Merge pull request #476 from treeform/dev

* De-flake some tests.
* Blit strategy.

src/pixie/images.nim

@@ -454,16 +454,63 @@ 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 | Mask, transform = mat3(), blendMode = NormalBlend, tiled = false
 ) {.raises: [PixieError].} =
   ## Draws one image onto another using matrix with color blending.
 
-  when type(a) is Image:
-    let blender = blendMode.blender()
-  else: # a is a Mask
-    let masker = blendMode.masker()
-
   var
     inverseTransform = transform.inverse()
     # Compute movement vectors
@@ -489,6 +536,25 @@ proc drawCorrect(
     filterBy2 *= 2
     inverseTransform = scale(vec2(2, 2)) * inverseTransform
 
+  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
+    )
+
+  when type(a) is Image and type(b) is Image:
+    if not hasRotationOrScaling and not smooth and not tiled:
+      blitRect(a, b, ivec2(transform[2, 0].int32, transform[2, 1].int32), blendMode)
+      return
+
+  when type(a) is Image:
+    let blender = blendMode.blender()
+  else: # a is a Mask
+    let masker = blendMode.masker()
+
   for y in 0 ..< a.height:
     for x in 0 ..< a.width:
       let

src/pixie/simd/sse2.nim

@@ -417,5 +417,50 @@ proc applyOpacitySse2*(mask: Mask, opacity: float32) {.simd.} =
   for i in i ..< mask.data.len:
     mask.data[i] = ((mask.data[i] * opacity) div 255).uint8
 
+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))
+
+    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)
+
+    let done = mm_add_epi8(
+      source,
+      mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
+    )
+
+    mm_storeu_si128(a[i].addr, done)
+
+    i += 4
+
+  # TODO last 1-3 pixels
+  # for i in i ..< len:
+  #   a[i] = blendNormal(a[i], b[i])
+
 when defined(release):
   {.pop.}

tests/bench_blends.nim

@@ -1,8 +1,8 @@
 import benchy, chroma, pixie/blends, pixie/images, vmath
 
 let
-  backdrop = newImage(512, 512)
-  source = newImage(512, 512)
+  backdrop = newImage(256, 256)
+  source = newImage(256, 256)
 source.fill(rgba(100, 100, 100, 100))
 
 template reset() =

tests/bench_images.nim

@@ -134,9 +134,9 @@ timeIt "mix integers":
     doAssert z > 0
 
 timeIt "mix floats":
-  for i in 0 ..< 100000:
-    let c = image[0, 0]
+  for i in 0 ..< 1000:
+    let c = image[0, 0].color
     var z: int
     for t in 0 .. 100:
-      z += mix(c.color, c.color, t.float32 / 100).rgba().a.int
+      z += mix(c, c, t.float32 / 100).rgba().a.int
     doAssert z > 0

tests/bench_images_draw_correct.nim

@@ -0,0 +1,37 @@
+import pixie, strformat, xrays
+import pixie/images {.all.}
+
+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")
+
+import benchy
+
+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)