Re: Inline DOT_PRODUCT revisited - to libgfortran/m4 gurus

[Paul Thomas <paulthomas2 at wanadoo dot fr>]

Steven,

How are you?  It's good to hear from you.

> See the original message from November 2005 here:
> http://gcc.gnu.org/ml/gcc-patches/2005-11/msg00686.html
>
> On Thursday 10 November 2005 09:13, Paul Thomas wrote:
>
> > an inline version of DOT_PRODUCT, which is never slower that the
> > original library version and is very much faster for small vectors.
>
>
> Amazingly your patch still applied almost cleanly, and it seems to do
> what it is supposed to do. The test case you had for dot_product was
> useless because it got optimized to an empty loop (which for whatever
> reason was not removed itself). The new test case is attached below.
> As is the updated patch, which is just a re-diff.

The empty loop was forced and the timing subtracted from the time to do 
the loop with dot_product in it.  I agree that the difference is small 
but it was all part of a cunning plan.....

.... the correct outcome of which is attached and some results 
below(obtained on the system that the regtesting is being carried out 
on).  It may be seen that gfortran-4.1(with library dot_product) begins 
to win out for larger arrays but that gfortran-4.2(with inline) is a 
clear winner for small arrays; ie. the inline has a smaller overhead but 
seems to be slightly less well optimised for each product and sum.  I 
will submit with figures obtained on a quiet system.

I agree that the patch should be applied - I was a bit surprised that 
folk lost interest at the time and that my call for a bit of help in 
expunging all memory of dot_product from the library went unheeded.  I 
have done all the necessary on the library, seem to have regenerated 
correctly and am in the midst of regtesting.  I am not entirely sure 
that I'll have time to write the changelog entries this evening but will 
commit a definitive patch tomorrow morning.

Paul

[prt@localhost dot_product]# /svn-4.2/bin/gfortran  -O3 
-fdump-tree-original dottest.f90;./a.out
DOT_PRODUCT test
  array length        time(ns)

       4               32.80
       8               16.40
      16              115.60
      32              219.00
      64              427.80
     128              818.00
     256             1715.70
     512             2783.90
    1024             6160.70
[prt@localhost dot_product]# ifort  -O3 -fdump-tree-original 
dottest.f90;./a.out
ifort: Command line warning: ignoring unknown option '-fdump-tree-original'
DOT_PRODUCT test
  array length        time(ns)

       4               21.90
       8               31.40
      16               95.80
      32              188.70
      64              342.40
     128              713.00
     256             1285.90
     512             2768.60
    1024             5358.10
[prt@localhost dot_product]# export 
LD_LIBRARY_PATH=/svn-4.1/lib:/opt/intel/fc/9.0/lib
[prt@localhost dot_product]# /svn-4.1/bin/gfortran  -O3 
-fdump-tree-original dottest.f90;./a.out
DOT_PRODUCT test
  array length        time(ns)

       4               71.00
       8               77.40
      16              153.20
      32              189.50
      64              388.10
     128              704.40
     256             1414.40
     512             2904.90
    1024             5477.70


  implicit real(8) (x-z)
  integer(8), parameter :: n= 10000000
  real(8)               :: t1 = 0.0, t2 = 0.0, t3 = 0.0
  real(8)               :: dt1
  real(4)               :: xin(1024), yin(1024), z(n)
  integer(8)            :: i, j, m

  xin = (/(dble(j),j=1,1024)/)
  yin = xin

  print '(a/a/)', "DOT_PRODUCT test", "   array length        time(ns)"
  do l = 2,10

    m = 2**l
    t1 = tim ()

    do i = 1, n
      z(i) = dot_product (xin(1:m), yin(1:m))
    end do

    t2 = tim ()

    do i = 1, n
      z(i) = 0d0
    end do

    t3 = tim ()

    dt1 = 1e9_8 * (2d0*t2 - t1 - t3) / (dble (n))
    print '(5x,i4,12x,f8.2,12x,f8.2)', m, dt1
  end do

contains

! Do not use secnds - various compilers do not support it.
  real(8) function tim ()
    character(12)         :: ddate, dtime, zone
    integer(4)            :: elems (8)
    call date_and_time (ddate, dtime, zone, elems)
    tim = 60.0_8 * elems(6) + elems(7) + 0.001_8 * elems(8)
  end function tim

  end