From 82c7c8b8643078b5b48416e665ce7d7b04631724 Mon Sep 17 00:00:00 2001
From: Ryan Oldenburg <ryan@guzba.com>
Date: Wed, 22 Jun 2022 01:03:32 -0500
Subject: [PATCH] faster

---
 src/pixie/internal.nim | 16 ++++++++--------
 src/pixie/paths.nim    | 42 ++++++++++++++++++++----------------------
 2 files changed, 28 insertions(+), 30 deletions(-)

diff --git a/src/pixie/internal.nim b/src/pixie/internal.nim
index d850a36..f9b255d 100644
--- a/src/pixie/internal.nim
+++ b/src/pixie/internal.nim
@@ -52,7 +52,7 @@ proc `*`*(color: ColorRGBX, opacity: float32): ColorRGBX {.raises: [].} =
 
 proc fillUnsafe*(
   data: var seq[uint8], value: uint8, start, len: int
-) {.raises: [].} =
+) {.inline, raises: [].} =
   ## Fills the mask data with the value starting at index start and
   ## continuing for len indices.
   nimSetMem(data[start].addr, value.cint, len)
@@ -62,9 +62,7 @@ proc fillUnsafe*(
 ) {.raises: [].} =
   ## Fills the image data with the color starting at index start and
   ## continuing for len indices.
-
   let rgbx = color.asRgbx()
-
   # Use memset when every byte has the same value
   if rgbx.r == rgbx.g and rgbx.r == rgbx.b and rgbx.r == rgbx.a:
     nimSetMem(data[start].addr, rgbx.r.cint, len * 4)
@@ -78,11 +76,13 @@ proc fillUnsafe*(
       # When supported, SIMD fill until we run out of room
       let
         colorVec = mm_set1_epi32(cast[int32](rgbx))
-        remaining = start + len - i
-      for _ in 0 ..< remaining div 8:
-        mm_store_si128(data[i + 0].addr, colorVec)
-        mm_store_si128(data[i + 4].addr, colorVec)
-        i += 8
+        iterations = (start + len - i) div 8
+      var p = cast[uint](data[i].addr)
+      for _ in 0 ..< iterations:
+        mm_store_si128(cast[pointer](p), colorVec)
+        mm_store_si128(cast[pointer](p + 16), colorVec)
+        p += 32
+      i += iterations * 8
     else:
       when sizeof(int) == 8:
         # Fill 8 bytes at a time when possible
diff --git a/src/pixie/paths.nim b/src/pixie/paths.nim
index 1a2c90e..22afc69 100644
--- a/src/pixie/paths.nim
+++ b/src/pixie/paths.nim
@@ -1133,9 +1133,9 @@ proc partitionSegments(
   result.partitionHeight = height.uint32 div numPartitions
 
   for (segment, winding) in segments:
-    let entry = initPartitionEntry(segment, winding)
+    var entry = initPartitionEntry(segment, winding)
     if result.partitionHeight == 0:
-      result.partitions[0].entries.add(entry)
+      result.partitions[0].entries.add(move entry)
     else:
       var
         atPartition = max(0, segment.at.y - result.startY.float32).uint32
@@ -1619,16 +1619,15 @@ proc fillHits(
   template simdBlob(image: Image, x: var int, len: int, blendProc: untyped) =
     when allowSimd:
       when defined(amd64):
-        let colorVec = mm_set1_epi32(cast[int32](rgbx))
-        var dataIndex = image.dataIndex(x, y)
-        for _ in 0 ..< len div 4:
-          let backdrop = mm_loadu_si128(image.data[dataIndex].addr)
-          mm_storeu_si128(
-            image.data[dataIndex].addr,
-            blendProc(backdrop, colorVec)
-          )
-          x += 4
-          dataIndex += 4
+        var p = cast[uint](image.data[image.dataIndex(x, y)].addr)
+        let
+          iterations = len div 4
+          colorVec = mm_set1_epi32(cast[int32](rgbx))
+        for _ in 0 ..< iterations:
+          let backdrop = mm_loadu_si128(cast[pointer](p))
+          mm_storeu_si128(cast[pointer](p), blendProc(backdrop, colorVec))
+          p += 16
+        x += iterations * 4
 
   case blendMode:
   of OverwriteBlend:
@@ -1714,16 +1713,15 @@ proc fillHits(
   template simdBlob(mask: Mask, x: var int, len: int, blendProc: untyped) =
     when allowSimd:
       when defined(amd64):
-        let vec255 = mm_set1_epi8(255)
-        var dataIndex = mask.dataIndex(x, y)
-        for _ in 0 ..< len div 16:
-          let backdrop = mm_loadu_si128(mask.data[dataIndex].addr)
-          mm_storeu_si128(
-            mask.data[dataIndex].addr,
-            blendProc(backdrop, vec255)
-          )
-          x += 16
-          dataIndex += 16
+        var p = cast[uint](mask.data[mask.dataIndex(x, y)].addr)
+        let
+          iterations = len div 16
+          vec255 = mm_set1_epi8(255)
+        for _ in 0 ..< iterations:
+          let backdrop = mm_loadu_si128(cast[pointer](p))
+          mm_storeu_si128(cast[pointer](p), blendProc(backdrop, vec255))
+          p += 16
+        x += iterations * 16
 
   case blendMode:
   of NormalBlend, OverwriteBlend: