You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Associating a ur_queue_handle_t with a ur_context_handle_t is sort of meaningless if we are using a multi device context, as the context doesn't tell us which device in the context the queue should be associated with. As such I think we should not allow the hDevice param to be null.
The text was updated successfully, but these errors were encountered:
hdelan
changed the title
urQueueCreateWithNativeHandle Shouldn't allow hDevice to be null
urQueueCreateWithNativeHandle shouldn't allow hDevice to be null
Sep 2, 2024
If we do change this it should be a relatively trivial change on the UR side, but it will almost certainly break sycl RT's make_queue implementation https://github.com/intel/llvm/blob/sycl/sycl/include/sycl/backend.hpp#L354https://github.com/intel/llvm/blob/sycl/sycl/source/backend.cpp#L116 we had a lot of that kind of issue when we ported to UR and turned on the validation layer: sycl loves to pass nullptr to CreateWithNative type entry points - which is probably why null was allowed here in the first place. To my untrained eye it looks like you'd need to do a "get first device in context" thing on the sycl side to remediate, which is really just shuffling the nastiness up a layer rather than doing away with it.
Associating a
ur_queue_handle_t
with aur_context_handle_t
is sort of meaningless if we are using a multi device context, as the context doesn't tell us which device in the context the queue should be associated with. As such I think we should not allow thehDevice
param to be null.The text was updated successfully, but these errors were encountered: