This library provides code for performing robust geometric computations using adaptive-precision floating-point arithmetic. It is based on code by Jonathan Richard Shewchuk; the original source is available here. The theory behind adaptive-precision arithmetic used in the original source was described in a paper also by Shewchuk. These geometric predicates were used to write the mesh generation program Triangle, which received the 2003 Wilkinson Prize for Numerical Software.
Things that have changed since the original source code was written in 1996:
- The C99 standard was ratified and implemented in free, open-source compilers (GCC, clang). This introduces many syntactic niceties not available in K&R C, the language used in the original code.
- Compilers have gotten a whole lot smarter about optimizations like loop unrolling and function inlining.
- The most common CPU was the Intel Pentium series, which used 80-bit extended precision floating-point registers. These were intended to give greater numerical accuracy, but instead made FP computations notoriously unpredictable; x87 floating-point arithmetic has since been deprecated in favor of SSE instructions.
- The C standard library and headers have grown substantially, obviating the need for some of the routines in the original source code.
For example, one no longer needs to calculate the machine epsilon; this is now in the standard library header
float.h, although one has to take care which definition of epsilon is meant. - The C math library now includes the routines
nextafterandnextafterf, which return the next representable floating-point number after an input number; these procedures make it much easier to write stress tests for geometric predicates.
This repository is intended to contain an updated version of Shewchuk's original source code accounting for these developments. The original source is one 4.5k-line C file, although full of many helpful comments. This is somewhat hard to digest, so I'm breaking it up into several smaller files. Additionally, I'm adding unit tests, which will demonstrate both robustness of geometric predicates implemented with adaptive-precision arithmetic and the non-robustness of naive implementations. There will be no substantive modifications to the main routines.
Unit tests are in the directory test/; demonstrations of the failure of geometric predicates in floating-point arithmetic are in the directory demos.
The code is built using CMake as follows:
mkdir build
cd build
cmake ..
make
To run the unit tests:
make test
Demo programs are optionally built if you have zlib and libpng.
This code has not been released under a well-known free software license such as the BSD License or the GNU Public License. Instead, Shewchuk released the original predicates code stating only that it was "Placed in the public domain". Lawrence Rosen, general counsel for the Open Source Initiative, has argued that public domain is not a license. I am not a lawyer, but if you're thinking of using this code for proprietary software you should consult a lawyer first. I do not own the copyright on this code and cannot grant a license to use it.