Yup, Matt is working with a student on exactly that, though in the context of pipelined GMRES (which might be more important, the CG recurrence is more sensitive to numerical stability).<div class="gmail_extra"><br><br><div class="gmail_quote">
On Wed, Dec 5, 2012 at 9:33 AM, Karl Rupp <span dir="ltr"><<a href="mailto:rupp@mcs.anl.gov" target="_blank">rupp@mcs.anl.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Hi Jed,<br>
<br>
great, thanks! I think these versions might also work in a multi-GPU setting, where global synchronization is even more expensive. I should be able to give that a try soon :-)<br>
<br>
Best regards,<br>
Karli<div class="HOEnZb"><div class="h5"><br>
<br>
<br>
<br>
On 12/05/2012 11:08 AM, Jed Brown wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
I just merged several pipelined conjugate gradient methods from Pieter<br>
Ghysels. If you are interested, here's the paper.<br>
<br>
<a href="http://www.exascience.com/wp-content/uploads/2012/12/Ghysels_12.2012.1.pdf" target="_blank">http://www.exascience.com/wp-<u></u>content/uploads/2012/12/<u></u>Ghysels_12.2012.1.pdf</a><br>
<br>
Access the methods with -ksp_type <pipecg,groppcg,pipecr><br>
</blockquote>
<br>
</div></div></blockquote></div><br></div>