Enabling sycl_queue for USMNdArray in dpnp overloads #1007
Conversation
chudur-budur
force-pushed
the
fix/usmndarray-enable-sycl-queue
branch
3 times, most recently
from
0f037ea to
9c0900a
Compare
| ) | ||
|
|
||
| if queue is not None: | ||
| if queue and not isinstance(queue, types.misc.Omitted): |
There was a problem hiding this comment.
If we have a non-None queue argument can it be anything but a dpctl.SyclQueue? Where can we end up with the queue being a types.misc.Omitted?
There was a problem hiding this comment.
If we have a non-None queue argument can it be anything but a
dpctl.SyclQueue?
Then we fail on these tests numba_dpex/tests/dpjit_tests/parfors/test_builtin_ops.py::test_built_in_operators2
Where can we end up with the queue being a
types.misc.Omitted?
The type of queue becomes types.misc.Omitted when dpnp.epmty() is being called by dpnp functions internally.
numba_dpex/dpnp_iface/arrayobj.py
Outdated
| """ | ||
| if queue and not isinstance(queue, types.misc.Omitted): | ||
|
|
||
| if sycl_queue and not isinstance(sycl_queue, types.misc.Omitted) and device: |
There was a problem hiding this comment.
Can be removed as the check is already inside the UsmNdArray constructor.
There was a problem hiding this comment.
Can be removed as the check is already inside the
UsmNdArrayconstructor.
We can actually remove these checks in all ol_dpnp_<functions>() too, we are doing this same check on three levels, which we shouldn't do.
chudur-budur
force-pushed
the
fix/usmndarray-enable-sycl-queue
branch
from
dadf286 to
f5d155e
Compare
diptorupd
force-pushed
the
fix/usmndarray-enable-sycl-queue
branch
from
f5d155e to
1802d7c
Compare
Remove device='gpu' in tests/types/USMNdArray/test_array_creation_errors.py
chudur-budur
force-pushed
the
fix/usmndarray-enable-sycl-queue
branch
from
1802d7c to
4ac7a87
Compare
|
Superceded by #1022 |
2 participants
This is PR enables
dpctl.SyclQueueforUSMNdArrayindpnpoverloads.This PR does not pass SYCL queue to the functions in
numba_depx/core/runtime/_dpexrt_python.c(yet).Since this PR, we won't have
"unknown"as a default device string anymore, we will useNoneinstead, when there is no need to specify the device type.