hydro_iact.h 39.1 KB
Newer Older
1
2
/*******************************************************************************
 * This file is part of SWIFT.
3
 * Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
4
 *                    Matthieu Schaller (matthieu.schaller@durham.ac.uk)
5
 *
6
7
8
9
 * 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.
10
 *
11
12
13
14
 * 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.
15
 *
16
17
 * 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/>.
18
 *
19
 ******************************************************************************/
20
21
#ifndef SWIFT_GADGET2_HYDRO_IACT_H
#define SWIFT_GADGET2_HYDRO_IACT_H
22
23

/**
24
 * @file Gadget2/hydro_iact.h
25
26
 * @brief SPH interaction functions following the Gadget-2 version of SPH.
 *
27
 * The interactions computed here are the ones presented in the Gadget-2 paper
28
29
 * Springel, V., MNRAS, Volume 364, Issue 4, pp. 1105-1134.
 * We use the same numerical coefficients as the Gadget-2 code. When used with
30
31
32
 * the Spline-3 kernel, the results should be equivalent to the ones obtained
 * with Gadget-2 up to the rounding errors and interactions missed by the
 * Gadget-2 tree-code neighbours search.
33
34
 */

35
36
#include "minmax.h"

37
38
39
/**
 * @brief Density loop
 */
40
41
42
__attribute__((always_inline)) INLINE static void runner_iact_density(
    float r2, float *dx, float hi, float hj, struct part *pi, struct part *pj) {

43
44
  float wi, wi_dx;
  float wj, wj_dx;
45
  float dv[3], curlvr[3];
46

47
  /* Get the masses. */
48
  const float mi = pi->mass;
49
50
51
52
53
54
55
56
57
58
59
60
61
  const float mj = pj->mass;

  /* Get r and r inverse. */
  const float r = sqrtf(r2);
  const float r_inv = 1.0f / r;

  /* Compute the kernel function for pi */
  const float hi_inv = 1.f / hi;
  const float ui = r * hi_inv;
  kernel_deval(ui, &wi, &wi_dx);

  /* Compute contribution to the density */
  pi->rho += mj * wi;
62
  pi->density.rho_dh -= mj * (hydro_dimension * wi + ui * wi_dx);
63

64
65
66
67
68
69
70
71
72
73
74
  /* Compute contribution to the number of neighbours */
  pi->density.wcount += wi;
  pi->density.wcount_dh -= ui * wi_dx;

  /* Compute the kernel function for pj */
  const float hj_inv = 1.f / hj;
  const float uj = r * hj_inv;
  kernel_deval(uj, &wj, &wj_dx);

  /* Compute contribution to the density */
  pj->rho += mi * wj;
75
  pj->density.rho_dh -= mi * (hydro_dimension * wj + uj * wj_dx);
76

77
78
79
  /* Compute contribution to the number of neighbours */
  pj->density.wcount += wj;
  pj->density.wcount_dh -= uj * wj_dx;
80

81
82
  const float faci = mj * wi_dx * r_inv;
  const float facj = mi * wj_dx * r_inv;
83

84
85
86
87
  /* Compute dv dot r */
  dv[0] = pi->v[0] - pj->v[0];
  dv[1] = pi->v[1] - pj->v[1];
  dv[2] = pi->v[2] - pj->v[2];
88
89
  const float dvdr = dv[0] * dx[0] + dv[1] * dx[1] + dv[2] * dx[2];

90
91
  pi->density.div_v -= faci * dvdr;
  pj->density.div_v -= facj * dvdr;
92
93
94
95
96
97

  /* Compute dv cross r */
  curlvr[0] = dv[1] * dx[2] - dv[2] * dx[1];
  curlvr[1] = dv[2] * dx[0] - dv[0] * dx[2];
  curlvr[2] = dv[0] * dx[1] - dv[1] * dx[0];

98
99
100
  pi->density.rot_v[0] += faci * curlvr[0];
  pi->density.rot_v[1] += faci * curlvr[1];
  pi->density.rot_v[2] += faci * curlvr[2];
101

102
103
104
  pj->density.rot_v[0] += facj * curlvr[0];
  pj->density.rot_v[1] += facj * curlvr[1];
  pj->density.rot_v[2] += facj * curlvr[2];
105
106
}

107
108
109
110
111
112
/**
 * @brief Density loop (Vectorized version)
 */
__attribute__((always_inline)) INLINE static void runner_iact_vec_density(
    float *R2, float *Dx, float *Hi, float *Hj, struct part **pi,
    struct part **pj) {
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137

#ifdef WITH_VECTORIZATION

  vector r, ri, r2, xi, xj, hi, hj, hi_inv, hj_inv, wi, wj, wi_dx, wj_dx;
  vector rhoi, rhoj, rhoi_dh, rhoj_dh, wcounti, wcountj, wcounti_dh, wcountj_dh;
  vector mi, mj;
  vector dx[3], dv[3];
  vector vi[3], vj[3];
  vector dvdr, div_vi, div_vj;
  vector curlvr[3], curl_vi[3], curl_vj[3];
  int k, j;

#if VEC_SIZE == 8
  /* Get the masses. */
  mi.v = vec_set(pi[0]->mass, pi[1]->mass, pi[2]->mass, pi[3]->mass,
                 pi[4]->mass, pi[5]->mass, pi[6]->mass, pi[7]->mass);
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass,
                 pj[4]->mass, pj[5]->mass, pj[6]->mass, pj[7]->mass);
  /* Get each velocity component. */
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k],
                      pi[4]->v[k], pi[5]->v[k], pi[6]->v[k], pi[7]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k],
                      pj[4]->v[k], pj[5]->v[k], pj[6]->v[k], pj[7]->v[k]);
  }
Matthieu Schaller's avatar
Matthieu Schaller committed
138
139
  /* Get each component of particle separation.
   * (Dx={dx1,dy1,dz1,dx2,dy2,dz2,...,dxn,dyn,dzn})*/
140
141
142
143
144
145
146
147
148
149
150
151
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k], Dx[12 + k],
                      Dx[15 + k], Dx[18 + k], Dx[21 + k]);
#elif VEC_SIZE == 4
  mi.v = vec_set(pi[0]->mass, pi[1]->mass, pi[2]->mass, pi[3]->mass);
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass);
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k]);
  }
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k]);
152
153
#else
  error("Unknown vector size.")
154
155
156
157
158
#endif

  /* Get the radius and inverse radius. */
  r2.v = vec_load(R2);
  ri.v = vec_rsqrt(r2.v);
Matthieu Schaller's avatar
Matthieu Schaller committed
159
160
  /*vec_rsqrt does not have the level of accuracy we need, so an extra term is
   * added below.*/
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
189
190
191
192
193
194
  ri.v = ri.v - vec_set1(0.5f) * ri.v * (r2.v * ri.v * ri.v - vec_set1(1.0f));
  r.v = r2.v * ri.v;

  hi.v = vec_load(Hi);
  hi_inv.v = vec_rcp(hi.v);
  hi_inv.v = hi_inv.v - hi_inv.v * (hi_inv.v * hi.v - vec_set1(1.0f));
  xi.v = r.v * hi_inv.v;

  hj.v = vec_load(Hj);
  hj_inv.v = vec_rcp(hj.v);
  hj_inv.v = hj_inv.v - hj_inv.v * (hj_inv.v * hj.v - vec_set1(1.0f));
  xj.v = r.v * hj_inv.v;

  /* Compute the kernel function. */
  kernel_deval_vec(&xi, &wi, &wi_dx);
  kernel_deval_vec(&xj, &wj, &wj_dx);

  /* Compute dv. */
  dv[0].v = vi[0].v - vj[0].v;
  dv[1].v = vi[1].v - vj[1].v;
  dv[2].v = vi[2].v - vj[2].v;

  /* Compute dv dot r */
  dvdr.v = (dv[0].v * dx[0].v) + (dv[1].v * dx[1].v) + (dv[2].v * dx[2].v);
  dvdr.v = dvdr.v * ri.v;

  /* Compute dv cross r */
  curlvr[0].v = dv[1].v * dx[2].v - dv[2].v * dx[1].v;
  curlvr[1].v = dv[2].v * dx[0].v - dv[0].v * dx[2].v;
  curlvr[2].v = dv[0].v * dx[1].v - dv[1].v * dx[0].v;
  for (k = 0; k < 3; k++) curlvr[k].v *= ri.v;

  /* Compute density of pi. */
  rhoi.v = mj.v * wi.v;
195
  rhoi_dh.v = mj.v * (vec_set1(hydro_dimension) * wi.v + xi.v * wi_dx.v);
196
197
198
199
200
201
202
  wcounti.v = wi.v;
  wcounti_dh.v = xi.v * wi_dx.v;
  div_vi.v = mj.v * dvdr.v * wi_dx.v;
  for (k = 0; k < 3; k++) curl_vi[k].v = mj.v * curlvr[k].v * wi_dx.v;

  /* Compute density of pj. */
  rhoj.v = mi.v * wj.v;
203
  rhoj_dh.v = mi.v * (vec_set1(hydro_dimension) * wj.v + xj.v * wj_dx.v);
204
205
206
207
208
209
210
211
  wcountj.v = wj.v;
  wcountj_dh.v = xj.v * wj_dx.v;
  div_vj.v = mi.v * dvdr.v * wj_dx.v;
  for (k = 0; k < 3; k++) curl_vj[k].v = mi.v * curlvr[k].v * wj_dx.v;

  /* Update particles. */
  for (k = 0; k < VEC_SIZE; k++) {
    pi[k]->rho += rhoi.f[k];
212
    pi[k]->density.rho_dh -= rhoi_dh.f[k];
213
214
    pi[k]->density.wcount += wcounti.f[k];
    pi[k]->density.wcount_dh -= wcounti_dh.f[k];
215
    pi[k]->density.div_v -= div_vi.f[k];
216
217
    for (j = 0; j < 3; j++) pi[k]->density.rot_v[j] += curl_vi[j].f[k];
    pj[k]->rho += rhoj.f[k];
218
    pj[k]->density.rho_dh -= rhoj_dh.f[k];
219
220
    pj[k]->density.wcount += wcountj.f[k];
    pj[k]->density.wcount_dh -= wcountj_dh.f[k];
221
    pj[k]->density.div_v -= div_vj.f[k];
222
223
224
225
226
    for (j = 0; j < 3; j++) pj[k]->density.rot_v[j] += curl_vj[j].f[k];
  }

#else

Matthieu Schaller's avatar
Matthieu Schaller committed
227
228
  error(
      "The Gadget2 serial version of runner_iact_density was called when the "
229
      "vectorised version should have been used.");
230
231

#endif
232
233
}

234
235
236
/**
 * @brief Density loop (non-symmetric version)
 */
237
238
239
240
241
242
243
__attribute__((always_inline)) INLINE static void runner_iact_nonsym_density(
    float r2, float *dx, float hi, float hj, struct part *pi, struct part *pj) {

  float wi, wi_dx;
  float dv[3], curlvr[3];

  /* Get the masses. */
244
  const float mj = pj->mass;
245
246

  /* Get r and r inverse. */
247
248
  const float r = sqrtf(r2);
  const float ri = 1.0f / r;
249

250
  /* Compute the kernel function */
251
252
253
  const float hi_inv = 1.0f / hi;
  const float ui = r * hi_inv;
  kernel_deval(ui, &wi, &wi_dx);
254
255
256

  /* Compute contribution to the density */
  pi->rho += mj * wi;
257
  pi->density.rho_dh -= mj * (hydro_dimension * wi + ui * wi_dx);
258
259
260

  /* Compute contribution to the number of neighbours */
  pi->density.wcount += wi;
261
  pi->density.wcount_dh -= ui * wi_dx;
262

263
  const float fac = mj * wi_dx * ri;
264

265
266
267
268
269
  /* Compute dv dot r */
  dv[0] = pi->v[0] - pj->v[0];
  dv[1] = pi->v[1] - pj->v[1];
  dv[2] = pi->v[2] - pj->v[2];
  const float dvdr = dv[0] * dx[0] + dv[1] * dx[1] + dv[2] * dx[2];
270
  pi->density.div_v -= fac * dvdr;
271

272
273
274
275
276
  /* Compute dv cross r */
  curlvr[0] = dv[1] * dx[2] - dv[2] * dx[1];
  curlvr[1] = dv[2] * dx[0] - dv[0] * dx[2];
  curlvr[2] = dv[0] * dx[1] - dv[1] * dx[0];

277
278
279
  pi->density.rot_v[0] += fac * curlvr[0];
  pi->density.rot_v[1] += fac * curlvr[1];
  pi->density.rot_v[2] += fac * curlvr[2];
280
281
}

282
283
284
285
286
287
/**
 * @brief Density loop (non-symmetric vectorized version)
 */
__attribute__((always_inline)) INLINE static void
runner_iact_nonsym_vec_density(float *R2, float *Dx, float *Hi, float *Hj,
                               struct part **pi, struct part **pj) {
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310

#ifdef WITH_VECTORIZATION

  vector r, ri, r2, xi, hi, hi_inv, wi, wi_dx;
  vector rhoi, rhoi_dh, wcounti, wcounti_dh, div_vi;
  vector mj;
  vector dx[3], dv[3];
  vector vi[3], vj[3];
  vector dvdr;
  vector curlvr[3], curl_vi[3];
  int k, j;

#if VEC_SIZE == 8
  /* Get the masses. */
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass,
                 pj[4]->mass, pj[5]->mass, pj[6]->mass, pj[7]->mass);
  /* Get each velocity component. */
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k],
                      pi[4]->v[k], pi[5]->v[k], pi[6]->v[k], pi[7]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k],
                      pj[4]->v[k], pj[5]->v[k], pj[6]->v[k], pj[7]->v[k]);
  }
Matthieu Schaller's avatar
Matthieu Schaller committed
311
312
  /* Get each component of particle separation.
   * (Dx={dx1,dy1,dz1,dx2,dy2,dz2,...,dxn,dyn,dzn})*/
313
314
315
316
317
318
319
320
321
322
323
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k], Dx[12 + k],
                      Dx[15 + k], Dx[18 + k], Dx[21 + k]);
#elif VEC_SIZE == 4
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass);
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k]);
  }
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k]);
324
325
#else
  error("Unknown vector size.")
326
327
328
329
330
#endif

  /* Get the radius and inverse radius. */
  r2.v = vec_load(R2);
  ri.v = vec_rsqrt(r2.v);
Matthieu Schaller's avatar
Matthieu Schaller committed
331
332
  /*vec_rsqrt does not have the level of accuracy we need, so an extra term is
   * added below.*/
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
  ri.v = ri.v - vec_set1(0.5f) * ri.v * (r2.v * ri.v * ri.v - vec_set1(1.0f));
  r.v = r2.v * ri.v;

  hi.v = vec_load(Hi);
  hi_inv.v = vec_rcp(hi.v);
  hi_inv.v = hi_inv.v - hi_inv.v * (hi_inv.v * hi.v - vec_set1(1.0f));
  xi.v = r.v * hi_inv.v;

  kernel_deval_vec(&xi, &wi, &wi_dx);

  /* Compute dv. */
  dv[0].v = vi[0].v - vj[0].v;
  dv[1].v = vi[1].v - vj[1].v;
  dv[2].v = vi[2].v - vj[2].v;

  /* Compute dv dot r */
  dvdr.v = (dv[0].v * dx[0].v) + (dv[1].v * dx[1].v) + (dv[2].v * dx[2].v);
  dvdr.v = dvdr.v * ri.v;

  /* Compute dv cross r */
  curlvr[0].v = dv[1].v * dx[2].v - dv[2].v * dx[1].v;
  curlvr[1].v = dv[2].v * dx[0].v - dv[0].v * dx[2].v;
  curlvr[2].v = dv[0].v * dx[1].v - dv[1].v * dx[0].v;
  for (k = 0; k < 3; k++) curlvr[k].v *= ri.v;

  /* Compute density of pi. */
  rhoi.v = mj.v * wi.v;
360
  rhoi_dh.v = mj.v * (vec_set1(hydro_dimension) * wi.v + xi.v * wi_dx.v);
361
362
363
364
365
366
367
368
  wcounti.v = wi.v;
  wcounti_dh.v = xi.v * wi_dx.v;
  div_vi.v = mj.v * dvdr.v * wi_dx.v;
  for (k = 0; k < 3; k++) curl_vi[k].v = mj.v * curlvr[k].v * wi_dx.v;

  /* Update particles. */
  for (k = 0; k < VEC_SIZE; k++) {
    pi[k]->rho += rhoi.f[k];
369
    pi[k]->density.rho_dh -= rhoi_dh.f[k];
370
371
    pi[k]->density.wcount += wcounti.f[k];
    pi[k]->density.wcount_dh -= wcounti_dh.f[k];
372
    pi[k]->density.div_v -= div_vi.f[k];
373
374
375
376
377
    for (j = 0; j < 3; j++) pi[k]->density.rot_v[j] += curl_vi[j].f[k];
  }

#else

Matthieu Schaller's avatar
Matthieu Schaller committed
378
379
  error(
      "The Gadget2 serial version of runner_iact_nonsym_density was called "
380
      "when the vectorised version should have been used.");
381
382

#endif
383
384
}

385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
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
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
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
502
503
504
505
506
507
508
/**
 * @brief Density interaction computed using 2 interleaved vectors (non-symmetric vectorized version).
 */
__attribute__((always_inline)) INLINE static void
runner_iact_nonsym_2_vec_density(float *R2, float *Dx, float *Dy, float *Dz, vector hi_inv, 
                                vector vix, vector viy, vector viz, float *Vjx, float *Vjy, float *Vjz, float *Mj, vector *rhoSum, vector *rho_dhSum, vector *wcountSum, vector *wcount_dhSum, vector *div_vSum, vector *curlvxSum,vector *curlvySum, vector *curlvzSum, vector mask, vector mask2, int knlMask, int knlMask2) {

  vector r, ri, r2, xi, wi, wi_dx;
  vector mj;
  vector dx, dy, dz, dvx, dvy, dvz;
  vector vjx, vjy, vjz;
  vector dvdr;
  vector curlvrx, curlvry, curlvrz;
  vector r_2, ri2, r2_2, xi2, wi2, wi_dx2;
  vector mj2;
  vector dx2, dy2, dz2, dvx2, dvy2, dvz2;
  vector vjx2, vjy2, vjz2;
  vector dvdr2;
  vector curlvrx2, curlvry2, curlvrz2;

  /* Get the masses. */
  mj.v = vec_load(Mj);
  mj2.v = vec_load(&Mj[VEC_SIZE]);
  vjx.v = vec_load(Vjx);
  vjx2.v = vec_load(&Vjx[VEC_SIZE]);
  vjy.v = vec_load(Vjy);
  vjy2.v = vec_load(&Vjy[VEC_SIZE]);
  vjz.v = vec_load(Vjz);
  vjz2.v = vec_load(&Vjz[VEC_SIZE]);
  dx.v = vec_load(Dx);
  dx2.v = vec_load(&Dx[VEC_SIZE]);
  dy.v = vec_load(Dy);
  dy2.v = vec_load(&Dy[VEC_SIZE]);
  dz.v = vec_load(Dz);
  dz2.v = vec_load(&Dz[VEC_SIZE]);

  /* Get the radius and inverse radius. */
  r2.v = vec_load(R2);
  r2_2.v = vec_load(&R2[VEC_SIZE]);
  VEC_RECIPROCAL_SQRT(r2.v, ri.v);
  VEC_RECIPROCAL_SQRT(r2_2.v, ri2.v);
  r.v = vec_mul(r2.v, ri.v);
  r_2.v = vec_mul(r2_2.v, ri2.v);

  xi.v = vec_mul(r.v, hi_inv.v);
  xi2.v = vec_mul(r_2.v, hi_inv.v);

  //kernel_deval_2_vec(&xi, &wi, &wi_dx,&xi2, &wi2, &wi_dx2);
  kernel_deval_vec(&xi, &wi, &wi_dx);
  kernel_deval_vec(&xi2, &wi2, &wi_dx2);

  /* Compute dv. */
  dvx.v = vec_sub(vix.v, vjx.v);
  dvx2.v = vec_sub(vix.v, vjx2.v);
  dvy.v = vec_sub(viy.v, vjy.v);
  dvy2.v = vec_sub(viy.v, vjy2.v);
  dvz.v = vec_sub(viz.v, vjz.v);
  dvz2.v = vec_sub(viz.v, vjz2.v);

  /* Compute dv dot r */
  dvdr.v = vec_fma(dvx.v, dx.v, vec_fma(dvy.v, dy.v, vec_mul(dvz.v, dz.v)));
  dvdr2.v = vec_fma(dvx2.v, dx2.v, vec_fma(dvy2.v, dy2.v, vec_mul(dvz2.v, dz2.v)));
  dvdr.v = vec_mul(dvdr.v, ri.v);
  dvdr2.v = vec_mul(dvdr2.v, ri2.v);

  /* Compute dv cross r */
  curlvrx.v = vec_fma(dvy.v, dz.v, vec_mul(vec_set1(-1.0f),vec_mul(dvz.v, dy.v)));
  curlvrx2.v = vec_fma(dvy2.v, dz2.v, vec_mul(vec_set1(-1.0f),vec_mul(dvz2.v, dy2.v)));
  curlvry.v = vec_fma(dvz.v, dx.v, vec_mul(vec_set1(-1.0f), vec_mul(dvx.v, dz.v)));
  curlvry2.v = vec_fma(dvz2.v, dx2.v, vec_mul(vec_set1(-1.0f), vec_mul(dvx2.v, dz2.v)));
  curlvrz.v = vec_fma(dvx.v, dy.v, vec_mul(vec_set1(-1.0f), vec_mul(dvy.v, dx.v)));
  curlvrz2.v = vec_fma(dvx2.v, dy2.v, vec_mul(vec_set1(-1.0f), vec_mul(dvy2.v, dx2.v)));
  curlvrx.v = vec_mul(curlvrx.v,ri.v);
  curlvrx2.v = vec_mul(curlvrx2.v,ri2.v);
  curlvry.v = vec_mul(curlvry.v,ri.v);
  curlvry2.v = vec_mul(curlvry2.v,ri2.v);
  curlvrz.v = vec_mul(curlvrz.v,ri.v);
  curlvrz2.v = vec_mul(curlvrz2.v,ri2.v);

  /* Mask updates to intermediate vector sums for particle pi. */
#ifdef HAVE_AVX512_F
  rhoSum->v = _mm512_mask_add_ps(rhoSum->v, knlMask, vec_mul(mj.v, wi.v), rhoSum->v);
  rhoSum->v = _mm512_mask_add_ps(rhoSum->v, knlMask2, vec_mul(mj2.v, wi2.v), rhoSum->v);

  rho_dhSum->v = _mm512_mask_sub_ps(rho_dhSum->v, knlMask, rho_dhSum->v, vec_mul(mj.v, vec_fma(vec_set1(hydro_dimension), wi.v, vec_mul(xi.v, wi_dx.v))));
  rho_dhSum->v = _mm512_mask_sub_ps(rho_dhSum->v, knlMask2, rho_dhSum->v, vec_mul(mj2.v, vec_fma(vec_set1(hydro_dimension), wi2.v, vec_mul(xi2.v, wi_dx2.v))));

  wcountSum->v = _mm512_mask_add_ps(wcountSum->v, knlMask, wi.v, wcountSum->v);
  wcountSum->v = _mm512_mask_add_ps(wcountSum->v, knlMask2, wi2.v, wcountSum->v);

  wcount_dhSum->v = _mm512_mask_sub_ps(wcount_dhSum->v, knlMask, wcount_dhSum->v, vec_mul(xi.v, wi_dx.v));
  wcount_dhSum->v = _mm512_mask_sub_ps(wcount_dhSum->v, knlMask2, wcount_dhSum->v, vec_mul(xi2.v, wi_dx2.v));

  div_vSum->v = _mm512_mask_sub_ps(div_vSum->v, knlMask, div_vSum->v, vec_mul(mj.v, vec_mul(dvdr.v, wi_dx.v)));
  div_vSum->v = _mm512_mask_sub_ps(div_vSum->v, knlMask2, div_vSum->v, vec_mul(mj2.v, vec_mul(dvdr2.v, wi_dx2.v)));

  curlvxSum->v = _mm512_mask_add_ps(curlvxSum->v, knlMask, vec_mul(mj.v, vec_mul(curlvrx.v, wi_dx.v)), curlvxSum->v);
  curlvxSum->v = _mm512_mask_add_ps(curlvxSum->v, knlMask2, vec_mul(mj2.v, vec_mul(curlvrx2.v, wi_dx2.v)), curlvxSum->v);

  curlvySum->v = _mm512_mask_add_ps(curlvySum->v, knlMask, vec_mul(mj.v, vec_mul(curlvry.v, wi_dx.v)), curlvySum->v);
  curlvySum->v = _mm512_mask_add_ps(curlvySum->v, knlMask2, vec_mul(mj2.v, vec_mul(curlvry2.v, wi_dx2.v)), curlvySum->v);

  curlvzSum->v = _mm512_mask_add_ps(curlvzSum->v, knlMask, vec_mul(mj.v, vec_mul(curlvrz.v, wi_dx.v)), curlvzSum->v);
  curlvzSum->v = _mm512_mask_add_ps(curlvzSum->v, knlMask2, vec_mul(mj2.v, vec_mul(curlvrz2.v, wi_dx2.v)), curlvzSum->v);
#else
  rhoSum->v += vec_and(vec_mul(mj.v, wi.v),mask.v);
  rhoSum->v += vec_and(vec_mul(mj2.v, wi2.v),mask2.v);
  rho_dhSum->v -= vec_and(vec_mul(mj.v, vec_fma(vec_set1(hydro_dimension), wi.v, vec_mul(xi.v, wi_dx.v))),mask.v);
  rho_dhSum->v -= vec_and(vec_mul(mj2.v, vec_fma(vec_set1(hydro_dimension), wi2.v, vec_mul(xi2.v, wi_dx2.v))),mask2.v);
  wcountSum->v += vec_and(wi.v,mask.v);
  wcountSum->v += vec_and(wi2.v,mask2.v);
  wcount_dhSum->v -= vec_and(vec_mul(xi.v, wi_dx.v),mask.v);
  wcount_dhSum->v -= vec_and(vec_mul(xi2.v, wi_dx2.v),mask2.v);
  div_vSum->v -= vec_and(vec_mul(mj.v, vec_mul(dvdr.v, wi_dx.v)),mask.v);
  div_vSum->v -= vec_and(vec_mul(mj2.v, vec_mul(dvdr2.v, wi_dx2.v)),mask2.v);
  curlvxSum->v += vec_and(vec_mul(mj.v, vec_mul(curlvrx.v, wi_dx.v)),mask.v);
  curlvxSum->v += vec_and(vec_mul(mj2.v, vec_mul(curlvrx2.v, wi_dx2.v)),mask2.v);
  curlvySum->v += vec_and(vec_mul(mj.v, vec_mul(curlvry.v, wi_dx.v)),mask.v);
  curlvySum->v += vec_and(vec_mul(mj2.v, vec_mul(curlvry2.v, wi_dx2.v)),mask2.v);
  curlvzSum->v += vec_and(vec_mul(mj.v, vec_mul(curlvrz.v, wi_dx.v)),mask.v);
  curlvzSum->v += vec_and(vec_mul(mj2.v, vec_mul(curlvrz2.v, wi_dx2.v)),mask2.v);
#endif
}

509
510
511
/**
 * @brief Force loop
 */
512
513
514
__attribute__((always_inline)) INLINE static void runner_iact_force(
    float r2, float *dx, float hi, float hj, struct part *pi, struct part *pj) {

515
516
517
  float wi, wj, wi_dx, wj_dx;

  const float fac_mu = 1.f; /* Will change with cosmological integration */
518

519
520
521
522
  const float r = sqrtf(r2);
  const float r_inv = 1.0f / r;

  /* Get some values in local variables. */
523
  const float mi = pi->mass;
524
525
526
527
528
529
  const float mj = pj->mass;
  const float rhoi = pi->rho;
  const float rhoj = pj->rho;

  /* Get the kernel for hi. */
  const float hi_inv = 1.0f / hi;
530
  const float hid_inv = pow_dimension_plus_one(hi_inv); /* 1/h^(d+1) */
531
532
  const float ui = r * hi_inv;
  kernel_deval(ui, &wi, &wi_dx);
533
  const float wi_dr = hid_inv * wi_dx;
534
535
536

  /* Get the kernel for hj. */
  const float hj_inv = 1.0f / hj;
537
  const float hjd_inv = pow_dimension_plus_one(hj_inv); /* 1/h^(d+1) */
538
539
  const float xj = r * hj_inv;
  kernel_deval(xj, &wj, &wj_dx);
540
  const float wj_dr = hjd_inv * wj_dx;
541

542
543
544
545
546
  /* Compute h-gradient terms */
  const float f_i = pi->force.f;
  const float f_j = pj->force.f;

  /* Compute pressure terms */
547
548
  const float P_over_rho2_i = pi->force.P_over_rho2;
  const float P_over_rho2_j = pj->force.P_over_rho2;
549
550

  /* Compute sound speeds */
551
552
  const float ci = pi->force.soundspeed;
  const float cj = pj->force.soundspeed;
553

554
  /* Compute dv dot r. */
555
556
557
  const float dvdr = (pi->v[0] - pj->v[0]) * dx[0] +
                     (pi->v[1] - pj->v[1]) * dx[1] +
                     (pi->v[2] - pj->v[2]) * dx[2];
558

559
  /* Balsara term */
560
561
  const float balsara_i = pi->force.balsara;
  const float balsara_j = pj->force.balsara;
Matthieu Schaller's avatar
Matthieu Schaller committed
562

563
  /* Are the particles moving towards each others ? */
564
  const float omega_ij = (dvdr < 0.f) ? dvdr : 0.f;
565
566
567
568
569
570
571
572
573
  const float mu_ij = fac_mu * r_inv * omega_ij; /* This is 0 or negative */

  /* Signal velocity */
  const float v_sig = ci + cj - 3.f * mu_ij;

  /* Now construct the full viscosity term */
  const float rho_ij = 0.5f * (rhoi + rhoj);
  const float visc = -0.25f * const_viscosity_alpha * v_sig * mu_ij *
                     (balsara_i + balsara_j) / rho_ij;
574
575

  /* Now, convolve with the kernel */
576
  const float visc_term = 0.5f * visc * (wi_dr + wj_dr) * r_inv;
577
  const float sph_term =
578
      (f_i * P_over_rho2_i * wi_dr + f_j * P_over_rho2_j * wj_dr) * r_inv;
579
580
581
582
583

  /* Eventually got the acceleration */
  const float acc = visc_term + sph_term;

  /* Use the force Luke ! */
584
585
586
  pi->a_hydro[0] -= mj * acc * dx[0];
  pi->a_hydro[1] -= mj * acc * dx[1];
  pi->a_hydro[2] -= mj * acc * dx[2];
587

588
589
590
  pj->a_hydro[0] += mi * acc * dx[0];
  pj->a_hydro[1] += mi * acc * dx[1];
  pj->a_hydro[2] += mi * acc * dx[2];
591

592
  /* Get the time derivative for h. */
593
594
  pi->force.h_dt -= mj * dvdr * r_inv / rhoj * wi_dr;
  pj->force.h_dt -= mi * dvdr * r_inv / rhoi * wj_dr;
595

596
  /* Update the signal velocity. */
597
598
  pi->force.v_sig = (pi->force.v_sig > v_sig) ? pi->force.v_sig : v_sig;
  pj->force.v_sig = (pj->force.v_sig > v_sig) ? pj->force.v_sig : v_sig;
599

600
  /* Change in entropy */
601
602
  pi->entropy_dt += mj * visc_term * dvdr;
  pj->entropy_dt += mi * visc_term * dvdr;
603
}
604

605
606
607
608
609
610
/**
 * @brief Force loop (Vectorized version)
 */
__attribute__((always_inline)) INLINE static void runner_iact_vec_force(
    float *R2, float *Dx, float *Hi, float *Hj, struct part **pi,
    struct part **pj) {
611
612
613
614
615
616

#ifdef WITH_VECTORIZATION

  vector r, r2, ri;
  vector xi, xj;
  vector hi, hj, hi_inv, hj_inv;
617
  vector hid_inv, hjd_inv;
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
  vector wi, wj, wi_dx, wj_dx, wi_dr, wj_dr, dvdr;
  vector piPOrho, pjPOrho, pirho, pjrho;
  vector mi, mj;
  vector f;
  vector dx[3];
  vector vi[3], vj[3];
  vector pia[3], pja[3];
  vector pih_dt, pjh_dt;
  vector ci, cj, v_sig;
  vector omega_ij, mu_ij, fac_mu, balsara;
  vector rho_ij, visc, visc_term, sph_term, acc, entropy_dt;
  int j, k;

  fac_mu.v = vec_set1(1.f); /* Will change with cosmological integration */

Matthieu Schaller's avatar
Matthieu Schaller committed
633
/* Load stuff. */
634
635
636
637
638
#if VEC_SIZE == 8
  mi.v = vec_set(pi[0]->mass, pi[1]->mass, pi[2]->mass, pi[3]->mass,
                 pi[4]->mass, pi[5]->mass, pi[6]->mass, pi[7]->mass);
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass,
                 pj[4]->mass, pj[5]->mass, pj[6]->mass, pj[7]->mass);
Matthieu Schaller's avatar
Matthieu Schaller committed
639
640
641
642
643
644
645
646
  piPOrho.v = vec_set(pi[0]->force.P_over_rho2, pi[1]->force.P_over_rho2,
                      pi[2]->force.P_over_rho2, pi[3]->force.P_over_rho2,
                      pi[4]->force.P_over_rho2, pi[5]->force.P_over_rho2,
                      pi[6]->force.P_over_rho2, pi[7]->force.P_over_rho2);
  pjPOrho.v = vec_set(pj[0]->force.P_over_rho2, pj[1]->force.P_over_rho2,
                      pj[2]->force.P_over_rho2, pj[3]->force.P_over_rho2,
                      pj[4]->force.P_over_rho2, pj[5]->force.P_over_rho2,
                      pj[6]->force.P_over_rho2, pj[7]->force.P_over_rho2);
647
648
649
650
  pirho.v = vec_set(pi[0]->rho, pi[1]->rho, pi[2]->rho, pi[3]->rho, pi[4]->rho,
                    pi[5]->rho, pi[6]->rho, pi[7]->rho);
  pjrho.v = vec_set(pj[0]->rho, pj[1]->rho, pj[2]->rho, pj[3]->rho, pj[4]->rho,
                    pj[5]->rho, pj[6]->rho, pj[7]->rho);
Matthieu Schaller's avatar
Matthieu Schaller committed
651
652
653
654
655
656
657
658
  ci.v = vec_set(pi[0]->force.soundspeed, pi[1]->force.soundspeed,
                 pi[2]->force.soundspeed, pi[3]->force.soundspeed,
                 pi[4]->force.soundspeed, pi[5]->force.soundspeed,
                 pi[6]->force.soundspeed, pi[7]->force.soundspeed);
  cj.v = vec_set(pj[0]->force.soundspeed, pj[1]->force.soundspeed,
                 pj[2]->force.soundspeed, pj[3]->force.soundspeed,
                 pj[4]->force.soundspeed, pj[5]->force.soundspeed,
                 pj[6]->force.soundspeed, pj[7]->force.soundspeed);
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k],
                      pi[4]->v[k], pi[5]->v[k], pi[6]->v[k], pi[7]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k],
                      pj[4]->v[k], pj[5]->v[k], pj[6]->v[k], pj[7]->v[k]);
  }
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k], Dx[12 + k],
                      Dx[15 + k], Dx[18 + k], Dx[21 + k]);
  balsara.v =
      vec_set(pi[0]->force.balsara, pi[1]->force.balsara, pi[2]->force.balsara,
              pi[3]->force.balsara, pi[4]->force.balsara, pi[5]->force.balsara,
              pi[6]->force.balsara, pi[7]->force.balsara) +
      vec_set(pj[0]->force.balsara, pj[1]->force.balsara, pj[2]->force.balsara,
              pj[3]->force.balsara, pj[4]->force.balsara, pj[5]->force.balsara,
              pj[6]->force.balsara, pj[7]->force.balsara);
#elif VEC_SIZE == 4
676
  mi.v = vec_set(pi[0]->mass, pi[1]->mass, pi[2]->mass, pi[3]->mass);
677
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass);
678
  piPOrho.v = vec_set(pi[0]->force.P_over_rho2, pi[1]->force.P_over_rho2,
Matthieu Schaller's avatar
Matthieu Schaller committed
679
                      pi[2]->force.P_over_rho2, pi[3]->force.P_over_rho2);
680
  pjPOrho.v = vec_set(pj[0]->force.P_over_rho2, pj[1]->force.P_over_rho2,
Matthieu Schaller's avatar
Matthieu Schaller committed
681
                      pj[2]->force.P_over_rho2, pj[3]->force.P_over_rho2);
682
683
  pirho.v = vec_set(pi[0]->rho, pi[1]->rho, pi[2]->rho, pi[3]->rho);
  pjrho.v = vec_set(pj[0]->rho, pj[1]->rho, pj[2]->rho, pj[3]->rho);
Matthieu Schaller's avatar
Matthieu Schaller committed
684
685
686
687
  ci.v = vec_set(pi[0]->force.soundspeed, pi[1]->force.soundspeed,
                 pi[2]->force.soundspeed, pi[3]->force.soundspeed);
  cj.v = vec_set(pj[0]->force.soundspeed, pj[1]->force.soundspeed,
                 pj[2]->force.soundspeed, pj[3]->force.soundspeed);
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k]);
  }
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k]);
  balsara.v = vec_set(pi[0]->force.balsara, pi[1]->force.balsara,
                      pi[2]->force.balsara, pi[3]->force.balsara) +
              vec_set(pj[0]->force.balsara, pj[1]->force.balsara,
                      pj[2]->force.balsara, pj[3]->force.balsara);
#else
  error("Unknown vector size.")
#endif

  /* Get the radius and inverse radius. */
  r2.v = vec_load(R2);
  ri.v = vec_rsqrt(r2.v);
  ri.v = ri.v - vec_set1(0.5f) * ri.v * (r2.v * ri.v * ri.v - vec_set1(1.0f));
  r.v = r2.v * ri.v;

  /* Get the kernel for hi. */
  hi.v = vec_load(Hi);
  hi_inv.v = vec_rcp(hi.v);
  hi_inv.v = hi_inv.v - hi_inv.v * (hi.v * hi_inv.v - vec_set1(1.0f));
712
  hid_inv = pow_dimension_plus_one_vec(hi_inv); /* 1/h^(d+1) */
713
714
  xi.v = r.v * hi_inv.v;
  kernel_deval_vec(&xi, &wi, &wi_dx);
715
  wi_dr.v = hid_inv.v * wi_dx.v;
716
717
718
719
720

  /* Get the kernel for hj. */
  hj.v = vec_load(Hj);
  hj_inv.v = vec_rcp(hj.v);
  hj_inv.v = hj_inv.v - hj_inv.v * (hj.v * hj_inv.v - vec_set1(1.0f));
721
  hjd_inv = pow_dimension_plus_one_vec(hj_inv); /* 1/h^(d+1) */
722
723
  xj.v = r.v * hj_inv.v;
  kernel_deval_vec(&xj, &wj, &wj_dx);
724
  wj_dr.v = hjd_inv.v * wj_dx.v;
725
726
727
728

  /* Compute dv dot r. */
  dvdr.v = ((vi[0].v - vj[0].v) * dx[0].v) + ((vi[1].v - vj[1].v) * dx[1].v) +
           ((vi[2].v - vj[2].v) * dx[2].v);
Matthieu Schaller's avatar
Matthieu Schaller committed
729
  // dvdr.v = dvdr.v * ri.v;
730
731
732
733
734

  /* Compute the relative velocity. (This is 0 if the particles move away from
   * each other and negative otherwise) */
  omega_ij.v = vec_fmin(dvdr.v, vec_set1(0.0f));
  mu_ij.v = fac_mu.v * ri.v * omega_ij.v; /* This is 0 or negative */
Matthieu Schaller's avatar
Matthieu Schaller committed
735

736
737
  /* Compute signal velocity */
  v_sig.v = ci.v + cj.v - vec_set1(3.0f) * mu_ij.v;
Matthieu Schaller's avatar
Matthieu Schaller committed
738

739
740
  /* Now construct the full viscosity term */
  rho_ij.v = vec_set1(0.5f) * (pirho.v + pjrho.v);
Matthieu Schaller's avatar
Matthieu Schaller committed
741
742
  visc.v = vec_set1(-0.25f) * vec_set1(const_viscosity_alpha) * v_sig.v *
           mu_ij.v * balsara.v / rho_ij.v;
743
744
745
746
747
748
749

  /* Now, convolve with the kernel */
  visc_term.v = vec_set1(0.5f) * visc.v * (wi_dr.v + wj_dr.v) * ri.v;
  sph_term.v = (piPOrho.v * wi_dr.v + pjPOrho.v * wj_dr.v) * ri.v;

  /* Eventually get the acceleration */
  acc.v = visc_term.v + sph_term.v;
Matthieu Schaller's avatar
Matthieu Schaller committed
750

751
752
753
754
755
756
757
758
759
760
761
762
  /* Use the force, Luke! */
  for (k = 0; k < 3; k++) {
    f.v = dx[k].v * acc.v;
    pia[k].v = mj.v * f.v;
    pja[k].v = mi.v * f.v;
  }

  /* Get the time derivative for h. */
  pih_dt.v = mj.v * dvdr.v * ri.v / pjrho.v * wi_dr.v;
  pjh_dt.v = mi.v * dvdr.v * ri.v / pirho.v * wj_dr.v;

  /* Change in entropy */
763
  entropy_dt.v = visc_term.v * dvdr.v;
Matthieu Schaller's avatar
Matthieu Schaller committed
764

765
766
767
768
769
770
  /* Store the forces back on the particles. */
  for (k = 0; k < VEC_SIZE; k++) {
    for (j = 0; j < 3; j++) {
      pi[k]->a_hydro[j] -= pia[j].f[k];
      pj[k]->a_hydro[j] += pja[j].f[k];
    }
771
772
    pi[k]->force.h_dt -= pih_dt.f[k];
    pj[k]->force.h_dt -= pjh_dt.f[k];
773
774
    pi[k]->force.v_sig = max(pi[k]->force.v_sig, v_sig.f[k]);
    pj[k]->force.v_sig = max(pj[k]->force.v_sig, v_sig.f[k]);
775
    pi[k]->entropy_dt += entropy_dt.f[k] * mj.f[k];
776
    pj[k]->entropy_dt += entropy_dt.f[k] * mi.f[k];
777
778
  }

Matthieu Schaller's avatar
Matthieu Schaller committed
779
#else
780

Matthieu Schaller's avatar
Matthieu Schaller committed
781
782
  error(
      "The Gadget2 serial version of runner_iact_nonsym_force was called when "
783
      "the vectorised version should have been used.");
784
785

#endif
786
787
}

788
789
790
/**
 * @brief Force loop (non-symmetric version)
 */
791
792
793
__attribute__((always_inline)) INLINE static void runner_iact_nonsym_force(
    float r2, float *dx, float hi, float hj, struct part *pi, struct part *pj) {

794
795
796
  float wi, wj, wi_dx, wj_dx;

  const float fac_mu = 1.f; /* Will change with cosmological integration */
797

798
799
800
801
  const float r = sqrtf(r2);
  const float r_inv = 1.0f / r;

  /* Get some values in local variables. */
802
  // const float mi = pi->mass;
803
804
805
806
807
808
  const float mj = pj->mass;
  const float rhoi = pi->rho;
  const float rhoj = pj->rho;

  /* Get the kernel for hi. */
  const float hi_inv = 1.0f / hi;
809
  const float hid_inv = pow_dimension_plus_one(hi_inv); /* 1/h^(d+1) */
810
811
  const float ui = r * hi_inv;
  kernel_deval(ui, &wi, &wi_dx);
812
  const float wi_dr = hid_inv * wi_dx;
813
814
815

  /* Get the kernel for hj. */
  const float hj_inv = 1.0f / hj;
816
  const float hjd_inv = pow_dimension_plus_one(hj_inv); /* 1/h^(d+1) */
817
818
  const float xj = r * hj_inv;
  kernel_deval(xj, &wj, &wj_dx);
819
  const float wj_dr = hjd_inv * wj_dx;
820

821
822
823
824
825
  /* Compute h-gradient terms */
  const float f_i = pi->force.f;
  const float f_j = pj->force.f;

  /* Compute pressure terms */
826
827
  const float P_over_rho2_i = pi->force.P_over_rho2;
  const float P_over_rho2_j = pj->force.P_over_rho2;
828
829

  /* Compute sound speeds */
830
831
  const float ci = pi->force.soundspeed;
  const float cj = pj->force.soundspeed;
832

833
  /* Compute dv dot r. */
834
835
836
  const float dvdr = (pi->v[0] - pj->v[0]) * dx[0] +
                     (pi->v[1] - pj->v[1]) * dx[1] +
                     (pi->v[2] - pj->v[2]) * dx[2];
837

838
  /* Balsara term */
839
840
  const float balsara_i = pi->force.balsara;
  const float balsara_j = pj->force.balsara;
841
842

  /* Are the particles moving towards each others ? */
843
  const float omega_ij = (dvdr < 0.f) ? dvdr : 0.f;
844
845
846
847
848
849
850
851
852
  const float mu_ij = fac_mu * r_inv * omega_ij; /* This is 0 or negative */

  /* Signal velocity */
  const float v_sig = ci + cj - 3.f * mu_ij;

  /* Now construct the full viscosity term */
  const float rho_ij = 0.5f * (rhoi + rhoj);
  const float visc = -0.25f * const_viscosity_alpha * v_sig * mu_ij *
                     (balsara_i + balsara_j) / rho_ij;
853
854

  /* Now, convolve with the kernel */
855
  const float visc_term = 0.5f * visc * (wi_dr + wj_dr) * r_inv;
856
  const float sph_term =
857
      (f_i * P_over_rho2_i * wi_dr + f_j * P_over_rho2_j * wj_dr) * r_inv;
858
859
860

  /* Eventually got the acceleration */
  const float acc = visc_term + sph_term;
861

862
  /* Use the force Luke ! */
863
864
865
  pi->a_hydro[0] -= mj * acc * dx[0];
  pi->a_hydro[1] -= mj * acc * dx[1];
  pi->a_hydro[2] -= mj * acc * dx[2];
866

867
  /* Get the time derivative for h. */
868
  pi->force.h_dt -= mj * dvdr * r_inv / rhoj * wi_dr;
869

870
  /* Update the signal velocity. */
871
  pi->force.v_sig = (pi->force.v_sig > v_sig) ? pi->force.v_sig : v_sig;
872

873
  /* Change in entropy */
874
  pi->entropy_dt += mj * visc_term * dvdr;
875
}
876

877
878
879
880
881
882
/**
 * @brief Force loop (Vectorized non-symmetric version)
 */
__attribute__((always_inline)) INLINE static void runner_iact_nonsym_vec_force(
    float *R2, float *Dx, float *Hi, float *Hj, struct part **pi,
    struct part **pj) {
883

Matthieu Schaller's avatar
Matthieu Schaller committed
884
#ifdef WITH_VECTORIZATION
885
886
887
888

  vector r, r2, ri;
  vector xi, xj;
  vector hi, hj, hi_inv, hj_inv;
889
  vector hid_inv, hjd_inv;
890
891
892
893
894
895
896
897
  vector wi, wj, wi_dx, wj_dx, wi_dr, wj_dr, dvdr;
  vector piPOrho, pjPOrho, pirho, pjrho;
  vector mj;
  vector f;
  vector dx[3];
  vector vi[3], vj[3];
  vector pia[3];
  vector pih_dt;
898
899
  vector ci, cj, v_sig;
  vector omega_ij, mu_ij, fac_mu, balsara;
900
901
902
903
904
  vector rho_ij, visc, visc_term, sph_term, acc, entropy_dt;
  int j, k;

  fac_mu.v = vec_set1(1.f); /* Will change with cosmological integration */

Matthieu Schaller's avatar
Matthieu Schaller committed
905
/* Load stuff. */
906
907
908
#if VEC_SIZE == 8
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass,
                 pj[4]->mass, pj[5]->mass, pj[6]->mass, pj[7]->mass);
Matthieu Schaller's avatar
Matthieu Schaller committed
909
910
911
912
913
914
915
916
  piPOrho.v = vec_set(pi[0]->force.P_over_rho2, pi[1]->force.P_over_rho2,
                      pi[2]->force.P_over_rho2, pi[3]->force.P_over_rho2,
                      pi[4]->force.P_over_rho2, pi[5]->force.P_over_rho2,
                      pi[6]->force.P_over_rho2, pi[7]->force.P_over_rho2);
  pjPOrho.v = vec_set(pj[0]->force.P_over_rho2, pj[1]->force.P_over_rho2,
                      pj[2]->force.P_over_rho2, pj[3]->force.P_over_rho2,
                      pj[4]->force.P_over_rho2, pj[5]->force.P_over_rho2,
                      pj[6]->force.P_over_rho2, pj[7]->force.P_over_rho2);
917
918
919
920
  pirho.v = vec_set(pi[0]->rho, pi[1]->rho, pi[2]->rho, pi[3]->rho, pi[4]->rho,
                    pi[5]->rho, pi[6]->rho, pi[7]->rho);
  pjrho.v = vec_set(pj[0]->rho, pj[1]->rho, pj[2]->rho, pj[3]->rho, pj[4]->rho,
                    pj[5]->rho, pj[6]->rho, pj[7]->rho);
Matthieu Schaller's avatar
Matthieu Schaller committed
921
922
923
924
925
926
927
928
  ci.v = vec_set(pi[0]->force.soundspeed, pi[1]->force.soundspeed,
                 pi[2]->force.soundspeed, pi[3]->force.soundspeed,
                 pi[4]->force.soundspeed, pi[5]->force.soundspeed,
                 pi[6]->force.soundspeed, pi[7]->force.soundspeed);
  cj.v = vec_set(pj[0]->force.soundspeed, pj[1]->force.soundspeed,
                 pj[2]->force.soundspeed, pj[3]->force.soundspeed,
                 pj[4]->force.soundspeed, pj[5]->force.soundspeed,
                 pj[6]->force.soundspeed, pj[7]->force.soundspeed);
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k],
                      pi[4]->v[k], pi[5]->v[k], pi[6]->v[k], pi[7]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k],
                      pj[4]->v[k], pj[5]->v[k], pj[6]->v[k], pj[7]->v[k]);
  }
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k], Dx[12 + k],
                      Dx[15 + k], Dx[18 + k], Dx[21 + k]);
  balsara.v =
      vec_set(pi[0]->force.balsara, pi[1]->force.balsara, pi[2]->force.balsara,
              pi[3]->force.balsara, pi[4]->force.balsara, pi[5]->force.balsara,
              pi[6]->force.balsara, pi[7]->force.balsara) +
      vec_set(pj[0]->force.balsara, pj[1]->force.balsara, pj[2]->force.balsara,
              pj[3]->force.balsara, pj[4]->force.balsara, pj[5]->force.balsara,
              pj[6]->force.balsara, pj[7]->force.balsara);
#elif VEC_SIZE == 4
  mj.v = vec_set(pj[0]->mass, pj[1]->mass, pj[2]->mass, pj[3]->mass);
947
  piPOrho.v = vec_set(pi[0]->force.P_over_rho2, pi[1]->force.P_over_rho2,
Matthieu Schaller's avatar
Matthieu Schaller committed
948
                      pi[2]->force.P_over_rho2, pi[3]->force.P_over_rho2);
949
  pjPOrho.v = vec_set(pj[0]->force.P_over_rho2, pj[1]->force.P_over_rho2,
Matthieu Schaller's avatar
Matthieu Schaller committed
950
                      pj[2]->force.P_over_rho2, pj[3]->force.P_over_rho2);
951
952
  pirho.v = vec_set(pi[0]->rho, pi[1]->rho, pi[2]->rho, pi[3]->rho);
  pjrho.v = vec_set(pj[0]->rho, pj[1]->rho, pj[2]->rho, pj[3]->rho);
Matthieu Schaller's avatar
Matthieu Schaller committed
953
954
955
956
  ci.v = vec_set(pi[0]->force.soundspeed, pi[1]->force.soundspeed,
                 pi[2]->force.soundspeed, pi[3]->force.soundspeed);
  cj.v = vec_set(pj[0]->force.soundspeed, pj[1]->force.soundspeed,
                 pj[2]->force.soundspeed, pj[3]->force.soundspeed);
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
  for (k = 0; k < 3; k++) {
    vi[k].v = vec_set(pi[0]->v[k], pi[1]->v[k], pi[2]->v[k], pi[3]->v[k]);
    vj[k].v = vec_set(pj[0]->v[k], pj[1]->v[k], pj[2]->v[k], pj[3]->v[k]);
  }
  for (k = 0; k < 3; k++)
    dx[k].v = vec_set(Dx[0 + k], Dx[3 + k], Dx[6 + k], Dx[9 + k]);
  balsara.v = vec_set(pi[0]->force.balsara, pi[1]->force.balsara,
                      pi[2]->force.balsara, pi[3]->force.balsara) +
              vec_set(pj[0]->force.balsara, pj[1]->force.balsara,
                      pj[2]->force.balsara, pj[3]->force.balsara);
#else
  error("Unknown vector size.")
#endif

  /* Get the radius and inverse radius. */
  r2.v = vec_load(R2);
  ri.v = vec_rsqrt(r2.v);
  ri.v = ri.v - vec_set1(0.5f) * ri.v * (r2.v * ri.v * ri.v - vec_set1(1.0f));
  r.v = r2.v * ri.v;

  /* Get the kernel for hi. */
  hi.v = vec_load(Hi);
  hi_inv.v = vec_rcp(hi.v);
  hi_inv.v = hi_inv.v - hi_inv.v * (hi.v * hi_inv.v - vec_set1(1.0f));
981
  hid_inv = pow_dimension_plus_one_vec(hi_inv);
982
983
  xi.v = r.v * hi_inv.v;
  kernel_deval_vec(&xi, &wi, &wi_dx);
984
  wi_dr.v = hid_inv.v * wi_dx.v;
985
986
987
988
989

  /* Get the kernel for hj. */
  hj.v = vec_load(Hj);
  hj_inv.v = vec_rcp(hj.v);
  hj_inv.v = hj_inv.v - hj_inv.v * (hj.v * hj_inv.v - vec_set1(1.0f));
990
  hjd_inv = pow_dimension_plus_one_vec(hj_inv);
991
992
  xj.v = r.v * hj_inv.v;
  kernel_deval_vec(&xj, &wj, &wj_dx);
993
  wj_dr.v = hjd_inv.v * wj_dx.v;
994
995
996
997

  /* Compute dv dot r. */
  dvdr.v = ((vi[0].v - vj[0].v) * dx[0].v) + ((vi[1].v - vj[1].v) * dx[1].v) +
           ((vi[2].v - vj[2].v) * dx[2].v);
Matthieu Schaller's avatar
Matthieu Schaller committed
998
  // dvdr.v = dvdr.v * ri.v;
999
1000
1001
1002
1003

  /* Compute the relative velocity. (This is 0 if the particles move away from
   * each other and negative otherwise) */
  omega_ij.v = vec_fmin(dvdr.v, vec_set1(0.0f));
  mu_ij.v = fac_mu.v * ri.v * omega_ij.v; /* This is 0 or negative */
Matthieu Schaller's avatar
Matthieu Schaller committed
1004

1005
1006
  /* Compute signal velocity */
  v_sig.v = ci.v + cj.v - vec_set1(3.0f) * mu_ij.v;
Matthieu Schaller's avatar
Matthieu Schaller committed
1007

1008
1009
  /* Now construct the full viscosity term */
  rho_ij.v = vec_set1(0.5f) * (pirho.v + pjrho.v);
Matthieu Schaller's avatar
Matthieu Schaller committed
1010
1011
  visc.v = vec_set1(-0.25f) * vec_set1(const_viscosity_alpha) * v_sig.v *
           mu_ij.v * balsara.v / rho_ij.v;
1012
1013
1014
1015
1016
1017
1018

  /* Now, convolve with the kernel */
  visc_term.v = vec_set1(0.5f) * visc.v * (wi_dr.v + wj_dr.v) * ri.v;
  sph_term.v = (piPOrho.v * wi_dr.v + pjPOrho.v * wj_dr.v) * ri.v;

  /* Eventually get the acceleration */
  acc.v = visc_term.v + sph_term.v;
Matthieu Schaller's avatar
Matthieu Schaller committed
1019

1020
1021
1022
1023
1024
1025
1026
  /* Use the force, Luke! */
  for (k = 0; k < 3; k++) {
    f.v = dx[k].v * acc.v;
    pia[k].v = mj.v * f.v;
  }

  /* Get the time derivative for h. */
1027
  pih_dt.v = mj.v * dvdr.v * ri.v / pjrho.v * wi_dr.v;
1028
1029

  /* Change in entropy */
1030
  entropy_dt.v = mj.v * visc_term.v * dvdr.v;
Matthieu Schaller's avatar
Matthieu Schaller committed
1031

1032
1033
1034
  /* Store the forces back on the particles. */
  for (k = 0; k < VEC_SIZE; k++) {
    for (j = 0; j < 3; j++) pi[k]->a_hydro[j] -= pia[j].f[k];
1035
    pi[k]->force.h_dt -= pih_dt.f[k];
1036
    pi[k]->force.v_sig = max(pi[k]->force.v_sig, v_sig.f[k]);