b)
 
 block:
   let
@@ -31,7 +49,6 @@ block:
 
   timeIt "scale x2":
     a.draw(b, translate(vec2(25, 25)) * scale(vec2(2, 2)), NormalBlend)
-    keep(b)
 
 block:
   let
@@ -42,7 +59,6 @@ block:
 
   timeIt "smooth x-translate":
     a.draw(b, translate(vec2(25.2, 0)), NormalBlend)
-    keep(b)
 
 block:
   let
@@ -53,7 +69,6 @@ block:
 
   timeIt "smooth y-translate":
     a.draw(b, translate(vec2(0, 25.2)), NormalBlend)
-    keep(b)
 
 block:
   let
@@ -64,7 +79,6 @@ block:
 
   timeIt "smooth translate":
     a.draw(b, translate(vec2(25.2, 25.2)), NormalBlend)
-    keep(b)
 
 block:
   let
@@ -75,7 +89,6 @@ block:
 
   timeIt "smooth rotate 45":
     a.draw(b, translate(vec2(0, 500)) * rotate(toRadians(45)), NormalBlend)
-    keep(b)
 
 block:
   let
@@ -85,14 +98,12 @@ block:
   timeIt "shadow no offset":
     b.fill(rgba(0, 0, 0, 255))
     a.draw(b, translate(vec2(25, 25)))
-
+    discard a.shadow(
-    let shadow = a.shadow(
       offset = vec2(0, 0),
       spread = 10,
       blur = 10,
       color = rgba(0, 0, 0, 255)
     )
-    keep(shadow)
 
 block:
   let
@@ -102,11 +113,9 @@ block:
   timeIt "shadow with offset":
     b.fill(rgba(0, 0, 0, 255))
     a.draw(b, translate(vec2(25, 25)))
-
+    discard a.shadow(
-    let shadow = a.shadow(
       offset = vec2(10, 10),
       spread = 10,
       blur = 10,
       color = rgba(0, 0, 0, 255)
     )
-    keep(shadow)

tests/bench_images_draw_correct.nim

@@ -1,33 +0,0 @@
-import benchy, pixie, pixie/images {.all.}, strformat, xrays
-
-block:
-  let
-    a = newImage(1000, 1000)
-    b = newImage(500, 500)
-  a.fill(rgba(255, 0, 0, 255))
-  b.fill(rgba(0, 255, 0, 255))
-
-  a.drawCorrect(b, translate(vec2(250, 250)), blendMode = OverwriteBlend)
-  a.writeFile("tests/images/rotate0.png")
-
-block:
-  let
-    a = newImage(1000, 1000)
-    b = newImage(500, 500)
-
-  timeIt "drawCorrect":
-    a.fill(rgba(255, 0, 0, 255))
-    b.fill(rgba(0, 255, 0, 255))
-
-    a.drawCorrect(b, translate(vec2(250, 250)), blendMode = OverwriteBlend)
-
-block:
-  let
-    a = newImage(1000, 1000)
-    b = newImage(500, 500)
-
-  timeIt "draw":
-    a.fill(rgba(255, 0, 0, 255))
-    b.fill(rgba(0, 255, 0, 255))
-
-    a.draw(b, translate(vec2(250, 250)), blendMode = OverwriteBlend)

tests/bench_paths.nim

@@ -3,7 +3,7 @@ import benchy, pixie
 let pathStr = "m57.611-8.591c-1.487,1.851-4.899,4.42-1.982,6.348,0.194,0.129,0.564,0.133,0.737-0.001,2.021-1.565,4.024-2.468,6.46-3.05,0.124-0.029,0.398,0.438,0.767,0.277,1.613-0.703,3.623-0.645,4.807-1.983,3.767,0.224,7.332-0.892,10.723-2.2,1.161-0.448,2.431-1.007,3.632-1.509,1.376-0.576,2.58-1.504,3.692-2.645,0.133-0.136,0.487-0.046,0.754-0.046-0.04-0.863,0.922-0.99,1.169-1.612,0.092-0.232-0.058-0.628,0.075-0.73,2.138-1.63,3.058-3.648,1.889-6.025-0.285-0.578-0.534-1.196-1.1-1.672-1.085-0.911-2.187-0.057-3.234-0.361-0.159,0.628-0.888,0.456-1.274,0.654-0.859,0.439-2.192-0.146-3.051,0.292-1.362,0.695-2.603,0.864-4.025,1.241-0.312,0.082-1.09-0.014-1.25,0.613-0.134-0.134-0.282-0.368-0.388-0.346-1.908,0.396-3.168,0.61-4.469,2.302-0.103,0.133-0.545-0.046-0.704,0.089-0.957,0.808-1.362,2.042-2.463,2.714-0.201,0.123-0.553-0.045-0.747,0.084-0.646,0.431-1.013,1.072-1.655,1.519-0.329,0.229-0.729-0.096-0.697-0.352,0.245-1.947,0.898-3.734,0.323-5.61,2.077-2.52,4.594-4.469,6.4-7.2,0.015-2.166,0.707-4.312,0.594-6.389-0.01-0.193-0.298-0.926-0.424-1.273-0.312-0.854,0.594-1.92-0.25-2.644-1.404-1.203-2.696-0.327-3.52,1.106-1.838,0.39-3.904,1.083-5.482-0.151-1.007-0.787-1.585-1.693-2.384-2.749-0.985-1.302-0.65-2.738-0.58-4.302,0.006-0.128-0.309-0.264-0.309-0.398,0.001-0.135,0.221-0.266,0.355-0.4-0.706-0.626-0.981-1.684-2-2,0.305-1.092-0.371-1.976-1.242-2.278-1.995-0.691-3.672,1.221-5.564,1.294-0.514,0.019-0.981-1.019-1.63-1.344-0.432-0.216-1.136-0.249-1.498,0.017-0.688,0.504-1.277,0.618-2.035,0.823-1.617,0.436-2.895,1.53-4.375,2.385-1.485,0.857-2.44,2.294-3.52,3.614-0.941,1.152-1.077,3.566,0.343,4.066,1.843,0.65,3.147-2.053,5.113-1.727,0.312,0.051,0.518,0.362,0.408,0.75,0.389,0.109,0.607-0.12,0.8-0.4,0.858,1.019,2.022,1.356,2.96,2.229,0.97,0.904,2.716,0.486,3.731,1.483,1.529,1.502,0.97,4.183,2.909,5.488-0.586,1.313-1.193,2.59-1.528,4.017-0.282,1.206,0.712,2.403,1.923,2.312,1.258-0.094,1.52-0.853,2.005-1.929,0.267,0.267,0.736,0.564,0.695,0.78-0.457,2.387-1.484,4.38-1.942,6.811-0.059,0.317-0.364,0.519-0.753,0.409-0.468,4.149-4.52,6.543-7.065,9.708-0.403,0.502-0.407,1.751,0.002,2.154,1.403,1.387,3.363-0.159,5.063-0.662,0.213-1.206,1.072-2.148,2.404-2.092,0.256,0.01,0.491-0.532,0.815-0.662,0.348-0.138,0.85,0.086,1.136-0.112,1.729-1.195,3.137-2.301,4.875-3.49,0.192-0.131,0.536,0.028,0.752-0.08,0.325-0.162,0.512-0.549,0.835-0.734,0.348-0.2,0.59,0.09,0.783,0.37-0.646,0.349-0.65,1.306-1.232,1.508-0.775,0.268-1.336,0.781-2.01,1.228-0.292,0.193-0.951-0.055-1.055,0.124-0.598,1.028-1.782,1.466-2.492,2.349z"
 
 timeIt "parsePath":
-  keep parsePath(pathStr)
+  discard parsePath(pathStr)
 
 block:
   let path = parsePath("""

tests/bench_qoi.nim

@@ -3,4 +3,4 @@ import benchy, pixie/fileformats/qoi
 let data = readFile("tests/fileformats/qoi/testcard_rgba.qoi")
 
 timeIt "pixie decode":
-  keep decodeQoi(data)
+  discard decodeQoi(data)

tests/fuzz_image_draw.nim

@@ -2,30 +2,26 @@ import pixie, random
 
 randomize()
 
-for i in 0 ..< 25:
+for i in 0 ..< 250:
   let a = newImage(rand(1 .. 20), rand(1 .. 20))
   for j in 0 ..< 25:
     let b = newImage(rand(1 .. 20), rand(1 .. 20))
 
-    let
+    let translation = vec2(rand(-25..25).float32, rand(-25..25).float32)
-      translation = vec2(rand(25.0), rand(25.0)) - vec2(5, 5)
-      rotation = rand(2 * PI).float32
 
-    echo a, " ", b, " ", translation, " ", rotation
+    echo a, " ", b, " ", translation
 
     a.draw(b, translate(vec2(translation.x.trunc, translation.y.trunc)))
-    a.draw(b, translate(translation) * rotate(rotation))
+    a.draw(b, translate(translation))
 
-for i in 0 ..< 25:
+for i in 0 ..< 250:
   let a = newImage(rand(1 .. 2000), rand(1 .. 2000))
   for j in 0 ..< 25:
     let b = newImage(rand(1 .. 1000), rand(1 .. 1000))
 
-    let
+    let translation = vec2(rand(-2500..2500).float32, rand(-2500..2500).float32)
-      translation = vec2(rand(2500.0), rand(2500.0)) - vec2(500, 500)
-      rotation = rand(2 * PI).float32
 
-    echo a, " ", b, " ", translation, " ", rotation
+    echo a, " ", b, " ", translation
 
     a.draw(b, translate(vec2(translation.x.trunc, translation.y.trunc)))
-    a.draw(b, translate(translation) * rotate(rotation))
+    a.draw(b, translate(translation))

tests/fuzz_image_draw_smooth.nim

@@ -0,0 +1,31 @@
+import pixie, random
+
+randomize()
+
+for i in 0 ..< 25:
+  let a = newImage(rand(1 .. 20), rand(1 .. 20))
+  for j in 0 ..< 25:
+    let b = newImage(rand(1 .. 20), rand(1 .. 20))
+
+    let
+      translation = vec2(rand(25.0), rand(25.0)) - vec2(5, 5)
+      rotation = rand(2 * PI).float32
+
+    echo a, " ", b, " ", translation, " ", rotation
+
+    a.draw(b, translate(vec2(translation.x.trunc, translation.y.trunc)))
+    a.draw(b, translate(translation) * rotate(rotation))
+
+for i in 0 ..< 25:
+  let a = newImage(rand(1 .. 2000), rand(1 .. 2000))
+  for j in 0 ..< 25:
+    let b = newImage(rand(1 .. 1000), rand(1 .. 1000))
+
+    let
+      translation = vec2(rand(2500.0), rand(2500.0)) - vec2(500, 500)
+      rotation = rand(2 * PI).float32
+
+    echo a, " ", b, " ", translation, " ", rotation
+
+    a.draw(b, translate(vec2(translation.x.trunc, translation.y.trunc)))
+    a.draw(b, translate(translation) * rotate(rotation))

tests/test_images_draw.nim

@@ -307,3 +307,54 @@ block:
     a.draw(b, translate(translation))
 
   a.xray("tests/images/fillOptimization2.png")
+
+block:
+  let transforms = [
+    translate(vec2(0, 0)),
+    translate(vec2(50, -50)),
+    translate(vec2(50, 50)),
+    translate(vec2(-50, 50)),
+    translate(vec2(-50, -50))
+  ]
+
+  let path = newPath()
+  path.rect(10, 10, 80, 80)
+
+  let mask = newImage(100, 100)
+  mask.fillPath(path, color(1, 1, 1, 1))
+
+  let a = newImage(100, 100)
+
+  for i, transform in transforms:
+    a.fill(color(0, 0, 1, 1))
+    a.draw(mask, transform, blendMode = MaskBlend)
+    a.xray("tests/images/maskClearsOnDraw" & $i & ".png")
+
+block:
+  let transforms = [
+    translate(vec2(100, 100)),
+    translate(vec2(100, -100)),
+    translate(vec2(-100, 100)),
+    translate(vec2(-100, -100)),
+    translate(vec2(0, 1000)),
+    translate(vec2(0, -1000)),
+    translate(vec2(1000, 0)),
+    translate(vec2(-1000, 0)),
+    translate(vec2(1000, 1000)),
+    translate(vec2(1000, -1000)),
+    translate(vec2(-1000, 1000)),
+    translate(vec2(-1000, -1000)),
+  ]
+
+  let path = newPath()
+  path.rect(10, 10, 80, 80)
+
+  let mask = newImage(100, 100)
+  mask.fillPath(path, color(1, 1, 1, 1))
+
+  let a = newImage(100, 100)
+
+  for i, transform in transforms:
+    a.fill(color(0, 0, 1, 1))
+    a.draw(mask, transform, blendMode = MaskBlend)
+    doAssert a.isTransparent()