A DType and scalar which uses MPFR for multi precision floating point math. MPFR itself has an LGPL license.
A very basic example::
import numpy as np
from mpfdtype import MPFDType, MPFloat
# create an array with 200 bits precision:
arr = np.arange(3).astype(MPFDType(200))
print(repr(arr))
# array(['0E+00', '1.0E+00', '2.0E+00'], dtype=MPFDType(200))
# Math uses the input precision (wraps almost all math functions):
res = arr**2 + np.sqrt(arr)
print(repr(res))
# array(['0E+00', '2.0E+00',
# '5.4142135623730950488016887242096980785696718753769480731766784E+00'],
# dtype=MPFDType(200))
# cast to a different precision:
arr2 = arr.astype(MPFDType(500))
print(repr(arr2))
# array(['0E+00', '1.0E+00', '2.0E+00'], dtype=MPFDType(500))
res = arr + arr2
print(repr(res)) # uses the larger precision now
# array(['0E+00', '2.0E+00', '4.0E+00'], dtype=MPFDType(500))
There also is an mpf.MPFloat(value, prec=None). There is no "context"
as most libraries like this (including mpfr itself) typically have.
The rounding mode is always the normal one.