[Libclc-dev] [PATCH 2/2] amdgcn--mesa3d: Implement get_num_groups()

[Matt Arsenault via Libclc-dev]

> On Jul 26, 2016, at 08:08, Jan Vesely via Libclc-dev <libclc-dev at lists.llvm.org> wrote:
> 
>>> 
>>> I planned to switch r600 to reading these (and global size and
>>> local
>>> size) from implicitarg ptr. Is there an advantage to gcn reading
>>> this
>>> from dispatch info, rather than implicitarg ptr?
>>> 
>> 
>> These intrinsics read these values from the user SGPRs and not the
>> dispatch info.  This is really the only difference between what Mesa
>> does and what HSA does.
> 
> The questions are mostly out of my curiosity, I can't really comment on
> HSA patches.
> I take it SGPRs are faster?
> If these intrinsics follow HSA ABI shouldn't they be in amdgcn-amdhsa
> directory?
> The first patch introduced amdgcn-mesa3d directory, but if mesa/clover
> is switched to use HSA ABI shouldn't libclc produced for amdgcn-amdhsa
> work for clover as well?
> 
> thanks,
> Jan

These are defined in the HSA ABI, but are never actually used since they are not really implemented in the microcode. They also introduce the possibility of using more user SGPRs than the maximum of 16

-Matt
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/libclc-dev/attachments/20160726/0949cdcc/attachment.html>