From ce7507cd6cd6acc84e6af636492a25dae27da7df Mon Sep 17 00:00:00 2001
From: Ryan Oldenburg <guzba8@gmail.com>
Date: Wed, 6 Jan 2021 14:18:47 -0600
Subject: [PATCH] fast fill, remove benchmark weirdness

---
 src/pixie/images.nim       | 34 +++++++++++----
 tests/benchmark_images.nim | 87 +-------------------------------------
 2 files changed, 26 insertions(+), 95 deletions(-)

diff --git a/src/pixie/images.nim b/src/pixie/images.nim
index ba472a0..af6c8ee 100644
--- a/src/pixie/images.nim
+++ b/src/pixie/images.nim
@@ -1,4 +1,4 @@
-import chroma, blends, bumpy, vmath, common, nimsimd/sse2
+import chroma, blends, bumpy, vmath, common, nimsimd/sse2, system/memory
 
 const h = 0.5.float32
 
@@ -72,14 +72,30 @@ proc `[]=`*(image: Image, x, y: int, rgba: ColorRGBA) {.inline.} =
 
 proc fill*(image: Image, rgba: ColorRgba) =
   ## Fills the image with a solid color.
-  # 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
+
+  # Use memset whene very byte has the same value
+  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:
+    var i: int
+    when defined(amd64):
+      # When supported, SIMD fill until we run out of room.
+      let m = mm_set1_epi32(cast[int32](rgba))
+      while i < image.data.len - 4:
+        mm_store_si128(image.data[i].addr, m)
+        i += 4
+    else:
+      when sizeof(int) == 8:
+        # Fill 8 bytes at a time when possible
+        let
+          u32 = cast[uint32](rgba)
+          u64 = cast[uint64]([u32, u32])
+        while i < image.data.len - 2:
+          cast[ptr uint64](image.data[i].addr)[] = u64
+          i += 2
+    # Fill whatever is left the slow way
+    for j in i ..< image.data.len:
+      image.data[j] = rgba
 
 proc invert*(image: Image) =
   ## Inverts all of the colors and alpha.
diff --git a/tests/benchmark_images.nim b/tests/benchmark_images.nim
index 606b4f3..c9e3d43 100644
--- a/tests/benchmark_images.nim
+++ b/tests/benchmark_images.nim
@@ -1,110 +1,25 @@
 import chroma, pixie, benchy, system/memory
 
-proc fill1(a: Image, rgba: ColorRGBA) =
-  for y in 0 ..< a.height:
-    for x in 0 ..< a.width:
-      a.setRgbaUnsafe(x, y, rgba)
-
-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)
-      rgba.r = 255 - rgba.r
-      rgba.g = 255 - rgba.g
-      rgba.b = 255 - rgba.b
-      rgba.a = 255 - rgba.a
-      a.setRgbaUnsafe(x, y, rgba)
-
-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:
-    for x in 0 ..< a.width:
-      var rgba = a.getRgbaUnsafe(x, y)
-      rgba.a = ((rgba.a.uint32 * op) div 255).clamp(0, 255).uint8
-      result.setRgbaUnsafe(x, y, rgba)
-
-proc sharpOpacity1(a: Image): Image =
-  result = newImage(a.width, a.height)
-  for y in 0 ..< a.height:
-    for x in 0 ..< a.width:
-      var rgba = a.getRgbaUnsafe(x, y)
-      if rgba.a == 0:
-        result.setRgbaUnsafe(x, y, rgba(0, 0, 0, 0))
-      else:
-        result.setRgbaUnsafe(x, y, rgba(255, 255, 255, 255))
-
-var a = newImage(2560, 1440)
-
-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)
+let a = newImage(2560, 1440)
 
 timeIt "fill":
   a.fill(rgba(255, 255, 255, 255))
   doAssert a[0, 0] == rgba(255, 255, 255, 255)
   keep(a)
 
-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 "applyOpacity1":
-  a = a.applyOpacity1(0.5)
-  keep(a)
-
 timeIt "applyOpacity":
   a.applyOpacity(0.5)
   keep(a)
 
-timeIt "sharpOpacity1":
-  a = a.sharpOpacity1()
-  keep(a)
-
 timeIt "sharpOpacity":
   a.sharpOpacity()
   keep(a)