Skip to content

[TST] Adjust tests to be more tolerant to floating point math operations being imprecise - #25813

Merged
oscargus merged 1 commit into
matplotlib:mainfrom
ksunden:npy_prec_robustness
May 6, 2023
Merged

oscargus merged 1 commit into
matplotlib:mainfrom
ksunden:npy_prec_robustness

Conversation

@ksunden

@ksunden ksunden commented May 4, 2023

Copy link
Copy Markdown
Member

PR summary

Closes #25789, see there for detailed discussion of what changed

This is addressing the extremely small (literally Unit of Least Precision, ULP) changes to how numpy computes sin/cos on main.
Through a combination of ratios of small numbers and relying on precise edge case behavior in tests of floats, we saw 3 cases where the miniscule difference compounded to test failures.

Two cases of projections having degenerate points where the edge case is inherently unstable:

  • Mollweide projection is degenerate at the poles, we were lucky that the floating point operations precisely canceled
    • Fixed by just not testing round tripping of the poles, as they shouldn't necessarily round trip
  • Formatting of polar coordinates
    • Polar coordinates are degenerate at r=0, thus we were relying on boundary behavior to get zero decimals of precision
    • Does an absolute difference mod 2 pi, thus the maximum possible is pi, and then divides by pi, thus the only way to get the tested precision was to get exactly np.pi
      • The only place this can actually happen in adjacent pixels is around the origin
    • The method is used for tool tip in interactive plotting, and precisely placing the mouse at 0,0 is unlikely to impossible depending on resolution, etc
    • Fixed by moving the radius out far enough that we are no longer degenerate and have stable expected precision

one case of attempting to do equality checks on processed floats:

  • Round tripping of 3d axis colors with light source at 90 degrees
  • since pi/2 is imprecise (because all floats except sums of powers of two are) there should be some allowance for float precision error
  • The tolerance is still really tight (4 ULPs, or a relative tolerance of ~1e-15)

PR checklist

…ons being imprecise

Two cases of projections having degenerate points where the edge case is inherently unstable, one case of attempting to do equality checks on processed floats
@ksunden

ksunden commented May 5, 2023

Copy link
Copy Markdown
Member Author

Not sure why azure didn't run here (I was initially hoping that our tk/macos azure image problems were resolved when I saw a green checkmark instead of a red X, but no, no azure at all here...)

@oscargus

oscargus commented May 5, 2023 •

Copy link
Copy Markdown
Member

(because all floats except sums of powers of two, the power differing at most 53 are)

@tacaswell tacaswell added this to the v3.7.2 milestone May 5, 2023
@oscargus
oscargus merged commit 4bdae2e into matplotlib:main May 6, 2023
meeseeksmachine pushed a commit to meeseeksmachine/matplotlib that referenced this pull request May 6, 2023
…to floating point math operations being imprecise
QuLogic added a commit that referenced this pull request May 6, 2023
…813-on-v3.7.x

Backport PR #25813 on branch v3.7.x ([TST] Adjust tests to be more tolerant to floating point math operations being imprecise)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[TST] Upcoming dependency test failures

3 participants