Release GIL around blocking operations in libtensor

[oleksandr-pavlyk]

Copy from NumPy ndarray to usm_ndarray is blocking, so release GIL.

mask-positions (cumulative value) which returns a total is blocking, so release the GIL.

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• If this PR is a work in progress, are you opening the PR as a draft?

[github-actions]

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

[github-actions]

Array API standard conformance tests for dpctl=0.18.0dev0=py310ha798474_160 ran successfully.
Passed: 894
Failed: 1
Skipped: 119

[coveralls]

coverage: 87.609%. remained the same
when pulling 69aa1ed on fix-for-deadlocks
into ef9ebc8 on master.

[github-actions]

Array API standard conformance tests for dpctl=0.18.0dev0=py310ha798474_161 ran successfully.
Passed: 895
Failed: 0
Skipped: 119

[antonwolfy]

The dpnp tests are passed with that changes. The example also work now without any hang.
Thank you @oleksandr-pavlyk

[oleksandr-pavlyk]

The last commit fixed a crash in the following snippet:

import dpctl
import dpctl.utils as du
import dpctl.tensor as dpt

class A:
    def __init__(self, arr):
        self._array_obj = arr

    @property
    def __sycl_usm_array_interface__(self):
        return self._array_obj.__sycl_usm_array_interface__

a = dpt.linspace(0, dpt.pi, num=2**27, dtype="f8")

a_dummy = A(a)

a_sin = dpt.sin(a)
print(1)
b = dpt.asarray(a_dummy, dtype="f8")
b /= 1
print(2)

Technically such a code is a usage error, since script ends with pending asynchronous tasks. It used to crash due to such tasks requiring GIL which can not be acquired during the interpreter shutdown and caused seg-fault in the thread executing the cleanup host_task which attempted to acquire GIL.

The fix was to used weakref.finalize to perform synchronization of the dpctl.utils.SequentialOrderManager at shutdown.

The test is not being added here, but we must document that user is advised to perform the synchronization explicitly before the script ends, and the recommended usage would be

import dpctl
import dpctl.utils as du
import dpctl.tensor as dpt

class A:
    def __init__(self, arr):
        self._array_obj = arr

    @property
    def __sycl_usm_array_interface__(self):
        return self._array_obj.__sycl_usm_array_interface__

a = dpt.linspace(0, dpt.pi, num=2**27, dtype="f8")

a_dummy = A(a)

a_sin = dpt.sin(a)
print(1)
b = dpt.asarray(a_dummy, dtype="f8")
b /= 1
print(2)

# ensure all submitted tasks complete
b.sycl_queue.wait()

[github-actions]

Array API standard conformance tests for dpctl=0.18.0dev0=py310ha798474_172 ran successfully.
Passed: 894
Failed: 1
Skipped: 119

[github-actions]

Array API standard conformance tests for dpctl=0.18.0dev0=py310ha798474_172 ran successfully.
Passed: 894
Failed: 1
Skipped: 119

[github-actions]

Array API standard conformance tests for dpctl=0.18.0dev0=py310ha798474_179 ran successfully.
Passed: 895
Failed: 0
Skipped: 119

[github-actions]

Array API standard conformance tests for dpctl=0.18.0dev0=py310ha798474_179 ran successfully.
Passed: 895
Failed: 0
Skipped: 119

[github-actions]

Array API standard conformance tests for dpctl=0.18.0dev0=py310ha798474_180 ran successfully.
Passed: 894
Failed: 1
Skipped: 119

[ndgrigorian]

With tests added and CI passing this LGTM!