accumulated Carpet updates

Pull requests https://bitbucket.org/eschnett/carpet/pull-requests/22/rhaas-carpetupdates/diff https://bitbucket.org/cactuscode/cactus/pull-requests/53/rhaas-updates/diff https://bitbucket.org/cactuscode/cactuspugh/pull-requests/3/pugh-initialize-new-cgh-fields/diff contain accumulated updates to Carpet, Cactus and PUGH in the funhpc branches.

I have verified that the three pull requests produce bit-identical output for the qc0-mclachlan.par parfile using 5 MPI ranks and 2 threads per rank on my workstation (debian-buster, gcc 8.2).

A summary of changes:

add support for very large grids where 64bit integer are needed for grid indices and sizes of transfer buffers
fix how how physical_time_per_hour is computed
add functionality to align interior of grid functions to cache boundaries. This requires changes to Cactus and PUGH as well.
add a parameter "granularity" to make sure the interior of components is a multiple of N points in each direction
always align data if VECTORISE_ALIGN_FOR_CACHE is set in option list indep. of pad_to_cachelines runtime parameter
cleanup, reformatting, minor speedups

To compare the changes it is not advisable to look at the diff directly. It is much better to run both the current and the new version through clang-format first:

cd repos/carpet
git checkout master
git checkout -b master-formatted
clang-format -style=file -i */src/*.{cc,hh}
git commit -a -m 'Beautify code'
git checkout rhaas/carpetUpdates
git checkout -b rhaas/carpetUpdates-formatted
clang-format -style=file -i */src/*.{cc,hh}
git commit -a -m 'Beautify code'
git diff master-formatted..rhaas/carpetUpdates-formatted

which is more manageable (but still multiple thousands of lines)

Keyword: backport

Comments (8)