isOpaque, faster nonsimd draw, drawBasic

experiments/benchmark_cairo_draw.nim

@@ -1,4 +1,41 @@
-import benchy, cairo, pixie
+import benchy, cairo, pixie, pixie/blends
+
+when defined(amd64) and not defined(pixieNoSimd):
+  import nimsimd/sse2
+
+when defined(release):
+  {.push checks: off.}
+
+proc drawBasic(backdrop, source: Image) =
+  let sourceIsOpaque = source.isOpaque()
+
+  for y in 0 ..< min(backdrop.height, source.height):
+    if sourceIsOpaque:
+      copyMem(
+        backdrop.data[backdrop.dataIndex(0, y)].addr,
+        source.data[source.dataIndex(0, y)].addr,
+        min(backdrop.width, source.width) * 4
+      )
+    else:
+      var x: int
+      when defined(amd64) and not defined(pixieNoSimd):
+        let vec255 = mm_set1_epi32(cast[int32](uint32.high))
+        for _ in 0 ..< min(backdrop.width, source.width) div 4:
+          let sourceVec = mm_loadu_si128(source.data[source.dataIndex(x, y)].addr)
+          if mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, mm_setzero_si128())) != 0xffff:
+            if (mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, vec255)) and 0x8888) == 0x8888:
+              mm_storeu_si128(backdrop.data[backdrop.dataIndex(x, y)].addr, sourceVec)
+            else:
+              let backdropVec = mm_loadu_si128(backdrop.data[backdrop.dataIndex(x, y)].addr)
+              mm_storeu_si128(
+                backdrop.data[backdrop.dataIndex(x, y)].addr,
+                blendNormalInlineSimd(backdropVec, sourceVec)
+              )
+          x += 4
+      # No scalar for now
+
+when defined(release):
+  {.pop.}
 
 block:
   let
@@ -7,7 +44,13 @@ block:
     tmp = imageSurfaceCreate(FORMAT_ARGB32, 1568, 940)
     ctx = tmp.create()
 
-  timeIt "cairo draw basic":
+  timeIt "cairo draw normal":
+    # ctx.setSourceRgba(0.5, 0.5, 0.5, 1)
+    # let operator = ctx.getOperator()
+    # ctx.setOperator(OperatorSource)
+    # ctx.paint()
+    # ctx.setOperator(operator)
+
     ctx.setSource(backdrop, 0, 0)
     ctx.paint()
     ctx.setSource(source, 0, 0)
@@ -22,15 +65,71 @@ block:
     source = readImage("tests/fileformats/svg/masters/Ghostscript_Tiger.png")
     tmp = newImage(1568, 940)
 
-  timeIt "isOneColor":
-    doAssert not backdrop.isOneColor()
-
-  timeIt "pixie draw basic":
+  timeIt "pixie draw normal":
+    # tmp.fill(rgbx(127, 127, 127, 255))
     tmp.draw(backdrop)
     tmp.draw(source)
 
   # tmp.writeFile("tmp2.png")
 
+block:
+  let
+    backdrop = readImage("tests/fileformats/svg/masters/dragon2.png")
+    source = readImage("tests/fileformats/svg/masters/Ghostscript_Tiger.png")
+    tmp = newImage(1568, 940)
+
+  timeIt "pixie draw overwrite":
+    # tmp.fill(rgbx(127, 127, 127, 255))
+    tmp.draw(backdrop, blendMode = bmOverwrite)
+    tmp.draw(source)
+
+  # tmp.writeFile("tmp2.png")
+
+block:
+  let
+    backdrop = readImage("tests/fileformats/svg/masters/dragon2.png")
+    source = readImage("tests/fileformats/svg/masters/Ghostscript_Tiger.png")
+    tmp = newImage(1568, 940)
+
+  timeIt "pixie draw basic":
+    # tmp.fill(rgbx(127, 127, 127, 255))
+    tmp.drawBasic(backdrop)
+    tmp.drawBasic(source)
+
+  # tmp.writeFile("tmp2.png")
+
+block:
+  let
+    backdrop = imageSurfaceCreateFromPng("tests/fileformats/svg/masters/dragon2.png")
+    source = imageSurfaceCreateFromPng("tests/fileformats/svg/masters/Ghostscript_Tiger.png")
+    tmp = imageSurfaceCreate(FORMAT_ARGB32, 1568, 940)
+    ctx = tmp.create()
+
+  timeIt "cairo draw mask":
+    ctx.setSourceRgba(1, 1, 1, 1)
+    let operator = ctx.getOperator()
+    ctx.setOperator(OperatorSource)
+    ctx.paint()
+    ctx.setOperator(operator)
+
+    ctx.setSource(backdrop, 0, 0)
+    ctx.mask(source, 0, 0)
+    tmp.flush()
+
+  # echo tmp.writeToPng("tmp_masked.png")
+
+block:
+  let
+    backdrop = readImage("tests/fileformats/svg/masters/dragon2.png")
+    source = readImage("tests/fileformats/svg/masters/Ghostscript_Tiger.png")
+    tmp = newImage(1568, 940)
+
+  timeIt "pixie draw mask":
+    tmp.draw(backdrop)
+    tmp.draw(source, blendMode = bmMask)
+
+  # tmp.writeFile("tmp_masked2.png")
+
 block:
   let
     backdrop = imageSurfaceCreateFromPng("tests/fileformats/svg/masters/dragon2.png")

src/pixie.nim

@@ -1,10 +1,9 @@
-import bumpy, chroma, flatty/binny, os, pixie/blends, pixie/common,
-    pixie/contexts, pixie/fileformats/bmp, pixie/fileformats/gif,
-    pixie/fileformats/jpg, pixie/fileformats/png, pixie/fileformats/svg,
-    pixie/fonts, pixie/images, pixie/masks, pixie/paints, pixie/paths, strutils, vmath
+import bumpy, chroma, flatty/binny, os, pixie/common, pixie/contexts,
+    pixie/fileformats/bmp, pixie/fileformats/gif, pixie/fileformats/jpg,
+    pixie/fileformats/png, pixie/fileformats/svg, pixie/fonts, pixie/images,
+    pixie/masks, pixie/paints, pixie/paths, strutils, vmath
 
-export blends, bumpy, chroma, common, contexts, fonts, images, masks, paints,
-    paths, vmath
+export bumpy, chroma, common, contexts, fonts, images, masks, paints, paths, vmath
 
 type
   FileFormat* = enum

src/pixie/blends.nim

@@ -9,31 +9,6 @@ when defined(amd64) and not defined(pixieNoSimd):
 # See https://www.khronos.org/registry/OpenGL/extensions/KHR/KHR_blend_equation_advanced.txt
 
 type
-  BlendMode* = enum
-    bmNormal
-    bmDarken
-    bmMultiply
-    # bmLinearBurn
-    bmColorBurn
-    bmLighten
-    bmScreen
-    # bmLinearDodge
-    bmColorDodge
-    bmOverlay
-    bmSoftLight
-    bmHardLight
-    bmDifference
-    bmExclusion
-    bmHue
-    bmSaturation
-    bmColor
-    bmLuminosity
-
-    bmMask         ## Special blend mode that is used for masking
-    bmOverwrite    ## Special blend mode that just copies pixels
-    bmSubtractMask ## Inverse mask
-    bmExcludeMask
-
   Blender* = proc(backdrop, source: ColorRGBX): ColorRGBX {.gcsafe, raises: [].}
     ## Function signature returned by blender.
   Masker* = proc(backdrop, source: uint8): uint8 {.gcsafe, raises: [].}
@@ -165,7 +140,7 @@ proc SetSat(C: Color, s: float32): Color {.inline.} =
   if satC > 0:
     result = (C - min([C.r, C.g, C.b])) * s / satC
 
-proc blendNormal(backdrop, source: ColorRGBX): ColorRGBX =
+proc blendNormal*(backdrop, source: ColorRGBX): ColorRGBX =
   if backdrop.a == 0 or source.a == 255:
     return source
   if source.a == 0:
@@ -419,7 +394,7 @@ proc blendSaturation(backdrop, source: ColorRGBX): ColorRGBX =
     blended = SetLum(SetSat(backdrop, Sat(source)), Lum(backdrop))
   result = alphaFix(backdrop, source, blended).rgba.rgbx()
 
-proc blendMask(backdrop, source: ColorRGBX): ColorRGBX =
+proc blendMask*(backdrop, source: ColorRGBX): ColorRGBX =
   let k = source.a.uint32
   result.r = ((backdrop.r * k) div 255).uint8
   result.g = ((backdrop.g * k) div 255).uint8
@@ -477,10 +452,13 @@ proc maskNormal(backdrop, source: uint8): uint8 =
   ## Blending masks
   blendAlpha(backdrop, source)
 
-proc maskMask(backdrop, source: uint8): uint8 =
+proc maskMaskInline*(backdrop, source: uint8): uint8 {.inline.} =
   ## Masking masks
   ((backdrop.uint32 * source) div 255).uint8
 
+proc maskMask(backdrop, source: uint8): uint8 =
+  maskMaskInline(backdrop, source)
+
 proc maskSubtract(backdrop, source: uint8): uint8 =
   ((backdrop.uint32 * (255 - source)) div 255).uint8
 

src/pixie/common.nim

@@ -3,6 +3,31 @@ import bumpy, chroma, vmath
 type
   PixieError* = object of ValueError ## Raised if an operation fails.
 
+  BlendMode* = enum
+    bmNormal
+    bmDarken
+    bmMultiply
+    # bmLinearBurn
+    bmColorBurn
+    bmLighten
+    bmScreen
+    # bmLinearDodge
+    bmColorDodge
+    bmOverlay
+    bmSoftLight
+    bmHardLight
+    bmDifference
+    bmExclusion
+    bmHue
+    bmSaturation
+    bmColor
+    bmLuminosity
+
+    bmMask         ## Special blend mode that is used for masking
+    bmOverwrite    ## Special blend mode that just copies pixels
+    bmSubtractMask ## Inverse mask
+    bmExcludeMask
+
 proc mix*(a, b: uint8, t: float32): uint8 {.inline, raises: [].} =
   ## Linearly interpolate between a and b using t.
   let t = round(t * 255).uint32

src/pixie/contexts.nim

@@ -1,5 +1,5 @@
-import bumpy, chroma, pixie/blends, pixie/common, pixie/fonts, pixie/images,
-    pixie/masks, pixie/paints, pixie/paths, tables, vmath
+import bumpy, chroma, pixie/common, pixie/fonts, pixie/images, pixie/masks,
+    pixie/paints, pixie/paths, tables, vmath
 
 ## This file provides a Nim version of the Canvas 2D API commonly used on the
 ## web. The goal is to make picking up Pixie easy for developers familiar with