Merge pull request #454 from guzba/master
simd changes started as discussed
This commit is contained in:
Andre von Houck
GitHub
commit
fc3b834e62
7 changed files with 508 additions and 384 deletions
|
|
@ -1,7 +1,10 @@
|
|||
import blends, bumpy, chroma, common, masks, pixie/internal, vmath
|
||||
import blends, bumpy, chroma, common, internal, masks, vmath
|
||||
|
||||
when defined(amd64) and allowSimd:
|
||||
import nimsimd/sse2, runtimechecked/avx2
|
||||
when allowSimd:
|
||||
import simd
|
||||
|
||||
when defined(amd64):
|
||||
import nimsimd/sse2
|
||||
|
||||
const h = 0.5.float32
|
||||
|
||||
|
|
@ -28,21 +31,18 @@ proc newImage*(width, height: int): Image {.raises: [PixieError].} =
|
|||
|
||||
proc newImage*(mask: Mask): Image {.raises: [PixieError].} =
|
||||
result = newImage(mask.width, mask.height)
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
for _ in 0 ..< mask.data.len div 16:
|
||||
var alphas = mm_loadu_si128(mask.data[i].addr)
|
||||
for j in 0 ..< 4:
|
||||
var unpacked = unpackAlphaValues(alphas)
|
||||
unpacked = mm_or_si128(unpacked, mm_srli_epi32(unpacked, 8))
|
||||
unpacked = mm_or_si128(unpacked, mm_srli_epi32(unpacked, 16))
|
||||
mm_storeu_si128(result.data[i + j * 4].addr, unpacked)
|
||||
alphas = mm_srli_si128(alphas, 4)
|
||||
i += 16
|
||||
|
||||
for j in i ..< mask.data.len:
|
||||
let v = mask.data[j]
|
||||
result.data[j] = rgbx(v, v, v, v)
|
||||
when allowSimd and compiles(newImageFromMaskSimd):
|
||||
newImageFromMaskSimd(
|
||||
cast[ptr UncheckedArray[ColorRGBX]](result.data[0].addr),
|
||||
cast[ptr UncheckedArray[uint8]](mask.data[0].addr),
|
||||
mask.data.len
|
||||
)
|
||||
return
|
||||
|
||||
for i in 0 ..< mask.data.len:
|
||||
let v = mask.data[i]
|
||||
result.data[i] = rgbx(v, v, v, v)
|
||||
|
||||
proc copy*(image: Image): Image {.raises: [PixieError].} =
|
||||
## Copies the image data into a new image.
|
||||
|
|
@ -101,83 +101,30 @@ proc fill*(image: Image, color: SomeColor) {.inline, raises: [].} =
|
|||
|
||||
proc isOneColor*(image: Image): bool {.raises: [].} =
|
||||
## Checks if the entire image is the same color.
|
||||
when defined(amd64) and allowSimd:
|
||||
if cpuHasAvx2:
|
||||
return isOneColorAvx2(image.data, 0, image.data.len)
|
||||
when allowSimd and compiles(isOneColorSimd):
|
||||
return isOneColorSimd(
|
||||
cast[ptr UncheckedArray[ColorRGBX]](image.data[0].addr),
|
||||
image.data.len
|
||||
)
|
||||
|
||||
result = true
|
||||
|
||||
let color = image.data[0]
|
||||
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
# Align to 16 bytes
|
||||
var p = cast[uint](image.data[i].addr)
|
||||
while i < image.data.len and (p and 15) != 0:
|
||||
if image.data[i] != color:
|
||||
return false
|
||||
inc i
|
||||
p += 4
|
||||
|
||||
let
|
||||
colorVec = mm_set1_epi32(cast[int32](color))
|
||||
iterations = (image.data.len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm_load_si128(cast[pointer](p))
|
||||
values1 = mm_load_si128(cast[pointer](p + 16))
|
||||
values2 = mm_load_si128(cast[pointer](p + 32))
|
||||
values3 = mm_load_si128(cast[pointer](p + 48))
|
||||
eq0 = mm_cmpeq_epi8(values0, colorVec)
|
||||
eq1 = mm_cmpeq_epi8(values1, colorVec)
|
||||
eq2 = mm_cmpeq_epi8(values2, colorVec)
|
||||
eq3 = mm_cmpeq_epi8(values3, colorVec)
|
||||
eq0123 = mm_and_si128(mm_and_si128(eq0, eq1), mm_and_si128(eq2, eq3))
|
||||
if mm_movemask_epi8(eq0123) != 0xffff:
|
||||
return false
|
||||
p += 64
|
||||
i += 16 * iterations
|
||||
|
||||
for i in i ..< image.data.len:
|
||||
if image.data[i] != color:
|
||||
let color = cast[uint32](image.data[0])
|
||||
for i in 0 ..< image.data.len:
|
||||
if cast[uint32](image.data[i]) != color:
|
||||
return false
|
||||
|
||||
proc isTransparent*(image: Image): bool {.raises: [].} =
|
||||
## Checks if this image is fully transparent or not.
|
||||
when defined(amd64) and allowSimd:
|
||||
if cpuHasAvx2:
|
||||
return isTransparentAvx2(image.data, 0, image.data.len)
|
||||
when allowSimd and compiles(isTransparentSimd):
|
||||
return isTransparentSimd(
|
||||
cast[ptr UncheckedArray[ColorRGBX]](image.data[0].addr),
|
||||
image.data.len
|
||||
)
|
||||
|
||||
result = true
|
||||
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
# Align to 16 bytes
|
||||
var p = cast[uint](image.data[i].addr)
|
||||
while i < image.data.len and (p and 15) != 0:
|
||||
if image.data[i].a != 0:
|
||||
return false
|
||||
inc i
|
||||
p += 4
|
||||
|
||||
let
|
||||
vecZero = mm_setzero_si128()
|
||||
iterations = (image.data.len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm_load_si128(cast[pointer](p))
|
||||
values1 = mm_load_si128(cast[pointer](p + 16))
|
||||
values2 = mm_load_si128(cast[pointer](p + 32))
|
||||
values3 = mm_load_si128(cast[pointer](p + 48))
|
||||
values01 = mm_or_si128(values0, values1)
|
||||
values23 = mm_or_si128(values2, values3)
|
||||
values0123 = mm_or_si128(values01, values23)
|
||||
if mm_movemask_epi8(mm_cmpeq_epi8(values0123, vecZero)) != 0xffff:
|
||||
return false
|
||||
p += 64
|
||||
i += 16 * iterations
|
||||
|
||||
for i in i ..< image.data.len:
|
||||
for i in 0 ..< image.data.len:
|
||||
if image.data[i].a != 0:
|
||||
return false
|
||||
|
||||
|
|
@ -410,89 +357,48 @@ proc magnifyBy2*(image: Image, power = 1): Image {.raises: [PixieError].} =
|
|||
result.width * 4
|
||||
)
|
||||
|
||||
proc applyOpacity*(target: Image | Mask, opacity: float32) {.raises: [].} =
|
||||
proc applyOpacity*(image: Image, opacity: float32) {.raises: [].} =
|
||||
## Multiplies alpha of the image by opacity.
|
||||
let opacity = round(255 * opacity).uint16
|
||||
if opacity == 255:
|
||||
return
|
||||
|
||||
if opacity == 0:
|
||||
when type(target) is Image:
|
||||
target.fill(rgbx(0, 0, 0, 0))
|
||||
else:
|
||||
target.fill(0)
|
||||
image.fill(rgbx(0, 0, 0, 0))
|
||||
return
|
||||
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
when type(target) is Image:
|
||||
let byteLen = target.data.len * 4
|
||||
else:
|
||||
let byteLen = target.data.len
|
||||
when allowSimd and compiles(applyOpacitySimd):
|
||||
applyOpacitySimd(
|
||||
cast[ptr UncheckedArray[uint8]](image.data[0].addr),
|
||||
image.data.len * 4,
|
||||
opacity
|
||||
)
|
||||
return
|
||||
|
||||
let
|
||||
oddMask = mm_set1_epi16(cast[int16](0xff00))
|
||||
div255 = mm_set1_epi16(cast[int16](0x8081))
|
||||
zeroVec = mm_setzero_si128()
|
||||
opacityVec = mm_slli_epi16(mm_set1_epi16(cast[int16](opacity)), 8)
|
||||
for _ in 0 ..< byteLen div 16:
|
||||
when type(target) is Image:
|
||||
let index = i div 4
|
||||
else:
|
||||
let index = i
|
||||
|
||||
let values = mm_loadu_si128(target.data[index].addr)
|
||||
if mm_movemask_epi8(mm_cmpeq_epi16(values, zeroVec)) != 0xffff:
|
||||
var
|
||||
valuesEven = mm_slli_epi16(values, 8)
|
||||
valuesOdd = mm_and_si128(values, oddMask)
|
||||
valuesEven = mm_mulhi_epu16(valuesEven, opacityVec)
|
||||
valuesOdd = mm_mulhi_epu16(valuesOdd, opacityVec)
|
||||
valuesEven = mm_srli_epi16(mm_mulhi_epu16(valuesEven, div255), 7)
|
||||
valuesOdd = mm_srli_epi16(mm_mulhi_epu16(valuesOdd, div255), 7)
|
||||
mm_storeu_si128(
|
||||
target.data[index].addr,
|
||||
mm_or_si128(valuesEven, mm_slli_epi16(valuesOdd, 8))
|
||||
)
|
||||
|
||||
i += 16
|
||||
|
||||
when type(target) is Image:
|
||||
for j in i div 4 ..< target.data.len:
|
||||
var rgbx = target.data[j]
|
||||
rgbx.r = ((rgbx.r * opacity) div 255).uint8
|
||||
rgbx.g = ((rgbx.g * opacity) div 255).uint8
|
||||
rgbx.b = ((rgbx.b * opacity) div 255).uint8
|
||||
rgbx.a = ((rgbx.a * opacity) div 255).uint8
|
||||
target.data[j] = rgbx
|
||||
else:
|
||||
for j in i ..< target.data.len:
|
||||
target.data[j] = ((target.data[j] * opacity) div 255).uint8
|
||||
for i in 0 ..< image.data.len:
|
||||
var rgbx = image.data[i]
|
||||
rgbx.r = ((rgbx.r * opacity) div 255).uint8
|
||||
rgbx.g = ((rgbx.g * opacity) div 255).uint8
|
||||
rgbx.b = ((rgbx.b * opacity) div 255).uint8
|
||||
rgbx.a = ((rgbx.a * opacity) div 255).uint8
|
||||
image.data[i] = rgbx
|
||||
|
||||
proc invert*(image: Image) {.raises: [].} =
|
||||
## Inverts all of the colors and alpha.
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
let vec255 = mm_set1_epi8(cast[int8](255))
|
||||
for _ in 0 ..< image.data.len div 16:
|
||||
let
|
||||
a = mm_loadu_si128(image.data[i + 0].addr)
|
||||
b = mm_loadu_si128(image.data[i + 4].addr)
|
||||
c = mm_loadu_si128(image.data[i + 8].addr)
|
||||
d = mm_loadu_si128(image.data[i + 12].addr)
|
||||
mm_storeu_si128(image.data[i + 0].addr, mm_sub_epi8(vec255, a))
|
||||
mm_storeu_si128(image.data[i + 4].addr, mm_sub_epi8(vec255, b))
|
||||
mm_storeu_si128(image.data[i + 8].addr, mm_sub_epi8(vec255, c))
|
||||
mm_storeu_si128(image.data[i + 12].addr, mm_sub_epi8(vec255, d))
|
||||
i += 16
|
||||
when allowSimd and compiles(invertImageSimd):
|
||||
invertImageSimd(
|
||||
cast[ptr UncheckedArray[ColorRGBX]](image.data[0].addr),
|
||||
image.data.len
|
||||
)
|
||||
return
|
||||
|
||||
for j in i ..< image.data.len:
|
||||
var rgbx = image.data[j]
|
||||
for i in 0 ..< image.data.len:
|
||||
var rgbx = image.data[i]
|
||||
rgbx.r = 255 - rgbx.r
|
||||
rgbx.g = 255 - rgbx.g
|
||||
rgbx.b = 255 - rgbx.b
|
||||
rgbx.a = 255 - rgbx.a
|
||||
image.data[j] = rgbx
|
||||
image.data[i] = rgbx
|
||||
|
||||
# Inverting rgbx(50, 100, 150, 200) becomes rgbx(205, 155, 105, 55). This
|
||||
# is not a valid premultiplied alpha color.
|
||||
|
|
@ -564,22 +470,16 @@ proc newMask*(image: Image): Mask {.raises: [PixieError].} =
|
|||
## Returns a new mask using the alpha values of the image.
|
||||
result = newMask(image.width, image.height)
|
||||
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
for _ in 0 ..< image.data.len div 16:
|
||||
let
|
||||
a = mm_loadu_si128(image.data[i + 0].addr)
|
||||
b = mm_loadu_si128(image.data[i + 4].addr)
|
||||
c = mm_loadu_si128(image.data[i + 8].addr)
|
||||
d = mm_loadu_si128(image.data[i + 12].addr)
|
||||
mm_storeu_si128(
|
||||
result.data[i].addr,
|
||||
pack4xAlphaValues(a, b, c, d)
|
||||
)
|
||||
i += 16
|
||||
when allowSimd and compiles(newMaskFromImageSimd):
|
||||
newMaskFromImageSimd(
|
||||
cast[ptr UncheckedArray[uint8]](result.data[0].addr),
|
||||
cast[ptr UncheckedArray[ColorRGBX]](image.data[0].addr),
|
||||
image.data.len
|
||||
)
|
||||
return
|
||||
|
||||
for j in i ..< image.data.len:
|
||||
result.data[j] = image.data[j].a
|
||||
for i in 0 ..< image.data.len:
|
||||
result.data[i] = image.data[i].a
|
||||
|
||||
proc getRgbaSmooth*(
|
||||
image: Image, x, y: float32, wrapped = false
|
||||
|
|
|
|||
|
|
@ -2,11 +2,11 @@ import bumpy, chroma, common, system/memory, vmath
|
|||
|
||||
const allowSimd* = not defined(pixieNoSimd) and not defined(tcc)
|
||||
|
||||
when defined(amd64) and allowSimd:
|
||||
import nimsimd/runtimecheck, nimsimd/sse2, runtimechecked/avx, runtimechecked/avx2
|
||||
let
|
||||
cpuHasAvx* = checkInstructionSets({AVX})
|
||||
cpuHasAvx2* = checkInstructionSets({AVX, AVX2})
|
||||
when allowSimd:
|
||||
import simd
|
||||
|
||||
when defined(amd64):
|
||||
import nimsimd/sse2
|
||||
|
||||
template currentExceptionAsPixieError*(): untyped =
|
||||
## Gets the current exception and returns it as a PixieError with stack trace.
|
||||
|
|
@ -81,45 +81,20 @@ proc fillUnsafe*(
|
|||
## continuing for len indices.
|
||||
let rgbx = color.asRgbx()
|
||||
|
||||
# If we can use AVX, do so
|
||||
when defined(amd64) and allowSimd:
|
||||
if cpuHasAvx and len >= 64:
|
||||
fillUnsafeAvx(data, rgbx, start, len)
|
||||
return
|
||||
when allowSimd and compiles(fillUnsafeSimd):
|
||||
fillUnsafeSimd(
|
||||
cast[ptr UncheckedArray[ColorRGBX]](data[start].addr),
|
||||
len,
|
||||
rgbx
|
||||
)
|
||||
return
|
||||
|
||||
# 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)
|
||||
else:
|
||||
var i = start
|
||||
when defined(amd64) and allowSimd:
|
||||
# Align to 16 bytes
|
||||
var p = cast[uint](data[i].addr)
|
||||
while i < (start + len) and (p and 15) != 0:
|
||||
data[i] = rgbx
|
||||
inc i
|
||||
p += 4
|
||||
# When supported, SIMD fill until we run out of room
|
||||
let
|
||||
colorVec = mm_set1_epi32(cast[int32](rgbx))
|
||||
iterations = (start + len - i) div 8
|
||||
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
|
||||
var
|
||||
u32 = cast[uint32](rgbx)
|
||||
u64 = cast[uint64]([u32, u32])
|
||||
for _ in 0 ..< len div 2:
|
||||
copyMem(data[i].addr, u64.addr, 8)
|
||||
i += 2
|
||||
# Fill whatever is left the slow way
|
||||
for i in i ..< start + len:
|
||||
data[i] = rgbx
|
||||
for color in data.mitems:
|
||||
color = rgbx
|
||||
|
||||
const straightAlphaTable = block:
|
||||
var table: array[256, array[256, uint8]]
|
||||
|
|
@ -141,39 +116,14 @@ proc toStraightAlpha*(data: var seq[ColorRGBA | ColorRGBX]) {.raises: [].} =
|
|||
|
||||
proc toPremultipliedAlpha*(data: var seq[ColorRGBA | ColorRGBX]) {.raises: [].} =
|
||||
## Converts an image to premultiplied alpha from straight alpha.
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
if cpuHasAvx2:
|
||||
i = toPremultipliedAlphaAvx2(data)
|
||||
else:
|
||||
let
|
||||
alphaMask = mm_set1_epi32(cast[int32](0xff000000))
|
||||
oddMask = mm_set1_epi16(cast[int16](0xff00))
|
||||
div255 = mm_set1_epi16(cast[int16](0x8081))
|
||||
for _ in 0 ..< data.len div 4:
|
||||
let
|
||||
values = mm_loadu_si128(data[i].addr)
|
||||
alpha = mm_and_si128(values, alphaMask)
|
||||
eq = mm_cmpeq_epi8(values, alphaMask)
|
||||
if (mm_movemask_epi8(eq) and 0x00008888) != 0x00008888:
|
||||
let
|
||||
evenMultiplier = mm_or_si128(alpha, mm_srli_epi32(alpha, 16))
|
||||
oddMultiplier = mm_or_si128(evenMultiplier, alphaMask)
|
||||
var
|
||||
colorsEven = mm_slli_epi16(values, 8)
|
||||
colorsOdd = mm_and_si128(values, oddMask)
|
||||
colorsEven = mm_mulhi_epu16(colorsEven, evenMultiplier)
|
||||
colorsOdd = mm_mulhi_epu16(colorsOdd, oddMultiplier)
|
||||
colorsEven = mm_srli_epi16(mm_mulhi_epu16(colorsEven, div255), 7)
|
||||
colorsOdd = mm_srli_epi16(mm_mulhi_epu16(colorsOdd, div255), 7)
|
||||
mm_storeu_si128(
|
||||
data[i].addr,
|
||||
mm_or_si128(colorsEven, mm_slli_epi16(colorsOdd, 8))
|
||||
)
|
||||
i += 4
|
||||
when allowSimd and compiles(toPremultipliedAlphaSimd):
|
||||
toPremultipliedAlphaSimd(
|
||||
cast[ptr UncheckedArray[uint32]](data[0].addr),
|
||||
data.len
|
||||
)
|
||||
return
|
||||
|
||||
# Convert whatever is left
|
||||
for i in i ..< data.len:
|
||||
for i in 0 ..< data.len:
|
||||
var c = data[i]
|
||||
if c.a != 255:
|
||||
c.r = ((c.r.uint32 * c.a) div 255).uint8
|
||||
|
|
@ -182,41 +132,15 @@ proc toPremultipliedAlpha*(data: var seq[ColorRGBA | ColorRGBX]) {.raises: [].}
|
|||
data[i] = c
|
||||
|
||||
proc isOpaque*(data: var seq[ColorRGBX], start, len: int): bool =
|
||||
when defined(amd64) and allowSimd:
|
||||
if cpuHasAvx2 and len >= 64:
|
||||
return isOpaqueAvx2(data, start, len)
|
||||
when allowSimd and compiles(isOpaqueSimd):
|
||||
return isOpaqueSimd(
|
||||
cast[ptr UncheckedArray[ColorRGBX]](data[start].addr),
|
||||
len
|
||||
)
|
||||
|
||||
result = true
|
||||
|
||||
var i = start
|
||||
when defined(amd64) and allowSimd:
|
||||
# Align to 16 bytes
|
||||
var p = cast[uint](data[i].addr)
|
||||
while i < (start + len) and (p and 15) != 0:
|
||||
if data[i].a != 255:
|
||||
return false
|
||||
inc i
|
||||
p += 4
|
||||
|
||||
let
|
||||
vec255 = mm_set1_epi8(255)
|
||||
iterations = (start + len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm_load_si128(cast[pointer](p))
|
||||
values1 = mm_load_si128(cast[pointer](p + 16))
|
||||
values2 = mm_load_si128(cast[pointer](p + 32))
|
||||
values3 = mm_load_si128(cast[pointer](p + 48))
|
||||
values01 = mm_and_si128(values0, values1)
|
||||
values23 = mm_and_si128(values2, values3)
|
||||
values0123 = mm_and_si128(values01, values23)
|
||||
eq = mm_cmpeq_epi8(values0123, vec255)
|
||||
if (mm_movemask_epi8(eq) and 0x00008888) != 0x00008888:
|
||||
return false
|
||||
p += 64
|
||||
i += 16 * iterations
|
||||
|
||||
for i in i ..< start + len:
|
||||
for i in start ..< start + len:
|
||||
if data[i].a != 255:
|
||||
return false
|
||||
|
||||
|
|
@ -228,24 +152,7 @@ when defined(amd64) and allowSimd:
|
|||
finalColor = mm_packus_epi16(finalColor, mm_setzero_si128())
|
||||
cast[ColorRGBX](mm_cvtsi128_si32(finalColor))
|
||||
|
||||
proc packAlphaValues(v: M128i): M128i {.inline, raises: [].} =
|
||||
## Shuffle the alpha values for these 4 colors to the first 4 bytes
|
||||
result = mm_srli_epi32(v, 24)
|
||||
result = mm_packus_epi16(result, mm_setzero_si128())
|
||||
result = mm_packus_epi16(result, mm_setzero_si128())
|
||||
|
||||
proc pack4xAlphaValues*(i, j, k, l: M128i): M128i {.inline, raises: [].} =
|
||||
let
|
||||
i = packAlphaValues(i)
|
||||
j = mm_slli_si128(packAlphaValues(j), 4)
|
||||
k = mm_slli_si128(packAlphaValues(k), 8)
|
||||
l = mm_slli_si128(packAlphaValues(l), 12)
|
||||
mm_or_si128(mm_or_si128(i, j), mm_or_si128(k, l))
|
||||
|
||||
proc unpackAlphaValues*(v: M128i): M128i {.inline, raises: [].} =
|
||||
## Unpack the first 32 bits into 4 rgba(0, 0, 0, value)
|
||||
result = mm_unpacklo_epi8(mm_setzero_si128(), v)
|
||||
result = mm_unpacklo_epi8(mm_setzero_si128(), result)
|
||||
export pack4xAlphaValues, unpackAlphaValues
|
||||
|
||||
when defined(release):
|
||||
{.pop.}
|
||||
|
|
|
|||
|
|
@ -75,6 +75,10 @@ proc setValue*(mask: Mask, x, y: int, value: uint8) {.inline, raises: [].} =
|
|||
## Sets a value at (x, y) or does nothing if outside of bounds.
|
||||
mask[x, y] = value
|
||||
|
||||
proc fill*(mask: Mask, value: uint8) {.inline, raises: [].} =
|
||||
## Fills the mask with the value.
|
||||
fillUnsafe(mask.data, value, 0, mask.data.len)
|
||||
|
||||
proc minifyBy2*(mask: Mask, power = 1): Mask {.raises: [PixieError].} =
|
||||
## Scales the mask down by an integer scale.
|
||||
if power < 0:
|
||||
|
|
@ -179,9 +183,26 @@ proc magnifyBy2*(mask: Mask, power = 1): Mask {.raises: [PixieError].} =
|
|||
result.width * 4
|
||||
)
|
||||
|
||||
proc fill*(mask: Mask, value: uint8) {.inline, raises: [].} =
|
||||
## Fills the mask with the value.
|
||||
fillUnsafe(mask.data, value, 0, mask.data.len)
|
||||
proc applyOpacity*(mask: Mask, opacity: float32) {.raises: [].} =
|
||||
## Multiplies alpha of the image by opacity.
|
||||
let opacity = round(255 * opacity).uint16
|
||||
if opacity == 255:
|
||||
return
|
||||
|
||||
if opacity == 0:
|
||||
mask.fill(0)
|
||||
return
|
||||
|
||||
when allowSimd and compiles(applyOpacitySimd):
|
||||
applyOpacitySimd(
|
||||
cast[ptr UncheckedArray[uint8]](mask.data[0].addr),
|
||||
mask.data.len,
|
||||
opacity
|
||||
)
|
||||
return
|
||||
|
||||
for i in 0 ..< mask.data.len:
|
||||
mask.data[i] = ((mask.data[i] * opacity) div 255).uint8
|
||||
|
||||
proc getValueSmooth*(mask: Mask, x, y: float32): uint8 {.raises: [].} =
|
||||
## Gets a interpolated value with float point coordinates.
|
||||
|
|
@ -213,17 +234,15 @@ proc getValueSmooth*(mask: Mask, x, y: float32): uint8 {.raises: [].} =
|
|||
|
||||
proc invert*(mask: Mask) {.raises: [].} =
|
||||
## Inverts all of the values - creates a negative of the mask.
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
let vec255 = mm_set1_epi8(255)
|
||||
for _ in 0 ..< mask.data.len div 16:
|
||||
var values = mm_loadu_si128(mask.data[i].addr)
|
||||
values = mm_sub_epi8(vec255, values)
|
||||
mm_storeu_si128(mask.data[i].addr, values)
|
||||
i += 16
|
||||
when allowSimd and compiles(invertImageSimd):
|
||||
invertMaskSimd(
|
||||
cast[ptr UncheckedArray[uint8]](mask.data[0].addr),
|
||||
mask.data.len
|
||||
)
|
||||
return
|
||||
|
||||
for j in i ..< mask.data.len:
|
||||
mask.data[j] = 255 - mask.data[j]
|
||||
for i in 0 ..< mask.data.len:
|
||||
mask.data[i] = 255 - mask.data[i]
|
||||
|
||||
proc spread*(mask: Mask, spread: float32) {.raises: [PixieError].} =
|
||||
## Grows the mask by spread.
|
||||
|
|
@ -288,21 +307,16 @@ proc spread*(mask: Mask, spread: float32) {.raises: [PixieError].} =
|
|||
|
||||
proc ceil*(mask: Mask) {.raises: [].} =
|
||||
## A value of 0 stays 0. Anything else turns into 255.
|
||||
var i: int
|
||||
when defined(amd64) and allowSimd:
|
||||
let
|
||||
zeroVec = mm_setzero_si128()
|
||||
vec255 = mm_set1_epi8(255)
|
||||
for _ in 0 ..< mask.data.len div 16:
|
||||
var values = mm_loadu_si128(mask.data[i].addr)
|
||||
values = mm_cmpeq_epi8(values, zeroVec)
|
||||
values = mm_andnot_si128(values, vec255)
|
||||
mm_storeu_si128(mask.data[i].addr, values)
|
||||
i += 16
|
||||
when allowSimd and compiles(invertImageSimd):
|
||||
ceilMaskSimd(
|
||||
cast[ptr UncheckedArray[uint8]](mask.data[0].addr),
|
||||
mask.data.len
|
||||
)
|
||||
return
|
||||
|
||||
for j in i ..< mask.data.len:
|
||||
if mask.data[j] != 0:
|
||||
mask.data[j] = 255
|
||||
for i in 0 ..< mask.data.len:
|
||||
if mask.data[i] != 0:
|
||||
mask.data[i] = 255
|
||||
|
||||
proc blur*(mask: Mask, radius: float32, outOfBounds: uint8 = 0) {.raises: [PixieError].} =
|
||||
## Applies Gaussian blur to the image given a radius.
|
||||
|
|
|
|||
|
|
@ -1823,7 +1823,7 @@ proc fillHits(
|
|||
|
||||
proc fillShapes(
|
||||
image: Image,
|
||||
shapes: seq[Polygon],
|
||||
shapes: var seq[Polygon],
|
||||
color: SomeColor,
|
||||
windingRule: WindingRule,
|
||||
blendMode: BlendMode
|
||||
|
|
@ -1852,8 +1852,10 @@ proc fillShapes(
|
|||
var
|
||||
partitions = partitionSegments(segments, startY, pathHeight - startY)
|
||||
partitionIndex: int
|
||||
entryIndices = newSeq[int](partitions.maxEntryCount)
|
||||
numEntryIndices: int
|
||||
coverages = newSeq[uint8](pathWidth)
|
||||
hits = newSeq[(Fixed32, int16)](partitions.maxEntryCount)
|
||||
hits = newSeq[(Fixed32, int16)](entryIndices.len)
|
||||
numHits: int
|
||||
aa: bool
|
||||
|
||||
|
|
@ -1895,13 +1897,13 @@ proc fillShapes(
|
|||
y += partitionHeight
|
||||
continue
|
||||
|
||||
var
|
||||
allEntriesInScanlineSpanIt = true
|
||||
tmp: int
|
||||
entryIndices: array[2, int]
|
||||
var allEntriesInScanlineSpanIt = true
|
||||
numEntryIndices = 0
|
||||
|
||||
if partitions[partitionIndex].twoNonintersectingSpanningSegments:
|
||||
tmp = 2
|
||||
entryIndices = [0, 1]
|
||||
numEntryIndices = 2
|
||||
entryIndices[0] = 0
|
||||
entryIndices[1] = 1
|
||||
else:
|
||||
for i in 0 ..< partitions[partitionIndex].entries.len:
|
||||
if partitions[partitionIndex].entries[i].segment.to.y < y.float32 or
|
||||
|
|
@ -1911,14 +1913,10 @@ proc fillShapes(
|
|||
partitions[partitionIndex].entries[i].segment.to.y < (y + 1).float32:
|
||||
allEntriesInScanlineSpanIt = false
|
||||
break
|
||||
if tmp < 2:
|
||||
entryIndices[tmp] = i
|
||||
inc tmp
|
||||
else:
|
||||
tmp = 0
|
||||
break
|
||||
entryIndices[numEntryIndices] = i
|
||||
inc numEntryIndices
|
||||
|
||||
if allEntriesInScanlineSpanIt and tmp == 2:
|
||||
if allEntriesInScanlineSpanIt and numEntryIndices == 2:
|
||||
var
|
||||
left = partitions[partitionIndex].entries[entryIndices[0]]
|
||||
right = partitions[partitionIndex].entries[entryIndices[1]]
|
||||
|
|
|
|||
|
|
@ -7,28 +7,23 @@ when defined(release):
|
|||
{.push checks: off.}
|
||||
|
||||
proc fillUnsafeAvx*(
|
||||
data: var seq[ColorRGBX],
|
||||
rgbx: ColorRGBX,
|
||||
start, len: int
|
||||
data: ptr UncheckedArray[ColorRGBX],
|
||||
len: int,
|
||||
rgbx: ColorRGBX
|
||||
) =
|
||||
var
|
||||
i = start
|
||||
p = cast[uint](data[i].addr)
|
||||
# Align to 32 bytes
|
||||
while i < (start + len) and (p and 31) != 0:
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 31) != 0: # Align to 32 bytes
|
||||
data[i] = rgbx
|
||||
inc i
|
||||
p += 4
|
||||
# When supported, SIMD fill until we run out of room
|
||||
|
||||
let
|
||||
iterations = (start + len - i) div 8
|
||||
iterations = (len - i) div 8
|
||||
colorVec = mm256_set1_epi32(cast[int32](rgbx))
|
||||
for _ in 0 ..< iterations:
|
||||
mm256_store_si256(cast[pointer](p), colorVec)
|
||||
p += 32
|
||||
i += iterations * 8
|
||||
mm256_store_si256(data[i].addr, colorVec)
|
||||
i += 8
|
||||
# Fill whatever is left the slow way
|
||||
for i in i ..< start + len:
|
||||
for i in i ..< len:
|
||||
data[i] = rgbx
|
||||
|
||||
when defined(release):
|
||||
|
|
|
|||
|
|
@ -6,108 +6,96 @@ when defined(gcc) or defined(clang):
|
|||
when defined(release):
|
||||
{.push checks: off.}
|
||||
|
||||
proc isOneColorAvx2*(data: var seq[ColorRGBX], start, len: int): bool =
|
||||
proc isOneColorAvx2*(data: ptr UncheckedArray[ColorRGBX], len: int): bool =
|
||||
result = true
|
||||
|
||||
let color = data[0]
|
||||
|
||||
var
|
||||
i = start
|
||||
p = cast[uint](data[i].addr)
|
||||
# Align to 32 bytes
|
||||
while i < (start + len) and (p and 31) != 0:
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 31) != 0: # Align to 32 bytes
|
||||
if data[i] != color:
|
||||
return false
|
||||
inc i
|
||||
p += 4
|
||||
|
||||
let
|
||||
colorVec = mm256_set1_epi32(cast[int32](color))
|
||||
iterations = (start + len - i) div 16
|
||||
iterations = (len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm256_load_si256(cast[pointer](p))
|
||||
values1 = mm256_load_si256(cast[pointer](p + 32))
|
||||
values0 = mm256_load_si256(data[i].addr)
|
||||
values1 = mm256_load_si256(data[i + 8].addr)
|
||||
eq0 = mm256_cmpeq_epi8(values0, colorVec)
|
||||
eq1 = mm256_cmpeq_epi8(values1, colorVec)
|
||||
eq01 = mm256_and_si256(eq0, eq1)
|
||||
if mm256_movemask_epi8(eq01) != cast[int32](0xffffffff):
|
||||
return false
|
||||
p += 64
|
||||
i += 16 * iterations
|
||||
i += 16
|
||||
|
||||
for i in i ..< start + len:
|
||||
for i in i ..< len:
|
||||
if data[i] != color:
|
||||
return false
|
||||
|
||||
proc isTransparentAvx2*(data: var seq[ColorRGBX], start, len: int): bool =
|
||||
proc isTransparentAvx2*(data: ptr UncheckedArray[ColorRGBX], len: int): bool =
|
||||
result = true
|
||||
|
||||
var
|
||||
i = start
|
||||
p = cast[uint](data[i].addr)
|
||||
# Align to 32 bytes
|
||||
while i < (start + len) and (p and 31) != 0:
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 31) != 0: # Align to 32 bytes
|
||||
if data[i].a != 0:
|
||||
return false
|
||||
inc i
|
||||
p += 4
|
||||
|
||||
let
|
||||
vecZero = mm256_setzero_si256()
|
||||
iterations = (start + len - i) div 16
|
||||
iterations = (len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm256_load_si256(cast[pointer](p))
|
||||
values1 = mm256_load_si256(cast[pointer](p + 32))
|
||||
values0 = mm256_load_si256(data[i].addr)
|
||||
values1 = mm256_load_si256(data[i + 8].addr)
|
||||
values01 = mm256_or_si256(values0, values1)
|
||||
eq = mm256_cmpeq_epi8(values01, vecZero)
|
||||
if mm256_movemask_epi8(eq) != cast[int32](0xffffffff):
|
||||
return false
|
||||
p += 64
|
||||
i += 16 * iterations
|
||||
i += 16
|
||||
|
||||
for i in i ..< start + len:
|
||||
for i in i ..< len:
|
||||
if data[i].a != 0:
|
||||
return false
|
||||
|
||||
proc isOpaqueAvx2*(data: var seq[ColorRGBX], start, len: int): bool =
|
||||
proc isOpaqueAvx2*(data: ptr UncheckedArray[ColorRGBX], len: int): bool =
|
||||
result = true
|
||||
|
||||
var
|
||||
i = start
|
||||
p = cast[uint](data[i].addr)
|
||||
# Align to 32 bytes
|
||||
while i < (start + len) and (p and 31) != 0:
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 31) != 0: # Align to 32 bytes
|
||||
if data[i].a != 255:
|
||||
return false
|
||||
inc i
|
||||
p += 4
|
||||
|
||||
let
|
||||
vec255 = mm256_set1_epi8(255)
|
||||
iterations = (start + len - i) div 16
|
||||
iterations = (len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm256_load_si256(cast[pointer](p))
|
||||
values1 = mm256_load_si256(cast[pointer](p + 32))
|
||||
values0 = mm256_load_si256(data[i].addr)
|
||||
values1 = mm256_load_si256(data[i + 8].addr)
|
||||
values01 = mm256_and_si256(values0, values1)
|
||||
eq = mm256_cmpeq_epi8(values01, vec255)
|
||||
if (mm256_movemask_epi8(eq) and 0x88888888) != 0x88888888:
|
||||
return false
|
||||
p += 64
|
||||
i += 16 * iterations
|
||||
i += 16
|
||||
|
||||
for i in i ..< start + len:
|
||||
for i in i ..< len:
|
||||
if data[i].a != 255:
|
||||
return false
|
||||
|
||||
proc toPremultipliedAlphaAvx2*(data: var seq[ColorRGBA | ColorRGBX]): int =
|
||||
proc toPremultipliedAlphaAvx2*(
|
||||
data: ptr UncheckedArray[uint32],
|
||||
len: int
|
||||
): int =
|
||||
let
|
||||
alphaMask = mm256_set1_epi32(cast[int32](0xff000000))
|
||||
oddMask = mm256_set1_epi16(cast[int16](0xff00))
|
||||
div255 = mm256_set1_epi16(cast[int16](0x8081))
|
||||
for _ in 0 ..< data.len div 8:
|
||||
for _ in 0 ..< len div 8:
|
||||
let
|
||||
values = mm256_loadu_si256(data[result].addr)
|
||||
alpha = mm256_and_si256(values, alphaMask)
|
||||
|
|
|
|||
322
src/pixie/simd.nim
Normal file
322
src/pixie/simd.nim
Normal file
|
|
@ -0,0 +1,322 @@
|
|||
import chroma, vmath
|
||||
|
||||
when defined(release):
|
||||
{.push checks: off.}
|
||||
|
||||
when defined(amd64):
|
||||
import nimsimd/runtimecheck, nimsimd/sse2, runtimechecked/avx,
|
||||
runtimechecked/avx2
|
||||
|
||||
let
|
||||
cpuHasAvx* = checkInstructionSets({AVX})
|
||||
cpuHasAvx2* = checkInstructionSets({AVX, AVX2})
|
||||
|
||||
proc packAlphaValues(v: M128i): M128i {.inline.} =
|
||||
## Shuffle the alpha values for these 4 colors to the first 4 bytes.
|
||||
result = mm_srli_epi32(v, 24)
|
||||
result = mm_packus_epi16(result, mm_setzero_si128())
|
||||
result = mm_packus_epi16(result, mm_setzero_si128())
|
||||
|
||||
proc pack4xAlphaValues*(i, j, k, l: M128i): M128i {.inline.} =
|
||||
let
|
||||
i = packAlphaValues(i)
|
||||
j = mm_slli_si128(packAlphaValues(j), 4)
|
||||
k = mm_slli_si128(packAlphaValues(k), 8)
|
||||
l = mm_slli_si128(packAlphaValues(l), 12)
|
||||
mm_or_si128(mm_or_si128(i, j), mm_or_si128(k, l))
|
||||
|
||||
proc unpackAlphaValues*(v: M128i): M128i {.inline, raises: [].} =
|
||||
## Unpack the first 32 bits into 4 rgba(0, 0, 0, value).
|
||||
result = mm_unpacklo_epi8(mm_setzero_si128(), v)
|
||||
result = mm_unpacklo_epi8(mm_setzero_si128(), result)
|
||||
|
||||
proc fillUnsafeSimd*(
|
||||
data: ptr UncheckedArray[ColorRGBX],
|
||||
len: int,
|
||||
rgbx: ColorRGBX
|
||||
) =
|
||||
if cpuHasAvx and len >= 64:
|
||||
fillUnsafeAvx(data, len, rgbx)
|
||||
else:
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 15) != 0: # Align to 16 bytes
|
||||
data[i] = rgbx
|
||||
inc i
|
||||
|
||||
let
|
||||
colorVec = mm_set1_epi32(cast[int32](rgbx))
|
||||
iterations = (len - i) div 8
|
||||
for _ in 0 ..< iterations:
|
||||
mm_store_si128(data[i].addr, colorVec)
|
||||
mm_store_si128(data[i + 4].addr, colorVec)
|
||||
i += 8
|
||||
|
||||
for i in i ..< len:
|
||||
data[i] = rgbx
|
||||
|
||||
proc isOneColorSimd*(data: ptr UncheckedArray[ColorRGBX], len: int): bool =
|
||||
if cpuHasAvx2:
|
||||
return isOneColorAvx2(data, len)
|
||||
|
||||
result = true
|
||||
|
||||
let color = data[0]
|
||||
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 15) != 0: # Align to 16 bytes
|
||||
if data[i] != color:
|
||||
return false
|
||||
inc i
|
||||
|
||||
let
|
||||
colorVec = mm_set1_epi32(cast[int32](color))
|
||||
iterations = (len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm_load_si128(data[i].addr)
|
||||
values1 = mm_load_si128(data[i + 4].addr)
|
||||
values2 = mm_load_si128(data[i + 8].addr)
|
||||
values3 = mm_load_si128(data[i + 12].addr)
|
||||
eq0 = mm_cmpeq_epi8(values0, colorVec)
|
||||
eq1 = mm_cmpeq_epi8(values1, colorVec)
|
||||
eq2 = mm_cmpeq_epi8(values2, colorVec)
|
||||
eq3 = mm_cmpeq_epi8(values3, colorVec)
|
||||
eq0123 = mm_and_si128(mm_and_si128(eq0, eq1), mm_and_si128(eq2, eq3))
|
||||
if mm_movemask_epi8(eq0123) != 0xffff:
|
||||
return false
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
if data[i] != color:
|
||||
return false
|
||||
|
||||
proc isTransparentSimd*(data: ptr UncheckedArray[ColorRGBX], len: int): bool =
|
||||
if cpuHasAvx2:
|
||||
return isTransparentAvx2(data, len)
|
||||
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 15) != 0: # Align to 16 bytes
|
||||
if data[i].a != 0:
|
||||
return false
|
||||
inc i
|
||||
|
||||
result = true
|
||||
|
||||
let
|
||||
vecZero = mm_setzero_si128()
|
||||
iterations = (len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm_load_si128(data[i].addr)
|
||||
values1 = mm_load_si128(data[i + 4].addr)
|
||||
values2 = mm_load_si128(data[i + 8].addr)
|
||||
values3 = mm_load_si128(data[i + 12].addr)
|
||||
values01 = mm_or_si128(values0, values1)
|
||||
values23 = mm_or_si128(values2, values3)
|
||||
values0123 = mm_or_si128(values01, values23)
|
||||
if mm_movemask_epi8(mm_cmpeq_epi8(values0123, vecZero)) != 0xffff:
|
||||
return false
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
if data[i].a != 0:
|
||||
return false
|
||||
|
||||
proc isOpaqueSimd*(data: ptr UncheckedArray[ColorRGBX], len: int): bool =
|
||||
if cpuHasAvx2:
|
||||
return isOpaqueAvx2(data, len)
|
||||
|
||||
result = true
|
||||
|
||||
var i: int
|
||||
while i < len and (cast[uint](data[i].addr) and 15) != 0: # Align to 16 bytes
|
||||
if data[i].a != 255:
|
||||
return false
|
||||
inc i
|
||||
|
||||
let
|
||||
vec255 = mm_set1_epi8(255)
|
||||
iterations = (len - i) div 16
|
||||
for _ in 0 ..< iterations:
|
||||
let
|
||||
values0 = mm_load_si128(data[i].addr)
|
||||
values1 = mm_load_si128(data[i + 4].addr)
|
||||
values2 = mm_load_si128(data[i + 8].addr)
|
||||
values3 = mm_load_si128(data[i + 12].addr)
|
||||
values01 = mm_and_si128(values0, values1)
|
||||
values23 = mm_and_si128(values2, values3)
|
||||
values0123 = mm_and_si128(values01, values23)
|
||||
eq = mm_cmpeq_epi8(values0123, vec255)
|
||||
if (mm_movemask_epi8(eq) and 0x00008888) != 0x00008888:
|
||||
return false
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
if data[i].a != 255:
|
||||
return false
|
||||
|
||||
proc toPremultipliedAlphaSimd*(data: ptr UncheckedArray[uint32], len: int) =
|
||||
var i: int
|
||||
if cpuHasAvx2:
|
||||
i = toPremultipliedAlphaAvx2(data, len)
|
||||
else:
|
||||
let
|
||||
alphaMask = mm_set1_epi32(cast[int32](0xff000000))
|
||||
oddMask = mm_set1_epi16(cast[int16](0xff00))
|
||||
div255 = mm_set1_epi16(cast[int16](0x8081))
|
||||
for _ in 0 ..< len div 4:
|
||||
let
|
||||
values = mm_loadu_si128(data[i].addr)
|
||||
alpha = mm_and_si128(values, alphaMask)
|
||||
eq = mm_cmpeq_epi8(values, alphaMask)
|
||||
if (mm_movemask_epi8(eq) and 0x00008888) != 0x00008888:
|
||||
let
|
||||
evenMultiplier = mm_or_si128(alpha, mm_srli_epi32(alpha, 16))
|
||||
oddMultiplier = mm_or_si128(evenMultiplier, alphaMask)
|
||||
var
|
||||
colorsEven = mm_slli_epi16(values, 8)
|
||||
colorsOdd = mm_and_si128(values, oddMask)
|
||||
colorsEven = mm_mulhi_epu16(colorsEven, evenMultiplier)
|
||||
colorsOdd = mm_mulhi_epu16(colorsOdd, oddMultiplier)
|
||||
colorsEven = mm_srli_epi16(mm_mulhi_epu16(colorsEven, div255), 7)
|
||||
colorsOdd = mm_srli_epi16(mm_mulhi_epu16(colorsOdd, div255), 7)
|
||||
mm_storeu_si128(
|
||||
data[i].addr,
|
||||
mm_or_si128(colorsEven, mm_slli_epi16(colorsOdd, 8))
|
||||
)
|
||||
i += 4
|
||||
|
||||
for i in i ..< len:
|
||||
var c: ColorRGBX
|
||||
copyMem(c.addr, data[i].addr, 4)
|
||||
c.r = ((c.r.uint32 * c.a) div 255).uint8
|
||||
c.g = ((c.g.uint32 * c.a) div 255).uint8
|
||||
c.b = ((c.b.uint32 * c.a) div 255).uint8
|
||||
copyMem(data[i].addr, c.addr, 4)
|
||||
|
||||
proc newImageFromMaskSimd*(
|
||||
dst: ptr UncheckedArray[ColorRGBX],
|
||||
src: ptr UncheckedArray[uint8],
|
||||
len: int
|
||||
) =
|
||||
var i: int
|
||||
for _ in 0 ..< len div 16:
|
||||
var alphas = mm_loadu_si128(src[i].addr)
|
||||
for j in 0 ..< 4:
|
||||
var unpacked = unpackAlphaValues(alphas)
|
||||
unpacked = mm_or_si128(unpacked, mm_srli_epi32(unpacked, 8))
|
||||
unpacked = mm_or_si128(unpacked, mm_srli_epi32(unpacked, 16))
|
||||
mm_storeu_si128(dst[i + j * 4].addr, unpacked)
|
||||
alphas = mm_srli_si128(alphas, 4)
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
let v = src[i]
|
||||
dst[i] = rgbx(v, v, v, v)
|
||||
|
||||
proc newMaskFromImageSimd*(
|
||||
dst: ptr UncheckedArray[uint8],
|
||||
src: ptr UncheckedArray[ColorRGBX],
|
||||
len: int
|
||||
) =
|
||||
var i: int
|
||||
for _ in 0 ..< len div 16:
|
||||
let
|
||||
a = mm_loadu_si128(src[i + 0].addr)
|
||||
b = mm_loadu_si128(src[i + 4].addr)
|
||||
c = mm_loadu_si128(src[i + 8].addr)
|
||||
d = mm_loadu_si128(src[i + 12].addr)
|
||||
mm_storeu_si128(
|
||||
dst[i].addr,
|
||||
pack4xAlphaValues(a, b, c, d)
|
||||
)
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
dst[i] = src[i].a
|
||||
|
||||
proc invertImageSimd*(data: ptr UncheckedArray[ColorRGBX], len: int) =
|
||||
var i: int
|
||||
let vec255 = mm_set1_epi8(cast[int8](255))
|
||||
for _ in 0 ..< len div 16:
|
||||
let
|
||||
a = mm_loadu_si128(data[i + 0].addr)
|
||||
b = mm_loadu_si128(data[i + 4].addr)
|
||||
c = mm_loadu_si128(data[i + 8].addr)
|
||||
d = mm_loadu_si128(data[i + 12].addr)
|
||||
mm_storeu_si128(data[i + 0].addr, mm_sub_epi8(vec255, a))
|
||||
mm_storeu_si128(data[i + 4].addr, mm_sub_epi8(vec255, b))
|
||||
mm_storeu_si128(data[i + 8].addr, mm_sub_epi8(vec255, c))
|
||||
mm_storeu_si128(data[i + 12].addr, mm_sub_epi8(vec255, d))
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
var rgbx = data[i]
|
||||
rgbx.r = 255 - rgbx.r
|
||||
rgbx.g = 255 - rgbx.g
|
||||
rgbx.b = 255 - rgbx.b
|
||||
rgbx.a = 255 - rgbx.a
|
||||
data[i] = rgbx
|
||||
|
||||
toPremultipliedAlphaSimd(cast[ptr UncheckedArray[uint32]](data), len)
|
||||
|
||||
proc invertMaskSimd*(data: ptr UncheckedArray[uint8], len: int) =
|
||||
var i: int
|
||||
let vec255 = mm_set1_epi8(255)
|
||||
for _ in 0 ..< len div 16:
|
||||
var values = mm_loadu_si128(data[i].addr)
|
||||
values = mm_sub_epi8(vec255, values)
|
||||
mm_storeu_si128(data[i].addr, values)
|
||||
i += 16
|
||||
|
||||
for j in i ..< len:
|
||||
data[j] = 255 - data[j]
|
||||
|
||||
proc ceilMaskSimd*(data: ptr UncheckedArray[uint8], len: int) =
|
||||
var i: int
|
||||
let
|
||||
zeroVec = mm_setzero_si128()
|
||||
vec255 = mm_set1_epi8(255)
|
||||
for _ in 0 ..< len div 16:
|
||||
var values = mm_loadu_si128(data[i].addr)
|
||||
values = mm_cmpeq_epi8(values, zeroVec)
|
||||
values = mm_andnot_si128(values, vec255)
|
||||
mm_storeu_si128(data[i].addr, values)
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
if data[i] != 0:
|
||||
data[i] = 255
|
||||
|
||||
proc applyOpacitySimd*(
|
||||
data: ptr UncheckedArray[uint8],
|
||||
len: int,
|
||||
opacity: uint16
|
||||
) =
|
||||
var i: int
|
||||
let
|
||||
oddMask = mm_set1_epi16(cast[int16](0xff00))
|
||||
div255 = mm_set1_epi16(cast[int16](0x8081))
|
||||
zeroVec = mm_setzero_si128()
|
||||
opacityVec = mm_slli_epi16(mm_set1_epi16(cast[int16](opacity)), 8)
|
||||
for _ in 0 ..< len div 16:
|
||||
let values = mm_loadu_si128(data[i].addr)
|
||||
if mm_movemask_epi8(mm_cmpeq_epi16(values, zeroVec)) != 0xffff:
|
||||
var
|
||||
valuesEven = mm_slli_epi16(values, 8)
|
||||
valuesOdd = mm_and_si128(values, oddMask)
|
||||
valuesEven = mm_mulhi_epu16(valuesEven, opacityVec)
|
||||
valuesOdd = mm_mulhi_epu16(valuesOdd, opacityVec)
|
||||
valuesEven = mm_srli_epi16(mm_mulhi_epu16(valuesEven, div255), 7)
|
||||
valuesOdd = mm_srli_epi16(mm_mulhi_epu16(valuesOdd, div255), 7)
|
||||
mm_storeu_si128(
|
||||
data[i].addr,
|
||||
mm_or_si128(valuesEven, mm_slli_epi16(valuesOdd, 8))
|
||||
)
|
||||
i += 16
|
||||
|
||||
for i in i ..< len:
|
||||
data[i] = ((data[i] * opacity) div 255).uint8
|
||||
|
||||
when defined(release):
|
||||
{.pop.}
|
||||
Loading…
Reference in a new issue