?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Remove all unnecessary HAVE_CONFIG_H
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_avx_128_fma_double / nb_kernel_ElecRF_VdwNone_GeomW3W3_avx_128_fma_double.c
blobb4320c04156d7a099b4b8ee1a20df4c4bdb5a918
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014, 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 "config.h"
39
40 #include <math.h>
41
42 #include "../nb_kernel.h"
43 #include "types/simple.h"
44 #include "gromacs/math/vec.h"
45 #include "nrnb.h"
46
47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
48 #include "kernelutil_x86_avx_128_fma_double.h"
49
50 /*
51 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_avx_128_fma_double
52 * Electrostatics interaction: ReactionField
53 * VdW interaction: None
54 * Geometry: Water3-Water3
55 * Calculate force/pot: PotentialAndForce
56 */
57 void
58 nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_avx_128_fma_double
59 (t_nblist * gmx_restrict nlist,
60 rvec * gmx_restrict xx,
61 rvec * gmx_restrict ff,
62 t_forcerec * gmx_restrict fr,
63 t_mdatoms * gmx_restrict mdatoms,
64 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
65 t_nrnb * gmx_restrict nrnb)
66 {
67 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
68 * just 0 for non-waters.
69 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
70 * jnr indices corresponding to data put in the four positions in the SIMD register.
71 */
72 int i_shift_offset,i_coord_offset,outeriter,inneriter;
73 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
74 int jnrA,jnrB;
75 int j_coord_offsetA,j_coord_offsetB;
76 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
77 real rcutoff_scalar;
78 real *shiftvec,*fshift,*x,*f;
79 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
80 int vdwioffset0;
81 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
82 int vdwioffset1;
83 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
84 int vdwioffset2;
85 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
86 int vdwjidx0A,vdwjidx0B;
87 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
88 int vdwjidx1A,vdwjidx1B;
89 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
90 int vdwjidx2A,vdwjidx2B;
91 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
92 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
93 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
94 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
95 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
96 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
97 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
98 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
99 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
100 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
101 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
102 real *charge;
103 __m128d dummy_mask,cutoff_mask;
104 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
105 __m128d one = _mm_set1_pd(1.0);
106 __m128d two = _mm_set1_pd(2.0);
107 x = xx[0];
108 f = ff[0];
109
110 nri = nlist->nri;
111 iinr = nlist->iinr;
112 jindex = nlist->jindex;
113 jjnr = nlist->jjnr;
114 shiftidx = nlist->shift;
115 gid = nlist->gid;
116 shiftvec = fr->shift_vec[0];
117 fshift = fr->fshift[0];
118 facel = _mm_set1_pd(fr->epsfac);
119 charge = mdatoms->chargeA;
120 krf = _mm_set1_pd(fr->ic->k_rf);
121 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
122 crf = _mm_set1_pd(fr->ic->c_rf);
123
124 /* Setup water-specific parameters */
125 inr = nlist->iinr[0];
126 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
127 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
128 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
129
130 jq0 = _mm_set1_pd(charge[inr+0]);
131 jq1 = _mm_set1_pd(charge[inr+1]);
132 jq2 = _mm_set1_pd(charge[inr+2]);
133 qq00 = _mm_mul_pd(iq0,jq0);
134 qq01 = _mm_mul_pd(iq0,jq1);
135 qq02 = _mm_mul_pd(iq0,jq2);
136 qq10 = _mm_mul_pd(iq1,jq0);
137 qq11 = _mm_mul_pd(iq1,jq1);
138 qq12 = _mm_mul_pd(iq1,jq2);
139 qq20 = _mm_mul_pd(iq2,jq0);
140 qq21 = _mm_mul_pd(iq2,jq1);
141 qq22 = _mm_mul_pd(iq2,jq2);
142
143 /* Avoid stupid compiler warnings */
144 jnrA = jnrB = 0;
145 j_coord_offsetA = 0;
146 j_coord_offsetB = 0;
147
148 outeriter = 0;
149 inneriter = 0;
150
151 /* Start outer loop over neighborlists */
152 for(iidx=0; iidx<nri; iidx++)
153 {
154 /* Load shift vector for this list */
155 i_shift_offset = DIM*shiftidx[iidx];
156
157 /* Load limits for loop over neighbors */
158 j_index_start = jindex[iidx];
159 j_index_end = jindex[iidx+1];
160
161 /* Get outer coordinate index */
162 inr = iinr[iidx];
163 i_coord_offset = DIM*inr;
164
165 /* Load i particle coords and add shift vector */
166 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
167 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
168
169 fix0 = _mm_setzero_pd();
170 fiy0 = _mm_setzero_pd();
171 fiz0 = _mm_setzero_pd();
172 fix1 = _mm_setzero_pd();
173 fiy1 = _mm_setzero_pd();
174 fiz1 = _mm_setzero_pd();
175 fix2 = _mm_setzero_pd();
176 fiy2 = _mm_setzero_pd();
177 fiz2 = _mm_setzero_pd();
178
179 /* Reset potential sums */
180 velecsum = _mm_setzero_pd();
181
182 /* Start inner kernel loop */
183 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
184 {
185
186 /* Get j neighbor index, and coordinate index */
187 jnrA = jjnr[jidx];
188 jnrB = jjnr[jidx+1];
189 j_coord_offsetA = DIM*jnrA;
190 j_coord_offsetB = DIM*jnrB;
191
192 /* load j atom coordinates */
193 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
194 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
195
196 /* Calculate displacement vector */
197 dx00 = _mm_sub_pd(ix0,jx0);
198 dy00 = _mm_sub_pd(iy0,jy0);
199 dz00 = _mm_sub_pd(iz0,jz0);
200 dx01 = _mm_sub_pd(ix0,jx1);
201 dy01 = _mm_sub_pd(iy0,jy1);
202 dz01 = _mm_sub_pd(iz0,jz1);
203 dx02 = _mm_sub_pd(ix0,jx2);
204 dy02 = _mm_sub_pd(iy0,jy2);
205 dz02 = _mm_sub_pd(iz0,jz2);
206 dx10 = _mm_sub_pd(ix1,jx0);
207 dy10 = _mm_sub_pd(iy1,jy0);
208 dz10 = _mm_sub_pd(iz1,jz0);
209 dx11 = _mm_sub_pd(ix1,jx1);
210 dy11 = _mm_sub_pd(iy1,jy1);
211 dz11 = _mm_sub_pd(iz1,jz1);
212 dx12 = _mm_sub_pd(ix1,jx2);
213 dy12 = _mm_sub_pd(iy1,jy2);
214 dz12 = _mm_sub_pd(iz1,jz2);
215 dx20 = _mm_sub_pd(ix2,jx0);
216 dy20 = _mm_sub_pd(iy2,jy0);
217 dz20 = _mm_sub_pd(iz2,jz0);
218 dx21 = _mm_sub_pd(ix2,jx1);
219 dy21 = _mm_sub_pd(iy2,jy1);
220 dz21 = _mm_sub_pd(iz2,jz1);
221 dx22 = _mm_sub_pd(ix2,jx2);
222 dy22 = _mm_sub_pd(iy2,jy2);
223 dz22 = _mm_sub_pd(iz2,jz2);
224
225 /* Calculate squared distance and things based on it */
226 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
227 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
228 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
229 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
230 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
231 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
232 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
233 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
234 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
235
236 rinv00 = gmx_mm_invsqrt_pd(rsq00);
237 rinv01 = gmx_mm_invsqrt_pd(rsq01);
238 rinv02 = gmx_mm_invsqrt_pd(rsq02);
239 rinv10 = gmx_mm_invsqrt_pd(rsq10);
240 rinv11 = gmx_mm_invsqrt_pd(rsq11);
241 rinv12 = gmx_mm_invsqrt_pd(rsq12);
242 rinv20 = gmx_mm_invsqrt_pd(rsq20);
243 rinv21 = gmx_mm_invsqrt_pd(rsq21);
244 rinv22 = gmx_mm_invsqrt_pd(rsq22);
245
246 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
247 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
248 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
249 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
250 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
251 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
252 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
253 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
254 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
255
256 fjx0 = _mm_setzero_pd();
257 fjy0 = _mm_setzero_pd();
258 fjz0 = _mm_setzero_pd();
259 fjx1 = _mm_setzero_pd();
260 fjy1 = _mm_setzero_pd();
261 fjz1 = _mm_setzero_pd();
262 fjx2 = _mm_setzero_pd();
263 fjy2 = _mm_setzero_pd();
264 fjz2 = _mm_setzero_pd();
265
266 /**************************
267 * CALCULATE INTERACTIONS *
268 **************************/
269
270 /* REACTION-FIELD ELECTROSTATICS */
271 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
272 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
273
274 /* Update potential sum for this i atom from the interaction with this j atom. */
275 velecsum = _mm_add_pd(velecsum,velec);
276
277 fscal = felec;
278
279 /* Update vectorial force */
280 fix0 = _mm_macc_pd(dx00,fscal,fix0);
281 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
282 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
283
284 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
285 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
286 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
287
288 /**************************
289 * CALCULATE INTERACTIONS *
290 **************************/
291
292 /* REACTION-FIELD ELECTROSTATICS */
293 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
294 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
295
296 /* Update potential sum for this i atom from the interaction with this j atom. */
297 velecsum = _mm_add_pd(velecsum,velec);
298
299 fscal = felec;
300
301 /* Update vectorial force */
302 fix0 = _mm_macc_pd(dx01,fscal,fix0);
303 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
304 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
305
306 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
307 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
308 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
309
310 /**************************
311 * CALCULATE INTERACTIONS *
312 **************************/
313
314 /* REACTION-FIELD ELECTROSTATICS */
315 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
316 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
317
318 /* Update potential sum for this i atom from the interaction with this j atom. */
319 velecsum = _mm_add_pd(velecsum,velec);
320
321 fscal = felec;
322
323 /* Update vectorial force */
324 fix0 = _mm_macc_pd(dx02,fscal,fix0);
325 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
326 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
327
328 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
329 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
330 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
331
332 /**************************
333 * CALCULATE INTERACTIONS *
334 **************************/
335
336 /* REACTION-FIELD ELECTROSTATICS */
337 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
338 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
339
340 /* Update potential sum for this i atom from the interaction with this j atom. */
341 velecsum = _mm_add_pd(velecsum,velec);
342
343 fscal = felec;
344
345 /* Update vectorial force */
346 fix1 = _mm_macc_pd(dx10,fscal,fix1);
347 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
348 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
349
350 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
351 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
352 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
353
354 /**************************
355 * CALCULATE INTERACTIONS *
356 **************************/
357
358 /* REACTION-FIELD ELECTROSTATICS */
359 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
360 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
361
362 /* Update potential sum for this i atom from the interaction with this j atom. */
363 velecsum = _mm_add_pd(velecsum,velec);
364
365 fscal = felec;
366
367 /* Update vectorial force */
368 fix1 = _mm_macc_pd(dx11,fscal,fix1);
369 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
370 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
371
372 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
373 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
374 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
375
376 /**************************
377 * CALCULATE INTERACTIONS *
378 **************************/
379
380 /* REACTION-FIELD ELECTROSTATICS */
381 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
382 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
383
384 /* Update potential sum for this i atom from the interaction with this j atom. */
385 velecsum = _mm_add_pd(velecsum,velec);
386
387 fscal = felec;
388
389 /* Update vectorial force */
390 fix1 = _mm_macc_pd(dx12,fscal,fix1);
391 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
392 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
393
394 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
395 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
396 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
397
398 /**************************
399 * CALCULATE INTERACTIONS *
400 **************************/
401
402 /* REACTION-FIELD ELECTROSTATICS */
403 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
404 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
405
406 /* Update potential sum for this i atom from the interaction with this j atom. */
407 velecsum = _mm_add_pd(velecsum,velec);
408
409 fscal = felec;
410
411 /* Update vectorial force */
412 fix2 = _mm_macc_pd(dx20,fscal,fix2);
413 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
414 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
415
416 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
417 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
418 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
419
420 /**************************
421 * CALCULATE INTERACTIONS *
422 **************************/
423
424 /* REACTION-FIELD ELECTROSTATICS */
425 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
426 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
427
428 /* Update potential sum for this i atom from the interaction with this j atom. */
429 velecsum = _mm_add_pd(velecsum,velec);
430
431 fscal = felec;
432
433 /* Update vectorial force */
434 fix2 = _mm_macc_pd(dx21,fscal,fix2);
435 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
436 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
437
438 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
439 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
440 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
441
442 /**************************
443 * CALCULATE INTERACTIONS *
444 **************************/
445
446 /* REACTION-FIELD ELECTROSTATICS */
447 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
448 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
449
450 /* Update potential sum for this i atom from the interaction with this j atom. */
451 velecsum = _mm_add_pd(velecsum,velec);
452
453 fscal = felec;
454
455 /* Update vectorial force */
456 fix2 = _mm_macc_pd(dx22,fscal,fix2);
457 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
458 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
459
460 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
461 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
462 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
463
464 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
465
466 /* Inner loop uses 315 flops */
467 }
468
469 if(jidx<j_index_end)
470 {
471
472 jnrA = jjnr[jidx];
473 j_coord_offsetA = DIM*jnrA;
474
475 /* load j atom coordinates */
476 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
477 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
478
479 /* Calculate displacement vector */
480 dx00 = _mm_sub_pd(ix0,jx0);
481 dy00 = _mm_sub_pd(iy0,jy0);
482 dz00 = _mm_sub_pd(iz0,jz0);
483 dx01 = _mm_sub_pd(ix0,jx1);
484 dy01 = _mm_sub_pd(iy0,jy1);
485 dz01 = _mm_sub_pd(iz0,jz1);
486 dx02 = _mm_sub_pd(ix0,jx2);
487 dy02 = _mm_sub_pd(iy0,jy2);
488 dz02 = _mm_sub_pd(iz0,jz2);
489 dx10 = _mm_sub_pd(ix1,jx0);
490 dy10 = _mm_sub_pd(iy1,jy0);
491 dz10 = _mm_sub_pd(iz1,jz0);
492 dx11 = _mm_sub_pd(ix1,jx1);
493 dy11 = _mm_sub_pd(iy1,jy1);
494 dz11 = _mm_sub_pd(iz1,jz1);
495 dx12 = _mm_sub_pd(ix1,jx2);
496 dy12 = _mm_sub_pd(iy1,jy2);
497 dz12 = _mm_sub_pd(iz1,jz2);
498 dx20 = _mm_sub_pd(ix2,jx0);
499 dy20 = _mm_sub_pd(iy2,jy0);
500 dz20 = _mm_sub_pd(iz2,jz0);
501 dx21 = _mm_sub_pd(ix2,jx1);
502 dy21 = _mm_sub_pd(iy2,jy1);
503 dz21 = _mm_sub_pd(iz2,jz1);
504 dx22 = _mm_sub_pd(ix2,jx2);
505 dy22 = _mm_sub_pd(iy2,jy2);
506 dz22 = _mm_sub_pd(iz2,jz2);
507
508 /* Calculate squared distance and things based on it */
509 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
510 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
511 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
512 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
513 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
514 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
515 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
516 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
517 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
518
519 rinv00 = gmx_mm_invsqrt_pd(rsq00);
520 rinv01 = gmx_mm_invsqrt_pd(rsq01);
521 rinv02 = gmx_mm_invsqrt_pd(rsq02);
522 rinv10 = gmx_mm_invsqrt_pd(rsq10);
523 rinv11 = gmx_mm_invsqrt_pd(rsq11);
524 rinv12 = gmx_mm_invsqrt_pd(rsq12);
525 rinv20 = gmx_mm_invsqrt_pd(rsq20);
526 rinv21 = gmx_mm_invsqrt_pd(rsq21);
527 rinv22 = gmx_mm_invsqrt_pd(rsq22);
528
529 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
530 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
531 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
532 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
533 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
534 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
535 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
536 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
537 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
538
539 fjx0 = _mm_setzero_pd();
540 fjy0 = _mm_setzero_pd();
541 fjz0 = _mm_setzero_pd();
542 fjx1 = _mm_setzero_pd();
543 fjy1 = _mm_setzero_pd();
544 fjz1 = _mm_setzero_pd();
545 fjx2 = _mm_setzero_pd();
546 fjy2 = _mm_setzero_pd();
547 fjz2 = _mm_setzero_pd();
548
549 /**************************
550 * CALCULATE INTERACTIONS *
551 **************************/
552
553 /* REACTION-FIELD ELECTROSTATICS */
554 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
555 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
556
557 /* Update potential sum for this i atom from the interaction with this j atom. */
558 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
559 velecsum = _mm_add_pd(velecsum,velec);
560
561 fscal = felec;
562
563 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
564
565 /* Update vectorial force */
566 fix0 = _mm_macc_pd(dx00,fscal,fix0);
567 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
568 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
569
570 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
571 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
572 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
573
574 /**************************
575 * CALCULATE INTERACTIONS *
576 **************************/
577
578 /* REACTION-FIELD ELECTROSTATICS */
579 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
580 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
581
582 /* Update potential sum for this i atom from the interaction with this j atom. */
583 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
584 velecsum = _mm_add_pd(velecsum,velec);
585
586 fscal = felec;
587
588 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
589
590 /* Update vectorial force */
591 fix0 = _mm_macc_pd(dx01,fscal,fix0);
592 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
593 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
594
595 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
596 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
597 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
598
599 /**************************
600 * CALCULATE INTERACTIONS *
601 **************************/
602
603 /* REACTION-FIELD ELECTROSTATICS */
604 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
605 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
606
607 /* Update potential sum for this i atom from the interaction with this j atom. */
608 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
609 velecsum = _mm_add_pd(velecsum,velec);
610
611 fscal = felec;
612
613 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
614
615 /* Update vectorial force */
616 fix0 = _mm_macc_pd(dx02,fscal,fix0);
617 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
618 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
619
620 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
621 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
622 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
623
624 /**************************
625 * CALCULATE INTERACTIONS *
626 **************************/
627
628 /* REACTION-FIELD ELECTROSTATICS */
629 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
630 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
631
632 /* Update potential sum for this i atom from the interaction with this j atom. */
633 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
634 velecsum = _mm_add_pd(velecsum,velec);
635
636 fscal = felec;
637
638 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
639
640 /* Update vectorial force */
641 fix1 = _mm_macc_pd(dx10,fscal,fix1);
642 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
643 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
644
645 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
646 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
647 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
648
649 /**************************
650 * CALCULATE INTERACTIONS *
651 **************************/
652
653 /* REACTION-FIELD ELECTROSTATICS */
654 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
655 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
656
657 /* Update potential sum for this i atom from the interaction with this j atom. */
658 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
659 velecsum = _mm_add_pd(velecsum,velec);
660
661 fscal = felec;
662
663 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
664
665 /* Update vectorial force */
666 fix1 = _mm_macc_pd(dx11,fscal,fix1);
667 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
668 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
669
670 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
671 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
672 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
673
674 /**************************
675 * CALCULATE INTERACTIONS *
676 **************************/
677
678 /* REACTION-FIELD ELECTROSTATICS */
679 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
680 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
681
682 /* Update potential sum for this i atom from the interaction with this j atom. */
683 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
684 velecsum = _mm_add_pd(velecsum,velec);
685
686 fscal = felec;
687
688 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
689
690 /* Update vectorial force */
691 fix1 = _mm_macc_pd(dx12,fscal,fix1);
692 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
693 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
694
695 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
696 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
697 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
698
699 /**************************
700 * CALCULATE INTERACTIONS *
701 **************************/
702
703 /* REACTION-FIELD ELECTROSTATICS */
704 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
705 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
706
707 /* Update potential sum for this i atom from the interaction with this j atom. */
708 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
709 velecsum = _mm_add_pd(velecsum,velec);
710
711 fscal = felec;
712
713 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
714
715 /* Update vectorial force */
716 fix2 = _mm_macc_pd(dx20,fscal,fix2);
717 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
718 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
719
720 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
721 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
722 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
723
724 /**************************
725 * CALCULATE INTERACTIONS *
726 **************************/
727
728 /* REACTION-FIELD ELECTROSTATICS */
729 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
730 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
731
732 /* Update potential sum for this i atom from the interaction with this j atom. */
733 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
734 velecsum = _mm_add_pd(velecsum,velec);
735
736 fscal = felec;
737
738 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
739
740 /* Update vectorial force */
741 fix2 = _mm_macc_pd(dx21,fscal,fix2);
742 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
743 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
744
745 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
746 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
747 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
748
749 /**************************
750 * CALCULATE INTERACTIONS *
751 **************************/
752
753 /* REACTION-FIELD ELECTROSTATICS */
754 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
755 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
756
757 /* Update potential sum for this i atom from the interaction with this j atom. */
758 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
759 velecsum = _mm_add_pd(velecsum,velec);
760
761 fscal = felec;
762
763 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
764
765 /* Update vectorial force */
766 fix2 = _mm_macc_pd(dx22,fscal,fix2);
767 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
768 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
769
770 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
771 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
772 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
773
774 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
775
776 /* Inner loop uses 315 flops */
777 }
778
779 /* End of innermost loop */
780
781 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
782 f+i_coord_offset,fshift+i_shift_offset);
783
784 ggid = gid[iidx];
785 /* Update potential energies */
786 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
787
788 /* Increment number of inner iterations */
789 inneriter += j_index_end - j_index_start;
790
791 /* Outer loop uses 19 flops */
792 }
793
794 /* Increment number of outer iterations */
795 outeriter += nri;
796
797 /* Update outer/inner flops */
798
799 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*315);
800 }
801 /*
802 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomW3W3_F_avx_128_fma_double
803 * Electrostatics interaction: ReactionField
804 * VdW interaction: None
805 * Geometry: Water3-Water3
806 * Calculate force/pot: Force
807 */
808 void
809 nb_kernel_ElecRF_VdwNone_GeomW3W3_F_avx_128_fma_double
810 (t_nblist * gmx_restrict nlist,
811 rvec * gmx_restrict xx,
812 rvec * gmx_restrict ff,
813 t_forcerec * gmx_restrict fr,
814 t_mdatoms * gmx_restrict mdatoms,
815 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
816 t_nrnb * gmx_restrict nrnb)
817 {
818 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
819 * just 0 for non-waters.
820 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
821 * jnr indices corresponding to data put in the four positions in the SIMD register.
822 */
823 int i_shift_offset,i_coord_offset,outeriter,inneriter;
824 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
825 int jnrA,jnrB;
826 int j_coord_offsetA,j_coord_offsetB;
827 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
828 real rcutoff_scalar;
829 real *shiftvec,*fshift,*x,*f;
830 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
831 int vdwioffset0;
832 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
833 int vdwioffset1;
834 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
835 int vdwioffset2;
836 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
837 int vdwjidx0A,vdwjidx0B;
838 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
839 int vdwjidx1A,vdwjidx1B;
840 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
841 int vdwjidx2A,vdwjidx2B;
842 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
843 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
844 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
845 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
846 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
847 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
848 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
849 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
850 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
851 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
852 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
853 real *charge;
854 __m128d dummy_mask,cutoff_mask;
855 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
856 __m128d one = _mm_set1_pd(1.0);
857 __m128d two = _mm_set1_pd(2.0);
858 x = xx[0];
859 f = ff[0];
860
861 nri = nlist->nri;
862 iinr = nlist->iinr;
863 jindex = nlist->jindex;
864 jjnr = nlist->jjnr;
865 shiftidx = nlist->shift;
866 gid = nlist->gid;
867 shiftvec = fr->shift_vec[0];
868 fshift = fr->fshift[0];
869 facel = _mm_set1_pd(fr->epsfac);
870 charge = mdatoms->chargeA;
871 krf = _mm_set1_pd(fr->ic->k_rf);
872 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
873 crf = _mm_set1_pd(fr->ic->c_rf);
874
875 /* Setup water-specific parameters */
876 inr = nlist->iinr[0];
877 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
878 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
879 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
880
881 jq0 = _mm_set1_pd(charge[inr+0]);
882 jq1 = _mm_set1_pd(charge[inr+1]);
883 jq2 = _mm_set1_pd(charge[inr+2]);
884 qq00 = _mm_mul_pd(iq0,jq0);
885 qq01 = _mm_mul_pd(iq0,jq1);
886 qq02 = _mm_mul_pd(iq0,jq2);
887 qq10 = _mm_mul_pd(iq1,jq0);
888 qq11 = _mm_mul_pd(iq1,jq1);
889 qq12 = _mm_mul_pd(iq1,jq2);
890 qq20 = _mm_mul_pd(iq2,jq0);
891 qq21 = _mm_mul_pd(iq2,jq1);
892 qq22 = _mm_mul_pd(iq2,jq2);
893
894 /* Avoid stupid compiler warnings */
895 jnrA = jnrB = 0;
896 j_coord_offsetA = 0;
897 j_coord_offsetB = 0;
898
899 outeriter = 0;
900 inneriter = 0;
901
902 /* Start outer loop over neighborlists */
903 for(iidx=0; iidx<nri; iidx++)
904 {
905 /* Load shift vector for this list */
906 i_shift_offset = DIM*shiftidx[iidx];
907
908 /* Load limits for loop over neighbors */
909 j_index_start = jindex[iidx];
910 j_index_end = jindex[iidx+1];
911
912 /* Get outer coordinate index */
913 inr = iinr[iidx];
914 i_coord_offset = DIM*inr;
915
916 /* Load i particle coords and add shift vector */
917 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
918 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
919
920 fix0 = _mm_setzero_pd();
921 fiy0 = _mm_setzero_pd();
922 fiz0 = _mm_setzero_pd();
923 fix1 = _mm_setzero_pd();
924 fiy1 = _mm_setzero_pd();
925 fiz1 = _mm_setzero_pd();
926 fix2 = _mm_setzero_pd();
927 fiy2 = _mm_setzero_pd();
928 fiz2 = _mm_setzero_pd();
929
930 /* Start inner kernel loop */
931 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
932 {
933
934 /* Get j neighbor index, and coordinate index */
935 jnrA = jjnr[jidx];
936 jnrB = jjnr[jidx+1];
937 j_coord_offsetA = DIM*jnrA;
938 j_coord_offsetB = DIM*jnrB;
939
940 /* load j atom coordinates */
941 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
942 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
943
944 /* Calculate displacement vector */
945 dx00 = _mm_sub_pd(ix0,jx0);
946 dy00 = _mm_sub_pd(iy0,jy0);
947 dz00 = _mm_sub_pd(iz0,jz0);
948 dx01 = _mm_sub_pd(ix0,jx1);
949 dy01 = _mm_sub_pd(iy0,jy1);
950 dz01 = _mm_sub_pd(iz0,jz1);
951 dx02 = _mm_sub_pd(ix0,jx2);
952 dy02 = _mm_sub_pd(iy0,jy2);
953 dz02 = _mm_sub_pd(iz0,jz2);
954 dx10 = _mm_sub_pd(ix1,jx0);
955 dy10 = _mm_sub_pd(iy1,jy0);
956 dz10 = _mm_sub_pd(iz1,jz0);
957 dx11 = _mm_sub_pd(ix1,jx1);
958 dy11 = _mm_sub_pd(iy1,jy1);
959 dz11 = _mm_sub_pd(iz1,jz1);
960 dx12 = _mm_sub_pd(ix1,jx2);
961 dy12 = _mm_sub_pd(iy1,jy2);
962 dz12 = _mm_sub_pd(iz1,jz2);
963 dx20 = _mm_sub_pd(ix2,jx0);
964 dy20 = _mm_sub_pd(iy2,jy0);
965 dz20 = _mm_sub_pd(iz2,jz0);
966 dx21 = _mm_sub_pd(ix2,jx1);
967 dy21 = _mm_sub_pd(iy2,jy1);
968 dz21 = _mm_sub_pd(iz2,jz1);
969 dx22 = _mm_sub_pd(ix2,jx2);
970 dy22 = _mm_sub_pd(iy2,jy2);
971 dz22 = _mm_sub_pd(iz2,jz2);
972
973 /* Calculate squared distance and things based on it */
974 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
975 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
976 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
977 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
978 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
979 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
980 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
981 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
982 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
983
984 rinv00 = gmx_mm_invsqrt_pd(rsq00);
985 rinv01 = gmx_mm_invsqrt_pd(rsq01);
986 rinv02 = gmx_mm_invsqrt_pd(rsq02);
987 rinv10 = gmx_mm_invsqrt_pd(rsq10);
988 rinv11 = gmx_mm_invsqrt_pd(rsq11);
989 rinv12 = gmx_mm_invsqrt_pd(rsq12);
990 rinv20 = gmx_mm_invsqrt_pd(rsq20);
991 rinv21 = gmx_mm_invsqrt_pd(rsq21);
992 rinv22 = gmx_mm_invsqrt_pd(rsq22);
993
994 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
995 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
996 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
997 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
998 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
999 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1000 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1001 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1002 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1003
1004 fjx0 = _mm_setzero_pd();
1005 fjy0 = _mm_setzero_pd();
1006 fjz0 = _mm_setzero_pd();
1007 fjx1 = _mm_setzero_pd();
1008 fjy1 = _mm_setzero_pd();
1009 fjz1 = _mm_setzero_pd();
1010 fjx2 = _mm_setzero_pd();
1011 fjy2 = _mm_setzero_pd();
1012 fjz2 = _mm_setzero_pd();
1013
1014 /**************************
1015 * CALCULATE INTERACTIONS *
1016 **************************/
1017
1018 /* REACTION-FIELD ELECTROSTATICS */
1019 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1020
1021 fscal = felec;
1022
1023 /* Update vectorial force */
1024 fix0 = _mm_macc_pd(dx00,fscal,fix0);
1025 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
1026 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
1027
1028 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
1029 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
1030 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
1031
1032 /**************************
1033 * CALCULATE INTERACTIONS *
1034 **************************/
1035
1036 /* REACTION-FIELD ELECTROSTATICS */
1037 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1038
1039 fscal = felec;
1040
1041 /* Update vectorial force */
1042 fix0 = _mm_macc_pd(dx01,fscal,fix0);
1043 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
1044 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
1045
1046 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
1047 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
1048 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
1049
1050 /**************************
1051 * CALCULATE INTERACTIONS *
1052 **************************/
1053
1054 /* REACTION-FIELD ELECTROSTATICS */
1055 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1056
1057 fscal = felec;
1058
1059 /* Update vectorial force */
1060 fix0 = _mm_macc_pd(dx02,fscal,fix0);
1061 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
1062 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
1063
1064 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
1065 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
1066 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
1067
1068 /**************************
1069 * CALCULATE INTERACTIONS *
1070 **************************/
1071
1072 /* REACTION-FIELD ELECTROSTATICS */
1073 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1074
1075 fscal = felec;
1076
1077 /* Update vectorial force */
1078 fix1 = _mm_macc_pd(dx10,fscal,fix1);
1079 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
1080 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
1081
1082 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
1083 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
1084 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
1085
1086 /**************************
1087 * CALCULATE INTERACTIONS *
1088 **************************/
1089
1090 /* REACTION-FIELD ELECTROSTATICS */
1091 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1092
1093 fscal = felec;
1094
1095 /* Update vectorial force */
1096 fix1 = _mm_macc_pd(dx11,fscal,fix1);
1097 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
1098 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
1099
1100 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
1101 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
1102 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
1103
1104 /**************************
1105 * CALCULATE INTERACTIONS *
1106 **************************/
1107
1108 /* REACTION-FIELD ELECTROSTATICS */
1109 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1110
1111 fscal = felec;
1112
1113 /* Update vectorial force */
1114 fix1 = _mm_macc_pd(dx12,fscal,fix1);
1115 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
1116 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
1117
1118 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
1119 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
1120 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
1121
1122 /**************************
1123 * CALCULATE INTERACTIONS *
1124 **************************/
1125
1126 /* REACTION-FIELD ELECTROSTATICS */
1127 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1128
1129 fscal = felec;
1130
1131 /* Update vectorial force */
1132 fix2 = _mm_macc_pd(dx20,fscal,fix2);
1133 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
1134 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
1135
1136 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
1137 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
1138 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
1139
1140 /**************************
1141 * CALCULATE INTERACTIONS *
1142 **************************/
1143
1144 /* REACTION-FIELD ELECTROSTATICS */
1145 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1146
1147 fscal = felec;
1148
1149 /* Update vectorial force */
1150 fix2 = _mm_macc_pd(dx21,fscal,fix2);
1151 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
1152 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
1153
1154 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
1155 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
1156 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
1157
1158 /**************************
1159 * CALCULATE INTERACTIONS *
1160 **************************/
1161
1162 /* REACTION-FIELD ELECTROSTATICS */
1163 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1164
1165 fscal = felec;
1166
1167 /* Update vectorial force */
1168 fix2 = _mm_macc_pd(dx22,fscal,fix2);
1169 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
1170 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
1171
1172 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
1173 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
1174 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
1175
1176 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1177
1178 /* Inner loop uses 270 flops */
1179 }
1180
1181 if(jidx<j_index_end)
1182 {
1183
1184 jnrA = jjnr[jidx];
1185 j_coord_offsetA = DIM*jnrA;
1186
1187 /* load j atom coordinates */
1188 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1189 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1190
1191 /* Calculate displacement vector */
1192 dx00 = _mm_sub_pd(ix0,jx0);
1193 dy00 = _mm_sub_pd(iy0,jy0);
1194 dz00 = _mm_sub_pd(iz0,jz0);
1195 dx01 = _mm_sub_pd(ix0,jx1);
1196 dy01 = _mm_sub_pd(iy0,jy1);
1197 dz01 = _mm_sub_pd(iz0,jz1);
1198 dx02 = _mm_sub_pd(ix0,jx2);
1199 dy02 = _mm_sub_pd(iy0,jy2);
1200 dz02 = _mm_sub_pd(iz0,jz2);
1201 dx10 = _mm_sub_pd(ix1,jx0);
1202 dy10 = _mm_sub_pd(iy1,jy0);
1203 dz10 = _mm_sub_pd(iz1,jz0);
1204 dx11 = _mm_sub_pd(ix1,jx1);
1205 dy11 = _mm_sub_pd(iy1,jy1);
1206 dz11 = _mm_sub_pd(iz1,jz1);
1207 dx12 = _mm_sub_pd(ix1,jx2);
1208 dy12 = _mm_sub_pd(iy1,jy2);
1209 dz12 = _mm_sub_pd(iz1,jz2);
1210 dx20 = _mm_sub_pd(ix2,jx0);
1211 dy20 = _mm_sub_pd(iy2,jy0);
1212 dz20 = _mm_sub_pd(iz2,jz0);
1213 dx21 = _mm_sub_pd(ix2,jx1);
1214 dy21 = _mm_sub_pd(iy2,jy1);
1215 dz21 = _mm_sub_pd(iz2,jz1);
1216 dx22 = _mm_sub_pd(ix2,jx2);
1217 dy22 = _mm_sub_pd(iy2,jy2);
1218 dz22 = _mm_sub_pd(iz2,jz2);
1219
1220 /* Calculate squared distance and things based on it */
1221 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1222 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1223 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1224 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1225 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1226 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1227 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1228 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1229 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1230
1231 rinv00 = gmx_mm_invsqrt_pd(rsq00);
1232 rinv01 = gmx_mm_invsqrt_pd(rsq01);
1233 rinv02 = gmx_mm_invsqrt_pd(rsq02);
1234 rinv10 = gmx_mm_invsqrt_pd(rsq10);
1235 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1236 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1237 rinv20 = gmx_mm_invsqrt_pd(rsq20);
1238 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1239 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1240
1241 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1242 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1243 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1244 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1245 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1246 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1247 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1248 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1249 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1250
1251 fjx0 = _mm_setzero_pd();
1252 fjy0 = _mm_setzero_pd();
1253 fjz0 = _mm_setzero_pd();
1254 fjx1 = _mm_setzero_pd();
1255 fjy1 = _mm_setzero_pd();
1256 fjz1 = _mm_setzero_pd();
1257 fjx2 = _mm_setzero_pd();
1258 fjy2 = _mm_setzero_pd();
1259 fjz2 = _mm_setzero_pd();
1260
1261 /**************************
1262 * CALCULATE INTERACTIONS *
1263 **************************/
1264
1265 /* REACTION-FIELD ELECTROSTATICS */
1266 felec = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1267
1268 fscal = felec;
1269
1270 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1271
1272 /* Update vectorial force */
1273 fix0 = _mm_macc_pd(dx00,fscal,fix0);
1274 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
1275 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
1276
1277 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
1278 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
1279 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
1280
1281 /**************************
1282 * CALCULATE INTERACTIONS *
1283 **************************/
1284
1285 /* REACTION-FIELD ELECTROSTATICS */
1286 felec = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1287
1288 fscal = felec;
1289
1290 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1291
1292 /* Update vectorial force */
1293 fix0 = _mm_macc_pd(dx01,fscal,fix0);
1294 fiy0 = _mm_macc_pd(dy01,fscal,fiy0);
1295 fiz0 = _mm_macc_pd(dz01,fscal,fiz0);
1296
1297 fjx1 = _mm_macc_pd(dx01,fscal,fjx1);
1298 fjy1 = _mm_macc_pd(dy01,fscal,fjy1);
1299 fjz1 = _mm_macc_pd(dz01,fscal,fjz1);
1300
1301 /**************************
1302 * CALCULATE INTERACTIONS *
1303 **************************/
1304
1305 /* REACTION-FIELD ELECTROSTATICS */
1306 felec = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1307
1308 fscal = felec;
1309
1310 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1311
1312 /* Update vectorial force */
1313 fix0 = _mm_macc_pd(dx02,fscal,fix0);
1314 fiy0 = _mm_macc_pd(dy02,fscal,fiy0);
1315 fiz0 = _mm_macc_pd(dz02,fscal,fiz0);
1316
1317 fjx2 = _mm_macc_pd(dx02,fscal,fjx2);
1318 fjy2 = _mm_macc_pd(dy02,fscal,fjy2);
1319 fjz2 = _mm_macc_pd(dz02,fscal,fjz2);
1320
1321 /**************************
1322 * CALCULATE INTERACTIONS *
1323 **************************/
1324
1325 /* REACTION-FIELD ELECTROSTATICS */
1326 felec = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1327
1328 fscal = felec;
1329
1330 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1331
1332 /* Update vectorial force */
1333 fix1 = _mm_macc_pd(dx10,fscal,fix1);
1334 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
1335 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
1336
1337 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
1338 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
1339 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
1340
1341 /**************************
1342 * CALCULATE INTERACTIONS *
1343 **************************/
1344
1345 /* REACTION-FIELD ELECTROSTATICS */
1346 felec = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1347
1348 fscal = felec;
1349
1350 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1351
1352 /* Update vectorial force */
1353 fix1 = _mm_macc_pd(dx11,fscal,fix1);
1354 fiy1 = _mm_macc_pd(dy11,fscal,fiy1);
1355 fiz1 = _mm_macc_pd(dz11,fscal,fiz1);
1356
1357 fjx1 = _mm_macc_pd(dx11,fscal,fjx1);
1358 fjy1 = _mm_macc_pd(dy11,fscal,fjy1);
1359 fjz1 = _mm_macc_pd(dz11,fscal,fjz1);
1360
1361 /**************************
1362 * CALCULATE INTERACTIONS *
1363 **************************/
1364
1365 /* REACTION-FIELD ELECTROSTATICS */
1366 felec = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1367
1368 fscal = felec;
1369
1370 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1371
1372 /* Update vectorial force */
1373 fix1 = _mm_macc_pd(dx12,fscal,fix1);
1374 fiy1 = _mm_macc_pd(dy12,fscal,fiy1);
1375 fiz1 = _mm_macc_pd(dz12,fscal,fiz1);
1376
1377 fjx2 = _mm_macc_pd(dx12,fscal,fjx2);
1378 fjy2 = _mm_macc_pd(dy12,fscal,fjy2);
1379 fjz2 = _mm_macc_pd(dz12,fscal,fjz2);
1380
1381 /**************************
1382 * CALCULATE INTERACTIONS *
1383 **************************/
1384
1385 /* REACTION-FIELD ELECTROSTATICS */
1386 felec = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1387
1388 fscal = felec;
1389
1390 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1391
1392 /* Update vectorial force */
1393 fix2 = _mm_macc_pd(dx20,fscal,fix2);
1394 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
1395 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
1396
1397 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
1398 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
1399 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
1400
1401 /**************************
1402 * CALCULATE INTERACTIONS *
1403 **************************/
1404
1405 /* REACTION-FIELD ELECTROSTATICS */
1406 felec = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1407
1408 fscal = felec;
1409
1410 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1411
1412 /* Update vectorial force */
1413 fix2 = _mm_macc_pd(dx21,fscal,fix2);
1414 fiy2 = _mm_macc_pd(dy21,fscal,fiy2);
1415 fiz2 = _mm_macc_pd(dz21,fscal,fiz2);
1416
1417 fjx1 = _mm_macc_pd(dx21,fscal,fjx1);
1418 fjy1 = _mm_macc_pd(dy21,fscal,fjy1);
1419 fjz1 = _mm_macc_pd(dz21,fscal,fjz1);
1420
1421 /**************************
1422 * CALCULATE INTERACTIONS *
1423 **************************/
1424
1425 /* REACTION-FIELD ELECTROSTATICS */
1426 felec = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1427
1428 fscal = felec;
1429
1430 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1431
1432 /* Update vectorial force */
1433 fix2 = _mm_macc_pd(dx22,fscal,fix2);
1434 fiy2 = _mm_macc_pd(dy22,fscal,fiy2);
1435 fiz2 = _mm_macc_pd(dz22,fscal,fiz2);
1436
1437 fjx2 = _mm_macc_pd(dx22,fscal,fjx2);
1438 fjy2 = _mm_macc_pd(dy22,fscal,fjy2);
1439 fjz2 = _mm_macc_pd(dz22,fscal,fjz2);
1440
1441 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1442
1443 /* Inner loop uses 270 flops */
1444 }
1445
1446 /* End of innermost loop */
1447
1448 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1449 f+i_coord_offset,fshift+i_shift_offset);
1450
1451 /* Increment number of inner iterations */
1452 inneriter += j_index_end - j_index_start;
1453
1454 /* Outer loop uses 18 flops */
1455 }
1456
1457 /* Increment number of outer iterations */
1458 outeriter += nri;
1459
1460 /* Update outer/inner flops */
1461
1462 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
1463 }
The ultimate molecular dynamics simulation package