[group] Refine scan_over_group for sub-group

[Nuullll]

Use global linear id to index input data. Store both the sub-group id and item linear id within the sub-group to recover sub-group construction when verifying, so that we don't make any assumption on the implementation-defined sub-group partitioning and ordering.

[steffenlarsen]

This looks right. It is easy to forget not all the groups have the same size.

[Nuullll]

This looks right. It is easy to forget not all the groups have the same size.

Thanks for reviewing. Please allow me to finish some extra testings.

[Nuullll]

Neither sub_group.get_local_linear_range() nor sub_group.get_max_local_range() is reliable:

1. Not all the sub-groups are guaranteed to have the same size, so we shouldn't use get_local_linear_range().
2. Considering the 2-dimensional NDRange (3, 4) with reqd_sub_group_size=16, then the implementation may construct 3 sub-groups (in the highest dimension) where get_max_local_range() == 16 for all the sub-groups. It's a little counter-intuitive, but as noted in OpenCL C Spec - cl_intel_required_subgroup_size:

The optional attribute((intel_reqd_sub_group_size())) can be used to indicate that the kernel must be compiled and executed with the specified sub-group size. When this attribute is present, get_max_sub_group_size() is guaranteed to return the specified integer value. This is important for the correctness of many sub-group algorithms, and in some cases may be used by the compiler to generate more optimal code.

So using the global linear id directly is probably safer.

[steffenlarsen]

Blocking this to avoid regression.

[steffenlarsen]

Thank you, @Nuullll ! LGTM!

[Nuullll]

Any further comments? Can we merge this?

[keryell]

Great!