Flexible Accelerator Library: Approximate Computing Executer (AxC Executer)

AxC Executer

The AxC Executer is a C++-based software suite that performs inference of Convolutional Neural Networks. The execution backend is fully written in C++ and intensively uses C++ Templates for kernel description. The executer backend is based on inheritance, with fallback execution modules that perform exact SW computation. Nevertheless, each execution module can be overridden to use custom execution modules.

Execution modules available:

• Convolution 2D
• Dense
• Reduction-based Activations: Softmax
• Addition 1D and 2D
• Activations: TanH

Since each module is replaceable, it is possible to implement custom modules to test optimisations and approximations and offload the computation to acceleration units. Likewise, it is possible to replace the arithmetic units (adders and multiplier) for a higher degree of approximation.

The models are represented through arrays of layers in C++. The models available are:

• LeNet5

Moreover, the current examples involve:

• Float32 computation
• Quantisation computation with arbitrary precision `ap_fixed<W, I>`
• Approximated arithmetic with Approximate Math Library.

Official Repository: GitLab

This project is part of the Flexible Accelerators Library (FAL), a work from:

• Costa Rica Institute of Technology
• University of Trieste
• Karlsruhe Institute of Technology

Authors:

• David Cordero Chavarría <dcorderoch@ieee.org>: student
• Luis G. Leon Vega <luis.leon@ieee.org>: supervisor

This repository includes the following:

• Running results
• Source code