[gmx-users] single X double precision
szilard.pall at cbr.su.se
Tue Apr 17 20:43:08 CEST 2012
2012/4/17 Pedro Alexandre de Araújo Gomes Lapido Loureiro <palaplou at gmail.com>:
> I've come across a discussion about the "single X double precision" issue in
> a NAMD users list e-mail
> I would like to know what do you think about these 3 specific points:
> 1) "there are some places where single precision (FFT, force/energy
> computation), where single precision can be applied, but in other
> places (summations) where even double precision may result in
> artefacts, due to limited numerical accuracy for large enough
> systems. there are some applications that show little sensitivity
> to single/double precision issues (homogeneous bulk lennard-jones
> systems), but others are notoriously difficult (systems with
> significant potential drops in one or two dimensions, e.g. metal
> surface slabs or lipid bilayers) since you don't have that much
> error cancellation anymore."
> 2) "if you store coordinates in single precision, you have to take
> extra precautions for handling very large systems, since you
> would store coordinates with different relative precision,
> depending on how close you are to the origin. e.g., with domain
> decomposition, you can define per domain offsets, but that
> would work only well in case of a large enough number of domains. "
> 3) "the best way to study the impact of single vs. double would be
> by running tests with the gromacs code. gromacs _can_ be compiled
> in single or double precision. for a proper comparison, you'll have
> to turn off the assembly innerloops though, since they may use single
> precision even in double precision mode."
> Specifically this last point startled me. In fact, is this true?
AFAIK no, the assembly (SSE intrinsic in 4.6) kernels have both single
and double versions. If double precision is requested that's always
respected in the code.
Additionally, single precision in GROMACS is in fact not *pure* single
precision, in some parts of the code where needed double precision is
used (e.g. constraints). Therefore, technically what we call single
corresponds to what some MD packages call mixed precision.
> gmx-users mailing list gmx-users at gromacs.org
> Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
More information about the gromacs.org_gmx-users