Fix touching window

[dachengx]

What is the problem / what does the code in this PR do

In some insane cases, where both endtime of things and container are not sorted, like:

things = np.array([(0, 1, 1, 0),
                   (1, 5, 1, 0),
                   (2, 1, 1, 0)],
                   dtype=strax.interval_dtype)

containers = np.array([(0, 4, 1, 0), (0, 1, 1, 0)],
                       dtype=strax.interval_dtype)

The current algorithm will give wrong results:

[[0, 3], [0, 3]]

The correct answer is:

[[0, 3], [0, 1]]

Can you briefly describe how it works?

The new algorithm only search for interval when the time and endtime of things and containers are sorted. If not, we first sort them and record the argsort, then later recover the order following argsort of input.

When things' endtime are not sorted, we will return the indices of the first and last things which are touching the container. So we return the minimum index in argsort.

Can you give a minimal working example (or illustrate with a figure)?

Please include the following if applicable:

• Update the docstring(s)
• Update the documentation
• Tests to check the (new) code is working as desired.
• Does it solve one of the open issues on github?

Please make sure that all automated tests have passed before asking for a review (you can save the PR as a draft otherwise).

[coveralls]

coverage: 91.553% (-0.06%) from 91.609% when pulling 772cea8 on fix_touching_window into 4f19632 on master.

[WenzDaniel]

Hej @dachengx I am a little bit hesitant in merging your PR as the code gets much more complex. I do not know if we really need to cover every single edge case which is actually forbidden based on the provided doc-string. Did not we just add the consistency check for this very same reason?

In addition, I think one should try to avoid using np.append in a numba function as this makes things slow. At least it did in the past.

[dachengx]

Hej @dachengx I am a little bit hesitant in merging your PR as the code gets much more complex. I do not know if we really need to cover every single edge case which is actually forbidden based on the provided doc-string. Did not we just add the consistency check for this very same reason?

In addition, I think one should try to avoid using np.append in a numba function as this makes things slow. At least it did in the past.

The append is removed. Currently, the doc-string allows non-sorted endtime of things and containers. So I think we should cover all of these cases in the function. I do not concern too much about the speed, the new function is ~100% slower than the original one, tested by %timeit. But the original function is very fast.

[WenzDaniel]

Thanks Dacheng, to me it looks fine. Although I am a little bit concerned performance wise. However, this should now lead to correct results.