treeform · treeform · Jul 29, 2022 · Jul 29, 2022 · Jul 29, 2022 · Jul 29, 2022
diff --git a/src/pixie/common.nim b/src/pixie/common.nim
@@ -56,29 +56,24 @@ proc copy*(image: Image): Image {.raises: [].} =
 template dataIndex*(image: Image, x, y: int): int =
   image.width * y + x
 
-proc mix*(a, b: uint8, t: float32): uint8 {.inline, raises: [].} =
-  ## Linearly interpolate between a and b using t.
-  let t = round(t * 255).uint32
-  ((a * (255 - t) + b * t) div 255).uint8
-
 proc mix*(a, b: ColorRGBX, t: float32): ColorRGBX {.inline, raises: [].} =
   ## Linearly interpolate between a and b using t.
   let x = round(t * 255).uint32
-  result.r = ((a.r.uint32 * (255 - x) + b.r.uint32 * x) div 255).uint8
-  result.g = ((a.g.uint32 * (255 - x) + b.g.uint32 * x) div 255).uint8
-  result.b = ((a.b.uint32 * (255 - x) + b.b.uint32 * x) div 255).uint8
-  result.a = ((a.a.uint32 * (255 - x) + b.a.uint32 * x) div 255).uint8
+  result.r = ((a.r.uint32 * (255 - x) + b.r.uint32 * x + 127) div 255).uint8
+  result.g = ((a.g.uint32 * (255 - x) + b.g.uint32 * x + 127) div 255).uint8
+  result.b = ((a.b.uint32 * (255 - x) + b.b.uint32 * x + 127) div 255).uint8
+  result.a = ((a.a.uint32 * (255 - x) + b.a.uint32 * x + 127) div 255).uint8
 
 proc `*`*(color: ColorRGBX, opacity: float32): ColorRGBX {.raises: [].} =
   if opacity == 0:
     rgbx(0, 0, 0, 0)
   else:
     let
       x = round(opacity * 255).uint32
-      r = ((color.r * x) div 255).uint8
-      g = ((color.g * x) div 255).uint8
-      b = ((color.b * x) div 255).uint8
-      a = ((color.a * x) div 255).uint8
+      r = ((color.r * x + 127) div 255).uint8
+      g = ((color.g * x + 127) div 255).uint8
+      b = ((color.b * x + 127) div 255).uint8
+      a = ((color.a * x + 127) div 255).uint8
     rgbx(r, g, b, a)
 
 proc snapToPixels*(rect: Rect): Rect {.raises: [].} =

diff --git a/src/pixie/images.nim b/src/pixie/images.nim
@@ -192,10 +192,10 @@ proc minifyBy2*(
           c = src.data[bottomRowStart + x * 2 + 1]
           d = src.data[bottomRowStart + x * 2]
           mixed = rgbx(
-            ((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
+            ((a.r.uint32 + b.r + c.r + d.r + 2) div 4).uint8,
+            ((a.g.uint32 + b.g + c.g + d.g + 2) div 4).uint8,
+            ((a.b.uint32 + b.b + c.b + d.b + 2) div 4).uint8,
+            ((a.a.uint32 + b.a + c.a + d.a + 2) div 4).uint8
           )
         result.data[result.dataIndex(x, y)] = mixed
 

diff --git a/src/pixie/simd/avx2.nim b/src/pixie/simd/avx2.nim
@@ -296,6 +296,7 @@ proc minifyBy2Avx2*(image: Image, power = 1): Image {.simd.} =
     )
     let
       oddMask = mm256_set1_epi16(0xff00)
+      vec2 = mm256_set1_epi16(2)
       permuteControl = mm256_set_epi32(7, 7, 7, 7, 6, 4, 2, 0)
     for y in 0 ..< resultEvenHeight:
       let
@@ -323,8 +324,10 @@ proc minifyBy2Avx2*(image: Image, power = 1): Image {.simd.} =
           bottomAddedOdd = mm256_add_epi16(bottomOdd, bottomShiftedOdd)
           addedEven = mm256_add_epi16(topAddedEven, bottomAddedEven)
           addedOdd = mm256_add_epi16(topAddedOdd, bottomAddedOdd)
-          addedEvenDiv4 = mm256_srli_epi16(addedEven, 2)
-          addedOddDiv4 = mm256_srli_epi16(addedOdd, 2)
+          addedEvenRounding = mm256_add_epi16(addedEven, vec2)
+          addedOddRounding = mm256_add_epi16(addedOdd, vec2)
+          addedEvenDiv4 = mm256_srli_epi16(addedEvenRounding, 2)
+          addedOddDiv4 = mm256_srli_epi16(addedOddRounding, 2)
           merged = mm256_or_si256(addedEvenDiv4, mm256_slli_epi16(addedOddDiv4, 8))
           # Merged has the correct values for the next two pixels at
           # index 0, 2, 4, 6 so permute into position and store
@@ -342,10 +345,10 @@ proc minifyBy2Avx2*(image: Image, power = 1): Image {.simd.} =
           c = src.data[bottomRowStart + x * 2 + 1]
           d = src.data[bottomRowStart + x * 2]
           mixed = rgbx(
-            ((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
+            ((a.r.uint32 + b.r + c.r + d.r + 2) div 4).uint8,
+            ((a.g.uint32 + b.g + c.g + d.g + 2) div 4).uint8,
+            ((a.b.uint32 + b.b + c.b + d.b + 2) div 4).uint8,
+            ((a.a.uint32 + b.a + c.a + d.a + 2) div 4).uint8
           )
         result.data[result.dataIndex(x, y)] = mixed
 

diff --git a/src/pixie/simd/sse2.nim b/src/pixie/simd/sse2.nim
@@ -350,19 +350,33 @@ proc minifyBy2Sse2*(image: Image, power = 1): Image {.simd.} =
       if srcWidthIsOdd: resultEvenWidth + 1 else: resultEvenWidth,
       if srcHeightIsOdd: resultEvenHeight + 1 else: resultEvenHeight
     )
-    let oddMask = mm_set1_epi16(0xff00)
+    let
+      oddMask = mm_set1_epi16(0xff00)
+      loMask = mm_set_epi32(0, 0, uint32.high, uint32.high)
+      hiMask = mm_set_epi32(uint32.high, uint32.high, 0, 0)
+      vec2 = mm_set1_epi16(2)
     for y in 0 ..< resultEvenHeight:
       let
         topRowStart = src.dataIndex(0, y * 2)
         bottomRowStart = src.dataIndex(0, y * 2 + 1)
 
+      template loadEven(src: Image, idx: int): M128i =
+        var
+          a = mm_loadu_si128(src.data[idx].addr)
+          b = mm_loadu_si128(src.data[idx + 4].addr)
+        a = mm_shuffle_epi32(a, MM_SHUFFLE(3, 3, 2, 0))
+        b = mm_shuffle_epi32(b, MM_SHUFFLE(2, 0, 3, 3))
+        a = mm_and_si128(a, loMask)
+        b = mm_and_si128(b, hiMask)
+        mm_or_si128(a, b)
+
       var x: int
-      while x <= resultEvenWidth - 4:
+      while x <= resultEvenWidth - 9:
         let
-          top = mm_loadu_si128(src.data[topRowStart + x * 2].addr)
-          bottom = mm_loadu_si128(src.data[bottomRowStart + x * 2].addr)
-          topShifted = mm_srli_si128(top, 4)
-          bottomShifted = mm_srli_si128(bottom, 4)
+          top = loadEven(src, topRowStart + x * 2)
+          bottom = loadEven(src, bottomRowStart + x * 2)
+          topShifted = loadEven(src, topRowStart + x * 2 + 1)
+          bottomShifted = loadEven(src, bottomRowStart + x * 2 + 1)
           topEven = mm_andnot_si128(oddMask, top)
           topOdd = mm_srli_epi16(top, 8)
           bottomEven = mm_andnot_si128(oddMask, bottom)
@@ -377,15 +391,13 @@ proc minifyBy2Sse2*(image: Image, power = 1): Image {.simd.} =
           bottomAddedOdd = mm_add_epi16(bottomOdd, bottomShiftedOdd)
           addedEven = mm_add_epi16(topAddedEven, bottomAddedEven)
           addedOdd = mm_add_epi16(topAddedOdd, bottomAddedOdd)
-          addedEvenDiv4 = mm_srli_epi16(addedEven, 2)
-          addedOddDiv4 = mm_srli_epi16(addedOdd, 2)
+          addedEvenRounding = mm_add_epi16(addedEven, vec2)
+          addedOddRounding = mm_add_epi16(addedOdd, vec2)
+          addedEvenDiv4 = mm_srli_epi16(addedEvenRounding, 2)
+          addedOddDiv4 = mm_srli_epi16(addedOddRounding, 2)
           merged = mm_or_si128(addedEvenDiv4, mm_slli_epi16(addedOddDiv4, 8))
-          # Merged has the correct values for the next two pixels at
-          # index 0 and 2 so shift 0 and 2 into position and store
-          shuffled = mm_shuffle_epi32(merged, MM_SHUFFLE(3, 3, 2, 0))
-          lower = mm_cvtsi128_si64(shuffled)
-        copyMem(result.data[result.dataIndex(x, y)].addr, lower.unsafeAddr, 8)
-        x += 2
+        mm_storeu_si128(result.data[result.dataIndex(x, y)].addr, merged)
+        x += 4
 
       for x in x ..< resultEvenWidth:
         let
@@ -394,10 +406,10 @@ proc minifyBy2Sse2*(image: Image, power = 1): Image {.simd.} =
           c = src.data[bottomRowStart + x * 2 + 1]
           d = src.data[bottomRowStart + x * 2]
           mixed = rgbx(
-            ((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
+            ((a.r.uint32 + b.r + c.r + d.r + 2) div 4).uint8,
+            ((a.g.uint32 + b.g + c.g + d.g + 2) div 4).uint8,
+            ((a.b.uint32 + b.b + c.b + d.b + 2) div 4).uint8,
+            ((a.a.uint32 + b.a + c.a + d.a + 2) div 4).uint8
           )
         result.data[result.dataIndex(x, y)] = mixed
 

diff --git a/tests/contexts/beginPath_1.png b/tests/contexts/beginPath_1.png
diff --git a/tests/contexts/bezierCurveTo_1.png b/tests/contexts/bezierCurveTo_1.png
diff --git a/tests/contexts/bezierCurveTo_2.png b/tests/contexts/bezierCurveTo_2.png
diff --git a/tests/contexts/blendmode_1.png b/tests/contexts/blendmode_1.png
diff --git a/tests/contexts/clearRect_1.png b/tests/contexts/clearRect_1.png
diff --git a/tests/contexts/clip_1.png b/tests/contexts/clip_1.png
diff --git a/tests/contexts/clip_1b.png b/tests/contexts/clip_1b.png
diff --git a/tests/contexts/clip_1c.png b/tests/contexts/clip_1c.png
diff --git a/tests/contexts/clip_1d.png b/tests/contexts/clip_1d.png
diff --git a/tests/contexts/clip_1e.png b/tests/contexts/clip_1e.png
diff --git a/tests/contexts/clip_1f.png b/tests/contexts/clip_1f.png
diff --git a/tests/contexts/clip_2.png b/tests/contexts/clip_2.png
diff --git a/tests/contexts/clip_3.png b/tests/contexts/clip_3.png
diff --git a/tests/contexts/clip_text.png b/tests/contexts/clip_text.png
diff --git a/tests/contexts/closePath_1.png b/tests/contexts/closePath_1.png
diff --git a/tests/contexts/draw_image.png b/tests/contexts/draw_image.png
diff --git a/tests/contexts/draw_image_rhino.png b/tests/contexts/draw_image_rhino.png
diff --git a/tests/contexts/draw_image_rhino2.png b/tests/contexts/draw_image_rhino2.png
diff --git a/tests/contexts/draw_image_scaled.png b/tests/contexts/draw_image_scaled.png
diff --git a/tests/contexts/draw_image_self_scaled.png b/tests/contexts/draw_image_self_scaled.png
diff --git a/tests/contexts/draw_image_translated.png b/tests/contexts/draw_image_translated.png
diff --git a/tests/contexts/ellipse_1.png b/tests/contexts/ellipse_1.png
diff --git a/tests/contexts/fillText_1.png b/tests/contexts/fillText_1.png
diff --git a/tests/contexts/fill_1.png b/tests/contexts/fill_1.png
diff --git a/tests/contexts/globalAlpha_1.png b/tests/contexts/globalAlpha_1.png
diff --git a/tests/contexts/moveTo_1.png b/tests/contexts/moveTo_1.png
diff --git a/tests/contexts/paintSaveRestore.png b/tests/contexts/paintSaveRestore.png
diff --git a/tests/contexts/quadracticCurveTo_1.png b/tests/contexts/quadracticCurveTo_1.png
diff --git a/tests/contexts/quadracticCurveTo_2.png b/tests/contexts/quadracticCurveTo_2.png
diff --git a/tests/contexts/resetTransform_1.png b/tests/contexts/resetTransform_1.png
diff --git a/tests/contexts/resetTransform_2.png b/tests/contexts/resetTransform_2.png
diff --git a/tests/contexts/rotate_1.png b/tests/contexts/rotate_1.png
diff --git a/tests/contexts/save_1.png b/tests/contexts/save_1.png
diff --git a/tests/contexts/scale_1.png b/tests/contexts/scale_1.png
diff --git a/tests/contexts/setLineDash_1.png b/tests/contexts/setLineDash_1.png
diff --git a/tests/contexts/setTransform_1.png b/tests/contexts/setTransform_1.png
diff --git a/tests/contexts/strokeRect_1.png b/tests/contexts/strokeRect_1.png
diff --git a/tests/contexts/strokeRect_2.png b/tests/contexts/strokeRect_2.png
diff --git a/tests/contexts/strokeText_1.png b/tests/contexts/strokeText_1.png
diff --git a/tests/contexts/stroke_1.png b/tests/contexts/stroke_1.png
diff --git a/tests/contexts/stroke_2.png b/tests/contexts/stroke_2.png
diff --git a/tests/contexts/stroke_3.png b/tests/contexts/stroke_3.png
diff --git a/tests/contexts/textBaseline_1.png b/tests/contexts/textBaseline_1.png
diff --git a/tests/contexts/translate_1.png b/tests/contexts/translate_1.png
diff --git a/tests/fileformats/gif/3x5.png b/tests/fileformats/gif/3x5.png
diff --git a/tests/fileformats/gif/audrey.png b/tests/fileformats/gif/audrey.png
diff --git a/tests/fileformats/gif/sunflower.png b/tests/fileformats/gif/sunflower.png
diff --git a/tests/images/drawEllipse.png b/tests/images/drawEllipse.png
diff --git a/tests/images/drawPolygon.png b/tests/images/drawPolygon.png
diff --git a/tests/images/drawRect.png b/tests/images/drawRect.png
diff --git a/tests/images/drawRoundedRect.png b/tests/images/drawRoundedRect.png
diff --git a/tests/images/drawSegment.png b/tests/images/drawSegment.png
diff --git a/tests/images/fillOptimization.png b/tests/images/fillOptimization.png
diff --git a/tests/images/fillOptimization2.png b/tests/images/fillOptimization2.png
diff --git a/tests/images/flipped1.png b/tests/images/flipped1.png
diff --git a/tests/images/flipped2.png b/tests/images/flipped2.png
diff --git a/tests/images/flipped3.png b/tests/images/flipped3.png
diff --git a/tests/images/imageblur20.png b/tests/images/imageblur20.png
diff --git a/tests/images/imageblur20oob.png b/tests/images/imageblur20oob.png
diff --git a/tests/images/magnifiedBy2.png b/tests/images/magnifiedBy2.png
diff --git a/tests/images/magnifiedBy4.png b/tests/images/magnifiedBy4.png
diff --git a/tests/images/minifiedBy2.png b/tests/images/minifiedBy2.png
diff --git a/tests/images/minifiedBy4.png b/tests/images/minifiedBy4.png
diff --git a/tests/images/minifiedMandrill.png b/tests/images/minifiedMandrill.png
diff --git a/tests/images/rotate180.png b/tests/images/rotate180.png
diff --git a/tests/images/rotate270.png b/tests/images/rotate270.png
diff --git a/tests/images/rotate360.png b/tests/images/rotate360.png
diff --git a/tests/images/rotate90.png b/tests/images/rotate90.png
diff --git a/tests/images/scaleHalf.png b/tests/images/scaleHalf.png
diff --git a/tests/images/strokeEllipse.png b/tests/images/strokeEllipse.png
diff --git a/tests/images/strokePolygon.png b/tests/images/strokePolygon.png
diff --git a/tests/images/strokeRect.png b/tests/images/strokeRect.png
diff --git a/tests/images/strokeRoundedRect.png b/tests/images/strokeRoundedRect.png
diff --git a/tests/images/superimage1.png b/tests/images/superimage1.png
diff --git a/tests/images/superimage2.png b/tests/images/superimage2.png
diff --git a/tests/images/superimage3.png b/tests/images/superimage3.png
diff --git a/tests/images/superimage4.png b/tests/images/superimage4.png
diff --git a/tests/images/superimage5.png b/tests/images/superimage5.png
diff --git a/tests/images/superimage6.png b/tests/images/superimage6.png
diff --git a/tests/paths/ButtCap.png b/tests/paths/ButtCap.png
diff --git a/tests/paths/RoundCap.png b/tests/paths/RoundCap.png
diff --git a/tests/paths/SquareCap.png b/tests/paths/SquareCap.png
diff --git a/tests/paths/arc.png b/tests/paths/arc.png
diff --git a/tests/paths/arcTo1.png b/tests/paths/arcTo1.png
diff --git a/tests/paths/arcTo2.png b/tests/paths/arcTo2.png
diff --git a/tests/paths/arcTo3.png b/tests/paths/arcTo3.png
diff --git a/tests/paths/boxBevel.png b/tests/paths/boxBevel.png
diff --git a/tests/paths/boxMiter.png b/tests/paths/boxMiter.png
diff --git a/tests/paths/boxRound.png b/tests/paths/boxRound.png
diff --git a/tests/paths/dashes.png b/tests/paths/dashes.png
diff --git a/tests/paths/fillImagePaint.png b/tests/paths/fillImagePaint.png
diff --git a/tests/paths/fillTiledImagePaint.png b/tests/paths/fillTiledImagePaint.png
diff --git a/tests/paths/gradientAngular.png b/tests/paths/gradientAngular.png
diff --git a/tests/paths/gradientAngularOpacity.png b/tests/paths/gradientAngularOpacity.png
diff --git a/tests/paths/gradientLinear.png b/tests/paths/gradientLinear.png
diff --git a/tests/paths/gradientLinear2.png b/tests/paths/gradientLinear2.png
diff --git a/tests/paths/gradientRadial.png b/tests/paths/gradientRadial.png
diff --git a/tests/paths/miterLimit_10deg_2.00num.png b/tests/paths/miterLimit_10deg_2.00num.png
diff --git a/tests/paths/miterLimit_145deg_2.00num.png b/tests/paths/miterLimit_145deg_2.00num.png
diff --git a/tests/paths/miterLimit_145deg_3.32num.png b/tests/paths/miterLimit_145deg_3.32num.png
diff --git a/tests/paths/miterLimit_145deg_3.33num.png b/tests/paths/miterLimit_145deg_3.33num.png
diff --git a/tests/paths/miterLimit_155deg_2.00num.png b/tests/paths/miterLimit_155deg_2.00num.png
diff --git a/tests/paths/miterLimit_165deg_10.00num.png b/tests/paths/miterLimit_165deg_10.00num.png
diff --git a/tests/paths/miterLimit_165deg_2.00num.png b/tests/paths/miterLimit_165deg_2.00num.png
diff --git a/tests/paths/opacityFill.png b/tests/paths/opacityFill.png
diff --git a/tests/paths/opacityStroke.png b/tests/paths/opacityStroke.png
diff --git a/tests/paths/paintImage.png b/tests/paths/paintImage.png
diff --git a/tests/paths/paintImageOpacity.png b/tests/paths/paintImageOpacity.png
diff --git a/tests/paths/paintImageTiled.png b/tests/paths/paintImageTiled.png
diff --git a/tests/paths/paintImageTiledOpacity.png b/tests/paths/paintImageTiledOpacity.png
diff --git a/tests/paths/paintSolid.png b/tests/paths/paintSolid.png
diff --git a/tests/paths/path0pxCover.png b/tests/paths/path0pxCover.png
diff --git a/tests/paths/path1pxCover.png b/tests/paths/path1pxCover.png
diff --git a/tests/paths/pathBlackRectangle.png b/tests/paths/pathBlackRectangle.png
diff --git a/tests/paths/pathBlackRectangleZ.png b/tests/paths/pathBlackRectangleZ.png
diff --git a/tests/paths/pathBottomArc.png b/tests/paths/pathBottomArc.png
diff --git a/tests/paths/pathCornerArc.png b/tests/paths/pathCornerArc.png
diff --git a/tests/paths/pathHeart.png b/tests/paths/pathHeart.png
diff --git a/tests/paths/pathInvertedCornerArc.png b/tests/paths/pathInvertedCornerArc.png
diff --git a/tests/paths/pathRedRectangle.png b/tests/paths/pathRedRectangle.png
diff --git a/tests/paths/pathRotatedArc.png b/tests/paths/pathRotatedArc.png
diff --git a/tests/paths/pathStroke1.png b/tests/paths/pathStroke1.png
diff --git a/tests/paths/pathStroke1Big.png b/tests/paths/pathStroke1Big.png
diff --git a/tests/paths/pathStroke2.png b/tests/paths/pathStroke2.png
diff --git a/tests/paths/pathStroke3.png b/tests/paths/pathStroke3.png
diff --git a/tests/paths/pathSwish.png b/tests/paths/pathSwish.png
diff --git a/tests/paths/pathYellowRectangle.png b/tests/paths/pathYellowRectangle.png
diff --git a/tests/paths/pixelScale.png b/tests/paths/pixelScale.png
diff --git a/tests/paths/rectExcludeMask.png b/tests/paths/rectExcludeMask.png
diff --git a/tests/paths/rectExcludeMaskAA.png b/tests/paths/rectExcludeMaskAA.png
diff --git a/tests/paths/rectMask.png b/tests/paths/rectMask.png
diff --git a/tests/paths/rectMaskAA.png b/tests/paths/rectMaskAA.png
diff --git a/tests/paths/rectMaskStroke.png b/tests/paths/rectMaskStroke.png
diff --git a/tests/paths/selfclosing.png b/tests/paths/selfclosing.png