[Openframeworkwg] Some concerns about the new Fabric/IBverbs API

Hefty, Sean sean.hefty at intel.com
Tue Dec 10 13:27:50 PST 2013


> > However, if we support optimized calls, such as
> > rc_qp->post_rdma_write_with_immediate(), it would seem that the
> > benefits of inlining would be minimal.  Ideally, there shouldn't be
> > much that the compiler could do.
> 
> Even branch free inlining should save load/store traffic in terms of
> stack frame setup and lost optimization opportunities due to alias
> analysis (restrict helps with this a bit, but...).
> 
> I'd say this is especially true on the send path - I have several
> code cases where the send path is doing a batch of work in a
> loop.
> 
> The rx path, maybe not so much since you generally would rx one thing
> and then process..

I will add it as a requirement to support provider code inlining for highly-optimized use cases.



More information about the ofiwg mailing list