Fix issues with overflows #73
Conversation
phe/util.py
Outdated
|
|
||
|
|
||
| def round_m_mul_2_pow_e_mul_b_pow_c(m, e, b, c): | ||
| """Compute round(m * 2 ** e * b ** e) for m, e, b, c integers. |
There was a problem hiding this comment.
Typo in the equation (e instead of c).
Would be good to add a few regression tests checking that this function equals round(m * 2 ** e * b ** e) where it should...
phe/encoding.py
Outdated
| # a perfectly valid float. We represent scalar as integers | ||
| # and avoid coercing to float. | ||
| m, e = frexp_int(scalar) | ||
| int_rep = round_m_mul_2_pow_e_mul_b_pow_c(m, e, |
There was a problem hiding this comment.
Variable names should be more explicit - use mantissa and exponent instead of m and e.
Also the util function is a bit of a mouthful, I'd probably go for something more along the lines of compute_float_encoding.
phe/util.py
Outdated
| @@ -427,3 +429,68 @@ def is_prime(n, mr_rounds=25): | |||
| return False | |||
| # the actual generic test; give a false prime with probability 2⁻⁵⁰ | |||
| return miller_rabin(n, mr_rounds) | |||
|
|
|||
|
|
|||
| def int_div_round(a, b): | |||
phe/encoding.py
Outdated
| int_rep = scalar * cls.BASE ** -exponent | ||
| else: | ||
| # Integer division with correct rounding. | ||
| # Can't just to `scalar * cls.BASE ** -exponent` as it |
phe/encoding.py
Outdated
| # It is possible for scalar and BASE ** -exponent to not be | ||
| # representable as float, even when the encoded value is | ||
| # a perfectly valid float. We represent scalar as integers | ||
| # and avoid coercing to float. |
There was a problem hiding this comment.
Possible to misunderstand your comment here as the scalar you mention is different from the current scoped variable scalar which at this point is a float.
|
Oh one final comment, there needs to be tests that take each possible encoding path. |
|
I’ve got rid of a bunch of helper functions by using the |
Fixes an issue in
EncodedNumber.encodeand another inEncodedNumber.decodewhereOverflowErrors occur for valid inputs. This happens when a very bigintis coerced to afloatbut can’t fit. (This may occur even when the intended result is a perfectly valid, finitefloat.)We work around this by performing exact arithmetic for as long as possible.
May be related to #62.