Fix incorrect all_of validation when sub-group size is 1

[frasercrmck]

When the sub-group size is 1, the 'one_true' and 'some_true' predicates return true, so we should instead be checking that all_of returns true.

[CLAassistant]

All committers have signed the CLA.

[frasercrmck]

This problem technically also applies to the corresponding work-group tests, but I'm assuming we'll never run those tests with a work-group size of 1.

[bader]

Thanks for fixing this!

I'm assuming we'll never run those tests with a work-group size of 1.

Probably it's worth adding an assertion for that + a comment which of existing checks not work for a work-group size of 1.

[keryell]

Good catch!
Could you add a comment in front of all your changes explaining the problem?

[frasercrmck]

Thanks for comments, will address later in the day.

I have another question, though. I was looking at this again with your feedback in mind, but is it valid for the predicate_function_of_sub_group test to be using get_global_linear_id as it does? Should we not be using the sub-group local linear ID ?

I ask because I thought it was invalid to make assumptions about the mapping of work-items to sub-group items. The predicate_function_of_sub_group test seems to rely on the first sub-group with ID 0 having all of the work-item IDs from 0 to (sub-group-size)-1. The bool_function_of_sub_group test doesn't do this, as it runs on all work-items of all sub-groups and uses the sub-group linear ID.

But also (noticed while writing this comment) is the bool_function_of_sub_group not overwriting its results for every sub-group? It seems to run over all sub-groups but only write to a fixed set of indices. Does that mean that some sub-groups might report incorrect results but get ignored?

[frasercrmck]

I've updated with the comments requested.

[keryell]

@frasercrmck do you plan to dive into your discoveries in your last comment?

[frasercrmck]

@frasercrmck do you plan to dive into your discoveries in your last comment?

If you agree they're real issues, yeah I can certainly take a look. I would prefer to keep that to a separate task/PR though if that's alright; I might need some guidance and a bit more back-and-forth as I'm not actually very that used to writing SYCL. Is that okay?

[keryell]

@frasercrmck do you plan to dive into your discoveries in your last comment?

If you agree they're real issues, yeah I can certainly take a look. I would prefer to keep that to a separate task/PR though if that's alright; I might need some guidance and a bit more back-and-forth as I'm not actually very that used to writing SYCL. Is that okay?

sub-groups are very confusing to me. After looking at the SYCL specification, I have the feeling that the problem you expose is real.
T local_var(item.get_global_linear_id() + 1); should be T local_var = sub_group.get_local_linear_id()
and so on.