Improve Axis3 interactivity performance

[ffreyer]

Description

Change in text

I noticed that Axis3 can be rather sluggish at times and one reason for it seem to be text updates. A large part of the time spent on updating the angles in Axis3 was in the update of camera.projectionview. From there I intercepted the updates with timed updates with this:

using GLMakie, LaTeXStrings
fig = Figure()
ax = Axis3(fig[1, 1], xlabel = L"Test $\frac{1}{1+x}$")

function print_time(f, i)
    x -> begin
        t = time()
        val = f(x)
        dt = time() - t
        if dt > 1e-3
            printstyled("\nTime: $dt\n", color = :red)
            println("@ index $i")
            flush(stdout)
        end
        return val
    end
end

src = ax.scene.camera.projectionview
for i in eachindex(src.listeners)
    src.listeners[i] = print_time(src.listeners[i], i)
end

fig

and found that most of the time was spend in add_ticks_and_ticklabels!, in (and after) label_pos_rot_valign. (You can find that out by doing autocompletes on the functions held by the observable.)

Well, long story shot - I removed the position trigger from text layouting and that reduced this particular timing from ~0.06s to ~0.006s. I also changed the method with (string, pos) pairs to only update when necessary to avoid rebuilding glyphcollections too much.

In Axis3

For this I tested with src = ax.azimuth. After the text changes the first 3 listeners took ~0.006, ~0.006 and ~0.012s. The first two are mi1 and mi2 and the last is the calculation of projectionview. I changed mi1, mi2 and mi3 to only update on changes which practically eliminates the first two calls.

I also split up some observables for the axis label to avoid triggering text layouting unnecessarily. That (or the changes text! for latexstrings) reduced the last timing to ~0.006s for me.

Type of change

• Performance

[ffreyer]

With the changes in the pr this doesn't work anymore for CairoMakie, but updating the nodes in either order normally works

[SimonDanisch]

what does that mean?

[ffreyer]

I think I made the textnode.val = ... -> textnode[] = ... change explicitly for CairoMakie tests to pass

[SimonDanisch]

ah that way around...

[SimonDanisch]

As far as I can tell, this works pretty well :) tested a few latex + text updates and Axis3 interaction

[ffreyer]

It would probably be worth testing updates to attributes like xticklabelalign as well. I think those are the most likely to stop working from this

[jkrumbiegel]

Looks good to me although I didn't have time to run it.

[SimonDanisch]

Hm,

f, ax, pl = scatter(1:4; axis=(xticklabelalign=(:left, :bottom),))
ax.xticklabelalign = (:right, :top)

Doesn't update, but also doesn't update on last tagged version..Also, the block refactor actually broke setting it completely, but I'll make a PR to fix that.

[ffreyer]

https://github.com/JuliaPlots/Makie.jl/blob/b3b0c6daaff5a7c5d5ed36310e5a65fc11832b67/GLMakie/src/drawing_primitives.jl#L313-L327

This is what makes updates of textsize, align, font and probably everything else that isn't position, offset or transfunc fail with Vector{GlyphCollection}. Why is it there?

[ffreyer]

Good old synchronization issues is why it's there...

[ffreyer]

Didn't mean to link the second issue here lol. If tests pass the issue with updates to text attributes not propagating properly with vector positions should be fixed.

[ffreyer]

Somehow this changed the colors in "Dynamically adjusting number of particles in a meshscatter"... It looks fine though. Should we merge this with the attribute update fixes anyway?

[SimonDanisch]

Somehow this changed the colors in "Dynamically adjusting number of particles in a meshscatter".

ColorTypes.jl changed their rand implementation, which lead to this... I rebased against master and updated the ref images, so hopefully should turn green now!