# GTIR Q3 24

## For DaCe

- Lowering FOAST to GTIR
- Transformation passes update:
- Inline Functions
- Infer domain (trace shift followed by propagate domain)
- Merge Type Inference

### in DaCe backend

- deal with let-lambdas?

### Optimization work

- Back to nabla4 with torus against Ioannis' experiments?

## For GTIR

- Update remaining transformation passes including temporary placement

--------------------

# Roadmap for next quarter

## Milestones
1. Use one of the combined programs in the dycore to test the full pipeline to go to DaCe using the new GTIR and test performance. (cycle 23)
2. Full dycore working with the new DaCe backend (maybe ignoring some exotic features and corner cases) (cycle 24)
3. Validated dycore with better performance than current gt4py performance. (end of cycle 24 or cycle 25)

## Problems
- It's not 100% clear how easy could be to find and apply valid DaCe transformations for basic performance optimizations life fusing maps.
- Implementing the exotic features in DaCe might take longer than expected.
- external depedendency on DaCe team
- Need to get a reliable baseline for icon4py programs (should be done outside of GT4Py core to parallelize work)