Installation and Basic Usage

Installation

Binary packages are avialable from PyPI. Use pip to install the binary package.

pip install numpy-flint

Numpy-flint is a c-extension python module, and must be built for each python version, each operating system, and each hardware architechure. Binary packages have been built for most recent python versions and common architechures. If you find that your particular system is not supported you can try and build the project form source

Usage

To use in python, import the flint package and declare the number as a flint type.

from flint import flint

# Floating point numbers sometimes don't make sense
a = 0.2
b = 0.6
# This evaluate to False
print( (a+a+a) == b )

# Rounded floating point intervals will fix these issues
x = flint(0.2)
y = flint(0.6)
# This evalautes to True
print( (x+x+x) == y )

The flint type renders as a single number: the tracked value. To see the full interval you need to use the interval property which returns a tuple with the lower bound, upper bound. If all you need is the size of the interval, you can use the eps property.

f = flint(1.5)
print(f) # shows 1.5
print(f.interval) # shows (1.4999999999999998, 1.5000000000000002)
print(f.eps) # Shows 4.440892098500626e-16

In addition to arithmetic, many math functions have been implemented as methods on the flint object. To see a full list of supported math functions look at the python api.

f = flint(2)
print(f.sqrt()) # 1.4142135623730951

In addition to a new python class, flint``s have been added as a new custom dtype. To use with NumPy, import NumPy as well. Then you can create your array as usual, simply mark the array's dtype as ``flint.

import numpy as np
from flint import flint

a = np.fill((3,), 0.2, dtype=flint)
b = flint(0.6)
# This evaluates to True
print( np.sum(a) == b ) # True

Finally all math functions have been added as recognized data-types to NumPy’s math functions, so you can write:

a = np.arange(5, dtype=flint)
print(np.sqrt(a)) # np.array([0.0, 1.0,  1.4142135623730951, 1.7320508075688772, 2.0], dtype=flint)

Caution

Working with floating point intervals is much slower than standard floating point operations, and should only be used where the added guarantee of ‘could be equal’ is required. Also note that the project is in an early and mostly untested state. As of right now, numpy-flint implements most real functions defined in the C99 math.h header file, but lacks some of the functions for interacting with integers such as round, ceil, floor, and fmod.