<div class="gmail_quote">On Fri, Nov 11, 2011 at 08:11, <a href="mailto:markus.sons@gmail.com">markus.sons@gmail.com</a> <span dir="ltr"><<a href="mailto:markus.sons@gmail.com">markus.sons@gmail.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">We want to use PETSc to solve this problem in parallel and would have hoped to be able to simultaneously drop the wrapper. I guess a mixed approach would be perfect then?</blockquote>
</div><br><div>If it's performance-sensitive and smaller than dimension 10 or 20, you want to avoid BLAS (and PETSc Mat/Vec). You also don't want to parallelize super small problems; solve them redundantly if necessary. For larger problems, use whatever abstraction you like. MatDense has minimal overhead forwarding into BLAS/Lapack, but not every function is wrapped.</div>