how to disable vectorization

how to disable vectorization

[Tim Prince]

Now that gfortran 6.0 has become aggressive enough on vectorization to
reduce performance in some cases, is there an equivalent to:
gcc (by function): __attribute__((optimize("no-tree-vectorize")))
or ifort (by DO loop) !dir$ novector ?

Re: how to disable vectorization

[Tobias Burnus]

Tim Prince wrote:
> Now that gfortran 6.0 has become aggressive enough on vectorization to
> reduce performance in some cases, is there an equivalent to:
> gcc (by function): __attribute__((optimize("no-tree-vectorize")))
> or ifort (by DO loop) !dir$ novector ?

I think there isn't but I wonder whether the OpenMP's "!$omp simd 
savelen(1)" directive is able to do this.

  (Requires -fopenmp-simd or -fopenmp, the former only enables the SIMD 
directives, ignoring other OpenMP directives, without linking libgomp 
and without  implying -frecursive.)

Cheers,

Tobias

Re: how to disable vectorization

[Steve Kargl]

On Sat, Nov 14, 2015 at 04:33:27PM -0500, Tim Prince wrote:
> Now that gfortran 6.0 has become aggressive enough on vectorization to
> reduce performance in some cases, is there an equivalent to:
> gcc (by function): __attribute__((optimize("no-tree-vectorize")))
> or ifort (by DO loop) !dir$ novector ?

No.

There is -fno-tree-vectorize.

-- 
Steve

Re: how to disable vectorization

[Tim Prince]

On 11/14/2015 5:00 PM, Tobias Burnus wrote:
> Tim Prince wrote:
>> Now that gfortran 6.0 has become aggressive enough on vectorization to
>> reduce performance in some cases, is there an equivalent to:
>> gcc (by function): __attribute__((optimize("no-tree-vectorize")))
>> or ifort (by DO loop) !dir$ novector ?
>
> I think there isn't but I wonder whether the OpenMP's "!$omp simd
> savelen(1)" directive is able to do this.
>
>  (Requires -fopenmp-simd or -fopenmp, the former only enables the SIMD
> directives, ignoring other OpenMP directives, without linking libgomp
> and without  implying -frecursive.)
>
> Cheers,
>
> Tobias
>
>
This works for a DO loop (allowing a compiler targeting Xeon Phi to
optimize):

#ifndef __MIC__
!$omp simd safelen(1)
#endif

leaving no way to control the optimization under f90 array assignment. 
This makes 14 omp simd loops, along with 17 omp parallel regions, in my
benchmark.
I brought up omp simd safelen last month when rehearsing for my Intel
sponsored webinar and it was the first time they had seen an example of
it; now you've pointed out another use.  It would also reconcile the C
source code between Intel and gnu compilers.

Thanks,

-- 
Tim Prince

Re: how to disable vectorization

[Richard Biener]

On November 14, 2015 10:33:27 PM GMT+01:00, Tim Prince <tprince818.tp@gmail.com> wrote:
>Now that gfortran 6.0 has become aggressive enough on vectorization to
>reduce performance in some cases, is there an equivalent to:
>gcc (by function): __attribute__((optimize("no-tree-vectorize")))
>or ifort (by DO loop) !dir$ novector ?

I'd be interested in such cases.  Can you file a bugreport please?

Richard.

Re: how to disable vectorization

[Tim Prince]

On 11/15/2015 2:44 AM, Richard Biener wrote:
> On November 14, 2015 10:33:27 PM GMT+01:00, Tim Prince <tprince818.tp@gmail.com> wrote:
>> Now that gfortran 6.0 has become aggressive enough on vectorization to
>> reduce performance in some cases, is there an equivalent to:
>> gcc (by function): __attribute__((optimize("no-tree-vectorize")))
>> or ifort (by DO loop) !dir$ novector ?
> I'd be interested in such cases.  Can you file a bugreport please?
>
> Richard.
>
>
The need for pragma to prevent vectorization (with ifdef guard to allow
it for Intel Xeon Phi) has been more pervasive in the past with Intel
compilers.  Recent Intel Fortran has improved its cost model so as to
make the right decision more consistently, but Intel C++ requires it inn
the same cases as the recent g++/gfortran update.   I will first work to
reconcile my test suite with the omp simd safelen(1) then I will post on
bugzilla.

Thanks,

-- 
Tim Prince