COJAC - Boosting arithmetic capabilities of Java numbers
Java numbers with a fresh eye and a cool tool...
... Don't miss the opportunity:
Numerical Sniffer, and other floating point gadgets
Automatic detection of all those poisonous problems with numbers, such as:
- integer overflow
- absorption and cancellation with float/double
- NaN or Infinite results
- offending type casting
New behaviors for floating point numbers, such as doubles-are-floats, or redefinition of the default rounding mode.
No recompilation needed, easy to use (
java -javaagent:cojac.jar ...)
Automatic replacement of every float/double with more powerful number types, such as:
- high precision numbers: you choose the number of significant digits for the standard float/double types! What a super-power for a runtime flag!
- interval computation, to keep rounding errors under control and signal when it grows too much. Again, without any source code modification!
- automatic differentiation, to bring you (for free!) the derivative of any encoded function. No other port of AutoDiff is that elegant!
- symbolic processing, and even Chebfun-like representation
No recompilation needed, easy to use (you want 100 significant digits?
java -javaagent:cojac.jar="-Rb 100" ...)
The following features have been implemented and recently included in the current release:
symbolic expressions: we rebuild the formulas from the bytecode, and then apply simplification rules, just like Mathematica and similar great tools are able to do. Hard to believe, but yes we did it!
symbolic functions: a step further, where you can tag a number as being the unknown.
Chebfun representation: a completely different way to represent symbolic functions, with a good compromise of accuracy/efficiency (see Matlab's Chebfun). But here completely transparently: with COJAC the same computation code can be run numerically, symbolically, or as Chebfuns.
a couple of well-known debugging tools useful to track possible instabilities: forcing doubles to act like floats, changing the rounding mode, negating the comparisons when the numbers are too close together.
a Delta-Debugging infrastructure for the debugging features above, so that you can automatically find the portion of code that causes the instabilities.
Developed at the School of Engineering of Fribourg, Switzerland.
Distributed under the "Apache License, v2.0".