[ODK] Meeting 2018 11 14

Winter Days

• a priori 4,5,7 december
• Montpellier still needs to send its dates

Study of Rational CRA machinery

• is it possible to separate the CRA reconstruction with the Rational Reconstruction

  • no because early termination requires to be able to do rational reconstruction to detect stabilization

• rational-cra.h et cra-domain.h are almost identical except for the nb of output args (num,den vs value)

  • should be merged and have only 1 value, which could possibly be a struct of num,den

• rational-cra2.h is a more refined machinery for matrices over QQ. No idea how well it works for the moment (probably not much tested).

  • let it as is for the moment, longer term objective

• Once the merge of rational-cra and cra-domain are merged

  • rewrite cra-omp.h from it (using FFLAS-FFPACK's Paladin)
  • rewrite cra-mpi.h from it

• masked primeIterator: the operator() of the rational-cra:

  • should have a constructor from (j, max, primeiter), to get all information passed to operator(). Should also get its tag (heuristic) from the main primeiterator.

• thought on the cra-omp manager

  • we should avoid as much as possible the critical sections where each worker updates the builder with a call to progress
  • implement a specific manager where progress is void and all reconstruction is done at result
  • replace the while loop with a parallel for (from paladin)