[yt-dev] Problems with periodicity in 2.x and 3.0

[Matthew Turk]

Hi all,

(This message is particularly relevant for Doug, Stephen and Nathan.)

I've been tracking down some failing tests.  The script I have used is here:

http://paste.yt-project.org/show/3266/

Here are the four images:

v2 Density: http://i.imgur.com/mdmJqw6.png
v2 invrad: http://i.imgur.com/XmBmHZw.png

v3 Density: http://i.imgur.com/INlyBdr.png
v3 invrad: http://i.imgur.com/qu3tFpc.png

They're both wrong.  But, in related ways.  Note that there's only
*one* grid in the fake_random_pf, so the grids may also be getting
selected wrong, but that's being masked here by the other problem.

v2 => The wrong cells are being selected.  This is related to the v3
problem, as v2 uses the RadiusCode field to select cells that are
within the data object.

v3 => The correct cells are being selected, but the problem with
RadiusCode persists.  Fortunately, we have a different cell-selection
method.  (I also think we may have too loose a restriction on cell
selection in v3, but that's another topic; once we can sort out this
script we can bring that up.)

So I think the ultimate problem here is RadiusCode, and underneath
that, get_radius.  I know I provided comments on this, but I am now
having a hard time figuring out how to fix it.  Here's the source:

def get_radius(data, field_prefix):
    center = data.get_field_parameter("center")
    DW = data.pf.domain_right_edge - data.pf.domain_left_edge
    radius = np.zeros(data[field_prefix+"x"].shape, dtype='float64')
    r = radius.copy()
    if any(data.pf.periodicity):
        rdw = radius.copy()
    for i, ax in enumerate('xyz'):
        np.subtract(data["%s%s" % (field_prefix, ax)], center[i], r)
        if data.pf.periodicity[i] == True:
            np.subtract(DW[i], r, rdw)
            np.abs(r, r)
            np.minimum(r, rdw, r)
        np.power(r, 2.0, r)
        np.add(radius, r, radius)
    np.sqrt(radius, radius)
    return radius

The periodicity seems to be the issue here.  As a note, this actually
does work with center on 0, 0, 0, which led me to realize it's an
asymmetric problem of wrapping-right, and not wrapping-left.  I
*believe* this can be fixed by changing inside the periodicity check
to be:

        if data.pf.periodicity[i] == True:
            np.abs(r, r)
            np.subtract(r, DW[i], rdw)
            np.abs(rdw, rdw)
            np.minimum(r, rdw, r)

I have issued a pull request for this, but I felt it appropriate to
email the list because I would like to make sure this issue is raised
for a somewhat larger audience than those who get the PR
notifications.

https://bitbucket.org/yt_analysis/yt/pull-request/465/get_radius-periodicity-fix

-Matt