Vector et emergo: On the vectorization of the HIRLAM Weather Prediction code.

[Toon Moene]

L.S.,

At the last GCC summit I showed that complete compilation of the
HIRLAM Numerical Weather Prediction suite by gfortran/gcc was near
completion - only a few compiler bugs were between us and running
programs.

This (autumn) holiday, I was able to convince myself that the problems
on the part of gfortran/gcc were confined to gfortran not knowing the
LOC (extension) intrinsic, which will be part of LLNL's "Cray" pointers
package, and an "unclassifiable statement" when trying to parse a
five part concatenation of substrings (for which I'll file a bug report
shortly).

All other problems are due to HIRLAM people not following standards
closely enough, which I will remedy effectively ;-)

Therefore, the next challenge presented itself:  Given that the HIRLAM
code always has performed well on vector machines (as of '85), it's a
natural to compile it with -ftree-vectorize -ftree-vectorizer-verbose=n
compiler options, to see what the vectorizer can make of it.

Here are the top level results:

Loops considered by the vectorizer as potentially vectorizable: 10420.
Loops actually vectorized:                                       2645.
Loops not vectorized:                                            7775.
   of which:
   - unhandled data-ref:                                         3479.
   - complicated access pattern:                                 1500.
   - can't determine dependence:                                 1456.
   - unsupported use in stmt:                                     662.
   - no vectype for type (complex8):                              235.
   - relevant stmt not supported:                                 120.
   - mixed data-types:                                            149.

I'll send examples of the top three of this list, plus the ultimate
example why 235 occurrences of "no vectype for type (complex8)" is an
unnessary thing.

Kind regards,

-- 
Toon Moene - e-mail: toon@moene.indiv.nluug.nl - phone: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
A maintainer of GNU Fortran 95: http://gcc.gnu.org/fortran/