-O3 and -ftree-vectorize

[Tim Prince]

On 2/7/2014 11:09 AM, Tim Prince wrote:
>
> On 02/07/2014 10:22 AM, Jakub Jelinek wrote:
>> On Thu, Feb 06, 2014 at 05:21:00PM -0500, Tim Prince wrote:
>>> I'm seeing vectorization  but no output from
>>> -ftree-vectorizer-verbose, and no dot product vectorization inside
>>> omp parallel regions, with gcc g++ or gfortran 4.9.  Primary targets
>>> are cygwin64 and linux x86_64.
>>> I've been unable to use -O3 vectorization with gcc, although it
>>> works with gfortran and g++, so use gcc -O2 -ftree-vectorize
>>> together with additional optimization flags which don't break.
>> Can you file a GCC bugzilla PR with minimal testcases for this (or 
>> point us
>> at already filed bugreports)?
> The question of problems with gcc -O3 (called from gfortran) have 
> eluded me as to finding a minimal test case.  When I run under debug, 
> it appears that somewhere prior to the crash some gfortran code is 
> over-written with data by the gcc code, overwhelming my debugging 
> skill.  I can get full performance with -O2 plus a bunch of 
> intermediate flags.
> As to non-vectorization of dot product in omp parallel region, 
> -fopt-info (which I didn't know about) is reporting vectorization, but 
> there are no parallel simd instructions in the generated code for the 
> omp_fn.  I'll file a PR on that if it's still reproduced in a minimal 
> case.
>
>>
>>> I've made source code changes to take advantage of the new
>>> vectorization with merge() and ? operators; while it's useful for
>>> -march=core-avx2, it's sometimes a loss for -msse4.1.
>>> gcc vectorization with #pragma omp parallel for simd is reasonably
>>> effective in my tests only on 12 or more cores.
>> Likewise.
> Those are cases of 2 levels of loops from netlib "vector" benchmark 
> where only one level is vectorizable and parallelizable. By putting 
> the vectorizable loop on the outside the parallelization scales to a 
> large number of cores.  I don't expect it to out-perform single thread 
> optimized avx vectorization until 8 or more cores are in use, but it 
> needs more than expected number of threads even relative to SSE 
> vectorization.
>>
>>> #pragma omp simd reduction(max: ) is giving correct results but poor
>>> performance in my tests.
>> Likewise.
> I'll file a PR on this, didn't know if there might be interest.  I 
> have an Intel compiler issue "closed, will not be fixed" so the simd 
> reduction(max: ) isn't viable for icc in the near term.
> Thanks,
>
With further investigation, my case with reverse_copy outside and 
inner_product inside an omp parallel region is working very well with 
-O3 -ffast-math for double data type.  There seems a possible 
performance problem with reverse_copy for float data type, so much so 
that gfortran does better with the loop reversal pushed down into the 
parallel dot_products.  I have seen at least 2 cases where the new gcc 
vectorization of stride -1 with vpermd is superior to other compilers, 
even for float data type.
For the cases where omp parallel for simd is set in expectation of 
gaining outer loop parallel simd, gcc is ignoring the simd clause. So it 
is understandable that a large number of cores is needed to overcome the 
lack of parallel simd (other than by simd intrinsics coding).
I'll choose an example of omp simd reduction(max: ) for a PR.
Thanks.

-- 
Tim Prince