Vincenzo Innocente
Vincenzo Innocente
Is really so different than ```assert()```? should not we try to implement some sort of ```LogError/LogWarning``` using unified memory?
we could reduce the number of error-checking in the code leaving assert for debug and trap only specific places... (which?)
GPU code expects data from a single module to be contiguous. It does not expect rawid to be sorted. The sequence printed by @ferencek and @nothingface0 should give problems on...
Not sure what to do. The input format is not what expected (duplicated modules?). It's a bit like duplicated pixels. The numbering of the clusters is not consistent in any...
here https://cmssdt.cern.ch/dxr/CMSSW/source/RecoLocalTracker/SiPixelClusterizer/plugins/gpuClustering.h#386 something for sure goes "wrong". Last arrived is served...
In my opinion if one switch ```GPU_DEBUG``` some assert will fail. As is RecHits "maybe" messed up in the affected modules (some missing, some "added"). A crash on gpu is...
ok. let's protect the conversion from SoA and go on: (rechits module most probably did not run yet)
detecting on GPU is trivial enough to count the number of appearance of a given module at https://cmssdt.cern.ch/dxr/CMSSW/source/RecoLocalTracker/SiPixelClusterizer/plugins/gpuClustering.h#87 What to do next is unclear to me. One could invalidate "single"...
large number of duplicate pixels I suppose
the detection of duplicate pixels has never been integrated, not even in debug mode...