fix contour boundingbox with NaNs

[t-bltg]

Description

Fixes discord issue (fix #3211):

let
    a = b = 1:3
    c = [0 1 2; 1 2 3; 4 5 6]
    ct1 = contour(a, b, c; levels = [1, 2, 3], labels = true)   # OK
    d = [0 1 2; 1 2 3; 4 5 NaN]
    ct2 = contour(a, b, d; levels = [1, 2, 3], labels = true)   # NOK
end

Fix a CairoMakie bug when glyphs positions contained a NaN, corrupting the Cairo drawing state / context.

Type of change

Delete options that do not apply:

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)

Checklist

• Added an entry in NEWS.md (for new features and breaking changes)
• Added or changed relevant sections in the documentation
• Added unit tests for new algorithms, conversion methods, etc.
• Added reference image tests for new plotting functions, recipes, visual options, etc.

[Lightup1]

Hi @t-bltg ,
I tried this branch.
There is no error message now, but there are also no contour lines now.

[Lightup1]

[jkrumbiegel]

Which backend was this?

[Lightup1]

@jkrumbiegel CairoMakie

[t-bltg]

Let me have a look this morning.

[jkrumbiegel]

Could be a bug in CairoMakie, too, because there is one little line segment and afterwards nothing, also the axis lines are missing. That usually happens when Cairo silently corrupts the surface state, which can happen with NaNs sometimes.

[t-bltg]

Let's reason first with NaN-less arrays: c = [0 1 2; 1 2 3; 4 5 6].

What is visually represented is reverse(c; dims=1):

3×3 Matrix{Float64}:
 4.0  5.0    6.0
 1.0  2.0    3.0
 0.0  1.0    2.0

So this is contourf(a, b, c; levels):

With GLMakie (or CairoMakie), contour(a, b, c; levels, labels=true):
. Something could probably be done to enhance the label position for 2.

Now with NaNs, d = [0 1 2; 1 2 3; 4 5 NaN], contourf(a, b, d; levels):
.

So I will expect at least the 1 and 2 levels lines with labels to appear (3 is more tricky because I use NaNs internally to mask contour segments hidden by the label).

It is also difficult to work with such small arrays implying short contour segments.

I'll try to rework the PR.

[Lightup1]

GLMakie works fine now (except missing 2 labels).
CairoMakie produces no contour lines.

[t-bltg]

With GLMakie:

Lines appear, but missing 2 contour labels.

Missing lines on CairoMakie.

[t-bltg]

I'm marking as ready for review because I have no ideas left currently.
The contour algorithm output is different whether there is a NaN or not.

[Lightup1]

Dear @t-bltg ,
Is there any quick fixing way to make Makie ignore NaN?
I tried to replace NaN with 1e6 in some cases it works fine, but in some other cases (when the border is smaller than the specified levels) it will draw multiple lines near the border between normal values and NaN.
Does julia support void element in a matrix which is ignored by Makie?

[t-bltg]

Does julia support void element in a matrix which is ignored by Makie?

I would want missing (or nothing) to behave how you describe unfortunately it doesn't work either in this case.

EDIT: on hold since it breaks the examples of #2496.

EDIT2: fixed.

[jkrumbiegel]

I don't think contour algorithms work with values you want to "ignore", NaNs always cause these gaps because in their surroundings the interpolation is not defined. However, if you want that behavior you can linearly interpolate the points yourself, as linear interpolation is what the contour algorithm does anyway.

[jkrumbiegel]

Aside from the NaN-interpolation question discussed above, is this now in mergeable state? With NaN causing the appropriate gaps?

[t-bltg]

Current state, output from the code snippet in #3210 (comment).

CairoMakie - no Nans c

CairoMakie - Nans d

GLMakie - no Nans c

GLMakie - Nans d

[jkrumbiegel]

Ok so CairoMakie is still buggy

[t-bltg]

Is this inconsistency between CairoMakie and GLMakie backends happening at backend level (the way every backend deals with segments with NaNs) or is this something we can solve in Makie ?

[jkrumbiegel]

This must be a NaN-handling bug in CairoMakie. It may not actually draw NaNs, they have to be skipped (but one has to be careful to create the right paths in the process). And here it seems it's drawn (or something else that doesn't work) because the axis lines are missing, so Cairo broke in the process of drawing this.

[t-bltg]

Oh right I didn't notice the axis lines missing.
Do you have any hint where this might occur ?
I'll investigate before the end of the week ...

[jkrumbiegel]

Yeah probably in the code I refactored a while ago, it's a bit more complex because it tries to draw lines together as long as they have the same color and width. https://github.com/MakieOrg/Makie.jl/blob/master/CairoMakie/src/primitives.jl#L207-L288

[t-bltg]

@jkrumbiegel, this is the new output, the bug was lying in the routine drawing glyphs.

The output is now consistent across CairoMakie and GLMakie, so I think this PR can be reviewed again (pending CI state).

GLMakie - no NaNs c

GLMakie - NaNs d

CairoMakie - no NaNs c

CairoMakie - NaNs d

[jkrumbiegel]

Could you remove all the non functionality related formatting changes? Makes it harder to see what's up. Also the empty ifs with comments can be left as they are, I know they technically don't do anything but they helped me understand what the code was doing.

[t-bltg]

Could you remove all the non functionality related formatting changes?

Done.

Makes it harder to see what's up.

I agree for the review, although these changes were make to simplify and make the current code consistent.

Also the empty ifs with comments can be left as they are

Done.

[jkrumbiegel]

Thank you! Looks good to me