/stream_aie

Multi-core HW accelerator mapping optimization framework for layer-fused ML workloads.

Primary LanguagePythonMIT LicenseMIT

Stream Documentation

Stream is a HW architecture-mapping design space exploration (DSE) framework for multi-core deep learning accelerators. The mapping can be explored at different granularities, ranging from classical layer-by-layer processing to fine-grained layer-fused processing. Stream builds on top of the ZigZag DSE framework, found here.

More information with respect to the capabilities of Stream can be found in the following paper:

A. Symons, L. Mei, S. Colleman, P. Houshmand, S. Karl and M. Verhelst, “Towards Heterogeneous Multi-core Accelerators Exploiting Fine-grained Scheduling of Layer-Fused Deep Neural Networks”, arXiv e-prints, 2022. doi:10.48550/arXiv.2212.10612.

Install required packages:

> pip install -r requirements.txt

The first run

> cd stream
> python api.py

Documentation

Documentation for Stream is underway!