clustersize is clarified subgroup_rotate, but not for subgroup_reduce operations

[karolherbst]

I wonder if this is intentional.

For sub_group_clustered_rotate the CLC spec states:

clustersize must be an integer constant expression and a power of two, smaller than or equal to the maximum sub-group size, otherwise the behavior is undefined.

but not for any other subgroup operation with a clustersize parameter.

[karolherbst]

Especially given the SPIR-V side has the same constraints for OpGroupNonUniform* operations

[bashbaug]

I think the same restriction is documented for the clustered reductions, it's just in a different place. See the header paragraph here:

https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#sub-group-clustered-reduction-functions

Behavior is undefined if the specified cluster size is not an integer constant expression, is not a power-of-two, or is greater than the maximum size of a sub-group within the dispatch.

The restriction for sub_group_clustered_rotate is in the description of the function, though.

https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#sub-group-rotate-functions

I agree this is confusing. Should we document the restriction in description of the functions in all cases? Editorially, the table cell descriptions have plenty of room, so this wouldn't significantly increase the length of the spec.

[karolherbst]

Yeah I think it would be better to have written down in each entry even if it's more verbose. At least I'd have spotted it more easily then. I actually tried to search for clustersize but that also didn't made me find it. Anyway, as it is right now it's a bit buried.

[bashbaug]

Have a look at #1535.