Harmonize device selection and allow SYCL to circumvent it when no GPU is available #5222
Conversation
dalg24
force-pushed
the
harmonize_device_selection
branch
from
1b1a13d
to
0c2eca6
Compare
dalg24
force-pushed
the
harmonize_device_selection
branch
from
0c2eca6
to
f2ffb32
Compare
There was a problem hiding this comment.
Testing current develop, the behavior for initializing SYCL seems as I intended it to be. If you are requesting a specific device it is assumed to be a GPU and we fail if no GPUs are available. This pull request changes that behavior to silently ignore the device id.
I don't necessarily think this change in behavior is a problem; we should just be conscious of it.
I updated the PR to only consider the fallback when no device id was specified |
There was a problem hiding this comment.
Now, this pull request changes the error message when requesting a specific device id when there are no GPUs from
terminate called after throwing an instance of 'std::runtime_error'
what(): Requested GPU with id 0 but only 0 GPU(s) available!
to
Error: no GPU available for execution.
which is fine with me.
#5211 may have broken SYCL running on CPUs
This PR make sure backends do not try to interpret on their own the settings and use directly the GPU that is selected by Core. It allows SYCL to not call
Impl::get_gpuwhen it detects there are none.