[BUG] Fix off-axis rendering when center is not [0.5, 0.5, 0.5], fix periodicity

[cphyc]

PR Summary

The off-axis renderer introduced in #4741 has an issue when the center and the center in rotated frame do not coincide (which happens when it isn't 0.5, 0.5, 0.5 afaik).

In addition, it wraps around the boundaries of the width of the plotting window rather than the simulation domain (if required).

This PR fixes both.

PR Checklist

• Adds a test for any bugs fixed. Adds tests for new features.

Bug reproduction:

import numpy as np
import yt

ds = yt.load_sample("output_00080")
ad = ds.all_data()

center = ds.arr(ad.argmax("density"))

sp = ds.sphere(center, (200, "kpc"))

n = np.array([1, 1, 1.])
n /= np.linalg.norm(n)
p = yt.ProjectionPlot(ds, n, "density", center=center, data_source=sp, width=(400, "kpc"), weight_field=("gas", "density"))
p.set_zlim(("gas", "density"), 1e-4, 5)
p.save("frames/")

Before	With this PR

[cphyc]

@yt-fido test this please

[chrishavlin]

LGTM now assuming tests are green again.

[chrishavlin]

well hmmm....

FAILED yt/visualization/tests/test_offaxisprojection.py::test_off_axis_octree - AssertionError: assert unyt_quantity(0.38448323, '(dimensionless)') < 0.2

Maybe that check I made you add is too restrictive?

[chrishavlin]

So part of the trickiness is that the test is highly variable between runs: I just ran the code from test_off_axis_octree locally, recording the maximum of the absolute value in the difference and over 10 runs, the max abs fractional difference was [0.17469636745540013, 0.22734252938752475, 0.5588387477439833, 0.06742315690744759, 0.19014792348598422, 0.32429411325319835, 0.24292832837751724, 0.20697617456665565, 0.3351829388729374, 0.09990611014962943]. Setting a seed would help make it deterministic... but the large variation might be indicative of some other issue? Or maybe it's expected and you should just revert the comparison of the absolute value in difference?

I did check what the diffs look like, the large differences are always at cell edges. e.g., for one case:

for which the standard axis aligned ProjectionPlot with annotated cells looks like:

[cphyc]

OK - I've actually uncovered a funny one.

import numpy as np
import yt
import matplotlib.pyplot as plt
ds = yt.load("output_00080")
p1 = yt.ProjectionPlot(
    ds,
    "x",
    ("gas", "density"),
    center=[0.5] * 3,
    width=.8,
    #weight_field=("gas", "density"),
)
p2 = yt.OffAxisProjectionPlot(
    ds,
    [1, 0, 0],
    ("gas", "density"),
    center=[0.5] * 3,
    width=.8,
    #weight_field=("gas", "density"),
)

# Note: due to our implementation, the off-axis projection will have a
# slightly blurred cell edges so we can't do an exact comparison
v1, v2 = p1.frb["gas", "density"], p2.frb["gas", "density"]

fig = plt.figure()
plt.imshow(v1, origin="lower", norm=plt.matplotlib.colors.LogNorm())
plt.colorbar()
fig.savefig("/tmp/1.png")

fig = plt.figure()
plt.imshow(v2, origin="lower", norm=plt.matplotlib.colors.LogNorm())
plt.colorbar()
fig.savefig("/tmp/2.png")

fig = plt.figure()
plt.imshow(v1 / v2, origin="lower")
plt.colorbar()
fig.savefig("/tmp/3.png")
print(np.median(v1 / v2), np.mean(v1 / v2), np.std(v1 / v2))

This returns:

Projection	Off-axis projection	Difference

The funny thing is that the ratio between the on-axis and the off-axis projection is equal (within some uncertainties) to plot_width * 0.5675 (above, plot_width = 0.8). This is the case for a RAMSES dataset (here), but the same can be obtained using fake_random_ds instead, yielding the same value.

As much as I see how I could have missed a plot_width factor, what the heck is 0.5675?

[chrishavlin]

ya, I'm thoroughly confused now. that is a weird number to be off by... and looking at projections of index fields like ('index', 'ones') or ('index', 'y'), the cell edges pop out very clearly and the projected values are all larger than they should be by a pretty big fraction, even in cell interiors (example notebook).

[cphyc]

@chrishavlin I think fixing this bug above is definitely important, but may be part of another PR. Would you be fine with the following plan?

1. Roll back the test on the absolute value of the difference, since this is hard to get stable,
2. Create an issue with milestone 4.4.1 describing the mismatch when using a projection with no weight,
3. Have this PR merged as is?

Regarding point 1, it think it is possible to correct the cell-edges effects, but I don't have the time to think carefully about it. If you are adamant this should be done, I'm happy to open another issue to keep track of it.

[chrishavlin]

@cphyc sounds like a good plan to me!

[cphyc]

@chrishavlin, I have opened the issue and rebased this PR.