SC21 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

SODA-OPT: System-Level Design in MLIR for HLS

Authors: Nicolas Bohm Agostini (Pacific Northwest National Laboratory (PNNL), Northeastern University); David Kaeli (Northeastern University); and Antonino Tumeo (Pacific Northwest National Laboratory (PNNL))

Abstract: High-level-synthesis (HLS) enables the generation of hardware descriptions from applications implemented with high-level languages. State-of-the-art tools, however, typically require the application to be manually translated to C/C++ and carefully annotated to improve final design performance. This cumbersome process prevents scientists and researchers from tapping into the power of HLS, as many of their applications are written in higher-level languages, such as python, and require significant effort to be ported. To overcome this challenge, we propose SODA-OPT, a front-end compiler tool that leverages the MLIR framework to provide host binaries and pre-optimized accelerator code for a back-end tool of choice. SODA-OPT requires no manual code translations and presents a novel approach to automate the outlining, mapping and generation of custom accelerators. Thanks to our selection of MLIR optimizations at the appropriate level of abstraction, our tool creates high-quality optimized kernels for HLS that expose code structures resulting in 1.6x faster runtime.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF

Back to Poster Archive Listing