Colloquium Computer Science - Final MSc Luc van den Brand

Title:  L-systems as a graph of operations

Abstract:
   L-systems allow for the procedural generation of complex structures
by iteratively rewriting symbols using a given set of rewrite rules.
L-systems are non-trivial to implement on parallel architectures, due
to their implicit and sometimes even explicit context sensitivity and
divergent behaviour. This thesis both implements and discusses some
known approaches to efficiently generate L-systems on multiple
processing cores, and focusses primarily on the viability of the
recently published approach of constructing an Operator Graph (OG)
from L-systems. We show that the provided OG implementation does not
support L-system generation, and we consequently reimplement the
theory using TensorFlow. TensorFlow implements the same concept of
constructing a graph of operations which can automatically be
scheduled for efficient parallel execution. The results of this thesis
put into question the practical use of constructing a graph of
operations as it pertains to efficiently generating L-systems. All our
reference implementations outperform our TensorFlow implementation,
due to their relatively small overhead and the limits of directly
implementing L-system generation as a graph of operations. The thesis
concludes by providing some open questions on the use of L-systems
through TensorFlow that have yet to be addressed.