Analyzing missing vectorization speedup

[Tim_S_4]

Hello,

I am using the icc v.16 compiler to parallelize this part of my program:

#pragma simd assert
for(int i=0; i<nParticlesInUse; i++) {
  if (particles.id == INVALID)
    continue;
  particles.ax = GRAVITY_X;
  particles.ay = GRAVITY_Y;

  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 ) {

      } 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;
          particles.ay = 0;
        }
      }
    }
  }
}

gprof indicated that this is the most used function by far.
So, the vectorization report indicates a speedup of 2.29x:

LOOP BEGIN at noc8x8.cpp(1813,6)
   remark #15328: vectorization support: gather was emulated for the variable this:  strided by 14   [ noc8x8.cpp(1815,7) ]
   remark #15329: vectorization support: scatter was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1817,3) ]
   remark #15329: vectorization support: scatter was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1818,3) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1827,33) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1828,33) ]
   remark #15329: vectorization support: scatter was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1840,5) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1840,5) ]
   remark #15329: vectorization support: scatter was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1841,5) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1841,5) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1843,9) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1843,25) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1843,43) ]
   remark #15328: vectorization support: gather was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1843,59) ]
   remark #15329: vectorization support: scatter was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1845,6) ]
   remark #15329: vectorization support: scatter was emulated for the variable this: masked, strided by 7   [ noc8x8.cpp(1846,6) ]
   remark #15305: vectorization support: vector length 4
   remark #15309: vectorization support: normalized vectorization overhead 0.019
   remark #15301: SIMD LOOP WAS VECTORIZED
   remark #15452: unmasked strided loads: 8
   remark #15453: unmasked strided stores: 6
   remark #15460: masked strided loads: 1
   remark #15475: --- begin vector loop cost summary ---
   remark #15476: scalar loop cost: 679
   remark #15477: vector loop cost: 296.000
   remark #15478: estimated potential speedup: 2.290
   remark #15482: vectorized math library calls: 1
   remark #15488: --- end vector loop cost summary ---
   remark #25015: Estimate of max trip count of loop=15000

   LOOP BEGIN at noc8x8.cpp(1819,3)
      remark #25460: No loop optimizations reported
      remark #25015: Estimate of max trip count of loop=60000
   LOOP END
LOOP END

Unfortunately, the vectorization slows down the execution: without simd: 22.26 seconds; with simd: 24.28 seconds.
Can somebody give me some pointers what is best practice to debug this behavior?

Best regards,

Tim
 

[TimP]

Beyond testing a change from array of structures to structure of arrays?