# Brainstorming Cycle 33 11/25

## Blueline

- [ ] Finish GHEX changes (Mikael)
- asynchronous scheduling
- NCCL support
- [x] Coupled simulation using blue line (Christoph, Hannes)
- merge blueline into coupled branch
- run (and compile) atmosphere at scale
- Prepare upstreaming of blue line, start discussion with ICON community (maybe use icon-cscs fork as basis)
- [ ]The current integration via py2fgen (Christoph)
- fix dimension order for sparse fields
- how do we upstream
- [ ] Investigate Community Interface! (Hannes)
- CI for blue line on balfrin and santis (Daniel)
(- merge icon:icon into icon-exclaim if necessary)
- finally can use upstream CI in dkrz (again by using icon-cscs)
- discuss icon-cscs idea during shaping
- Nightly builds for key science configurations (involve Anurag in shaping)
- add one torus grid
- Use at least small global experiments `exp.exclaim_uncoupled_R02B05L120` to benchmark on santis and the MCH production experiments to benchmark on balfrin (fix current bencher setup to have these experiments)
- Investigate balfrin dycore performance (Christoph)
- Building and deployment - involve C2SM (Matthieu, Daniel)
- document how to use it for users and developers
- tech talk?
- AMD MI300 port of ICON, with at least the dycore as a component, possibly some physics, which could include ecRAD-DaCe which @Andreas Jocksch (CSCS) mentions. (Will, Andreas)
- sketch the strategy

## Greenline
- integrate Configuration in driver (continue from cycle 32) (Magdalena)
- standalone driver for jw test case (continue from cycle 32) (Yilu, Chia Rui)
- solve the crash
- test on r2b6 or r2b7 grids
- add IO driver/scientific validation
- multinode (Magdalena, Nikki)
- finish Halo Construction
- add halo exchanges where needed (outside dycore) (continue from cycle 32)
- fix configuration (only do exchanges where needed)
- check for more issues like topography
- global reduction operations (nflatgradp, mean values, ...)
- multinode IO
- custom PyPI index server for experimental packages (dace, ghex, ...) (Enrique)
- Finish torus support
- Muphys graupel comparison (Will, Hannes)
- Preparation for paper
- Fix issue with lowering to SDFG of let-lambdas (also in GT4Py backlog)
- GT4Py code style?
- DaCe optimizations
- Start investigating AMD performance on ICON4Py dycore (Ioannis, Hannes)
- Profiling + NVIDIA micro-optimizations (Ioannis, Christoph)
- DaCe optimizations (Philip)
- Inlining pass (https://hackmd.io/H2mXKtePQIaa35RHTylQ9w)
- DaCe Optimization XII.a (https://hackmd.io/SuiIF_DpQ9G7kzSmCuMAxg)
- k-blocking; Why does it not results in speedups in diffusion (no need to shape)
- Library Node for Broadcast (https://hackmd.io/rxWTk1QxTAS_BihLu7ofnA)
- DaCe Miscellaneous refactor tasks (https://hackmd.io/Y8_80IBBQmOd32sy_zothA)
- uint16 tables (maybe not needed to shape)
- Look into ICON4Py test infrastruture regarding memory usage (Enrique, Magdalena)
- Handover document by Magdalena (MAGDALENA)

## GT4Py
- Next features workshop? (Hannes)
- do some preparation this cycle
- involve people in planning to set reasonable goals
- length, date, etc
- features:
- generics
- Instrumentation API (including traces and timers) (Enrique)
- Staggering: full vertical level vs half vertical level feature (https://hackmd.io/w2vmjApmRemddVeVLtfQOw)
- Dace backend:
- Fix issue with lowering to SDFG of let-lambdas (ongoing)
- Refactoring of GTIR visitor to SDFG for better maintanability (Edoardo)
- GT4Py usability (Enrique)
- Support read-only filesystems for cache folder / multiple cache folders (Enrique, maybe no need to shape)
- Drop support for FrozenNamespace and use nicer mechanisms
- use Enums instead
- final objects
- Python 3.14 support (Hannes)
- drop Python 3.10, 3.11?
- update icon uenv?