LAPACK machine constants

Thread View: sci.math.num-analysis

2 messages

[T] Back to Threads [V] Tree View [G] Group Articles

2 total messages Started by Keith Briggs rba8 7088 Wed, 27 Jan 1993 06:31

#1106

Author: Keith Briggs rba8 7088
Date: Wed, 27 Jan 1993 06:31

36 lines
2010 bytes

The routine DLAMCH in LAPACK is supposed to return machine constants.
Here is the output when compiled and run with xlf on a IBM RISC/6000:

 eps      = 0.111022302462515654E-15  (machine eps)
 sfmin    = 0.22250738585072015E-307  (safe minimum)
 base     =  2.00000000000000000     
 eps*base = 0.222044604925031308E-15 
 t        =  53.0000000000000000      (digits)
 rnd      =  1.00000000000000000     
 emin     = -1021.00000000000000     
 rmin     = 0.22250738585072015E-307 
 emax     =  1024.00000000000000     
 rmax     = 0.17976931348623157E+309 

No problems!   However, after converting the type declarations to 
quadruple precision we get

 eps      =-0.1112536929253600691545116358666202E-307 
 sfmin    =-0.46566128730773925781250000000000000E-09 
 base     = -2147483648.0000000000000000000000000     
 eps*base = 0.2389154863368240330214629234492972E-298 
 t        =  34.000000000000000000000000000000000     
 rnd      = 0.00000000000000000000000000000000000E+00 
 emin     = 0.00000000000000000000000000000000000E+00 
 rmin     =-0.46566128730773925781250000000000000E-09 
 emax     =  1.0000000000000000000000000000000000     
 rmax     = -2147483648.0000000000000000000000000     


The weird (and unusable) results are caused by the IBM representation of
quadruple as two doubles.    In other words, DLAMCH has not been written
in a sufficiently portable way.

DOES ANYBODY KNOW THE CORRECT VALUES?

Keith.

Re: LAPACK machine constants

#1108

Author: Victor Eijkhout
Date: Wed, 27 Jan 1993 14:57

21 lines
1553 bytes

In article <9302717.22929@mulga.cs.mu.OZ.AU> kbriggs@mundoe.maths.mu.OZ.AU (Keith Briggs rba8 7088) writes:

   The weird (and unusable) results are caused by the IBM representation of
   quadruple as two doubles.    In other words, DLAMCH has not been written
   in a sufficiently portable way.

   DOES ANYBODY KNOW THE CORRECT VALUES?

I'm sure that that information is IBM proprietary :-)

(The source of dlamch refers to a few papers with algorithms.
You might try to find out what assumptions about the architecture
they use that are not applicable to the RS/6000)



--
Victor Eijkhout ................................ `There are also a few bugs,
Department of Computer Science .......... though not as many as I've come to
University of Tennessee ............... expect in new [MS]Windows products.'
Knoxville TN 37919 ................................ (from a software review)

Thread Navigation

This is a paginated view of messages in the thread with full content displayed inline.

Messages are displayed in chronological order, with the original post highlighted in green.

Use pagination controls to navigate through all messages in large threads.

Back to All Threads