Skip to content

Refactor the FieldCache::getB in order to enable multiversioning for the vectorized part

Christos Anastopoulos requested to merge ATLAS-EGamma/athena:getB_FMV_AVX2 into main

Refactor the FieldCache::getB in order to enable multiversioning for the vectorized part. Note that I keep it avx2 and avoid avx2+fma as fma could affect the numerical stability here

On what I am looking at this plays a role in the InDet fitters that use the full field (not just fast/ ZR).

I have run the with/without 4 times in the usual ttbar and calo +tracking + egamma for a skylake and the PerfMonMT seem to be

before

PerfMonMTSvc                                         INFO Execute        99        61886              InDetAmbiguitySolver
PerfMonMTSvc                                         INFO Execute        99        57843              InDetExtensionProcessor   
PerfMonMTSvc                                         INFO Execute        99        9740               EMBremCollectionBuilder

after

PerfMonMTSvc                                         INFO Execute        99        60181              InDetAmbiguitySolver
PerfMonMTSvc                                         INFO Execute        99        56359              InDetExtensionProcessor              
PerfMonMTSvc                                         INFO Execute        99        9300               EMBremCollectionBuilder

So it seems it actualy help...

Edited by Christos Anastopoulos

Merge request reports