From ec5bc9ba1eea55b5190d4bfefa3810c932a60351 Mon Sep 17 00:00:00 2001
From: Ryan Oldenburg <guzba8@gmail.com>
Date: Tue, 8 Dec 2020 18:10:20 -0600
Subject: [PATCH] simd helping

---
 pixie.nimble               |  1 +
 src/pixie/images.nim       | 23 +++++++++++---
 tests/benchmark_images.nim | 65 ++++++++++++++++++++++++++++++--------
 3 files changed, 72 insertions(+), 17 deletions(-)

diff --git a/pixie.nimble b/pixie.nimble
index c759ac4..8c24e54 100644
--- a/pixie.nimble
+++ b/pixie.nimble
@@ -10,3 +10,4 @@ requires "vmath >= 0.3.3"
 requires "chroma >= 0.1.5"
 requires "zippy >= 0.3.5"
 requires "flatty >= 0.1.2"
+requires "nimsimd >= 0.4.5"
diff --git a/src/pixie/images.nim b/src/pixie/images.nim
index aee7451..8e12fa8 100644
--- a/src/pixie/images.nim
+++ b/src/pixie/images.nim
@@ -1,4 +1,4 @@
-import chroma, blends, vmath, common, system/memory
+import chroma, blends, vmath, common, system/memory, nimsimd/sse2
 
 type
   Image* = ref object
@@ -73,16 +73,31 @@ proc fill*(image: Image, rgba: ColorRgba) =
   if rgba.r == rgba.g and rgba.r == rgba.b and rgba.r == rgba.a:
     nimSetMem(image.data[0].addr, rgba.r.cint, image.data.len * 4)
   else:
-    for c in image.data.mitems:
-      c = rgba
+    # SIMD fill until we run out of room.
+    let m = mm_set1_epi32(cast[int32](rgba))
+    var i: int
+    while i < image.data.len - 4:
+      mm_store_si128(image.data[i].addr, m)
+      i += 4
+    for j in i ..< image.data.len:
+      image.data[j] = rgba
 
 proc invert*(image: Image) =
   ## Inverts all of the colors and alpha.
-  for rgba in image.data.mitems:
+  let vec255 = mm_set1_epi8(255)
+  var i: int
+  while i < image.data.len - 4:
+    var m = mm_loadu_si128(image.data[i].addr)
+    m = mm_sub_epi8(vec255, m)
+    mm_store_si128(image.data[i].addr, m)
+    i += 4
+  for j in i ..< image.data.len:
+    var rgba = image.data[j]
     rgba.r = 255 - rgba.r
     rgba.g = 255 - rgba.g
     rgba.b = 255 - rgba.b
     rgba.a = 255 - rgba.a
+    image.data[j] = rgba
 
 proc subImage*(image: Image, x, y, w, h: int): Image =
   ## Gets a sub image of the main image.
diff --git a/tests/benchmark_images.nim b/tests/benchmark_images.nim
index 0e9b9d9..606b4f3 100644
--- a/tests/benchmark_images.nim
+++ b/tests/benchmark_images.nim
@@ -1,11 +1,19 @@
-import chroma, pixie, benchy
+import chroma, pixie, benchy, system/memory
 
-proc fillOriginal(a: Image, rgba: ColorRGBA) =
+proc fill1(a: Image, rgba: ColorRGBA) =
   for y in 0 ..< a.height:
     for x in 0 ..< a.width:
       a.setRgbaUnsafe(x, y, rgba)
 
-proc invertOriginal(a: Image) =
+proc fill2*(image: Image, rgba: ColorRgba) =
+  ## Fills the image with a solid color.
+  if rgba.r == rgba.g and rgba.r == rgba.b and rgba.r == rgba.a:
+    nimSetMem(image.data[0].addr, rgba.r.cint, image.data.len * 4)
+  else:
+    for c in image.data.mitems:
+      c = rgba
+
+proc invert1(a: Image) =
   for y in 0 ..< a.height:
     for x in 0 ..< a.width:
       var rgba = a.getRgbaUnsafe(x, y)
@@ -15,7 +23,14 @@ proc invertOriginal(a: Image) =
       rgba.a = 255 - rgba.a
       a.setRgbaUnsafe(x, y, rgba)
 
-proc applyOpacityOriginal(a: Image, opacity: float32): Image =
+proc invert2*(image: Image) =
+  for rgba in image.data.mitems:
+    rgba.r = 255 - rgba.r
+    rgba.g = 255 - rgba.g
+    rgba.b = 255 - rgba.b
+    rgba.a = 255 - rgba.a
+
+proc applyOpacity1(a: Image, opacity: float32): Image =
   result = newImage(a.width, a.height)
   let op = (255 * opacity).uint32
   for y in 0 ..< a.height:
@@ -24,7 +39,7 @@ proc applyOpacityOriginal(a: Image, opacity: float32): Image =
       rgba.a = ((rgba.a.uint32 * op) div 255).clamp(0, 255).uint8
       result.setRgbaUnsafe(x, y, rgba)
 
-proc sharpOpacityOriginal(a: Image): Image =
+proc sharpOpacity1(a: Image): Image =
   result = newImage(a.width, a.height)
   for y in 0 ..< a.height:
     for x in 0 ..< a.width:
@@ -36,8 +51,13 @@ proc sharpOpacityOriginal(a: Image): Image =
 
 var a = newImage(2560, 1440)
 
-timeIt "fillOriginal":
-  a.fillOriginal(rgba(255, 255, 255, 255))
+timeIt "fill1":
+  a.fill1(rgba(255, 255, 255, 255))
+  doAssert a[0, 0] == rgba(255, 255, 255, 255)
+  keep(a)
+
+timeIt "fill2":
+  a.fill2(rgba(255, 255, 255, 255))
   doAssert a[0, 0] == rgba(255, 255, 255, 255)
   keep(a)
 
@@ -46,24 +66,43 @@ timeIt "fill":
   doAssert a[0, 0] == rgba(255, 255, 255, 255)
   keep(a)
 
-timeIt "invertOriginal":
-  a.invertOriginal()
+timeIt "fill1_rgba":
+  a.fill1(rgba(63, 127, 191, 255))
+  doAssert a[0, 0] == rgba(63, 127, 191, 255)
+  keep(a)
+
+timeIt "fill2_rgba":
+  a.fill2(rgba(63, 127, 191, 255))
+  doAssert a[0, 0] == rgba(63, 127, 191, 255)
+  keep(a)
+
+timeIt "fill_rgba":
+  a.fill(rgba(63, 127, 191, 255))
+  doAssert a[0, 0] == rgba(63, 127, 191, 255)
+  keep(a)
+
+timeIt "invert1":
+  a.invert1()
+  keep(a)
+
+timeIt "invert2":
+  a.invert2()
   keep(a)
 
 timeIt "invert":
   a.invert()
   keep(a)
 
-timeIt "applyOpacityOriginal":
-  a = a.applyOpacityOriginal(0.5)
+timeIt "applyOpacity1":
+  a = a.applyOpacity1(0.5)
   keep(a)
 
 timeIt "applyOpacity":
   a.applyOpacity(0.5)
   keep(a)
 
-timeIt "sharpOpacityOriginal":
-  a = a.sharpOpacityOriginal()
+timeIt "sharpOpacity1":
+  a = a.sharpOpacity1()
   keep(a)
 
 timeIt "sharpOpacity":