Genetic improvement of GPU software

Abstract

We survey Genetic Improvement (GI) of general purpose computing on graphics cards. We summarise several experiments which demonstrate four themes. Experiments with the gzip program show that genetic programming (GP) can automatically port sequential C code to parallel code. Experiments with the StereoCamera program show that GI can upgrade legacy parallel code for new hardware and software. Experiments with NiftyReg and BarraCUDA show that GI can make substantial improvements to current parallel CUDA applications. Finally, experiments with the pknotsRG program show that with semi-automated approaches, enormous speed ups can sometimes be had by growing and grafting new code with genetic programming in combination with human input.

Improving SSE parallel code with grow and graft genetic programming

“…Next we return to the traditional role of software engineering: making better programs. We selectively review some examples of using evolutionary computation, a biology inspired technique, not just to make existing programs faster, but also as a relatively automated way of exploring different trade-offs, particularly between speed and quality, and also of exploiting parallel hardware (see also [14]).…”

Mutation Testing Advances: An Analysis and Survey

“…By realising this idea, one can form methods that assist activities outside software testing. Examples of this line of research are methods that automatically localise faults [17], automatically repair software [18], automatically improve programs' non-functional properties such as security [19], memory consumption [20] and execution speed [20,21].…”

A survey of genetic improvement search spaces

“…Langdon et al [18,21] presented fitness landscapes for first-order GI mutations (see Definition 2.6) applied to three large real-world programs. A subset of the codebase was considered for each piece of software, ranging from a few hundred to a few thousand lines of code.…”