llvm dev - Apr 2013 - [LLVMdev] Loop vectorizer behaviour for 2D arrays and parallel annotation

Hello,

I am trying to vectorize the following loop but the vectorizer says:
"Found a possible write-write reorder" and does not vectorize.

Why?



for (j=0; j < 8; j++)
  {
    jj = j << 3;
    m2[j][0] = diff[jj  ] + diff[jj+4];
    m2[j][1] = diff[jj+1] + diff[jj+5];
    m2[j][2] = diff[jj+2] + diff[jj+6];
    m2[j][3] = diff[jj+3] + diff[jj+7];
    m2[j][4] = diff[jj  ] - diff[jj+4];
    m2[j][5] = diff[jj+1] - diff[jj+5];
    m2[j][6] = diff[jj+2] - diff[jj+6];
    m2[j][7] = diff[jj+3] - diff[jj+7];
  }


Another question is regarding the isannotatedparallel() check. Is
there a way to make clang (or any other frontend) to generate parallel
annotated IR?


Best,
Anadi.

On 04/17/2013 04:55 AM, Anadi Mishra wrote:
> Hello,
>
> I am trying to vectorize the following loop but the vectorizer says:
> "Found a possible write-write reorder" and does not vectorize.
> Why?
To my knowledge, the dependence analysis in the loop vectorizer is not 
yet able to prove the absence of dependences here.
> for (j=0; j < 8; j++)
>    {
>      jj = j << 3;
>      m2[j][0] = diff[jj  ] + diff[jj+4];
>      m2[j][1] = diff[jj+1] + diff[jj+5];
>      m2[j][2] = diff[jj+2] + diff[jj+6];
>      m2[j][3] = diff[jj+3] + diff[jj+7];
>      m2[j][4] = diff[jj  ] - diff[jj+4];
>      m2[j][5] = diff[jj+1] - diff[jj+5];
>      m2[j][6] = diff[jj+2] - diff[jj+6];
>      m2[j][7] = diff[jj+3] - diff[jj+7];
>    }
>
>
> Another question is regarding the isannotatedparallel() check. Is
> there a way to make clang (or any other frontend) to generate parallel
> annotated IR?
Did you try to put '#pragma ivdep' before the loop.

Tobias

P.S.: Please attach a full C file as test case. The way the different 
data structures are declared my influence the analysis.

Hi Anadi Mishra,

On 04/17/2013 05:55 AM, Anadi Mishra wrote:
> Another question is regarding the isannotatedparallel() check. Is
> there a way to make clang (or any other frontend) to generate parallel
> annotated IR?
Paul Redmond was adding support for "#pragma ivdep" that would use the
parallel metadata, but I haven't been able to follow its progress lately.

FWIW, pocl's OpenCL kernel compiler adds the metadata to work-item loops.
That is, if your loop body was an OpenCL kernel with each work-item
executing a single iteration, it *might* get "horizontally vectorized"
using the loop vectorizer if you use pocl's 'loopvec' work group
method and
if the memory access pattern is suitable. This is quite fresh code which
I'm still optimizing, but I've already managed to autovectorize some
work
groups using it.

BR,
-- 
Pekka

On 2013-04-17, at 6:31 AM, Pekka Jääskeläinen wrote:
> 
> Paul Redmond was adding support for "#pragma ivdep" that would
use the
> parallel metadata, but I haven't been able to follow its progress
lately.
> 
I'm still working on it--just slowly :P I'm hoping to have some more
patches in the next week or two.

paul

On Wed, Apr 17, 2013 at 8:08 AM, Tobias Grosser <tobias at grosser.es>
wrote:
> On 04/17/2013 04:55 AM, Anadi Mishra wrote:
>>
>> Hello,
>>
>> I am trying to vectorize the following loop but the vectorizer says:
>> "Found a possible write-write reorder" and does not
vectorize.
>> Why?
>
>
> To my knowledge, the dependence analysis in the loop vectorizer is not yet
> able to prove the absence of dependences here.
While that is true, the debug message printed by the vectorizer is
misleading, which should not be.
>>
>> Another question is regarding the isannotatedparallel() check. Is
>> there a way to make clang (or any other frontend) to generate parallel
>> annotated IR?
>
>
> Did you try to put '#pragma ivdep' before the loop.
Thanks for the suggestion, it worked using the latest llvm from svn.
Thanks Pekka and Paul for your inputs.
>
> Tobias
>
> P.S.: Please attach a full C file as test case. The way the different data
> structures are declared my influence the analysis.
PFA the example.

-Best,
Anadi.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: example.c
Type: text/x-csrc
Size: 611 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130417/a101a71e/attachment.c>