Kernels for GAP8 implementing the layer fusion of depthwise convolution and pointwise convolutions. If you intend to use or reference this library for an academic publication, please consider citing it:
@inproceedings{accelerating,
title={Accelerating Depthwise Separable Convolutions on Ultra-Low-Power Devices},
author={Daghero, Francesco and Burrello, Alessio and Poncino, Massimo and Macii, Enrico and Jahier Pagliari, Daniele},
booktitle={International Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation},
pages={46--58},
year={2024},
organization={Springer}
}
The source file can be found under src/, divided in three different directories.
Kernels of base depthwise and pointwise with different layouts (CHW/HWC) than the ones available in pulp-nn.
Fused depthwise separable kernels.
Base matmul kernel with CHW layout.