Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!usc!cs.utexas.edu!samsung!caen!news.cs.indiana.edu!ariel.unm.edu!triton.unm.edu!prentice
From: prentice@triton.unm.edu (John Prentice)
Newsgroups: comp.lang.fortran
Subject: vectorization question reposed
Message-ID: <1991Mar30.142903.5225@ariel.unm.edu>
Date: 30 Mar 91 14:29:03 GMT
Organization: University of New Mexico, Albuquerque
Lines: 54
Originator: prentice@triton.unm.edu

My thanks to those people who responsed to my query about how to
collapse a three tier array into one.  Unforunately, none of the
responses were quite what I needed (I neglected to specify the constraints:
1) Fortran 77, 2) clean, 3) portable, 4) must calculate i,j,k, and 5)
vectorizeable).  The most serious of my ommissions was to neglect to
say that I need the i,j,k indices.  A solution I came up with that is
vectorizeable is:


      program test
      parameter (imax=100,jmax=150,kmax=200)
      dimension a(imax,jmax,kmax)
      do 10 n=1,imax*jmax*kmax
          i=mod(n,imax)
          if(i.eq.0) i=imax
          j=n/imax+1
          if(j*imax.gt.n) j=j-1
          k=n/(imax*jmax)+1
          if(k*imax*jmax.gt.n) k=k-1
          a(i,j,k)= some function of i,j,k
   10 continue
      end

I would now appreciate it if people would tell me any pitfalls they
see to this.  In particular, is the compiler (cft77 4.0.2 is what I
am using on a YMP) smart enough to realize that my stride is 1 here
and therefore not do a gather/scatter operation when it is not necessary?

One other question.  What is the size of the memory banks on the 
YMP and hence what strides should you avoid?  Also, suppose you 
have the following code:


      subroutine ralph (a,b,c,indx)
      dimension a(100),b(100),c(100),indx(100)
      do 10 n=1,100
	  a(n)=b(indx(n))*c(indx(n))
   10 continue
      end

where indx(n) maps into 1 to 100.  To vectorize this, the system
has to do a gather/scatter.  If indx(i) and indx(i+1) happen to 
point to locations in the same memory bank, will the system do
anything to reorder the references or will this indeed cause
a bank conflict?  What about cache?  Does the YMP have a cache and
if so, what rules are there for maximizing performance with it?

Many thanks.

John
-- 
John K. Prentice    john@unmfys.unm.edu (Internet)
Dept. of Physics and Astronomy, University of New Mexico, Albuquerque, NM, USA
Computational Physics Group, Amparo Corporation, Albuquerque, NM, USA