[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

your implementation of L'Ecuyer's MRG32k3a generator

This page is part of the web mail archives of SRFI 27 from before July 7th, 2015. The new archives for SRFI 27 contain all messages, not just those from before July 7th, 2015.




Brad,

At the moment I am redesigning the SRFI-27 on random sources.

I have adopted your proposal of using the MRG32k3a method of P. L'Ecuyer
as the default implementation and have implemented it just using (allocated)
integers. Now I am putting together a fixnum/flonum implementation, too,
as the generic implementation could be terribly slow.

I have the following question regarding your code (which seems to me
quite efficient):

* Why do you use an f64vector for the seed? Is this more efficient than
creating six different variables? (for example because flonums are
allocated rather than immediate)

* Why does the fixnum/flonum implementation need flonums at all?
Wouldn't it be more efficient to use 54-bit fixnums throughout? (Which
should be there in a Scheme running on a DEC Alpha)

* What does (declare (flonum)) actually mean here?

* Why have you chosen to write a nested let _expression_ instead of a let*?
Is this a special style recognized by the compiler to get better code?

* I have installed GAMBIT here at a HP system. The definition
flonum-m-bits-plus-1 in the 'header.scm' file shows that the actual
mantissa width of FLONUMs are just 53 bits; which is exactly one bit short
of what is needed for MRG32k3a theoretically. Do you have the same
problem on your platform? If not, what platform (HW/SW) are you using?

Sebastian.

----
Dr. Sebastian Egner
Senior Scientist
Philips Research Laboratories
Prof. Holstlaan 4 (WY2)
5656 AA Eindhoven
The Netherlands
tel:       +31 40 27-43309
fax:      +31 40 27-44918
email: sebastian.egner@xxxxxxxxxxx