On Tuesday, 10 June 2014 at 23:08:33 UTC, Chris Cain wrote:
I had an opportunity to give the entire code a good once over read and I have a few comments.

Thanks! :-)

1. Biggest thing about the new hap.random is how much nicer it is to actually READ. The first few times I went through the current std.random, I remember basically running out of breath. hap.random was almost a refreshing read, in contrast. I'm guessing it has a lot to do with breaking it down into smaller, more manageable pieces. Regardless, good work on that. I suspect it'll make it easier to contribute to in the future.

That's great to hear, as it was a design goal. I think there will probably at some point be a need to separate things further (e.g. std.random.generator will probably have to be separated as will std.random.distribution) but always keeping the principle of implementing packages to make it possible to just "import hap.random" (or "import hap.random.generator", or whatever).

2. Something I'd really like to see is for the seed-by-range functions to take the range by reference instead of by value to ensure that the seed values used are less likely to be used in another RNG inadvertently later. Basically, I envision a similar problem with seedRanges as we currently have with RNGs where we have to make sure people are careful with what they do with the ranges in the end. This should cover use cases where users do things like `blah.seed(myEntropyRange.take(3))` as well, so that might take some investigation to figure out how realistic it would be to support.

Yea, that's an interesting point. I mean, you'd hope that myEntropyRange would be a reference type anyway, but every little helps :-)

3. I'd also REALLY like to see seed support ranges/values giving ANY type of integer and guarantee that few bytes are wasted (so, if it supplies 64-bit ints and the generator's internal state array only accepts 32-bit ints, it should spread the 64-bit int across two cells in the array). I have working code in another language that does this, and I wouldn't mind porting it to D for the standard library. I think this would greatly simplify the seeding process in user code (since they wouldn't have to care what the internal representation of the Random state is, then).

That would be very cool.  Can you point me at your code examples?

4. I'd just like to say the idea of using ranges for seeds gets me giddy because I could totally see a range that queries https://random.org for true random bits to seed with, wrapped by a range that zeroes out the memory on popFront. Convenient and safe (possibly? Needs review before I get excited, obviously) for crypto purposes!

The paranoiac in me feels that anything that involves getting random data via HTTPS is probably insecure crypto-wise :-) However, I think sourcing random.org is a perfect case for an entry in hap.random.device. I think the best thing to do would probably be to offer a RandomOrgClient (which offers a very thin API around the random.org HTTP API) and then to wrap that in a range type that uses the client internally to generate random numbers with particular properties.

5. Another possible improvement would be something akin to a "remix" function. It should work identically to reseeding, but instead of setting the internal state to match the seed (as I see in https://github.com/WebDrake/hap/blob/master/source/hap/random/generator.d#L485), remixing should probably be XOR'd into the current state. That way if you have a state based on some real entropy, you can slowly, over time, drip in more entropy into the state.

Also a very interesting suggestion. Is there a standard name for this kind of procedure?

6. I'd like to see about supporting xorshift1024 as well (described here: http://xorshift.di.unimi.it/ and it's public domain code, so very convenient to port ... I'd do it too, of course, if that seems like an okay idea). This is a really small thing because xorshift1024 isn't really much better than xorshift128 (but some people might like the idea of it having significantly longer period).

Fantastic, I will see about implementing those. I wasn't previously aware of that work, but I _was_ aware that the standard Xorshift generators have some statistical flaws, so it's great to have some alternatives. It should be straightforward to implement things like XorshiftP128 or XorshiftS1024 and XorshiftS4096 (using P and S in place of + and *).

With these in place we might even be able to deprecate the old Xorshift generators.

Just for clarity, here's how I see things rolling out for the future:

* First goal is to ensure the existing codebase "plays nice" with people's programs and that it works OK with dub, rdmd, etc. and doesn't have any serious architectural or other bugs. The 1.0.0 release will not have any new functionality compared to what is
    in place now.

  * Once it seems to be reasonably stable then work can begin on a
    1.x release series that brings in successive pieces of new

Done :) ... if I get a response, I'll make sure to incorporate everything said.

Great, let me know how that goes. :-)

Reply via email to