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

bit of neon in png

Browse source
This commit is contained in:
Ryan Oldenburg 2022-07-31 23:10:19 -05:00
parent 93e9ef79b0
commit e4afb30f2d
1 changed files with 17 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

25
src/pixie/fileformats/png.nim
View file

@ -129,16 +129,25 @@ proc unfilter(
uncompressedStartIdx = uncompressedIdx(1, y) uncompressedStartIdx = uncompressedIdx(1, y)
unfilteredStartIx = unfiteredIdx(0, y) unfilteredStartIx = unfiteredIdx(0, y)
var x: int var x: int
when allowSimd and defined(amd64): when allowSimd and (defined(amd64) or defined(arm64)):
if y - 1 >= 0: if y - 1 >= 0:
for _ in 0 ..< rowBytes div 16: for _ in 0 ..< rowBytes div 16:
let when defined(amd64):
bytes = mm_loadu_si128(uncompressed[uncompressedStartIdx + x].addr) let
up = mm_loadu_si128(result[unfilteredStartIx + x - rowBytes].addr) bytes = mm_loadu_si128(uncompressed[uncompressedStartIdx + x].addr)
mm_storeu_si128( up = mm_loadu_si128(result[unfilteredStartIx + x - rowBytes].addr)
result[unfilteredStartIx + x].addr, mm_storeu_si128(
mm_add_epi8(bytes, up) result[unfilteredStartIx + x].addr,
) mm_add_epi8(bytes, up)
)
else: # arm64
let
bytes = vld1q_u8(uncompressed[uncompressedStartIdx + x].addr)
up = vld1q_u8(result[unfilteredStartIx + x - rowBytes].addr)
vst1q_u8(
result[unfilteredStartIx + x].addr,
vaddq_u8(bytes, up)
)
x += 16 x += 16
for x in x ..< rowBytes: for x in x ..< rowBytes:
var value = uncompressed[uncompressedStartIdx + x] var value = uncompressed[uncompressedStartIdx + x]