cgtblas: sycl backend does not handle nullptr case

[bd4]

cgtblas: sycl backend does not handle nullptr case

[germasch]

So I think there's been some subtlety in the type used for sycl, anyway, in that a reference type (queue&) has been used, if I even understood things correctly. It seems to me that sycl could be made to behave more cuda/hip like by having some

struct sycl_stream_t
{
   sycl::queue *q;
}

And that could hopefully be used to eliminate those subtle differences at the interface level?

[bd4]

The queue is the gateway to all operations in SYCL, so it can't be null - something needs to intercept the attempt to set a null queue and replace it with the default for that device, or the interface needs to change so setting a null "stream" is not alowed, and resetting to the default is done another way.

[germasch]

Right, the queue can't be null, but a pointer to it can be. So if that's what we're passing from Fortran, we can still use C_NULL_PTR. Currently, we directly use queue&, which never can be officially NULL (though in practice if one passes is from Fortran, with Fortran thinking it's a pointer, it will be null. On the C side, one basically has to go, somewhere,

if (queue_pointer) return *queue; else return default queue;