Interop with Dask

[saulshanabrook]

I am opening this issue to track how xnd/gumath could work with Dask based on talking with @mrocklin.

We can use the dask.array.from_array on an xnd object, to have Dask chunk up that array and execute operations in parallel. The requirements dask has for an array like object are listed here.

We were going back and forth on whether the numpy-ish API should be implemented on the xnd object directly or if we should create a wrapper class, like this, that adds numpy methods.

The most basic requirement is to have shape and dtype attributes on the object. We can forward those from the type attribute:

In [1]: from xnd import xnd

In [3]: x = xnd(list(range(10)))

In [4]: import dask.array as da

In [7]: x.shape = x.type.shape

In [15]: x.dtype = str(x.type.hidden_dtype)

In [17]: d = da.from_array(x, chunks=(5,))

In [19]: d.sum()
Out[19]: dask.array<sum-aggregate, shape=(), dtype=int64, chunksize=()>

In [21]: d.sum().compute()
Out[21]: 45

In [23]: np.exp(d).sum().compute()
Out[23]: 12818.308050524603

In [24]: d[::3]
Out[24]: dask.array<getitem, shape=(4,), dtype=int64, chunksize=(2,)>

In [25]: d[::3].compute()
Out[25]: array([0, 3, 6, 9])

We could also implement __array_ufunc__ so that when we call np.sin on an xnd array, we could get back an xnd array by using the gumath.sin function. This is how cupy implements it.

We can also create a concat and register it with dask.array.core.concatenate_lookup. I don't think a concat gufunc exists yet.

[hameerabbasi]

There is an issue with this: Dask casts to a NumPy array by default in da.from_array unless you pass the asarray=False kwarg. So internally, all calculations are happening as a NumPy array.

[teoliphant]

I think we need to write a dask.xnd that mirrors dask.array but uses xnd. Also, we should look at porting NumPyWren to XNDWren Travis

…

On Mon, Oct 29, 2018, 11:57 AM Hameer Abbasi ***@***.***> wrote: There is an issue with this: Dask casts to a NumPy array by default in da.from_array <https://docs.dask.org/en/latest/array-api.html#dask.array.from_array> unless you pass the asarray=False kwarg. So internally, all calculations are happening as a NumPy array. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#22 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAPjoCM-2LCHAthrNupW5qd_PdV9chUMks5upyVYgaJpZM4VKDdp> .

[skrah]

I've implemented __array_ufunc__ and __array__. Unfortunately dask calls np.dtype(obj), which does not use __array_typestr__.

I'm not sure if __array_typestr__ is deprecated, but it is exactly what is needed here.

@saulshanabrook's example works now if xnd.array.dtype returns a PEP-3118 string, but that is a bit of a waste, because many ndt types cannot be expressed as PEP-3118.

By "works", I mean that d.sum().compute() also returns an xnd.array again.

So it is only __array_typestr__ or a similar protocol that is missing.

[skrah]

I think I can try the __array_interface__, though that seems like a lot of code just for the typestring.

[teoliphant]

Can you try looking at __array_function__ now (your xnd.array should also implement that).

…

-Travis

On Tue, Mar 26, 2019 at 11:08 AM Stefan Krah ***@***.***> wrote: I think I can try the __array_interface__ though that seems like a lot of code just for the typestring. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#22 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAPjoA-cED_S_ggdpZELI8oXmcxQ2ksaks5vakXygaJpZM4VKDdp> .

-- *Travis Oliphant* CEO 512 826 7480 <https://www.quansight.com/>

[skrah]

Yes, I've added __array_function__ now. It seems to work with NumPy.

EDIT: deleted the CuPy example, which was wrong.

[mrocklin]

cc @pentschev for visibility. He has been doing the Dask+CuPy work and might have suggestions for you all

[skrah]

Thanks @mrocklin @pentschev for suggestions. For xnd the problem is (but perhaps I'm doing something silly):

The following results in array(15.0, type='float32') <class 'xnd.array'>, which is what I want:

from xnd import array
import dask.array as da

x = array([1,2,3,4,5], dtype="float32")
d = da.from_array(x, chunks=(5,), asarray=False)
ans = d.sum().compute()
print(ans, type(ans))

This gives 15.0 <class 'numpy.float32'>:

from xnd import array
import dask.array as da

x = array([1,2,3,4,5], dtype="float32")
d = da.from_array(x, chunks=(2,), asarray=False)
ans = d.sum().compute()
print(ans, type(ans))

So the question is essentially if conversion to ndarray is expected somewhere along the way if the chunks are not trivial.

[skrah]

Here are hopefully correct examples from the website. For xnd the problem is the same, the cupy example has a ValueError.

Cupy and dask are the latest versions from PyPI, installed today.

$ cat xnd_numpy.py 
import os
os.environ["NUMPY_EXPERIMENTAL_ARRAY_FUNCTION"] = "1"
import numpy as np
import dask.array as da
import dask
import xnd

x = np.random.random((10000, 1000))
y = xnd.array.from_buffer(x)
print(type(y))

u, s, v = np.linalg.svd(y)
print(type(u))
print(type(s))
print(type(v))
$ 
$ /home/stefan/rel/bin/python3 xnd_numpy.py 
<class 'xnd.array'>
<class 'xnd.array'>
<class 'xnd.array'>
<class 'xnd.array'>

$ cat xnd_dask.py 
import os
os.environ["NUMPY_EXPERIMENTAL_ARRAY_FUNCTION"] = "1"
import numpy as np
import dask.array as da
import dask
import xnd

x = np.random.random((10000, 1000))
y = xnd.array.from_buffer(x)
print(type(y))

dy = da.from_array(y, chunks=(5000, 1000), asarray=False)
u, s, v = np.linalg.svd(dy)
u, s, v = dask.compute(u, s, v)
print(type(u))
print(type(s))
print(type(v))
$ 
$ /home/stefan/rel/bin/python3 xnd_dask.py 
<class 'xnd.array'>
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>

$ cat cupy_dask.py 
import os
os.environ["NUMPY_EXPERIMENTAL_ARRAY_FUNCTION"] = "1"
import numpy as np
import dask.array as da
import dask
import cupy

x = np.random.random((10000, 1000))
y = cupy.array(x)
print(type(y))

dy = da.from_array(y, chunks=(5000, 1000), asarray=False)
u, s, v = np.linalg.svd(dy)
u, s, v = dask.compute(u, s, v)
print(type(u))
print(type(s))
print(type(v))
$ 
$ /home/stefan/rel/bin/python3 cupy_dask.py 
<class 'cupy.core.core.ndarray'>
Traceback (most recent call last):
  File "cupy_dask.py", line 13, in <module>
    u, s, v = np.linalg.svd(dy)
  File "/home/stefan/rel/lib/python3.6/site-packages/numpy/core/overrides.py", line 165, in public_api
    implementation, public_api, relevant_args, args, kwargs)
  File "/home/stefan/rel/lib/python3.6/site-packages/numpy/linalg/linalg.py", line 1592, in svd
    a, wrap = _makearray(a)
  File "/home/stefan/rel/lib/python3.6/site-packages/numpy/linalg/linalg.py", line 117, in _makearray
    new = asarray(a)
  File "/home/stefan/rel/lib/python3.6/site-packages/numpy/core/numeric.py", line 538, in asarray
    return array(a, dtype, copy=False, order=order)
  File "/home/stefan/rel/lib/python3.6/site-packages/dask/array/core.py", line 1002, in __array__
    x = np.array(x)
ValueError: object __array__ method not producing an array

[pentschev]

For __array_function__ to work with CuPy you need at least version 6.0.0b3, which is not in PyPI yet, I think the only way to get that is to install it from source. And that last code on CuPy should work as you are expecting it to once you have CuPy 6.0.0b3 and Dask 1.1.4.

I don't see any issues with your xnd example, the issue must be somewhere in the implementation.

[pentschev]

Ah, sorry, you're also gonna need Dask master. Release 1.1.4 doesn't yet include __array_function__ support. It was only merged a couple of weeks ago.

dask/dask#4567

[skrah]

The cupy example works here now with numpy, dask and cupy master branches. I had tried dask and cupy from source before, but numpy was from PyPI (1.16.2).

[skrah]

@pentschev The issue for xnd.array is that explicit registering of backends is required:

backends.py:8:    concatenate_lookup.register(cupy.ndarray, cupy.concatenate)

If concatenate is not registered, numpy.empty() is called here (the traceback is not an error but an explicit instrumentation of some functions in the example):

  File "/home/stefan/rel/lib/python3.8/site-packages/dask-1.1.4+22.g96c3381-py3.8.egg/dask/array/core.py", line 3520, in concatenate3
    result = np.empty(shape=shape, dtype=dtype(deepfirst(arrays)))

One way for unregistered backends would be to try to call numpy.concatenate directly, which works on xnd.array.

[pentschev]

Ah, right. I think in a post-__array_function__ world, calling numpy.concatenate is the preferred/expected way.

@mrocklin correct me if I'm wrong, but I think once __array_function__ is consolidated, we want to get rid of these lookups, right?

[mrocklin]

Yes, that's probably correct, at least when the user has a modern version of NumPy. We maintain support for NumPy going back a few versions, so we probably won't be able to drop these completely for a little while.

…

On Fri, Mar 29, 2019 at 2:08 PM Peter Andreas Entschev < ***@***.***> wrote: Ah, right. I think in a post-__array_function__ world, calling numpy.concatenate is the preferred/expected way. @mrocklin <https://github.com/mrocklin> correct me if I'm wrong, but I think once __array_function__ is consolidated, we want to get rid of these lookups, right? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#22 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AASszJFvkpRUcXjuxstz3LRcVVvU8pFoks5vboDGgaJpZM4VKDdp> .