Broken model transformations in GLMakie and WGLMakie but works in CairoMakie

[mashu]

Hi,

Consider following example

using Makie.KernelDensity

x = randn(1000)
k = kde(x, npoints = 200)

f = Figure()
ax = Axis3(f[1, 1], limits = (-4, 4, -4, 4, 0, 1))
p = poly!(ax, Point2f.(k.x, k.density),
    model = Makie.Mat4f([   1 0 0 0
                            0 0 1 4
                            0 1 0 0
                            0 0 0 1
                        ])
    )
p = poly!(ax, Point2f.(k.x, k.density),
    model = Makie.Mat4f([   0 0 1 4
                            1 0 0 0
                            0 1 0 0
                            0 0 0 1
                        ])
    )
f

Picture shows up correctly in the CairoMakie backend, but is broken in GLMakie/WGLMakie showing distribution unaffected in the middle of the plot, shown in second picture

[jkrumbiegel]

I assume the reason this happens is that the model transform of parent plots is not combined with the children transforms in GLMakie and WGLMakie.

[ffreyer]

This happens because model isn't passed down in poly. CairoMakie uses poly directly, WGLMakie and GLMakie look at mesh (and lines).

julia> p.model[]
4×4 StaticArrays.SMatrix{4, 4, Float32, 16} with indices SOneTo(4)×SOneTo(4):
 0.0  0.0  1.0  4.0
 1.0  0.0  0.0  0.0
 0.0  1.0  0.0  0.0
 0.0  0.0  0.0  1.0

julia> p.plots[1].model[]
4×4 StaticArrays.SMatrix{4, 4, Float32, 16} with indices SOneTo(4)×SOneTo(4):
 1.0  0.0  0.0  0.0
 0.0  1.0  0.0  0.0
 0.0  0.0  1.0  0.0
 0.0  0.0  0.0  1.0

[jkrumbiegel]

Hm yes but do you think all child plots should have completely separate model matrices? It seems reasonable to me that one might put a child element with a special model matrix into a parent object, and then wants to give the parent a different model matrix, which should leave the relative look of the child vs the parent intact.

[ffreyer]

That's what happens with plot.transformation.model. plot.attributes.model acts as an overwrite atm, which I think is probably good. With transformation you can only affect translation, scale and rotation, while the model matrix could also apply shearing.

The better solution to the above is probably using rotate!(plot, axis, angle) + translate!(plot, pos) to keep this transformation inheritance stuff. That should also not have issues between different backends. (Though it does get annoying when different coordinate spaces get mixed)

[jkrumbiegel]

Ok I probably just didn't understand the system well, then.

Can confirm that using this interface, I could make it work with GLMakie:

using Makie.KernelDensity

x = randn(1000)
k = kde(x, npoints = 200)

f = Figure()
ax = Axis3(f[1, 1], limits = (-4, 4, -4, 4, 0, 1))
p = poly!(ax, Point2f.(k.x, k.density))
translate!(p, Point3f(0, 4, 0))
rotate!(p, Point3f(1, 0, 0), pi/2)
p = poly!(ax, Point2f.(k.x, k.density))
rotate!(p, Point3f(0, 0, 1), pi/2)
rotate!(Accum, p, Point3f(1, 0, 0), pi/2)
translate!(p, Point3f(4, 0, 0))
f

[SimonDanisch]

Can we actually close this?

[ffreyer]

I think the question here is: Should the model attribute be passed to children?

If yes we may need to add model = :automatic as a default attribute so that it can be replaced with transformation.model if it's not set and so that we don't get this issue when model is set later.

If no, maybe we should stop advertising it in the plot example and instead advertise transformations.

[ffreyer]

CairoMakie behaves like GLMakie and WGLMakie here now.

After #3246 the rescaling Axis3 does has also moved to transformation.model, which makes overwriting the model matrix through plot attributes more error-prone.