?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Removed simple.h from nb_kernel_sse2_XX
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse2_double / nb_kernel_ElecRF_VdwNone_GeomW3W3_sse2_double.c
blobc167eabae3b8d6854a679644a65f0a3fbb97bbf6
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014,2015, 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 sse2_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/math/vec.h"
46 #include "gromacs/legacyheaders/nrnb.h"
47
48 #include "gromacs/simd/math_x86_sse2_double.h"
49 #include "kernelutil_x86_sse2_double.h"
50
51 /*
52 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sse2_double
53 * Electrostatics interaction: ReactionField
54 * VdW interaction: None
55 * Geometry: Water3-Water3
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sse2_double
60 (t_nblist * gmx_restrict nlist,
61 rvec * gmx_restrict xx,
62 rvec * gmx_restrict ff,
63 t_forcerec * gmx_restrict fr,
64 t_mdatoms * gmx_restrict mdatoms,
65 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
66 t_nrnb * gmx_restrict nrnb)
67 {
68 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
69 * just 0 for non-waters.
70 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
71 * jnr indices corresponding to data put in the four positions in the SIMD register.
72 */
73 int i_shift_offset,i_coord_offset,outeriter,inneriter;
74 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
75 int jnrA,jnrB;
76 int j_coord_offsetA,j_coord_offsetB;
77 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
78 real rcutoff_scalar;
79 real *shiftvec,*fshift,*x,*f;
80 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
81 int vdwioffset0;
82 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
83 int vdwioffset1;
84 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
85 int vdwioffset2;
86 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
87 int vdwjidx0A,vdwjidx0B;
88 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
89 int vdwjidx1A,vdwjidx1B;
90 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
91 int vdwjidx2A,vdwjidx2B;
92 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
93 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
94 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
95 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
96 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
97 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
98 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
99 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
100 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
101 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
102 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
103 real *charge;
104 __m128d dummy_mask,cutoff_mask;
105 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
106 __m128d one = _mm_set1_pd(1.0);
107 __m128d two = _mm_set1_pd(2.0);
108 x = xx[0];
109 f = ff[0];
110
111 nri = nlist->nri;
112 iinr = nlist->iinr;
113 jindex = nlist->jindex;
114 jjnr = nlist->jjnr;
115 shiftidx = nlist->shift;
116 gid = nlist->gid;
117 shiftvec = fr->shift_vec[0];
118 fshift = fr->fshift[0];
119 facel = _mm_set1_pd(fr->epsfac);
120 charge = mdatoms->chargeA;
121 krf = _mm_set1_pd(fr->ic->k_rf);
122 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
123 crf = _mm_set1_pd(fr->ic->c_rf);
124
125 /* Setup water-specific parameters */
126 inr = nlist->iinr[0];
127 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
128 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
129 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
130
131 jq0 = _mm_set1_pd(charge[inr+0]);
132 jq1 = _mm_set1_pd(charge[inr+1]);
133 jq2 = _mm_set1_pd(charge[inr+2]);
134 qq00 = _mm_mul_pd(iq0,jq0);
135 qq01 = _mm_mul_pd(iq0,jq1);
136 qq02 = _mm_mul_pd(iq0,jq2);
137 qq10 = _mm_mul_pd(iq1,jq0);
138 qq11 = _mm_mul_pd(iq1,jq1);
139 qq12 = _mm_mul_pd(iq1,jq2);
140 qq20 = _mm_mul_pd(iq2,jq0);
141 qq21 = _mm_mul_pd(iq2,jq1);
142 qq22 = _mm_mul_pd(iq2,jq2);
143
144 /* Avoid stupid compiler warnings */
145 jnrA = jnrB = 0;
146 j_coord_offsetA = 0;
147 j_coord_offsetB = 0;
148
149 outeriter = 0;
150 inneriter = 0;
151
152 /* Start outer loop over neighborlists */
153 for(iidx=0; iidx<nri; iidx++)
154 {
155 /* Load shift vector for this list */
156 i_shift_offset = DIM*shiftidx[iidx];
157
158 /* Load limits for loop over neighbors */
159 j_index_start = jindex[iidx];
160 j_index_end = jindex[iidx+1];
161
162 /* Get outer coordinate index */
163 inr = iinr[iidx];
164 i_coord_offset = DIM*inr;
165
166 /* Load i particle coords and add shift vector */
167 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
168 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
169
170 fix0 = _mm_setzero_pd();
171 fiy0 = _mm_setzero_pd();
172 fiz0 = _mm_setzero_pd();
173 fix1 = _mm_setzero_pd();
174 fiy1 = _mm_setzero_pd();
175 fiz1 = _mm_setzero_pd();
176 fix2 = _mm_setzero_pd();
177 fiy2 = _mm_setzero_pd();
178 fiz2 = _mm_setzero_pd();
179
180 /* Reset potential sums */
181 velecsum = _mm_setzero_pd();
182
183 /* Start inner kernel loop */
184 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
185 {
186
187 /* Get j neighbor index, and coordinate index */
188 jnrA = jjnr[jidx];
189 jnrB = jjnr[jidx+1];
190 j_coord_offsetA = DIM*jnrA;
191 j_coord_offsetB = DIM*jnrB;
192
193 /* load j atom coordinates */
194 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
195 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
196
197 /* Calculate displacement vector */
198 dx00 = _mm_sub_pd(ix0,jx0);
199 dy00 = _mm_sub_pd(iy0,jy0);
200 dz00 = _mm_sub_pd(iz0,jz0);
201 dx01 = _mm_sub_pd(ix0,jx1);
202 dy01 = _mm_sub_pd(iy0,jy1);
203 dz01 = _mm_sub_pd(iz0,jz1);
204 dx02 = _mm_sub_pd(ix0,jx2);
205 dy02 = _mm_sub_pd(iy0,jy2);
206 dz02 = _mm_sub_pd(iz0,jz2);
207 dx10 = _mm_sub_pd(ix1,jx0);
208 dy10 = _mm_sub_pd(iy1,jy0);
209 dz10 = _mm_sub_pd(iz1,jz0);
210 dx11 = _mm_sub_pd(ix1,jx1);
211 dy11 = _mm_sub_pd(iy1,jy1);
212 dz11 = _mm_sub_pd(iz1,jz1);
213 dx12 = _mm_sub_pd(ix1,jx2);
214 dy12 = _mm_sub_pd(iy1,jy2);
215 dz12 = _mm_sub_pd(iz1,jz2);
216 dx20 = _mm_sub_pd(ix2,jx0);
217 dy20 = _mm_sub_pd(iy2,jy0);
218 dz20 = _mm_sub_pd(iz2,jz0);
219 dx21 = _mm_sub_pd(ix2,jx1);
220 dy21 = _mm_sub_pd(iy2,jy1);
221 dz21 = _mm_sub_pd(iz2,jz1);
222 dx22 = _mm_sub_pd(ix2,jx2);
223 dy22 = _mm_sub_pd(iy2,jy2);
224 dz22 = _mm_sub_pd(iz2,jz2);
225
226 /* Calculate squared distance and things based on it */
227 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
228 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
229 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
230 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
231 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
232 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
233 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
234 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
235 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
236
237 rinv00 = gmx_mm_invsqrt_pd(rsq00);
238 rinv01 = gmx_mm_invsqrt_pd(rsq01);
239 rinv02 = gmx_mm_invsqrt_pd(rsq02);
240 rinv10 = gmx_mm_invsqrt_pd(rsq10);
241 rinv11 = gmx_mm_invsqrt_pd(rsq11);
242 rinv12 = gmx_mm_invsqrt_pd(rsq12);
243 rinv20 = gmx_mm_invsqrt_pd(rsq20);
244 rinv21 = gmx_mm_invsqrt_pd(rsq21);
245 rinv22 = gmx_mm_invsqrt_pd(rsq22);
246
247 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
248 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
249 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
250 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
251 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
252 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
253 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
254 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
255 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
256
257 fjx0 = _mm_setzero_pd();
258 fjy0 = _mm_setzero_pd();
259 fjz0 = _mm_setzero_pd();
260 fjx1 = _mm_setzero_pd();
261 fjy1 = _mm_setzero_pd();
262 fjz1 = _mm_setzero_pd();
263 fjx2 = _mm_setzero_pd();
264 fjy2 = _mm_setzero_pd();
265 fjz2 = _mm_setzero_pd();
266
267 /**************************
268 * CALCULATE INTERACTIONS *
269 **************************/
270
271 /* REACTION-FIELD ELECTROSTATICS */
272 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
273 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
274
275 /* Update potential sum for this i atom from the interaction with this j atom. */
276 velecsum = _mm_add_pd(velecsum,velec);
277
278 fscal = felec;
279
280 /* Calculate temporary vectorial force */
281 tx = _mm_mul_pd(fscal,dx00);
282 ty = _mm_mul_pd(fscal,dy00);
283 tz = _mm_mul_pd(fscal,dz00);
284
285 /* Update vectorial force */
286 fix0 = _mm_add_pd(fix0,tx);
287 fiy0 = _mm_add_pd(fiy0,ty);
288 fiz0 = _mm_add_pd(fiz0,tz);
289
290 fjx0 = _mm_add_pd(fjx0,tx);
291 fjy0 = _mm_add_pd(fjy0,ty);
292 fjz0 = _mm_add_pd(fjz0,tz);
293
294 /**************************
295 * CALCULATE INTERACTIONS *
296 **************************/
297
298 /* REACTION-FIELD ELECTROSTATICS */
299 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
300 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
301
302 /* Update potential sum for this i atom from the interaction with this j atom. */
303 velecsum = _mm_add_pd(velecsum,velec);
304
305 fscal = felec;
306
307 /* Calculate temporary vectorial force */
308 tx = _mm_mul_pd(fscal,dx01);
309 ty = _mm_mul_pd(fscal,dy01);
310 tz = _mm_mul_pd(fscal,dz01);
311
312 /* Update vectorial force */
313 fix0 = _mm_add_pd(fix0,tx);
314 fiy0 = _mm_add_pd(fiy0,ty);
315 fiz0 = _mm_add_pd(fiz0,tz);
316
317 fjx1 = _mm_add_pd(fjx1,tx);
318 fjy1 = _mm_add_pd(fjy1,ty);
319 fjz1 = _mm_add_pd(fjz1,tz);
320
321 /**************************
322 * CALCULATE INTERACTIONS *
323 **************************/
324
325 /* REACTION-FIELD ELECTROSTATICS */
326 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
327 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
328
329 /* Update potential sum for this i atom from the interaction with this j atom. */
330 velecsum = _mm_add_pd(velecsum,velec);
331
332 fscal = felec;
333
334 /* Calculate temporary vectorial force */
335 tx = _mm_mul_pd(fscal,dx02);
336 ty = _mm_mul_pd(fscal,dy02);
337 tz = _mm_mul_pd(fscal,dz02);
338
339 /* Update vectorial force */
340 fix0 = _mm_add_pd(fix0,tx);
341 fiy0 = _mm_add_pd(fiy0,ty);
342 fiz0 = _mm_add_pd(fiz0,tz);
343
344 fjx2 = _mm_add_pd(fjx2,tx);
345 fjy2 = _mm_add_pd(fjy2,ty);
346 fjz2 = _mm_add_pd(fjz2,tz);
347
348 /**************************
349 * CALCULATE INTERACTIONS *
350 **************************/
351
352 /* REACTION-FIELD ELECTROSTATICS */
353 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
354 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
355
356 /* Update potential sum for this i atom from the interaction with this j atom. */
357 velecsum = _mm_add_pd(velecsum,velec);
358
359 fscal = felec;
360
361 /* Calculate temporary vectorial force */
362 tx = _mm_mul_pd(fscal,dx10);
363 ty = _mm_mul_pd(fscal,dy10);
364 tz = _mm_mul_pd(fscal,dz10);
365
366 /* Update vectorial force */
367 fix1 = _mm_add_pd(fix1,tx);
368 fiy1 = _mm_add_pd(fiy1,ty);
369 fiz1 = _mm_add_pd(fiz1,tz);
370
371 fjx0 = _mm_add_pd(fjx0,tx);
372 fjy0 = _mm_add_pd(fjy0,ty);
373 fjz0 = _mm_add_pd(fjz0,tz);
374
375 /**************************
376 * CALCULATE INTERACTIONS *
377 **************************/
378
379 /* REACTION-FIELD ELECTROSTATICS */
380 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
381 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),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 /* Calculate temporary vectorial force */
389 tx = _mm_mul_pd(fscal,dx11);
390 ty = _mm_mul_pd(fscal,dy11);
391 tz = _mm_mul_pd(fscal,dz11);
392
393 /* Update vectorial force */
394 fix1 = _mm_add_pd(fix1,tx);
395 fiy1 = _mm_add_pd(fiy1,ty);
396 fiz1 = _mm_add_pd(fiz1,tz);
397
398 fjx1 = _mm_add_pd(fjx1,tx);
399 fjy1 = _mm_add_pd(fjy1,ty);
400 fjz1 = _mm_add_pd(fjz1,tz);
401
402 /**************************
403 * CALCULATE INTERACTIONS *
404 **************************/
405
406 /* REACTION-FIELD ELECTROSTATICS */
407 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
408 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
409
410 /* Update potential sum for this i atom from the interaction with this j atom. */
411 velecsum = _mm_add_pd(velecsum,velec);
412
413 fscal = felec;
414
415 /* Calculate temporary vectorial force */
416 tx = _mm_mul_pd(fscal,dx12);
417 ty = _mm_mul_pd(fscal,dy12);
418 tz = _mm_mul_pd(fscal,dz12);
419
420 /* Update vectorial force */
421 fix1 = _mm_add_pd(fix1,tx);
422 fiy1 = _mm_add_pd(fiy1,ty);
423 fiz1 = _mm_add_pd(fiz1,tz);
424
425 fjx2 = _mm_add_pd(fjx2,tx);
426 fjy2 = _mm_add_pd(fjy2,ty);
427 fjz2 = _mm_add_pd(fjz2,tz);
428
429 /**************************
430 * CALCULATE INTERACTIONS *
431 **************************/
432
433 /* REACTION-FIELD ELECTROSTATICS */
434 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
435 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
436
437 /* Update potential sum for this i atom from the interaction with this j atom. */
438 velecsum = _mm_add_pd(velecsum,velec);
439
440 fscal = felec;
441
442 /* Calculate temporary vectorial force */
443 tx = _mm_mul_pd(fscal,dx20);
444 ty = _mm_mul_pd(fscal,dy20);
445 tz = _mm_mul_pd(fscal,dz20);
446
447 /* Update vectorial force */
448 fix2 = _mm_add_pd(fix2,tx);
449 fiy2 = _mm_add_pd(fiy2,ty);
450 fiz2 = _mm_add_pd(fiz2,tz);
451
452 fjx0 = _mm_add_pd(fjx0,tx);
453 fjy0 = _mm_add_pd(fjy0,ty);
454 fjz0 = _mm_add_pd(fjz0,tz);
455
456 /**************************
457 * CALCULATE INTERACTIONS *
458 **************************/
459
460 /* REACTION-FIELD ELECTROSTATICS */
461 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
462 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
463
464 /* Update potential sum for this i atom from the interaction with this j atom. */
465 velecsum = _mm_add_pd(velecsum,velec);
466
467 fscal = felec;
468
469 /* Calculate temporary vectorial force */
470 tx = _mm_mul_pd(fscal,dx21);
471 ty = _mm_mul_pd(fscal,dy21);
472 tz = _mm_mul_pd(fscal,dz21);
473
474 /* Update vectorial force */
475 fix2 = _mm_add_pd(fix2,tx);
476 fiy2 = _mm_add_pd(fiy2,ty);
477 fiz2 = _mm_add_pd(fiz2,tz);
478
479 fjx1 = _mm_add_pd(fjx1,tx);
480 fjy1 = _mm_add_pd(fjy1,ty);
481 fjz1 = _mm_add_pd(fjz1,tz);
482
483 /**************************
484 * CALCULATE INTERACTIONS *
485 **************************/
486
487 /* REACTION-FIELD ELECTROSTATICS */
488 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
489 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
490
491 /* Update potential sum for this i atom from the interaction with this j atom. */
492 velecsum = _mm_add_pd(velecsum,velec);
493
494 fscal = felec;
495
496 /* Calculate temporary vectorial force */
497 tx = _mm_mul_pd(fscal,dx22);
498 ty = _mm_mul_pd(fscal,dy22);
499 tz = _mm_mul_pd(fscal,dz22);
500
501 /* Update vectorial force */
502 fix2 = _mm_add_pd(fix2,tx);
503 fiy2 = _mm_add_pd(fiy2,ty);
504 fiz2 = _mm_add_pd(fiz2,tz);
505
506 fjx2 = _mm_add_pd(fjx2,tx);
507 fjy2 = _mm_add_pd(fjy2,ty);
508 fjz2 = _mm_add_pd(fjz2,tz);
509
510 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
511
512 /* Inner loop uses 288 flops */
513 }
514
515 if(jidx<j_index_end)
516 {
517
518 jnrA = jjnr[jidx];
519 j_coord_offsetA = DIM*jnrA;
520
521 /* load j atom coordinates */
522 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
523 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
524
525 /* Calculate displacement vector */
526 dx00 = _mm_sub_pd(ix0,jx0);
527 dy00 = _mm_sub_pd(iy0,jy0);
528 dz00 = _mm_sub_pd(iz0,jz0);
529 dx01 = _mm_sub_pd(ix0,jx1);
530 dy01 = _mm_sub_pd(iy0,jy1);
531 dz01 = _mm_sub_pd(iz0,jz1);
532 dx02 = _mm_sub_pd(ix0,jx2);
533 dy02 = _mm_sub_pd(iy0,jy2);
534 dz02 = _mm_sub_pd(iz0,jz2);
535 dx10 = _mm_sub_pd(ix1,jx0);
536 dy10 = _mm_sub_pd(iy1,jy0);
537 dz10 = _mm_sub_pd(iz1,jz0);
538 dx11 = _mm_sub_pd(ix1,jx1);
539 dy11 = _mm_sub_pd(iy1,jy1);
540 dz11 = _mm_sub_pd(iz1,jz1);
541 dx12 = _mm_sub_pd(ix1,jx2);
542 dy12 = _mm_sub_pd(iy1,jy2);
543 dz12 = _mm_sub_pd(iz1,jz2);
544 dx20 = _mm_sub_pd(ix2,jx0);
545 dy20 = _mm_sub_pd(iy2,jy0);
546 dz20 = _mm_sub_pd(iz2,jz0);
547 dx21 = _mm_sub_pd(ix2,jx1);
548 dy21 = _mm_sub_pd(iy2,jy1);
549 dz21 = _mm_sub_pd(iz2,jz1);
550 dx22 = _mm_sub_pd(ix2,jx2);
551 dy22 = _mm_sub_pd(iy2,jy2);
552 dz22 = _mm_sub_pd(iz2,jz2);
553
554 /* Calculate squared distance and things based on it */
555 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
556 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
557 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
558 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
559 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
560 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
561 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
562 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
563 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
564
565 rinv00 = gmx_mm_invsqrt_pd(rsq00);
566 rinv01 = gmx_mm_invsqrt_pd(rsq01);
567 rinv02 = gmx_mm_invsqrt_pd(rsq02);
568 rinv10 = gmx_mm_invsqrt_pd(rsq10);
569 rinv11 = gmx_mm_invsqrt_pd(rsq11);
570 rinv12 = gmx_mm_invsqrt_pd(rsq12);
571 rinv20 = gmx_mm_invsqrt_pd(rsq20);
572 rinv21 = gmx_mm_invsqrt_pd(rsq21);
573 rinv22 = gmx_mm_invsqrt_pd(rsq22);
574
575 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
576 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
577 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
578 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
579 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
580 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
581 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
582 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
583 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
584
585 fjx0 = _mm_setzero_pd();
586 fjy0 = _mm_setzero_pd();
587 fjz0 = _mm_setzero_pd();
588 fjx1 = _mm_setzero_pd();
589 fjy1 = _mm_setzero_pd();
590 fjz1 = _mm_setzero_pd();
591 fjx2 = _mm_setzero_pd();
592 fjy2 = _mm_setzero_pd();
593 fjz2 = _mm_setzero_pd();
594
595 /**************************
596 * CALCULATE INTERACTIONS *
597 **************************/
598
599 /* REACTION-FIELD ELECTROSTATICS */
600 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
601 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
602
603 /* Update potential sum for this i atom from the interaction with this j atom. */
604 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
605 velecsum = _mm_add_pd(velecsum,velec);
606
607 fscal = felec;
608
609 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
610
611 /* Calculate temporary vectorial force */
612 tx = _mm_mul_pd(fscal,dx00);
613 ty = _mm_mul_pd(fscal,dy00);
614 tz = _mm_mul_pd(fscal,dz00);
615
616 /* Update vectorial force */
617 fix0 = _mm_add_pd(fix0,tx);
618 fiy0 = _mm_add_pd(fiy0,ty);
619 fiz0 = _mm_add_pd(fiz0,tz);
620
621 fjx0 = _mm_add_pd(fjx0,tx);
622 fjy0 = _mm_add_pd(fjy0,ty);
623 fjz0 = _mm_add_pd(fjz0,tz);
624
625 /**************************
626 * CALCULATE INTERACTIONS *
627 **************************/
628
629 /* REACTION-FIELD ELECTROSTATICS */
630 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
631 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
632
633 /* Update potential sum for this i atom from the interaction with this j atom. */
634 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
635 velecsum = _mm_add_pd(velecsum,velec);
636
637 fscal = felec;
638
639 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
640
641 /* Calculate temporary vectorial force */
642 tx = _mm_mul_pd(fscal,dx01);
643 ty = _mm_mul_pd(fscal,dy01);
644 tz = _mm_mul_pd(fscal,dz01);
645
646 /* Update vectorial force */
647 fix0 = _mm_add_pd(fix0,tx);
648 fiy0 = _mm_add_pd(fiy0,ty);
649 fiz0 = _mm_add_pd(fiz0,tz);
650
651 fjx1 = _mm_add_pd(fjx1,tx);
652 fjy1 = _mm_add_pd(fjy1,ty);
653 fjz1 = _mm_add_pd(fjz1,tz);
654
655 /**************************
656 * CALCULATE INTERACTIONS *
657 **************************/
658
659 /* REACTION-FIELD ELECTROSTATICS */
660 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
661 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
662
663 /* Update potential sum for this i atom from the interaction with this j atom. */
664 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
665 velecsum = _mm_add_pd(velecsum,velec);
666
667 fscal = felec;
668
669 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
670
671 /* Calculate temporary vectorial force */
672 tx = _mm_mul_pd(fscal,dx02);
673 ty = _mm_mul_pd(fscal,dy02);
674 tz = _mm_mul_pd(fscal,dz02);
675
676 /* Update vectorial force */
677 fix0 = _mm_add_pd(fix0,tx);
678 fiy0 = _mm_add_pd(fiy0,ty);
679 fiz0 = _mm_add_pd(fiz0,tz);
680
681 fjx2 = _mm_add_pd(fjx2,tx);
682 fjy2 = _mm_add_pd(fjy2,ty);
683 fjz2 = _mm_add_pd(fjz2,tz);
684
685 /**************************
686 * CALCULATE INTERACTIONS *
687 **************************/
688
689 /* REACTION-FIELD ELECTROSTATICS */
690 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
691 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
692
693 /* Update potential sum for this i atom from the interaction with this j atom. */
694 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
695 velecsum = _mm_add_pd(velecsum,velec);
696
697 fscal = felec;
698
699 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
700
701 /* Calculate temporary vectorial force */
702 tx = _mm_mul_pd(fscal,dx10);
703 ty = _mm_mul_pd(fscal,dy10);
704 tz = _mm_mul_pd(fscal,dz10);
705
706 /* Update vectorial force */
707 fix1 = _mm_add_pd(fix1,tx);
708 fiy1 = _mm_add_pd(fiy1,ty);
709 fiz1 = _mm_add_pd(fiz1,tz);
710
711 fjx0 = _mm_add_pd(fjx0,tx);
712 fjy0 = _mm_add_pd(fjy0,ty);
713 fjz0 = _mm_add_pd(fjz0,tz);
714
715 /**************************
716 * CALCULATE INTERACTIONS *
717 **************************/
718
719 /* REACTION-FIELD ELECTROSTATICS */
720 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
721 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
722
723 /* Update potential sum for this i atom from the interaction with this j atom. */
724 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
725 velecsum = _mm_add_pd(velecsum,velec);
726
727 fscal = felec;
728
729 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
730
731 /* Calculate temporary vectorial force */
732 tx = _mm_mul_pd(fscal,dx11);
733 ty = _mm_mul_pd(fscal,dy11);
734 tz = _mm_mul_pd(fscal,dz11);
735
736 /* Update vectorial force */
737 fix1 = _mm_add_pd(fix1,tx);
738 fiy1 = _mm_add_pd(fiy1,ty);
739 fiz1 = _mm_add_pd(fiz1,tz);
740
741 fjx1 = _mm_add_pd(fjx1,tx);
742 fjy1 = _mm_add_pd(fjy1,ty);
743 fjz1 = _mm_add_pd(fjz1,tz);
744
745 /**************************
746 * CALCULATE INTERACTIONS *
747 **************************/
748
749 /* REACTION-FIELD ELECTROSTATICS */
750 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
751 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
752
753 /* Update potential sum for this i atom from the interaction with this j atom. */
754 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
755 velecsum = _mm_add_pd(velecsum,velec);
756
757 fscal = felec;
758
759 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
760
761 /* Calculate temporary vectorial force */
762 tx = _mm_mul_pd(fscal,dx12);
763 ty = _mm_mul_pd(fscal,dy12);
764 tz = _mm_mul_pd(fscal,dz12);
765
766 /* Update vectorial force */
767 fix1 = _mm_add_pd(fix1,tx);
768 fiy1 = _mm_add_pd(fiy1,ty);
769 fiz1 = _mm_add_pd(fiz1,tz);
770
771 fjx2 = _mm_add_pd(fjx2,tx);
772 fjy2 = _mm_add_pd(fjy2,ty);
773 fjz2 = _mm_add_pd(fjz2,tz);
774
775 /**************************
776 * CALCULATE INTERACTIONS *
777 **************************/
778
779 /* REACTION-FIELD ELECTROSTATICS */
780 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
781 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
782
783 /* Update potential sum for this i atom from the interaction with this j atom. */
784 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
785 velecsum = _mm_add_pd(velecsum,velec);
786
787 fscal = felec;
788
789 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
790
791 /* Calculate temporary vectorial force */
792 tx = _mm_mul_pd(fscal,dx20);
793 ty = _mm_mul_pd(fscal,dy20);
794 tz = _mm_mul_pd(fscal,dz20);
795
796 /* Update vectorial force */
797 fix2 = _mm_add_pd(fix2,tx);
798 fiy2 = _mm_add_pd(fiy2,ty);
799 fiz2 = _mm_add_pd(fiz2,tz);
800
801 fjx0 = _mm_add_pd(fjx0,tx);
802 fjy0 = _mm_add_pd(fjy0,ty);
803 fjz0 = _mm_add_pd(fjz0,tz);
804
805 /**************************
806 * CALCULATE INTERACTIONS *
807 **************************/
808
809 /* REACTION-FIELD ELECTROSTATICS */
810 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
811 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
812
813 /* Update potential sum for this i atom from the interaction with this j atom. */
814 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
815 velecsum = _mm_add_pd(velecsum,velec);
816
817 fscal = felec;
818
819 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
820
821 /* Calculate temporary vectorial force */
822 tx = _mm_mul_pd(fscal,dx21);
823 ty = _mm_mul_pd(fscal,dy21);
824 tz = _mm_mul_pd(fscal,dz21);
825
826 /* Update vectorial force */
827 fix2 = _mm_add_pd(fix2,tx);
828 fiy2 = _mm_add_pd(fiy2,ty);
829 fiz2 = _mm_add_pd(fiz2,tz);
830
831 fjx1 = _mm_add_pd(fjx1,tx);
832 fjy1 = _mm_add_pd(fjy1,ty);
833 fjz1 = _mm_add_pd(fjz1,tz);
834
835 /**************************
836 * CALCULATE INTERACTIONS *
837 **************************/
838
839 /* REACTION-FIELD ELECTROSTATICS */
840 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
841 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
842
843 /* Update potential sum for this i atom from the interaction with this j atom. */
844 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
845 velecsum = _mm_add_pd(velecsum,velec);
846
847 fscal = felec;
848
849 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
850
851 /* Calculate temporary vectorial force */
852 tx = _mm_mul_pd(fscal,dx22);
853 ty = _mm_mul_pd(fscal,dy22);
854 tz = _mm_mul_pd(fscal,dz22);
855
856 /* Update vectorial force */
857 fix2 = _mm_add_pd(fix2,tx);
858 fiy2 = _mm_add_pd(fiy2,ty);
859 fiz2 = _mm_add_pd(fiz2,tz);
860
861 fjx2 = _mm_add_pd(fjx2,tx);
862 fjy2 = _mm_add_pd(fjy2,ty);
863 fjz2 = _mm_add_pd(fjz2,tz);
864
865 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
866
867 /* Inner loop uses 288 flops */
868 }
869
870 /* End of innermost loop */
871
872 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
873 f+i_coord_offset,fshift+i_shift_offset);
874
875 ggid = gid[iidx];
876 /* Update potential energies */
877 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
878
879 /* Increment number of inner iterations */
880 inneriter += j_index_end - j_index_start;
881
882 /* Outer loop uses 19 flops */
883 }
884
885 /* Increment number of outer iterations */
886 outeriter += nri;
887
888 /* Update outer/inner flops */
889
890 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*288);
891 }
892 /*
893 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse2_double
894 * Electrostatics interaction: ReactionField
895 * VdW interaction: None
896 * Geometry: Water3-Water3
897 * Calculate force/pot: Force
898 */
899 void
900 nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse2_double
901 (t_nblist * gmx_restrict nlist,
902 rvec * gmx_restrict xx,
903 rvec * gmx_restrict ff,
904 t_forcerec * gmx_restrict fr,
905 t_mdatoms * gmx_restrict mdatoms,
906 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
907 t_nrnb * gmx_restrict nrnb)
908 {
909 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
910 * just 0 for non-waters.
911 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
912 * jnr indices corresponding to data put in the four positions in the SIMD register.
913 */
914 int i_shift_offset,i_coord_offset,outeriter,inneriter;
915 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
916 int jnrA,jnrB;
917 int j_coord_offsetA,j_coord_offsetB;
918 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
919 real rcutoff_scalar;
920 real *shiftvec,*fshift,*x,*f;
921 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
922 int vdwioffset0;
923 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
924 int vdwioffset1;
925 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
926 int vdwioffset2;
927 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
928 int vdwjidx0A,vdwjidx0B;
929 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
930 int vdwjidx1A,vdwjidx1B;
931 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
932 int vdwjidx2A,vdwjidx2B;
933 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
934 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
935 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
936 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
937 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
938 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
939 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
940 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
941 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
942 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
943 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
944 real *charge;
945 __m128d dummy_mask,cutoff_mask;
946 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
947 __m128d one = _mm_set1_pd(1.0);
948 __m128d two = _mm_set1_pd(2.0);
949 x = xx[0];
950 f = ff[0];
951
952 nri = nlist->nri;
953 iinr = nlist->iinr;
954 jindex = nlist->jindex;
955 jjnr = nlist->jjnr;
956 shiftidx = nlist->shift;
957 gid = nlist->gid;
958 shiftvec = fr->shift_vec[0];
959 fshift = fr->fshift[0];
960 facel = _mm_set1_pd(fr->epsfac);
961 charge = mdatoms->chargeA;
962 krf = _mm_set1_pd(fr->ic->k_rf);
963 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
964 crf = _mm_set1_pd(fr->ic->c_rf);
965
966 /* Setup water-specific parameters */
967 inr = nlist->iinr[0];
968 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
969 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
970 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
971
972 jq0 = _mm_set1_pd(charge[inr+0]);
973 jq1 = _mm_set1_pd(charge[inr+1]);
974 jq2 = _mm_set1_pd(charge[inr+2]);
975 qq00 = _mm_mul_pd(iq0,jq0);
976 qq01 = _mm_mul_pd(iq0,jq1);
977 qq02 = _mm_mul_pd(iq0,jq2);
978 qq10 = _mm_mul_pd(iq1,jq0);
979 qq11 = _mm_mul_pd(iq1,jq1);
980 qq12 = _mm_mul_pd(iq1,jq2);
981 qq20 = _mm_mul_pd(iq2,jq0);
982 qq21 = _mm_mul_pd(iq2,jq1);
983 qq22 = _mm_mul_pd(iq2,jq2);
984
985 /* Avoid stupid compiler warnings */
986 jnrA = jnrB = 0;
987 j_coord_offsetA = 0;
988 j_coord_offsetB = 0;
989
990 outeriter = 0;
991 inneriter = 0;
992
993 /* Start outer loop over neighborlists */
994 for(iidx=0; iidx<nri; iidx++)
995 {
996 /* Load shift vector for this list */
997 i_shift_offset = DIM*shiftidx[iidx];
998
999 /* Load limits for loop over neighbors */
1000 j_index_start = jindex[iidx];
1001 j_index_end = jindex[iidx+1];
1002
1003 /* Get outer coordinate index */
1004 inr = iinr[iidx];
1005 i_coord_offset = DIM*inr;
1006
1007 /* Load i particle coords and add shift vector */
1008 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1009 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1010
1011 fix0 = _mm_setzero_pd();
1012 fiy0 = _mm_setzero_pd();
1013 fiz0 = _mm_setzero_pd();
1014 fix1 = _mm_setzero_pd();
1015 fiy1 = _mm_setzero_pd();
1016 fiz1 = _mm_setzero_pd();
1017 fix2 = _mm_setzero_pd();
1018 fiy2 = _mm_setzero_pd();
1019 fiz2 = _mm_setzero_pd();
1020
1021 /* Start inner kernel loop */
1022 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1023 {
1024
1025 /* Get j neighbor index, and coordinate index */
1026 jnrA = jjnr[jidx];
1027 jnrB = jjnr[jidx+1];
1028 j_coord_offsetA = DIM*jnrA;
1029 j_coord_offsetB = DIM*jnrB;
1030
1031 /* load j atom coordinates */
1032 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1033 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1034
1035 /* Calculate displacement vector */
1036 dx00 = _mm_sub_pd(ix0,jx0);
1037 dy00 = _mm_sub_pd(iy0,jy0);
1038 dz00 = _mm_sub_pd(iz0,jz0);
1039 dx01 = _mm_sub_pd(ix0,jx1);
1040 dy01 = _mm_sub_pd(iy0,jy1);
1041 dz01 = _mm_sub_pd(iz0,jz1);
1042 dx02 = _mm_sub_pd(ix0,jx2);
1043 dy02 = _mm_sub_pd(iy0,jy2);
1044 dz02 = _mm_sub_pd(iz0,jz2);
1045 dx10 = _mm_sub_pd(ix1,jx0);
1046 dy10 = _mm_sub_pd(iy1,jy0);
1047 dz10 = _mm_sub_pd(iz1,jz0);
1048 dx11 = _mm_sub_pd(ix1,jx1);
1049 dy11 = _mm_sub_pd(iy1,jy1);
1050 dz11 = _mm_sub_pd(iz1,jz1);
1051 dx12 = _mm_sub_pd(ix1,jx2);
1052 dy12 = _mm_sub_pd(iy1,jy2);
1053 dz12 = _mm_sub_pd(iz1,jz2);
1054 dx20 = _mm_sub_pd(ix2,jx0);
1055 dy20 = _mm_sub_pd(iy2,jy0);
1056 dz20 = _mm_sub_pd(iz2,jz0);
1057 dx21 = _mm_sub_pd(ix2,jx1);
1058 dy21 = _mm_sub_pd(iy2,jy1);
1059 dz21 = _mm_sub_pd(iz2,jz1);
1060 dx22 = _mm_sub_pd(ix2,jx2);
1061 dy22 = _mm_sub_pd(iy2,jy2);
1062 dz22 = _mm_sub_pd(iz2,jz2);
1063
1064 /* Calculate squared distance and things based on it */
1065 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1066 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1067 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1068 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1069 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1070 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1071 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1072 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1073 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1074
1075 rinv00 = gmx_mm_invsqrt_pd(rsq00);
1076 rinv01 = gmx_mm_invsqrt_pd(rsq01);
1077 rinv02 = gmx_mm_invsqrt_pd(rsq02);
1078 rinv10 = gmx_mm_invsqrt_pd(rsq10);
1079 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1080 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1081 rinv20 = gmx_mm_invsqrt_pd(rsq20);
1082 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1083 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1084
1085 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1086 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1087 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1088 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1089 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1090 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1091 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1092 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1093 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1094
1095 fjx0 = _mm_setzero_pd();
1096 fjy0 = _mm_setzero_pd();
1097 fjz0 = _mm_setzero_pd();
1098 fjx1 = _mm_setzero_pd();
1099 fjy1 = _mm_setzero_pd();
1100 fjz1 = _mm_setzero_pd();
1101 fjx2 = _mm_setzero_pd();
1102 fjy2 = _mm_setzero_pd();
1103 fjz2 = _mm_setzero_pd();
1104
1105 /**************************
1106 * CALCULATE INTERACTIONS *
1107 **************************/
1108
1109 /* REACTION-FIELD ELECTROSTATICS */
1110 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1111
1112 fscal = felec;
1113
1114 /* Calculate temporary vectorial force */
1115 tx = _mm_mul_pd(fscal,dx00);
1116 ty = _mm_mul_pd(fscal,dy00);
1117 tz = _mm_mul_pd(fscal,dz00);
1118
1119 /* Update vectorial force */
1120 fix0 = _mm_add_pd(fix0,tx);
1121 fiy0 = _mm_add_pd(fiy0,ty);
1122 fiz0 = _mm_add_pd(fiz0,tz);
1123
1124 fjx0 = _mm_add_pd(fjx0,tx);
1125 fjy0 = _mm_add_pd(fjy0,ty);
1126 fjz0 = _mm_add_pd(fjz0,tz);
1127
1128 /**************************
1129 * CALCULATE INTERACTIONS *
1130 **************************/
1131
1132 /* REACTION-FIELD ELECTROSTATICS */
1133 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1134
1135 fscal = felec;
1136
1137 /* Calculate temporary vectorial force */
1138 tx = _mm_mul_pd(fscal,dx01);
1139 ty = _mm_mul_pd(fscal,dy01);
1140 tz = _mm_mul_pd(fscal,dz01);
1141
1142 /* Update vectorial force */
1143 fix0 = _mm_add_pd(fix0,tx);
1144 fiy0 = _mm_add_pd(fiy0,ty);
1145 fiz0 = _mm_add_pd(fiz0,tz);
1146
1147 fjx1 = _mm_add_pd(fjx1,tx);
1148 fjy1 = _mm_add_pd(fjy1,ty);
1149 fjz1 = _mm_add_pd(fjz1,tz);
1150
1151 /**************************
1152 * CALCULATE INTERACTIONS *
1153 **************************/
1154
1155 /* REACTION-FIELD ELECTROSTATICS */
1156 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1157
1158 fscal = felec;
1159
1160 /* Calculate temporary vectorial force */
1161 tx = _mm_mul_pd(fscal,dx02);
1162 ty = _mm_mul_pd(fscal,dy02);
1163 tz = _mm_mul_pd(fscal,dz02);
1164
1165 /* Update vectorial force */
1166 fix0 = _mm_add_pd(fix0,tx);
1167 fiy0 = _mm_add_pd(fiy0,ty);
1168 fiz0 = _mm_add_pd(fiz0,tz);
1169
1170 fjx2 = _mm_add_pd(fjx2,tx);
1171 fjy2 = _mm_add_pd(fjy2,ty);
1172 fjz2 = _mm_add_pd(fjz2,tz);
1173
1174 /**************************
1175 * CALCULATE INTERACTIONS *
1176 **************************/
1177
1178 /* REACTION-FIELD ELECTROSTATICS */
1179 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1180
1181 fscal = felec;
1182
1183 /* Calculate temporary vectorial force */
1184 tx = _mm_mul_pd(fscal,dx10);
1185 ty = _mm_mul_pd(fscal,dy10);
1186 tz = _mm_mul_pd(fscal,dz10);
1187
1188 /* Update vectorial force */
1189 fix1 = _mm_add_pd(fix1,tx);
1190 fiy1 = _mm_add_pd(fiy1,ty);
1191 fiz1 = _mm_add_pd(fiz1,tz);
1192
1193 fjx0 = _mm_add_pd(fjx0,tx);
1194 fjy0 = _mm_add_pd(fjy0,ty);
1195 fjz0 = _mm_add_pd(fjz0,tz);
1196
1197 /**************************
1198 * CALCULATE INTERACTIONS *
1199 **************************/
1200
1201 /* REACTION-FIELD ELECTROSTATICS */
1202 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1203
1204 fscal = felec;
1205
1206 /* Calculate temporary vectorial force */
1207 tx = _mm_mul_pd(fscal,dx11);
1208 ty = _mm_mul_pd(fscal,dy11);
1209 tz = _mm_mul_pd(fscal,dz11);
1210
1211 /* Update vectorial force */
1212 fix1 = _mm_add_pd(fix1,tx);
1213 fiy1 = _mm_add_pd(fiy1,ty);
1214 fiz1 = _mm_add_pd(fiz1,tz);
1215
1216 fjx1 = _mm_add_pd(fjx1,tx);
1217 fjy1 = _mm_add_pd(fjy1,ty);
1218 fjz1 = _mm_add_pd(fjz1,tz);
1219
1220 /**************************
1221 * CALCULATE INTERACTIONS *
1222 **************************/
1223
1224 /* REACTION-FIELD ELECTROSTATICS */
1225 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1226
1227 fscal = felec;
1228
1229 /* Calculate temporary vectorial force */
1230 tx = _mm_mul_pd(fscal,dx12);
1231 ty = _mm_mul_pd(fscal,dy12);
1232 tz = _mm_mul_pd(fscal,dz12);
1233
1234 /* Update vectorial force */
1235 fix1 = _mm_add_pd(fix1,tx);
1236 fiy1 = _mm_add_pd(fiy1,ty);
1237 fiz1 = _mm_add_pd(fiz1,tz);
1238
1239 fjx2 = _mm_add_pd(fjx2,tx);
1240 fjy2 = _mm_add_pd(fjy2,ty);
1241 fjz2 = _mm_add_pd(fjz2,tz);
1242
1243 /**************************
1244 * CALCULATE INTERACTIONS *
1245 **************************/
1246
1247 /* REACTION-FIELD ELECTROSTATICS */
1248 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1249
1250 fscal = felec;
1251
1252 /* Calculate temporary vectorial force */
1253 tx = _mm_mul_pd(fscal,dx20);
1254 ty = _mm_mul_pd(fscal,dy20);
1255 tz = _mm_mul_pd(fscal,dz20);
1256
1257 /* Update vectorial force */
1258 fix2 = _mm_add_pd(fix2,tx);
1259 fiy2 = _mm_add_pd(fiy2,ty);
1260 fiz2 = _mm_add_pd(fiz2,tz);
1261
1262 fjx0 = _mm_add_pd(fjx0,tx);
1263 fjy0 = _mm_add_pd(fjy0,ty);
1264 fjz0 = _mm_add_pd(fjz0,tz);
1265
1266 /**************************
1267 * CALCULATE INTERACTIONS *
1268 **************************/
1269
1270 /* REACTION-FIELD ELECTROSTATICS */
1271 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1272
1273 fscal = felec;
1274
1275 /* Calculate temporary vectorial force */
1276 tx = _mm_mul_pd(fscal,dx21);
1277 ty = _mm_mul_pd(fscal,dy21);
1278 tz = _mm_mul_pd(fscal,dz21);
1279
1280 /* Update vectorial force */
1281 fix2 = _mm_add_pd(fix2,tx);
1282 fiy2 = _mm_add_pd(fiy2,ty);
1283 fiz2 = _mm_add_pd(fiz2,tz);
1284
1285 fjx1 = _mm_add_pd(fjx1,tx);
1286 fjy1 = _mm_add_pd(fjy1,ty);
1287 fjz1 = _mm_add_pd(fjz1,tz);
1288
1289 /**************************
1290 * CALCULATE INTERACTIONS *
1291 **************************/
1292
1293 /* REACTION-FIELD ELECTROSTATICS */
1294 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1295
1296 fscal = felec;
1297
1298 /* Calculate temporary vectorial force */
1299 tx = _mm_mul_pd(fscal,dx22);
1300 ty = _mm_mul_pd(fscal,dy22);
1301 tz = _mm_mul_pd(fscal,dz22);
1302
1303 /* Update vectorial force */
1304 fix2 = _mm_add_pd(fix2,tx);
1305 fiy2 = _mm_add_pd(fiy2,ty);
1306 fiz2 = _mm_add_pd(fiz2,tz);
1307
1308 fjx2 = _mm_add_pd(fjx2,tx);
1309 fjy2 = _mm_add_pd(fjy2,ty);
1310 fjz2 = _mm_add_pd(fjz2,tz);
1311
1312 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1313
1314 /* Inner loop uses 243 flops */
1315 }
1316
1317 if(jidx<j_index_end)
1318 {
1319
1320 jnrA = jjnr[jidx];
1321 j_coord_offsetA = DIM*jnrA;
1322
1323 /* load j atom coordinates */
1324 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1325 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1326
1327 /* Calculate displacement vector */
1328 dx00 = _mm_sub_pd(ix0,jx0);
1329 dy00 = _mm_sub_pd(iy0,jy0);
1330 dz00 = _mm_sub_pd(iz0,jz0);
1331 dx01 = _mm_sub_pd(ix0,jx1);
1332 dy01 = _mm_sub_pd(iy0,jy1);
1333 dz01 = _mm_sub_pd(iz0,jz1);
1334 dx02 = _mm_sub_pd(ix0,jx2);
1335 dy02 = _mm_sub_pd(iy0,jy2);
1336 dz02 = _mm_sub_pd(iz0,jz2);
1337 dx10 = _mm_sub_pd(ix1,jx0);
1338 dy10 = _mm_sub_pd(iy1,jy0);
1339 dz10 = _mm_sub_pd(iz1,jz0);
1340 dx11 = _mm_sub_pd(ix1,jx1);
1341 dy11 = _mm_sub_pd(iy1,jy1);
1342 dz11 = _mm_sub_pd(iz1,jz1);
1343 dx12 = _mm_sub_pd(ix1,jx2);
1344 dy12 = _mm_sub_pd(iy1,jy2);
1345 dz12 = _mm_sub_pd(iz1,jz2);
1346 dx20 = _mm_sub_pd(ix2,jx0);
1347 dy20 = _mm_sub_pd(iy2,jy0);
1348 dz20 = _mm_sub_pd(iz2,jz0);
1349 dx21 = _mm_sub_pd(ix2,jx1);
1350 dy21 = _mm_sub_pd(iy2,jy1);
1351 dz21 = _mm_sub_pd(iz2,jz1);
1352 dx22 = _mm_sub_pd(ix2,jx2);
1353 dy22 = _mm_sub_pd(iy2,jy2);
1354 dz22 = _mm_sub_pd(iz2,jz2);
1355
1356 /* Calculate squared distance and things based on it */
1357 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1358 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1359 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1360 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1361 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1362 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1363 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1364 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1365 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1366
1367 rinv00 = gmx_mm_invsqrt_pd(rsq00);
1368 rinv01 = gmx_mm_invsqrt_pd(rsq01);
1369 rinv02 = gmx_mm_invsqrt_pd(rsq02);
1370 rinv10 = gmx_mm_invsqrt_pd(rsq10);
1371 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1372 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1373 rinv20 = gmx_mm_invsqrt_pd(rsq20);
1374 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1375 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1376
1377 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1378 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1379 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1380 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1381 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1382 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1383 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1384 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1385 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1386
1387 fjx0 = _mm_setzero_pd();
1388 fjy0 = _mm_setzero_pd();
1389 fjz0 = _mm_setzero_pd();
1390 fjx1 = _mm_setzero_pd();
1391 fjy1 = _mm_setzero_pd();
1392 fjz1 = _mm_setzero_pd();
1393 fjx2 = _mm_setzero_pd();
1394 fjy2 = _mm_setzero_pd();
1395 fjz2 = _mm_setzero_pd();
1396
1397 /**************************
1398 * CALCULATE INTERACTIONS *
1399 **************************/
1400
1401 /* REACTION-FIELD ELECTROSTATICS */
1402 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1403
1404 fscal = felec;
1405
1406 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1407
1408 /* Calculate temporary vectorial force */
1409 tx = _mm_mul_pd(fscal,dx00);
1410 ty = _mm_mul_pd(fscal,dy00);
1411 tz = _mm_mul_pd(fscal,dz00);
1412
1413 /* Update vectorial force */
1414 fix0 = _mm_add_pd(fix0,tx);
1415 fiy0 = _mm_add_pd(fiy0,ty);
1416 fiz0 = _mm_add_pd(fiz0,tz);
1417
1418 fjx0 = _mm_add_pd(fjx0,tx);
1419 fjy0 = _mm_add_pd(fjy0,ty);
1420 fjz0 = _mm_add_pd(fjz0,tz);
1421
1422 /**************************
1423 * CALCULATE INTERACTIONS *
1424 **************************/
1425
1426 /* REACTION-FIELD ELECTROSTATICS */
1427 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1428
1429 fscal = felec;
1430
1431 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1432
1433 /* Calculate temporary vectorial force */
1434 tx = _mm_mul_pd(fscal,dx01);
1435 ty = _mm_mul_pd(fscal,dy01);
1436 tz = _mm_mul_pd(fscal,dz01);
1437
1438 /* Update vectorial force */
1439 fix0 = _mm_add_pd(fix0,tx);
1440 fiy0 = _mm_add_pd(fiy0,ty);
1441 fiz0 = _mm_add_pd(fiz0,tz);
1442
1443 fjx1 = _mm_add_pd(fjx1,tx);
1444 fjy1 = _mm_add_pd(fjy1,ty);
1445 fjz1 = _mm_add_pd(fjz1,tz);
1446
1447 /**************************
1448 * CALCULATE INTERACTIONS *
1449 **************************/
1450
1451 /* REACTION-FIELD ELECTROSTATICS */
1452 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1453
1454 fscal = felec;
1455
1456 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1457
1458 /* Calculate temporary vectorial force */
1459 tx = _mm_mul_pd(fscal,dx02);
1460 ty = _mm_mul_pd(fscal,dy02);
1461 tz = _mm_mul_pd(fscal,dz02);
1462
1463 /* Update vectorial force */
1464 fix0 = _mm_add_pd(fix0,tx);
1465 fiy0 = _mm_add_pd(fiy0,ty);
1466 fiz0 = _mm_add_pd(fiz0,tz);
1467
1468 fjx2 = _mm_add_pd(fjx2,tx);
1469 fjy2 = _mm_add_pd(fjy2,ty);
1470 fjz2 = _mm_add_pd(fjz2,tz);
1471
1472 /**************************
1473 * CALCULATE INTERACTIONS *
1474 **************************/
1475
1476 /* REACTION-FIELD ELECTROSTATICS */
1477 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1478
1479 fscal = felec;
1480
1481 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1482
1483 /* Calculate temporary vectorial force */
1484 tx = _mm_mul_pd(fscal,dx10);
1485 ty = _mm_mul_pd(fscal,dy10);
1486 tz = _mm_mul_pd(fscal,dz10);
1487
1488 /* Update vectorial force */
1489 fix1 = _mm_add_pd(fix1,tx);
1490 fiy1 = _mm_add_pd(fiy1,ty);
1491 fiz1 = _mm_add_pd(fiz1,tz);
1492
1493 fjx0 = _mm_add_pd(fjx0,tx);
1494 fjy0 = _mm_add_pd(fjy0,ty);
1495 fjz0 = _mm_add_pd(fjz0,tz);
1496
1497 /**************************
1498 * CALCULATE INTERACTIONS *
1499 **************************/
1500
1501 /* REACTION-FIELD ELECTROSTATICS */
1502 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1503
1504 fscal = felec;
1505
1506 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1507
1508 /* Calculate temporary vectorial force */
1509 tx = _mm_mul_pd(fscal,dx11);
1510 ty = _mm_mul_pd(fscal,dy11);
1511 tz = _mm_mul_pd(fscal,dz11);
1512
1513 /* Update vectorial force */
1514 fix1 = _mm_add_pd(fix1,tx);
1515 fiy1 = _mm_add_pd(fiy1,ty);
1516 fiz1 = _mm_add_pd(fiz1,tz);
1517
1518 fjx1 = _mm_add_pd(fjx1,tx);
1519 fjy1 = _mm_add_pd(fjy1,ty);
1520 fjz1 = _mm_add_pd(fjz1,tz);
1521
1522 /**************************
1523 * CALCULATE INTERACTIONS *
1524 **************************/
1525
1526 /* REACTION-FIELD ELECTROSTATICS */
1527 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1528
1529 fscal = felec;
1530
1531 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1532
1533 /* Calculate temporary vectorial force */
1534 tx = _mm_mul_pd(fscal,dx12);
1535 ty = _mm_mul_pd(fscal,dy12);
1536 tz = _mm_mul_pd(fscal,dz12);
1537
1538 /* Update vectorial force */
1539 fix1 = _mm_add_pd(fix1,tx);
1540 fiy1 = _mm_add_pd(fiy1,ty);
1541 fiz1 = _mm_add_pd(fiz1,tz);
1542
1543 fjx2 = _mm_add_pd(fjx2,tx);
1544 fjy2 = _mm_add_pd(fjy2,ty);
1545 fjz2 = _mm_add_pd(fjz2,tz);
1546
1547 /**************************
1548 * CALCULATE INTERACTIONS *
1549 **************************/
1550
1551 /* REACTION-FIELD ELECTROSTATICS */
1552 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1553
1554 fscal = felec;
1555
1556 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1557
1558 /* Calculate temporary vectorial force */
1559 tx = _mm_mul_pd(fscal,dx20);
1560 ty = _mm_mul_pd(fscal,dy20);
1561 tz = _mm_mul_pd(fscal,dz20);
1562
1563 /* Update vectorial force */
1564 fix2 = _mm_add_pd(fix2,tx);
1565 fiy2 = _mm_add_pd(fiy2,ty);
1566 fiz2 = _mm_add_pd(fiz2,tz);
1567
1568 fjx0 = _mm_add_pd(fjx0,tx);
1569 fjy0 = _mm_add_pd(fjy0,ty);
1570 fjz0 = _mm_add_pd(fjz0,tz);
1571
1572 /**************************
1573 * CALCULATE INTERACTIONS *
1574 **************************/
1575
1576 /* REACTION-FIELD ELECTROSTATICS */
1577 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1578
1579 fscal = felec;
1580
1581 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1582
1583 /* Calculate temporary vectorial force */
1584 tx = _mm_mul_pd(fscal,dx21);
1585 ty = _mm_mul_pd(fscal,dy21);
1586 tz = _mm_mul_pd(fscal,dz21);
1587
1588 /* Update vectorial force */
1589 fix2 = _mm_add_pd(fix2,tx);
1590 fiy2 = _mm_add_pd(fiy2,ty);
1591 fiz2 = _mm_add_pd(fiz2,tz);
1592
1593 fjx1 = _mm_add_pd(fjx1,tx);
1594 fjy1 = _mm_add_pd(fjy1,ty);
1595 fjz1 = _mm_add_pd(fjz1,tz);
1596
1597 /**************************
1598 * CALCULATE INTERACTIONS *
1599 **************************/
1600
1601 /* REACTION-FIELD ELECTROSTATICS */
1602 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1603
1604 fscal = felec;
1605
1606 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1607
1608 /* Calculate temporary vectorial force */
1609 tx = _mm_mul_pd(fscal,dx22);
1610 ty = _mm_mul_pd(fscal,dy22);
1611 tz = _mm_mul_pd(fscal,dz22);
1612
1613 /* Update vectorial force */
1614 fix2 = _mm_add_pd(fix2,tx);
1615 fiy2 = _mm_add_pd(fiy2,ty);
1616 fiz2 = _mm_add_pd(fiz2,tz);
1617
1618 fjx2 = _mm_add_pd(fjx2,tx);
1619 fjy2 = _mm_add_pd(fjy2,ty);
1620 fjz2 = _mm_add_pd(fjz2,tz);
1621
1622 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1623
1624 /* Inner loop uses 243 flops */
1625 }
1626
1627 /* End of innermost loop */
1628
1629 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1630 f+i_coord_offset,fshift+i_shift_offset);
1631
1632 /* Increment number of inner iterations */
1633 inneriter += j_index_end - j_index_start;
1634
1635 /* Outer loop uses 18 flops */
1636 }
1637
1638 /* Increment number of outer iterations */
1639 outeriter += nri;
1640
1641 /* Update outer/inner flops */
1642
1643 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*243);
1644 }
The ultimate molecular dynamics simulation package