|
The fdct.go and idct.go files were derived from the MPEG-SSG and
JPEG-IJG reference code and therefore carry licenses specific to those
groups. Various license checkers flag these files as potentially
problematic. The code is also not terribly well documented.
This CL fixes the license problem by adding a new, from-scratch
implementation using a different algorithm. As a bonus, the new code
is both faster and more accurate than the old encumbered code.
On speed, the new code is up to 20% faster; benchmarks below.
On accuracy, in the set of blocks used in the test, we can measure
the number of output values that are off-by-one from the exact rounded answer.
The old FDCT was off in 8.6% of values; the new one is off in 2.5%.
The old IDCT was off in 1.4% of values; the new one is off in 1.2%.
goos: darwin
goarch: arm64
pkg: image/jpeg
cpu: Apple M3 Pro
│ old │ new │
│ sec/op │ sec/op vs base │
FDCT-12 619.6n ± 3% 586.5n ± 1% -5.34% (p=0.000 n=10)
IDCT-12 752.4n ± 4% 628.0n ± 1% -16.54% (p=0.000 n=10)
goos: linux
goarch: amd64
cpu: Intel(R) Xeon(R) CPU @ 2.30GHz
│ old │ new │
│ sec/op │ sec/op vs base │
FDCT-16 1.817µ ± 0% 1.542µ ± 0% -15.11% (p=0.000 n=10)
IDCT-16 1.897µ ± 0% 1.514µ ± 0% -20.22% (p=0.000 n=10)
goos: linux
goarch: arm64
cpu: whatever gotip-linux-arm64 has
│ old │ new │
│ sec/op │ sec/op vs base │
FDCT-8 1.844µ ± 0% 1.847µ ± 0% +0.14% (p=0.000 n=10)
IDCT-8 2.127µ ± 0% 1.973µ ± 0% -7.26% (p=0.000 n=10)
Change-Id: Ie6d14103c8478ba5a779f234da84f345828eb925
Reviewed-on: https://go-review.googlesource.com/c/go/+/705518
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Alan Donovan <adonovan@google.com>
Auto-Submit: Russ Cox <rsc@golang.org>
Reviewed-by: Nigel Tao <nigeltao@google.com>
|
|
Before, calling the RGBA method of YCbCr color would return red values
in the range [0x0080, 0xff80]. After, the range is [0x0000, 0xffff] and
is consistent with what Gray colors' RGBA method returns. In particular,
pure black, pure white and every Gray color in between are now exactly
representable as a YCbCr color.
This fixes a regression from Go 1.4 (where YCbCr{0x00, 0x80, 0x80} was
no longer equivalent to pure black), introduced by golang.org/cl/8073 in
the Go 1.5 development cycle. In Go 1.4, the +0x80 rounding was not
noticable when Cb == 0x80 && Cr == 0x80, because the YCbCr to RGBA
conversion truncated to 8 bits before multiplying by 0x101, so the
output range was [0x0000, 0xffff].
The TestYCbCrRoundtrip fuzzy-match tolerance grows from 1 to 2 because
the YCbCr to RGB conversion now maps to an ever-so-slightly larger
range, along with the usual imprecision of accumulating rounding errors.
Also s/int/int32/ in ycbcr.go. The conversion shouldn't overflow either
way, as int is always at least 32 bits, but it does make it clearer that
the computation doesn't depend on sizeof(int).
Fixes #11691
Change-Id: I538ca0adf7e040fa96c5bc8b3aef4454535126b9
Reviewed-on: https://go-review.googlesource.com/12220
Reviewed-by: Rob Pike <r@golang.org>
|
