MueLu/Tpetra local-ordinal/offset type errornous conversion issues
Created by: mndevec
I am having issues whenever I "serially" run experiments with large matrices that have larger than 2B nonzeroes. For example, I am running MueLu test driver, with 200M rows and 5.4B rows, which should be okay for single rank, as local ordinals are int, and offset type is size_t. This segfaults without printing anything, which suggests overflow issues as early as during input generation.
I am trying to trackdown the issue using the compile flag -Wconversion to capture where the local ordinal types are used instead of offset types. Many of the warnings are over cautious, but still there are various errors that arise from the confusion of LO and offset_type, possibly causing segfaults in this case.
This conflicts with the use of separate types for offset and local ordinals. That would be good to fix these issues, and if possible to have a "stress" test for such cases. @kddevin