Electronic Design Automation Project

This repository accompanies our solution for the course Electronic Design Automation (5SIB0) at Eindhoven University of Technology.

Usage

The design automation tool synthesises a Verilog implementation of a cross-correlation function for a convolutional layer in a neural network. It does this based on parameters supplied by the user:

• input_size: The size (width and height) of the input matrix
• channels: The amount of input channels (e.g. 3 for an RGB input "image")
• kernel_size: The size (width and height) of the convolution kernel
• filters: The amount of convoliton sets, i.e. the amount dimentions on the output

Modules

The solution consists of following modules (see the src directory):

• cmd: the entrypoint of the script which performs the optimization algorithm
• operations: contains all Operation subclasses, used to make an abstract representation of the cross-correlation loop
• generator: generates an abstract representation of the network and performs the loop unrolling optimization

Autors

K.H.W. Stolle k.h.w.stolle@student.tue.nl

M. van Dijke m.v.dijke@student.tue.nl

S.L. Kamp s.l.kamp@student.tue.nl