PR: Improve CFI 2017 and TM30-18 Performance #1120

tjdcs · 2023-03-14T03:09:13Z

Summary

This PR implements several performance improvements, again, that should improve the speed and quality of the entire code base. These improvements mainly revolve around caching in certain places during sd_to_XYZ in order to prevent excessive copies of CMFs etc...

Because these features change caching behavior there is a small risk that there will be regressions, despite the tests passing so it is requested that it is tried in some other developer applications before merging.

2X performance improvement in CFI 2017 workloads

Preflight

Code Style and Quality

Unit tests have been implemented and passed.
Pyright static checking has been run and passed.
Pre-commit hooks have been run and passed.
~~New transformations have been added to the Automatic Colour Conversion Graph.~~
~~New transformations have been exported to the relevant namespaces, e.g. colour, colour.models.~~

Documentation

~~New features are documented along with examples if relevant.~~
~~The documentation is Sphinx and numpydoc compliant.~~

tjdcs · 2023-03-14T03:19:15Z

Looks like I did actually mess up the tests some how... I was really hoping that was my local environment quarks.

tjdcs · 2023-03-14T18:17:41Z

@KelSolaar looks like the only issue left is the PCA issue. We'll have to figure out a plan for those PCA tests. FYI three of them are failing on my local machine. I suspect that small changes in the code base are affecting floating point errors and that's feeding forward into PCA results.

coveralls · 2023-04-14T19:27:47Z

coverage: 99.773% (-0.008%) from 99.781% when pulling 074721b on performance/cfi2017 into 6b8b5d4 on develop.

tjdcs · 2023-04-24T18:04:52Z

Status: around 42x faster than v.0.4.2 release

tjdcs · 2023-04-25T00:56:05Z

@KelSolaar we have about an 100x performance improvement for TM30 in this branch now. I think it should be merged. I'd like to get another 2x faster than this... but that will require more significant refactoring. First targets if anyone wants to pick it up are the Ohno 2013 code and the CAM code...

I might introduce / write a paper on a highly compute optimized CAM. That could be nice in the future.

Anyway... The only failing tests look like some of those precision errors. The tests pass on my local machine. Let me know if you want me to fix the tests, or what you'd like to do to pass this.

And of course... looking forward to your review.

KelSolaar

Thanks a lot @tjdcs!

A lot of good and exciting stuff there. I did a first review pass. I would be keen to clean up the commits a bit as discussed on Slack.

KelSolaar · 2023-04-28T21:21:49Z

colour/appearance/ciecam02.py

@@ -231,6 +230,7 @@ def XYZ_to_CIECAM02(
        "Average"
    ],
    discount_illuminant: bool = False,
+    compute_HQ: bool = True,


Major

Wondering if we should not generalise which correlates to compute. Having modified the CIECAM02 signature begs for modifying other CAMs for consistency. We could also have an inlined version of the model that is optimised for speed. We have a Hellwig et al. (2022) implementation that is inlined here for example: https://github.com/colour-science/colour/blob/feature/flattened_hellwig2022/colour/appearance/hellwig2022.py#L154

My personal preference actually is that all color models are flattened, or close to it. Even models that are substantially similar.

As for generalizing a "which components to compute" that sounds like a great feature... but I don't want to address that in this PR. I like the proposal though.

colour/appearance/ciecam02.py

colour/appearance/hunt.py

colour/algebra/common.py

colour/temperature/tests/test_ohno2013.py

colour/temperature/ohno2013.py

KelSolaar · 2023-05-01T09:39:26Z

colour/temperature/ohno2013.py

+    -------
+    :class:`numpy.ndarray`
+        *CIE UCS* colourspace *uv* chromaticity coordinates.
+


Minor

Ditto for example.

.gitignore

tjdcs · 2023-05-04T03:28:15Z

@KelSolaar This is greatly cleaned up now if you want to give it another pass.

tjdcs · 2023-05-04T05:53:03Z

ad67a6b adds a dependency on xxhash which is an extremely fast hashing algorithm specifically suited for dictionary keys. https://github.com/Cyan4973/xxHash#quality

KelSolaar · 2023-05-26T21:53:23Z

@tjdcs : Where are we with this one?

tjdcs · 2023-05-26T22:39:30Z

Bizarre corner cases in the initialization of SpectralDistribution and Signal appear to be failing tests. And there is a mountain of type errors that I can't wade through because my development environment and the CI environment are different. CI complains there are only a few type errors, but pyright on my machine reports more than 600.

It's going to be a small while before I can address these on my own. :/ i have to prep for cinegear, and then infocomm. I have urgent work projects taking the few days I have in between.

Probably the end of June is when I will have serious time to sit down with this again.

KelSolaar · 2023-05-26T22:41:01Z

Thanks, sounds good! I'm going to update Pyright on our side so that we run the latest and then we can rebase your branch on top of develop.

KelSolaar · 2023-05-28T20:57:46Z

develop is up-to-date against the latest pyright and a few fixes have been rolled for static analysis.

KelSolaar · 2023-06-27T20:31:39Z

I looked quickly why the docs build is failing and I think that we are destroying the blackbody spectral distributions that we defined here :

colour/utilities/generate_plots.py

Line 187 in 36d9255

blackbody_sds = [

, there:

colour/colour/colorimetry/tristimulus_values.py

Line 858 in 36d9255

if sd.shape.boundaries != cmfs.shape.boundaries:

This is one of the reason why reshape_sd was always operating on a copy, now we have introduced those kind of hard to debug side effects. I think we need to use a context manager or expose the copy argument in the definitions that use reshape_sd. Food for thoughts!

I would be keen to drop the Poetry commit from this PR as it is a separate problem and just created a huge mess on my end, took me a while to realise what had happened.

Repair documenation build error fix test due to default interpolation change [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci fix peculier init case in MaxOS environment fix type checking Auto stash before revert of "Improve default interpolation performance"

fix domain / range copy on signal init fix doctests

tjdcs · 2023-06-29T19:36:37Z

@KelSolaar ready for a final review, assuming that you are OK if I hurt coverage by 3 lines...

- Remove unnecessary interpolator change. - Improve documentation around spectral generator interpolation changes.

KelSolaar · 2023-07-16T00:45:11Z

Just pushed a small varnishing pass, hopefully we can merge right after if everything builds.

KelSolaar · 2023-07-16T04:49:55Z

LGTM, happy to merge?

tjdcs · 2023-07-16T04:53:47Z

Make it so!

KelSolaar · 2023-07-16T07:33:32Z

Merged, thanks for the great work @tjdcs! Do you remember roughly the performance improvements for the release notes?

tjdcs · 2023-07-16T07:47:34Z

45x from pre-merge commit. From when the work started I think it was around 100x I can check that tomorrow. Some other performance stuff got merged in early or there were other changes.

I backed out some of the hashing improvements which is the next big thing, but in order to do it with good code style we need to make xxhash a default install. (need to access private signal variables in multisignal __hash__)

tjdcs · 2023-07-16T16:18:33Z

colour_fidelity_index_ANSIIESTM3018 100x compared to v0.4.2

uv_to_CCT_Ohno2013 5x faster than v.0.4.2 which might also be of interest in a release note

tjdcs requested a review from KelSolaar March 14, 2023 03:09

tjdcs force-pushed the performance/cfi2017 branch from 6d4e98a to 283e544 Compare March 14, 2023 15:26

tjdcs force-pushed the performance/cfi2017 branch from ee2899b to 8045c1c Compare April 14, 2023 19:02

tjdcs mentioned this pull request Apr 26, 2023

Feature/lru cache #1142

Closed

7 tasks

KelSolaar mentioned this pull request Apr 30, 2023

PR: Implement support for callbacks in colour.continuous.AbstractContinuousFunction class. #1145

Closed

5 tasks

KelSolaar requested changes May 1, 2023

View reviewed changes

tjdcs force-pushed the performance/cfi2017 branch from b1ca2a0 to a7d8c37 Compare May 4, 2023 02:00

tjdcs force-pushed the performance/cfi2017 branch 2 times, most recently from c832828 to ad67a6b Compare May 4, 2023 05:49

KelSolaar mentioned this pull request May 5, 2023

PR: Implement support for fast integer digest using "xxhash". #1149

Merged

5 tasks

tjdcs force-pushed the performance/cfi2017 branch from 382a8c9 to 009700b Compare May 9, 2023 12:34

KelSolaar mentioned this pull request May 13, 2023

PR: Implement support for better ndarray copy management in colour.continuous.Signal class. #1151

Merged

5 tasks

tjdcs force-pushed the performance/cfi2017 branch 2 times, most recently from 44dc34a to f83b0f1 Compare May 17, 2023 22:19

KelSolaar changed the title ~~Performance/cfi2017~~ PR: Improve CFI 2017 and TM30-18 Performance May 26, 2023

tjdcs force-pushed the performance/cfi2017 branch 3 times, most recently from 9dc7021 to 36d9255 Compare June 27, 2023 07:41

tjdcs added 6 commits June 29, 2023 14:51

Remove illuminant modifications side effect

46082a9

re-architect output to include CAM

d306a3f

Re-evaluate speed in planck generation

bbae504

fix tests

52c688f

fix domain / range copy on signal init fix doctests

trim could cause an incorrect result in some edge cases (neq interval)

0558836

tjdcs force-pushed the performance/cfi2017 branch from b9ad315 to 0558836 Compare June 29, 2023 18:51

tjdcs and others added 12 commits June 29, 2023 16:23

Hash signal objects now that we have the power of xxhash

766fa39

resolve old comments on PR

d8834a9

Improve hue quadrature computation consistency in CAMs.

979b80d

Improve spectral generator default interpolation changes.

f2387dc

- Remove unnecessary interpolator change. - Improve documentation around spectral generator interpolation changes.

Improve unit tests for colour.planck_law definition.

352ce1a

Remove colour quality scale unnecessary unit test changes.

1357e33

Revert DataColorimetry_TCS_CIE2017 fields to "Colour" 0.4.2 ordering.

97e1a4a

Update various doctests.

b58e982

Update various imports.

580b114

Varnish some performance code changes.

bc1c254

Imeplement or simplify directory change in various utility scripts.

807032e

Update various docstrings.

074721b

tjdcs mentioned this pull request Jul 16, 2023

[FEATURE]: Performance improvements for ohno2013 #1053

Closed

KelSolaar merged commit ca0bba3 into develop Jul 16, 2023
25 of 26 checks passed

KelSolaar deleted the performance/cfi2017 branch July 16, 2023 07:32

KelSolaar added this to the v0.4.3 milestone Jul 16, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

PR: Improve CFI 2017 and TM30-18 Performance #1120

PR: Improve CFI 2017 and TM30-18 Performance #1120

tjdcs commented Mar 14, 2023 •

edited

tjdcs commented Mar 14, 2023

tjdcs commented Mar 14, 2023 •

edited

coveralls commented Apr 14, 2023 •

edited

tjdcs commented Apr 24, 2023

tjdcs commented Apr 25, 2023 •

edited

KelSolaar left a comment

KelSolaar Apr 28, 2023

tjdcs May 4, 2023

KelSolaar May 1, 2023

tjdcs commented May 4, 2023

tjdcs commented May 4, 2023

KelSolaar commented May 26, 2023

tjdcs commented May 26, 2023

KelSolaar commented May 26, 2023

KelSolaar commented May 28, 2023

KelSolaar commented Jun 27, 2023

tjdcs commented Jun 29, 2023

KelSolaar commented Jul 16, 2023

KelSolaar commented Jul 16, 2023

tjdcs commented Jul 16, 2023

KelSolaar commented Jul 16, 2023

tjdcs commented Jul 16, 2023

tjdcs commented Jul 16, 2023 •

edited

PR: Improve CFI 2017 and TM30-18 Performance #1120

PR: Improve CFI 2017 and TM30-18 Performance #1120

Conversation

tjdcs commented Mar 14, 2023 • edited

Summary

Preflight

tjdcs commented Mar 14, 2023

tjdcs commented Mar 14, 2023 • edited

coveralls commented Apr 14, 2023 • edited

tjdcs commented Apr 24, 2023

tjdcs commented Apr 25, 2023 • edited

KelSolaar left a comment

Choose a reason for hiding this comment

KelSolaar Apr 28, 2023

Choose a reason for hiding this comment

tjdcs May 4, 2023

Choose a reason for hiding this comment

KelSolaar May 1, 2023

Choose a reason for hiding this comment

tjdcs commented May 4, 2023

tjdcs commented May 4, 2023

KelSolaar commented May 26, 2023

tjdcs commented May 26, 2023

KelSolaar commented May 26, 2023

KelSolaar commented May 28, 2023

KelSolaar commented Jun 27, 2023

tjdcs commented Jun 29, 2023

KelSolaar commented Jul 16, 2023

KelSolaar commented Jul 16, 2023

tjdcs commented Jul 16, 2023

KelSolaar commented Jul 16, 2023

tjdcs commented Jul 16, 2023

tjdcs commented Jul 16, 2023 • edited

tjdcs commented Mar 14, 2023 •

edited

tjdcs commented Mar 14, 2023 •

edited

coveralls commented Apr 14, 2023 •

edited

tjdcs commented Apr 25, 2023 •

edited

tjdcs commented Jul 16, 2023 •

edited