Skip to content

64-bit programming

4 messages · Helgi Tomasson, Brian Ripley

Original
Helgi Tomasson
# 
Has R been ported to  64-bit platforms?  E.g. as in FORTRAN when you
define
double precision number as:

REAL*16

I have heard that some SUN computers already offer 64-bit architecture
but I am unsure
about the GNU support  for that hardware.  I am seeking increased
accuracy  in floating
point operations.

regards

Helgi

--
Helgi Tomasson                                       FAX:  354-552-6806
University of Iceland                                PHONE:354-525-4571
Faculty of Economics and Business Administration     email:helgito at rhi.hi.is
Oddi v/ Sturlugotu
IS-101 Reykjavik
ICELAND



-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
Brian Ripley
#
On Tue, 11 Sep 2001 helgito at hi.is wrote:

            

      
      
      
    

        
We run R in 64-bit mode on Suns under the native compiler. However,
the double type in 64-bit mode is 8 bytes, not 16 bytes, and that's
part of the agreement between manufacturers for 64-but platforms as we
understand it:

int - 4 bytes
long - 8 bytes
double - 8 bytes

The 64-bit refers to the pointer size and perhaps to buses and to the
preferred size for integer arithmetic. It does not refer to the FPU,
and FPUs have been 64 bits or greater (e.g. i686) for about 15 years.

Why do you think REAL*16 will give you increased accuracy?  Accuracy has a
lot more to do with the algorithms used and their convergence criteria.
As R is a statistical package, numerical accuracy is almost always
secondary to data uncertainty.

At least the build of gcc 3.0.1 we have does not run in -m64 mode.
The potential benefit of 64-but mode is to address large amounts of
virtual memory from a single process, but R has internal limits there
still, I believe.

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Helgi Tomasson
    

    

      
      #
    

  


  

          

      
      
      
    

        
I tested a Solaris fotran compiler on a Sun and it allowed REAL*16 whereas
a GNU fortran on the same machine did not.  It seems to me that that the
increased
numerical accuracy compared to REAL*8 on an Intel-86 machine was not
proportional. The Intel machines have a numerical stack of lenght 80 so if
the Fortran program is well optimized you should not loose as as many digits
as you might expect.
REAL*8 gives about 15 digits in floating-point.  It was my hope that REAL*16
would give around 30.  In my simple experiment it seemed that I only got
about 20-25.
Maybe the Sun or the compiler does not use a stack as the Intel-86 machines.

I want to add several hundred thousand numbers, and their squares, which some
are range from  maybe 10^(-6) to 1 or more.
With crude programming I fear that I am loosing important digits.

I was wondering whether I should apply for a Sun station or wait further
development.

regards

Helgi

            

      
      
      
    

        
--
Helgi Tomasson                                       FAX:  354-552-6806
University of Iceland                                PHONE:354-525-4571
Faculty of Economics and Business Administration     email:helgito at rhi.hi.is
Oddi v/ Sturlugotu
IS-101 Reykjavik
ICELAND



-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

  
  


  


      

    

    

      
      

        


  

        

      Brian Ripley
    

    

      
      #
    

  


  

      
On Tue, 11 Sep 2001 helgito at hi.is wrote:

        

            

      
      
      
    

        
Better to learn some numerical analysis.  For example you could sum the
smallest first.  If you want centred moments, centre first ....
There is a whole science out there of getting accurate results with low
precision (it was needed where the only fast arithmetic was a few
hexadecimal digits).  That could be tricky and even needed iterative
refinement, but double precision swept all that away at least of
statistical purposes (except perhaps for people who think the normal
equations are the way to solve least-squares problems).