Emeraude-MLIR
Multi-level arithmetic compilation from real-valued intent to synthesizable hardware
emeraude-mlir is my research compiler for keeping arithmetic intent explicit from high-level kernels down to hardware generation.
What It Adds
- Custom dialects for real-valued and fixed-point arithmetic (
RealArith,FixedPointArith, plus Faust integration). - Conversion passes between standard MLIR dialects and arithmetic-aware representations.
- Approximation hooks (FloPoCo/Sollya) for controlled polynomial materialization.
- End-to-end HLS driver pipelines to run reproducible experiments.
Scope
The compiler exposes precision, circuit structure, and lowering choices before RTL generation.
This project feeds my EuroLLVM/MLIR workshop work and the TinyTapeout demonstrators where the path from software kernel to silicon artifact remains inspectable.
Links
- Availability: the compiler repository is not public. The ttsky26a demonstrator and publications document parts of the flow.
- Related post: My Latest TinyTapeout makes MLIR Arithmetic Silicon-Visible
- Related publication entry: HAL 05594483