Miscellaneous of updates and adaptations #13
Conversation
There was a problem hiding this comment.
Many thanks for these changes and the cleanup @stubbiali . I can confirm the Meluxina arch works as expected, and the other bug fixes and clean-ups are also very nice.
The only thing I don't think should be in this PR is the switch to single precision, as this requires separate validation and a specific mention in the README etc. With a quick trial on our host nodes I already got this to not validate (Intel) or fail (Nvidia), and since CLOUDSC2 was specifically release for TL/AD exploration I'm reluctant to enable this without further investigation of results. Could you please remove this change, and file this as a separate PR/issue for a dedicated discussion?
Once this is addressed, and @reuterbal confirms the LUMI setup I think we're good to go here.
| ecbuild_find_package( NAME loki ) | ||
|
|
||
| # Add option for single-precision builds | ||
| ecbuild_add_option( FEATURE SINGLE_PRECISION | ||
| DESCRIPTION "Build CLOUDSC in single precision" DEFAULT OFF |
There was a problem hiding this comment.
I don't think this should be packed in with a general housekeeping PR, as single-precision is effectively not working for the Fortran variants. In my quick and easy experimentation I got floating point errors with Nvidia compiler on host side. Would you mind taking this to a separate PR for proper due diligence?
There was a problem hiding this comment.
I have tested the 15.0.1 arch on LUMI. Building works fine, apart from ad/tl plain SCC variants, which crash the compiler (what we have seen with CLOUDSC as well).
NL runs fine, too. However, tl/ad crash with a segfault on-device for me, suggesting something is not quite right there. This requires some further investigation but is out of scope for this PR.
loki_transform()rather thanloki_transform_convert()in CMakeLists.VALIDATE_TAYLOR_TEST.