Extending MLIR to support SDFGs

Autor: Ates, Berke
Přispěvatelé: Hoefler, Torsten, Ben-Nun, Tal, Calotoiu, Alexandru
Jazyk: angličtina
Rok vydání: 2021
Předmět:
DOI: 10.3929/ethz-b-000520054
Popis: At present, the optimization of code boils down to minimizing the movement of data. DaCe achieves this by creating an intermediate representation and applying different optimizations to it. However, DaCe is currently only available for source code written in Python. Here we show that using DaCe to optimize intermediate representations results in significant performance gains. We found that the Multi- Level Intermediate Representation (hereafter denoted as MLIR) is well suited to support more source languages. Our results demonstrate that creating a dialect in MLIR could potentially allow any source language to leverage DaCe and profit from the increased performance. We anticipate this project to result in an optimization pipeline that decouples the domain scientist from the performance engineer and increases performance for all existing software projects without manual optimization. The performance in scientific computing, as well as personal computing, could be significantly improved.
Databáze: OpenAIRE