[cfe-dev] libc++: Implementation of std::poisson_distribution::operator()

[Marshall Clow via cfe-dev]

On Wed, Jan 10, 2018 at 11:25 AM, Nagurne, James via cfe-dev <
cfe-dev at lists.llvm.org> wrote:

> Back in 2010, the algorithm for this function was updated, adding the
> following clause:
>
>
>
>             if (__x < 10)
>
>             {
>
>                 const result_type __fac[] = {1, 1, 2, 6, 24, 120, 720,
> 5040,
>
>                                              40320, 362880};
>
>                 __px = -__pr.__mean_;
>
>                 __py = _VSTD::pow(__pr.__mean_, (double)__x) / __fac[__x];
>
>             }
>
>
>
> In this case, result_type is the single template parameter to the
> instantiation of std::poisson_distribution.
>
> In the case that std::poisson_distribution is instantiated with a 16-bit
> signed integral type, __fac[8] and __fac[9] are out of range of
> result_type, and are truncated or result in a change of sign.
>
>
>
> Since the only use of __fac is as an operand to a double-precision
> floating point division, would it not make sense to instead declare it
> ‘double’ instead of ‘result_type’, and then modify the values populating it
> to be doubles?
>
>
>
------------------------------------------------------------------------
r322556 | marshall | 2018-01-16 06:54:36 -0800 (Tue, 16 Jan 2018) | 7 lines
Changed paths:
   M /libcxx/trunk/include/random

Change an internal table of constants for the poisson distribution from
type 'result_type' to 'double'. The only thing that we ever do with
these numbers is to promote them to 'double' and use them in a division.
For small result_types, the values were getting truncated, skewing the
results. Thanks to James Nagurne for the suggestion.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20180116/c5229ae7/attachment.html>