3.3.0

Key Highlights

LFRic

• Generalisation of the previous GlobalSum functionality to support other types of reduction (minimum/maximum value of a field);
• Support for new ScalarArray Kernel arguments;
• Extensive new training material added;
• Improvements for kernel extraction;
• Improvements to value-range check functionality, including new example;
• Access mapping moved from configuration file to LFRic constants;
• Shortening of generated symbol names plus additional checks to avoid invalid names;
• PSyclone integration tests updated to use LFRic 3.1;
• Functionality to write and re-name kernels removed - all kernel transformations must now be applied to kernels that have been brought into the same module as the call site (using KernelModuleInlineTrans);

NEMO

• Integration tests extended to exercise netcdf output and check the results;
• Transformation to increase rank of work arrays applied to solfrac and zdftke;
• Lists of NEMO functions removed from transformation scripts in favour of resolving imports;
• All supported reduction-replacement transformations are now applied when processing code (MIN/MAX with arguments other than default real now supported);
• Example transformation scripts have been unified and simplified;
• Transformations now applied to extract and then parallelise array expressions in arguments to iom_put.

General

• New profiling wrappers for the ROCTx (AMD) library;
• Use of the Python logging infrastructure has been introduced (but not yet fully adopted);
• Significant improvement/generalisation for handling existing code (change the default assumption such that unresolved symbols with parentheses are now assumed to be function calls rather than array accesses);
• Improved support for Fortran intrinsics, including keyword arguments and the ability to reason about their return type;
• New transformation to replace an expression with a temporary (DataNodeToTempTrans);
• New meta-transformation to construct the largest possible transformed region (MaximalRegionTrans), and largest possible OpenMP Parallel region (MaximalOMPParallelRegionTrans);
• Improvements to the handling of existing directives, including in the presence of comments;
• Various bug fixes and improvements to the handling of scalar variables in parallel regions, including the ability to explicitly specify private variables (was previously restricted to Loops);
• Initial support added for OMPCritical;
• Support for OpenMP conditional compilation;
• Improvements to the handling of symbols within WHERE constructs;
• Improvements and fixes to Reference2ArrayRangeTrans (including derived-type accessors and nested WHEREs);
• Bug fixes to ArrayAssignment2LoopTrans (allowing for dependencies and making use of Call/RoutineSymbol type information);
• Improvements to handling of Fortran derived types containing unsupported components;
• More code re-structuring - moving Transformation sub-classes out of transformations.py and deprecation of TransInfo;
• Performance optimisations, especially for dependence analysis;
• New command-line option to override configuration-file settings;
• Bug-fix for line-length limiting of inline comments;
• Null (integer) operations removed when constructing PSyIR;
• Support for Fortran character variables declared with a length attribute;
• Compatibility with OpenMP 5.2: SUB reduction replaced with ADD;
• Fortran backend extended to generate else if clauses to reduce deeply nested structures in output;
• Support added for preserving ACC ROUTINE/OMP DECLARE TARGET directives in existing code;
• Initial support for Fortran array constructors;

New Contributors

• @addy419 made their first contribution in #3282
• @thomasgibson made their first contribution in #3317

Full Changelog: 3.2.2...v3.3.0-rc1