?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Remove nb-parameters from t_forcerec
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse4_1_double / nb_kernel_ElecRF_VdwLJ_GeomW3W3_sse4_1_double.c
blob3052784fd1733f789dc8a27f5cf5f42a68dab48f
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014,2015,2017, 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 sse4_1_double kernel generator.
37 */
38 #include "gmxpre.h"
39
40 #include "config.h"
41
42 #include <math.h>
43
44 #include "../nb_kernel.h"
45 #include "gromacs/gmxlib/nrnb.h"
46
47 #include "kernelutil_x86_sse4_1_double.h"
48
49 /*
50 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_sse4_1_double
51 * Electrostatics interaction: ReactionField
52 * VdW interaction: LennardJones
53 * Geometry: Water3-Water3
54 * Calculate force/pot: PotentialAndForce
55 */
56 void
57 nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_sse4_1_double
58 (t_nblist * gmx_restrict nlist,
59 rvec * gmx_restrict xx,
60 rvec * gmx_restrict ff,
61 struct t_forcerec * gmx_restrict fr,
62 t_mdatoms * gmx_restrict mdatoms,
63 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
64 t_nrnb * gmx_restrict nrnb)
65 {
66 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
67 * just 0 for non-waters.
68 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
69 * jnr indices corresponding to data put in the four positions in the SIMD register.
70 */
71 int i_shift_offset,i_coord_offset,outeriter,inneriter;
72 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
73 int jnrA,jnrB;
74 int j_coord_offsetA,j_coord_offsetB;
75 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
76 real rcutoff_scalar;
77 real *shiftvec,*fshift,*x,*f;
78 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
79 int vdwioffset0;
80 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
81 int vdwioffset1;
82 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
83 int vdwioffset2;
84 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
85 int vdwjidx0A,vdwjidx0B;
86 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
87 int vdwjidx1A,vdwjidx1B;
88 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
89 int vdwjidx2A,vdwjidx2B;
90 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
91 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
92 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
93 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
94 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
95 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
96 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
97 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
98 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
99 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
100 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
101 real *charge;
102 int nvdwtype;
103 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
104 int *vdwtype;
105 real *vdwparam;
106 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
107 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
108 __m128d dummy_mask,cutoff_mask;
109 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
110 __m128d one = _mm_set1_pd(1.0);
111 __m128d two = _mm_set1_pd(2.0);
112 x = xx[0];
113 f = ff[0];
114
115 nri = nlist->nri;
116 iinr = nlist->iinr;
117 jindex = nlist->jindex;
118 jjnr = nlist->jjnr;
119 shiftidx = nlist->shift;
120 gid = nlist->gid;
121 shiftvec = fr->shift_vec[0];
122 fshift = fr->fshift[0];
123 facel = _mm_set1_pd(fr->ic->epsfac);
124 charge = mdatoms->chargeA;
125 krf = _mm_set1_pd(fr->ic->k_rf);
126 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
127 crf = _mm_set1_pd(fr->ic->c_rf);
128 nvdwtype = fr->ntype;
129 vdwparam = fr->nbfp;
130 vdwtype = mdatoms->typeA;
131
132 /* Setup water-specific parameters */
133 inr = nlist->iinr[0];
134 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
135 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
136 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
137 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
138
139 jq0 = _mm_set1_pd(charge[inr+0]);
140 jq1 = _mm_set1_pd(charge[inr+1]);
141 jq2 = _mm_set1_pd(charge[inr+2]);
142 vdwjidx0A = 2*vdwtype[inr+0];
143 qq00 = _mm_mul_pd(iq0,jq0);
144 c6_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
145 c12_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
146 qq01 = _mm_mul_pd(iq0,jq1);
147 qq02 = _mm_mul_pd(iq0,jq2);
148 qq10 = _mm_mul_pd(iq1,jq0);
149 qq11 = _mm_mul_pd(iq1,jq1);
150 qq12 = _mm_mul_pd(iq1,jq2);
151 qq20 = _mm_mul_pd(iq2,jq0);
152 qq21 = _mm_mul_pd(iq2,jq1);
153 qq22 = _mm_mul_pd(iq2,jq2);
154
155 /* Avoid stupid compiler warnings */
156 jnrA = jnrB = 0;
157 j_coord_offsetA = 0;
158 j_coord_offsetB = 0;
159
160 outeriter = 0;
161 inneriter = 0;
162
163 /* Start outer loop over neighborlists */
164 for(iidx=0; iidx<nri; iidx++)
165 {
166 /* Load shift vector for this list */
167 i_shift_offset = DIM*shiftidx[iidx];
168
169 /* Load limits for loop over neighbors */
170 j_index_start = jindex[iidx];
171 j_index_end = jindex[iidx+1];
172
173 /* Get outer coordinate index */
174 inr = iinr[iidx];
175 i_coord_offset = DIM*inr;
176
177 /* Load i particle coords and add shift vector */
178 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
179 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
180
181 fix0 = _mm_setzero_pd();
182 fiy0 = _mm_setzero_pd();
183 fiz0 = _mm_setzero_pd();
184 fix1 = _mm_setzero_pd();
185 fiy1 = _mm_setzero_pd();
186 fiz1 = _mm_setzero_pd();
187 fix2 = _mm_setzero_pd();
188 fiy2 = _mm_setzero_pd();
189 fiz2 = _mm_setzero_pd();
190
191 /* Reset potential sums */
192 velecsum = _mm_setzero_pd();
193 vvdwsum = _mm_setzero_pd();
194
195 /* Start inner kernel loop */
196 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
197 {
198
199 /* Get j neighbor index, and coordinate index */
200 jnrA = jjnr[jidx];
201 jnrB = jjnr[jidx+1];
202 j_coord_offsetA = DIM*jnrA;
203 j_coord_offsetB = DIM*jnrB;
204
205 /* load j atom coordinates */
206 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
207 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
208
209 /* Calculate displacement vector */
210 dx00 = _mm_sub_pd(ix0,jx0);
211 dy00 = _mm_sub_pd(iy0,jy0);
212 dz00 = _mm_sub_pd(iz0,jz0);
213 dx01 = _mm_sub_pd(ix0,jx1);
214 dy01 = _mm_sub_pd(iy0,jy1);
215 dz01 = _mm_sub_pd(iz0,jz1);
216 dx02 = _mm_sub_pd(ix0,jx2);
217 dy02 = _mm_sub_pd(iy0,jy2);
218 dz02 = _mm_sub_pd(iz0,jz2);
219 dx10 = _mm_sub_pd(ix1,jx0);
220 dy10 = _mm_sub_pd(iy1,jy0);
221 dz10 = _mm_sub_pd(iz1,jz0);
222 dx11 = _mm_sub_pd(ix1,jx1);
223 dy11 = _mm_sub_pd(iy1,jy1);
224 dz11 = _mm_sub_pd(iz1,jz1);
225 dx12 = _mm_sub_pd(ix1,jx2);
226 dy12 = _mm_sub_pd(iy1,jy2);
227 dz12 = _mm_sub_pd(iz1,jz2);
228 dx20 = _mm_sub_pd(ix2,jx0);
229 dy20 = _mm_sub_pd(iy2,jy0);
230 dz20 = _mm_sub_pd(iz2,jz0);
231 dx21 = _mm_sub_pd(ix2,jx1);
232 dy21 = _mm_sub_pd(iy2,jy1);
233 dz21 = _mm_sub_pd(iz2,jz1);
234 dx22 = _mm_sub_pd(ix2,jx2);
235 dy22 = _mm_sub_pd(iy2,jy2);
236 dz22 = _mm_sub_pd(iz2,jz2);
237
238 /* Calculate squared distance and things based on it */
239 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
240 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
241 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
242 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
243 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
244 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
245 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
246 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
247 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
248
249 rinv00 = sse41_invsqrt_d(rsq00);
250 rinv01 = sse41_invsqrt_d(rsq01);
251 rinv02 = sse41_invsqrt_d(rsq02);
252 rinv10 = sse41_invsqrt_d(rsq10);
253 rinv11 = sse41_invsqrt_d(rsq11);
254 rinv12 = sse41_invsqrt_d(rsq12);
255 rinv20 = sse41_invsqrt_d(rsq20);
256 rinv21 = sse41_invsqrt_d(rsq21);
257 rinv22 = sse41_invsqrt_d(rsq22);
258
259 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
260 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
261 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
262 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
263 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
264 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
265 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
266 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
267 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
268
269 fjx0 = _mm_setzero_pd();
270 fjy0 = _mm_setzero_pd();
271 fjz0 = _mm_setzero_pd();
272 fjx1 = _mm_setzero_pd();
273 fjy1 = _mm_setzero_pd();
274 fjz1 = _mm_setzero_pd();
275 fjx2 = _mm_setzero_pd();
276 fjy2 = _mm_setzero_pd();
277 fjz2 = _mm_setzero_pd();
278
279 /**************************
280 * CALCULATE INTERACTIONS *
281 **************************/
282
283 /* REACTION-FIELD ELECTROSTATICS */
284 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
285 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
286
287 /* LENNARD-JONES DISPERSION/REPULSION */
288
289 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
290 vvdw6 = _mm_mul_pd(c6_00,rinvsix);
291 vvdw12 = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
292 vvdw = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
293 fvdw = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
294
295 /* Update potential sum for this i atom from the interaction with this j atom. */
296 velecsum = _mm_add_pd(velecsum,velec);
297 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
298
299 fscal = _mm_add_pd(felec,fvdw);
300
301 /* Calculate temporary vectorial force */
302 tx = _mm_mul_pd(fscal,dx00);
303 ty = _mm_mul_pd(fscal,dy00);
304 tz = _mm_mul_pd(fscal,dz00);
305
306 /* Update vectorial force */
307 fix0 = _mm_add_pd(fix0,tx);
308 fiy0 = _mm_add_pd(fiy0,ty);
309 fiz0 = _mm_add_pd(fiz0,tz);
310
311 fjx0 = _mm_add_pd(fjx0,tx);
312 fjy0 = _mm_add_pd(fjy0,ty);
313 fjz0 = _mm_add_pd(fjz0,tz);
314
315 /**************************
316 * CALCULATE INTERACTIONS *
317 **************************/
318
319 /* REACTION-FIELD ELECTROSTATICS */
320 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
321 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
322
323 /* Update potential sum for this i atom from the interaction with this j atom. */
324 velecsum = _mm_add_pd(velecsum,velec);
325
326 fscal = felec;
327
328 /* Calculate temporary vectorial force */
329 tx = _mm_mul_pd(fscal,dx01);
330 ty = _mm_mul_pd(fscal,dy01);
331 tz = _mm_mul_pd(fscal,dz01);
332
333 /* Update vectorial force */
334 fix0 = _mm_add_pd(fix0,tx);
335 fiy0 = _mm_add_pd(fiy0,ty);
336 fiz0 = _mm_add_pd(fiz0,tz);
337
338 fjx1 = _mm_add_pd(fjx1,tx);
339 fjy1 = _mm_add_pd(fjy1,ty);
340 fjz1 = _mm_add_pd(fjz1,tz);
341
342 /**************************
343 * CALCULATE INTERACTIONS *
344 **************************/
345
346 /* REACTION-FIELD ELECTROSTATICS */
347 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
348 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
349
350 /* Update potential sum for this i atom from the interaction with this j atom. */
351 velecsum = _mm_add_pd(velecsum,velec);
352
353 fscal = felec;
354
355 /* Calculate temporary vectorial force */
356 tx = _mm_mul_pd(fscal,dx02);
357 ty = _mm_mul_pd(fscal,dy02);
358 tz = _mm_mul_pd(fscal,dz02);
359
360 /* Update vectorial force */
361 fix0 = _mm_add_pd(fix0,tx);
362 fiy0 = _mm_add_pd(fiy0,ty);
363 fiz0 = _mm_add_pd(fiz0,tz);
364
365 fjx2 = _mm_add_pd(fjx2,tx);
366 fjy2 = _mm_add_pd(fjy2,ty);
367 fjz2 = _mm_add_pd(fjz2,tz);
368
369 /**************************
370 * CALCULATE INTERACTIONS *
371 **************************/
372
373 /* REACTION-FIELD ELECTROSTATICS */
374 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
375 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
376
377 /* Update potential sum for this i atom from the interaction with this j atom. */
378 velecsum = _mm_add_pd(velecsum,velec);
379
380 fscal = felec;
381
382 /* Calculate temporary vectorial force */
383 tx = _mm_mul_pd(fscal,dx10);
384 ty = _mm_mul_pd(fscal,dy10);
385 tz = _mm_mul_pd(fscal,dz10);
386
387 /* Update vectorial force */
388 fix1 = _mm_add_pd(fix1,tx);
389 fiy1 = _mm_add_pd(fiy1,ty);
390 fiz1 = _mm_add_pd(fiz1,tz);
391
392 fjx0 = _mm_add_pd(fjx0,tx);
393 fjy0 = _mm_add_pd(fjy0,ty);
394 fjz0 = _mm_add_pd(fjz0,tz);
395
396 /**************************
397 * CALCULATE INTERACTIONS *
398 **************************/
399
400 /* REACTION-FIELD ELECTROSTATICS */
401 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
402 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
403
404 /* Update potential sum for this i atom from the interaction with this j atom. */
405 velecsum = _mm_add_pd(velecsum,velec);
406
407 fscal = felec;
408
409 /* Calculate temporary vectorial force */
410 tx = _mm_mul_pd(fscal,dx11);
411 ty = _mm_mul_pd(fscal,dy11);
412 tz = _mm_mul_pd(fscal,dz11);
413
414 /* Update vectorial force */
415 fix1 = _mm_add_pd(fix1,tx);
416 fiy1 = _mm_add_pd(fiy1,ty);
417 fiz1 = _mm_add_pd(fiz1,tz);
418
419 fjx1 = _mm_add_pd(fjx1,tx);
420 fjy1 = _mm_add_pd(fjy1,ty);
421 fjz1 = _mm_add_pd(fjz1,tz);
422
423 /**************************
424 * CALCULATE INTERACTIONS *
425 **************************/
426
427 /* REACTION-FIELD ELECTROSTATICS */
428 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
429 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
430
431 /* Update potential sum for this i atom from the interaction with this j atom. */
432 velecsum = _mm_add_pd(velecsum,velec);
433
434 fscal = felec;
435
436 /* Calculate temporary vectorial force */
437 tx = _mm_mul_pd(fscal,dx12);
438 ty = _mm_mul_pd(fscal,dy12);
439 tz = _mm_mul_pd(fscal,dz12);
440
441 /* Update vectorial force */
442 fix1 = _mm_add_pd(fix1,tx);
443 fiy1 = _mm_add_pd(fiy1,ty);
444 fiz1 = _mm_add_pd(fiz1,tz);
445
446 fjx2 = _mm_add_pd(fjx2,tx);
447 fjy2 = _mm_add_pd(fjy2,ty);
448 fjz2 = _mm_add_pd(fjz2,tz);
449
450 /**************************
451 * CALCULATE INTERACTIONS *
452 **************************/
453
454 /* REACTION-FIELD ELECTROSTATICS */
455 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
456 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
457
458 /* Update potential sum for this i atom from the interaction with this j atom. */
459 velecsum = _mm_add_pd(velecsum,velec);
460
461 fscal = felec;
462
463 /* Calculate temporary vectorial force */
464 tx = _mm_mul_pd(fscal,dx20);
465 ty = _mm_mul_pd(fscal,dy20);
466 tz = _mm_mul_pd(fscal,dz20);
467
468 /* Update vectorial force */
469 fix2 = _mm_add_pd(fix2,tx);
470 fiy2 = _mm_add_pd(fiy2,ty);
471 fiz2 = _mm_add_pd(fiz2,tz);
472
473 fjx0 = _mm_add_pd(fjx0,tx);
474 fjy0 = _mm_add_pd(fjy0,ty);
475 fjz0 = _mm_add_pd(fjz0,tz);
476
477 /**************************
478 * CALCULATE INTERACTIONS *
479 **************************/
480
481 /* REACTION-FIELD ELECTROSTATICS */
482 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
483 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
484
485 /* Update potential sum for this i atom from the interaction with this j atom. */
486 velecsum = _mm_add_pd(velecsum,velec);
487
488 fscal = felec;
489
490 /* Calculate temporary vectorial force */
491 tx = _mm_mul_pd(fscal,dx21);
492 ty = _mm_mul_pd(fscal,dy21);
493 tz = _mm_mul_pd(fscal,dz21);
494
495 /* Update vectorial force */
496 fix2 = _mm_add_pd(fix2,tx);
497 fiy2 = _mm_add_pd(fiy2,ty);
498 fiz2 = _mm_add_pd(fiz2,tz);
499
500 fjx1 = _mm_add_pd(fjx1,tx);
501 fjy1 = _mm_add_pd(fjy1,ty);
502 fjz1 = _mm_add_pd(fjz1,tz);
503
504 /**************************
505 * CALCULATE INTERACTIONS *
506 **************************/
507
508 /* REACTION-FIELD ELECTROSTATICS */
509 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
510 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
511
512 /* Update potential sum for this i atom from the interaction with this j atom. */
513 velecsum = _mm_add_pd(velecsum,velec);
514
515 fscal = felec;
516
517 /* Calculate temporary vectorial force */
518 tx = _mm_mul_pd(fscal,dx22);
519 ty = _mm_mul_pd(fscal,dy22);
520 tz = _mm_mul_pd(fscal,dz22);
521
522 /* Update vectorial force */
523 fix2 = _mm_add_pd(fix2,tx);
524 fiy2 = _mm_add_pd(fiy2,ty);
525 fiz2 = _mm_add_pd(fiz2,tz);
526
527 fjx2 = _mm_add_pd(fjx2,tx);
528 fjy2 = _mm_add_pd(fjy2,ty);
529 fjz2 = _mm_add_pd(fjz2,tz);
530
531 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
532
533 /* Inner loop uses 300 flops */
534 }
535
536 if(jidx<j_index_end)
537 {
538
539 jnrA = jjnr[jidx];
540 j_coord_offsetA = DIM*jnrA;
541
542 /* load j atom coordinates */
543 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
544 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
545
546 /* Calculate displacement vector */
547 dx00 = _mm_sub_pd(ix0,jx0);
548 dy00 = _mm_sub_pd(iy0,jy0);
549 dz00 = _mm_sub_pd(iz0,jz0);
550 dx01 = _mm_sub_pd(ix0,jx1);
551 dy01 = _mm_sub_pd(iy0,jy1);
552 dz01 = _mm_sub_pd(iz0,jz1);
553 dx02 = _mm_sub_pd(ix0,jx2);
554 dy02 = _mm_sub_pd(iy0,jy2);
555 dz02 = _mm_sub_pd(iz0,jz2);
556 dx10 = _mm_sub_pd(ix1,jx0);
557 dy10 = _mm_sub_pd(iy1,jy0);
558 dz10 = _mm_sub_pd(iz1,jz0);
559 dx11 = _mm_sub_pd(ix1,jx1);
560 dy11 = _mm_sub_pd(iy1,jy1);
561 dz11 = _mm_sub_pd(iz1,jz1);
562 dx12 = _mm_sub_pd(ix1,jx2);
563 dy12 = _mm_sub_pd(iy1,jy2);
564 dz12 = _mm_sub_pd(iz1,jz2);
565 dx20 = _mm_sub_pd(ix2,jx0);
566 dy20 = _mm_sub_pd(iy2,jy0);
567 dz20 = _mm_sub_pd(iz2,jz0);
568 dx21 = _mm_sub_pd(ix2,jx1);
569 dy21 = _mm_sub_pd(iy2,jy1);
570 dz21 = _mm_sub_pd(iz2,jz1);
571 dx22 = _mm_sub_pd(ix2,jx2);
572 dy22 = _mm_sub_pd(iy2,jy2);
573 dz22 = _mm_sub_pd(iz2,jz2);
574
575 /* Calculate squared distance and things based on it */
576 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
577 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
578 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
579 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
580 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
581 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
582 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
583 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
584 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
585
586 rinv00 = sse41_invsqrt_d(rsq00);
587 rinv01 = sse41_invsqrt_d(rsq01);
588 rinv02 = sse41_invsqrt_d(rsq02);
589 rinv10 = sse41_invsqrt_d(rsq10);
590 rinv11 = sse41_invsqrt_d(rsq11);
591 rinv12 = sse41_invsqrt_d(rsq12);
592 rinv20 = sse41_invsqrt_d(rsq20);
593 rinv21 = sse41_invsqrt_d(rsq21);
594 rinv22 = sse41_invsqrt_d(rsq22);
595
596 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
597 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
598 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
599 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
600 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
601 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
602 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
603 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
604 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
605
606 fjx0 = _mm_setzero_pd();
607 fjy0 = _mm_setzero_pd();
608 fjz0 = _mm_setzero_pd();
609 fjx1 = _mm_setzero_pd();
610 fjy1 = _mm_setzero_pd();
611 fjz1 = _mm_setzero_pd();
612 fjx2 = _mm_setzero_pd();
613 fjy2 = _mm_setzero_pd();
614 fjz2 = _mm_setzero_pd();
615
616 /**************************
617 * CALCULATE INTERACTIONS *
618 **************************/
619
620 /* REACTION-FIELD ELECTROSTATICS */
621 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
622 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
623
624 /* LENNARD-JONES DISPERSION/REPULSION */
625
626 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
627 vvdw6 = _mm_mul_pd(c6_00,rinvsix);
628 vvdw12 = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
629 vvdw = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
630 fvdw = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
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 vvdw = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
636 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
637
638 fscal = _mm_add_pd(felec,fvdw);
639
640 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
641
642 /* Calculate temporary vectorial force */
643 tx = _mm_mul_pd(fscal,dx00);
644 ty = _mm_mul_pd(fscal,dy00);
645 tz = _mm_mul_pd(fscal,dz00);
646
647 /* Update vectorial force */
648 fix0 = _mm_add_pd(fix0,tx);
649 fiy0 = _mm_add_pd(fiy0,ty);
650 fiz0 = _mm_add_pd(fiz0,tz);
651
652 fjx0 = _mm_add_pd(fjx0,tx);
653 fjy0 = _mm_add_pd(fjy0,ty);
654 fjz0 = _mm_add_pd(fjz0,tz);
655
656 /**************************
657 * CALCULATE INTERACTIONS *
658 **************************/
659
660 /* REACTION-FIELD ELECTROSTATICS */
661 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
662 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
663
664 /* Update potential sum for this i atom from the interaction with this j atom. */
665 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
666 velecsum = _mm_add_pd(velecsum,velec);
667
668 fscal = felec;
669
670 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
671
672 /* Calculate temporary vectorial force */
673 tx = _mm_mul_pd(fscal,dx01);
674 ty = _mm_mul_pd(fscal,dy01);
675 tz = _mm_mul_pd(fscal,dz01);
676
677 /* Update vectorial force */
678 fix0 = _mm_add_pd(fix0,tx);
679 fiy0 = _mm_add_pd(fiy0,ty);
680 fiz0 = _mm_add_pd(fiz0,tz);
681
682 fjx1 = _mm_add_pd(fjx1,tx);
683 fjy1 = _mm_add_pd(fjy1,ty);
684 fjz1 = _mm_add_pd(fjz1,tz);
685
686 /**************************
687 * CALCULATE INTERACTIONS *
688 **************************/
689
690 /* REACTION-FIELD ELECTROSTATICS */
691 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
692 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
693
694 /* Update potential sum for this i atom from the interaction with this j atom. */
695 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
696 velecsum = _mm_add_pd(velecsum,velec);
697
698 fscal = felec;
699
700 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
701
702 /* Calculate temporary vectorial force */
703 tx = _mm_mul_pd(fscal,dx02);
704 ty = _mm_mul_pd(fscal,dy02);
705 tz = _mm_mul_pd(fscal,dz02);
706
707 /* Update vectorial force */
708 fix0 = _mm_add_pd(fix0,tx);
709 fiy0 = _mm_add_pd(fiy0,ty);
710 fiz0 = _mm_add_pd(fiz0,tz);
711
712 fjx2 = _mm_add_pd(fjx2,tx);
713 fjy2 = _mm_add_pd(fjy2,ty);
714 fjz2 = _mm_add_pd(fjz2,tz);
715
716 /**************************
717 * CALCULATE INTERACTIONS *
718 **************************/
719
720 /* REACTION-FIELD ELECTROSTATICS */
721 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
722 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
723
724 /* Update potential sum for this i atom from the interaction with this j atom. */
725 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
726 velecsum = _mm_add_pd(velecsum,velec);
727
728 fscal = felec;
729
730 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
731
732 /* Calculate temporary vectorial force */
733 tx = _mm_mul_pd(fscal,dx10);
734 ty = _mm_mul_pd(fscal,dy10);
735 tz = _mm_mul_pd(fscal,dz10);
736
737 /* Update vectorial force */
738 fix1 = _mm_add_pd(fix1,tx);
739 fiy1 = _mm_add_pd(fiy1,ty);
740 fiz1 = _mm_add_pd(fiz1,tz);
741
742 fjx0 = _mm_add_pd(fjx0,tx);
743 fjy0 = _mm_add_pd(fjy0,ty);
744 fjz0 = _mm_add_pd(fjz0,tz);
745
746 /**************************
747 * CALCULATE INTERACTIONS *
748 **************************/
749
750 /* REACTION-FIELD ELECTROSTATICS */
751 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
752 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
753
754 /* Update potential sum for this i atom from the interaction with this j atom. */
755 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
756 velecsum = _mm_add_pd(velecsum,velec);
757
758 fscal = felec;
759
760 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
761
762 /* Calculate temporary vectorial force */
763 tx = _mm_mul_pd(fscal,dx11);
764 ty = _mm_mul_pd(fscal,dy11);
765 tz = _mm_mul_pd(fscal,dz11);
766
767 /* Update vectorial force */
768 fix1 = _mm_add_pd(fix1,tx);
769 fiy1 = _mm_add_pd(fiy1,ty);
770 fiz1 = _mm_add_pd(fiz1,tz);
771
772 fjx1 = _mm_add_pd(fjx1,tx);
773 fjy1 = _mm_add_pd(fjy1,ty);
774 fjz1 = _mm_add_pd(fjz1,tz);
775
776 /**************************
777 * CALCULATE INTERACTIONS *
778 **************************/
779
780 /* REACTION-FIELD ELECTROSTATICS */
781 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
782 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
783
784 /* Update potential sum for this i atom from the interaction with this j atom. */
785 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
786 velecsum = _mm_add_pd(velecsum,velec);
787
788 fscal = felec;
789
790 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
791
792 /* Calculate temporary vectorial force */
793 tx = _mm_mul_pd(fscal,dx12);
794 ty = _mm_mul_pd(fscal,dy12);
795 tz = _mm_mul_pd(fscal,dz12);
796
797 /* Update vectorial force */
798 fix1 = _mm_add_pd(fix1,tx);
799 fiy1 = _mm_add_pd(fiy1,ty);
800 fiz1 = _mm_add_pd(fiz1,tz);
801
802 fjx2 = _mm_add_pd(fjx2,tx);
803 fjy2 = _mm_add_pd(fjy2,ty);
804 fjz2 = _mm_add_pd(fjz2,tz);
805
806 /**************************
807 * CALCULATE INTERACTIONS *
808 **************************/
809
810 /* REACTION-FIELD ELECTROSTATICS */
811 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
812 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
813
814 /* Update potential sum for this i atom from the interaction with this j atom. */
815 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
816 velecsum = _mm_add_pd(velecsum,velec);
817
818 fscal = felec;
819
820 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
821
822 /* Calculate temporary vectorial force */
823 tx = _mm_mul_pd(fscal,dx20);
824 ty = _mm_mul_pd(fscal,dy20);
825 tz = _mm_mul_pd(fscal,dz20);
826
827 /* Update vectorial force */
828 fix2 = _mm_add_pd(fix2,tx);
829 fiy2 = _mm_add_pd(fiy2,ty);
830 fiz2 = _mm_add_pd(fiz2,tz);
831
832 fjx0 = _mm_add_pd(fjx0,tx);
833 fjy0 = _mm_add_pd(fjy0,ty);
834 fjz0 = _mm_add_pd(fjz0,tz);
835
836 /**************************
837 * CALCULATE INTERACTIONS *
838 **************************/
839
840 /* REACTION-FIELD ELECTROSTATICS */
841 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
842 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
843
844 /* Update potential sum for this i atom from the interaction with this j atom. */
845 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
846 velecsum = _mm_add_pd(velecsum,velec);
847
848 fscal = felec;
849
850 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
851
852 /* Calculate temporary vectorial force */
853 tx = _mm_mul_pd(fscal,dx21);
854 ty = _mm_mul_pd(fscal,dy21);
855 tz = _mm_mul_pd(fscal,dz21);
856
857 /* Update vectorial force */
858 fix2 = _mm_add_pd(fix2,tx);
859 fiy2 = _mm_add_pd(fiy2,ty);
860 fiz2 = _mm_add_pd(fiz2,tz);
861
862 fjx1 = _mm_add_pd(fjx1,tx);
863 fjy1 = _mm_add_pd(fjy1,ty);
864 fjz1 = _mm_add_pd(fjz1,tz);
865
866 /**************************
867 * CALCULATE INTERACTIONS *
868 **************************/
869
870 /* REACTION-FIELD ELECTROSTATICS */
871 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
872 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
873
874 /* Update potential sum for this i atom from the interaction with this j atom. */
875 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
876 velecsum = _mm_add_pd(velecsum,velec);
877
878 fscal = felec;
879
880 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
881
882 /* Calculate temporary vectorial force */
883 tx = _mm_mul_pd(fscal,dx22);
884 ty = _mm_mul_pd(fscal,dy22);
885 tz = _mm_mul_pd(fscal,dz22);
886
887 /* Update vectorial force */
888 fix2 = _mm_add_pd(fix2,tx);
889 fiy2 = _mm_add_pd(fiy2,ty);
890 fiz2 = _mm_add_pd(fiz2,tz);
891
892 fjx2 = _mm_add_pd(fjx2,tx);
893 fjy2 = _mm_add_pd(fjy2,ty);
894 fjz2 = _mm_add_pd(fjz2,tz);
895
896 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
897
898 /* Inner loop uses 300 flops */
899 }
900
901 /* End of innermost loop */
902
903 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
904 f+i_coord_offset,fshift+i_shift_offset);
905
906 ggid = gid[iidx];
907 /* Update potential energies */
908 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
909 gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
910
911 /* Increment number of inner iterations */
912 inneriter += j_index_end - j_index_start;
913
914 /* Outer loop uses 20 flops */
915 }
916
917 /* Increment number of outer iterations */
918 outeriter += nri;
919
920 /* Update outer/inner flops */
921
922 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*300);
923 }
924 /*
925 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse4_1_double
926 * Electrostatics interaction: ReactionField
927 * VdW interaction: LennardJones
928 * Geometry: Water3-Water3
929 * Calculate force/pot: Force
930 */
931 void
932 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse4_1_double
933 (t_nblist * gmx_restrict nlist,
934 rvec * gmx_restrict xx,
935 rvec * gmx_restrict ff,
936 struct t_forcerec * gmx_restrict fr,
937 t_mdatoms * gmx_restrict mdatoms,
938 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
939 t_nrnb * gmx_restrict nrnb)
940 {
941 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
942 * just 0 for non-waters.
943 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
944 * jnr indices corresponding to data put in the four positions in the SIMD register.
945 */
946 int i_shift_offset,i_coord_offset,outeriter,inneriter;
947 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
948 int jnrA,jnrB;
949 int j_coord_offsetA,j_coord_offsetB;
950 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
951 real rcutoff_scalar;
952 real *shiftvec,*fshift,*x,*f;
953 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
954 int vdwioffset0;
955 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
956 int vdwioffset1;
957 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
958 int vdwioffset2;
959 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
960 int vdwjidx0A,vdwjidx0B;
961 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
962 int vdwjidx1A,vdwjidx1B;
963 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
964 int vdwjidx2A,vdwjidx2B;
965 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
966 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
967 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
968 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
969 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
970 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
971 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
972 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
973 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
974 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
975 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
976 real *charge;
977 int nvdwtype;
978 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
979 int *vdwtype;
980 real *vdwparam;
981 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
982 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
983 __m128d dummy_mask,cutoff_mask;
984 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
985 __m128d one = _mm_set1_pd(1.0);
986 __m128d two = _mm_set1_pd(2.0);
987 x = xx[0];
988 f = ff[0];
989
990 nri = nlist->nri;
991 iinr = nlist->iinr;
992 jindex = nlist->jindex;
993 jjnr = nlist->jjnr;
994 shiftidx = nlist->shift;
995 gid = nlist->gid;
996 shiftvec = fr->shift_vec[0];
997 fshift = fr->fshift[0];
998 facel = _mm_set1_pd(fr->ic->epsfac);
999 charge = mdatoms->chargeA;
1000 krf = _mm_set1_pd(fr->ic->k_rf);
1001 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
1002 crf = _mm_set1_pd(fr->ic->c_rf);
1003 nvdwtype = fr->ntype;
1004 vdwparam = fr->nbfp;
1005 vdwtype = mdatoms->typeA;
1006
1007 /* Setup water-specific parameters */
1008 inr = nlist->iinr[0];
1009 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1010 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1011 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1012 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1013
1014 jq0 = _mm_set1_pd(charge[inr+0]);
1015 jq1 = _mm_set1_pd(charge[inr+1]);
1016 jq2 = _mm_set1_pd(charge[inr+2]);
1017 vdwjidx0A = 2*vdwtype[inr+0];
1018 qq00 = _mm_mul_pd(iq0,jq0);
1019 c6_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1020 c12_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1021 qq01 = _mm_mul_pd(iq0,jq1);
1022 qq02 = _mm_mul_pd(iq0,jq2);
1023 qq10 = _mm_mul_pd(iq1,jq0);
1024 qq11 = _mm_mul_pd(iq1,jq1);
1025 qq12 = _mm_mul_pd(iq1,jq2);
1026 qq20 = _mm_mul_pd(iq2,jq0);
1027 qq21 = _mm_mul_pd(iq2,jq1);
1028 qq22 = _mm_mul_pd(iq2,jq2);
1029
1030 /* Avoid stupid compiler warnings */
1031 jnrA = jnrB = 0;
1032 j_coord_offsetA = 0;
1033 j_coord_offsetB = 0;
1034
1035 outeriter = 0;
1036 inneriter = 0;
1037
1038 /* Start outer loop over neighborlists */
1039 for(iidx=0; iidx<nri; iidx++)
1040 {
1041 /* Load shift vector for this list */
1042 i_shift_offset = DIM*shiftidx[iidx];
1043
1044 /* Load limits for loop over neighbors */
1045 j_index_start = jindex[iidx];
1046 j_index_end = jindex[iidx+1];
1047
1048 /* Get outer coordinate index */
1049 inr = iinr[iidx];
1050 i_coord_offset = DIM*inr;
1051
1052 /* Load i particle coords and add shift vector */
1053 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1054 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1055
1056 fix0 = _mm_setzero_pd();
1057 fiy0 = _mm_setzero_pd();
1058 fiz0 = _mm_setzero_pd();
1059 fix1 = _mm_setzero_pd();
1060 fiy1 = _mm_setzero_pd();
1061 fiz1 = _mm_setzero_pd();
1062 fix2 = _mm_setzero_pd();
1063 fiy2 = _mm_setzero_pd();
1064 fiz2 = _mm_setzero_pd();
1065
1066 /* Start inner kernel loop */
1067 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1068 {
1069
1070 /* Get j neighbor index, and coordinate index */
1071 jnrA = jjnr[jidx];
1072 jnrB = jjnr[jidx+1];
1073 j_coord_offsetA = DIM*jnrA;
1074 j_coord_offsetB = DIM*jnrB;
1075
1076 /* load j atom coordinates */
1077 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1078 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1079
1080 /* Calculate displacement vector */
1081 dx00 = _mm_sub_pd(ix0,jx0);
1082 dy00 = _mm_sub_pd(iy0,jy0);
1083 dz00 = _mm_sub_pd(iz0,jz0);
1084 dx01 = _mm_sub_pd(ix0,jx1);
1085 dy01 = _mm_sub_pd(iy0,jy1);
1086 dz01 = _mm_sub_pd(iz0,jz1);
1087 dx02 = _mm_sub_pd(ix0,jx2);
1088 dy02 = _mm_sub_pd(iy0,jy2);
1089 dz02 = _mm_sub_pd(iz0,jz2);
1090 dx10 = _mm_sub_pd(ix1,jx0);
1091 dy10 = _mm_sub_pd(iy1,jy0);
1092 dz10 = _mm_sub_pd(iz1,jz0);
1093 dx11 = _mm_sub_pd(ix1,jx1);
1094 dy11 = _mm_sub_pd(iy1,jy1);
1095 dz11 = _mm_sub_pd(iz1,jz1);
1096 dx12 = _mm_sub_pd(ix1,jx2);
1097 dy12 = _mm_sub_pd(iy1,jy2);
1098 dz12 = _mm_sub_pd(iz1,jz2);
1099 dx20 = _mm_sub_pd(ix2,jx0);
1100 dy20 = _mm_sub_pd(iy2,jy0);
1101 dz20 = _mm_sub_pd(iz2,jz0);
1102 dx21 = _mm_sub_pd(ix2,jx1);
1103 dy21 = _mm_sub_pd(iy2,jy1);
1104 dz21 = _mm_sub_pd(iz2,jz1);
1105 dx22 = _mm_sub_pd(ix2,jx2);
1106 dy22 = _mm_sub_pd(iy2,jy2);
1107 dz22 = _mm_sub_pd(iz2,jz2);
1108
1109 /* Calculate squared distance and things based on it */
1110 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1111 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1112 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1113 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1114 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1115 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1116 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1117 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1118 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1119
1120 rinv00 = sse41_invsqrt_d(rsq00);
1121 rinv01 = sse41_invsqrt_d(rsq01);
1122 rinv02 = sse41_invsqrt_d(rsq02);
1123 rinv10 = sse41_invsqrt_d(rsq10);
1124 rinv11 = sse41_invsqrt_d(rsq11);
1125 rinv12 = sse41_invsqrt_d(rsq12);
1126 rinv20 = sse41_invsqrt_d(rsq20);
1127 rinv21 = sse41_invsqrt_d(rsq21);
1128 rinv22 = sse41_invsqrt_d(rsq22);
1129
1130 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1131 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1132 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1133 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1134 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1135 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1136 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1137 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1138 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1139
1140 fjx0 = _mm_setzero_pd();
1141 fjy0 = _mm_setzero_pd();
1142 fjz0 = _mm_setzero_pd();
1143 fjx1 = _mm_setzero_pd();
1144 fjy1 = _mm_setzero_pd();
1145 fjz1 = _mm_setzero_pd();
1146 fjx2 = _mm_setzero_pd();
1147 fjy2 = _mm_setzero_pd();
1148 fjz2 = _mm_setzero_pd();
1149
1150 /**************************
1151 * CALCULATE INTERACTIONS *
1152 **************************/
1153
1154 /* REACTION-FIELD ELECTROSTATICS */
1155 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1156
1157 /* LENNARD-JONES DISPERSION/REPULSION */
1158
1159 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1160 fvdw = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1161
1162 fscal = _mm_add_pd(felec,fvdw);
1163
1164 /* Calculate temporary vectorial force */
1165 tx = _mm_mul_pd(fscal,dx00);
1166 ty = _mm_mul_pd(fscal,dy00);
1167 tz = _mm_mul_pd(fscal,dz00);
1168
1169 /* Update vectorial force */
1170 fix0 = _mm_add_pd(fix0,tx);
1171 fiy0 = _mm_add_pd(fiy0,ty);
1172 fiz0 = _mm_add_pd(fiz0,tz);
1173
1174 fjx0 = _mm_add_pd(fjx0,tx);
1175 fjy0 = _mm_add_pd(fjy0,ty);
1176 fjz0 = _mm_add_pd(fjz0,tz);
1177
1178 /**************************
1179 * CALCULATE INTERACTIONS *
1180 **************************/
1181
1182 /* REACTION-FIELD ELECTROSTATICS */
1183 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1184
1185 fscal = felec;
1186
1187 /* Calculate temporary vectorial force */
1188 tx = _mm_mul_pd(fscal,dx01);
1189 ty = _mm_mul_pd(fscal,dy01);
1190 tz = _mm_mul_pd(fscal,dz01);
1191
1192 /* Update vectorial force */
1193 fix0 = _mm_add_pd(fix0,tx);
1194 fiy0 = _mm_add_pd(fiy0,ty);
1195 fiz0 = _mm_add_pd(fiz0,tz);
1196
1197 fjx1 = _mm_add_pd(fjx1,tx);
1198 fjy1 = _mm_add_pd(fjy1,ty);
1199 fjz1 = _mm_add_pd(fjz1,tz);
1200
1201 /**************************
1202 * CALCULATE INTERACTIONS *
1203 **************************/
1204
1205 /* REACTION-FIELD ELECTROSTATICS */
1206 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1207
1208 fscal = felec;
1209
1210 /* Calculate temporary vectorial force */
1211 tx = _mm_mul_pd(fscal,dx02);
1212 ty = _mm_mul_pd(fscal,dy02);
1213 tz = _mm_mul_pd(fscal,dz02);
1214
1215 /* Update vectorial force */
1216 fix0 = _mm_add_pd(fix0,tx);
1217 fiy0 = _mm_add_pd(fiy0,ty);
1218 fiz0 = _mm_add_pd(fiz0,tz);
1219
1220 fjx2 = _mm_add_pd(fjx2,tx);
1221 fjy2 = _mm_add_pd(fjy2,ty);
1222 fjz2 = _mm_add_pd(fjz2,tz);
1223
1224 /**************************
1225 * CALCULATE INTERACTIONS *
1226 **************************/
1227
1228 /* REACTION-FIELD ELECTROSTATICS */
1229 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1230
1231 fscal = felec;
1232
1233 /* Calculate temporary vectorial force */
1234 tx = _mm_mul_pd(fscal,dx10);
1235 ty = _mm_mul_pd(fscal,dy10);
1236 tz = _mm_mul_pd(fscal,dz10);
1237
1238 /* Update vectorial force */
1239 fix1 = _mm_add_pd(fix1,tx);
1240 fiy1 = _mm_add_pd(fiy1,ty);
1241 fiz1 = _mm_add_pd(fiz1,tz);
1242
1243 fjx0 = _mm_add_pd(fjx0,tx);
1244 fjy0 = _mm_add_pd(fjy0,ty);
1245 fjz0 = _mm_add_pd(fjz0,tz);
1246
1247 /**************************
1248 * CALCULATE INTERACTIONS *
1249 **************************/
1250
1251 /* REACTION-FIELD ELECTROSTATICS */
1252 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1253
1254 fscal = felec;
1255
1256 /* Calculate temporary vectorial force */
1257 tx = _mm_mul_pd(fscal,dx11);
1258 ty = _mm_mul_pd(fscal,dy11);
1259 tz = _mm_mul_pd(fscal,dz11);
1260
1261 /* Update vectorial force */
1262 fix1 = _mm_add_pd(fix1,tx);
1263 fiy1 = _mm_add_pd(fiy1,ty);
1264 fiz1 = _mm_add_pd(fiz1,tz);
1265
1266 fjx1 = _mm_add_pd(fjx1,tx);
1267 fjy1 = _mm_add_pd(fjy1,ty);
1268 fjz1 = _mm_add_pd(fjz1,tz);
1269
1270 /**************************
1271 * CALCULATE INTERACTIONS *
1272 **************************/
1273
1274 /* REACTION-FIELD ELECTROSTATICS */
1275 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1276
1277 fscal = felec;
1278
1279 /* Calculate temporary vectorial force */
1280 tx = _mm_mul_pd(fscal,dx12);
1281 ty = _mm_mul_pd(fscal,dy12);
1282 tz = _mm_mul_pd(fscal,dz12);
1283
1284 /* Update vectorial force */
1285 fix1 = _mm_add_pd(fix1,tx);
1286 fiy1 = _mm_add_pd(fiy1,ty);
1287 fiz1 = _mm_add_pd(fiz1,tz);
1288
1289 fjx2 = _mm_add_pd(fjx2,tx);
1290 fjy2 = _mm_add_pd(fjy2,ty);
1291 fjz2 = _mm_add_pd(fjz2,tz);
1292
1293 /**************************
1294 * CALCULATE INTERACTIONS *
1295 **************************/
1296
1297 /* REACTION-FIELD ELECTROSTATICS */
1298 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1299
1300 fscal = felec;
1301
1302 /* Calculate temporary vectorial force */
1303 tx = _mm_mul_pd(fscal,dx20);
1304 ty = _mm_mul_pd(fscal,dy20);
1305 tz = _mm_mul_pd(fscal,dz20);
1306
1307 /* Update vectorial force */
1308 fix2 = _mm_add_pd(fix2,tx);
1309 fiy2 = _mm_add_pd(fiy2,ty);
1310 fiz2 = _mm_add_pd(fiz2,tz);
1311
1312 fjx0 = _mm_add_pd(fjx0,tx);
1313 fjy0 = _mm_add_pd(fjy0,ty);
1314 fjz0 = _mm_add_pd(fjz0,tz);
1315
1316 /**************************
1317 * CALCULATE INTERACTIONS *
1318 **************************/
1319
1320 /* REACTION-FIELD ELECTROSTATICS */
1321 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1322
1323 fscal = felec;
1324
1325 /* Calculate temporary vectorial force */
1326 tx = _mm_mul_pd(fscal,dx21);
1327 ty = _mm_mul_pd(fscal,dy21);
1328 tz = _mm_mul_pd(fscal,dz21);
1329
1330 /* Update vectorial force */
1331 fix2 = _mm_add_pd(fix2,tx);
1332 fiy2 = _mm_add_pd(fiy2,ty);
1333 fiz2 = _mm_add_pd(fiz2,tz);
1334
1335 fjx1 = _mm_add_pd(fjx1,tx);
1336 fjy1 = _mm_add_pd(fjy1,ty);
1337 fjz1 = _mm_add_pd(fjz1,tz);
1338
1339 /**************************
1340 * CALCULATE INTERACTIONS *
1341 **************************/
1342
1343 /* REACTION-FIELD ELECTROSTATICS */
1344 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1345
1346 fscal = felec;
1347
1348 /* Calculate temporary vectorial force */
1349 tx = _mm_mul_pd(fscal,dx22);
1350 ty = _mm_mul_pd(fscal,dy22);
1351 tz = _mm_mul_pd(fscal,dz22);
1352
1353 /* Update vectorial force */
1354 fix2 = _mm_add_pd(fix2,tx);
1355 fiy2 = _mm_add_pd(fiy2,ty);
1356 fiz2 = _mm_add_pd(fiz2,tz);
1357
1358 fjx2 = _mm_add_pd(fjx2,tx);
1359 fjy2 = _mm_add_pd(fjy2,ty);
1360 fjz2 = _mm_add_pd(fjz2,tz);
1361
1362 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1363
1364 /* Inner loop uses 250 flops */
1365 }
1366
1367 if(jidx<j_index_end)
1368 {
1369
1370 jnrA = jjnr[jidx];
1371 j_coord_offsetA = DIM*jnrA;
1372
1373 /* load j atom coordinates */
1374 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1375 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1376
1377 /* Calculate displacement vector */
1378 dx00 = _mm_sub_pd(ix0,jx0);
1379 dy00 = _mm_sub_pd(iy0,jy0);
1380 dz00 = _mm_sub_pd(iz0,jz0);
1381 dx01 = _mm_sub_pd(ix0,jx1);
1382 dy01 = _mm_sub_pd(iy0,jy1);
1383 dz01 = _mm_sub_pd(iz0,jz1);
1384 dx02 = _mm_sub_pd(ix0,jx2);
1385 dy02 = _mm_sub_pd(iy0,jy2);
1386 dz02 = _mm_sub_pd(iz0,jz2);
1387 dx10 = _mm_sub_pd(ix1,jx0);
1388 dy10 = _mm_sub_pd(iy1,jy0);
1389 dz10 = _mm_sub_pd(iz1,jz0);
1390 dx11 = _mm_sub_pd(ix1,jx1);
1391 dy11 = _mm_sub_pd(iy1,jy1);
1392 dz11 = _mm_sub_pd(iz1,jz1);
1393 dx12 = _mm_sub_pd(ix1,jx2);
1394 dy12 = _mm_sub_pd(iy1,jy2);
1395 dz12 = _mm_sub_pd(iz1,jz2);
1396 dx20 = _mm_sub_pd(ix2,jx0);
1397 dy20 = _mm_sub_pd(iy2,jy0);
1398 dz20 = _mm_sub_pd(iz2,jz0);
1399 dx21 = _mm_sub_pd(ix2,jx1);
1400 dy21 = _mm_sub_pd(iy2,jy1);
1401 dz21 = _mm_sub_pd(iz2,jz1);
1402 dx22 = _mm_sub_pd(ix2,jx2);
1403 dy22 = _mm_sub_pd(iy2,jy2);
1404 dz22 = _mm_sub_pd(iz2,jz2);
1405
1406 /* Calculate squared distance and things based on it */
1407 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1408 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1409 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1410 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1411 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1412 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1413 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1414 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1415 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1416
1417 rinv00 = sse41_invsqrt_d(rsq00);
1418 rinv01 = sse41_invsqrt_d(rsq01);
1419 rinv02 = sse41_invsqrt_d(rsq02);
1420 rinv10 = sse41_invsqrt_d(rsq10);
1421 rinv11 = sse41_invsqrt_d(rsq11);
1422 rinv12 = sse41_invsqrt_d(rsq12);
1423 rinv20 = sse41_invsqrt_d(rsq20);
1424 rinv21 = sse41_invsqrt_d(rsq21);
1425 rinv22 = sse41_invsqrt_d(rsq22);
1426
1427 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1428 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1429 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1430 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1431 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1432 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1433 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1434 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1435 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1436
1437 fjx0 = _mm_setzero_pd();
1438 fjy0 = _mm_setzero_pd();
1439 fjz0 = _mm_setzero_pd();
1440 fjx1 = _mm_setzero_pd();
1441 fjy1 = _mm_setzero_pd();
1442 fjz1 = _mm_setzero_pd();
1443 fjx2 = _mm_setzero_pd();
1444 fjy2 = _mm_setzero_pd();
1445 fjz2 = _mm_setzero_pd();
1446
1447 /**************************
1448 * CALCULATE INTERACTIONS *
1449 **************************/
1450
1451 /* REACTION-FIELD ELECTROSTATICS */
1452 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1453
1454 /* LENNARD-JONES DISPERSION/REPULSION */
1455
1456 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1457 fvdw = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1458
1459 fscal = _mm_add_pd(felec,fvdw);
1460
1461 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1462
1463 /* Calculate temporary vectorial force */
1464 tx = _mm_mul_pd(fscal,dx00);
1465 ty = _mm_mul_pd(fscal,dy00);
1466 tz = _mm_mul_pd(fscal,dz00);
1467
1468 /* Update vectorial force */
1469 fix0 = _mm_add_pd(fix0,tx);
1470 fiy0 = _mm_add_pd(fiy0,ty);
1471 fiz0 = _mm_add_pd(fiz0,tz);
1472
1473 fjx0 = _mm_add_pd(fjx0,tx);
1474 fjy0 = _mm_add_pd(fjy0,ty);
1475 fjz0 = _mm_add_pd(fjz0,tz);
1476
1477 /**************************
1478 * CALCULATE INTERACTIONS *
1479 **************************/
1480
1481 /* REACTION-FIELD ELECTROSTATICS */
1482 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1483
1484 fscal = felec;
1485
1486 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1487
1488 /* Calculate temporary vectorial force */
1489 tx = _mm_mul_pd(fscal,dx01);
1490 ty = _mm_mul_pd(fscal,dy01);
1491 tz = _mm_mul_pd(fscal,dz01);
1492
1493 /* Update vectorial force */
1494 fix0 = _mm_add_pd(fix0,tx);
1495 fiy0 = _mm_add_pd(fiy0,ty);
1496 fiz0 = _mm_add_pd(fiz0,tz);
1497
1498 fjx1 = _mm_add_pd(fjx1,tx);
1499 fjy1 = _mm_add_pd(fjy1,ty);
1500 fjz1 = _mm_add_pd(fjz1,tz);
1501
1502 /**************************
1503 * CALCULATE INTERACTIONS *
1504 **************************/
1505
1506 /* REACTION-FIELD ELECTROSTATICS */
1507 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1508
1509 fscal = felec;
1510
1511 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1512
1513 /* Calculate temporary vectorial force */
1514 tx = _mm_mul_pd(fscal,dx02);
1515 ty = _mm_mul_pd(fscal,dy02);
1516 tz = _mm_mul_pd(fscal,dz02);
1517
1518 /* Update vectorial force */
1519 fix0 = _mm_add_pd(fix0,tx);
1520 fiy0 = _mm_add_pd(fiy0,ty);
1521 fiz0 = _mm_add_pd(fiz0,tz);
1522
1523 fjx2 = _mm_add_pd(fjx2,tx);
1524 fjy2 = _mm_add_pd(fjy2,ty);
1525 fjz2 = _mm_add_pd(fjz2,tz);
1526
1527 /**************************
1528 * CALCULATE INTERACTIONS *
1529 **************************/
1530
1531 /* REACTION-FIELD ELECTROSTATICS */
1532 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1533
1534 fscal = felec;
1535
1536 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1537
1538 /* Calculate temporary vectorial force */
1539 tx = _mm_mul_pd(fscal,dx10);
1540 ty = _mm_mul_pd(fscal,dy10);
1541 tz = _mm_mul_pd(fscal,dz10);
1542
1543 /* Update vectorial force */
1544 fix1 = _mm_add_pd(fix1,tx);
1545 fiy1 = _mm_add_pd(fiy1,ty);
1546 fiz1 = _mm_add_pd(fiz1,tz);
1547
1548 fjx0 = _mm_add_pd(fjx0,tx);
1549 fjy0 = _mm_add_pd(fjy0,ty);
1550 fjz0 = _mm_add_pd(fjz0,tz);
1551
1552 /**************************
1553 * CALCULATE INTERACTIONS *
1554 **************************/
1555
1556 /* REACTION-FIELD ELECTROSTATICS */
1557 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1558
1559 fscal = felec;
1560
1561 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1562
1563 /* Calculate temporary vectorial force */
1564 tx = _mm_mul_pd(fscal,dx11);
1565 ty = _mm_mul_pd(fscal,dy11);
1566 tz = _mm_mul_pd(fscal,dz11);
1567
1568 /* Update vectorial force */
1569 fix1 = _mm_add_pd(fix1,tx);
1570 fiy1 = _mm_add_pd(fiy1,ty);
1571 fiz1 = _mm_add_pd(fiz1,tz);
1572
1573 fjx1 = _mm_add_pd(fjx1,tx);
1574 fjy1 = _mm_add_pd(fjy1,ty);
1575 fjz1 = _mm_add_pd(fjz1,tz);
1576
1577 /**************************
1578 * CALCULATE INTERACTIONS *
1579 **************************/
1580
1581 /* REACTION-FIELD ELECTROSTATICS */
1582 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1583
1584 fscal = felec;
1585
1586 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1587
1588 /* Calculate temporary vectorial force */
1589 tx = _mm_mul_pd(fscal,dx12);
1590 ty = _mm_mul_pd(fscal,dy12);
1591 tz = _mm_mul_pd(fscal,dz12);
1592
1593 /* Update vectorial force */
1594 fix1 = _mm_add_pd(fix1,tx);
1595 fiy1 = _mm_add_pd(fiy1,ty);
1596 fiz1 = _mm_add_pd(fiz1,tz);
1597
1598 fjx2 = _mm_add_pd(fjx2,tx);
1599 fjy2 = _mm_add_pd(fjy2,ty);
1600 fjz2 = _mm_add_pd(fjz2,tz);
1601
1602 /**************************
1603 * CALCULATE INTERACTIONS *
1604 **************************/
1605
1606 /* REACTION-FIELD ELECTROSTATICS */
1607 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1608
1609 fscal = felec;
1610
1611 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1612
1613 /* Calculate temporary vectorial force */
1614 tx = _mm_mul_pd(fscal,dx20);
1615 ty = _mm_mul_pd(fscal,dy20);
1616 tz = _mm_mul_pd(fscal,dz20);
1617
1618 /* Update vectorial force */
1619 fix2 = _mm_add_pd(fix2,tx);
1620 fiy2 = _mm_add_pd(fiy2,ty);
1621 fiz2 = _mm_add_pd(fiz2,tz);
1622
1623 fjx0 = _mm_add_pd(fjx0,tx);
1624 fjy0 = _mm_add_pd(fjy0,ty);
1625 fjz0 = _mm_add_pd(fjz0,tz);
1626
1627 /**************************
1628 * CALCULATE INTERACTIONS *
1629 **************************/
1630
1631 /* REACTION-FIELD ELECTROSTATICS */
1632 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1633
1634 fscal = felec;
1635
1636 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1637
1638 /* Calculate temporary vectorial force */
1639 tx = _mm_mul_pd(fscal,dx21);
1640 ty = _mm_mul_pd(fscal,dy21);
1641 tz = _mm_mul_pd(fscal,dz21);
1642
1643 /* Update vectorial force */
1644 fix2 = _mm_add_pd(fix2,tx);
1645 fiy2 = _mm_add_pd(fiy2,ty);
1646 fiz2 = _mm_add_pd(fiz2,tz);
1647
1648 fjx1 = _mm_add_pd(fjx1,tx);
1649 fjy1 = _mm_add_pd(fjy1,ty);
1650 fjz1 = _mm_add_pd(fjz1,tz);
1651
1652 /**************************
1653 * CALCULATE INTERACTIONS *
1654 **************************/
1655
1656 /* REACTION-FIELD ELECTROSTATICS */
1657 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1658
1659 fscal = felec;
1660
1661 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1662
1663 /* Calculate temporary vectorial force */
1664 tx = _mm_mul_pd(fscal,dx22);
1665 ty = _mm_mul_pd(fscal,dy22);
1666 tz = _mm_mul_pd(fscal,dz22);
1667
1668 /* Update vectorial force */
1669 fix2 = _mm_add_pd(fix2,tx);
1670 fiy2 = _mm_add_pd(fiy2,ty);
1671 fiz2 = _mm_add_pd(fiz2,tz);
1672
1673 fjx2 = _mm_add_pd(fjx2,tx);
1674 fjy2 = _mm_add_pd(fjy2,ty);
1675 fjz2 = _mm_add_pd(fjz2,tz);
1676
1677 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1678
1679 /* Inner loop uses 250 flops */
1680 }
1681
1682 /* End of innermost loop */
1683
1684 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1685 f+i_coord_offset,fshift+i_shift_offset);
1686
1687 /* Increment number of inner iterations */
1688 inneriter += j_index_end - j_index_start;
1689
1690 /* Outer loop uses 18 flops */
1691 }
1692
1693 /* Increment number of outer iterations */
1694 outeriter += nri;
1695
1696 /* Update outer/inner flops */
1697
1698 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1699 }
The ultimate molecular dynamics simulation package