With the latest code, mixed precision seems to be working fine, but my DPDP
run fails on startup with the following message:
Error: invalid argument launching kernel kNLBuildNeighborList
This happens on two very different simulations, each of which had run several
hundred nanoseconds with the earlier code. Last thing printed to mdout are
the tests of the CUBIC SPLINE -- it never gets to the first energy printout.
This is with GNU compilers. So, before I dig further, I just wanted to check
with Scott: does the DPDP option work for you with the latest code release?
Second question: although the new code for SPDP is running, it seems to be
about 3 times *slower* than the previous code -- although I am comparing intel
(earlier) to gnu (now). This is an NPT run, if that makes a difference to
what one should expect. For the jac benchmark, on a C2050 board: the previous
timings (since about last November) were 415 sec for nonsetup wall time for
50,000; now I am getting 255 sec., a 1.6x speedup, throughput of 34 nsec/day.
Of course, I have more work to do, but again just a question of whether NPT
might be involved....has the new code been tested (for accuracy and speed)
with NPT?
...thx...dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Thu Feb 10 2011 - 09:30:04 PST
