[gmx-users] Running multiple kernels

[Erik Lindahl]

Hi,

On Apr 10, 2007, at 9:11 PM, Justin M. Shorb wrote:

> Greetings,
>
> I have been running a modified version of gromacs 3.3 and would  
> like to run two kernel loops: one which is 330 and one which is  
> 300. In the src/gmxlib/nonbonded/nb_kerneltype.h file it strictly  
> says to use wrapper routines that are included in the  
> gmx_nonbonded.h file, which I cannot find. I assume that in  
> switching versions the filename has changed, and perhaps was meant  
> to point me to nonbonded.c ? Given that nb_kerneltype.h says that  
> it is part of the gromacs 4.0 version, I would not be surprised if  
> a version misnomer occured. However, I would be more than happy if  
> you can point me to where this file should be located.
>
> In a similar light, is it possible to build and call two separate  
> kernels within one run of gromacs, or would the tabulated files end  
> up on top of each other? The nb_kerneltype.h seems to say you can  
> use a build routine (probably setup_kernels() in nonbonded.c, but I  
> want to make sure I'm not missing the gmx_nonbonded.h file somewhere).

Certainly. Each neighborlist is attached to a specific loop, but you  
can have as many neighborlists as you want. Your best option is  
likely to create the (different) neighborlists with the nonbonded  
setups they should have, and then call do_nonbonded() as declared in  
gromacs-x.y.z/include/nonbonded.h (sorry, the gmx_ prefix was lost  
when backported).

>
> My plan currently is to use setup_kernels() to explicitly setup  
> both kernels needed in the nb_kernel_list (which I believe is the  
> list of kernels to be setup ranging from 1 to eNB_NBKERNEL_NR) and  
> then simply force the kernel 300 to be created as well as the  
> generic propagation kernel (typically 330). Then it should be a  
> matter of choosing a different nrnb_ind index to run the alternate  
> kernel in the do_nonbonded() routine.
>

As you might already have found out, setup_kernels() is a really  
trivial routine that sets kernel pointers based primarily on your  
architecture (SSE/SSE2/Altivec/ia64/etc.), and assigns NULL to the  
solvent-optimized kernel pointers if that type of optimization isn't  
present.

You can still call all kernels, although they might be generic  
(though optimized) C code if no acceleration has been implemented.

> Are there double counting things that should be taken into account?  
> And more importantly, is there an efficient subroutine written for  
> such a matter that is being overlooked (hidden in a *.h file  
> somewhere)?

There should be a routine in src/mdlib/ns.c that initalizes the  
kernel index based on the interactions in a particular neighborlist.

Cheers,

Erik


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20070410/8823b89d/attachment.html>