[libclc] Add v3 variants of async_work_group_copy/async_work_group_strided_copy/prefetch #137932
Conversation
…rided_copy/prefetch 3-component vector type is supported for them per OpenCL spec.
|
@frasercrmck please help to review, thanks |
frasercrmck
left a comment
frasercrmck
left a comment
There was a problem hiding this comment.
LGTM.
I see there's a note in the spec: async_work_group_copy and async_work_group_strided_copy for 3-component vector types behave as async_work_group_copy and async_work_group_strided_copy respectively for 4-component vector types. . I'm not sure what that really means for our implementation which does a loop and a store.
I also wonder why we have async/gentype.inc. With this change, couldn't we just use float/gentype.inc and integer/gentype.inc in succession?
OpenCL spec requires that 3-component vector type data has the same size as 4-component vector type data in memory layout.
done, thanks for the suggestions. Deleted |
frasercrmck
left a comment
frasercrmck
left a comment
There was a problem hiding this comment.
Thanks. We could probably come up with some kind of combined gentype.inc that does them both at once (though both currently undef __CLC_BODY after they finish so we'd need to be able to stop that, or preserve __CLC_BODY across gentypes).
Come to think of it, we do a lot of unnecessary undef __CLC_BODY around the codebase. I will try to clean that up in a separate PR.
good idea. |
…rided_copy/prefetch (llvm#137932) 3-component vector type is supported for them per OpenCL spec.
…rided_copy/prefetch (llvm#137932) 3-component vector type is supported for them per OpenCL spec.
2 participants
3-component vector type is supported for them per OpenCL spec.