switch to split draw, either pixel-aligned fast or smooth slower

experiments/bench_cairo_draw.nim

@@ -1,39 +1,4 @@
-import benchy, cairo, pixie, pixie/blends, pixie/internal
-
-when defined(amd64) and not defined(pixieNoSimd):
-  import nimsimd/sse2
-
-when defined(release):
-  {.push checks: off.}
-
-proc drawBasic(backdrop, source: Image) =
-  for y in 0 ..< min(backdrop.height, source.height):
-    if isOpaque(source.data, source.dataIndex(0, y), source.width):
-      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.}
+import benchy, cairo, pixie
 
 block:
   let
@@ -43,12 +8,6 @@ block:
     ctx = tmp.create()
 
   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)
@@ -64,7 +23,6 @@ block:
     tmp = newImage(1568, 940)
 
   timeIt "pixie draw normal":
-    # tmp.fill(rgbx(127, 127, 127, 255))
     tmp.draw(backdrop)
     tmp.draw(source)
 
@@ -77,25 +35,11 @@ block:
     tmp = newImage(1568, 940)
 
   timeIt "pixie draw overwrite":
-    # tmp.fill(rgbx(127, 127, 127, 255))
     tmp.draw(backdrop, blendMode = OverwriteBlend)
     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")
@@ -104,17 +48,11 @@ block:
     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")
+  # tmp.writeToPng("tmp_masked.png")
 
 block:
   let

src/pixie/blends.nim

@@ -1,6 +1,6 @@
 ## Blending modes.
 
-import chroma, common, simd, std/math
+import chroma, common, std/math
 
 # See https://www.w3.org/TR/compositing-1/
 # See https://www.khronos.org/registry/OpenGL/extensions/KHR/KHR_blend_equation_advanced.txt
@@ -412,82 +412,5 @@ proc blender*(blendMode: BlendMode): Blender {.raises: [].} =
   of SubtractMaskBlend: subtractMaskBlender
   of ExcludeMaskBlend: excludeMaskBlender
 
-when defined(amd64) and allowSimd:
-  type
-    BlenderSimd* = proc(blackdrop, source: M128i): M128i {.gcsafe, raises: [].}
-      ## Function signature returned by blenderSimd.
-
-  proc blendNormalSimd*(backdrop, source: M128i): M128i {.inline.} =
-    let
-      alphaMask = mm_set1_epi32(cast[int32](0xff000000))
-      oddMask = mm_set1_epi16(cast[int16](0xff00))
-      div255 = mm_set1_epi16(cast[int16](0x8081))
-
-    var
-      sourceAlpha = mm_and_si128(source, alphaMask)
-      backdropEven = mm_slli_epi16(backdrop, 8)
-      backdropOdd = mm_and_si128(backdrop, oddMask)
-
-    sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
-
-    let k = mm_sub_epi32(
-      mm_set1_epi32(cast[int32]([0.uint8, 255, 0, 255])),
-      sourceAlpha
-    )
-
-    backdropEven = mm_mulhi_epu16(backdropEven, k)
-    backdropOdd = mm_mulhi_epu16(backdropOdd, k)
-
-    backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
-    backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
-
-    mm_add_epi8(
-      source,
-      mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
-    )
-
-  proc blendMaskSimd*(backdrop, source: M128i): M128i {.inline.} =
-    let
-      alphaMask = mm_set1_epi32(cast[int32](0xff000000))
-      oddMask = mm_set1_epi16(cast[int16](0xff00))
-      div255 = mm_set1_epi16(cast[int16](0x8081))
-
-    var
-      sourceAlpha = mm_and_si128(source, alphaMask)
-      backdropEven = mm_slli_epi16(backdrop, 8)
-      backdropOdd = mm_and_si128(backdrop, oddMask)
-
-    sourceAlpha = mm_or_si128(sourceAlpha, mm_srli_epi32(sourceAlpha, 16))
-
-    backdropEven = mm_mulhi_epu16(backdropEven, sourceAlpha)
-    backdropOdd = mm_mulhi_epu16(backdropOdd, sourceAlpha)
-
-    backdropEven = mm_srli_epi16(mm_mulhi_epu16(backdropEven, div255), 7)
-    backdropOdd = mm_srli_epi16(mm_mulhi_epu16(backdropOdd, div255), 7)
-
-    mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
-
-  proc normalSimdBlender(backdrop, source: M128i): M128i =
-    blendNormalSimd(backdrop, source)
-
-  proc maskSimdBlender(backdrop, source: M128i): M128i =
-    blendMaskSimd(backdrop, source)
-
-  proc overwriteSimdBlender(backdrop, source: M128i): M128i =
-    source
-
-  proc blenderSimd*(blendMode: BlendMode): BlenderSimd {.raises: [PixieError].} =
-    ## Returns a blend function for a given blend mode with SIMD support.
-    case blendMode:
-    of NormalBlend: normalSimdBlender
-    of MaskBlend: maskSimdBlender
-    of OverwriteBlend: overwriteSimdBlender
-    else:
-      raise newException(PixieError, "No SIMD blender for " & $blendMode)
-
-  proc hasSimdBlender*(blendMode: BlendMode): bool {.inline, raises: [].} =
-    ## Is there a blend function for a given blend mode with SIMD support?
-    blendMode in {NormalBlend, MaskBlend, OverwriteBlend}
-
 when defined(release):
   {.pop.}