Performance optimization for NEC SX AURORA

What is the bug

The subroutines init_zero_4d_[dp,sp,i4] do not vectorize properly on NEC SX AURORA.

How do you fix it

Loop collapsing is enforced by a compiler directive for the subroutines init_zero_4d_[dp,sp,i4], in order to ensure proper vectorization on NEC SX AURORA.

How urgent is the bugfix

• I need it as soon as possible
• I can wait for a couple of days
• None of my current codes is directly affected

Mandatory steps before review

• Gitlab CI passes (Hint: use make format for linting)
• Bugfix is covered by additional unit tests
• Mark the merge request as ready by removing Draft:

Mandatory steps before merge

• Test coverage does not decrease
• Reviewed by a maintainer
• Incorporate review suggestions
• Prior to merging, please remove any boilerplate from the MR description, retaining only the What is the bug and How do you fix it section to maintain
• Remember to edit the commit message and select the proper changelog category (feature/bugfix/other)

You are not supposed to merge this request by yourself, the maintainers of fortan-support take care of this action!

Dear Yen-Chen @b382190 and Pradipta @k202170,

is it possible to include the following NEC-specific optimizations into the fortran-support library?

Best regards, Daniel

requested review from @k202170

changed title from Draft: Performance optimizatiojn for NEC SX AURORA to Draft: Performance optimization for NEC SX AURORA

marked the checklist item Gitlab CI passes (Hint: use make format for linting) as completed

marked this merge request as ready

marked the checklist item Mark the merge request as ready by removing Draft: as completed

changed the description

marked the checklist item Test coverage does not decrease as completed

marked the checklist item Reviewed by a maintainer as completed

approved this merge request

enabled an automatic merge when all merge checks for 3e5adc64 pass

resolved all threads

merged

mentioned in commit 35723fc9

Hi Yen-Chen, Hi Pradipta,

just a short update regarding my attempt to upgrade to the latest commit of libfortran-support. It seems that this is currently not straightforward for icon-nwp (and probably icon-main as well).

I encountered problems in the external icon-dace and in shared/mo_var_list.f90. There might be a few more problems, as I have aborted my attempt at this point.

It seems that icon-dace needs to be upgraded such that the variable nnml is no longer used from mo_namelist. With newer version of libfortran-support it must be used from mo_io_units instead. This is totally fine, but it requires icon-dace to be updated before we can move to newer versions of libfortran-support.

In shared/mo_var_list.f90 I encountered the error message

  380 |       CALL init_contiguous_sp(new_elem%s_ptr, PRODUCT(d(1:5)), ivals%sval)
      |                                                                          1
Error: Missing actual argument for argument 'lacc' at (1)

It is not super urgent for me to switch to the latest version of libfortran-support. I just wanted to share my experience.

Are you planning to move to the latest version of libfortran-support in icon-main in the near future? Please let me know, if I should talk to our DACE colleagues regarding the required changes in icon-dace.

Dear Daniel, thanks for finding this error.
This was an error related to an update where we got rid of the misused i_am_accel_node argument. We missed this function during the update. See issue in libfortran-support and version update in <code data-sourcepos="2:273-2:281">icon-mpim</code>
To fix this specific piece of code, please add lacc=lopenacc in the argument list, same for the other similar init_contiguous functions.

Yes, we plan to do a new version update very soon. I will talk to people from DACE and figure out the process. Thank you for posting the questions.

Hi Yen-Chen,

thanks you very much for this background information.

Hi Daniel @m300173,
For your information, we are doing library version updates here: https://gitlab.dkrz.de/icon/icon-mpim/-/merge_requests/563.
We do it in icon-mpim because it is currently the only repo with libiconmath.
I have fixed all the compile issues regarding the new update in the MR.

Because we update the code in icon-mpim, there might be files that we are missing in icon-nwp where the compile issue regarding the lacc arguments in the init and copy functions come up again.
I am happy to help if you encounter similar trouble in the future.

Hi Yen-Chen,

thank you for your update on this. I will try to pull your changes to icon-nwp as soon as they are available in icon.git and let you know if I run into trouble. I guess that the lacc issue should be rather straightforward to solve for me, but I am a bit concerned regarding the icon-dace issue. We'll see ;-)

resolved all threads