Re: [boost] [random] new threefry random engine

Subject: Re: [boost] [random] new threefry random engine
From: Steven Watanabe (watanabesj_at_[hidden])
Date: 2014-05-02 11:43:24

AMDG

On 04/29/2014 10:00 AM, John Salmon wrote:
> 
> I've dusted off the code I wrote a couple of years ago and republished it,
> now on github:
> 
> https://github.com/DEShawResearch/Random123-Boost
> 

For this to be included in Boost.Random, it must be
released under the Boost Software License.

> <snip>
> 
> CBRNGs overcome all these problems.  With CBRNGs, the code looks like
> this (see libs/random/examples/counter_based_example.cpp for a fully
> worked example):
> 
>     using namespace boost::random;
>     typedef threefry<4, uint32_t> Prf;
>     normal_distribution nd;
>     Prf::key_type key = {seed};
>     Prf prf(key);  // seed may come from command line
>     for(size_t i=0; i<atoms.size(); ++i){
>         float boltzmannfactor = sqrt(kT/atoms[i].mass);
>         Prf::domain_type d = {atoms[i].id, timestep, THERMALIZE_CTXT};
>         counter_based_urng<Prf> cbrng(prf, d);
>         nd.reset();
>         atoms[i].vx = boltzmannfactor*nd(cbrng);
>         atoms[i].vy = boltzmannfactor*nd(cbrng);
>         atoms[i].vz = boltzmannfactor*nd(cbrng);
>     }
> 

Is there a good reason why this shouldn't be written as:

for(...) {
  ...
  std::seed_seq s{ seed, atoms[i].id, timestep, THERMALIZE_CTXT };
  counter_based_engine<Prf> cbrng(s);
  ...
}

I don't really like making the details of the
algorithm (i.e. key and domain) part of the
primary interface.

> Let's consider the code changes between the two fragments:
> 
> <snip>
> 
> - Since it models a URNG, cbrng can be passed as an argument to the
> normal_distribution, nd.  In order to make each atom independent,
> 
>         nd.reset()
> 
>     is called each time through the loop.
> 
> <snip>

FWIW, I don't really think that reset was a good idea
in the first place, and it's currently a no-op in all
distributions.  To run the loop in parallel, you'd need
to make a copy of nd in every thread, anyway.

> counter_based_urng
> ------------------
> 
> The counter_based_urng class uses the pseudo-random property of PRFs
> to perform random number generation in accordance with the
> requirements of a UniformRandomNumberGenerator.  <snip>
> 
> counter_based_engine
> --------------------
> 
> The counter_based_engine class is a templated adapter class that
> models a bona fide RandomNumberEngine.  <snip>
> 

Is there really a good reason to make two separate
templates?  A RandomNumberEngine is a URNG, by definition.

In Christ,
Steven Watanabe