Improve performance of Fortran index calculations

Issue #235 closed
Erik Schnetter created an issue

In Fortran, Cactus currently declares grid functions e.g. as (this is the expansion of DECLARE_CCTK_ARGUMENTS)

REAL*8 gxx (X0metric,X1metric,X2metric)

where X0metric etc. are integers passed into the routine. Each grid function group has its own, independent size. This has two disadvantages:

  1. The compiler does not know that all grid functions have the same size (namely cctk_lsh), and thus has to perform array index calculations separately for each group
  2. The argument list is longer than neded

The enclosed patch declares grid functions via cctk_lsh. Grid arrays are still declared independently.

This reduces the code size of e.g. GRHydro/GRHydro_Tmunu.F90 from 6836 to 6241 bytes on my system. I have not attempted to measure a performance difference.

Keyword:

Comments (7)

  1. anonymous
    • removed comment

    Why is there the conditional in GridFuncStuff.pl (new line 394):

    + if ($dim<3) + { + $type .= "${sep}cctk_lsh$dim1"; + } + else + { + $type .= "${sep}cctk_lsh($dim1)"; + }

  2. Erik Schnetter reporter
    • removed comment

    I do not know how efficient Fortran compilers are when arrays sizes are defined via an array. I therefore introduced three local scalars cctk_lsh1 cctk_lsh2 cctk_lsh3 that are used to specify the first three dimensions. Since these are the "innermost" dimensions, they are most crucial for performance. These scalars are set via CCTK_FARGUMENTS as defined in cctk.h.

  3. Frank Löffler
    • removed comment

    So, the "if ($dim<3)" could also be an "if ($dim<=3)"? (You define cctk_lsh3 as well)

  4. Erik Schnetter reporter
    • removed comment

    No, $dim1 = $dim+1, since this loop counts dimensions from 0, while Fortran counts from 1.

  5. Frank Löffler
    • removed comment

    Right. Maybe we should use dim1 in the 'if' construct to avoid others thinking along the same wrong lines as me. You are right, the code is ok.

  6. Log in to comment