Closed henryleberre closed 1 month ago
When this works it will be my favorite PR to date, second to the Frontier merge and my LAPACK commit.
@sbryngelson I modified the 1D_qbmm example case to run on 8 A100s in 3D with different BCs. I compared the results against master using the ./mfc.sh packer
and the results are EXACTLY the same:
Comparing test/master.txt to test/master.txt:
OK: The two packs are within tolerance (abs: 1.00E-12, rel: 1.00E-12).
Average error: abs: 0.00E+00, rel: 0.00E+00.
This test was important because halo-exchange buffers are different when qbmm .and. .not. polytropic
evaluates to true.
wonderful. I think the Frontier case will run before maintenance and then I will merge. I won't merge anything else in the meantime anyway since all other PRs will also be waiting on Frontier CI.
Description
s_mpi_sendrecv_variables_buffers
has too much repetition and is therefore very hard to read and maintain. This PR refactors it, removing 1.2k LOC.Scope
How Has This Been Tested?