[Libclc-dev] [PATCH 1/1] r600: Add fence implementation, rework barrier

[Jeroen Ketema]

On 01 May 2014, at 02:21, Matt Arsenault <Matthew.Arsenault at amd.com> wrote:

> On 04/30/2014 06:12 PM, Jan Vesely wrote:
>> Matt, is no AS argument a problem? OpenCL allows fences only on local
>> and global AS, and since we don't need any for LDS, I thought having
>> implicit global AS was ok.
> 
> Yes. Fencing all address spaces always would be unfortunate. I would be most concerned about not allowing private loads and stores to be moved across fences, which should always be OK, and would help reduce register usage

I think the only proper solution would be to extend all the atomics/fences with an address space argument. This would be needed in any case if there are plan to implement OpenCL 2.0, which roughly does C++11 atomics with address spaces bolted on.

Jeroen