Better variables for updated masks in fill_miss_2d #602
Conversation
Hallberg-NOAA
added
documentation
There was a problem hiding this comment.
PGI (nVidia) has a regression in OM4_025 and other experiments. See https://gitlab.gfdl.noaa.gov/ogrp/MOM6/-/jobs/126178
|
I don't see how to revise this proposal to avoid these changes in answers with the pgi (nVidia) compiler. After looking into this in detail, the answer changes only occur in cases that use an older formulation that does a 5-point sum without any parentheses. This give the compiler the ability to rearrange this sum at will, and the nVidia compiler appears to be doing so based on considerations that are not apparent to me. However, when the newer (rotationally symmetric) but mathematically equivalent form with parentheses that specify the order of operations are used, the answers are preserved for all of the compilers. |
Within fill_miss_2d, there are 2 sets of masked values used for remapping - masked values in neighboring points in arbitrary units (with names like "south") and the updated valid mask with values of 0 or 1 (with names like "gs"). At one point in the code, the former were being used where the latter were more appropriate. This commit changes this to help with the understandability of the code. However, when the order of arithmetic is not specified with parentheses, the Nvidia compiler is changing answers with this change of variables. To avoid this change in answers, this commit is only changing which variables are used when the answer_date is larger than 20190101 and the order of the sums are specified with parentheses. This commit also now moves the logical test for ans_2018 outside of the i- and j- loops where non-zero values of a_chg are set. All answers and output are bitwise identical.
Hallberg-NOAA
force-pushed
the
hor_regrid_fill_goodval
branch
from
a2b5dfd to
8c161ec
Compare
|
I have revised this commit to avoid the Nvidia changes by only changing which variables we are using when the relevant |
|
Gaea regression: https://gitlab.gfdl.noaa.gov/ogrp/MOM6/-/pipelines/23564 ✔️ |
Within fill_miss_2d, there are 2 sets of masked values used for remapping - masked values in neighboring points in arbitrary units (with names like "south") and the updated valid mask with values of 0 or 1 (with names like "gs"). At one point in the code, the former were being used where the latter were more appropriate. This commit changes this to help with the understandability of the code. All answers and output are bitwise identical.