g77 3x faster than gfortran

VelocideX · 05-13-2009, 01:52 PM

Hi all,

I have some quite old fortran 77 code which is compatible with gfortran. The code is tens of thousands of lines, and is used for scientific purposes.

I would love to move to gfortran, but it seems exceptionally slow. Why does the code execute 3x faster with g77 than gfortran?!

Any help would be much appreciated

I am using the same compiler flags for both (-O2). Using -O3 makes no difference. I should mention I'm running OpenSuSE 11.1 64 bit (i.e. the x86-64 architecture).

VelocideX · 05-13-2009, 02:34 PM

OK I have gprof-ed the code, to see where the difference are. One routine grows massively in time expenditure, although I can't understand why as it's a simple convolution routine.

Faster version(g77):

Code:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name

 64.12      2.52     2.52    23679     0.00     0.00  vp_spvoigte__
 19.34      3.28     0.76 28574044     0.00     0.00  voigt_
  6.36      3.53     0.25       12     0.02     0.28  deriv_
  4.33      3.70     0.17 32747695     0.00     0.00  dexpf_
  3.05      3.82     0.12      877     0.00     0.00  vp_chspread__
  1.27      3.87     0.05        1     0.05     0.05  pr_sort__
  0.51      3.89     0.02        1     0.02     0.02  probks_
  0.25      3.90     0.01    23679     0.00     0.00  calcn_
  0.25      3.91     0.01      877     0.00     0.00  vp_chipconv__
  0.25      3.92     0.01       11     0.00     0.00  udchole_
  0.25      3.93     0.01        1     0.01     0.01  pldef_
  0.00      3.93     0.00   550174     0.00     0.00  ucase_

Slower version (gfortran):

Code:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name

 43.27      3.28     3.28    33885     0.00     0.00  vp_spvoigte_
 27.04      5.33     2.05     1255     0.00     0.00  vp_subchspread_
 16.36      6.57     1.24 41100608     0.00     0.00  voigt_
  6.99      7.10     0.53 528972139     0.00     0.00  dexpf_
  3.56      7.37     0.27       17     0.02     0.38  deriv_
  0.66      7.42     0.05        1     0.05     0.05  pr_sort_
  0.66      7.47     0.05        1     0.05     0.05  probks_
  0.40      7.50     0.03   478767     0.00     0.00  varythis_
  0.40      7.53     0.03        1     0.03     7.52  vp_ucoptv_
  0.26      7.55     0.02    33878     0.00     0.00  vp_archwav_
  0.13      7.56     0.01    33885     0.00     0.00  calcn_
  0.13      7.57     0.01     1255     0.00     0.00  vp_chipconv_
  0.13      7.58     0.01        1     0.01     0.01  vp_gwclinfits_
  0.00      7.58     0.00   793778     0.00     0.00  ucase_

Even dexpf, which is part of glibc, is massively slower with gfortran... =\

mostlyharmless · 05-15-2009, 04:12 PM

I don't have an answer for you, but I do have an observation.

My experience has been that difffernt compilers work well, or less well than others for specific pieces of code. Even a seemingly trivial change in the way the code is written can sometimes trigger less than ideal behaviour on the part of the compiler. It can be the most annoying kind of bug to track down.

Perhaps you can isolate the problem and bring it to the attention of the gfortran developers so that they can track down the problem. Since you know it is one routine, I'd try to make a shorter version that will run independently of the rest of the code to see whether you still see a big time lag. If it's a smaller piece of code you might get someone to look at it or run it on a differerent compiler to see what kind of problems they get or don't get.