all int blends, simd alphafix + softlight, benchmarks
This commit is contained in:
Ryan Oldenburg
parent
8ef4b3ecc8
commit
b895030088
2 changed files with 276 additions and 72 deletions
|
|
@ -111,6 +111,10 @@ proc SetSat(C: Color, s: float32): Color {.inline.} =
|
|||
result = (C - min([C.r, C.g, C.b])) * s / satC
|
||||
|
||||
proc alphaFix(backdrop, source, mixed: Color): Color =
|
||||
result.a = (source.a + backdrop.a * (1.0 - source.a))
|
||||
if result.a == 0:
|
||||
return
|
||||
|
||||
let
|
||||
t0 = source.a * (1 - backdrop.a)
|
||||
t1 = source.a * backdrop.a
|
||||
|
|
@ -120,7 +124,6 @@ proc alphaFix(backdrop, source, mixed: Color): Color =
|
|||
result.g = t0 * source.g + t1 * mixed.g + t2 * backdrop.g
|
||||
result.b = t0 * source.b + t1 * mixed.b + t2 * backdrop.b
|
||||
|
||||
result.a = (source.a + backdrop.a * (1.0 - source.a))
|
||||
result.r /= result.a
|
||||
result.g /= result.a
|
||||
result.b /= result.a
|
||||
|
|
@ -258,25 +261,22 @@ proc blendExcludeMaskFloats(backdrop, source: Color): Color {.inline.} =
|
|||
proc blendOverwriteFloats(backdrop, source: Color): Color {.inline.} =
|
||||
source
|
||||
|
||||
proc alphaFix(backdrop, source, mixed: ColorRGBA): ColorRGBA {.inline.} =
|
||||
proc alphaFix(backdrop, source: ColorRGBA, vb, vs, vm: M128): ColorRGBA =
|
||||
let
|
||||
sa = source.a.float32
|
||||
ba = backdrop.a.float32
|
||||
a = sa + ba * (255 - sa) / 255
|
||||
if a < 1:
|
||||
if a == 0:
|
||||
return
|
||||
|
||||
let
|
||||
vb = mm_setr_ps(backdrop.r.float32, backdrop.g.float32, backdrop.b.float32, 0)
|
||||
vs = mm_setr_ps(source.r.float32, source.g.float32, source.b.float32, 0)
|
||||
vm = mm_setr_ps(mixed.r.float32, mixed.g.float32, mixed.b.float32, 0)
|
||||
vt0 = mm_set1_ps(sa * (255 - ba))
|
||||
vt1 = mm_set1_ps(sa * ba)
|
||||
vt2 = mm_set1_ps((255 - sa) * ba)
|
||||
t0 = mm_set1_ps(sa * (255 - ba))
|
||||
t1 = mm_set1_ps(sa * ba)
|
||||
t2 = mm_set1_ps((255 - sa) * ba)
|
||||
va = mm_set1_ps(a)
|
||||
v255 = mm_set1_ps(255)
|
||||
values = cast[array[4, int32]](
|
||||
mm_cvtps_epi32((vt0 * vs + vt1 * vm + vt2 * vb) / va / v255)
|
||||
values = cast[array[4, uint32]](
|
||||
mm_cvtps_epi32((t0 * vs + t1 * vm + t2 * vb) / va / v255)
|
||||
)
|
||||
|
||||
result.r = values[0].uint8
|
||||
|
|
@ -284,48 +284,140 @@ proc alphaFix(backdrop, source, mixed: ColorRGBA): ColorRGBA {.inline.} =
|
|||
result.b = values[2].uint8
|
||||
result.a = a.uint8
|
||||
|
||||
proc alphaFix(backdrop, source, mixed: ColorRGBA): ColorRGBA {.inline.} =
|
||||
if backdrop.a == 0 and source.a == 0:
|
||||
return
|
||||
let
|
||||
vb = mm_setr_ps(backdrop.r.float32, backdrop.g.float32, backdrop.b.float32, 0)
|
||||
vs = mm_setr_ps(source.r.float32, source.g.float32, source.b.float32, 0)
|
||||
vm = mm_setr_ps(mixed.r.float32, mixed.g.float32, mixed.b.float32, 0)
|
||||
alphaFix(backdrop, source, vb, vs, vm)
|
||||
|
||||
proc min(a, b: uint32): uint32 {.inline.} =
|
||||
if a < b: a else: b
|
||||
|
||||
proc screen(backdrop, source: uint32): uint8 {.inline.} =
|
||||
(255 - ((255 - backdrop) * (255 - source)) div 255).uint8
|
||||
|
||||
proc hardLight(backdrop, source: uint32): uint8 {.inline.} =
|
||||
if source <= 127:
|
||||
((backdrop * 2 * source) div 255).uint8
|
||||
else:
|
||||
screen(backdrop, 2 * source - 255)
|
||||
|
||||
proc blendNormal(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
result = source
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendDarken(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendDarkenFloats(backdrop.color, source.color).rgba
|
||||
result.r = min(backdrop.r, source.r)
|
||||
result.g = min(backdrop.g, source.g)
|
||||
result.b = min(backdrop.b, source.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendMultiply(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendMultiplyFloats(backdrop.color, source.color).rgba
|
||||
result.r = ((backdrop.r.uint32 * source.r) div 255).uint8
|
||||
result.g = ((backdrop.g.uint32 * source.g) div 255).uint8
|
||||
result.b = ((backdrop.b.uint32 * source.b) div 255).uint8
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendLinearBurn(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendLinearBurnFloats(backdrop.color, source.color).rgba
|
||||
result.r = min(0, backdrop.r.int16 + source.r.int16 - 255).uint8
|
||||
result.g = min(0, backdrop.g.int16 + source.g.int16 - 255).uint8
|
||||
result.b = min(0, backdrop.b.int16 + source.b.int16 - 255).uint8
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendColorBurn(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendColorBurnFloats(backdrop.color, source.color).rgba
|
||||
proc blend(backdrop, source: uint32): uint8 {.inline.} =
|
||||
if backdrop == 255:
|
||||
255.uint8
|
||||
elif source == 0:
|
||||
0
|
||||
else:
|
||||
255 - min(255, (255 * (255 - backdrop)) div source).uint8
|
||||
result.r = blend(backdrop.r, source.r)
|
||||
result.g = blend(backdrop.g, source.g)
|
||||
result.b = blend(backdrop.b, source.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendLighten(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendLightenFloats(backdrop.color, source.color).rgba
|
||||
result.r = max(backdrop.r, source.r)
|
||||
result.g = max(backdrop.g, source.g)
|
||||
result.b = max(backdrop.b, source.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendScreen(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendScreenFloats(backdrop.color, source.color).rgba
|
||||
result.r = screen(backdrop.r, source.r)
|
||||
result.g = screen(backdrop.g, source.g)
|
||||
result.b = screen(backdrop.b, source.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendLinearDodge(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendLinearDodgeFloats(backdrop.color, source.color).rgba
|
||||
result.r = min(backdrop.r.uint32 + source.r, 255).uint8
|
||||
result.g = min(backdrop.g.uint32 + source.g, 255).uint8
|
||||
result.b = min(backdrop.b.uint32 + source.b, 255).uint8
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendColorDodge(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendColorDodgeFloats(backdrop.color, source.color).rgba
|
||||
proc blend(backdrop, source: uint32): uint8 {.inline.} =
|
||||
if backdrop == 0:
|
||||
0.uint8
|
||||
elif source == 255:
|
||||
255
|
||||
else:
|
||||
min(255, (255 * backdrop) div (255 - source)).uint8
|
||||
result.r = blend(backdrop.r, source.r)
|
||||
result.g = blend(backdrop.g, source.g)
|
||||
result.b = blend(backdrop.b, source.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendOverlay(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendOverlayFloats(backdrop.color, source.color).rgba
|
||||
result.r = hardLight(source.r, backdrop.r)
|
||||
result.g = hardLight(source.g, backdrop.g)
|
||||
result.b = hardLight(source.b, backdrop.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendHardLight(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendHardLightFloats(backdrop.color, source.color).rgba
|
||||
result.r = hardLight(backdrop.r, source.r)
|
||||
result.g = hardLight(backdrop.g, source.g)
|
||||
result.b = hardLight(backdrop.b, source.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendSoftLight(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendSoftLightFloats(backdrop.color, source.color).rgba
|
||||
# proc softLight(backdrop, source: int32): uint8 {.inline.} =
|
||||
# ## Pegtop
|
||||
# (
|
||||
# ((255 - 2 * source) * backdrop ^ 2) div 255 ^ 2 +
|
||||
# (2 * source * backdrop) div 255
|
||||
# ).uint8
|
||||
|
||||
let
|
||||
vb = mm_setr_ps(backdrop.r.float32, backdrop.g.float32, backdrop.b.float32, 0)
|
||||
vs = mm_setr_ps(source.r.float32, source.g.float32, source.b.float32, 0)
|
||||
v2 = mm_set1_ps(2)
|
||||
v255 = mm_set1_ps(255)
|
||||
v255sq = mm_set1_ps(255 * 255)
|
||||
vm = ((v255 - v2 * vs) * vb * vb) / v255sq + (v2 * vs * vb) / v255
|
||||
values = cast[array[4, uint32]](mm_cvtps_epi32(vm))
|
||||
|
||||
result.r = values[0].uint8
|
||||
result.g = values[1].uint8
|
||||
result.b = values[2].uint8
|
||||
result = alphaFix(backdrop, source, vb, vs, vm)
|
||||
|
||||
proc blendDifference(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendDifferenceFloats(backdrop.color, source.color).rgba
|
||||
result.r = max(backdrop.r, source.r) - min(backdrop.r, source.r)
|
||||
result.g = max(backdrop.g, source.g) - min(backdrop.g, source.g)
|
||||
result.b = max(backdrop.b, source.b) - min(backdrop.b, source.b)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendExclusion(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendExclusionFloats(backdrop.color, source.color).rgba
|
||||
proc blend(backdrop, source: int32): uint8 {.inline.} =
|
||||
max(0, backdrop + source - (2 * backdrop * source) div 255).uint8
|
||||
result.r = blend(backdrop.r.int32, source.r.int32)
|
||||
result.g = blend(backdrop.g.int32, source.g.int32)
|
||||
result.b = blend(backdrop.b.int32, source.b.int32)
|
||||
result = alphaFix(backdrop, source, result)
|
||||
|
||||
proc blendColor(backdrop, source: ColorRGBA): ColorRGBA =
|
||||
blendColorFloats(backdrop.color, source.color).rgba
|
||||
|
|
|
|||
|
|
@ -1,73 +1,185 @@
|
|||
import benchy, chroma, pixie, vmath
|
||||
import benchy, chroma, vmath
|
||||
|
||||
include pixie/blends
|
||||
|
||||
const iterations = 1_000_000
|
||||
|
||||
let
|
||||
a = newImage(1000, 1000)
|
||||
b = newImage(1000, 1000)
|
||||
a = rgba(100, 200, 100, 255)
|
||||
b = rgba(25, 33, 100, 127)
|
||||
|
||||
b.fill(rgba(127, 127, 127, 255))
|
||||
timeIt "blendNormal":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendNormal(a, b)
|
||||
|
||||
timeIt "bmNormal":
|
||||
a.draw(b, vec2(0, 0), bmNormal)
|
||||
timeIt "blendNormalFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendNormalFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmDarken":
|
||||
a.draw(b, vec2(0, 0), bmDarken)
|
||||
timeIt "blendDarken":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendDarken(a, b)
|
||||
|
||||
timeIt "bmMultiply":
|
||||
a.draw(b, vec2(0, 0), bmMultiply)
|
||||
timeIt "blendDarkenFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendDarkenFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmLinearBurn":
|
||||
a.draw(b, vec2(0, 0), bmLinearBurn)
|
||||
timeIt "blendMultiply":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendMultiply(a, b)
|
||||
|
||||
timeIt "bmColorBurn":
|
||||
a.draw(b, vec2(0, 0), bmColorBurn)
|
||||
timeIt "blendMultiplyFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendMultiplyFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmLighten":
|
||||
a.draw(b, vec2(0, 0), bmLighten)
|
||||
timeIt "blendLinearBurn":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLinearBurn(a, b)
|
||||
|
||||
timeIt "bmScreen":
|
||||
a.draw(b, vec2(0, 0), bmScreen)
|
||||
timeIt "blendLinearBurnFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLinearBurnFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmLinearDodge":
|
||||
a.draw(b, vec2(0, 0), bmLinearDodge)
|
||||
timeIt "blendColorBurn":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendColorBurn(a, b)
|
||||
|
||||
timeIt "bmColorDodge":
|
||||
a.draw(b, vec2(0, 0), bmColorDodge)
|
||||
timeIt "blendColorBurnFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendColorBurnFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmOverlay":
|
||||
a.draw(b, vec2(0, 0), bmOverlay)
|
||||
timeIt "blendLighten":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLighten(a, b)
|
||||
|
||||
timeIt "bmSoftLight":
|
||||
a.draw(b, vec2(0, 0), bmSoftLight)
|
||||
timeIt "blendLightenFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLightenFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmHardLight":
|
||||
a.draw(b, vec2(0, 0), bmHardLight)
|
||||
timeIt "blendScreen":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendScreen(a, b)
|
||||
|
||||
timeIt "bmDifference":
|
||||
a.draw(b, vec2(0, 0), bmDifference)
|
||||
timeIt "blendScreenFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendScreenFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmExclusion":
|
||||
a.draw(b, vec2(0, 0), bmExclusion)
|
||||
timeIt "blendLinearDodge":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLinearDodge(a, b)
|
||||
|
||||
timeIt "bmHue":
|
||||
a.draw(b, vec2(0, 0), bmHue)
|
||||
timeIt "blendLinearDodgeFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLinearDodgeFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmSaturation":
|
||||
a.draw(b, vec2(0, 0), bmSaturation)
|
||||
timeIt "blendColorDodge":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendColorDodge(a, b)
|
||||
|
||||
timeIt "bmColor":
|
||||
a.draw(b, vec2(0, 0), bmColor)
|
||||
timeIt "blendColorDodgeFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendColorDodgeFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmLuminosity":
|
||||
a.draw(b, vec2(0, 0), bmLuminosity)
|
||||
timeIt "blendOverlay":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendOverlay(a, b)
|
||||
|
||||
timeIt "bmMask":
|
||||
a.draw(b, vec2(0, 0), bmMask)
|
||||
timeIt "blendOverlayFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendOverlayFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmSubtractMask":
|
||||
a.draw(b, vec2(0, 0), bmSubtractMask)
|
||||
timeIt "blendSoftLight":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendSoftLight(a, b)
|
||||
|
||||
timeIt "bmIntersectMask":
|
||||
a.draw(b, vec2(0, 0), bmIntersectMask)
|
||||
timeIt "blendSoftLightFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendSoftLightFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "bmExcludeMask":
|
||||
a.draw(b, vec2(0, 0), bmExcludeMask)
|
||||
timeIt "blendHardLight":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendHardLight(a, b)
|
||||
|
||||
timeIt "blendHardLightFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendHardLightFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendDifference":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendDifference(a, b)
|
||||
|
||||
timeIt "blendDifferenceFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendDifferenceFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendExclusion":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendExclusion(a, b)
|
||||
|
||||
timeIt "blendExclusionFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendExclusionFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendHue":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendHue(a, b)
|
||||
|
||||
timeIt "blendHueFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendHueFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendSaturation":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendSaturation(a, b)
|
||||
|
||||
timeIt "blendSaturationFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendSaturationFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendColor":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendColor(a, b)
|
||||
|
||||
timeIt "blendColorFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendColorFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendLuminosity":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLuminosity(a, b)
|
||||
|
||||
timeIt "blendLuminosityFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendLuminosityFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendMask":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendMask(a, b)
|
||||
|
||||
timeIt "blendMaskFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendMaskFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendSubtractMask":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendSubtractMask(a, b)
|
||||
|
||||
timeIt "blendSubtractMaskFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendSubtractMaskFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendIntersectMask":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendIntersectMask(a, b)
|
||||
|
||||
timeIt "blendIntersectMaskFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendIntersectMaskFloats(a.color, b.color).rgba
|
||||
|
||||
timeIt "blendExcludeMask":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendExcludeMask(a, b)
|
||||
|
||||
timeIt "blendExcludeMaskFloats":
|
||||
for i in 0 ..< iterations:
|
||||
keep blendExcludeMaskFloats(a.color, b.color).rgba
|
||||
|
|
|
|||
Loading…
Reference in a new issue