sfpy

softfloat and softposit in Python

• support for softfloat float16, float32, and float64
• support for softposit posit8, quire8, posit16, quire16, posit32, and quire32

Installation

On most linux distros, sfpy should work out of the box:

pip install sfpy

Binary wheels (compatible with manylinux1) are available for CPython 2.7, 3.5, 3.6, and 3.7.

Under the hood, sfpy uses Cython to create bindings for the softposit and softfloat C libraries. These building instructions are tested on Ubuntu 18.04 - for other platforms they may need some adaptation. The cythonized C and compiled static libraries (.a), as well as necessary headers, are included in the source releases; installing these does not require Cython.

Demo

>>> import sfpy
>>> from sfpy import *
>>> a, b = Float16(1.3), Float16(1.4)
>>> a * b - a / b
Float16(0.89208984375)
>>> sfpy.float.flag_get_inexact()
True
>>> a += b
>>> a
Float16(2.69921875)
>>>
>>> x, y = Posit16(3.0), Posit16(3)
>>> x
Posit16(3.0)
>>> x.bits
22528
>>> y
Posit16(2.9802322387695312e-08)
>>> y.bits
3
>>> x * y
Posit16(8.940696716308594e-08)
>>> acc = Posit16(0)
>>> for i in range(10000):
...   acc = acc.fma(x, y)
... 
>>> acc
Posit16(1.9073486328125e-06)
>>> acc.bits
24
>>> q = Quire16(0)
>>> for i in range(10000):
...   q.iqma(x, y)
... 
>>> q
Quire16(0.00089263916015625)
>>> q.bits
64424509440000
>>> q.to_posit()
Posit16(0.00089263916015625)
>>> q.to_posit().bits
490

Building

See BUILDING.