darktable
darktable copied to clipboard
darktable-org
three regressions tonight
We have three regressions detected tonight, this is due to some merge done yesterday.
Errors 3
- 0016-lowpass-bilateral
- 0019-color-mapping
- 0043-dithering-fs
The culprit is:
d540794cc07a3531825b6b2b33c1c93e3a96e5a0 is the first bad commit
commit d540794cc07a3531825b6b2b33c1c93e3a96e5a0
Author: ralfbrown <[email protected]>
Date: Sat Jul 30 18:24:42 2022 -0400
rewrite dt_Lab_to_XYZ to permit vector instructions
src/common/colorspaces_inline_conversions.h | 13 +++++++------
1 file changed, 7 insertions(+), 6 deletions(-)
For test 0043 it is actually one of:
There are only 'skip'ped commits left to test.
The first bad commit could be any of:
e9079df756fb0f52cad213a0ecac1639cd1d406d
b30c22a0bccb0eef8242c69886494677b9e77255
We cannot bisect more!
So
commit b30c22a0bccb0eef8242c69886494677b9e77255 (refs/bisect/bad)
Author: Ralf Brown <[email protected]>
Date: Sat Jul 30 18:30:43 2022 -0400
remove colorout.c SSE codepath
commit e9079df756fb0f52cad213a0ecac1639cd1d406d (refs/bisect/skip-e9079df756fb0f52cad213a0ecac1639cd1d406d)
Author: Ralf Brown <[email protected]>
Date: Sat Jul 30 18:27:26 2022 -0400
cleanup colorout.c
Together with the rewrite of dt_Lab_to_XYZ, the compiler-generated
code is now essentially the same speed as the SSE codepath.
Thr AutoV SSE
1 0.1028 0.1014
2 0.0522 0.0510
4 0.0268 0.0260
8 0.0140 0.0130
12 0.0120 0.0100
16 0.0100 0.0090
24 0.0070 0.0070
32 0.0073 0.0070
Test 0016 & 0019 have only a very limited number of pixel changed (216 and 1952), but test 0043 has 935056 which is bit more worrisome.
0016:
0019:
0043:
@ralfbrown : Can you have a look at the three mentioned commits above? TIA.
Since the problematic commits are the entirety of #12258, it might be best to temporarily revert. I won't be able to investigate until the weekend.
@ralfbrown : Here is what I have done:
I have reverted:
Revert "cleanup colorout.c"
This reverts commit e9079df756fb0f52cad213a0ecac1639cd1d406d.
and
Revert "remove colorout.c SSE codepath"
This reverts commit b30c22a0bccb0eef8242c69886494677b9e77255.
To fix test 0043.
And I have fixed the dt_Lab_to_XYZ() implementation which fixes 0016 & 0019 (do not hesitate to double check my changes).
Checking into this, I've discovered that current master (after the reversion) gives maxDE of 0.00 for tests 0016 and 0019 only when building RelWithDebInfo. Build type Release still gives maxDE of 8.9 and 15.1, respectively. My guess is that my recent changes merely enabled the compiler to perform some optimization at -O2 which was previously performed only at -O3.
The output difference on 0043 is due to some pixels being pushed across the quantization threshold when rounding errors change due to different code generation. The Floyd-Steinberg error diffusion then propagates that difference right and down.
More to come.