Open dpo opened 6 years ago
The best way of updating y (in terms of accuracy) is not yet clear to me. cpgmres_y updates y after the inner loop by using the same strategy currently used by reg_cpkrylov to get the final y (reg_cpkrylov does not use the last m entries of the solution computed by the CP-Krylov solver). Initial experiments with random matrices seemed to show that this strategy led to more accurate vectors y, but this has never been analyzed in detail.
Ok, thank you. We'll revisit this after more extensive tests on non-random problems.
cpgmres_y
computesy
differently at the end of the loop, but needsG
explicitly. Does it have an advantage overcpgmres
?