Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!cis.ohio-state.edu!tut.cis.ohio-state.edu!ucbvax!WATSON.IBM.COM!jbs
From: jbs@WATSON.IBM.COM
Newsgroups: comp.arch
Subject: Fujitsu linear recurrence vector instruction
Message-ID: <9105140213.AA15244@ucbvax.Berkeley.EDU>
Date: 14 May 91 02:00:34 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Lines: 11


          I believe what Fujitsu has actually done is provide a vector
instruction to compute x(j)=a(j)*x(j-1)+b(j) (j=1,length vector register)
given a,b and x(0).  This runs slower than most vector instructions be-
cause it can't be pipelined (ie the add for x(j) feeds the multiply for
x(j+1)).  It still may be advantageous however if it lies in a loop which
can otherwise be vectorized since otherwise time is wasted moving stuff
between the vector and scalar units.
          Of course I may have this all wrong.  Perhaps someone from
Fujitsu can give us the real story.
                          James B. Shearer