Codecov Report

Merging #56 into master will increase coverage by 13.61%.
The diff coverage is 78.75%.

@@             Coverage Diff             @@
##           master      #56       +/-   ##
===========================================
+ Coverage   33.72%   47.33%   +13.61%     
===========================================
  Files          25       24        -1     
  Lines        6571     1126     -5445     
  Branches     1261      145     -1116     
===========================================
- Hits         2216      533     -1683     
+ Misses       4355      593     -3762

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update eb132af...dd3dbd2. Read the comment docs.

I won't be able to work on this for a day or two, but currently there is an issue partially related to this branch. The DP4 example fails with the new Minuit2 fitter for any backend with a NLL=0 error, or with the old Minuit1 fitter only with debugging enabled (which is making it hard to track down). This makes me suspect uninitialized memory access, perhaps in AddPdf. @galapaegos, this might be related to the MPI changes, I haven't tried running with pre-MPI code yet but plan to.

Just a comment on this, Minuit1 with Release+debug will solve but Minuit2 does not for DP4. Would the expectation be that Minuit1 and Minuit2 come to the same or very close result? I'm getting different results beyond 0.01 between Minuit1 and Minuit2, and the error is a larger error difference, with Minuit1 having the larger error in general.

Definitely possible there is an MPI bug; the non-MPI path should be exactly the same as the original excluding the RO_CACHE changes. The RO_CACHE changes will trigger a crash not a parameter assertion or sum assertion if used as a fakeEvent.

I believe the Minuit2 has better defaults; it seems to perform a HESSE like calculation automatically. If explicitly called, I believe you can perform the exact same procedures, but I tried to stay close to the default for FitManager. So it might be slightly better, but should not be very different.

(Will try to look into this further to ensure the results are correct)