I found this issue addressing a review comment in #84.

Currently, in the SPIR-V environment specification section 7.2.8, when describing behavior when the extensions cl_khr_in64_base_atomics and cl_khr_int64_extended_atomics are supported, there is the line:

When the WorkgroupMemory Memory Semantic is used the Scope must be Workgroup.

However, we don't include any similar validation rule for non-64-bit atomics. Should we?

Regardless, I think this should be described as a common validation rule, and not tied specifically to the 64-bit atomics extensions.

For reference, the OpenCL C spec says:

Atomic operations to local memory only guarantee memory ordering in the work-group not across work-groups and therefore ignore the memory_scope value.