?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Fix segmentation fault in minimize
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_avx_128_fma_double / nb_kernel_ElecRF_VdwNone_GeomW3W3_avx_128_fma_double.cpp
blobda6f3df822b22ae5006d765ce9e3d9ae5fdca6be
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
8 *
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
13 *
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
23 *
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
31 *
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
34 */
35 /*
36 * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
37 */
38 #include "gmxpre.h"
39
40 #include "config.h"
41
42 #include <math.h>
43
44 #include "../nb_kernel.h"
45 #include "gromacs/gmxlib/nrnb.h"
46
47 #include "kernelutil_x86_avx_128_fma_double.h"
48
49 /*
50 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_avx_128_fma_double
51 * Electrostatics interaction: ReactionField
52 * VdW interaction: None
53 * Geometry: Water3-Water3
54 * Calculate force/pot: PotentialAndForce
55 */
56 void
57 nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_avx_128_fma_double
58 (t_nblist * gmx_restrict nlist,
59 rvec * gmx_restrict xx,
60 rvec * gmx_restrict ff,
61 struct t_forcerec * gmx_restrict fr,
62 t_mdatoms * gmx_restrict mdatoms,
63 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
64 t_nrnb * gmx_restrict nrnb)
65 {
66 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
67 * just 0 for non-waters.
68 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
69 * jnr indices corresponding to data put in the four positions in the SIMD register.
70 */
71 int i_shift_offset,i_coord_offset,outeriter,inneriter;
72 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
73 int jnrA,jnrB;
74 int j_coord_offsetA,j_coord_offsetB;
75 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
76 real rcutoff_scalar;
77 real *shiftvec,*fshift,*x,*f;
78 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
79 int vdwioffset0;
80 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
81 int vdwioffset1;
82 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
83 int vdwioffset2;
84 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
85 int vdwjidx0A,vdwjidx0B;
86 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
87 int vdwjidx1A,vdwjidx1B;
88 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
89 int vdwjidx2A,vdwjidx2B;
90 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
91 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
92 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
93 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
94 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
95 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
96 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
97 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
98 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
99 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
100 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
101 real *charge;
102 __m128d dummy_mask,cutoff_mask;
103 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
104 __m128d one = _mm_set1_pd(1.0);
105 __m128d two = _mm_set1_pd(2.0);
106 x = xx[0];
107 f = ff[0];
108
109 nri = nlist->nri;
110 iinr = nlist->iinr;
111 jindex = nlist->jindex;
112 jjnr = nlist->jjnr;
113 shiftidx = nlist->shift;
114 gid = nlist->gid;
115 shiftvec = fr->shift_vec[0];
116 fshift = fr->fshift[0];
117 facel = _mm_set1_pd(fr->ic->epsfac);
118 charge = mdatoms->chargeA;
119 krf = _mm_set1_pd(fr->ic->k_rf);
120 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
121 crf = _mm_set1_pd(fr->ic->c_rf);
122
123 /* Setup water-specific parameters */
124 inr = nlist->iinr[0];
125 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
126 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
127 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
128
129 jq0 = _mm_set1_pd(charge[inr+0]);
130 jq1 = _mm_set1_pd(charge[inr+1]);
131 jq2 = _mm_set1_pd(charge[inr+2]);
132 qq00 = _mm_mul_pd(iq0,jq0);
133 qq01 = _mm_mul_pd(iq0,jq1);
134 qq02 = _mm_mul_pd(iq0,jq2);
135 qq10 = _mm_mul_pd(iq1,jq0);
136 qq11 = _mm_mul_pd(iq1,jq1);
137 qq12 = _mm_mul_pd(iq1,jq2);
138 qq20 = _mm_mul_pd(iq2,jq0);
139 qq21 = _mm_mul_pd(iq2,jq1);
140 qq22 = _mm_mul_pd(iq2,jq2);
141
142 /* Avoid stupid compiler warnings */
143 jnrA = jnrB = 0;
144 j_coord_offsetA = 0;
145 j_coord_offsetB = 0;
146
147 outeriter = 0;
148 inneriter = 0;
149
150 /* Start outer loop over neighborlists */
151 for(iidx=0; iidx<nri; iidx++)
152 {
153 /* Load shift vector for this list */
154 i_shift_offset = DIM*shiftidx[iidx];
155
156 /* Load limits for loop over neighbors */
157 j_index_start = jindex[iidx];
158 j_index_end = jindex[iidx+1];
159
160 /* Get outer coordinate index */
161 inr = iinr[iidx];
162 i_coord_offset = DIM*inr;
163
164 /* Load i particle coords and add shift vector */
165 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
166 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
167
168 fix0 = _mm_setzero_pd();
169 fiy0 = _mm_setzero_pd();
170 fiz0 = _mm_setzero_pd();
171 fix1 = _mm_setzero_pd();
172 fiy1 = _mm_setzero_pd();
173 fiz1 = _mm_setzero_pd();
174 fix2 = _mm_setzero_pd();
175 fiy2 = _mm_setzero_pd();
176 fiz2 = _mm_setzero_pd();
177
178 /* Reset potential sums */
179 velecsum = _mm_setzero_pd();
180
181 /* Start inner kernel loop */
182 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
183 {
184
185 /* Get j neighbor index, and coordinate index */
186 jnrA = jjnr[jidx];
187 jnrB = jjnr[jidx+1];
188 j_coord_offsetA = DIM*jnrA;
189 j_coord_offsetB = DIM*jnrB;
190
191 /* load j atom coordinates */
192 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
193 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
194
195 /* Calculate displacement vector */
196 dx00 = _mm_sub_pd(ix0,jx0);
197 dy00 = _mm_sub_pd(iy0,jy0);
198 dz00 = _mm_sub_pd(iz0,jz0);
199 dx01 = _mm_sub_pd(ix0,jx1);
200 dy01 = _mm_sub_pd(iy0,jy1);
201 dz01 = _mm_sub_pd(iz0,jz1);
202 dx02 = _mm_sub_pd(ix0,jx2);
203 dy02 = _mm_sub_pd(iy0,jy2);
204 dz02 = _mm_sub_pd(iz0,jz2);
205 dx10 = _mm_sub_pd(ix1,jx0);
206 dy10 = _mm_sub_pd(iy1,jy0);
207 dz10 = _mm_sub_pd(iz1,jz0);
208 dx11 = _mm_sub_pd(ix1,jx1);
209 dy11 = _mm_sub_pd(iy1,jy1);
210 dz11 = _mm_sub_pd(iz1,jz1);
211 dx12 = _mm_sub_pd(ix1,jx2);
212 dy12 = _mm_sub_pd(iy1,jy2);
213 dz12 = _mm_sub_pd(iz1,jz2);
214 dx20 = _mm_sub_pd(ix2,jx0);
215 dy20 = _mm_sub_pd(iy2,jy0);
216 dz20 = _mm_sub_pd(iz2,jz0);
217 dx21 = _mm_sub_pd(ix2,jx1);
218 dy21 = _mm_sub_pd(iy2,jy1);
219 dz21 = _mm_sub_pd(iz2,jz1);
220 dx22 = _mm_sub_pd(ix2,jx2);
221 dy22 = _mm_sub_pd(iy2,jy2);
222 dz22 = _mm_sub_pd(iz2,jz2);
223
224 /* Calculate squared distance and things based on it */
225 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
226 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
227 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
228 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
229 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
230 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
231 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
232 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
233 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
234
235 rinv00 = avx128fma_invsqrt_d(rsq00);
236 rinv01 = avx128fma_invsqrt_d(rsq01);
237 rinv02 = avx128fma_invsqrt_d(rsq02);
238 rinv10 = avx128fma_invsqrt_d(rsq10);
239 rinv11 = avx128fma_invsqrt_d(rsq11);
240 rinv12 = avx128fma_invsqrt_d(rsq12);
241 rinv20 = avx128fma_invsqrt_d(rsq20);
242 rinv21 = avx128fma_invsqrt_d(rsq21);
243 rinv22 = avx128fma_invsqrt_d(rsq22);
244
245 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
246 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
247 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
248 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
249 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
250 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
251 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
252 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
253 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
254
255 fjx0 = _mm_setzero_pd();
256 fjy0 = _mm_setzero_pd();
257 fjz0 = _mm_setzero_pd();
258 fjx1 = _mm_setzero_pd();
259 fjy1 = _mm_setzero_pd();
260 fjz1 = _mm_setzero_pd();
261 fjx2 = _mm_setzero_pd();
262 fjy2 = _mm_setzero_pd();
263 fjz2 = _mm_setzero_pd();
264
265 /**************************
266 * CALCULATE INTERACTIONS *
267 **************************/
268
269 /* REACTION-FIELD ELECTROSTATICS */
270 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
271 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
272
273 /* Update potential sum for this i atom from the interaction with this j atom. */
274 velecsum = _mm_add_pd(velecsum,velec);
275
276 fscal = felec;
277
278 /* Update vectorial force */
279 fix0 = _mm_macc_pd(dx00,fscal,fix0);
280 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
281 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
282
283 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
284 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
285 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
286
287 /**************************
288 * CALCULATE INTERACTIONS *
289 **************************/
290
291 /* REACTION-FIELD ELECTROSTATICS */
292 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
293 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
294
295 /* Update potential sum for this i atom from the interaction with this j atom. */
296 velecsum = _mm_add_pd(velecsum,velec);
297
298 fscal = felec;
299
300 /* Update vectorial force */
301 fix0 = _mm_macc_pd(dx01,fscal,fix0);
302 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
303 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
304
305 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
306 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
307 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
308
309 /**************************
310 * CALCULATE INTERACTIONS *
311 **************************/
312
313 /* REACTION-FIELD ELECTROSTATICS */
314 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
315 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
316
317 /* Update potential sum for this i atom from the interaction with this j atom. */
318 velecsum = _mm_add_pd(velecsum,velec);
319
320 fscal = felec;
321
322 /* Update vectorial force */
323 fix0 = _mm_macc_pd(dx02,fscal,fix0);
324 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
325 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
326
327 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
328 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
329 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
330
331 /**************************
332 * CALCULATE INTERACTIONS *
333 **************************/
334
335 /* REACTION-FIELD ELECTROSTATICS */
336 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
337 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
338
339 /* Update potential sum for this i atom from the interaction with this j atom. */
340 velecsum = _mm_add_pd(velecsum,velec);
341
342 fscal = felec;
343
344 /* Update vectorial force */
345 fix1 = _mm_macc_pd(dx10,fscal,fix1);
346 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
347 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
348
349 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
350 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
351 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
352
353 /**************************
354 * CALCULATE INTERACTIONS *
355 **************************/
356
357 /* REACTION-FIELD ELECTROSTATICS */
358 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
359 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
360
361 /* Update potential sum for this i atom from the interaction with this j atom. */
362 velecsum = _mm_add_pd(velecsum,velec);
363
364 fscal = felec;
365
366 /* Update vectorial force */
367 fix1 = _mm_macc_pd(dx11,fscal,fix1);
368 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
369 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
370
371 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
372 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
373 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
374
375 /**************************
376 * CALCULATE INTERACTIONS *
377 **************************/
378
379 /* REACTION-FIELD ELECTROSTATICS */
380 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
381 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
382
383 /* Update potential sum for this i atom from the interaction with this j atom. */
384 velecsum = _mm_add_pd(velecsum,velec);
385
386 fscal = felec;
387
388 /* Update vectorial force */
389 fix1 = _mm_macc_pd(dx12,fscal,fix1);
390 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
391 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
392
393 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
394 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
395 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
396
397 /**************************
398 * CALCULATE INTERACTIONS *
399 **************************/
400
401 /* REACTION-FIELD ELECTROSTATICS */
402 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
403 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
404
405 /* Update potential sum for this i atom from the interaction with this j atom. */
406 velecsum = _mm_add_pd(velecsum,velec);
407
408 fscal = felec;
409
410 /* Update vectorial force */
411 fix2 = _mm_macc_pd(dx20,fscal,fix2);
412 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
413 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
414
415 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
416 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
417 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
418
419 /**************************
420 * CALCULATE INTERACTIONS *
421 **************************/
422
423 /* REACTION-FIELD ELECTROSTATICS */
424 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
425 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
426
427 /* Update potential sum for this i atom from the interaction with this j atom. */
428 velecsum = _mm_add_pd(velecsum,velec);
429
430 fscal = felec;
431
432 /* Update vectorial force */
433 fix2 = _mm_macc_pd(dx21,fscal,fix2);
434 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
435 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
436
437 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
438 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
439 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
440
441 /**************************
442 * CALCULATE INTERACTIONS *
443 **************************/
444
445 /* REACTION-FIELD ELECTROSTATICS */
446 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
447 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
448
449 /* Update potential sum for this i atom from the interaction with this j atom. */
450 velecsum = _mm_add_pd(velecsum,velec);
451
452 fscal = felec;
453
454 /* Update vectorial force */
455 fix2 = _mm_macc_pd(dx22,fscal,fix2);
456 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
457 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
458
459 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
460 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
461 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
462
463 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
464
465 /* Inner loop uses 315 flops */
466 }
467
468 if(jidx<j_index_end)
469 {
470
471 jnrA = jjnr[jidx];
472 j_coord_offsetA = DIM*jnrA;
473
474 /* load j atom coordinates */
475 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
476 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
477
478 /* Calculate displacement vector */
479 dx00 = _mm_sub_pd(ix0,jx0);
480 dy00 = _mm_sub_pd(iy0,jy0);
481 dz00 = _mm_sub_pd(iz0,jz0);
482 dx01 = _mm_sub_pd(ix0,jx1);
483 dy01 = _mm_sub_pd(iy0,jy1);
484 dz01 = _mm_sub_pd(iz0,jz1);
485 dx02 = _mm_sub_pd(ix0,jx2);
486 dy02 = _mm_sub_pd(iy0,jy2);
487 dz02 = _mm_sub_pd(iz0,jz2);
488 dx10 = _mm_sub_pd(ix1,jx0);
489 dy10 = _mm_sub_pd(iy1,jy0);
490 dz10 = _mm_sub_pd(iz1,jz0);
491 dx11 = _mm_sub_pd(ix1,jx1);
492 dy11 = _mm_sub_pd(iy1,jy1);
493 dz11 = _mm_sub_pd(iz1,jz1);
494 dx12 = _mm_sub_pd(ix1,jx2);
495 dy12 = _mm_sub_pd(iy1,jy2);
496 dz12 = _mm_sub_pd(iz1,jz2);
497 dx20 = _mm_sub_pd(ix2,jx0);
498 dy20 = _mm_sub_pd(iy2,jy0);
499 dz20 = _mm_sub_pd(iz2,jz0);
500 dx21 = _mm_sub_pd(ix2,jx1);
501 dy21 = _mm_sub_pd(iy2,jy1);
502 dz21 = _mm_sub_pd(iz2,jz1);
503 dx22 = _mm_sub_pd(ix2,jx2);
504 dy22 = _mm_sub_pd(iy2,jy2);
505 dz22 = _mm_sub_pd(iz2,jz2);
506
507 /* Calculate squared distance and things based on it */
508 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
509 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
510 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
511 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
512 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
513 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
514 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
515 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
516 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
517
518 rinv00 = avx128fma_invsqrt_d(rsq00);
519 rinv01 = avx128fma_invsqrt_d(rsq01);
520 rinv02 = avx128fma_invsqrt_d(rsq02);
521 rinv10 = avx128fma_invsqrt_d(rsq10);
522 rinv11 = avx128fma_invsqrt_d(rsq11);
523 rinv12 = avx128fma_invsqrt_d(rsq12);
524 rinv20 = avx128fma_invsqrt_d(rsq20);
525 rinv21 = avx128fma_invsqrt_d(rsq21);
526 rinv22 = avx128fma_invsqrt_d(rsq22);
527
528 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
529 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
530 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
531 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
532 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
533 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
534 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
535 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
536 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
537
538 fjx0 = _mm_setzero_pd();
539 fjy0 = _mm_setzero_pd();
540 fjz0 = _mm_setzero_pd();
541 fjx1 = _mm_setzero_pd();
542 fjy1 = _mm_setzero_pd();
543 fjz1 = _mm_setzero_pd();
544 fjx2 = _mm_setzero_pd();
545 fjy2 = _mm_setzero_pd();
546 fjz2 = _mm_setzero_pd();
547
548 /**************************
549 * CALCULATE INTERACTIONS *
550 **************************/
551
552 /* REACTION-FIELD ELECTROSTATICS */
553 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
554 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
555
556 /* Update potential sum for this i atom from the interaction with this j atom. */
557 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
558 velecsum = _mm_add_pd(velecsum,velec);
559
560 fscal = felec;
561
562 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
563
564 /* Update vectorial force */
565 fix0 = _mm_macc_pd(dx00,fscal,fix0);
566 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
567 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
568
569 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
570 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
571 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
572
573 /**************************
574 * CALCULATE INTERACTIONS *
575 **************************/
576
577 /* REACTION-FIELD ELECTROSTATICS */
578 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
579 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
580
581 /* Update potential sum for this i atom from the interaction with this j atom. */
582 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
583 velecsum = _mm_add_pd(velecsum,velec);
584
585 fscal = felec;
586
587 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
588
589 /* Update vectorial force */
590 fix0 = _mm_macc_pd(dx01,fscal,fix0);
591 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
592 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
593
594 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
595 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
596 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
597
598 /**************************
599 * CALCULATE INTERACTIONS *
600 **************************/
601
602 /* REACTION-FIELD ELECTROSTATICS */
603 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
604 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
605
606 /* Update potential sum for this i atom from the interaction with this j atom. */
607 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
608 velecsum = _mm_add_pd(velecsum,velec);
609
610 fscal = felec;
611
612 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
613
614 /* Update vectorial force */
615 fix0 = _mm_macc_pd(dx02,fscal,fix0);
616 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
617 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
618
619 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
620 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
621 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
622
623 /**************************
624 * CALCULATE INTERACTIONS *
625 **************************/
626
627 /* REACTION-FIELD ELECTROSTATICS */
628 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
629 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
630
631 /* Update potential sum for this i atom from the interaction with this j atom. */
632 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
633 velecsum = _mm_add_pd(velecsum,velec);
634
635 fscal = felec;
636
637 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
638
639 /* Update vectorial force */
640 fix1 = _mm_macc_pd(dx10,fscal,fix1);
641 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
642 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
643
644 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
645 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
646 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
647
648 /**************************
649 * CALCULATE INTERACTIONS *
650 **************************/
651
652 /* REACTION-FIELD ELECTROSTATICS */
653 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
654 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
655
656 /* Update potential sum for this i atom from the interaction with this j atom. */
657 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
658 velecsum = _mm_add_pd(velecsum,velec);
659
660 fscal = felec;
661
662 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
663
664 /* Update vectorial force */
665 fix1 = _mm_macc_pd(dx11,fscal,fix1);
666 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
667 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
668
669 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
670 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
671 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
672
673 /**************************
674 * CALCULATE INTERACTIONS *
675 **************************/
676
677 /* REACTION-FIELD ELECTROSTATICS */
678 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
679 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
680
681 /* Update potential sum for this i atom from the interaction with this j atom. */
682 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
683 velecsum = _mm_add_pd(velecsum,velec);
684
685 fscal = felec;
686
687 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
688
689 /* Update vectorial force */
690 fix1 = _mm_macc_pd(dx12,fscal,fix1);
691 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
692 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
693
694 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
695 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
696 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
697
698 /**************************
699 * CALCULATE INTERACTIONS *
700 **************************/
701
702 /* REACTION-FIELD ELECTROSTATICS */
703 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
704 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
705
706 /* Update potential sum for this i atom from the interaction with this j atom. */
707 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
708 velecsum = _mm_add_pd(velecsum,velec);
709
710 fscal = felec;
711
712 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
713
714 /* Update vectorial force */
715 fix2 = _mm_macc_pd(dx20,fscal,fix2);
716 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
717 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
718
719 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
720 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
721 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
722
723 /**************************
724 * CALCULATE INTERACTIONS *
725 **************************/
726
727 /* REACTION-FIELD ELECTROSTATICS */
728 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
729 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
730
731 /* Update potential sum for this i atom from the interaction with this j atom. */
732 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
733 velecsum = _mm_add_pd(velecsum,velec);
734
735 fscal = felec;
736
737 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
738
739 /* Update vectorial force */
740 fix2 = _mm_macc_pd(dx21,fscal,fix2);
741 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
742 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
743
744 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
745 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
746 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
747
748 /**************************
749 * CALCULATE INTERACTIONS *
750 **************************/
751
752 /* REACTION-FIELD ELECTROSTATICS */
753 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
754 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
755
756 /* Update potential sum for this i atom from the interaction with this j atom. */
757 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
758 velecsum = _mm_add_pd(velecsum,velec);
759
760 fscal = felec;
761
762 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
763
764 /* Update vectorial force */
765 fix2 = _mm_macc_pd(dx22,fscal,fix2);
766 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
767 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
768
769 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
770 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
771 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
772
773 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
774
775 /* Inner loop uses 315 flops */
776 }
777
778 /* End of innermost loop */
779
780 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
781 f+i_coord_offset,fshift+i_shift_offset);
782
783 ggid = gid[iidx];
784 /* Update potential energies */
785 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
786
787 /* Increment number of inner iterations */
788 inneriter += j_index_end - j_index_start;
789
790 /* Outer loop uses 19 flops */
791 }
792
793 /* Increment number of outer iterations */
794 outeriter += nri;
795
796 /* Update outer/inner flops */
797
798 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*315);
799 }
800 /*
801 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomW3W3_F_avx_128_fma_double
802 * Electrostatics interaction: ReactionField
803 * VdW interaction: None
804 * Geometry: Water3-Water3
805 * Calculate force/pot: Force
806 */
807 void
808 nb_kernel_ElecRF_VdwNone_GeomW3W3_F_avx_128_fma_double
809 (t_nblist * gmx_restrict nlist,
810 rvec * gmx_restrict xx,
811 rvec * gmx_restrict ff,
812 struct t_forcerec * gmx_restrict fr,
813 t_mdatoms * gmx_restrict mdatoms,
814 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
815 t_nrnb * gmx_restrict nrnb)
816 {
817 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
818 * just 0 for non-waters.
819 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
820 * jnr indices corresponding to data put in the four positions in the SIMD register.
821 */
822 int i_shift_offset,i_coord_offset,outeriter,inneriter;
823 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
824 int jnrA,jnrB;
825 int j_coord_offsetA,j_coord_offsetB;
826 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
827 real rcutoff_scalar;
828 real *shiftvec,*fshift,*x,*f;
829 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
830 int vdwioffset0;
831 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
832 int vdwioffset1;
833 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
834 int vdwioffset2;
835 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
836 int vdwjidx0A,vdwjidx0B;
837 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
838 int vdwjidx1A,vdwjidx1B;
839 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
840 int vdwjidx2A,vdwjidx2B;
841 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
842 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
843 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
844 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
845 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
846 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
847 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
848 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
849 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
850 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
851 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
852 real *charge;
853 __m128d dummy_mask,cutoff_mask;
854 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
855 __m128d one = _mm_set1_pd(1.0);
856 __m128d two = _mm_set1_pd(2.0);
857 x = xx[0];
858 f = ff[0];
859
860 nri = nlist->nri;
861 iinr = nlist->iinr;
862 jindex = nlist->jindex;
863 jjnr = nlist->jjnr;
864 shiftidx = nlist->shift;
865 gid = nlist->gid;
866 shiftvec = fr->shift_vec[0];
867 fshift = fr->fshift[0];
868 facel = _mm_set1_pd(fr->ic->epsfac);
869 charge = mdatoms->chargeA;
870 krf = _mm_set1_pd(fr->ic->k_rf);
871 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
872 crf = _mm_set1_pd(fr->ic->c_rf);
873
874 /* Setup water-specific parameters */
875 inr = nlist->iinr[0];
876 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
877 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
878 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
879
880 jq0 = _mm_set1_pd(charge[inr+0]);
881 jq1 = _mm_set1_pd(charge[inr+1]);
882 jq2 = _mm_set1_pd(charge[inr+2]);
883 qq00 = _mm_mul_pd(iq0,jq0);
884 qq01 = _mm_mul_pd(iq0,jq1);
885 qq02 = _mm_mul_pd(iq0,jq2);
886 qq10 = _mm_mul_pd(iq1,jq0);
887 qq11 = _mm_mul_pd(iq1,jq1);
888 qq12 = _mm_mul_pd(iq1,jq2);
889 qq20 = _mm_mul_pd(iq2,jq0);
890 qq21 = _mm_mul_pd(iq2,jq1);
891 qq22 = _mm_mul_pd(iq2,jq2);
892
893 /* Avoid stupid compiler warnings */
894 jnrA = jnrB = 0;
895 j_coord_offsetA = 0;
896 j_coord_offsetB = 0;
897
898 outeriter = 0;
899 inneriter = 0;
900
901 /* Start outer loop over neighborlists */
902 for(iidx=0; iidx<nri; iidx++)
903 {
904 /* Load shift vector for this list */
905 i_shift_offset = DIM*shiftidx[iidx];
906
907 /* Load limits for loop over neighbors */
908 j_index_start = jindex[iidx];
909 j_index_end = jindex[iidx+1];
910
911 /* Get outer coordinate index */
912 inr = iinr[iidx];
913 i_coord_offset = DIM*inr;
914
915 /* Load i particle coords and add shift vector */
916 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
917 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
918
919 fix0 = _mm_setzero_pd();
920 fiy0 = _mm_setzero_pd();
921 fiz0 = _mm_setzero_pd();
922 fix1 = _mm_setzero_pd();
923 fiy1 = _mm_setzero_pd();
924 fiz1 = _mm_setzero_pd();
925 fix2 = _mm_setzero_pd();
926 fiy2 = _mm_setzero_pd();
927 fiz2 = _mm_setzero_pd();
928
929 /* Start inner kernel loop */
930 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
931 {
932
933 /* Get j neighbor index, and coordinate index */
934 jnrA = jjnr[jidx];
935 jnrB = jjnr[jidx+1];
936 j_coord_offsetA = DIM*jnrA;
937 j_coord_offsetB = DIM*jnrB;
938
939 /* load j atom coordinates */
940 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
941 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
942
943 /* Calculate displacement vector */
944 dx00 = _mm_sub_pd(ix0,jx0);
945 dy00 = _mm_sub_pd(iy0,jy0);
946 dz00 = _mm_sub_pd(iz0,jz0);
947 dx01 = _mm_sub_pd(ix0,jx1);
948 dy01 = _mm_sub_pd(iy0,jy1);
949 dz01 = _mm_sub_pd(iz0,jz1);
950 dx02 = _mm_sub_pd(ix0,jx2);
951 dy02 = _mm_sub_pd(iy0,jy2);
952 dz02 = _mm_sub_pd(iz0,jz2);
953 dx10 = _mm_sub_pd(ix1,jx0);
954 dy10 = _mm_sub_pd(iy1,jy0);
955 dz10 = _mm_sub_pd(iz1,jz0);
956 dx11 = _mm_sub_pd(ix1,jx1);
957 dy11 = _mm_sub_pd(iy1,jy1);
958 dz11 = _mm_sub_pd(iz1,jz1);
959 dx12 = _mm_sub_pd(ix1,jx2);
960 dy12 = _mm_sub_pd(iy1,jy2);
961 dz12 = _mm_sub_pd(iz1,jz2);
962 dx20 = _mm_sub_pd(ix2,jx0);
963 dy20 = _mm_sub_pd(iy2,jy0);
964 dz20 = _mm_sub_pd(iz2,jz0);
965 dx21 = _mm_sub_pd(ix2,jx1);
966 dy21 = _mm_sub_pd(iy2,jy1);
967 dz21 = _mm_sub_pd(iz2,jz1);
968 dx22 = _mm_sub_pd(ix2,jx2);
969 dy22 = _mm_sub_pd(iy2,jy2);
970 dz22 = _mm_sub_pd(iz2,jz2);
971
972 /* Calculate squared distance and things based on it */
973 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
974 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
975 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
976 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
977 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
978 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
979 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
980 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
981 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
982
983 rinv00 = avx128fma_invsqrt_d(rsq00);
984 rinv01 = avx128fma_invsqrt_d(rsq01);
985 rinv02 = avx128fma_invsqrt_d(rsq02);
986 rinv10 = avx128fma_invsqrt_d(rsq10);
987 rinv11 = avx128fma_invsqrt_d(rsq11);
988 rinv12 = avx128fma_invsqrt_d(rsq12);
989 rinv20 = avx128fma_invsqrt_d(rsq20);
990 rinv21 = avx128fma_invsqrt_d(rsq21);
991 rinv22 = avx128fma_invsqrt_d(rsq22);
992
993 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
994 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
995 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
996 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
997 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
998 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
999 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1000 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1001 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1002
1003 fjx0 = _mm_setzero_pd();
1004 fjy0 = _mm_setzero_pd();
1005 fjz0 = _mm_setzero_pd();
1006 fjx1 = _mm_setzero_pd();
1007 fjy1 = _mm_setzero_pd();
1008 fjz1 = _mm_setzero_pd();
1009 fjx2 = _mm_setzero_pd();
1010 fjy2 = _mm_setzero_pd();
1011 fjz2 = _mm_setzero_pd();
1012
1013 /**************************
1014 * CALCULATE INTERACTIONS *
1015 **************************/
1016
1017 /* REACTION-FIELD ELECTROSTATICS */
1018 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1019
1020 fscal = felec;
1021
1022 /* Update vectorial force */
1023 fix0 = _mm_macc_pd(dx00,fscal,fix0);
1024 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
1025 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
1026
1027 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
1028 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
1029 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
1030
1031 /**************************
1032 * CALCULATE INTERACTIONS *
1033 **************************/
1034
1035 /* REACTION-FIELD ELECTROSTATICS */
1036 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1037
1038 fscal = felec;
1039
1040 /* Update vectorial force */
1041 fix0 = _mm_macc_pd(dx01,fscal,fix0);
1042 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
1043 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
1044
1045 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
1046 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
1047 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
1048
1049 /**************************
1050 * CALCULATE INTERACTIONS *
1051 **************************/
1052
1053 /* REACTION-FIELD ELECTROSTATICS */
1054 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1055
1056 fscal = felec;
1057
1058 /* Update vectorial force */
1059 fix0 = _mm_macc_pd(dx02,fscal,fix0);
1060 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
1061 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
1062
1063 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
1064 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
1065 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
1066
1067 /**************************
1068 * CALCULATE INTERACTIONS *
1069 **************************/
1070
1071 /* REACTION-FIELD ELECTROSTATICS */
1072 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1073
1074 fscal = felec;
1075
1076 /* Update vectorial force */
1077 fix1 = _mm_macc_pd(dx10,fscal,fix1);
1078 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
1079 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
1080
1081 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
1082 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
1083 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
1084
1085 /**************************
1086 * CALCULATE INTERACTIONS *
1087 **************************/
1088
1089 /* REACTION-FIELD ELECTROSTATICS */
1090 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1091
1092 fscal = felec;
1093
1094 /* Update vectorial force */
1095 fix1 = _mm_macc_pd(dx11,fscal,fix1);
1096 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
1097 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
1098
1099 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
1100 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
1101 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
1102
1103 /**************************
1104 * CALCULATE INTERACTIONS *
1105 **************************/
1106
1107 /* REACTION-FIELD ELECTROSTATICS */
1108 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1109
1110 fscal = felec;
1111
1112 /* Update vectorial force */
1113 fix1 = _mm_macc_pd(dx12,fscal,fix1);
1114 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
1115 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
1116
1117 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
1118 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
1119 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
1120
1121 /**************************
1122 * CALCULATE INTERACTIONS *
1123 **************************/
1124
1125 /* REACTION-FIELD ELECTROSTATICS */
1126 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1127
1128 fscal = felec;
1129
1130 /* Update vectorial force */
1131 fix2 = _mm_macc_pd(dx20,fscal,fix2);
1132 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
1133 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
1134
1135 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
1136 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
1137 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
1138
1139 /**************************
1140 * CALCULATE INTERACTIONS *
1141 **************************/
1142
1143 /* REACTION-FIELD ELECTROSTATICS */
1144 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1145
1146 fscal = felec;
1147
1148 /* Update vectorial force */
1149 fix2 = _mm_macc_pd(dx21,fscal,fix2);
1150 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
1151 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
1152
1153 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
1154 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
1155 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
1156
1157 /**************************
1158 * CALCULATE INTERACTIONS *
1159 **************************/
1160
1161 /* REACTION-FIELD ELECTROSTATICS */
1162 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1163
1164 fscal = felec;
1165
1166 /* Update vectorial force */
1167 fix2 = _mm_macc_pd(dx22,fscal,fix2);
1168 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
1169 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
1170
1171 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
1172 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
1173 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
1174
1175 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1176
1177 /* Inner loop uses 270 flops */
1178 }
1179
1180 if(jidx<j_index_end)
1181 {
1182
1183 jnrA = jjnr[jidx];
1184 j_coord_offsetA = DIM*jnrA;
1185
1186 /* load j atom coordinates */
1187 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1188 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1189
1190 /* Calculate displacement vector */
1191 dx00 = _mm_sub_pd(ix0,jx0);
1192 dy00 = _mm_sub_pd(iy0,jy0);
1193 dz00 = _mm_sub_pd(iz0,jz0);
1194 dx01 = _mm_sub_pd(ix0,jx1);
1195 dy01 = _mm_sub_pd(iy0,jy1);
1196 dz01 = _mm_sub_pd(iz0,jz1);
1197 dx02 = _mm_sub_pd(ix0,jx2);
1198 dy02 = _mm_sub_pd(iy0,jy2);
1199 dz02 = _mm_sub_pd(iz0,jz2);
1200 dx10 = _mm_sub_pd(ix1,jx0);
1201 dy10 = _mm_sub_pd(iy1,jy0);
1202 dz10 = _mm_sub_pd(iz1,jz0);
1203 dx11 = _mm_sub_pd(ix1,jx1);
1204 dy11 = _mm_sub_pd(iy1,jy1);
1205 dz11 = _mm_sub_pd(iz1,jz1);
1206 dx12 = _mm_sub_pd(ix1,jx2);
1207 dy12 = _mm_sub_pd(iy1,jy2);
1208 dz12 = _mm_sub_pd(iz1,jz2);
1209 dx20 = _mm_sub_pd(ix2,jx0);
1210 dy20 = _mm_sub_pd(iy2,jy0);
1211 dz20 = _mm_sub_pd(iz2,jz0);
1212 dx21 = _mm_sub_pd(ix2,jx1);
1213 dy21 = _mm_sub_pd(iy2,jy1);
1214 dz21 = _mm_sub_pd(iz2,jz1);
1215 dx22 = _mm_sub_pd(ix2,jx2);
1216 dy22 = _mm_sub_pd(iy2,jy2);
1217 dz22 = _mm_sub_pd(iz2,jz2);
1218
1219 /* Calculate squared distance and things based on it */
1220 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1221 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1222 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1223 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1224 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1225 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1226 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1227 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1228 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1229
1230 rinv00 = avx128fma_invsqrt_d(rsq00);
1231 rinv01 = avx128fma_invsqrt_d(rsq01);
1232 rinv02 = avx128fma_invsqrt_d(rsq02);
1233 rinv10 = avx128fma_invsqrt_d(rsq10);
1234 rinv11 = avx128fma_invsqrt_d(rsq11);
1235 rinv12 = avx128fma_invsqrt_d(rsq12);
1236 rinv20 = avx128fma_invsqrt_d(rsq20);
1237 rinv21 = avx128fma_invsqrt_d(rsq21);
1238 rinv22 = avx128fma_invsqrt_d(rsq22);
1239
1240 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1241 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1242 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1243 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1244 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1245 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1246 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1247 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1248 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1249
1250 fjx0 = _mm_setzero_pd();
1251 fjy0 = _mm_setzero_pd();
1252 fjz0 = _mm_setzero_pd();
1253 fjx1 = _mm_setzero_pd();
1254 fjy1 = _mm_setzero_pd();
1255 fjz1 = _mm_setzero_pd();
1256 fjx2 = _mm_setzero_pd();
1257 fjy2 = _mm_setzero_pd();
1258 fjz2 = _mm_setzero_pd();
1259
1260 /**************************
1261 * CALCULATE INTERACTIONS *
1262 **************************/
1263
1264 /* REACTION-FIELD ELECTROSTATICS */
1265 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1266
1267 fscal = felec;
1268
1269 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1270
1271 /* Update vectorial force */
1272 fix0 = _mm_macc_pd(dx00,fscal,fix0);
1273 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
1274 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
1275
1276 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
1277 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
1278 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
1279
1280 /**************************
1281 * CALCULATE INTERACTIONS *
1282 **************************/
1283
1284 /* REACTION-FIELD ELECTROSTATICS */
1285 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1286
1287 fscal = felec;
1288
1289 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1290
1291 /* Update vectorial force */
1292 fix0 = _mm_macc_pd(dx01,fscal,fix0);
1293 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
1294 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
1295
1296 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
1297 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
1298 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
1299
1300 /**************************
1301 * CALCULATE INTERACTIONS *
1302 **************************/
1303
1304 /* REACTION-FIELD ELECTROSTATICS */
1305 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1306
1307 fscal = felec;
1308
1309 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1310
1311 /* Update vectorial force */
1312 fix0 = _mm_macc_pd(dx02,fscal,fix0);
1313 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
1314 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
1315
1316 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
1317 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
1318 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
1319
1320 /**************************
1321 * CALCULATE INTERACTIONS *
1322 **************************/
1323
1324 /* REACTION-FIELD ELECTROSTATICS */
1325 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1326
1327 fscal = felec;
1328
1329 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1330
1331 /* Update vectorial force */
1332 fix1 = _mm_macc_pd(dx10,fscal,fix1);
1333 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
1334 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
1335
1336 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
1337 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
1338 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
1339
1340 /**************************
1341 * CALCULATE INTERACTIONS *
1342 **************************/
1343
1344 /* REACTION-FIELD ELECTROSTATICS */
1345 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1346
1347 fscal = felec;
1348
1349 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1350
1351 /* Update vectorial force */
1352 fix1 = _mm_macc_pd(dx11,fscal,fix1);
1353 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
1354 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
1355
1356 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
1357 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
1358 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
1359
1360 /**************************
1361 * CALCULATE INTERACTIONS *
1362 **************************/
1363
1364 /* REACTION-FIELD ELECTROSTATICS */
1365 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1366
1367 fscal = felec;
1368
1369 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1370
1371 /* Update vectorial force */
1372 fix1 = _mm_macc_pd(dx12,fscal,fix1);
1373 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
1374 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
1375
1376 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
1377 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
1378 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
1379
1380 /**************************
1381 * CALCULATE INTERACTIONS *
1382 **************************/
1383
1384 /* REACTION-FIELD ELECTROSTATICS */
1385 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1386
1387 fscal = felec;
1388
1389 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1390
1391 /* Update vectorial force */
1392 fix2 = _mm_macc_pd(dx20,fscal,fix2);
1393 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
1394 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
1395
1396 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
1397 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
1398 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
1399
1400 /**************************
1401 * CALCULATE INTERACTIONS *
1402 **************************/
1403
1404 /* REACTION-FIELD ELECTROSTATICS */
1405 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1406
1407 fscal = felec;
1408
1409 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1410
1411 /* Update vectorial force */
1412 fix2 = _mm_macc_pd(dx21,fscal,fix2);
1413 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
1414 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
1415
1416 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
1417 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
1418 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
1419
1420 /**************************
1421 * CALCULATE INTERACTIONS *
1422 **************************/
1423
1424 /* REACTION-FIELD ELECTROSTATICS */
1425 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1426
1427 fscal = felec;
1428
1429 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1430
1431 /* Update vectorial force */
1432 fix2 = _mm_macc_pd(dx22,fscal,fix2);
1433 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
1434 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
1435
1436 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
1437 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
1438 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
1439
1440 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1441
1442 /* Inner loop uses 270 flops */
1443 }
1444
1445 /* End of innermost loop */
1446
1447 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1448 f+i_coord_offset,fshift+i_shift_offset);
1449
1450 /* Increment number of inner iterations */
1451 inneriter += j_index_end - j_index_start;
1452
1453 /* Outer loop uses 18 flops */
1454 }
1455
1456 /* Increment number of outer iterations */
1457 outeriter += nri;
1458
1459 /* Update outer/inner flops */
1460
1461 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
1462 }
The ultimate molecular dynamics simulation package