From 788b5a0821a63dd46123957e69469b8d9273b9e1 Mon Sep 17 00:00:00 2001
From: Ryan Oldenburg <guzba8@gmail.com>
Date: Mon, 8 Feb 2021 20:06:07 -0600
Subject: [PATCH 1/2] consistent versions arcoss all combos

---
 src/pixie/images.nim | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

diff --git a/src/pixie/images.nim b/src/pixie/images.nim
index fef0bba..44a3850 100644
--- a/src/pixie/images.nim
+++ b/src/pixie/images.nim
@@ -365,12 +365,20 @@ proc draw*(
 ) {.inline.} =
   image.drawCorrect(mask, translate(pos), blendMode)
 
-proc draw*(a, b: Mask, mat = mat3(), blendMode = bmMask) =
+proc draw*(a, b: Mask, mat: Mat3, blendMode = bmMask) =
   a.drawCorrect(b, mat, blendMode)
 
-proc draw*(mask: Mask, image: Image, mat = mat3(), blendMode = bmMask) =
+proc draw*(a, b: Mask, pos = vec2(0, 0), blendMode = bmMask) {.inline.} =
+  a.draw(b, translate(pos), blendMode)
+
+proc draw*(mask: Mask, image: Image, mat: Mat3, blendMode = bmMask) =
   mask.drawCorrect(image, mat, blendMode)
 
+proc draw*(
+  mask: Mask, image: Image, pos = vec2(0, 0), blendMode = bmMask
+) {.inline.} =
+  mask.draw(image, translate(pos), blendMode)
+
 when defined(release):
   {.pop.}
 

From 6c5fa78ed968fd2463b388f6c211e0d01f2f45ac Mon Sep 17 00:00:00 2001
From: Ryan Oldenburg <guzba8@gmail.com>
Date: Mon, 8 Feb 2021 21:12:28 -0600
Subject: [PATCH 2/2] mask applyOpacity 10x faster

---
 src/pixie/masks.nim  | 43 ++++++++++++++++++++++++++++++++++++++++---
 tests/test_masks.nim |  7 +++++++
 2 files changed, 47 insertions(+), 3 deletions(-)

diff --git a/src/pixie/masks.nim b/src/pixie/masks.nim
index 93cf1af..805d621 100644
--- a/src/pixie/masks.nim
+++ b/src/pixie/masks.nim
@@ -1,5 +1,8 @@
 import blends, 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.
@@ -111,9 +114,43 @@ proc getValueSmooth*(mask: Mask, x, y: float32): uint8 =
 
 proc applyOpacity*(mask: Mask, opacity: float32) =
   ## Multiplies the values of the mask by opacity.
-  let opacity = round(255 * opacity).uint32
-  for value in mask.data.mitems:
-    value = ((value * opacity) div 255).uint8
+  let opacity = round(255 * opacity).uint16
+
+  var i: int
+  when defined(amd64) and not defined(pixieNoSimd):
+    let
+      oddMask = mm_set1_epi16(cast[int16](0xff00))
+      div255 = mm_set1_epi16(cast[int16](0x8081))
+      vOpacity = mm_slli_epi16(mm_set1_epi16(cast[int16](opacity)), 8)
+
+    for _ in countup(i, mask.data.len - 16, 16):
+      var values = mm_loadu_si128(mask.data[i].addr)
+
+      let eqZero = mm_cmpeq_epi16(values, mm_setzero_si128())
+      if mm_movemask_epi8(eqZero) != 0xffff:
+        var
+          valuesEven = mm_slli_epi16(mm_andnot_si128(oddMask, values), 8)
+          valuesOdd = mm_and_si128(values, oddMask)
+
+        # values * opacity
+        valuesEven = mm_mulhi_epu16(valuesEven, vOpacity)
+        valuesOdd = mm_mulhi_epu16(valuesOdd, vOpacity)
+
+        # div 255
+        valuesEven = mm_srli_epi16(mm_mulhi_epu16(valuesEven, div255), 7)
+        valuesOdd = mm_srli_epi16(mm_mulhi_epu16(valuesOdd, div255), 7)
+
+        valuesOdd = mm_slli_epi16(valuesOdd, 8)
+
+        mm_storeu_si128(
+          mask.data[i].addr,
+          mm_or_si128(valuesEven, valuesOdd)
+        )
+
+      i += 16
+
+  for j in i ..< mask.data.len:
+    mask.data[j] = ((mask.data[j] * opacity) div 255).uint8
 
 when defined(release):
   {.pop.}
diff --git a/tests/test_masks.nim b/tests/test_masks.nim
index e487a5b..02b5aaa 100644
--- a/tests/test_masks.nim
+++ b/tests/test_masks.nim
@@ -1,5 +1,12 @@
 import chroma, pixie, pixie/fileformats/png
 
+block:
+  let mask = newMask(100, 100)
+  mask.fill(200)
+  mask.applyOpacity(0.5)
+  doAssert mask[0, 0] == 100
+  doAssert mask[88, 88] == 100
+
 block:
   let
     mask = newMask(100, 100)