mirror of
https://github.com/ddidderr/wuffs.git
synced 2026-11-19 16:47:28 +01:00
std/jpeg: decode truncated JPEGs on MCU boundaries
This only affects truncated (malformed) JPEGs. It moves Wuffs' std/jpeg decoder closer to libjpeg-turbo, as seen by the "hash of the decoded output" for test/data/peacock.default.truncated.jpeg now matching test/data/peacock.default.truncated.decoded-by-djpeg.png instead of test/data/peacock.default.truncated.decoded-by-wuffs.png The decoding of test/data/peacock.progressive.truncated.jpeg is not yet affected, requiring further investigation, but a future commit could similarly make that match djpeg. The two libraries don't use exactly the same algorithm because, for performance, Wuffs' decoder reads from the bitstream 8 bytes at a time where libjpeg-turbo reads only 1 byte at a time. While this change moves Wuffs' output closer to libjpeg-turbo (to repeat, only for truncated JPEGs; well-formed JPEGs are unaffected), the benchmarks below show a negative performance impact, and visual inspection of test/data/peacock.*.truncated.decoded-by-*.png shows that matching libjpeg-turbo *exactly* on truncated JPEGs doesn't actually always look better. On balance, it doesn't seem worth it to keep this change (and spend more time on that further investigation). This commit will soon be followed by a rollback, but it is committed anyway so that we can refer to these numbers in the git log. name old speed new speed delta wuffs_jpeg_decode_19k_8bpp/clang11 378MB/s ± 0% 364MB/s ± 0% -3.53% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_progressive/clang11 325MB/s ± 0% 340MB/s ± 0% +4.62% (p=0.016 n=5+4) wuffs_jpeg_decode_30k_24bpp_sequential/clang11 1.02GB/s ± 0% 0.99GB/s ± 0% -2.64% (p=0.008 n=5+5) wuffs_jpeg_decode_77k_24bpp/clang11 842MB/s ± 0% 815MB/s ± 0% -3.16% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_420/clang11 916MB/s ± 0% 887MB/s ± 0% -3.08% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_444/clang11 1.10GB/s ± 0% 1.07GB/s ± 2% -3.39% (p=0.008 n=5+5) wuffs_jpeg_decode_4002k_24bpp/clang11 919MB/s ± 0% 892MB/s ± 0% -2.95% (p=0.008 n=5+5) wuffs_jpeg_decode_19k_8bpp/gcc10 360MB/s ± 0% 348MB/s ± 0% -3.34% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_progressive/gcc10 290MB/s ± 0% 283MB/s ± 0% -2.46% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_sequential/gcc10 925MB/s ± 0% 909MB/s ± 0% -1.70% (p=0.008 n=5+5) wuffs_jpeg_decode_77k_24bpp/gcc10 758MB/s ± 0% 732MB/s ± 0% -3.43% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_420/gcc10 829MB/s ± 0% 799MB/s ± 0% -3.54% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_444/gcc10 981MB/s ± 0% 964MB/s ± 0% -1.72% (p=0.008 n=5+5) wuffs_jpeg_decode_4002k_24bpp/gcc10 835MB/s ± 0% 811MB/s ± 0% -2.89% (p=0.008 n=5+5)
This commit is contained in:
1 parent
721463be95
commit
a3cecef69a
4 files changed
+41
-10
No files matched your search
@@ -280,6 +280,15 @@ pub struct decoder? implements base.image_decoder(
|
||||
bitstream_ri : base.u32[..= 0x800],
|
||||
bitstream_wi : base.u32[..= 0x800],
|
||||
|
||||
// This is the index (like bitstream_ri and bitstream_wi) where "EOF
|
||||
// means implicit zeroes" kicked in. It only affects truncated
|
||||
// (malformed) JPEGs, and in that case, encourages decoding to finish
|
||||
// on an MCU boundary. This behavior is closer to how libjpeg-turbo
|
||||
// works (again, on truncated JPEGs; well-formed JPEGs are unaffected),
|
||||
// but it does have a performance penalty. It's not clear that it's
|
||||
// worth keeping.
|
||||
bitstream_zi : base.u32[..= 0x800],
|
||||
|
||||
bitstream_is_closed : base.bool,
|
||||
|
||||
// bitstream_padding is the number (per restart) of implicit zeroes
|
||||
@@ -1278,7 +1287,7 @@ pri func decoder.decode_sos?(src: base.io_reader, workbuf: slice base.u8) {
|
||||
this.fill_bitstream!(src: args.src)
|
||||
|
||||
my = 0
|
||||
while my < this.scan_height_in_mcus {
|
||||
while.outer my < this.scan_height_in_mcus {
|
||||
assert my < 0x2000 via "a < b: a < c; c <= b"(c: this.scan_height_in_mcus)
|
||||
mx = 0
|
||||
while mx < this.scan_width_in_mcus,
|
||||
@@ -1315,9 +1324,11 @@ pri func decoder.decode_sos?(src: base.io_reader, workbuf: slice base.u8) {
|
||||
return "#bad SOS marker"
|
||||
} else if args.src.is_closed() and (not this.bitstream_is_closed) {
|
||||
if this.bitstream_wi < (0x800 / 2) {
|
||||
assert this.bitstream_wi <= (this.bitstream_wi + 264) via "a <= (a + b): 0 <= b"()
|
||||
this.bitstream_buffer[this.bitstream_wi .. (this.bitstream_wi + 264)].bulk_memset!(byte_value: 0)
|
||||
this.bitstream_wi += 264
|
||||
// 999 is arbitrary, but bigger than (2 * 264).
|
||||
assert this.bitstream_wi <= (this.bitstream_wi + 999) via "a <= (a + b): 0 <= b"()
|
||||
this.bitstream_zi = this.bitstream_wi
|
||||
this.bitstream_buffer[this.bitstream_wi .. (this.bitstream_wi + 999)].bulk_memset!(byte_value: 0)
|
||||
this.bitstream_wi += 999
|
||||
this.bitstream_is_closed = true
|
||||
}
|
||||
break.fill_bitstream
|
||||
@@ -1351,10 +1362,14 @@ pri func decoder.decode_sos?(src: base.io_reader, workbuf: slice base.u8) {
|
||||
}
|
||||
}
|
||||
|
||||
if this.bitstream_is_closed and (this.bitstream_ri >= (this.bitstream_zi + (this.bitstream_n_bits / 8))) {
|
||||
break.outer
|
||||
}
|
||||
|
||||
mx += 1
|
||||
} endwhile
|
||||
my += 1
|
||||
} endwhile
|
||||
} endwhile.outer
|
||||
|
||||
this.scan_count ~sat+= 1
|
||||
}
|
||||
@@ -1643,6 +1658,10 @@ pri func decoder.fill_bitstream!(src: base.io_reader) {
|
||||
var c : base.u8
|
||||
var new_wi : base.u32[..= 0x800]
|
||||
|
||||
if this.bitstream_is_closed {
|
||||
return nothing
|
||||
}
|
||||
|
||||
// Compact unread bytes to the start of the buffer.
|
||||
if this.bitstream_ri <= 0 {
|
||||
// No-op.
|
||||
|
||||
Reference in new issue
Block a user