rocapp / fast_prng

A modified Ziggurat Algorithm for efficiently generating exponentially- and normally-distributed PseudoRandom Numbers (PRNs).

Fast PRNG: an Exponentially- and Normally-distributed PseudoRandom Number Generator

OVERVIEW

---

This code is the implementation of A modified ziggurat algorithm for generating exponentially and normally distributed pseudorandom numbers.

INSTALLATION & USAGE

---

C/C++:

Source code and functions are found in exponential.h and normal.h. Simply include this directory in the path of your compiler. For an example of usage, see benchmarks/profile.c, which profiles the exponential PRNG with the compilation command:

$ gcc -O2 -DEXPONENTIAL profile.c -lm -o exponential_profiler.out

Python:

Install via the Python Package Index:

$ pip install fast_prng

The module's functions mimic numpy.random functions of the same name.

Matlab:

Install via the command:

$ matlab/install_matlab_functions.sh

This script will install to the default MATLAB userpath, unless $USERPATH is redefined. The installed functions: fast_exprnd, fast_randn, and fast_rand behave identically to the native MATLAB functions that share the same name. Ensure that the mex compiler is configured and accessibles in the system PATH, or specify the PATH in the install_matlab_functions.sh script.

CONTACT

---

Feedback is most appreciated. If you have any questions or suggestions, please contact me at christopherdmcfarland [at] gmail [dot] com

WEBSITE

Documentation and source code can be found at https://github.com/cd-mcfarland/fast_prng

FILE MANIFEST

---

MT19937.h

Uniform PRNG based on Super-Fast Mersene Twister. Some code that supported older architectures was removed. Functions for seeding the PRNG and accessing PRNs without unnecessary bounds-checking were added.

exponential.h

Exponential PRNG to be used with C/C++ code.

normal.h

Normal PRNG to be used with C/C++ code.

shared.h

Shared constants and MACROS used by exponential.h and normal.h.

create_layers.py

Python script that calculates X, L_max, A, f(X), and epsilon (described in the main text). All calculations are done to long double precision, and then rounded to double precision.

erfl.pyx

Long double precision Error function for create_layers.py

quality_test.c

Returns raw moments of a set of PRNs. An example of usage is described in the Appendix of the main text.

benchmarks/

A collection of profiling scripts. profile.c requires a definition at compile-time (e.g. EXPONENTIAL, NORMAL) that selects the algorithm to profile. For example:

$ gcc -O2 -DNORMAL profile.c -lm -o normal.out
$ ./normal.out
Created 1000000000 Standard Normal (Modified Ziggurat) distributed PRNs with mean 2.39616e-05.
Startup time: 55 (us).
Mean execution time (per PRN): 7.987 (ns).

Competing algorithms are also provided.

matlab/

Extensions for Matlab/Octave.