<p>1. I think it would be enough to have a way to run the "level 3" dense matrix kernels in the GPU. That is where it would make the biggest difference. Dense does not reuse the sequential kernels, so I guess it requires some code for both Seq and MPI. My hope was that since the dense API is fairly small and CUBlas is mature that it would be simple. I think fairly few people use Dense, but if it's sufficiently easy to support...</p>
<p>2. Is it as easy as PCSetDM? Or provide a coarse DM and get a hierarchy? Is there an example? I agree about having FieldSplit forward the pieces. I recall starting on that. Who is responsible for assembling rediscretized coarse operators?</p>
<p>2b. Can we decide on an interface for plumbing extra information into the Schur splits? Maybe composing the matrix like in PCLSC is the way and it just needs to be wrapped in a decent interface. Or maybe there is a better way.</p>
<p>Jed</p>
<p><blockquote type="cite">On Dec 19, 2010 2:54 PM, "Barry Smith" <<a href="mailto:bsmith@mcs.anl.gov">bsmith@mcs.anl.gov</a>> wrote:<br><br><p><font color="#500050"><br>On Dec 19, 2010, at 4:47 PM, Jed Brown wrote:<br>
<br>> Decent preallocation for DMComposite.<br>> <br>> CUDA fo...</font></p> What does this mean? Dense sequential matrices? Parallel?<br>
<p><font color="#500050"><br>> <br>> Move linear DMMG into PCMG. (I don't have a sense for how long this will take.)<br></font></p> Except for "grid sequencing" (not sure if I care about grid sequencing for linear problems) this is pretty much done. But we should still do the fieldsplit part of DMMG for linear problems into PCFIELDSPLIT<br>
<p><font color="#500050"><br>> <br>> Grid sequencing in SNES, this might be too much for 3.2.<br></font></p> I think this is too much.<br>
<font color="#888888"><br>
Barry<br>
</font><p><font color="#500050">> <br>> Jed<br>> <br>> <br>>> On Dec 19, 2010 10:05 AM, "Barry Smith" <<a href="mailto:bsmith@mcs.anl.gov">bsmith@mcs.anl.gov</a>> wrote:<br>>> <br>
>> <br>>> W...</font></p></blockquote></p>