Skip to content
/ FLiT Public

A project to quickly detect discrepancies in floating point computation across hardware, compilers, libraries and software.

License

Notifications You must be signed in to change notification settings

PRUNERS/FLiT

Repository files navigation

branch status
master Build Status
devel Build Status

PyPI - License PRs welcome

FLiT

FLiT Bird

FLiT is a C++ test infrastructure for detecting variability in floating-point code caused by variations in compiler code generation, hardware and execution environments.

Originally, FLiT stood for "Floating-point Litmus Tests", but has grown into a tool with much more flexability than to study simple litmus tests. However, it has always been the focus of FLiT to study the variability caused by compilers. That brings us to the other reason for the name, "flit" is defined by the Merriam Webster dictionary as "to pass quickly or abruptly from one place or condition to another". This fits in well with testing for various sources of variability.

Compilers are primarily focused on optimizing the speed of your code. However, when it comes to floating-point, compilers go a little further than some might want, to the point that you may not get the same result from your floating-point algorithms. For many applications, this may not matter as the difference is typically very small. But there are situations where

  1. The difference is not so small
  2. Your application cares even about small differences, or
  3. Your application is so large (such as a weather simulation) that a small change may propagate into very large result variability.

Note: It may be the variability observed is more a cause of bad programming practice than over-aggressive compiler optimizations. In this case, compiler optimizations may act as fuzzers for your code. It is up to you to determine if this is the case.

FLiT helps developers determine where reproducibility problems may occur due to compilers. The developer creates reproducibility tests with their code using the FLiT testing framework. Then FLiT takes those reproducibility tests and compiles them under a set of configured compilers and a large range of compiler flags. The results from the tests under different compilations are then compared against the results from a "ground truth" compilation (e.g. a completely unoptimized compilation).

More than simply comparing against a "ground truth" test result, the FLiT framework also measures runtime of your tests. Using this information, you can not only determine which compilations of your code are safe for your specific application, but you can also determine the fastest safe compilation. This ability helps the developer navigate the trade-off between reproducibility and performance.

It consists of the following components:

  • a C++ reproducibility test infrastructure
  • a dynamic make system to generate diverse compilations
  • an SQLite database containing results
  • tools to help analyze test results
  • a bisection tool that can isolate the file(s) and function(s) where variability was introduced by the compiler.

Contents:

About

A project to quickly detect discrepancies in floating point computation across hardware, compilers, libraries and software.

Resources

License

Stars

Watchers

Forks

Packages

No packages published