numfi 0.4.0

a numpy.ndarray subclass that does fixed-point arithmetic

numfi

numfi is a numpy.ndarray subclass that does fixed-point arithmetic.

Feature:

• Automatically perform fixed-point arithmetic through overloaded operators

• Maximum compatibility with numpy and other library, just like a normal numpy.ndarray

• Optimized calculation speed by minimizing quantization as much as possible

Install

Prerequisite: python3 and numpy

pip install numfi

or you can just copy numfi.py and do whatever you want, after all it's only 200+ lines of code

Quick start

import numfi
import numpy as np

# numfi(array=[], signed=1, bits_word=16, bits_frac=None, RoundingMethod='Nearest', OverflowAction='Saturate')
x = numfi(np.random.rand(3),1,16,8) 
# numfi.__repr__() return brief description of numfi object: x => s16/8-N/S
# s for 'signed', followed by word bits and fraction bits, N/S for 'Nearest' and 'Saturate` for rounding/overflow method

# any arithmetic operation with numfi will return a numfi object with proper precision and value
# By overloading operators, numfi object can do fixed-point arithmetic easily:

# normal arithmetic operation work with float form of x
y = x + 1
y = [1] - x
y = x * np.random.rand(3)
y = numfi([1,0,0.1234],1,21,15) / x
y = -x
y = x ** 0.5
y = x % 3
# comparison return np.array of bool, just like normal np.array
y = x > 0.5
y = x >= numfi([1,0,0.1234],1,21,15)
y = x == x
y = x <= np.ones(3)
y = x < [1,1,1]
# bitwise operation work with integer form of x
y = x & 0b101 
y = 0b100 | x   # order of operands doesn't matter
y = x ^ x       # two numfi object can also be used in bitwise operations
y = x << 4
y = x >> 2
...

# By inheriting from numpy.ndarray, numfi object can be used just like normal numpy array, and return same numfi object back
y = np.sin(x)
y = x[x>1]
y = x.sum()
y = x.reshape(3,1)
plt.plot(x)
pandas.DataFrame(x)
np.convolve(x,np.ones(4))
np.fft.fft(x,n=512)
for i in x:
    print(i)
...

Document

Details can be found here: https://numfi.readthedocs.io/en/latest/?

License

The project is licensed under the MIT license.