Lines plot artifacts with GLMakie v0.10.2 on julia 1.10.3 (win10)

[devel-chm]

Previously posted on discourse at
https://discourse.julialang.org/t/lines-plot-artifacts-with-glmakie-v0-10-2-on-julia-1-10-3-win10/114933

I was plotting some curves with a lot of low level noise with
GLMakie and the plots had a lot of extra line cruft as if the
line segments for the lines in the plot were not being correctly
handled.

A MWE for the above example is

using GLMakie;
data = [1 + 2*rand() + i/100 for i in 1:10000];
lines(data)

This should show a line of slope 1/100 with random noise as
a thickening strip of size 2 superimposed.

This is much less than the figure dimensions so the plot
should resemble a thick line. I find that if I zoom in enough that
some/most of the spilling line ends clean up.

[devel-chm]

I tried this on two other machines, one win11 with nvidia graphics and another with win10
and intel graphics. These both showed the correct image:

[SimonDanisch]

What is the GPU of the faulty plot?

[ffreyer]

Could you also try this with the faulty plot?

begin
    using GLMakie;
    data = [1 + 2*rand() + i/100 for i in 1:10000];
    f, a, p = lines(data)
    p.attributes[:debug] = Observable(true)
    f
end

[devel-chm]

@SimonDanisch: GPU is AMD Radeon Pro WX 3200 series

@ffreyer: The resulting image:

[ffreyer]

Does joinstyle = :bevel fix this?

[SimonDanisch]

Looks like art!

[ffreyer]

The green parts in the debug plot are lines that are hidden due to joint truncation. Since there is more green towards the top and less spikes, I'm thinking that always truncating may hide the issue.
But we shouldn't have such long line segments in the first place. Something is wrong with the quad positions. My guess would be that normalize(a - b) or (a - b) / length(a - b) is unstable here.

Makie.jl/GLMakie/assets/shader/lines.geom

Lines 275 to 285 in f3e9013

	vec3 v1 = (p2 - p1);
	float segment_length = length(v1.xy);
	v1 /= segment_length;
	// depth is irrelevant for these
	vec2 v0 = v1.xy;
	vec2 v2 = v1.xy;
	if (p1 != p0 && isvalid[0])
	v0 = normalize(p1.xy - p0.xy);
	if (p3 != p2 && isvalid[3])
	v2 = normalize(p3.xy - p2.xy);

Maybe another good test would be to make the line alternate so that we should always get truncated joints. If we don't then the failure has to happen pretty early, i.e. probably with the normalizations.

begin
    using GLMakie;
    data = [1 + 2*(i % 2) + i/100 for i in 1:10000];
    f, a, p = lines(data)
    p.attributes[:debug] = Observable(true)
    f
end

Makie.jl/GLMakie/assets/shader/lines.geom

Lines 295 to 307 in f3e9013

	vec2 miter = vec2(dot(v0, v1.xy), dot(v1.xy, v2));
	vec2 miter_n1 = miter.x < 0.0 ?
	sign_no_zero(dot(v0.xy, n1)) * normalize(v0.xy - v1.xy) : normalize(n0 + n1);
	vec2 miter_n2 = miter.y < 0.0 ?
	sign_no_zero(dot(v1.xy, n2)) * normalize(v1.xy - v2.xy) : normalize(n1 + n2);
	// Are we truncating the joint based on miter limit or joinstyle?
	// bevel / always truncate doesn't work with v1 == v2 (v0) so we use allow
	// miter joints a when v1 ≈ v2 (v0)
	bvec2 is_truncated = bvec2(
	(joinstyle == BEVEL) ? miter.x < 0.99 : miter.x < miter_limit,
	(joinstyle == BEVEL) ? miter.y < 0.99 : miter.y < miter_limit
	);

[devel-chm]

Does joinstyle = :bevel fix this?

No but :round did appear to fix it.
NOTE: I did not see :bevel as one of the documented joinstyles for GLMakie.