Hello all,
As some of you may already know, monarch_dodra and I have spent
quite a lot of time over the last year discussing the state of
std.random. To cut a long story short, there are significant
problems that arise because the current RNGs are value types
rather than reference types. We had quite a lot of back and
forth on different design ideas, with a lot of helpful input from
others in the community, but at the end of the day there are
really only two broad approaches: create structs that implement
reference semantics internally, or use classes. So, as an
exercise, I decided to create a class-based std.random.
The preliminary (but comprehensive) results of this are now
available here:
https://github.com/WebDrake/std.random2
Besides re-implementing random number generators as classes
rather than structs, the new code splits std.random2 into a
package of several different modules:
* std.random2.generator, pseudo-random number generators;
* std.random2.device, non-deterministic random sources;
* std.random2.distribution, random distributions such as
uniform,
normal, etc.;
* std.random2.adaptor, random "adaptors" such as randomShuffle,
randomSample, etc.
* std.random2.traits, RNG-specific traits such as isUniformRNG
and isSeedable.
A package.d file groups them together so one can still import all
together via "import std.random2". I've also taken the liberty
of following the new guideline to place import statements as
locally as possible; it was striking how easy and clean this made
things, and it should be easy to port that particular change back
to std.random.
The new package implements all of the functions, templates and
range objects from std.random except for the old
std.random.uniformDistribution, whose name I have cannibalized
for better purposes. Some have been updated: the
MersenneTwisterEngine has been tweaked to match the corresponding
code from Boost.Random, and this in turn has allowed the
definition of a 64-bit Mersenne Twister (Mt19937_64) and an
alternative 32-bit one (Mt11213b).
There are also a number of entirely new entries.
std.random2.distribution contains not just existing functions
such as dice and uniform, but also range-based random
distribution classes UniformDistribution, NormalDistribution and
DiscreteDistribution; the last of these is effectively a
range-based version of dice, and is based on Chris Cain's
excellent work here:
https://github.com/D-Programming-Language/phobos/pull/1702
The principal weak point in terms of functionality is
std.random2.device, where the implemented random devices (based
on Posix' /std/random and /std/urandom) are really very primitive
and just there to illustrate the principle. However, since their
API is pretty simple (they're just input ranges with min and max
defined) there should be plenty of opportunity to improve and
extend the internals in future. Advice and patches are welcome
for everything, but particularly here :-)
What's become quite apparent in the course of writing this
package is how much more natural it is for ranges implementing
randomness to be class objects. The basic fact that another
range can store a copy of an RNG internally without creating a
copy-by-value is merely the start: for example, in the case of
the class implementation of RandomSample, we no longer need to
have complications like,
@property auto ref front()
{
assert(!empty);
// The first sample point must be determined here to avoid
// having it always correspond to the first element of the
// input. The rest of the sample points are determined
each
// time we call popFront().
if (_skip == Skip.None)
{
initializeFront();
}
return _input.front;
}
that were necessary to avoid bugs like
https://d.puremagic.com/issues/show_bug.cgi?id=7936; because the
class-based implementation copies by reference, we can just
initialize everything in the constructor. Similarly, issues like
https://d.puremagic.com/issues/show_bug.cgi?id=7067 and
https://d.puremagic.com/issues/show_bug.cgi?id=8247 just vanish.
Obvious caveats about the approach include the fact that classes
need to be new'd, and questions over whether allocation on the
heap might create speed issues. The benchmarks I've run (code
available in the repo) seem to suggest that at least the latter
is not a worry, but these are obviously things that need to be
considered. My own feeling is that ultimately it is a
responsibility of the language to offer nice ways to allocate
classes without necessarily relying on new or the GC.
A few remarks on design and other factors:
* The new range objects have been implemented as final classes
for
speed purposes. However, I tried another approach where the
RNG
class templates were abstract classes, and the individual
parameterizations were final-class subclasses of those rather
than aliases. This was noticeably slower. My OO-fu is not
really
sufficient to explain this, so if anybody can offer a
reason, I'd
be happy to learn it.
* A design question I considered but have not yet pursued:
since at
least two functions require passing the RNG as the first
parameter
(dice and discreteDistribution), perhaps this should be made
a
general design pattern for everything? It would make it
harder to
adapt code using the existing std.random but would create a
useful
uniformity.
* I would have liked to ensure that every random distribution
had
both a range- and function-based version. However, I came
to the
conclusion that solely function-based versions should be
avoided
if either (i) the function would need to maintain internal
state
between calls, or (ii) the function would need to allocate
memory
per call. The first is why for example NormalDistribution
exists
only as a class/range. The second might in principle raise
some
objections to dice, but as dice seems to be a reasonably
standard
function, I kept it in.
* It might be good to implement helper functions for the
individual
RNGs (e.g. just as RandomSample has a randomSample helper
function
to deliver instances, so Mt19937 could have a corresponding
mt19937 helper function returning Mt19937 instances seeded
in line
with helper function parameters).
* Those with long memories may recall that when I originally
wrote
up my NormalDistribution code, it was written to allow
various
"normal engines" to be plugged in; mine was Box-Muller, but
jerro
also contributed a Ziggurat-based engine. This could still
be
provided here, although my own inclination is that it's
probably
best for Phobos to provide one single
good-for-general-purpose-use
implementation.
Known issues:
* While every bugfix I've made in the course of implementing
this
package has been propagated back to std.random where
possible,
this package is missing some of the more recent improvements
to
std.random by other people (e.g. I think it's missing Chris
Cain's
update to integer-based uniform()).
* The unittest coverage is overall pretty damn good, but there
are
weak spots in std.random.distribution and std.random2.device.
Some of the "unittests" in these cases are no more than basic
developer sanity checks that print results to console, and
need
to be replaced by well-defined, silent-unless-failed
alternatives.
* Some of the .save functions are implemented with the help of
rather
odd private constructors; it would probably be much better
to redo
these in terms of public this(typeof(this)) constructors.
* The random devices _really_ need to be better. Consider the
current
versions as placeholders ... :-)
Finally, a note on authorship: since this is still based very
substantially on std.random, I've made an effort to check git
logs and ensure that authors and copyright records (and dates of
contribution) are correct. My general principle here has been
that listed authors should only include those who've made a
substantial contribution (i.e. whole functions, large numbers of
unittests, ...), not just various 1-line tweaks. But if anyone
has any objection to any of the names, dates or other credits
given, or if anybody would like their name removed (!), just let
me know.
I owe a great debt of gratitude to many people here on the
forums, and monarch_dodra in particular, for a huge amount of
useful discussion, advice and feedback that has made its way into
the current code. Thank you all for your time, thoughts, ideas
and patience.
Anyway, please feel free to review, destroy and otherwise do fun
stuff with this module. I hope that some of you will find it
immediately useful, but please note that feedback and advice may
result in breaking changes -- this is intended to wind up in
Phobos, so it really needs to be perfect when it does so. Let's
review it really well and make it happen!
Thanks and best wishes,
-- Joe
