Merge pull request #285 from guzba/master

simd draw bugfix

README.md

@@ -282,16 +282,11 @@ image.fillPath(path, paint)
 ### Shadow
 [examples/shadow.nim](examples/shadow.nim)
 ```nim
+let path = newPath()
+path.polygon(vec2(100, 100), 70, sides = 8)
+
 let polygonImage = newImage(200, 200)
-
-let ctx = newContext(polygonImage)
-ctx.fillStyle = rgba(255, 255, 255, 255)
-
-ctx.fillPolygon(
-  vec2(100, 100),
-  70,
-  sides = 8
-)
+polygonImage.fillPath(path, rgba(255, 255, 255, 255))
 
 let shadow = polygonImage.shadow(
   offset = vec2(2, 2),

src/pixie/images.nim

@@ -757,24 +757,33 @@ proc drawUber(
           when type(a) is Image:
             if blendMode.hasSimdBlender():
               let blenderSimd = blendMode.blenderSimd()
-              for _ in countup(x, xMax - 4, 4):
+              for _ in countup(x, xMax - 16, 16):
                 let
                   srcPos = p + dx * x.float32 + dy * y.float32
                   sx = srcPos.x.int
                   sy = srcPos.y.int
-                  backdrop = mm_loadu_si128(a.data[a.dataIndex(x, y)].addr)
                 when type(b) is Image:
-                  let source = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
+                  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)
+                    )
                 else: # b is a Mask
-                  # Need to move 4 mask values into the alpha slots
-                  var source = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
-                  source = unpackAlphaValues(source)
-
-                mm_storeu_si128(
-                  a.data[a.dataIndex(x, y)].addr,
-                  blenderSimd(backdrop, source)
-                )
-                x += 4
+                  var values = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
+                  for q in [0, 4, 8, 12]:
+                    let
+                      backdrop = mm_loadu_si128(a.data[a.dataIndex(x + q, y)].addr)
+                      source = unpackAlphaValues(values)
+                    mm_storeu_si128(
+                      a.data[a.dataIndex(x + q, y)].addr,
+                      blenderSimd(backdrop, source)
+                    )
+                    # Shuffle 32 bits off for the next iteration
+                    values = mm_srli_si128(values, 4)
+                x += 16
           else: # is a Mask
             if blendMode.hasSimdMasker():
               let maskerSimd = blendMode.maskerSimd()

src/pixie/paths.nim

@@ -1401,9 +1401,11 @@ proc fillHits(
   windingRule: WindingRule,
   blendMode: BlendMode
 ) =
-  let blender = blendMode.blender()
+  let
+    blender = blendMode.blender()
+    width = image.width.float32
   var filledTo: int
-  for (prevAt, at, count) in hits.walk(numHits, windingRule, y, image.width.float32):
+  for (prevAt, at, count) in hits.walk(numHits, windingRule, y, width):
     let
       fillStart = prevAt.int
       fillLen = at.int - fillStart
@@ -1449,9 +1451,11 @@ proc fillHits(
   windingRule: WindingRule,
   blendMode: BlendMode
 ) =
-  let masker = blendMode.masker()
+  let
+    masker = blendMode.masker()
+    width = mask.width.float32
   var filledTo: int
-  for (prevAt, at, count) in hits.walk(numHits, windingRule, y, mask.width.float32):
+  for (prevAt, at, count) in hits.walk(numHits, windingRule, y, width):
     let
       fillStart = prevAt.int
       fillLen = at.int - fillStart

tests/benchmark_images_draw.nim

@@ -95,7 +95,7 @@ block:
 
   timeIt "shadow":
     b.fill(rgba(0, 0, 0, 255))
-    a.draw(b, vec2(25, 25))
+    a.draw(b, translate(vec2(25, 25)))
 
     let shadow = a.shadow(
       offset = vec2(0, 0),