runner_doiact_vec.c 33.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
/*******************************************************************************
 * This file is part of SWIFT.
 * Copyright (c) 2016 James Willis (james.s.willis@durham.ac.uk)
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 *
 ******************************************************************************/

/* Config parameters. */
#include "../config.h"

/* This object's header. */
#include "runner_doiact_vec.h"

#define array_align sizeof(float) * VEC_SIZE
#define NUM_VEC_PROC 2
#define C2_CACHE_SIZE (NUM_VEC_PROC * VEC_SIZE * 6) + (NUM_VEC_PROC * VEC_SIZE)

James Willis's avatar
James Willis committed
30
#ifdef WITH_VECTORIZATION
James Willis's avatar
James Willis committed
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
/**
 * @brief Compute the vector remainder interactions from the secondary cache.
 *
 * @param (return) r2q C2 cache of the separation between two particles squared.
 * @param (return) dxq C2 cache of the x separation between two particles.
 * @param (return) dyq C2 cache of the y separation between two particles.
 * @param (return) dzq C2 cache of the z separation between two particles.
 * @param (return) mq C2 cache of the mass of particle pj.
 * @param (return) vxq C2 cache of the x velocity of particle pj.
 * @param (return) vyq C2 cache of the y velocity of particle pj.
 * @param (return) vzq C2 cache of the z velocity of particle pj.
 * @param icount Interaction count.
 * @param (return) rhoSum #vector holding the cumulative sum of the density update on pi.
 * @param (return) rho_dhSum #vector holding the cumulative sum of the density gradient update on pi.
 * @param (return) wcountSum #vector holding the cumulative sum of the wcount update on pi.
 * @param (return) wcount_dhSum #vector holding the cumulative sum of the wcount gradient update on pi.
 * @param (return) div_vSum #vector holding the cumulative sum of the divergence update on pi.
 * @param (return) curlvxSum #vector holding the cumulative sum of the curl of vx update on pi.
 * @param (return) curlvySum #vector holding the cumulative sum of the curl of vy update on pi.
 * @param (return) curlvzSum #vector holding the cumulative sum of the curl of vz update on pi.
 * @param v_hi_inv #vector of 1/h for pi.
 * @param v_vix #vector of x velocity of pi.
 * @param v_viy #vector of y velocity of pi.
 * @param v_viz #vector of z velocity of pi.
 * @param (return) icount_align Interaction count after the remainder interactions have been performed, should be a multiple of the vector length.
 */
/__attribute__((always_inline)) INLINE static void calcRemInteractions(const struct cache *const cell_cache, float *r2q, float *dxq, float *dyq, float *dzq, float *mq, float *vxq, float *vyq, float *vzq, const int icount, vector *rhoSum, vector *rho_dhSum, vector *wcountSum, vector *wcount_dhSum, vector *div_vSum, vector *curlvxSum,vector *curlvySum, vector *curlvzSum, vector v_hi_inv, vector v_vix, vector v_viy, vector v_viz, int *icount_align) {
58
59
60
61
62

#ifdef HAVE_AVX512_F
  KNL_MASK_16 knl_mask, knl_mask2;
#endif
  vector int_mask, int_mask2;
James Willis's avatar
James Willis committed
63
64
 
  /* Work out the number of remainder interactions and pad secondary cache. */ 
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
  *icount_align = icount;
  int rem = icount % (NUM_VEC_PROC * VEC_SIZE);
  if (rem != 0) {
    int pad = (NUM_VEC_PROC * VEC_SIZE) - rem;
    *icount_align += pad;

    /* Initialise masks to true. */
#ifdef HAVE_AVX512_F
    knl_mask = 0xFFFF;
    knl_mask2 = 0xFFFF;
    int_mask.m = vec_setint1(0xFFFFFFFF);
    int_mask2.m = vec_setint1(0xFFFFFFFF);
#else
    int_mask.m = vec_setint1(0xFFFFFFFF);
    int_mask2.m = vec_setint1(0xFFFFFFFF);
#endif
James Willis's avatar
James Willis committed
81
    /* Pad secondary cache so that there are no contributions in the interaction function. */
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
    for(int i=icount; i<*icount_align; i++) {
      mq[i] = 0.f;
      r2q[i] = 1.f;
      dxq[i] = 0.f;
      dyq[i] = 0.f;
      dzq[i] = 0.f;
      vxq[i] = 0.f;
      vyq[i] = 0.f;
      vzq[i] = 0.f;
    }

    /* Zero parts of mask that represent the padded values.*/
    if (pad < VEC_SIZE) {
#ifdef HAVE_AVX512_F
      knl_mask2 = knl_mask2 >> pad;
#else
      for(int i=VEC_SIZE - pad; i<VEC_SIZE; i++) int_mask2.i[i] = 0;
#endif
    }
    else {
#ifdef HAVE_AVX512_F
      knl_mask = knl_mask >> (VEC_SIZE - rem);
      knl_mask2 = 0;
#else
      for(int i=rem; i<VEC_SIZE; i++) int_mask.i[i] = 0;
      int_mask2.v = vec_setzero();
#endif
    }

    /* Perform remainder interaction and remove remainder from aligned interaction count. */
    *icount_align = icount - rem;
    runner_iact_nonsym_2_vec_density(&r2q[*icount_align], &dxq[*icount_align], &dyq[*icount_align], &dzq[*icount_align], v_hi_inv, v_vix, v_viy, v_viz, &vxq[*icount_align], &vyq[*icount_align], &vzq[*icount_align], &mq[*icount_align], rhoSum, rho_dhSum, wcountSum, wcount_dhSum, div_vSum, curlvxSum, curlvySum, curlvzSum, int_mask, int_mask2,
#ifdef HAVE_AVX512_F
    knl_mask, knl_mask2);
#else
    0,0);
#endif
  }
}

James Willis's avatar
James Willis committed
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
/**
 * @brief Left-packs the values needed by an interaction into the secondary cache (Supports AVX, AVX2 and AVX512 instruction sets).
 *
 * @param mask Contains which particles need to interact.
 * @param v_r2 #vector of the separation between two particles squared.
 * @param v_dx #vector of the x separation between two particles.
 * @param v_dy #vector of the y separation between two particles.
 * @param v_dz #vector of the z separation between two particles.
 * @param v_mj #vector of the mass of particle pj.
 * @param v_vjx #vector of x velocity of pj.
 * @param v_vjy #vector of y velocity of pj.
 * @param v_vjz #vector of z velocity of pj.
 * @param cell_cache #cache of all particles in the cell.
 * @param r2q C2 cache of the separation between two particles squared.
 * @param dxq C2 cache of the x separation between two particles.
 * @param dyq C2 cache of the y separation between two particles.
 * @param dzq C2 cache of the z separation between two particles.
 * @param mq C2 cache of the mass of particle pj.
 * @param vxq C2 cache of the x velocity of particle pj.
 * @param vyq C2 cache of the y velocity of particle pj.
 * @param vzq C2 cache of the z velocity of particle pj.
 * @param icount Interaction count.
 * @param rhoSum #vector holding the cumulative sum of the density update on pi.
 * @param rho_dhSum #vector holding the cumulative sum of the density gradient update on pi.
 * @param wcountSum #vector holding the cumulative sum of the wcount update on pi.
 * @param wcount_dhSum #vector holding the cumulative sum of the wcount gradient update on pi.
 * @param div_vSum #vector holding the cumulative sum of the divergence update on pi.
 * @param curlvxSum #vector holding the cumulative sum of the curl of vx update on pi.
 * @param curlvySum #vector holding the cumulative sum of the curl of vy update on pi.
 * @param curlvzSum #vector holding the cumulative sum of the curl of vz update on pi.
 * @param v_hi_inv #vector of 1/h for pi.
 * @param v_vix #vector of x velocity of pi.
 * @param v_viy #vector of y velocity of pi.
 * @param v_viz #vector of z velocity of pi.
 */
157
158
__attribute__((always_inline)) INLINE static void storeInteractions(const int mask, const int pjd, vector *v_r2, vector *v_dx, vector *v_dy, vector *v_dz, vector *v_mj, vector *v_vjx, vector *v_vjy, vector *v_vjz, const struct cache *const cell_cache, float *r2q, float *dxq, float *dyq, float *dzq, float *mq, float *vxq, float *vyq, float *vzq, int *icount, vector *rhoSum, vector *rho_dhSum, vector *wcountSum, vector *wcount_dhSum, vector *div_vSum, vector *curlvxSum,vector *curlvySum, vector *curlvzSum, vector v_hi_inv, vector v_vix, vector v_viy, vector v_viz) {

James Willis's avatar
James Willis committed
159
/* Left-pack values needed into the secondary cache using the interaction mask. */
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
#if defined(HAVE_AVX2) || defined(HAVE_AVX512_F)
  int pack = 0;

#ifdef HAVE_AVX512_F
  pack += __builtin_popcount(mask);
  VEC_LEFT_PACK(v_r2->v,mask,&r2q[*icount]);
  VEC_LEFT_PACK(v_dx->v,mask,&dxq[*icount]);
  VEC_LEFT_PACK(v_dy->v,mask,&dyq[*icount]);
  VEC_LEFT_PACK(v_dz->v,mask,&dzq[*icount]);
  VEC_LEFT_PACK(v_mj->v,mask,&mq[*icount]);
  VEC_LEFT_PACK(v_vjx->v,mask,&vxq[*icount]);
  VEC_LEFT_PACK(v_vjy->v,mask,&vyq[*icount]);
  VEC_LEFT_PACK(v_vjz->v,mask,&vzq[*icount]);
#else
  vector v_mask;
  VEC_FORM_PACKED_MASK(mask,v_mask.m,pack);
  
  VEC_LEFT_PACK(v_r2->v,v_mask.m,&r2q[*icount]);
  VEC_LEFT_PACK(v_dx->v,v_mask.m,&dxq[*icount]);
  VEC_LEFT_PACK(v_dy->v,v_mask.m,&dyq[*icount]);
  VEC_LEFT_PACK(v_dz->v,v_mask.m,&dzq[*icount]);
  VEC_LEFT_PACK(v_mj->v,v_mask.m,&mq[*icount]);
  VEC_LEFT_PACK(v_vjx->v,v_mask.m,&vxq[*icount]);
  VEC_LEFT_PACK(v_vjy->v,v_mask.m,&vyq[*icount]);
  VEC_LEFT_PACK(v_vjz->v,v_mask.m,&vzq[*icount]);
#endif

  (*icount) += pack;
#else
James Willis's avatar
James Willis committed
189
  /* Quicker to do it serially in AVX rather than use intrinsics. */
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
  for(int bit_index = 0; bit_index<VEC_SIZE; bit_index++) {
    if (mask & (1 << bit_index)) {
      /* Add this interaction to the queue. */
      r2q[*icount] = v_r2->f[bit_index];
      dxq[*icount] = v_dx->f[bit_index];
      dyq[*icount] = v_dy->f[bit_index];
      dzq[*icount] = v_dz->f[bit_index];
      mq[*icount] = cell_cache->m[pjd + bit_index];
      vxq[*icount] = cell_cache->vx[pjd + bit_index];
      vyq[*icount] = cell_cache->vy[pjd + bit_index];
      vzq[*icount] = cell_cache->vz[pjd + bit_index];

      (*icount)++;
    }
  }
James Willis's avatar
James Willis committed
205
  /* Flush the c2 cache if it has reached capacity. */
206
207
208
  if(*icount >= (C2_CACHE_SIZE - (NUM_VEC_PROC * VEC_SIZE))) {

    int icount_align = *icount;
James Willis's avatar
James Willis committed
209
210
    
    /* Peform remainder interactions. */
211
212
213
214
215
    calcRemInteractions(cell_cache, r2q, dxq, dyq, dzq, mq, vxq, vyq, vzq, *icount, rhoSum, rho_dhSum, wcountSum, wcount_dhSum, div_vSum, curlvxSum, curlvySum, curlvzSum, v_hi_inv, v_vix, v_viy, v_viz, &icount_align);

    vector int_mask, int_mask2;
    int_mask.m = vec_setint1(0xFFFFFFFF);
    int_mask2.m = vec_setint1(0xFFFFFFFF);
James Willis's avatar
James Willis committed
216
217

    /* Perform interactions. */
218
219
220
    for (int pjd = 0; pjd < icount_align; pjd+=(NUM_VEC_PROC * VEC_SIZE)) {
      runner_iact_nonsym_2_vec_density(&r2q[pjd], &dxq[pjd], &dyq[pjd], &dzq[pjd], v_hi_inv, v_vix, v_viy, v_viz, &vxq[pjd], &vyq[pjd], &vzq[pjd], &mq[pjd], rhoSum, rho_dhSum, wcountSum, wcount_dhSum, div_vSum, curlvxSum, curlvySum, curlvzSum, int_mask, int_mask2, 0, 0);
    }
James Willis's avatar
James Willis committed
221
222

    /* Reset interaction count. */
223
224
225
226
227
    *icount = 0;
  }

#endif
}
James Willis's avatar
James Willis committed
228
#endif
229
230

/**
James Willis's avatar
James Willis committed
231
 * @brief Compute the cell self-interaction (non-symmetric) using vector intrinsics with one particle pi at a time.
232
233
234
235
 *
 * @param r The #runner.
 * @param c The #cell.
 */
James Willis's avatar
James Willis committed
236
__attribute__((always_inline)) INLINE void runner_doself1_density_vec(struct runner *r, struct cell *restrict c) {
237
238
239
240
241
242
243
244
245
246

#ifdef WITH_VECTORIZATION
  const int ti_current = r->e->ti_current;
  int doi_mask;
  struct part *restrict pi;
  int count_align;
  int num_vec_proc = NUM_VEC_PROC;

  struct part *restrict parts = c->parts;
  const int count = c->count;
James Willis's avatar
James Willis committed
247
248
249

  /* Get the particle cache from the runner and re-allocate 
   * the cache if it is not big enough for the cell. */
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
  struct cache *restrict cell_cache = &r->par_cache;
  
  int icount = 0, icount_align = 0;
  float r2q[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dxq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dyq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dzq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float mq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vxq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vyq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vzq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  
  vector v_hi, v_vix, v_viy, v_viz, v_hig2, v_r2;

  //TIMER_TIC

  if (c->ti_end_min > ti_current) return;
  if (c->ti_end_max < ti_current) error("Cell in an impossible time-zone");

  if(cell_cache->count < count) {
    cache_init(cell_cache,count);
  }

James Willis's avatar
James Willis committed
273
  /* Read the particles from the cell and store them locally in the cache. */
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
  cache_read_particles(c,cell_cache);

  /* Loop over the particles in the cell. */
  for (int pid = 0; pid < count; pid++) {

    /* Get a pointer to the ith particle. */
    pi = &parts[pid];

    /* Is the ith particle active? */
    if (pi->ti_end > ti_current) continue;

    vector pix, piy, piz;

    const float hi = cell_cache->h[pid];

James Willis's avatar
James Willis committed
289
    /* Fill particle pi vectors. */
290
291
292
293
294
295
296
297
298
299
300
    pix.v = vec_set1(cell_cache->x[pid]);
    piy.v = vec_set1(cell_cache->y[pid]);
    piz.v = vec_set1(cell_cache->z[pid]);
    v_hi.v = vec_set1(hi);
    v_vix.v = vec_set1(cell_cache->vx[pid]);
    v_viy.v = vec_set1(cell_cache->vy[pid]);
    v_viz.v = vec_set1(cell_cache->vz[pid]);

    const float hig2 = hi * hi * kernel_gamma2;
    v_hig2.v = vec_set1(hig2);

James Willis's avatar
James Willis committed
301
    /* Reset cumulative sums of update vectors. */
302
303
    vector rhoSum, rho_dhSum, wcountSum, wcount_dhSum, div_vSum, curlvxSum, curlvySum, curlvzSum;
    
James Willis's avatar
James Willis committed
304
    /* Get the inverse of hi. */
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
    vector v_hi_inv;
    
    VEC_RECIPROCAL(v_hi.v, v_hi_inv.v);
    
    rhoSum.v = vec_setzero();
    rho_dhSum.v = vec_setzero();
    wcountSum.v = vec_setzero();
    wcount_dhSum.v = vec_setzero();
    div_vSum.v = vec_setzero();
    curlvxSum.v = vec_setzero();
    curlvySum.v = vec_setzero();
    curlvzSum.v = vec_setzero();

    /* Pad cache if there is a serial remainder. */
    count_align = count;
    int rem = count % (num_vec_proc * VEC_SIZE);
    if (rem != 0) {
      int pad = (num_vec_proc * VEC_SIZE) - rem;

      count_align += pad;
      /* Set positions to the same as particle pi so when the r2 > 0 mask is applied these extra contributions are masked out.*/
      for(int i=count; i<count_align; i++) {
        cell_cache->x[i] = pix.f[0];
        cell_cache->y[i] = piy.f[0];
        cell_cache->z[i] = piz.f[0];
      }
    }

    vector pjx, pjy, pjz;
    vector pjvx, pjvy, pjvz, mj;
    vector pjx2, pjy2, pjz2;
    vector pjvx2, pjvy2, pjvz2, mj2;

James Willis's avatar
James Willis committed
338
    /* Find all of particle pi's interacions and store needed values in the secondary cache.*/
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
    for (int pjd = 0; pjd < count_align; pjd += (num_vec_proc * VEC_SIZE)) {

      /* Load 2 sets of vectors from the particle cache. */
      pjx.v = vec_load(&cell_cache->x[pjd]);
      pjy.v = vec_load(&cell_cache->y[pjd]);
      pjz.v = vec_load(&cell_cache->z[pjd]);
      pjvx.v = vec_load(&cell_cache->vx[pjd]);
      pjvy.v = vec_load(&cell_cache->vy[pjd]);
      pjvz.v = vec_load(&cell_cache->vz[pjd]);
      mj.v = vec_load(&cell_cache->m[pjd]);

      pjx2.v = vec_load(&cell_cache->x[pjd + VEC_SIZE]);
      pjy2.v = vec_load(&cell_cache->y[pjd + VEC_SIZE]);
      pjz2.v = vec_load(&cell_cache->z[pjd + VEC_SIZE]);
      pjvx2.v = vec_load(&cell_cache->vx[pjd + VEC_SIZE]);
      pjvy2.v = vec_load(&cell_cache->vy[pjd + VEC_SIZE]);
      pjvz2.v = vec_load(&cell_cache->vz[pjd + VEC_SIZE]);
      mj2.v = vec_load(&cell_cache->m[pjd +VEC_SIZE]);

      /* Compute the pairwise distance. */
      vector v_dx_tmp, v_dy_tmp, v_dz_tmp;
      vector v_dx_tmp2, v_dy_tmp2, v_dz_tmp2, v_r2_2;

      v_dx_tmp.v = vec_sub(pix.v,pjx.v);
      v_dy_tmp.v = vec_sub(piy.v,pjy.v);
      v_dz_tmp.v = vec_sub(piz.v,pjz.v);
      v_dx_tmp2.v = vec_sub(pix.v,pjx2.v);
      v_dy_tmp2.v = vec_sub(piy.v,pjy2.v);
      v_dz_tmp2.v = vec_sub(piz.v,pjz2.v);
      
      v_r2.v = vec_mul(v_dx_tmp.v,v_dx_tmp.v);
      v_r2.v = vec_fma(v_dy_tmp.v,v_dy_tmp.v,v_r2.v);
      v_r2.v = vec_fma(v_dz_tmp.v,v_dz_tmp.v,v_r2.v);
      v_r2_2.v = vec_mul(v_dx_tmp2.v,v_dx_tmp2.v);
      v_r2_2.v = vec_fma(v_dy_tmp2.v,v_dy_tmp2.v,v_r2_2.v);
      v_r2_2.v = vec_fma(v_dz_tmp2.v,v_dz_tmp2.v,v_r2_2.v);
      
      /* Form a mask from r2 < hig2 and r2 > 0.*/
#ifdef HAVE_AVX512_F
      //KNL_MASK_16 doi_mask, doi_mask_check, doi_mask2, doi_mask2_check;
      KNL_MASK_16 doi_mask_check, doi_mask2, doi_mask2_check;

      doi_mask_check = vec_cmp_gt(v_r2.v,vec_setzero());
      doi_mask = vec_cmp_lt(v_r2.v, v_hig2.v);

      doi_mask2_check = vec_cmp_gt(v_r2_2.v,vec_setzero());
      doi_mask2 = vec_cmp_lt(v_r2_2.v, v_hig2.v);

      doi_mask = doi_mask & doi_mask_check;
      doi_mask2 = doi_mask2 & doi_mask2_check;

#else
      vector v_doi_mask, v_doi_mask_check, v_doi_mask2, v_doi_mask2_check;
      int doi_mask2;

James Willis's avatar
James Willis committed
394
      /* Form r2 > 0 mask and r2 < hig2 mask. */
395
396
397
      v_doi_mask_check.v = vec_cmp_gt(v_r2.v,vec_setzero());
      v_doi_mask.v = vec_cmp_lt(v_r2.v, v_hig2.v);

James Willis's avatar
James Willis committed
398
      /* Form r2 > 0 mask and r2 < hig2 mask. */
399
400
401
      v_doi_mask2_check.v = vec_cmp_gt(v_r2_2.v,vec_setzero());
      v_doi_mask2.v = vec_cmp_lt(v_r2_2.v, v_hig2.v);

James Willis's avatar
James Willis committed
402
      /* Combine two masks and form integer mask. */
403
404
405
406
      doi_mask = vec_cmp_result(vec_and(v_doi_mask.v, v_doi_mask_check.v));
      doi_mask2 = vec_cmp_result(vec_and(v_doi_mask2.v, v_doi_mask2_check.v));
#endif

James Willis's avatar
James Willis committed
407
      /* If there are any interactions left pack interaction values into c2 cache. */
408
409
410
411
412
413
414
415
416
417
418
      if (doi_mask) {
        storeInteractions(doi_mask,pjd, &v_r2, &v_dx_tmp,&v_dy_tmp, &v_dz_tmp, &mj, &pjvx, &pjvy, &pjvz, cell_cache, &r2q[0], &dxq[0], &dyq[0], &dzq[0], &mq[0], &vxq[0], &vyq[0], &vzq[0], &icount, &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, v_hi_inv, v_vix, v_viy, v_viz);
      }
      if (doi_mask2) {
        storeInteractions(doi_mask2,pjd + VEC_SIZE, &v_r2_2, &v_dx_tmp2,&v_dy_tmp2, &v_dz_tmp2, &mj2, &pjvx2, &pjvy2, &pjvz2, cell_cache, &r2q[0], &dxq[0], &dyq[0], &dzq[0], &mq[0], &vxq[0], &vyq[0], &vzq[0], &icount, &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, v_hi_inv, v_vix, v_viy, v_viz);
      }
    }

    /* Perform padded vector remainder interactions if any are present. */    
    calcRemInteractions(cell_cache, &r2q[0], &dxq[0], &dyq[0], &dzq[0], &mq[0], &vxq[0], &vyq[0], &vzq[0], icount, &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, v_hi_inv, v_vix, v_viy, v_viz, &icount_align);
    
James Willis's avatar
James Willis committed
419
    /* Initialise masks to true in case remainder interactions have been performed. */
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
    vector int_mask, int_mask2;
#ifdef HAVE_AVX512_F
    KNL_MASK_16 knl_mask = 0xFFFF;
    KNL_MASK_16 knl_mask2 = 0xFFFF;
    int_mask.m = vec_setint1(0xFFFFFFFF);
    int_mask2.m = vec_setint1(0xFFFFFFFF);
#else
    int_mask.m = vec_setint1(0xFFFFFFFF);
    int_mask2.m = vec_setint1(0xFFFFFFFF);
#endif

    /* Perform interaction with 2 vectors. */
    for (int pjd = 0; pjd < icount_align; pjd+=(num_vec_proc * VEC_SIZE)) {
      runner_iact_nonsym_2_vec_density(&r2q[pjd], &dxq[pjd], &dyq[pjd], &dzq[pjd], v_hi_inv, v_vix, v_viy, v_viz, &vxq[pjd], &vyq[pjd], &vzq[pjd], &mq[pjd], &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, int_mask, int_mask2,
#ifdef HAVE_AVX512_F
      knl_mask, knl_mask2);
#else
      0, 0);      
#endif
    }

    /* Perform horizontal adds on vector sums and store result in particle pi. */
    VEC_HADD(rhoSum,pi->rho);
    VEC_HADD(rho_dhSum,pi->density.rho_dh);
    VEC_HADD(wcountSum,pi->density.wcount);
    VEC_HADD(wcount_dhSum,pi->density.wcount_dh);
    VEC_HADD(div_vSum,pi->density.div_v);
    VEC_HADD(curlvxSum,pi->density.rot_v[0]);
    VEC_HADD(curlvySum,pi->density.rot_v[1]);
    VEC_HADD(curlvzSum,pi->density.rot_v[2]);

    /* Reset interaction count. */
    icount = 0;
  } /* loop over all particles. */

  //TIMER_TOC(TIMER_DOSELF);
#endif
}

/**
James Willis's avatar
James Willis committed
460
461
 * @brief Compute the cell self-interaction (non-symmetric) using vector intrinsics with two particle pis at a time.
 * CURRENTLY BROKEN DO NOT USE.
462
463
464
465
 *
 * @param r The #runner.
 * @param c The #cell.
 */
James Willis's avatar
James Willis committed
466
__attribute__((always_inline)) INLINE void runner_doself1_density_vec_2(struct runner *r, struct cell *restrict c) {
467
468
469
470
471
472
473
474
475

#ifdef WITH_VECTORIZATION
  const int ti_current = r->e->ti_current;
  int doi_mask;
  int doi2_mask;
  struct part *restrict pi;
  struct part *restrict pi2;
  int count_align;

James Willis's avatar
James Willis committed
476
  /* Create two secondary caches. */  
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
  int icount = 0, icount_align = 0;
  float r2q[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dxq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dyq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dzq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float mq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vxq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vyq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vzq[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  
  int icount2 = 0, icount_align2 = 0;
  float r2q2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dxq2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dyq2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float dzq2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float mq2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vxq2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vyq2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));
  float vzq2[C2_CACHE_SIZE] __attribute__((aligned(array_align)));

  vector v_hi, v_vix, v_viy, v_viz, v_hig2, v_r2;
  vector v_hi2, v_vix2, v_viy2, v_viz2, v_hig2_2, v2_r2;

  //TIMER_TIC

James Willis's avatar
James Willis committed
502
  /* TODO: Need to find two active particles, not just one. */
503
504
505
506
507
  if (c->ti_end_min > ti_current) return;
  if (c->ti_end_max < ti_current) error("Cell in an impossible time-zone");

  struct part *restrict parts = c->parts;
  const int count = c->count;
James Willis's avatar
James Willis committed
508
509
510
  
  /* Get the particle cache from the runner and re-allocate 
   * the cache if it is not big enough for the cell. */
511
512
513
514
515
516
  struct cache *restrict cell_cache = &r->par_cache;

  if(cell_cache->count < count) {
    cache_init(cell_cache,count);
  }

James Willis's avatar
James Willis committed
517
  /* Read the particles from the cell and store them locally in the cache. */
518
519
520
521
522
  cache_read_particles(c,&r->par_cache);

  /* Loop over the particles in the cell. */
  for (int pid = 0; pid < count; pid+=2) {

James Willis's avatar
James Willis committed
523
    /* Get a pointer to the ith particle and next i particle. */
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
    pi = &parts[pid];
    pi2 = &parts[pid + 1];

    /* Is the ith particle active? */
    if (pi->ti_end > ti_current) continue;

    vector pix, piy, piz;
    vector pix2, piy2, piz2;

    const float hi = cell_cache->h[pid];
    const float hi2 = cell_cache->h[pid + 1];

    /* Fill pi position vector. */
    pix.v = vec_set1(cell_cache->x[pid]);
    piy.v = vec_set1(cell_cache->y[pid]);
    piz.v = vec_set1(cell_cache->z[pid]);
    v_hi.v = vec_set1(hi);
    v_vix.v = vec_set1(cell_cache->vx[pid]);
    v_viy.v = vec_set1(cell_cache->vy[pid]);
    v_viz.v = vec_set1(cell_cache->vz[pid]);

    pix2.v = vec_set1(cell_cache->x[pid + 1]);
    piy2.v = vec_set1(cell_cache->y[pid + 1]);
    piz2.v = vec_set1(cell_cache->z[pid + 1]);
    v_hi2.v = vec_set1(hi2);
    v_vix2.v = vec_set1(cell_cache->vx[pid + 1]);
    v_viy2.v = vec_set1(cell_cache->vy[pid + 1]);
    v_viz2.v = vec_set1(cell_cache->vz[pid + 1]);
    
    const float hig2 = hi * hi * kernel_gamma2;
    const float hig2_2 = hi2 * hi2 * kernel_gamma2;
    v_hig2.v = vec_set1(hig2);
    v_hig2_2.v = vec_set1(hig2_2);

    vector rhoSum, rho_dhSum, wcountSum, wcount_dhSum, div_vSum, curlvxSum, curlvySum, curlvzSum;
    vector rhoSum2, rho_dhSum2, wcountSum2, wcount_dhSum2, div_vSum2, curlvxSum2, curlvySum2, curlvzSum2;
    
    vector v_hi_inv, v_hi_inv2;
    
    VEC_RECIPROCAL(v_hi.v, v_hi_inv.v);
    VEC_RECIPROCAL(v_hi2.v, v_hi_inv2.v);
    
    rhoSum.v = vec_setzero();
    rho_dhSum.v = vec_setzero();
    wcountSum.v = vec_setzero();
    wcount_dhSum.v = vec_setzero();
    div_vSum.v = vec_setzero();
    curlvxSum.v = vec_setzero();
    curlvySum.v = vec_setzero();
    curlvzSum.v = vec_setzero();

    rhoSum2.v = vec_setzero();
    rho_dhSum2.v = vec_setzero();
    wcountSum2.v = vec_setzero();
    wcount_dhSum2.v = vec_setzero();
    div_vSum2.v = vec_setzero();
    curlvxSum2.v = vec_setzero();
    curlvySum2.v = vec_setzero();
    curlvzSum2.v = vec_setzero();

    /* Pad cache if there is a serial remainder. */
    count_align = count;
    int rem = count % (NUM_VEC_PROC * VEC_SIZE);
    if (rem != 0) {
      int pad = (NUM_VEC_PROC * VEC_SIZE) - rem;

      count_align += pad;
      /* Set positions to the same as particle pi so when the r2 > 0 mask is applied these extra contributions are masked out.*/
      for(int i=count; i<count_align; i++) {
        cell_cache->x[i] = pix.f[0];
        cell_cache->y[i] = piy.f[0];
        cell_cache->z[i] = piz.f[0];
      }
    }

    vector pjx, pjy, pjz;
    vector pjvx, pjvy, pjvz, mj;
    vector pjx2, pjy2, pjz2;
    vector pjvx2, pjvy2, pjvz2, mj2;

    /* Find all of particle pi's interacions and store needed values in secondary cache.*/
    for (int pjd = 0; pjd < count_align; pjd += (NUM_VEC_PROC * VEC_SIZE)) {

      /* Load 2 sets of vectors from the particle cache. */
      pjx.v = vec_load(&cell_cache->x[pjd]);
      pjy.v = vec_load(&cell_cache->y[pjd]);
      pjz.v = vec_load(&cell_cache->z[pjd]);
      pjvx.v = vec_load(&cell_cache->vx[pjd]);
      pjvy.v = vec_load(&cell_cache->vy[pjd]);
      pjvz.v = vec_load(&cell_cache->vz[pjd]);
      mj.v = vec_load(&cell_cache->m[pjd]);

      pjx2.v = vec_load(&cell_cache->x[pjd + VEC_SIZE]);
      pjy2.v = vec_load(&cell_cache->y[pjd + VEC_SIZE]);
      pjz2.v = vec_load(&cell_cache->z[pjd + VEC_SIZE]);
      pjvx2.v = vec_load(&cell_cache->vx[pjd + VEC_SIZE]);
      pjvy2.v = vec_load(&cell_cache->vy[pjd + VEC_SIZE]);
      pjvz2.v = vec_load(&cell_cache->vz[pjd + VEC_SIZE]);
      mj2.v = vec_load(&cell_cache->m[pjd +VEC_SIZE]);

      /* Compute the pairwise distance. */
      vector v_dx_tmp, v_dy_tmp, v_dz_tmp;
      vector v_dx_tmp2, v_dy_tmp2, v_dz_tmp2, v_r2_2;
      vector v_dx2_tmp, v_dy2_tmp, v_dz2_tmp;
      vector v_dx2_tmp2, v_dy2_tmp2, v_dz2_tmp2, v2_r2_2;

      v_dx_tmp.v = vec_sub(pix.v,pjx.v);
      v_dy_tmp.v = vec_sub(piy.v,pjy.v);
      v_dz_tmp.v = vec_sub(piz.v,pjz.v);
      v_dx_tmp2.v = vec_sub(pix.v,pjx2.v);
      v_dy_tmp2.v = vec_sub(piy.v,pjy2.v);
      v_dz_tmp2.v = vec_sub(piz.v,pjz2.v);
      
      v_dx2_tmp.v = vec_sub(pix2.v,pjx.v);
      v_dy2_tmp.v = vec_sub(piy2.v,pjy.v);
      v_dz2_tmp.v = vec_sub(piz2.v,pjz.v);
      v_dx2_tmp2.v = vec_sub(pix2.v,pjx2.v);
      v_dy2_tmp2.v = vec_sub(piy2.v,pjy2.v);
      v_dz2_tmp2.v = vec_sub(piz2.v,pjz2.v);

      v_r2.v = vec_mul(v_dx_tmp.v,v_dx_tmp.v);
      v_r2.v = vec_fma(v_dy_tmp.v,v_dy_tmp.v,v_r2.v);
      v_r2.v = vec_fma(v_dz_tmp.v,v_dz_tmp.v,v_r2.v);
      v_r2_2.v = vec_mul(v_dx_tmp2.v,v_dx_tmp2.v);
      v_r2_2.v = vec_fma(v_dy_tmp2.v,v_dy_tmp2.v,v_r2_2.v);
      v_r2_2.v = vec_fma(v_dz_tmp2.v,v_dz_tmp2.v,v_r2_2.v);
      
      v2_r2.v = vec_mul(v_dx2_tmp.v,v_dx2_tmp.v);
      v2_r2.v = vec_fma(v_dy2_tmp.v,v_dy2_tmp.v,v2_r2.v);
      v2_r2.v = vec_fma(v_dz2_tmp.v,v_dz2_tmp.v,v2_r2.v);
      v2_r2_2.v = vec_mul(v_dx2_tmp2.v,v_dx2_tmp2.v);
      v2_r2_2.v = vec_fma(v_dy2_tmp2.v,v_dy2_tmp2.v,v2_r2_2.v);
      v2_r2_2.v = vec_fma(v_dz2_tmp2.v,v_dz2_tmp2.v,v2_r2_2.v);

      /* Form a mask from r2 < hig2 and r2 > 0.*/
#ifdef HAVE_AVX512_F
      //KNL_MASK_16 doi_mask, doi_mask_check, doi_mask2, doi_mask2_check;
      KNL_MASK_16 doi_mask_check, doi_mask2, doi_mask2_check;
      KNL_MASK_16 doi2_mask_check, doi2_mask2, doi2_mask2_check;

      doi_mask_check = vec_cmp_gt(v_r2.v,vec_setzero());
      doi_mask = vec_cmp_lt(v_r2.v, v_hig2.v);

      doi2_mask_check = vec_cmp_gt(v2_r2.v,vec_setzero());
      doi2_mask = vec_cmp_lt(v2_r2.v, v_hig2_2.v);

      doi_mask2_check = vec_cmp_gt(v_r2_2.v,vec_setzero());
      doi_mask2 = vec_cmp_lt(v_r2_2.v, v_hig2.v);

      doi2_mask2_check = vec_cmp_gt(v2_r2_2.v,vec_setzero());
      doi2_mask2 = vec_cmp_lt(v2_r2_2.v, v_hig2_2.v);
      
      doi_mask = doi_mask & doi_mask_check;
      doi_mask2 = doi_mask2 & doi_mask2_check;

      doi2_mask = doi2_mask & doi2_mask_check;
      doi2_mask2 = doi2_mask2 & doi2_mask2_check;
#else
      vector v_doi_mask, v_doi_mask_check, v_doi_mask2, v_doi_mask2_check;
      int doi_mask2;

      vector v_doi2_mask, v_doi2_mask_check, v_doi2_mask2, v_doi2_mask2_check;
      int doi2_mask2;

      v_doi_mask_check.v = vec_cmp_gt(v_r2.v,vec_setzero());
      v_doi_mask.v = vec_cmp_lt(v_r2.v, v_hig2.v);

      v_doi2_mask_check.v = vec_cmp_gt(v2_r2.v,vec_setzero());
      v_doi2_mask.v = vec_cmp_lt(v2_r2.v, v_hig2_2.v);

      v_doi_mask2_check.v = vec_cmp_gt(v_r2_2.v,vec_setzero());
      v_doi_mask2.v = vec_cmp_lt(v_r2_2.v, v_hig2.v);

      v_doi2_mask2_check.v = vec_cmp_gt(v2_r2_2.v,vec_setzero());
      v_doi2_mask2.v = vec_cmp_lt(v2_r2_2.v, v_hig2_2.v);

      doi_mask = vec_cmp_result(vec_and(v_doi_mask.v, v_doi_mask_check.v));
      doi_mask2 = vec_cmp_result(vec_and(v_doi_mask2.v, v_doi_mask2_check.v));
      doi2_mask = vec_cmp_result(vec_and(v_doi2_mask.v, v_doi2_mask_check.v));
      doi2_mask2 = vec_cmp_result(vec_and(v_doi2_mask2.v, v_doi2_mask2_check.v));
#endif

      /* Hit or miss? */
      //if (doi_mask) {
        storeInteractions(doi_mask,pjd, &v_r2, &v_dx_tmp,&v_dy_tmp, &v_dz_tmp, &mj, &pjvx, &pjvy, &pjvz, cell_cache, &r2q[0], &dxq[0], &dyq[0], &dzq[0], &mq[0], &vxq[0], &vyq[0], &vzq[0], &icount, &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, v_hi_inv, v_vix, v_viy, v_viz);
      //}
      //if (doi2_mask) {
        storeInteractions(doi2_mask,pjd, &v2_r2, &v_dx2_tmp,&v_dy2_tmp, &v_dz2_tmp, &mj, &pjvx, &pjvy, &pjvz, cell_cache, &r2q2[0], &dxq2[0], &dyq2[0], &dzq2[0], &mq2[0], &vxq2[0], &vyq2[0], &vzq2[0], &icount2, &rhoSum2, &rho_dhSum2, &wcountSum2, &wcount_dhSum2, &div_vSum2, &curlvxSum2, &curlvySum2, &curlvzSum2, v_hi_inv2, v_vix2, v_viy2, v_viz2);
      //}       
      /* Hit or miss? */
      //if (doi_mask2) {
        storeInteractions(doi_mask2,pjd + VEC_SIZE, &v_r2_2, &v_dx_tmp2,&v_dy_tmp2, &v_dz_tmp2, &mj2, &pjvx2, &pjvy2, &pjvz2, cell_cache, &r2q[0], &dxq[0], &dyq[0], &dzq[0], &mq[0], &vxq[0], &vyq[0], &vzq[0], &icount, &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, v_hi_inv, v_vix, v_viy, v_viz);
      //}
      //if (doi2_mask2) {
        storeInteractions(doi2_mask2,pjd + VEC_SIZE, &v2_r2_2, &v_dx2_tmp2,&v_dy2_tmp2, &v_dz2_tmp2, &mj2, &pjvx2, &pjvy2, &pjvz2, cell_cache, &r2q2[0], &dxq2[0], &dyq2[0], &dzq2[0], &mq2[0], &vxq2[0], &vyq2[0], &vzq2[0], &icount2, &rhoSum2, &rho_dhSum2, &wcountSum2, &wcount_dhSum2, &div_vSum2, &curlvxSum2, &curlvySum2, &curlvzSum2, v_hi_inv2, v_vix2, v_viy2, v_viz2);
      //}
    }

    /* Perform padded vector remainder interactions if any are present. */    
    calcRemInteractions(cell_cache, &r2q[0], &dxq[0], &dyq[0], &dzq[0], &mq[0], &vxq[0], &vyq[0], &vzq[0], icount, &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, v_hi_inv, v_vix, v_viy, v_viz, &icount_align);
    
    calcRemInteractions(cell_cache, &r2q2[0], &dxq2[0], &dyq2[0], &dzq2[0], &mq2[0], &vxq2[0], &vyq2[0], &vzq2[0], icount2, &rhoSum2, &rho_dhSum2, &wcountSum2, &wcount_dhSum2, &div_vSum2, &curlvxSum2, &curlvySum2, &curlvzSum2, v_hi_inv2, v_vix2, v_viy2, v_viz2, &icount_align2);

    /* Initialise masks to true incase remainder interactions have been performed. */
    vector int_mask, int_mask2;
    vector int2_mask, int2_mask2;
#ifdef HAVE_AVX512_F
    KNL_MASK_16 knl_mask = 0xFFFF;
    KNL_MASK_16 knl_mask2 = 0xFFFF;
    int_mask.m = vec_setint1(0xFFFFFFFF);
    int_mask2.m = vec_setint1(0xFFFFFFFF);
    int2_mask.m = vec_setint1(0xFFFFFFFF);
    int2_mask2.m = vec_setint1(0xFFFFFFFF);
#else
    int_mask.m = vec_setint1(0xFFFFFFFF);
    int_mask2.m = vec_setint1(0xFFFFFFFF);
    
    int2_mask.m = vec_setint1(0xFFFFFFFF);
    int2_mask2.m = vec_setint1(0xFFFFFFFF);
#endif

    /* Perform interaction with 2 vectors. */
    for (int pjd = 0; pjd < icount_align; pjd+=(NUM_VEC_PROC * VEC_SIZE)) {
      runner_iact_nonsym_2_vec_density(&r2q[pjd], &dxq[pjd], &dyq[pjd], &dzq[pjd], v_hi_inv, v_vix, v_viy, v_viz, &vxq[pjd], &vyq[pjd], &vzq[pjd], &mq[pjd], &rhoSum, &rho_dhSum, &wcountSum, &wcount_dhSum, &div_vSum, &curlvxSum, &curlvySum, &curlvzSum, int_mask, int_mask2,
#ifdef HAVE_AVX512_F
      knl_mask, knl_mask2);
#else
      0, 0);      
#endif
    }

    for (int pjd = 0; pjd < icount_align2; pjd+=(NUM_VEC_PROC * VEC_SIZE)) {
      runner_iact_nonsym_2_vec_density(&r2q2[pjd], &dxq2[pjd], &dyq2[pjd], &dzq2[pjd], v_hi_inv2, v_vix2, v_viy2, v_viz2, &vxq2[pjd], &vyq2[pjd], &vzq2[pjd], &mq2[pjd], &rhoSum2, &rho_dhSum2, &wcountSum2, &wcount_dhSum2, &div_vSum2, &curlvxSum2, &curlvySum2, &curlvzSum2, int2_mask, int2_mask2,
#ifdef HAVE_AVX512_F
      knl_mask, knl_mask2);
#else
      0, 0);      
#endif
    }
    /* Perform horizontal adds on vector sums and store result in particle pi. */
    VEC_HADD(rhoSum,pi->rho);
    VEC_HADD(rho_dhSum,pi->density.rho_dh);
    VEC_HADD(wcountSum,pi->density.wcount);
    VEC_HADD(wcount_dhSum,pi->density.wcount_dh);
    VEC_HADD(div_vSum,pi->density.div_v);
    VEC_HADD(curlvxSum,pi->density.rot_v[0]);
    VEC_HADD(curlvySum,pi->density.rot_v[1]);
    VEC_HADD(curlvzSum,pi->density.rot_v[2]);

    VEC_HADD(rhoSum2,pi2->rho);
    VEC_HADD(rho_dhSum2,pi2->density.rho_dh);
    VEC_HADD(wcountSum2,pi2->density.wcount);
    VEC_HADD(wcount_dhSum2,pi2->density.wcount_dh);
    VEC_HADD(div_vSum2,pi2->density.div_v);
    VEC_HADD(curlvxSum2,pi2->density.rot_v[0]);
    VEC_HADD(curlvySum2,pi2->density.rot_v[1]);
    VEC_HADD(curlvzSum2,pi2->density.rot_v[2]);

    /* Reset interaction count. */
    icount = 0;
    icount2 = 0;
  } /* loop over all particles. */

  //TIMER_TOC(TIMER_DOSELF);
#endif

}