So, not surprisingly, *fixing* this GTI-dependent change, breaks 3 GTI
tests in DPFP suggesting that this is not a valid optimization to me.
Looking at the diffs, I propose updating the reference outputs. Any
objections?
On Tue, Nov 10, 2020 at 12:57 PM Scott Le Grand <varelse2005.gmail.com>
wrote:
> I suspect we could calculate a safe number of steps past a rebuild to skip
> in your turbo mode. But that's for another day. This code needs to go...
>
>
> On Tue, Nov 10, 2020 at 12:56 PM David Cerutti <dscerutti.gmail.com>
> wrote:
>
>> Yep. skin_permit should avoid rebuilds if the test passes the more
>> permissive threshold. And I do actually seem to remember seeing that
>> code,
>> or something to the effect "don't rebuild pair list if we just created it"
>> in the comments as I was working on the turbo mode. So whatever that code
>> is, it is pretty ancient and may affect test cases if we purge it, even
>> though that is probably a good idea.
>>
>> Dave
>>
>>
>> On Tue, Nov 10, 2020 at 3:51 PM Scott Le Grand <varelse2005.gmail.com>
>> wrote:
>>
>> > So looking at skin_permit, it looks like I can blow away this junk and
>> > skin_permit will do the rest?
>> >
>> > On Tue, Nov 10, 2020 at 12:16 PM Scott Le Grand <varelse2005.gmail.com>
>> > wrote:
>> >
>> > > ~line 2000 of kNeighborList.cu:
>> > >
>> > > extern "C" void kNLSkinTest(gpuContext gpu)
>> > > {
>> > > static unsigned counter = 0;
>> > > static unsigned lastCounter = gpu->sim.skinnb * 14;
>> > >
>> > > int counterCut = lastCounter*0.2;
>> > > unsigned interval = 1;
>> > > if (counter++ > counterCut && (counter % interval) == 0) {
>> > >
>> > > etc...
>> > >
>> > > Not only is this uncool, it's done with local static variables which
>> is
>> > > just horrible...
>> > >
>> > > On Tue, Nov 10, 2020 at 12:07 PM David Cerutti <dscerutti.gmail.com>
>> > > wrote:
>> > >
>> > >> I wasn't aware of the code doing anything like NB skin tests only
>> every
>> > >> other step. The "turbo" mode is enabled by setting skin_permit to a
>> > value
>> > >> between 0.5 and 1.0. See here:
>> > >>
>> > >> http://ambermd.org/GPUPerformance.php
>> > >>
>> > >> Where in the code is the call to the pairlist check being skipped on
>> > half
>> > >> the steps?
>> > >>
>> > >> Dave
>> > >>
>> > >>
>> > >> On Tue, Nov 10, 2020 at 2:03 PM Scott Le Grand <
>> varelse2005.gmail.com>
>> > >> wrote:
>> > >>
>> > >> > At some point, PMEMD stopped doing NB skin tests every step and it
>> now
>> > >> does
>> > >> > them every other step. No matter what your view on the validity of
>> > doing
>> > >> > so, doing this under the hood all the time without informing the
>> user
>> > is
>> > >> > uncool to the max.
>> > >> >
>> > >> > I am turning this off by default. Any objections? Dave, how does
>> one
>> > >> enable
>> > >> > your turbo mode? Because this seems cool there, just not on 100% of
>> > the
>> > >> > time with no option to turn it off.
>> > >> >
>> > >> > Scott
>> > >> > _______________________________________________
>> > >> > AMBER-Developers mailing list
>> > >> > AMBER-Developers.ambermd.org
>> > >> > http://lists.ambermd.org/mailman/listinfo/amber-developers
>> > >> >
>> > >> _______________________________________________
>> > >> AMBER-Developers mailing list
>> > >> AMBER-Developers.ambermd.org
>> > >> http://lists.ambermd.org/mailman/listinfo/amber-developers
>> > >>
>> > >
>> > _______________________________________________
>> > AMBER-Developers mailing list
>> > AMBER-Developers.ambermd.org
>> > http://lists.ambermd.org/mailman/listinfo/amber-developers
>> >
>> _______________________________________________
>> AMBER-Developers mailing list
>> AMBER-Developers.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber-developers
>>
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Nov 10 2020 - 15:00:01 PST
