RAN(3I) Last changed: 1-6-98
NAME
_ranf, RANF, RANGET, RANSET - Computes pseudo-random numbers
SYNOPSIS
C/C++:
#include <intrinsics.h>
double _ranf(void);
Fortran:
RANF()
RANGET ([I=]i)
RANSET ([K=]k)
Fortran on UNICOS/mk systems only:
RANSET ([K=]k [, [J=]j])
IMPLEMENTATION
C/C++: UNICOS and UNICOS/mk systems
Fortran: UNICOS, UNICOS/mk, and IRIX systems
STANDARDS
C/C++: Cray Research extension
Fortran: CF90 compiler extension to Fortran 90
DESCRIPTION
The _ranf and RAND functions return a pseudo-random floating-point
number in the range of 0.0 < x < 1.0.
The RANGET intrinsic procedure returns a seed from the random number
seed table.
The RANSET function establishes a seed in the random number seed
table.
C/C++ NOTES
Because _ranf is an intrinsic function, no externally visible library
function is available for it.
FORTRAN NOTESRANF obtains the first or next in a series of pseudo-random numbers.
Parenthesis are required, that is: var = RANF(). If an argument is
supplied, it is ignored.
The RANGET intrinsic function obtains a seed. It can be called as a
function or a subroutine; it is recommended that this routine be used
as a function. RANGET has an optional integer argument. If present,
the argument is set to the seed. The argument for RANGET is as
follows:
i An integer of default kind (KIND=8). If present, RANGET returns
the seed in i.
On UNICOS and UNICOS/mk systems, the RANSET intrinsic function
establishes a seed by using the lower 48 bits of the argument. The
result type is typeless. If no argument or a zero argument is
supplied, the seed is reset to an initial default value. When the
seed of the random number generator is reset, RANSET does not store
the supplied argument as the first value in the buffer of the random
number seeds. If an argument is supplied, the lower 48 bits are used
as the random-number seed. The rightmost bit is always set to 1.
The RANSET arguments are as follows:
k An optional integer, real, or Boolean argument of default kind
(KIND=8) The range of argument k is |k| < inf, where inf is as
follows:
2450
* On UNICOS systems, infinity is approximately 10 .
308
* On UNICOS/mk and IRIX systems, infinity is approximately 10 .
j An optional integer argument that, if specified, is used as a
count for skipping the first section of sequential random
numbers. You can use this to create a complete sequence of
random numbers while running on many PEs by breaking up the
sequence into subsequences and using RANSET() to get each
subsequence started in the correct location.
The names of these intrinsics cannot be passed as arguments.
THE RANF ALGORITHM
In Fortran on CRAY C90 systems, the random number generation algorithm
uses the following two equations:
* S(n+1) = M1*S(n) mod 2**48
* S(n+64) = M64*S(n) mod 2**48
Each S(i) is the ith seed.
The first equation is used to generate the first 128 random numbers
and store them in a table if a call to RANSET() was done. Otherwise,
the table contains the first 128 random numbers from the default seed.
The second equation is used because it vectorizes.
The default seed is S(0) = 1274321477413155 (octal).
The operations M1*S(n) and M64*S(n) are done as integer
multiplications in such a way to preserve the lower 48 bits. It is
the lower 48 bits that make the next random number seed and are used
in the return value.
The return value (random number) is the newly generated seed with an
exponent of 40000 (octal) added. This is normalized before exit.
The multiplier M1 is 1207264271730565 (octal) and is related to M64 by
the following expression:
M64 = lower 48 bits of M1**64 = 7027423346125401 (octal).
For example, the following Fortran program, when compiled with the
f90 -i 64 option on a CRAY C90 system, is the equivalent of the RANF
function:
cc
c THIS IS A FORTRAN VERSION OF RANF() FUNCTION
cc
REAL FUNCTION RANF()
REAL NORM
INTEGER MHI,MLO,EXPO,SEED,SEEDHI,SEEDLO
DATA SEED/1274321477413155B/
SAVE SEED
MHI = 12072642B
MLO = 71730565B
EXPO = SHIFTL(40000B,48)
SEEDHI = SHIFTR(AND(SHIFTL(77777777B,24),SEED),24)
SEEDLO = AND(77777777B,SEED)
SEED = AND(7777777777777777B,SEEDLO*MLO+
1 SHIFTL(SEEDLO*MHI+SEEDHI*MLO,24))
RANF=NORM(OR(EXPO,SEED),0.0)
RETURN
END
cc
c THIS IS HERE TO NORMALIZE THE FLOATING POINT RESULT
cc
REAL FUNCTION NORM(X,Y)
REAL X,Y
NORM = X+Y
RETURN
END
On IRIX systems, RANF uses a 64-bit linear congruential generator with
a default seed of 1274321477413155 (octal).
THE RANF REPEAT PERIOD
In Fortran, the period of RANF() is 2**46. If you need to insure that
two random number ranges do not overlap, you can determine this
empirically by generating the two sets of numbers and comparing them
against one another, and also against themselves, for an overlap. It
should be noted, however, that when using RANSET to set the random
number seed, the algorithm used always rounds up even-numbered seeds
to the nearest odd-numbered seed (that is, the right most bit is
always set to one). Some adjacent pairs of seeds will generate
exactly the same set of random numbers. For example, seeds 4 and 5
will generate the same set of random numbers.
RANF AND MULTITASKING
In Fortran, the random number generator uses static memory storage for
the random number seed table, so the RANF, RANSET, and RANGET
functions must be protected (locked) when called from a multitasked
program.
RANF generates a set of random numbers such that each random number
depends on the previous random number for its value. Thus, depending
on the order in which the tasks calling RANF execute, a different set
of random numbers will be returned to each task. It cannot be
guaranteed that each task will get a distinct and reproducible set of
random number values.
RETURN VALUES
_ranf and RANF return a 64-bit floating-point number in the range
0.0 < x < 1.0.
RANGET returns a 64-bit integer result.
RANSET returns a 64-bit typeless result.
EXAMPLES
The following examples are written in Fortran:
DO 10 I=1,10
10 RANDOM(I)=RANF()
CALL RANGET(iseed1)
C or
iseed=RANGET(ivalue)
CALL RANSET(ivalue)
C or
dummy=RANSET(ivalue)SEE ALSORANDOM_NUMBER(3I), RANDOM_SEED(3I)rand(3C) in the UNICOS System Libraries Reference Manual, publication
SR-2080
A complete list of C/C++ intrinsic functions available on Cray
Research systems is in the Cray C/C++ Reference Manual, publication SR
-2179.
Intrinsic Procedures Reference Manual, publication SR-2138, for the
printed version of this man page.
