?  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_GeomW4W4_sse2_double.c
blob04a06822c6be7aeaa4fc24a2b79ece6dcf8110b2
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_GeomW4W4_VF_sse2_double
53 * Electrostatics interaction: ReactionField
54 * VdW interaction: None
55 * Geometry: Water4-Water4
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecRF_VdwNone_GeomW4W4_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 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 vdwioffset3;
86 __m128d ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
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 int vdwjidx3A,vdwjidx3B;
92 __m128d jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
93 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
94 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
95 __m128d dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
96 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
97 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
98 __m128d dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
99 __m128d dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
100 __m128d dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
101 __m128d dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
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 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
128 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
129 iq3 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
130
131 jq1 = _mm_set1_pd(charge[inr+1]);
132 jq2 = _mm_set1_pd(charge[inr+2]);
133 jq3 = _mm_set1_pd(charge[inr+3]);
134 qq11 = _mm_mul_pd(iq1,jq1);
135 qq12 = _mm_mul_pd(iq1,jq2);
136 qq13 = _mm_mul_pd(iq1,jq3);
137 qq21 = _mm_mul_pd(iq2,jq1);
138 qq22 = _mm_mul_pd(iq2,jq2);
139 qq23 = _mm_mul_pd(iq2,jq3);
140 qq31 = _mm_mul_pd(iq3,jq1);
141 qq32 = _mm_mul_pd(iq3,jq2);
142 qq33 = _mm_mul_pd(iq3,jq3);
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+DIM,
168 &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
169
170 fix1 = _mm_setzero_pd();
171 fiy1 = _mm_setzero_pd();
172 fiz1 = _mm_setzero_pd();
173 fix2 = _mm_setzero_pd();
174 fiy2 = _mm_setzero_pd();
175 fiz2 = _mm_setzero_pd();
176 fix3 = _mm_setzero_pd();
177 fiy3 = _mm_setzero_pd();
178 fiz3 = _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+DIM,x+j_coord_offsetB+DIM,
195 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
196
197 /* Calculate displacement vector */
198 dx11 = _mm_sub_pd(ix1,jx1);
199 dy11 = _mm_sub_pd(iy1,jy1);
200 dz11 = _mm_sub_pd(iz1,jz1);
201 dx12 = _mm_sub_pd(ix1,jx2);
202 dy12 = _mm_sub_pd(iy1,jy2);
203 dz12 = _mm_sub_pd(iz1,jz2);
204 dx13 = _mm_sub_pd(ix1,jx3);
205 dy13 = _mm_sub_pd(iy1,jy3);
206 dz13 = _mm_sub_pd(iz1,jz3);
207 dx21 = _mm_sub_pd(ix2,jx1);
208 dy21 = _mm_sub_pd(iy2,jy1);
209 dz21 = _mm_sub_pd(iz2,jz1);
210 dx22 = _mm_sub_pd(ix2,jx2);
211 dy22 = _mm_sub_pd(iy2,jy2);
212 dz22 = _mm_sub_pd(iz2,jz2);
213 dx23 = _mm_sub_pd(ix2,jx3);
214 dy23 = _mm_sub_pd(iy2,jy3);
215 dz23 = _mm_sub_pd(iz2,jz3);
216 dx31 = _mm_sub_pd(ix3,jx1);
217 dy31 = _mm_sub_pd(iy3,jy1);
218 dz31 = _mm_sub_pd(iz3,jz1);
219 dx32 = _mm_sub_pd(ix3,jx2);
220 dy32 = _mm_sub_pd(iy3,jy2);
221 dz32 = _mm_sub_pd(iz3,jz2);
222 dx33 = _mm_sub_pd(ix3,jx3);
223 dy33 = _mm_sub_pd(iy3,jy3);
224 dz33 = _mm_sub_pd(iz3,jz3);
225
226 /* Calculate squared distance and things based on it */
227 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
228 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
229 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
230 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
231 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
232 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
233 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
234 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
235 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
236
237 rinv11 = gmx_mm_invsqrt_pd(rsq11);
238 rinv12 = gmx_mm_invsqrt_pd(rsq12);
239 rinv13 = gmx_mm_invsqrt_pd(rsq13);
240 rinv21 = gmx_mm_invsqrt_pd(rsq21);
241 rinv22 = gmx_mm_invsqrt_pd(rsq22);
242 rinv23 = gmx_mm_invsqrt_pd(rsq23);
243 rinv31 = gmx_mm_invsqrt_pd(rsq31);
244 rinv32 = gmx_mm_invsqrt_pd(rsq32);
245 rinv33 = gmx_mm_invsqrt_pd(rsq33);
246
247 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
248 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
249 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
250 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
251 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
252 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
253 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
254 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
255 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
256
257 fjx1 = _mm_setzero_pd();
258 fjy1 = _mm_setzero_pd();
259 fjz1 = _mm_setzero_pd();
260 fjx2 = _mm_setzero_pd();
261 fjy2 = _mm_setzero_pd();
262 fjz2 = _mm_setzero_pd();
263 fjx3 = _mm_setzero_pd();
264 fjy3 = _mm_setzero_pd();
265 fjz3 = _mm_setzero_pd();
266
267 /**************************
268 * CALCULATE INTERACTIONS *
269 **************************/
270
271 /* REACTION-FIELD ELECTROSTATICS */
272 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
273 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),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,dx11);
282 ty = _mm_mul_pd(fscal,dy11);
283 tz = _mm_mul_pd(fscal,dz11);
284
285 /* Update vectorial force */
286 fix1 = _mm_add_pd(fix1,tx);
287 fiy1 = _mm_add_pd(fiy1,ty);
288 fiz1 = _mm_add_pd(fiz1,tz);
289
290 fjx1 = _mm_add_pd(fjx1,tx);
291 fjy1 = _mm_add_pd(fjy1,ty);
292 fjz1 = _mm_add_pd(fjz1,tz);
293
294 /**************************
295 * CALCULATE INTERACTIONS *
296 **************************/
297
298 /* REACTION-FIELD ELECTROSTATICS */
299 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
300 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),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,dx12);
309 ty = _mm_mul_pd(fscal,dy12);
310 tz = _mm_mul_pd(fscal,dz12);
311
312 /* Update vectorial force */
313 fix1 = _mm_add_pd(fix1,tx);
314 fiy1 = _mm_add_pd(fiy1,ty);
315 fiz1 = _mm_add_pd(fiz1,tz);
316
317 fjx2 = _mm_add_pd(fjx2,tx);
318 fjy2 = _mm_add_pd(fjy2,ty);
319 fjz2 = _mm_add_pd(fjz2,tz);
320
321 /**************************
322 * CALCULATE INTERACTIONS *
323 **************************/
324
325 /* REACTION-FIELD ELECTROSTATICS */
326 velec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
327 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),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,dx13);
336 ty = _mm_mul_pd(fscal,dy13);
337 tz = _mm_mul_pd(fscal,dz13);
338
339 /* Update vectorial force */
340 fix1 = _mm_add_pd(fix1,tx);
341 fiy1 = _mm_add_pd(fiy1,ty);
342 fiz1 = _mm_add_pd(fiz1,tz);
343
344 fjx3 = _mm_add_pd(fjx3,tx);
345 fjy3 = _mm_add_pd(fjy3,ty);
346 fjz3 = _mm_add_pd(fjz3,tz);
347
348 /**************************
349 * CALCULATE INTERACTIONS *
350 **************************/
351
352 /* REACTION-FIELD ELECTROSTATICS */
353 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
354 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),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,dx21);
363 ty = _mm_mul_pd(fscal,dy21);
364 tz = _mm_mul_pd(fscal,dz21);
365
366 /* Update vectorial force */
367 fix2 = _mm_add_pd(fix2,tx);
368 fiy2 = _mm_add_pd(fiy2,ty);
369 fiz2 = _mm_add_pd(fiz2,tz);
370
371 fjx1 = _mm_add_pd(fjx1,tx);
372 fjy1 = _mm_add_pd(fjy1,ty);
373 fjz1 = _mm_add_pd(fjz1,tz);
374
375 /**************************
376 * CALCULATE INTERACTIONS *
377 **************************/
378
379 /* REACTION-FIELD ELECTROSTATICS */
380 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
381 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),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,dx22);
390 ty = _mm_mul_pd(fscal,dy22);
391 tz = _mm_mul_pd(fscal,dz22);
392
393 /* Update vectorial force */
394 fix2 = _mm_add_pd(fix2,tx);
395 fiy2 = _mm_add_pd(fiy2,ty);
396 fiz2 = _mm_add_pd(fiz2,tz);
397
398 fjx2 = _mm_add_pd(fjx2,tx);
399 fjy2 = _mm_add_pd(fjy2,ty);
400 fjz2 = _mm_add_pd(fjz2,tz);
401
402 /**************************
403 * CALCULATE INTERACTIONS *
404 **************************/
405
406 /* REACTION-FIELD ELECTROSTATICS */
407 velec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
408 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),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,dx23);
417 ty = _mm_mul_pd(fscal,dy23);
418 tz = _mm_mul_pd(fscal,dz23);
419
420 /* Update vectorial force */
421 fix2 = _mm_add_pd(fix2,tx);
422 fiy2 = _mm_add_pd(fiy2,ty);
423 fiz2 = _mm_add_pd(fiz2,tz);
424
425 fjx3 = _mm_add_pd(fjx3,tx);
426 fjy3 = _mm_add_pd(fjy3,ty);
427 fjz3 = _mm_add_pd(fjz3,tz);
428
429 /**************************
430 * CALCULATE INTERACTIONS *
431 **************************/
432
433 /* REACTION-FIELD ELECTROSTATICS */
434 velec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
435 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),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,dx31);
444 ty = _mm_mul_pd(fscal,dy31);
445 tz = _mm_mul_pd(fscal,dz31);
446
447 /* Update vectorial force */
448 fix3 = _mm_add_pd(fix3,tx);
449 fiy3 = _mm_add_pd(fiy3,ty);
450 fiz3 = _mm_add_pd(fiz3,tz);
451
452 fjx1 = _mm_add_pd(fjx1,tx);
453 fjy1 = _mm_add_pd(fjy1,ty);
454 fjz1 = _mm_add_pd(fjz1,tz);
455
456 /**************************
457 * CALCULATE INTERACTIONS *
458 **************************/
459
460 /* REACTION-FIELD ELECTROSTATICS */
461 velec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
462 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),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,dx32);
471 ty = _mm_mul_pd(fscal,dy32);
472 tz = _mm_mul_pd(fscal,dz32);
473
474 /* Update vectorial force */
475 fix3 = _mm_add_pd(fix3,tx);
476 fiy3 = _mm_add_pd(fiy3,ty);
477 fiz3 = _mm_add_pd(fiz3,tz);
478
479 fjx2 = _mm_add_pd(fjx2,tx);
480 fjy2 = _mm_add_pd(fjy2,ty);
481 fjz2 = _mm_add_pd(fjz2,tz);
482
483 /**************************
484 * CALCULATE INTERACTIONS *
485 **************************/
486
487 /* REACTION-FIELD ELECTROSTATICS */
488 velec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
489 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),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,dx33);
498 ty = _mm_mul_pd(fscal,dy33);
499 tz = _mm_mul_pd(fscal,dz33);
500
501 /* Update vectorial force */
502 fix3 = _mm_add_pd(fix3,tx);
503 fiy3 = _mm_add_pd(fiy3,ty);
504 fiz3 = _mm_add_pd(fiz3,tz);
505
506 fjx3 = _mm_add_pd(fjx3,tx);
507 fjy3 = _mm_add_pd(fjy3,ty);
508 fjz3 = _mm_add_pd(fjz3,tz);
509
510 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
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+DIM,
523 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
524
525 /* Calculate displacement vector */
526 dx11 = _mm_sub_pd(ix1,jx1);
527 dy11 = _mm_sub_pd(iy1,jy1);
528 dz11 = _mm_sub_pd(iz1,jz1);
529 dx12 = _mm_sub_pd(ix1,jx2);
530 dy12 = _mm_sub_pd(iy1,jy2);
531 dz12 = _mm_sub_pd(iz1,jz2);
532 dx13 = _mm_sub_pd(ix1,jx3);
533 dy13 = _mm_sub_pd(iy1,jy3);
534 dz13 = _mm_sub_pd(iz1,jz3);
535 dx21 = _mm_sub_pd(ix2,jx1);
536 dy21 = _mm_sub_pd(iy2,jy1);
537 dz21 = _mm_sub_pd(iz2,jz1);
538 dx22 = _mm_sub_pd(ix2,jx2);
539 dy22 = _mm_sub_pd(iy2,jy2);
540 dz22 = _mm_sub_pd(iz2,jz2);
541 dx23 = _mm_sub_pd(ix2,jx3);
542 dy23 = _mm_sub_pd(iy2,jy3);
543 dz23 = _mm_sub_pd(iz2,jz3);
544 dx31 = _mm_sub_pd(ix3,jx1);
545 dy31 = _mm_sub_pd(iy3,jy1);
546 dz31 = _mm_sub_pd(iz3,jz1);
547 dx32 = _mm_sub_pd(ix3,jx2);
548 dy32 = _mm_sub_pd(iy3,jy2);
549 dz32 = _mm_sub_pd(iz3,jz2);
550 dx33 = _mm_sub_pd(ix3,jx3);
551 dy33 = _mm_sub_pd(iy3,jy3);
552 dz33 = _mm_sub_pd(iz3,jz3);
553
554 /* Calculate squared distance and things based on it */
555 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
556 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
557 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
558 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
559 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
560 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
561 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
562 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
563 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
564
565 rinv11 = gmx_mm_invsqrt_pd(rsq11);
566 rinv12 = gmx_mm_invsqrt_pd(rsq12);
567 rinv13 = gmx_mm_invsqrt_pd(rsq13);
568 rinv21 = gmx_mm_invsqrt_pd(rsq21);
569 rinv22 = gmx_mm_invsqrt_pd(rsq22);
570 rinv23 = gmx_mm_invsqrt_pd(rsq23);
571 rinv31 = gmx_mm_invsqrt_pd(rsq31);
572 rinv32 = gmx_mm_invsqrt_pd(rsq32);
573 rinv33 = gmx_mm_invsqrt_pd(rsq33);
574
575 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
576 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
577 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
578 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
579 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
580 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
581 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
582 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
583 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
584
585 fjx1 = _mm_setzero_pd();
586 fjy1 = _mm_setzero_pd();
587 fjz1 = _mm_setzero_pd();
588 fjx2 = _mm_setzero_pd();
589 fjy2 = _mm_setzero_pd();
590 fjz2 = _mm_setzero_pd();
591 fjx3 = _mm_setzero_pd();
592 fjy3 = _mm_setzero_pd();
593 fjz3 = _mm_setzero_pd();
594
595 /**************************
596 * CALCULATE INTERACTIONS *
597 **************************/
598
599 /* REACTION-FIELD ELECTROSTATICS */
600 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
601 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),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,dx11);
613 ty = _mm_mul_pd(fscal,dy11);
614 tz = _mm_mul_pd(fscal,dz11);
615
616 /* Update vectorial force */
617 fix1 = _mm_add_pd(fix1,tx);
618 fiy1 = _mm_add_pd(fiy1,ty);
619 fiz1 = _mm_add_pd(fiz1,tz);
620
621 fjx1 = _mm_add_pd(fjx1,tx);
622 fjy1 = _mm_add_pd(fjy1,ty);
623 fjz1 = _mm_add_pd(fjz1,tz);
624
625 /**************************
626 * CALCULATE INTERACTIONS *
627 **************************/
628
629 /* REACTION-FIELD ELECTROSTATICS */
630 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
631 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),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,dx12);
643 ty = _mm_mul_pd(fscal,dy12);
644 tz = _mm_mul_pd(fscal,dz12);
645
646 /* Update vectorial force */
647 fix1 = _mm_add_pd(fix1,tx);
648 fiy1 = _mm_add_pd(fiy1,ty);
649 fiz1 = _mm_add_pd(fiz1,tz);
650
651 fjx2 = _mm_add_pd(fjx2,tx);
652 fjy2 = _mm_add_pd(fjy2,ty);
653 fjz2 = _mm_add_pd(fjz2,tz);
654
655 /**************************
656 * CALCULATE INTERACTIONS *
657 **************************/
658
659 /* REACTION-FIELD ELECTROSTATICS */
660 velec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
661 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),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,dx13);
673 ty = _mm_mul_pd(fscal,dy13);
674 tz = _mm_mul_pd(fscal,dz13);
675
676 /* Update vectorial force */
677 fix1 = _mm_add_pd(fix1,tx);
678 fiy1 = _mm_add_pd(fiy1,ty);
679 fiz1 = _mm_add_pd(fiz1,tz);
680
681 fjx3 = _mm_add_pd(fjx3,tx);
682 fjy3 = _mm_add_pd(fjy3,ty);
683 fjz3 = _mm_add_pd(fjz3,tz);
684
685 /**************************
686 * CALCULATE INTERACTIONS *
687 **************************/
688
689 /* REACTION-FIELD ELECTROSTATICS */
690 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
691 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),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,dx21);
703 ty = _mm_mul_pd(fscal,dy21);
704 tz = _mm_mul_pd(fscal,dz21);
705
706 /* Update vectorial force */
707 fix2 = _mm_add_pd(fix2,tx);
708 fiy2 = _mm_add_pd(fiy2,ty);
709 fiz2 = _mm_add_pd(fiz2,tz);
710
711 fjx1 = _mm_add_pd(fjx1,tx);
712 fjy1 = _mm_add_pd(fjy1,ty);
713 fjz1 = _mm_add_pd(fjz1,tz);
714
715 /**************************
716 * CALCULATE INTERACTIONS *
717 **************************/
718
719 /* REACTION-FIELD ELECTROSTATICS */
720 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
721 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),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,dx22);
733 ty = _mm_mul_pd(fscal,dy22);
734 tz = _mm_mul_pd(fscal,dz22);
735
736 /* Update vectorial force */
737 fix2 = _mm_add_pd(fix2,tx);
738 fiy2 = _mm_add_pd(fiy2,ty);
739 fiz2 = _mm_add_pd(fiz2,tz);
740
741 fjx2 = _mm_add_pd(fjx2,tx);
742 fjy2 = _mm_add_pd(fjy2,ty);
743 fjz2 = _mm_add_pd(fjz2,tz);
744
745 /**************************
746 * CALCULATE INTERACTIONS *
747 **************************/
748
749 /* REACTION-FIELD ELECTROSTATICS */
750 velec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
751 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),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,dx23);
763 ty = _mm_mul_pd(fscal,dy23);
764 tz = _mm_mul_pd(fscal,dz23);
765
766 /* Update vectorial force */
767 fix2 = _mm_add_pd(fix2,tx);
768 fiy2 = _mm_add_pd(fiy2,ty);
769 fiz2 = _mm_add_pd(fiz2,tz);
770
771 fjx3 = _mm_add_pd(fjx3,tx);
772 fjy3 = _mm_add_pd(fjy3,ty);
773 fjz3 = _mm_add_pd(fjz3,tz);
774
775 /**************************
776 * CALCULATE INTERACTIONS *
777 **************************/
778
779 /* REACTION-FIELD ELECTROSTATICS */
780 velec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
781 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),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,dx31);
793 ty = _mm_mul_pd(fscal,dy31);
794 tz = _mm_mul_pd(fscal,dz31);
795
796 /* Update vectorial force */
797 fix3 = _mm_add_pd(fix3,tx);
798 fiy3 = _mm_add_pd(fiy3,ty);
799 fiz3 = _mm_add_pd(fiz3,tz);
800
801 fjx1 = _mm_add_pd(fjx1,tx);
802 fjy1 = _mm_add_pd(fjy1,ty);
803 fjz1 = _mm_add_pd(fjz1,tz);
804
805 /**************************
806 * CALCULATE INTERACTIONS *
807 **************************/
808
809 /* REACTION-FIELD ELECTROSTATICS */
810 velec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
811 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),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,dx32);
823 ty = _mm_mul_pd(fscal,dy32);
824 tz = _mm_mul_pd(fscal,dz32);
825
826 /* Update vectorial force */
827 fix3 = _mm_add_pd(fix3,tx);
828 fiy3 = _mm_add_pd(fiy3,ty);
829 fiz3 = _mm_add_pd(fiz3,tz);
830
831 fjx2 = _mm_add_pd(fjx2,tx);
832 fjy2 = _mm_add_pd(fjy2,ty);
833 fjz2 = _mm_add_pd(fjz2,tz);
834
835 /**************************
836 * CALCULATE INTERACTIONS *
837 **************************/
838
839 /* REACTION-FIELD ELECTROSTATICS */
840 velec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
841 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),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,dx33);
853 ty = _mm_mul_pd(fscal,dy33);
854 tz = _mm_mul_pd(fscal,dz33);
855
856 /* Update vectorial force */
857 fix3 = _mm_add_pd(fix3,tx);
858 fiy3 = _mm_add_pd(fiy3,ty);
859 fiz3 = _mm_add_pd(fiz3,tz);
860
861 fjx3 = _mm_add_pd(fjx3,tx);
862 fjy3 = _mm_add_pd(fjy3,ty);
863 fjz3 = _mm_add_pd(fjz3,tz);
864
865 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
866
867 /* Inner loop uses 288 flops */
868 }
869
870 /* End of innermost loop */
871
872 gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
873 f+i_coord_offset+DIM,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_W4W4_VF,outeriter*19 + inneriter*288);
891 }
892 /*
893 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sse2_double
894 * Electrostatics interaction: ReactionField
895 * VdW interaction: None
896 * Geometry: Water4-Water4
897 * Calculate force/pot: Force
898 */
899 void
900 nb_kernel_ElecRF_VdwNone_GeomW4W4_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 vdwioffset1;
923 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
924 int vdwioffset2;
925 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
926 int vdwioffset3;
927 __m128d ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
928 int vdwjidx1A,vdwjidx1B;
929 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
930 int vdwjidx2A,vdwjidx2B;
931 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
932 int vdwjidx3A,vdwjidx3B;
933 __m128d jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
934 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
935 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
936 __m128d dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
937 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
938 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
939 __m128d dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
940 __m128d dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
941 __m128d dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
942 __m128d dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
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 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
969 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
970 iq3 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
971
972 jq1 = _mm_set1_pd(charge[inr+1]);
973 jq2 = _mm_set1_pd(charge[inr+2]);
974 jq3 = _mm_set1_pd(charge[inr+3]);
975 qq11 = _mm_mul_pd(iq1,jq1);
976 qq12 = _mm_mul_pd(iq1,jq2);
977 qq13 = _mm_mul_pd(iq1,jq3);
978 qq21 = _mm_mul_pd(iq2,jq1);
979 qq22 = _mm_mul_pd(iq2,jq2);
980 qq23 = _mm_mul_pd(iq2,jq3);
981 qq31 = _mm_mul_pd(iq3,jq1);
982 qq32 = _mm_mul_pd(iq3,jq2);
983 qq33 = _mm_mul_pd(iq3,jq3);
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+DIM,
1009 &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1010
1011 fix1 = _mm_setzero_pd();
1012 fiy1 = _mm_setzero_pd();
1013 fiz1 = _mm_setzero_pd();
1014 fix2 = _mm_setzero_pd();
1015 fiy2 = _mm_setzero_pd();
1016 fiz2 = _mm_setzero_pd();
1017 fix3 = _mm_setzero_pd();
1018 fiy3 = _mm_setzero_pd();
1019 fiz3 = _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+DIM,x+j_coord_offsetB+DIM,
1033 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1034
1035 /* Calculate displacement vector */
1036 dx11 = _mm_sub_pd(ix1,jx1);
1037 dy11 = _mm_sub_pd(iy1,jy1);
1038 dz11 = _mm_sub_pd(iz1,jz1);
1039 dx12 = _mm_sub_pd(ix1,jx2);
1040 dy12 = _mm_sub_pd(iy1,jy2);
1041 dz12 = _mm_sub_pd(iz1,jz2);
1042 dx13 = _mm_sub_pd(ix1,jx3);
1043 dy13 = _mm_sub_pd(iy1,jy3);
1044 dz13 = _mm_sub_pd(iz1,jz3);
1045 dx21 = _mm_sub_pd(ix2,jx1);
1046 dy21 = _mm_sub_pd(iy2,jy1);
1047 dz21 = _mm_sub_pd(iz2,jz1);
1048 dx22 = _mm_sub_pd(ix2,jx2);
1049 dy22 = _mm_sub_pd(iy2,jy2);
1050 dz22 = _mm_sub_pd(iz2,jz2);
1051 dx23 = _mm_sub_pd(ix2,jx3);
1052 dy23 = _mm_sub_pd(iy2,jy3);
1053 dz23 = _mm_sub_pd(iz2,jz3);
1054 dx31 = _mm_sub_pd(ix3,jx1);
1055 dy31 = _mm_sub_pd(iy3,jy1);
1056 dz31 = _mm_sub_pd(iz3,jz1);
1057 dx32 = _mm_sub_pd(ix3,jx2);
1058 dy32 = _mm_sub_pd(iy3,jy2);
1059 dz32 = _mm_sub_pd(iz3,jz2);
1060 dx33 = _mm_sub_pd(ix3,jx3);
1061 dy33 = _mm_sub_pd(iy3,jy3);
1062 dz33 = _mm_sub_pd(iz3,jz3);
1063
1064 /* Calculate squared distance and things based on it */
1065 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1066 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1067 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1068 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1069 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1070 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1071 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1072 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1073 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1074
1075 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1076 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1077 rinv13 = gmx_mm_invsqrt_pd(rsq13);
1078 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1079 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1080 rinv23 = gmx_mm_invsqrt_pd(rsq23);
1081 rinv31 = gmx_mm_invsqrt_pd(rsq31);
1082 rinv32 = gmx_mm_invsqrt_pd(rsq32);
1083 rinv33 = gmx_mm_invsqrt_pd(rsq33);
1084
1085 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1086 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1087 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
1088 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1089 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1090 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
1091 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
1092 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
1093 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
1094
1095 fjx1 = _mm_setzero_pd();
1096 fjy1 = _mm_setzero_pd();
1097 fjz1 = _mm_setzero_pd();
1098 fjx2 = _mm_setzero_pd();
1099 fjy2 = _mm_setzero_pd();
1100 fjz2 = _mm_setzero_pd();
1101 fjx3 = _mm_setzero_pd();
1102 fjy3 = _mm_setzero_pd();
1103 fjz3 = _mm_setzero_pd();
1104
1105 /**************************
1106 * CALCULATE INTERACTIONS *
1107 **************************/
1108
1109 /* REACTION-FIELD ELECTROSTATICS */
1110 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1111
1112 fscal = felec;
1113
1114 /* Calculate temporary vectorial force */
1115 tx = _mm_mul_pd(fscal,dx11);
1116 ty = _mm_mul_pd(fscal,dy11);
1117 tz = _mm_mul_pd(fscal,dz11);
1118
1119 /* Update vectorial force */
1120 fix1 = _mm_add_pd(fix1,tx);
1121 fiy1 = _mm_add_pd(fiy1,ty);
1122 fiz1 = _mm_add_pd(fiz1,tz);
1123
1124 fjx1 = _mm_add_pd(fjx1,tx);
1125 fjy1 = _mm_add_pd(fjy1,ty);
1126 fjz1 = _mm_add_pd(fjz1,tz);
1127
1128 /**************************
1129 * CALCULATE INTERACTIONS *
1130 **************************/
1131
1132 /* REACTION-FIELD ELECTROSTATICS */
1133 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1134
1135 fscal = felec;
1136
1137 /* Calculate temporary vectorial force */
1138 tx = _mm_mul_pd(fscal,dx12);
1139 ty = _mm_mul_pd(fscal,dy12);
1140 tz = _mm_mul_pd(fscal,dz12);
1141
1142 /* Update vectorial force */
1143 fix1 = _mm_add_pd(fix1,tx);
1144 fiy1 = _mm_add_pd(fiy1,ty);
1145 fiz1 = _mm_add_pd(fiz1,tz);
1146
1147 fjx2 = _mm_add_pd(fjx2,tx);
1148 fjy2 = _mm_add_pd(fjy2,ty);
1149 fjz2 = _mm_add_pd(fjz2,tz);
1150
1151 /**************************
1152 * CALCULATE INTERACTIONS *
1153 **************************/
1154
1155 /* REACTION-FIELD ELECTROSTATICS */
1156 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1157
1158 fscal = felec;
1159
1160 /* Calculate temporary vectorial force */
1161 tx = _mm_mul_pd(fscal,dx13);
1162 ty = _mm_mul_pd(fscal,dy13);
1163 tz = _mm_mul_pd(fscal,dz13);
1164
1165 /* Update vectorial force */
1166 fix1 = _mm_add_pd(fix1,tx);
1167 fiy1 = _mm_add_pd(fiy1,ty);
1168 fiz1 = _mm_add_pd(fiz1,tz);
1169
1170 fjx3 = _mm_add_pd(fjx3,tx);
1171 fjy3 = _mm_add_pd(fjy3,ty);
1172 fjz3 = _mm_add_pd(fjz3,tz);
1173
1174 /**************************
1175 * CALCULATE INTERACTIONS *
1176 **************************/
1177
1178 /* REACTION-FIELD ELECTROSTATICS */
1179 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1180
1181 fscal = felec;
1182
1183 /* Calculate temporary vectorial force */
1184 tx = _mm_mul_pd(fscal,dx21);
1185 ty = _mm_mul_pd(fscal,dy21);
1186 tz = _mm_mul_pd(fscal,dz21);
1187
1188 /* Update vectorial force */
1189 fix2 = _mm_add_pd(fix2,tx);
1190 fiy2 = _mm_add_pd(fiy2,ty);
1191 fiz2 = _mm_add_pd(fiz2,tz);
1192
1193 fjx1 = _mm_add_pd(fjx1,tx);
1194 fjy1 = _mm_add_pd(fjy1,ty);
1195 fjz1 = _mm_add_pd(fjz1,tz);
1196
1197 /**************************
1198 * CALCULATE INTERACTIONS *
1199 **************************/
1200
1201 /* REACTION-FIELD ELECTROSTATICS */
1202 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1203
1204 fscal = felec;
1205
1206 /* Calculate temporary vectorial force */
1207 tx = _mm_mul_pd(fscal,dx22);
1208 ty = _mm_mul_pd(fscal,dy22);
1209 tz = _mm_mul_pd(fscal,dz22);
1210
1211 /* Update vectorial force */
1212 fix2 = _mm_add_pd(fix2,tx);
1213 fiy2 = _mm_add_pd(fiy2,ty);
1214 fiz2 = _mm_add_pd(fiz2,tz);
1215
1216 fjx2 = _mm_add_pd(fjx2,tx);
1217 fjy2 = _mm_add_pd(fjy2,ty);
1218 fjz2 = _mm_add_pd(fjz2,tz);
1219
1220 /**************************
1221 * CALCULATE INTERACTIONS *
1222 **************************/
1223
1224 /* REACTION-FIELD ELECTROSTATICS */
1225 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1226
1227 fscal = felec;
1228
1229 /* Calculate temporary vectorial force */
1230 tx = _mm_mul_pd(fscal,dx23);
1231 ty = _mm_mul_pd(fscal,dy23);
1232 tz = _mm_mul_pd(fscal,dz23);
1233
1234 /* Update vectorial force */
1235 fix2 = _mm_add_pd(fix2,tx);
1236 fiy2 = _mm_add_pd(fiy2,ty);
1237 fiz2 = _mm_add_pd(fiz2,tz);
1238
1239 fjx3 = _mm_add_pd(fjx3,tx);
1240 fjy3 = _mm_add_pd(fjy3,ty);
1241 fjz3 = _mm_add_pd(fjz3,tz);
1242
1243 /**************************
1244 * CALCULATE INTERACTIONS *
1245 **************************/
1246
1247 /* REACTION-FIELD ELECTROSTATICS */
1248 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1249
1250 fscal = felec;
1251
1252 /* Calculate temporary vectorial force */
1253 tx = _mm_mul_pd(fscal,dx31);
1254 ty = _mm_mul_pd(fscal,dy31);
1255 tz = _mm_mul_pd(fscal,dz31);
1256
1257 /* Update vectorial force */
1258 fix3 = _mm_add_pd(fix3,tx);
1259 fiy3 = _mm_add_pd(fiy3,ty);
1260 fiz3 = _mm_add_pd(fiz3,tz);
1261
1262 fjx1 = _mm_add_pd(fjx1,tx);
1263 fjy1 = _mm_add_pd(fjy1,ty);
1264 fjz1 = _mm_add_pd(fjz1,tz);
1265
1266 /**************************
1267 * CALCULATE INTERACTIONS *
1268 **************************/
1269
1270 /* REACTION-FIELD ELECTROSTATICS */
1271 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1272
1273 fscal = felec;
1274
1275 /* Calculate temporary vectorial force */
1276 tx = _mm_mul_pd(fscal,dx32);
1277 ty = _mm_mul_pd(fscal,dy32);
1278 tz = _mm_mul_pd(fscal,dz32);
1279
1280 /* Update vectorial force */
1281 fix3 = _mm_add_pd(fix3,tx);
1282 fiy3 = _mm_add_pd(fiy3,ty);
1283 fiz3 = _mm_add_pd(fiz3,tz);
1284
1285 fjx2 = _mm_add_pd(fjx2,tx);
1286 fjy2 = _mm_add_pd(fjy2,ty);
1287 fjz2 = _mm_add_pd(fjz2,tz);
1288
1289 /**************************
1290 * CALCULATE INTERACTIONS *
1291 **************************/
1292
1293 /* REACTION-FIELD ELECTROSTATICS */
1294 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1295
1296 fscal = felec;
1297
1298 /* Calculate temporary vectorial force */
1299 tx = _mm_mul_pd(fscal,dx33);
1300 ty = _mm_mul_pd(fscal,dy33);
1301 tz = _mm_mul_pd(fscal,dz33);
1302
1303 /* Update vectorial force */
1304 fix3 = _mm_add_pd(fix3,tx);
1305 fiy3 = _mm_add_pd(fiy3,ty);
1306 fiz3 = _mm_add_pd(fiz3,tz);
1307
1308 fjx3 = _mm_add_pd(fjx3,tx);
1309 fjy3 = _mm_add_pd(fjy3,ty);
1310 fjz3 = _mm_add_pd(fjz3,tz);
1311
1312 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
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+DIM,
1325 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1326
1327 /* Calculate displacement vector */
1328 dx11 = _mm_sub_pd(ix1,jx1);
1329 dy11 = _mm_sub_pd(iy1,jy1);
1330 dz11 = _mm_sub_pd(iz1,jz1);
1331 dx12 = _mm_sub_pd(ix1,jx2);
1332 dy12 = _mm_sub_pd(iy1,jy2);
1333 dz12 = _mm_sub_pd(iz1,jz2);
1334 dx13 = _mm_sub_pd(ix1,jx3);
1335 dy13 = _mm_sub_pd(iy1,jy3);
1336 dz13 = _mm_sub_pd(iz1,jz3);
1337 dx21 = _mm_sub_pd(ix2,jx1);
1338 dy21 = _mm_sub_pd(iy2,jy1);
1339 dz21 = _mm_sub_pd(iz2,jz1);
1340 dx22 = _mm_sub_pd(ix2,jx2);
1341 dy22 = _mm_sub_pd(iy2,jy2);
1342 dz22 = _mm_sub_pd(iz2,jz2);
1343 dx23 = _mm_sub_pd(ix2,jx3);
1344 dy23 = _mm_sub_pd(iy2,jy3);
1345 dz23 = _mm_sub_pd(iz2,jz3);
1346 dx31 = _mm_sub_pd(ix3,jx1);
1347 dy31 = _mm_sub_pd(iy3,jy1);
1348 dz31 = _mm_sub_pd(iz3,jz1);
1349 dx32 = _mm_sub_pd(ix3,jx2);
1350 dy32 = _mm_sub_pd(iy3,jy2);
1351 dz32 = _mm_sub_pd(iz3,jz2);
1352 dx33 = _mm_sub_pd(ix3,jx3);
1353 dy33 = _mm_sub_pd(iy3,jy3);
1354 dz33 = _mm_sub_pd(iz3,jz3);
1355
1356 /* Calculate squared distance and things based on it */
1357 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1358 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1359 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1360 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1361 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1362 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1363 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1364 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1365 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1366
1367 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1368 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1369 rinv13 = gmx_mm_invsqrt_pd(rsq13);
1370 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1371 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1372 rinv23 = gmx_mm_invsqrt_pd(rsq23);
1373 rinv31 = gmx_mm_invsqrt_pd(rsq31);
1374 rinv32 = gmx_mm_invsqrt_pd(rsq32);
1375 rinv33 = gmx_mm_invsqrt_pd(rsq33);
1376
1377 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1378 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1379 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
1380 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1381 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1382 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
1383 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
1384 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
1385 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
1386
1387 fjx1 = _mm_setzero_pd();
1388 fjy1 = _mm_setzero_pd();
1389 fjz1 = _mm_setzero_pd();
1390 fjx2 = _mm_setzero_pd();
1391 fjy2 = _mm_setzero_pd();
1392 fjz2 = _mm_setzero_pd();
1393 fjx3 = _mm_setzero_pd();
1394 fjy3 = _mm_setzero_pd();
1395 fjz3 = _mm_setzero_pd();
1396
1397 /**************************
1398 * CALCULATE INTERACTIONS *
1399 **************************/
1400
1401 /* REACTION-FIELD ELECTROSTATICS */
1402 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),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,dx11);
1410 ty = _mm_mul_pd(fscal,dy11);
1411 tz = _mm_mul_pd(fscal,dz11);
1412
1413 /* Update vectorial force */
1414 fix1 = _mm_add_pd(fix1,tx);
1415 fiy1 = _mm_add_pd(fiy1,ty);
1416 fiz1 = _mm_add_pd(fiz1,tz);
1417
1418 fjx1 = _mm_add_pd(fjx1,tx);
1419 fjy1 = _mm_add_pd(fjy1,ty);
1420 fjz1 = _mm_add_pd(fjz1,tz);
1421
1422 /**************************
1423 * CALCULATE INTERACTIONS *
1424 **************************/
1425
1426 /* REACTION-FIELD ELECTROSTATICS */
1427 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),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,dx12);
1435 ty = _mm_mul_pd(fscal,dy12);
1436 tz = _mm_mul_pd(fscal,dz12);
1437
1438 /* Update vectorial force */
1439 fix1 = _mm_add_pd(fix1,tx);
1440 fiy1 = _mm_add_pd(fiy1,ty);
1441 fiz1 = _mm_add_pd(fiz1,tz);
1442
1443 fjx2 = _mm_add_pd(fjx2,tx);
1444 fjy2 = _mm_add_pd(fjy2,ty);
1445 fjz2 = _mm_add_pd(fjz2,tz);
1446
1447 /**************************
1448 * CALCULATE INTERACTIONS *
1449 **************************/
1450
1451 /* REACTION-FIELD ELECTROSTATICS */
1452 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),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,dx13);
1460 ty = _mm_mul_pd(fscal,dy13);
1461 tz = _mm_mul_pd(fscal,dz13);
1462
1463 /* Update vectorial force */
1464 fix1 = _mm_add_pd(fix1,tx);
1465 fiy1 = _mm_add_pd(fiy1,ty);
1466 fiz1 = _mm_add_pd(fiz1,tz);
1467
1468 fjx3 = _mm_add_pd(fjx3,tx);
1469 fjy3 = _mm_add_pd(fjy3,ty);
1470 fjz3 = _mm_add_pd(fjz3,tz);
1471
1472 /**************************
1473 * CALCULATE INTERACTIONS *
1474 **************************/
1475
1476 /* REACTION-FIELD ELECTROSTATICS */
1477 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),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,dx21);
1485 ty = _mm_mul_pd(fscal,dy21);
1486 tz = _mm_mul_pd(fscal,dz21);
1487
1488 /* Update vectorial force */
1489 fix2 = _mm_add_pd(fix2,tx);
1490 fiy2 = _mm_add_pd(fiy2,ty);
1491 fiz2 = _mm_add_pd(fiz2,tz);
1492
1493 fjx1 = _mm_add_pd(fjx1,tx);
1494 fjy1 = _mm_add_pd(fjy1,ty);
1495 fjz1 = _mm_add_pd(fjz1,tz);
1496
1497 /**************************
1498 * CALCULATE INTERACTIONS *
1499 **************************/
1500
1501 /* REACTION-FIELD ELECTROSTATICS */
1502 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),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,dx22);
1510 ty = _mm_mul_pd(fscal,dy22);
1511 tz = _mm_mul_pd(fscal,dz22);
1512
1513 /* Update vectorial force */
1514 fix2 = _mm_add_pd(fix2,tx);
1515 fiy2 = _mm_add_pd(fiy2,ty);
1516 fiz2 = _mm_add_pd(fiz2,tz);
1517
1518 fjx2 = _mm_add_pd(fjx2,tx);
1519 fjy2 = _mm_add_pd(fjy2,ty);
1520 fjz2 = _mm_add_pd(fjz2,tz);
1521
1522 /**************************
1523 * CALCULATE INTERACTIONS *
1524 **************************/
1525
1526 /* REACTION-FIELD ELECTROSTATICS */
1527 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),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,dx23);
1535 ty = _mm_mul_pd(fscal,dy23);
1536 tz = _mm_mul_pd(fscal,dz23);
1537
1538 /* Update vectorial force */
1539 fix2 = _mm_add_pd(fix2,tx);
1540 fiy2 = _mm_add_pd(fiy2,ty);
1541 fiz2 = _mm_add_pd(fiz2,tz);
1542
1543 fjx3 = _mm_add_pd(fjx3,tx);
1544 fjy3 = _mm_add_pd(fjy3,ty);
1545 fjz3 = _mm_add_pd(fjz3,tz);
1546
1547 /**************************
1548 * CALCULATE INTERACTIONS *
1549 **************************/
1550
1551 /* REACTION-FIELD ELECTROSTATICS */
1552 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),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,dx31);
1560 ty = _mm_mul_pd(fscal,dy31);
1561 tz = _mm_mul_pd(fscal,dz31);
1562
1563 /* Update vectorial force */
1564 fix3 = _mm_add_pd(fix3,tx);
1565 fiy3 = _mm_add_pd(fiy3,ty);
1566 fiz3 = _mm_add_pd(fiz3,tz);
1567
1568 fjx1 = _mm_add_pd(fjx1,tx);
1569 fjy1 = _mm_add_pd(fjy1,ty);
1570 fjz1 = _mm_add_pd(fjz1,tz);
1571
1572 /**************************
1573 * CALCULATE INTERACTIONS *
1574 **************************/
1575
1576 /* REACTION-FIELD ELECTROSTATICS */
1577 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),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,dx32);
1585 ty = _mm_mul_pd(fscal,dy32);
1586 tz = _mm_mul_pd(fscal,dz32);
1587
1588 /* Update vectorial force */
1589 fix3 = _mm_add_pd(fix3,tx);
1590 fiy3 = _mm_add_pd(fiy3,ty);
1591 fiz3 = _mm_add_pd(fiz3,tz);
1592
1593 fjx2 = _mm_add_pd(fjx2,tx);
1594 fjy2 = _mm_add_pd(fjy2,ty);
1595 fjz2 = _mm_add_pd(fjz2,tz);
1596
1597 /**************************
1598 * CALCULATE INTERACTIONS *
1599 **************************/
1600
1601 /* REACTION-FIELD ELECTROSTATICS */
1602 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),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,dx33);
1610 ty = _mm_mul_pd(fscal,dy33);
1611 tz = _mm_mul_pd(fscal,dz33);
1612
1613 /* Update vectorial force */
1614 fix3 = _mm_add_pd(fix3,tx);
1615 fiy3 = _mm_add_pd(fiy3,ty);
1616 fiz3 = _mm_add_pd(fiz3,tz);
1617
1618 fjx3 = _mm_add_pd(fjx3,tx);
1619 fjy3 = _mm_add_pd(fjy3,ty);
1620 fjz3 = _mm_add_pd(fjz3,tz);
1621
1622 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1623
1624 /* Inner loop uses 243 flops */
1625 }
1626
1627 /* End of innermost loop */
1628
1629 gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1630 f+i_coord_offset+DIM,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_W4W4_F,outeriter*18 + inneriter*243);
1644 }
The ultimate molecular dynamics simulation package