Vc at MIC

[Matthias Kretz]

On Thursday 22 August 2013 15:39:45 Kulakov, Igor wrote:
> > Q: The results below are for AVX?
> 
> We use -VC_IMPL=SSE only at the moment.
> I also checked float_v length it is 4 for both Vc-s.
> 
> Also I need to say that the tracker compiled in Release (with optimization)
> crashes with Vc-0.99. The results show where in Debug (without
> optimization). The behavior of the tracker with Vc-0.7 is same for both
> Release and Debug.

The crashes might be a first indicator as to what is going wrong. If it's a 
misaligned load/store or illegal instruction then it should be easy to fix. If 
it's out-of-bounds memory access then probably some index calculation went 
wrong...

> 
> Regards,
> Igor
> ________________________________________
>[please enable javascript to see the address]
>[please enable javascript to see the address]] on behalf of Matthias Kretz
>[please enable javascript to see the address]] Sent: Thursday, August 22, 2013 8:49 AM
>[please enable javascript to see the address]; Kisel, Ivan Prof. Dr.
> Subject: Re: Vc at MIC
> 
> On Wednesday 21 August 2013 21:48:19 Kulakov, Igor wrote:
> > We see significant differences in tracking results with Vc-0.99 and
> > Vc-0.7.2 using ICC-13.1.0. For example number of broken fast tracks
> > (clones for ref set) is factor of 7(!) different. (see numbers below)
> > Meanwhile with VC_IMPL=Scalar both Vc-0.99 and Vc-0.7.2 give same result.
> > 
> > Do you have any suggestions how to proceed?
> 
> Wow, that's bad.
> It's hard to guess where this is coming from so I need you to find the
> guilty function(s) in Vc.
> 
> You should look through the unit tests in detail whether any of the failures
> happen in functions that you use. If this is the case, let me know and I'll
> give priority to those.
> 
> If not, then something breaks that doesn't have a test or where the test
> isn't thorough enough. You'd have to find the place where it breaks and
> deduce the function in Vc that is broken. It might be a miscompilation in
> which case this may turn into a Heisenbug...
> 
> Q: The results below are for AVX?
> 
> Cheers,
>         Matthias
> 
> > Vc-0.99:
> >  ---- Global Performance 10 events Statistic ----
> > 
> > Track category         :  Eff   / Clones | All Reco | All MC
> > Allset    efficiency   : 0.920  / 0.290  | 7119  | 7740
> > Refset    efficiency   : 0.982  / 0.136  | 910  | 927
> > Extra     efficiency   : 0.911  / 0.310  | 6209  | 6813
> > Rest      efficiency   : 0.898  / 0.727  | 79  | 88
> > Ghost     probability  : 0.129 | 917
> > All reco tracks/ev : 711.900
> > Reconstruction Time Real =    767.606 ms, CPU =    767.000 ms,
> > parallelization speedup: 0.998
> > 
> > Vc-0.7.2:
> >  ---- Global Performance 10 events Statistic ----
> > 
> > Track category         :  Eff   / Clones | All Reco | All MC
> > Allset    efficiency   : 0.944  / 0.113  | 7304  | 7740
> > Refset    efficiency   : 0.980  / 0.020  | 908  | 927
> > Extra     efficiency   : 0.939  / 0.126  | 6396  | 6813
> > Rest      efficiency   : 0.909  / 0.773  | 80  | 88
> > Ghost     probability  : 0.127 | 929
> > All reco tracks/ev : 730.400
> > Reconstruction Time Real =    146.283 ms, CPU =    145.000 ms,
> > parallelization speedup: 0.981
-- 
Dipl.-Phys. Matthias Kretz

Web:   http://compeng.uni-frankfurt.de/?mkretz

SIMD easy and portable: http://compeng.uni-frankfurt.de/?vc