pgoelz/xfp

Synthesizer for accurate fixed-point arithmetic expressions

The xfp project

xfp is a tool that can synthesize rewrites of an arithmetic expressions which are more accurate when implemented in fixed-point arithmetic. The search is performed with genetic programming, for which we use ECJ (A Java-based Evolutionary Computation Research System). More details can be found in the paper Synthesis of Fixed-Point Programs, E. Darulova, V. Kuncak, R. Majumdar, I. Saha, EMSOFT'13.xf

This project consists of two parts:

• an analysis tool for estimating worst-case errors of fixed-point code
• genetic programming, implemented in the ECJ framework for synthesizing more precise fixed-point expressions.

Examples are located in the inputs/ directory.

Analysis tool

The analysis_tool/ directory contains the source files and a build file for sbt (simple-build-tool).

An easy way to compile, run and package the tool is using sbt. You can of course use any other tool.

Start sbt in interactive mode in the analysis_tool directory.

• to compile: type 'compile'
• to package: type 'package' (the jar file will be placed into the target/ directory)
• to run: type 'run' to see the options of the analysis tool.

Note: if you want to make changes visible for GP, you need to replace the xfp.jar in the lib/ directory.

GP

The xfp specific source code is in the ec/app/fixed/ directory and the source code for the expression parser from parser.jar is located in ec/utils/.

Settings (gp specific): ec/app/fixedpoint/fixedpoint.params

To build: make

To run the one example specified in fixedpoint.params: run.sh Results are written to the results.txt file.

To change what is printed in the results.txt output file, modify ec.simple.SimpleStatistics (check the end of the file).