Sparse arrays do not support all of Scipy's slicing options

I encountered this behavior in the most recent (CuPy 8.0.0 at the time of writing this) branch when attempting to call dask.array.from_array() with a csr_matrix. It appears that Dask is attempting to do some slicing that is supported in Scipy’s csr_matrix but is not supported by CuPy.

The following is a reproducible example:

>>> import cupy as cp
>>> import dask.array
>>> 
>>> a = cp.sparse.random(100, 100, format='csr')
>>> b = dask.array.from_array(a)
>>> b.compute()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/share/workspace/dask/dask/base.py", line 165, in compute
    (result,) = compute(self, traverse=False, **kwargs)
  File "/share/workspace/dask/dask/base.py", line 436, in compute
    results = schedule(dsk, keys, **kwargs)
  File "/share/workspace/dask/dask/threaded.py", line 81, in get
    **kwargs
  File "/share/workspace/dask/dask/local.py", line 486, in get_async
    raise_exception(exc, tb)
  File "/share/workspace/dask/dask/local.py", line 316, in reraise
    raise exc
  File "/share/workspace/dask/dask/local.py", line 222, in execute_task
    result = _execute_task(task, data)
  File "/share/workspace/dask/dask/core.py", line 119, in _execute_task
    return func(*args2)
  File "/share/workspace/dask/dask/array/core.py", line 104, in getter
    c = a[b]
  File "/share/workspace/cupy/cupyx/scipy/sparse/compressed.py", line 509, in __getitem__
    raise ValueError('unsupported indexing')
ValueError: unsupported indexing

Here’s the same code using Scipy:

>>> import scipy
>>> 
>>> a = scipy.sparse.random(100, 100, format='csr')
>>> b = dask.array.from_array(a)
>>> b.compute()
array(<100x100 sparse matrix of type '<class 'numpy.float64'>'
	with 100 stored elements in Compressed Sparse Row format>, dtype=object)

When using a CuPy coo_matrix:

>>> a = cp.sparse.random(100, 100, format='coo')
>>> b = dask.array.from_array(a)
>>> b.compute()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/share/workspace/dask/dask/base.py", line 165, in compute
    (result,) = compute(self, traverse=False, **kwargs)
  File "/share/workspace/dask/dask/base.py", line 436, in compute
    results = schedule(dsk, keys, **kwargs)
  File "/share/workspace/dask/dask/threaded.py", line 81, in get
    **kwargs
  File "/share/workspace/dask/dask/local.py", line 486, in get_async
    raise_exception(exc, tb)
  File "/share/workspace/dask/dask/local.py", line 316, in reraise
    raise exc
  File "/share/workspace/dask/dask/local.py", line 222, in execute_task
    result = _execute_task(task, data)
  File "/share/workspace/dask/dask/core.py", line 119, in _execute_task
    return func(*args2)
  File "/share/workspace/dask/dask/array/core.py", line 104, in getter
    c = a[b]
TypeError: 'coo_matrix' object is not subscriptable

As expected, the CuPy csc_matrix yields the same exception as CSR:

>>> a = cp.sparse.random(100, 100, format='csc')
>>> b = dask.array.from_array(a)
>>> b.compute()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/share/workspace/dask/dask/base.py", line 165, in compute
    (result,) = compute(self, traverse=False, **kwargs)
  File "/share/workspace/dask/dask/base.py", line 436, in compute
    results = schedule(dsk, keys, **kwargs)
  File "/share/workspace/dask/dask/threaded.py", line 81, in get
    **kwargs
  File "/share/workspace/dask/dask/local.py", line 486, in get_async
    raise_exception(exc, tb)
  File "/share/workspace/dask/dask/local.py", line 316, in reraise
    raise exc
  File "/share/workspace/dask/dask/local.py", line 222, in execute_task
    result = _execute_task(task, data)
  File "/share/workspace/dask/dask/core.py", line 119, in _execute_task
    return func(*args2)
  File "/share/workspace/dask/dask/array/core.py", line 104, in getter
    c = a[b]
  File "/share/workspace/cupy/cupyx/scipy/sparse/compressed.py", line 509, in __getitem__
    raise ValueError('unsupported indexing')
ValueError: unsupported indexing

[cjnolet]

@crumleyc, we are needing this feature in Dask and RAPIDS currently. I think it would be great if you worked on this issue! Ideally, this change will be included in the 8.0.0 release.

As a first step, it will be helpful to familiarize yourself with Cupy’s contribution guide, if you have not done so already.

I tend to create a pull request early, so that I can solicit feedback as I work through changes. An additional benefit to this is that I am able to reference my changes directly when asking for help.

[asi1024]

Current implementation raises a ValueError, but it should raise NotImplementedError. Sorry for being confusing.