# Brainstorming Cycle 34 02/26

Decision: 1 long 8-weeks cycle

Goals for the Quarter:
- Make ICON4Py granules (dycore and diffusion) available upstream by March end or early April.
- warm-bubble (submitted to ICCARUS March 16-20): because we can show the greenline is working (this quarter, but the blueline upstream can't be missed): JW multi-node has priority over Warm-Bubble single-node

PMAP:
- Move PMAP-LES to next

Other things happening in the quarter:
- muphys paper
- SIAM PP conference

## Blueline

- Upstreaming ICON4Py granule integration: ComIn (or py2fgen?) (Will, Christoph, Hannes)
- Figure out in the shaping if ComIn is a viable approach, and also document the py2fgen plan
- Upstream refactorings, e.g. swapped indices
- Ensure documentation and tests
- Session to explore, build and run BlueLine with C2SM (Annika, Matthieu, Mikael, Michael)
- Ensure commonly used ICON output features (e.g., lat/lon intp and meteogram) work with nblocks_e=1
- spack 1.X update ((easily) enables newer NVHPC, libfabric, MPI, NCCL) (Can Matthieu come up with a plan/timeline?)
- Nightly/weekly CI verification and performance runs of a real test case? (Christoph, Daniel, C2SM?)
- Continuations:
- Combine divergence damping coefficient calculation with other stencils ([PR#951](https://github.com/C2SM/icon4py/pull/951))

## Greenline

- [Warm bubble building blocks](https://hackmd.io/HvHaFPQrRP-8d9UzMA_Gkg?edit) (shaped by the group that made this document)
- Continuations:
- Halo construction
- MPI testing
- Likely finishable before next cycle
- Benchmarking
- Add AMD platform to benchmark-CI of standalone granules and stencil tests

[Warm Bubble Shaping](https://hackmd.io/5ceTe0y2SZWJGZBwFDdMiQ?both)

## GT4Py/DaCe

- Finish embedded implementation and add tests with `jax` / `jax.jit` (Enrique, Hannes)
- embedded concat_where + as_offset + bugfix for shift with connectivity of same dim + codim
- concat_where: infinite domain -> lazy field -> field refactoring???
- jax.numpy supported in testing
- jax.jit as backend
- Performance optimizations:
- AMD GPU performance (Ioannis)
- Someone tries to shape what we can do for balfrin performance (Christoph)
- Tracer/Expandable parameter use-case (shaping, but not necessarily schedule Till, Enrique, Hannes; we should be at least able to start optimizing this use case)
- caching bugs (Till, Enrique)
- Address cache stability issue, related to offset provider, observed in AMD reframe tests
- switching CPU <-> GPU backends
- Improve GT4Py compile time / lowering performance & predictability, stability
Motivation: Moving PMAP-LES to gt4py-next
- GTFN passes cleanup. Requires an investigation how can we can transform the IR not just pre or post-order.
- GTFN: Return K-only fields
- Improve testing infrastructure
- large programs that cover current weak points in gt4py which currently only surface in applications
- Fuzzing / automatically generated programs validated against embedded
- Generic types
- Any or variable length tuple (e.g. for containers)
- Union, Optional values
- Support for passing callable types