Polyhedron benchmark results on AMD64

[Steve Kargl]

On Fri, Feb 09, 2007 at 11:31:10PM +0100, Dominique Dhumieres wrote:
> 
> Steve Kargl wrote:
> 
> > I looked at gas_dyn a long time ago.  My conclusion was that
> > gfortran needed to do inter-procedure optimizations to make
> > any big gains.
> 
> I cannot comment on this, but I had a look to the code and did not find 
> any obvious way to hand optimize the code.
> 

A quick look with a profiler shows:

gfc4x -pg -O2 -march=opteron -funroll-loops -o z gas_dyn.f90

  %   cumulative   self              self     total           
 time   seconds   seconds    calls  ms/call  ms/call  name    
 46.3      11.67    11.67    20002     0.58     0.58  eos_ [3]
 15.4      15.55     3.88    20001     0.19     0.19  chozdt_ [4]
 11.0      18.31     2.76        0  100.00%           _gfortran_minloc0_4_r4 [5]
 10.3      20.90     2.59        0  100.00%           _mcount [6]
  6.7      22.59     1.69  5603564     0.00     0.00  area_ [7]

The guilty code in eos_ is this:

      IF (SHEAT>0.0 .AND. CGAMMA>0.0) THEN
          TEMP(:NODES) = IENER(:NODES)/SHEAT
          PRES(:NODES) = (CGAMMA - 1.0)*DENS(:NODES)*IENER(:NODES)
          GAMMA(:NODES) = CGAMMA
          CS(:NODES) = SQRT(CGAMMA*PRES(:NODES)/DENS(:NODES))
      ELSE

With naive and manual loop fusion:

      IF (SHEAT>0.0 .AND. CGAMMA>0.0) THEN
        do sk = 1, nodes
          TEMP(sk) = IENER(sk)/SHEAT
          PRES(sk) = (CGAMMA - 1.0)*DENS(sk)*IENER(sk)
          GAMMA(sk) = CGAMMA
          CS(sk) = SQRT(CGAMMA*PRES(sk)/DENS(sk))
        end do
      ELSE

I see 

  %   cumulative   self              self     total           
 time   seconds   seconds    calls  ms/call  ms/call  name    
 43.4      10.65    10.65    20002     0.53     0.53  eos_ [3]
 15.9      14.56     3.91    20001     0.20     0.20  chozdt_ [4]
 11.7      17.44     2.88        0  100.00%           _gfortran_minloc0_4_r4 [5]
 11.0      20.13     2.69        0  100.00%           _mcount [6]

Restoring the code to first snippet and adding -ftree-vectorize, I see

  %   cumulative   self              self     total           
 time   seconds   seconds    calls  ms/call  ms/call  name    
 36.8       6.63     6.63    20002     0.33     0.33  eos_ [3]
 15.4       9.42     2.78        0  100.00%           _gfortran_minloc0_4_r4 [4]
 14.3      11.99     2.58        0  100.00%           _mcount [5]
 10.7      13.92     1.92    20001     0.10     0.10  chozdt_ [6]

-- 
Steve