search

trackreco / mkFit

Vectorized, Parallelized Tracking
https://trackreco.github.io/
Apache License 2.0
17 stars 15 forks source link

back to BackwardFitBH for iters without bw search #349

Closed slava77 closed 2 years ago

slava77 commented 2 years ago

345 was not in an obvious way supposed to modify the iterations without the backward search. This PR recovers that behavior.

plots are available for mkfit=all option compiled with AVX2:=1 USE_INTRINSICS:=-DMPT_SIZE=1

black (this PR) vs the reference as of #344 are clearly the same based on the ratio plots, e.g. in ttbar initialStep built tracks

image

If the change from #345 for e.g. this initialStep case was an improvement, I guess we'd be fine to pick it up. Since this is a regression (a clear increase in fakes), I propose to take this PR and then perhaps investigate if the non-BH solution can be improved (IIUC, it is more flexible).

@osschar I've updated both the run_OneIteration and runBtpCe_MultiIter, but the CMSSW MTV test covers only run_OneIteration. Please check and if that seems correct, perhaps consider merging.

osschar commented 2 years ago

Can you please rather figure out where the difference comes from? We will need this for seed region rebuilding and will need to figure it out anyway.

slava77 commented 2 years ago

Can you please rather figure out where the difference comes from? We will need this for seed region rebuilding and will need to figure it out anyway.

OK, I'll give it a try (not before Monday), although I'd guess that you may be more familiar with this part of the code.

if me or someone gets to this before Friday, this PR can wait; otherwise for the next CMSSW update, I'd like to fall back to this revert PR.

slava77 commented 2 years ago

OK, I'll give it a try (not before Monday), although I'd guess that you may be more familiar with this part of the code.

it's likely later than Monday. My next item is the pixel bad modules.

osschar commented 2 years ago

OK, then it's on me, I can start looking into this Monday morning. I agree we fall back to this if we can't figure it out in time :)

slava77 commented 2 years ago

superseded by #351