Return input argument if no GPUs werde detected in get_gpu

[masterleinad]

Fixes #4621. @jczhang07 Can you please take a look if this works for you?

Essentially, the changes here are only relevant for the SYCL backend. If a specific device is requested (id>=0) we still error out if there are no GPUs. If that is not the case (id<0), we use the default selector to run on a host queue if no GPUs are detected. This (intended) behavior is not changed here but we have use_gpu return the input argument when no GPU is detected (instead of assigning GPU 0 which doesn't exist in that case).

[dalg24]

This makes that code even less readable

[masterleinad]

This makes that code even less readable

What is it specifically that you want me to change? Are you criticizing the intended behavior or do you merely want to change the implementation to be more readable?

In the latter case, we could copy

return sycl::device::get_devices(sycl::info::device_type::gpu).size();

before calling get_gpu in the SYCL case (although this function is then called twice for the regular case of running on a GPU). Would that make you accept the changes?

[jczhang07]

@masterleinad The fix works for me. Thanks.

[masterleinad]

@dalg24 I avoided modifying get_gpu now but moved the logic to the respective SYCL call site.

[masterleinad]

Retest this please.

[jczhang07]

Retest this please.
@masterleinad Retested and it worked. Thanks.

[dalg24]

Retest this please

[jczhang07]

Retest this please

Works here.

[masterleinad]

Retest this please.

[masterleinad]

Works here.

@jczhang07 The comments are not for you but for our CI. 🙂