MyouProject/pixie
3
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge pull request #347 from guzba/master

Browse source 
simd draw fast paths
This commit is contained in:
treeform 2021-12-15 10:47:59 -08:00 committed by GitHub
commit 3cfab49185
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 183 additions and 43 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

15
src/pixie/blends.nim
View file

@ -518,7 +518,7 @@ when defined(amd64) and not defined(pixieNoSimd):
proc blendNormalSimd(backdrop, source: M128i): M128i = proc blendNormalSimd(backdrop, source: M128i): M128i =
blendNormalInlineSimd(backdrop, source) blendNormalInlineSimd(backdrop, source)
proc blendMaskSimd(backdrop, source: M128i): M128i = proc blendMaskInlineSimd*(backdrop, source: M128i): M128i {.inline.} =
let let
alphaMask = mm_set1_epi32(cast[int32](0xff000000)) alphaMask = mm_set1_epi32(cast[int32](0xff000000))
oddMask = mm_set1_epi16(cast[int16](0xff00)) oddMask = mm_set1_epi16(cast[int16](0xff00))
@ -539,6 +539,9 @@ when defined(amd64) and not defined(pixieNoSimd):
mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8)) mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
proc blendMaskSimd(backdrop, source: M128i): M128i =
blendMaskInlineSimd(backdrop, source)
proc blendOverwriteSimd(backdrop, source: M128i): M128i = proc blendOverwriteSimd(backdrop, source: M128i): M128i =
source source
@ -555,7 +558,7 @@ when defined(amd64) and not defined(pixieNoSimd):
## Is there a blend function for a given blend mode with SIMD support? ## Is there a blend function for a given blend mode with SIMD support?
blendMode in {bmNormal, bmMask, bmOverwrite} blendMode in {bmNormal, bmMask, bmOverwrite}
proc maskNormalSimd(backdrop, source: M128i): M128i = proc maskNormalInlineSimd*(backdrop, source: M128i): M128i {.inline.} =
## Blending masks ## Blending masks
let let
oddMask = mm_set1_epi16(cast[int16](0xff00)) oddMask = mm_set1_epi16(cast[int16](0xff00))
@ -592,7 +595,10 @@ when defined(amd64) and not defined(pixieNoSimd):
mm_or_si128(blendedEven, mm_slli_epi16(blendedOdd, 8)) mm_or_si128(blendedEven, mm_slli_epi16(blendedOdd, 8))
proc maskMaskSimd(backdrop, source: M128i): M128i = proc maskNormalSimd(backdrop, source: M128i): M128i =
maskNormalInlineSimd(backdrop, source)
proc maskMaskInlineSimd*(backdrop, source: M128i): M128i =
let let
oddMask = mm_set1_epi16(cast[int16](0xff00)) oddMask = mm_set1_epi16(cast[int16](0xff00))
div255 = mm_set1_epi16(cast[int16](0x8081)) div255 = mm_set1_epi16(cast[int16](0x8081))
@ -613,6 +619,9 @@ when defined(amd64) and not defined(pixieNoSimd):
mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8)) mm_or_si128(backdropEven, mm_slli_epi16(backdropOdd, 8))
proc maskMaskSimd(backdrop, source: M128i): M128i =
maskMaskInlineSimd(backdrop, source)
proc maskerSimd*(blendMode: BlendMode): MaskerSimd {.raises: [PixieError].} = proc maskerSimd*(blendMode: BlendMode): MaskerSimd {.raises: [PixieError].} =
## Returns a blend masking function with SIMD support. ## Returns a blend masking function with SIMD support.
case blendMode: case blendMode:

133
src/pixie/images.nim
View file

@ -798,7 +798,138 @@ proc drawUber(
continue continue
when defined(amd64) and not defined(pixieNoSimd): when defined(amd64) and not defined(pixieNoSimd):
# Check we are not rotated case blendMode:
of bmOverwrite:
for _ in 0 ..< (xStop - xStart) div 16:
let
srcPos = p + dx * x.float32 + dy * y.float32
sx = srcPos.x.int
sy = srcPos.y.int
when type(a) is Image:
when type(b) is Image:
for q in [0, 4, 8, 12]:
let sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx + q, sy)].addr)
mm_storeu_si128(a.data[a.dataIndex(x + q, y)].addr, sourceVec)
else: # b is a Mask
var values = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
for q in [0, 4, 8, 12]:
let sourceVec = unpackAlphaValues(values)
mm_storeu_si128(a.data[a.dataIndex(x + q, y)].addr, sourceVec)
# Shuffle 32 bits off for the next iteration
values = mm_srli_si128(values, 4)
else: # a is a Mask
when type(b) is Image:
var
i = mm_loadu_si128(b.data[b.dataIndex(sx + 0, sy)].addr)
j = mm_loadu_si128(b.data[b.dataIndex(sx + 4, sy)].addr)
k = mm_loadu_si128(b.data[b.dataIndex(sx + 8, sy)].addr)
l = mm_loadu_si128(b.data[b.dataIndex(sx + 12, sy)].addr)
let sourceVec = pack4xAlphaValues(i, j, k, l)
else: # b is a Mask
let sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
mm_storeu_si128(a.data[a.dataIndex(x, y)].addr, sourceVec)
x += 16
of bmNormal:
let vec255 = mm_set1_epi32(cast[int32](uint32.high))
for _ in 0 ..< (xStop - xStart) div 16:
let
srcPos = p + dx * x.float32 + dy * y.float32
sx = srcPos.x.int
sy = srcPos.y.int
when type(a) is Image:
when type(b) is Image:
for q in [0, 4, 8, 12]:
let sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx + q, sy)].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(a.data[a.dataIndex(x + q, y)].addr, sourceVec)
else:
let backdropVec = mm_loadu_si128(a.data[a.dataIndex(x + q, y)].addr)
mm_storeu_si128(
a.data[a.dataIndex(x + q, y)].addr,
blendNormalInlineSimd(backdropVec, sourceVec)
)
else: # b is a Mask
var values = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
for q in [0, 4, 8, 12]:
let sourceVec = unpackAlphaValues(values)
if mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, mm_setzero_si128())) != 0xffff:
if (mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, vec255)) and 0x8888) == 0x8888:
discard
else:
let backdropVec = mm_loadu_si128(a.data[a.dataIndex(x + q, y)].addr)
mm_storeu_si128(
a.data[a.dataIndex(x + q, y)].addr,
blendNormalInlineSimd(backdropVec, sourceVec)
)
# Shuffle 32 bits off for the next iteration
values = mm_srli_si128(values, 4)
else: # a is a Mask
let backdropVec = mm_loadu_si128(a.data[a.dataIndex(x, y)].addr)
when type(b) is Image:
var
i = mm_loadu_si128(b.data[b.dataIndex(sx + 0, sy)].addr)
j = mm_loadu_si128(b.data[b.dataIndex(sx + 4, sy)].addr)
k = mm_loadu_si128(b.data[b.dataIndex(sx + 8, sy)].addr)
l = mm_loadu_si128(b.data[b.dataIndex(sx + 12, sy)].addr)
let sourceVec = pack4xAlphaValues(i, j, k, l)
else: # b is a Mask
let sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
mm_storeu_si128(
a.data[a.dataIndex(x, y)].addr,
maskNormalInlineSimd(backdropVec, sourceVec)
)
x += 16
of bmMask:
let vec255 = mm_set1_epi32(cast[int32](uint32.high))
for _ in 0 ..< (xStop - xStart) div 16:
let
srcPos = p + dx * x.float32 + dy * y.float32
sx = srcPos.x.int
sy = srcPos.y.int
when type(a) is Image:
when type(b) is Image:
for q in [0, 4, 8, 12]:
let sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx + q, sy)].addr)
if mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, mm_setzero_si128())) == 0xffff:
mm_storeu_si128(a.data[a.dataIndex(x + q, y)].addr, mm_setzero_si128())
elif mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, vec255)) != 0xffff:
let backdropVec = mm_loadu_si128(a.data[a.dataIndex(x + q, y)].addr)
mm_storeu_si128(
a.data[a.dataIndex(x + q, y)].addr,
blendMaskInlineSimd(backdropVec, sourceVec)
)
else: # b is a Mask
var values = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
for q in [0, 4, 8, 12]:
let sourceVec = unpackAlphaValues(values)
if mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, mm_setzero_si128())) == 0xffff:
mm_storeu_si128(a.data[a.dataIndex(x + q, y)].addr, mm_setzero_si128())
elif (mm_movemask_epi8(mm_cmpeq_epi8(sourceVec, vec255)) and 0x8888) != 0x8888:
let backdropVec = mm_loadu_si128(a.data[a.dataIndex(x + q, y)].addr)
mm_storeu_si128(
a.data[a.dataIndex(x + q, y)].addr,
blendMaskInlineSimd(backdropVec, sourceVec)
)
# Shuffle 32 bits off for the next iteration
values = mm_srli_si128(values, 4)
else: # a is a Mask
let backdropVec = mm_loadu_si128(a.data[a.dataIndex(x, y)].addr)
when type(b) is Image:
var
i = mm_loadu_si128(b.data[b.dataIndex(sx + 0, sy)].addr)
j = mm_loadu_si128(b.data[b.dataIndex(sx + 4, sy)].addr)
k = mm_loadu_si128(b.data[b.dataIndex(sx + 8, sy)].addr)
l = mm_loadu_si128(b.data[b.dataIndex(sx + 12, sy)].addr)
let sourceVec = pack4xAlphaValues(i, j, k, l)
else: # b is a Mask
let sourceVec = mm_loadu_si128(b.data[b.dataIndex(sx, sy)].addr)
mm_storeu_si128(
a.data[a.dataIndex(x, y)].addr,
maskMaskInlineSimd(backdropVec, sourceVec)
)
x += 16
else:
when type(a) is Image: when type(a) is Image:
if blendMode.hasSimdBlender(): if blendMode.hasSimdBlender():
let blenderSimd = blendMode.blenderSimd() let blenderSimd = blendMode.blenderSimd()