Home > Uncategorized > Tool for tuning the use of floating-point types

Tool for tuning the use of floating-point types

A common problem when writing code that performs floating-point arithmetic is figuring out which of the available three (usually) possible floating-point types to use (e.g., float, double or long double). Some language ‘solve’ this problem by only having one possibility (e.g., R) and some implementations of languages that offer three types use the same representation for all of them (e.g., 32 bits).

The type float often represents the least precision/range of values but occupies the smallest amount of storage and operations on it have traditionally been the fastest, type long double often represents the greatest precision/range of values but occupies the most storage and operations on it are generally the slowest. Applying the Goldilocks principle the type double is very often selected.

Anyone who has worked with floating-point values will be familiar with some of the ways they can bite very hard. Once a function that uses floating-point types is written the general advice is to leave it alone.

Precimonious is an interesting new tool that searches for possible performance/accuracy trade-offs; it randomly selects a floating-point declaration, changes the type used, executes the resulting program and compares the output against that produced by the original program. Users of the tool specify the maximum error (difference in output values) they are willing to accept and Precimonious searches for a combination of changes to the floating-point types contained within a program that result in a faster program that does not exceed this maximum error.

The performance improvements cited in the paper (which includes the doyen of floating-point in its long list of authors) cluster around zero and worthwhile double figure percentage (max 41.7%); sometimes no improvements were found until the maximum error was reduced from 10^{-10} to 10^{-4}.

Perhaps a combination of Precimonious and a tool that attempts to improve accuracy is the next step 🙂

There is resistance to using search based methods to fix faults. Perhaps tools like Precimonious will help developers get used to the idea of search assisted software development.

I wonder how long it will be before we see commentary in bug reports such as the following:

  • that combination of values was not in the Precimonious test set,
  • Precimonious cannot find a sufficiently optimized program within the desired error tolerance for that rarely seen combination of values. Won’t fix.
  1. Magnus
    February 3, 2014 04:04 | #1

    It’s worth pointing out that compilers support different floating point models independent of number of bits of representation, e.g. MSVC++ has /fp:precise, /fp:strict, or /fp:fast.

  1. No trackbacks yet.