john@ghostwheel.unm.edu (John Prentice) (02/24/91)
I have a question for the Cray guru's out there. When you execute
a conditional vector merge in Cray Fortran, what is really happening?
For example, consider:
a=cvmgp(f(x),g(x),h(x))
where f,g, and h are some arithmetic functions. Does it compute both
f(x) and g(x) and then return only one depending on what the value
of h(x) is or does it only compute f(x) or g(x) as required depending
on the value of h(x)?
Appreciate any help.
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, USAbloepfe@bernina.ethz.ch (Bruno Loepfe) (02/25/91)
In article <1991Feb24.050433.21694@ariel.unm.edu> john@ghostwheel.unm.edu (John Prentice) writes: >I have a question for the Cray guru's out there. When you execute >a conditional vector merge in Cray Fortran, what is really happening? >For example, consider: > > a=cvmgp(f(x),g(x),h(x)) > >where f,g, and h are some arithmetic functions. Does it compute both >f(x) and g(x) and then return only one depending on what the value >of h(x) is or does it only compute f(x) or g(x) as required depending >on the value of h(x)? > Yes, it does indeed. It evaluates f(x) to some vector register, evaluates g(x) to another vector register, creates a bit mask of (h(x) .GT. 0.0) and then stores the aproppriate value in a, according to the mask. This means, if for example h(x) > 0.0 for 90% or more of all x, f(x) = const and g(x) = VERY time consuming function of x, such a loop might execute slower in vector mode than in scalar mode... ------------------------------------------------------------------------------- Bruno Loepfe u36@czheth5a.bitnet Computing Center loepfe@rz.id.ethz.ch Federal Institute of Technology bloepfe@ethz.uucp Zuerich, Switzerland ..!uunet!mcsun!ethz!bloepfe (UUCP) -- ------------------------------------------------------------------------------- Bruno Loepfe u36@czheth5a.bitnet Computing Center loepfe@rz.id.ethz.ch Federal Institute of Technology bloepfe@ethz.uucp
jerry@violet.berkeley.edu (Jerry Berkman) (03/03/91)
In article <1991Feb24.050433.21694@ariel.unm.edu> john@ghostwheel.unm.edu (John Prentice) writes: >I have a question for the Cray guru's out there. When you execute >a conditional vector merge in Cray Fortran, what is really happening? >For example, consider: > > a=cvmgp(f(x),g(x),h(x)) > > ... >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 As others have noted, first all three expressions are computed, then a mask is formed, and then the result is returned (to be stored) depending on the mask. The cvmg* family of functions are all typeless, consider: do i = 1, n x(i) = cvmgt( x(i), i-5, y(i).gt.z ) end do this is almost the same as: do i = 1, n if( y(i).gt.z ) x(i) = i-5 end do except that since cvmgt() is typeless, no type conversion is made when storing i-5 in x(i). In a lot of cases, Cray's CFT77 compiler will compile if()s the same as cvmg*() calls. It's generally better to stay standard and avoid these except when the compiler gets confused and won't vectorize code unless you use them. - Jerry Berkman, U.C. Berkeley