adding support for array shape () passed as an index argument in advanced indexing

[ipdemes]

This PR making following work:

    arr[offsets[3]]=-1

[manopapad]

This has different behavior from NumPy on get_item, because we are always creating a dim-1 output.

This code prints out [1]:

arr = 4 - cn.arange(5)
offsets = cn.arange(5)
offset = offsets[3]
print(arr[offset])

while NumPy returns 1:

arr = 4 - np.arange(5)
offsets = np.arange(5)
offset = offsets[3]
print(arr[offset])

I naively just tried removing the changes on deferred.py (and keeping just the C++ changes), but that triggers nv-legate/legate.core#314.

[ipdemes]

@manopapad : Thank you for finding this! I will wait for Wonchan to update on #314

[ipdemes]

@manopapad I have added fixes for few other cases when either lhs or rhs are futures

[ipdemes]

@manopapad @magnatelee : do you think we can still merge this to 22.07 since if fixes some corner cases of advanced indexing? Or should I change this PR to be merged to 22.10?

[manopapad]

I added some further tests involving set/getitem on scalar arrays. It looks like there are issues when the index or value array is 0d and backed by a region-field:

arr = cn.full((5,), 42)
idx = cn.zeros((3,), dtype=np.int64)[2:3].reshape(())
arr[idx]

This example asserts inside of Legion:

[0 - 70000c362000]    2.048913 {5}{runtime}: [error 164] LEGION ERROR: Dynamic type mismatch in 'get_index_space_domain' (from file /Users/mpapadakis/legate.core/legion/runtime/legion/region_tree.inl:2843)

This might require changes in the core.

[ipdemes]

@manopapad I have reproduced the error and it seems it might take some time to fix it. Should we create a separate GitHub issue and PR for this use case?

[manopapad]

Yes, let me update the testcase to skip the problematic cases, and I'll open a new issue about this.

[ipdemes]

@manopapad : do you think we can merge it when CI pass?

[manopapad]

@ipdemes I rebased the PR onto branch-22.10, and reformatted some testcases following guidance from @bryevdv. Let's verify that these pass before merging. Note that you will now need the new pytest-lazy-fixture package to run these tests.