From: David Hardy (dhardy_at_ks.uiuc.edu)
Date: Wed Aug 08 2018 - 11:26:50 CDT
NAMD has recently been fixed to internally set bondedCUDA = 219 to calculate dihedrals and crossterms on CPU whenever either accelMDdihe or accelMDdual is set.
The change was made on Aug 2, so if your source build is older than that, you will need to explicitly set bondedCUDA when using aMD dihedral boost.
-- David J. Hardy, Ph.D. Beckman Institute University of Illinois at Urbana-Champaign 405 N. Mathews Ave., Urbana, IL 61801 dhardy_at_ks.uiuc.edu, http://www.ks.uiuc.edu/~dhardy/ > On Aug 7, 2018, at 11:43 AM, Alexander Adams <xadams_at_umich.edu> wrote: > > Dear NAMD developers, > > I was wondering if this aMD dihedral boost issue has been fixed or if we should still use the bondedCUDA workaround. > > Thanks in advance for your time, > Alex Adams > > University of Michigan, Ann Arbor > Ph.D. Candidate - Mayes, Glotzer Labs > College of Engineering - Chemical Engineering > > On Thu, May 17, 2018 at 11:46 AM, David Hardy <dhardy_at_ks.uiuc.edu <mailto:dhardy_at_ks.uiuc.edu>> wrote: > Dear Haochuan, > > Thanks for bringing this to our attention. > > Until we can get the CUDA kernel for dihedrals working correctly for accelerated MD, it is possible to disable just the dihedral part of the calculation. bondedCUDA is actually treated as a bit field that can enable (or disable) the individual term types: > 1 = bonds > 2 = angles > 4 = dihedrals > 8 = impropers > 16 = exclusions > 32 = crossterms > > You should be able to do your calculation correctly by setting bondedCUDA = 251 (i.e. 255-4) to disable just the dihedral calculation on the GPU but use the GPU for everything else. > > Apologies for the half-baked interface, which was never really intended to be exposed for normal NAMD use. > > Best regards, > Dave > > -- > David J. Hardy, Ph.D. > Beckman Institute > University of Illinois at Urbana-Champaign <https://maps.google.com/?q=Champaign+405+N.+Mathews+Ave.,+Urbana,+IL+61801&entry=gmail&source=g> > 405 N. Mathews Ave., Urbana, IL 61801 <https://maps.google.com/?q=Champaign+405+N.+Mathews+Ave.,+Urbana,+IL+61801&entry=gmail&source=g> > dhardy_at_ks.uiuc.edu <mailto:dhardy_at_ks.uiuc.edu>, http://www.ks.uiuc.edu/~dhardy/ <http://www.ks.uiuc.edu/~dhardy/> > >> On May 15, 2018, at 8:59 PM, yjcoshc <yjcoshc_at_gmail.com <mailto:yjcoshc_at_gmail.com>> wrote: >> >> Dear NAMD developers, >> >> Recently I have found the dihedral boost potential of accelerated MD (amd) doesn't affect the real distribution of dihedral angles. After some debugging the issue disappears if I add "bondedCUDA 0" to the NAMD configuration file. I have a look at the git commit e3e849639982dcba9a60c6072c81fdb7cba69759, which offloads the bonded terms to GPU but misses computing the aMD dihedral force. >> >> Thanks, >> >> Haochuan Chen >> > >
This archive was generated by hypermail 2.1.6 : Tue Dec 31 2019 - 23:20:10 CST