Parallelization of loops with if clause

[Tim_S_4]

Hello,

recently, I started to use the -avx feature and I am trying to transform my applications to simd.
So, I am running into some issues for parallelizing this loop:

#pragma omp simd
for (int j=0; j<nParticlesInUse; j++) {
  if (particles.id != INVALID) {

    double dx = particles.x - particles.x;
    double dy = particles.y - particles.y;
    double r2 = dx * dx + dy * dy;
    if( r2 > cutoff*cutoff ) {
      //return;
    } else {
      r2 = fmax( r2, min_r*min_r );
      double r = sqrt( r2 );
      double coef = ( 1 - cutoff / r ) / r2 / mass;
      particles.ax += coef * dx;
      particles.ay += coef * dy;

      if (particles.ax*particles.ax + particles.ay*particles.ay > 10000000) {
        particles.ax = 0; // here1
        particles.ay = 0; // here2
      }
    }
  }
}

I am getting an issue for the assignment in the if clause (here1 and here1):

      LOOP BEGIN at particle_simulation.cpp(1413,3)
         remark #15336: simd loop was not vectorized: conditional assignment to a scalar   [ test.cpp(1436,5) ]
         remark #13379: loop was not vectorized with "simd"
         remark #25015: Estimate of max trip count of loop=100
      LOOP END

I am using icc version 16.

In the Intel guide for vectorizing loops [1] I found this statement. However, I don't know how I can apply it to my situation:

The loop should consist primarily of straight-line code. There should be no jumps or branches such as switch statements, but masked assignments are allowed, including if-then-else constructs that can be interpreted as masked assignments.

Can somebody help me how to handle this if clause? Is there any best practice for these cases?
I am thankful for every pointer/help :)

Best regards

Tim

[1] https://software.intel.com/en-us/articles/requirements-for-vectorizable-loops

[Markus_W_Intel1]

Hello Tim,

Thank you for your question. From the issue you listed, it appears that the if clause was not vectorized with SIMD correctly. If you'll have a look at this SIMD article, you will see how to adjust your if clause.

Best Regards,

Markus

[Tim_S_4]

Hello Markus,

thank you for your feedback. I have read the article twice. Unfortunately,
I cannot find any pointer to my specific problem. You mentioned that the loop
was not vectorized correctly. Perhaps you can help me with these two questions:

a) What does the icc compiler means with scalar?

b) Can you describe the error message to me with some other words?
>> simd loop was not vectorized: conditional assignment to a scalar
   I don't understand this error message.

Best regards,

Tim

[Markus_W_Intel1]

Tim,

Both questions have already been addressed in this diagnostic, so please take a look at it. This link as well as related articles in it answer both of your questions.

For additional knowledge, I have included this Auto-Vectorization tutorial which I believe might be of more assistance to you.

Best Regards,

Markus

[TimP]

Among other issues, if you can use intel c++, std::max() is more amenable to optimization than fmax(). Unlike gcc which uses -ffast-math to discard the special requirements of fmax.

[Tim_S_4]

Hello Tim(P),

thank you very much for your pointer. I see that you are flying all around in this forum.
Perhaps you can help me with my parallelization issue. I read Markus's diagnostic article
and the "A Guide to Vectorization with Intel C++ Compilers". So, I believe that I got some
basic understanding about vectorization. Unfortunately, I am still not able to fix the loop.
Inserting lastprivate(particles.ax) in the pragma statement leads to
the error:

particle_simulation.cpp(1413): error: invalid entity for this variable list in omp clause
              #pragma omp simd lastprivate(particles.ax)

Do you know if the variable must have an arithmetic data type like int, float, double,... etc.?

Cheers!

Tim(S)

[Tim_S_4]

Okay, I could find the issue :) Problem solved :) thanks for all the pointers and help.

Best regards,

Tim