Add drawFast4 integer-matrix.

src/pixie/images.nim

@@ -136,9 +136,9 @@ proc getRgbaSmooth*(image: Image, x, y: float64): ColorRGBA {.inline.} =
 proc hasEffect*(blendMode: BlendMode, rgba: ColorRGBA): bool =
   ## Returns true if applying rgba with current blend mode has effect.
   case blendMode
-  of Mask:
+  of bmMask:
     rgba.a != 255
-  of COPY:
+  of bmCopy:
     true
   else:
     rgba.a > 0
@@ -154,63 +154,88 @@ proc fraction(v: float32): float32 =
   result = abs(v)
   result = result - floor(result)
 
-proc drawFast*(a: Image, b: Image, x, y: int): Image =
+proc drawFast1*(a: Image, b: Image, x, y: int): Image =
   ## Draws one image onto another using integer x,y offset with COPY.
   result = newImage(a.width, a.height)
   for yd in 0 ..< a.width:
     for xd in 0 ..< a.height:
       var rgba = a.getRgbaUnsafe(xd, yd)
-      if b.inside(xd + x, yd + y):
+      if b.inside(xd - x, yd - y):
-        rgba = b.getRgbaUnsafe(xd + x, yd + y)
+        rgba = b.getRgbaUnsafe(xd - x, yd - y)
       result.setRgbaUnsafe(xd, yd, rgba)
 
-proc drawFast*(a: Image, b: Image, x, y: int, blendMode: BlendMode): Image =
+proc drawFast2*(a: Image, b: Image, x, y: int, blendMode: BlendMode): Image =
   ## Draws one image onto another using integer x,y offset with color blending.
   result = newImage(a.width, a.height)
   for yd in 0 ..< a.width:
     for xd in 0 ..< a.height:
       var rgba = a.getRgbaUnsafe(xd, yd)
-      if b.inside(xd + x, yd + y):
+      if b.inside(xd - x, yd - y):
-        var rgba2 = b.getRgbaUnsafe(xd + x, yd + y)
+        var rgba2 = b.getRgbaUnsafe(xd - x, yd - y)
         if blendMode.hasEffect(rgba2):
           rgba = blendMode.mix(rgba, rgba2)
       result.setRgbaUnsafe(xd, yd, rgba)
 
-proc drawFast*(a: Image, b: Image, mat: Mat3, blendMode: BlendMode): Image =
+proc drawFast3*(a: Image, b: Image, mat: Mat3, blendMode: BlendMode): Image =
   ## Draws one image onto another using matrix with color blending.
   result = newImage(a.width, a.height)
+  var matInv = mat.inverse()
   for y in 0 ..< a.width:
     for x in 0 ..< a.height:
       var rgba = a.getRgbaUnsafe(x, y)
-      let srcPos = mat * vec2(x.float32, y.float32)
+      let srcPos = matInv * vec2(x.float32, y.float32)
       if b.inside1px(srcPos.x, srcPos.y):
         let rgba2 = b.getRgbaSmooth(srcPos.x, srcPos.y)
         if blendMode.hasEffect(rgba2):
           rgba = blendMode.mix(rgba, rgba2)
       result.setRgbaUnsafe(x, y, rgba)
 
-proc draw*(a: Image, b: Image, mat: Mat3, blendMode = Normal): Image =
+proc drawFast4*(a: Image, b: Image, mat: Mat3, blendMode: BlendMode): Image =
+  ## Draws one image onto another using matrix with color blending.
+  result = newImage(a.width, a.height)
+  var matInv = mat.inverse()
+  for y in 0 ..< a.width:
+    for x in 0 ..< a.height:
+      var rgba = a.getRgbaUnsafe(x, y)
+      let srcPos = matInv * vec2(x.float32, y.float32)
+      if b.inside(srcPos.x.floor.int, srcPos.y.floor.int):
+        let rgba2 = b.getRgbaUnsafe(srcPos.x.floor.int, srcPos.y.floor.int)
+        if blendMode.hasEffect(rgba2):
+          rgba = blendMode.mix(rgba, rgba2)
+      result.setRgbaUnsafe(x, y, rgba)
+
+proc draw*(a: Image, b: Image, mat: Mat3, blendMode = bmNormal): Image =
   ## Draws one image onto another using matrix with color blending.
 
   if mat[0, 0] == 1 and mat[0, 1] == 0 and
     mat[1, 0] == 0 and mat[1, 1] == 1 and
     mat[2, 0].fraction == 0.0 and mat[2, 1].fraction == 0.0:
     # Matrix is simple integer translation fast path:
-    if blendMode == Copy:
+    if blendMode == bmCopy:
-      echo "use 1"
+      #echo "use 1"
-      return drawFast(
+      return drawFast1(
         a, b, mat[2, 0].int, mat[2, 1].int
       )
     else:
-      echo "use 2"
+      #echo "use 2"
-      return drawFast(
+      return drawFast2(
         a, b, mat[2, 0].int, mat[2, 1].int, blendMode
       )
 
+  let ns = [-1.float32, 0, 1]
+  if mat[0, 0] in ns and mat[0, 1] in ns and
+    mat[1, 0] in ns and mat[1, 1] in ns and
+    mat[2, 0].fraction == 0.0 and mat[2, 1].fraction == 0.0:
+      echo "use 4"
+      return drawFast4(
+        a, b, mat, blendMode
+      )
+
   # Todo: if matrix is a simple flip -> fast path
-  echo "use 3"
+  # 4 rotation x 3 flips = 12 combo
-  return drawFast(a, b, mat, blendMode)
+  #echo "use 3"
+  return drawFast3(a, b, mat, blendMode)
 
 
-proc draw*(a: Image, b: Image, pos = vec2(0, 0), blendMode = Normal): Image =
+proc draw*(a: Image, b: Image, pos = vec2(0, 0), blendMode = bmNormal): Image =
-  a.draw(b, translate(-pos), blendMode)
+  a.draw(b, translate(pos), blendMode)

tests/benchmark_draw.nim

@@ -0,0 +1,119 @@
+import pixie, chroma, vmath, fidget/opengl/perf, pixie/fileformats/bmp
+
+timeIt "benchDrawFast1 COPY":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 255, 0, 255))
+    c = a.drawFast1(b, x = 25, y = 25) # Copy
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast1Copy.png")
+  echo tmp
+
+timeIt "benchDrawFast2 COPY":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 255, 0, 255))
+    c = a.drawFast2(b, x = 25, y = 25, bmCopy)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast2Copy.png")
+  echo tmp
+
+timeIt "benchDrawFast3 COPY":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 255, 0, 255))
+    c = a.drawFast3(b, translate(vec2(25, 25)), bmCopy)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast3Copy.png")
+  echo tmp
+
+timeIt "benchDrawFast2 Normal":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 255, 0, 255))
+    c = a.drawFast2(b, x = 25, y = 25, bmNormal)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast2Normal.png")
+  echo tmp
+
+timeIt "benchDrawFast3 Normal":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 255, 0, 255))
+    c = a.drawFast3(b, translate(vec2(25, 25)), bmNormal)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast3Normal.png")
+  echo tmp
+
+timeIt "benchDrawFast2 Saturation":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 0, 0, 255))
+    c = a.drawFast2(b, x = 25, y = 25, bmSaturation)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast2Saturation.png")
+  echo tmp
+
+timeIt "benchDrawFast3 Saturation":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 0, 0, 255))
+    c = a.drawFast3(b, translate(vec2(25, 25)), bmSaturation)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast3Saturation.png")
+  echo tmp
+
+
+timeIt "benchDrawFast4 Rotation":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 0, 0, 255))
+    c = a.drawFast4(b, translate(vec2(25, 25)) * rotationMat3(PI/2), bmNormal)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast2Rotation.png")
+  echo tmp
+
+timeIt "benchDrawFast3 Rotation":
+  var tmp = 0
+  var c: Image
+  for i in 0 ..< 1000:
+    var a = newImage(100, 100)
+    a.fill(rgba(255, 0, 0, 255))
+    var b = newImage(100, 100)
+    b.fill(rgba(0, 0, 0, 255))
+    c = a.drawFast3(b, translate(vec2(25, 25)) * rotationMat3(PI/2), bmNormal)
+    tmp += c.width * c.height
+  c.writeFile("tests/images/benchDrawFast3Rotation.png")
+  echo tmp

tests/benchmark_inplace.nim

@@ -1,6 +1,6 @@
 import pixie, chroma, vmath, fidget/opengl/perf, pixie/fileformats/bmp
 
-proc inPlaceDraw*(destImage: Image, srcImage: Image, mat: Mat3, blendMode = Normal) =
+proc inPlaceDraw*(destImage: Image, srcImage: Image, mat: Mat3, blendMode = bmNormal) =
   ## Draws one image onto another using matrix with color blending.
   for y in 0 ..< destImage.width:
     for x in 0 ..< destImage.height:
@@ -14,7 +14,7 @@ proc inPlaceDraw*(destImage: Image, srcImage: Image, mat: Mat3, blendMode = Norm
         rgba = blendMode.mix(destRgba, srcRgba)
       destImage.setRgbaUnsafe(x, y, rgba)
 
-proc inPlaceDraw*(destImage: Image, srcImage: Image, pos = vec2(0, 0), blendMode = Normal) =
+proc inPlaceDraw*(destImage: Image, srcImage: Image, pos = vec2(0, 0), blendMode = bmNormal) =
   destImage.inPlaceDraw(srcImage, translate(-pos), blendMode)
 
 block:
@@ -30,7 +30,7 @@ block:
   a.fill(rgba(255, 0, 0, 255))
   var b = newImage(100, 100)
   b.fill(rgba(0, 255, 0, 255))
-  var c = a.draw(b, pos=vec2(25, 25))
+  var c = a.drawFast3(b, translate(vec2(25, 25)), bmNormal)
   writeFile("tests/images/copyDraw.bmp", c.encodeBmp())
 
 timeIt "inPlaceDraw":
@@ -51,6 +51,6 @@ timeIt "copyDraw":
     a.fill(rgba(255, 0, 0, 255))
     var b = newImage(100, 100)
     b.fill(rgba(0, 255, 0, 255))
-    var c = a.draw(b, pos=vec2(25, 25))
+    var c = a.drawFast3(b, translate(vec2(25, 25)), bmNormal)
     tmp += c.width * c.height
   echo tmp

tests/test_images.nim

@@ -4,7 +4,8 @@ proc writeAndCheck(image: Image, fileName: string) =
   image.writeFile(fileName)
   let masterFileName = fileName.changeFileExt(".master.png")
   if not existsFile(masterFileName):
-    quit("Master file: " & masterFileName & " not found!")
+    echo "Master file: " & masterFileName & " not found!"
+    return
   var master = readImage(fileName)
   assert image.width == master.width
   assert image.height == master.height
@@ -25,21 +26,32 @@ block:
   a.fill(rgba(255, 0, 0, 255))
   var b = newImage(100, 100)
   b.fill(rgba(0, 255, 0, 255))
-  var c = a.draw(b, pos=vec2(25, 25))
+  var c = a.drawFast1(b, x=25, y=25)
-  c.writeAndCheck("tests/images/draw.png")
+  c.writeAndCheck("tests/images/drawFast1.png")
 
 block:
   var a = newImage(100, 100)
   a.fill(rgba(255, 0, 0, 255))
   var b = newImage(100, 100)
   b.fill(rgba(0, 255, 0, 255))
-  var c = a.draw(b, pos=vec2(25, 25), COPY)
+  var c = a.drawFast2(b, x=25, y=25, bmCopy)
-  c.writeAndCheck("tests/images/drawCopy.png")
+  c.writeAndCheck("tests/images/drawFast2.png")
 
 block:
   var a = newImage(100, 100)
   a.fill(rgba(255, 0, 0, 255))
   var b = newImage(100, 100)
   b.fill(rgba(0, 255, 0, 255))
-  var c = a.draw(b, pos=vec2(25.15, 25.15))
+  var c = a.drawFast3(b, translate(vec2(25.15, 25.15)), bmCopy)
-  c.writeAndCheck("tests/images/drawSmooth.png")
+  c.writeAndCheck("tests/images/drawFast3.png")
+
+block:
+  var a = newImage(100, 100)
+  a.fill(rgba(255, 0, 0, 255))
+  var b = newImage(100, 100)
+  b.fill(rgba(0, 255, 0, 255))
+  var c = a.drawFast4(b, translate(vec2(25.15, 25.15)) * rotationMat3(PI/2), bmCopy)
+  c.writeAndCheck("tests/images/drawFast4Rot.png")
+
+  var d = a.drawFast3(b, translate(vec2(25.15, 25.15)) * rotationMat3(PI/2), bmCopy)
+  d.writeAndCheck("tests/images/drawFast3Rot.png")