REFERENCE IMPLEMENTATIONS FOR SRFI-27 "Sources of Random Bits"
==============================================================
Sebastian.Egner@philips.com, 10-Apr-2002.
Files
-----
readme - this file
mrg32k3a.scm - generic parts of P. L' Ecuyer's MRG32k3a PRGN
mrg32k3a-a.scm - core generator in Scheme integers
mrg32k3a-b.c - core generator in C doubles for Scheme 48
mrg32k3a-c.scm - core generator in Gambit [Scheme] flonums
srfi-27-a.scm - Scheme 48 package definition for Scheme-only impl.
srfi-27-b.scm - Scheme 48 package definition for C/Scheme impl.
srfi-27-c.scm - Gambit definition for Scheme-only impl.
conftest.scm - confidence tests for the implementation
Implementations
---------------
The implementation has been factored into three parts.
One part implements the core generator, one part provides
the more generic functionality as specified in the SRFI,
and one part combines the parts and provides the interface
as specified in the SRFI.
a) A Scheme-only implementation for Scheme 48 0.57:
srfi-27-a.scm
mrg32k3a-a.scm
mrg32k3a.scm
This implementation uses 54-bit Scheme integers for all
arithmetics of the generator. The result are Scheme integers
and inexact Scheme numbers when floating point values are
requested.
The implementation is slow but tries to stay away from
unportable features as much as possible.
b) An implementation in Scheme 48 0.57 and ANSI-C:
srfi-27-b.scm
mrg32k3a-b.scm
mrg32k3a.scm
This is a more realistic implementation using C's (double)
datatype for the core generator and 54-bit Scheme integers
for the more infrequent operations on the state like the
random-source-pseudo-randomize! operation.
This implementation is meant as an example for a realistic
native code implementation of the SRFI. Performance is good.
c) A Scheme-only implementation for Gambit 3.0:
srfi-27-c.scm
mrg32k3a-c.scm
mrg32k3a.scm
This implementation uses Gambit's 64-bit flonums. It is
entirely written in Scheme but uses a few special features
of the Gambit system to tell the compiler.
This implementation is meant as an example for a realistic
Scheme implementation using flonums in Scheme and no C-code.
Performance is good when the code is used in compiled form;
the implementation has been optimized by Brad Lucier. This
has resulted in a subtantial performance gain.
