v1.5.2

A bugfix release to point to the correct version tag for FIELD_API and fix Loki processing when using the OMNI frontend instead of the (default) FParser2.

What's Changed

• Ensure yoecldp.F90 is parsed with FParser by @reuterbal in #82
• Mark projects no longer as optional and fix field-api to v0.3.0 by @reuterbal in #83
• Version 1.5.2 by @reuterbal in #84

Full Changelog: v1.5.1...v1.5.2

v1.5.1

This is a bugfix release that fixes a performance bug in the code of Loki-generated variants, restoring the expected throughput performance on GPUs.

What's Changed

• Loki: enable imports and mark headers as ignored instead of blocked by @reuterbal in #80
• v1.5.1 by @reuterbal in #81

Full Changelog: v1.5.0...v1.5.1

v1.5.0

What's New

• A new variant based on Atlas, supporting FieldSet and MultiField storage backends (#60)
• A new GPU-optimised OpenACC variant "k-caching" (#46)
• HIP implementations for SCC, SCC-HOIST, SCC-K-CACHING (#59)
• A new Python-interface variant with a Python driver that calls the Fortran code (#38)
• A Python implementation generated via Loki (#51)
• SYCL implementations for SCC, SCC-HOIST, SCC-K-CACHING (#64)
• HDF5 input support for C-based variants
• Update to the FIELD_API variant to use the new open-source FIELD_API library
• Several updated or new architecture files

What's Changed

• Remove obsolete architectures, add march=znver2 option to gnu by @piotrows in #45
• Cloudsc variant using SCC optimization with field_api (standalone repo) by @awnawab in #25
• Openacc k-caching variant by @MichaelSt98 in #46
• PyIface - Fortran-Python bridging with CMake compilation by @piotrows in #38
• Python: Adapt to latest GT4Py release by @stubbiali in #50
• Pure Python variant with test setup by @mlange05 in #51
• Python-gt4py: Removing GT4Py-based implementation of CLOUDSC by @mlange05 in #52
• Pool allocator via Loki by @reuterbal in #49
• Enable gvmode and add nvidia/22.11 arch files by @reuterbal in #55
• Switch to CTest for CI and add NVHPC 23.5 to CI testing by @reuterbal in #56
• Architecture files for Leonardo by @reuterbal in #58
• Bundle integration for Atlas by @reuterbal in #57
• New Atlas-based variant using BlockStructured FunctionSpace by @sbrdar in #54
• CLOUDSC HIP (SCC, SCC-HOIST, SCC-K-CACHING) by @MichaelSt98 in #59
• Loki config update by @mlange05 in #61
• HIP updates by @MichaelSt98 in #63
• SYCL-specific env and README entry by @mlange05 in #68
• Introducing SYCL implementations/variants by @MichaelSt98 in #64
• Cuda update: metric col/s by @MichaelSt98 in #67
• Update to new open-source FIELD_API by @awnawab in #70
• use Atlas structures with an IFS-like variable batching by @sbrdar in #60
• Loki F2C using loki_transform instead of loki_transform_transpile by @MichaelSt98 in #69
• Emancipation from serialbox via HDF5 for C-style variants by @MichaelSt98 in #62
• Fix LUMI-G archs for HDF5 in HIP variants by @reuterbal in #72
• HDF5 support for CUDA/HIP/SYCL variants by @MichaelSt98 in #71
• Adjust CUDA-specific loki config file to Loki v0.2.0 by @mlange05 in #75
• Add semantic for reading thread count from $OMP_NUM_THREADS by @antoine-morvan in #73
• Compatibility with Loki v0.2 by @reuterbal in #74
• Update Github actions versions to remove node.js warnings by @reuterbal in #76
• Update loki version to 0.2.0 and field_api to 0.3.0 by @reuterbal in #77
• Version 1.5.0 by @reuterbal in #78

New Contributors

• @awnawab made their first contribution in #25
• @sbrdar made their first contribution in #54
• @antoine-morvan made their first contribution in #73

Full Changelog: v1.4.0...v1.5.0

Version 1.4.0

What's Changed

• Remove dependence on the global modules from the cloudsc-fortran by @piotrows in #35
• Bug-fix for Loki-SCC/H array demotion by @mlange05 in #39
• New Loki SCC-CUF variant by @MichaelSt98 in #31
• CUDA Fortran (CUF) 'k-caching' version (further optimized version) by @MichaelSt98 in #33
• Add col/s throughput metric in performance reports, add vector_length(NPROMA) to gang loops by @reuterbal in #42
• Initial FIELD API and gpu-scc-field variant by @mlange05 in #41
• CLOUDSC SCC-CUDA-C (semi-automatic) implementations by @MichaelSt98 in #40

Full Changelog: v1.3.0...v1.4.0

Version 1.3.0

This release includes a significant number of changes and new developments:

• An OpenMP-offload variant (dwarf-cloudsc-gpu-omp-scc-hoist), derived from the OpenACC SCC-hoist variant, developed by L. Lucido (Atos).
• A new CUDA Fortran implementation (dwarf-cloudsc-gpu-scc-cuf), implementing the SCC loop layout. This requires the --with-cuda flag to build.
• A new Python implementation, based on GT4Py, capable of generating CPU and GPU code (using cupy and, optionally, DaCe)
• Deprecation of the dwarf-cloudsc-gpu-claw variant, which no longer works correctly on recent NVIDIA software stacks. Building this variant requires adding an explicit --with-claw flag to the build command.
• With Loki publicly available, the references to the Loki repository have been updated and testing of the source-to-source translation variants is now possible.
• A JUBE benchmark configuration has been added to ease testing across various platforms.
• Clean-up of the CMake scripts
• New arch files for:
  • HPC2020, the ECMWF Atos system in Bologna
  • LUMI
  • MeluXina
  • Additional Isambard partitions

Version 1.2.0

Merge pull request #9 from ecmwf-ifs/develop

Merge develop to main