On Tuesday, 10 June 2014 at 23:48:09 UTC, bearophile wrote:
Please stop, I am not worth that, and I don't even know how much good that generator is. So for you it's better to focus on more important matters of the new random module. Extra generators can be added later if needed.

After all the advice and help you've given me (and the rest of this community) over the course of years, it's really a pleasure to be able to offer you a small favour like this. But of course it could be fun to first run things through e.g. the TestU01 suite ...

Passing several cpu words by value for each generated value seems not very efficient. But this generator is for special situations, so a certain performance loss could be acceptable. And if the compiler is able to inline the functions, the data transfer overhead is removed, and most of the performance loss is restored (but I don't know if non-templated Phobos functions get inlined).

Well, I think it's worth experimenting with. For clarity, I wasn't suggesting modifying the existing Xorshift code, but creating a separate implementation in strongly pure style, and seeing how that differs performance-wise from what already exists.

I guess I might also consider finally getting my head round monads, and relating that to RNG design ... :-)

Reply via email to