RE: Vectorizer in GCC 4.0

["Balaji V. Iyer" <bviyer at ncsu dot edu>]

Thank you very much Mr. Naishlos.

-Balaji V. Iyer.

-----Original Message-----
From: Dorit Naishlos [mailto:DORIT@il.ibm.com] 
Sent: Monday, November 21, 2005 3:47 PM
To: Balaji V. Iyer
Cc: gcc@gcc.gnu.org
Subject: Re: Vectorizer in GCC 4.0





> Hello Everyone,
>     I am interested in knowing more about the vectorizer in GCC. Does
> anyone have or know of any statistics about the percentage of loops
> that can be vectorized in some benchmarks like MediaBench, SPEC2K and
> so forth?
>

I have some old Spec2000 statistics, from around June, using
autovect-branch:
226 loops where vectorized on powerpc970 out of 4506 inner-most
single-entry/exit single-bb countable loops. (note that on powerpc970 only
single-precision fp benchmarks can be vectorized, and most of Spec-fp
operates on doubles).
The more interesting question is whether we get any performance impact with
vectorization. There's one Spec-int benchmark for which vectorization
actually has an impact on the performance of the entire application - gzip
is improved by 8% due to vectorization.

Haven't tried MediaBench yet; I'd be curious to see the results if anyone
tries it out.

We vectorize 4 loops in linpack, getting ~36% improvement on the entire
benchmark.

Additional numbers on other benchmarks will be available in the near
future.

Also see:
- Vectorizing Hilaram:
http://gcc.gnu.org/ml/gcc/2005-10/msg01044.html

- Vectorizing libgfortran:
http://gcc.gnu.org/ml/gcc-bugs/2005-11/msg01722.html


dorit


> Thanking You,
>
> Yours Sincerely,
>
> Balaji V. Iyer.
>