Andy Porter
Andy Porter
Probably we are agreeing but just to be clear - it should be possible to parallelise (the loop over cells containing) a kernel but the metadata should contain enough information...
This will have to wait for #2461 to go on first as it's already 'ready to be merged'.
This is now ready to go once the CI completes.
Thanks @sergisiso, I like this proposal a lot. Presumably, if we were dealing with a named invoke (i.e. a `call invoke(name="big-science", kernel1(x,y))`) then that name would still be available in...
With the switch to the PSyIR backend has this issue gone away?
We had this problem with the PSyclone repo too - we had to downgrade the version of Ubuntu used for the runner in order to retain 3.7 support.
(Link-check failures appear to be because the NVIDIA website is a bit sluggish at the moment.)
Integration tests were all green.
Even simpler, you can extend the script section that records the names of various functions and then we refuse to put them in regions.
> ...as suggested over lunch last week :) I just didn't have the time to get to it, thanks for the reminder! Unless you are offering to take this over?...