Cached device store

[oleksandr-pavlyk]

This PR creates dedicate queue cache with context, device pair being a key.

Use of this cache in dpctl.tensor.Device, dpctl.tensor.from_dlpack and in dpctl.memory allows to returns the same queue for a given device, streamlining user experience with compute-follows-data model.

(dev_dpctl) opavlyk@opavlyk-mobl:~/repos/dpctl$ ipython
Python 3.9.12 (main, Jun  1 2022, 11:38:51)
Type 'copyright', 'credits' or 'license' for more information
IPython 8.4.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import dpctl.tensor as dpt

In [2]: dpt.Device.create_device("gpu") == dpt.Device.create_device("gpu")
Out[2]: True

In [3]: x = dpt.linspace(0, 3 + 2j, num=1000)

In [4]: y = dpt.from_dlpack(x)

In [5]: x.sycl_queue == y.sycl_queue
Out[5]: True

In [6]: dpt.Device.create_device("level_zero:gpu") == dpt.Device.create_device("opencl:gpu")
Out[6]: False

In [7]: dpt.Device.create_device("level_zero:gpu") == dpt.Device.create_device("gpu")
Out[7]: True

(dev_dpctl) opavlyk@opavlyk-mobl:~/repos/dpctl$ ipython
Python 3.9.12 (main, Jun  1 2022, 11:38:51)
Type 'copyright', 'credits' or 'license' for more information
IPython 8.4.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import dpctl.tensor as dpt

In [2]: import dpctl.memory as dpm

In [3]: m1 = dpm.MemoryUSMShared(100)

In [4]: m2 = dpm.MemoryUSMShared(100)

In [5]: m1.sycl_queue == m2.sycl_queue
Out[5]: True

In [6]: x = dpt.linspace(0, 3 + 2j, num=1000)

In [7]: y = dpt.from_dlpack(x)

In [8]: x.sycl_queue == y.sycl_queue
Out[8]: True

• Have you provided a meaningful PR description?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?

[github-actions]

View rendered docs @ https://intelpython.github.io/dpctl/pulls/1076/index.html

[coveralls]

Coverage: 82.296% (+0.03%) from 82.269% when pulling 0ed6e67 on cached-device-store into 0a8abd8 on master.

[github-actions]

Array API standard conformance tests for dpctl=0.14.1dev2=py310h41425db_21 ran successfully.
Passed: 33
Failed: 801
Skipped: 280

[github-actions]

Array API standard conformance tests for dpctl=0.14.1dev2=py310h41425db_29 ran successfully.
Passed: 33
Failed: 801
Skipped: 280

[antonwolfy]

I've tested together with dpnp and it works fine. Thank you! LGTM!

[vlad-perevezentsev]

Great job! Thank you @oleksandr-pavlyk

[github-actions]

Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞

[diptorupd]

@oleksandr-pavlyk the API should also allow passing in a filter string otherwise using this function from inside Numba will be challenging as we will need full typing support for SyclDevice in Numba/numba-dpex.

I do not expect numba-dpex to use this functionality. usm_ndarray ctors do permit using filter selector strings. Or am I missing the point?

[github-actions]

Array API standard conformance tests for dpctl=0.14.1dev2=py310h41425db_31 ran successfully.
Passed: 33
Failed: 801
Skipped: 280

[diptorupd]

This is the use case I had in mind:

@numba_dpex.dpjit
def foo():
    a = dpnp.ones(1024, device="gpu")
    b = dpnp.ones(1024, device="gpu")
    return a + b

To support these constructors inside the dpjit function, we will need to cache the queue. Will it be possible to use the cache in dpctl for that purpose?

[oleksandr-pavlyk]

@diptorupd Yes, and I can add support for string argument for sure, but without boxing/unboxing support for dpctl.SyclQueue object it would not be possible to support arrays allocated on sub-devices.

[diptorupd]

without boxing/unboxing support for dpctl.SyclQueue object it would not be possible to support arrays allocated on sub-devices.

yes, that is on my next set of todos