?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Removed simple.h from nb_kernel_sse4_1_XX
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse4_1_double / nb_kernel_ElecRFCut_VdwNone_GeomW4W4_sse4_1_double.c
blob434b89129be404b2fd952f1f46a0739071b31eed
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 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/math/vec.h"
46 #include "gromacs/legacyheaders/nrnb.h"
47
48 #include "gromacs/simd/math_x86_sse4_1_double.h"
49 #include "kernelutil_x86_sse4_1_double.h"
50
51 /*
52 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_double
53 * Electrostatics interaction: ReactionField
54 * VdW interaction: None
55 * Geometry: Water4-Water4
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_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 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
145 rcutoff_scalar = fr->rcoulomb;
146 rcutoff = _mm_set1_pd(rcutoff_scalar);
147 rcutoff2 = _mm_mul_pd(rcutoff,rcutoff);
148
149 /* Avoid stupid compiler warnings */
150 jnrA = jnrB = 0;
151 j_coord_offsetA = 0;
152 j_coord_offsetB = 0;
153
154 outeriter = 0;
155 inneriter = 0;
156
157 /* Start outer loop over neighborlists */
158 for(iidx=0; iidx<nri; iidx++)
159 {
160 /* Load shift vector for this list */
161 i_shift_offset = DIM*shiftidx[iidx];
162
163 /* Load limits for loop over neighbors */
164 j_index_start = jindex[iidx];
165 j_index_end = jindex[iidx+1];
166
167 /* Get outer coordinate index */
168 inr = iinr[iidx];
169 i_coord_offset = DIM*inr;
170
171 /* Load i particle coords and add shift vector */
172 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
173 &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
174
175 fix1 = _mm_setzero_pd();
176 fiy1 = _mm_setzero_pd();
177 fiz1 = _mm_setzero_pd();
178 fix2 = _mm_setzero_pd();
179 fiy2 = _mm_setzero_pd();
180 fiz2 = _mm_setzero_pd();
181 fix3 = _mm_setzero_pd();
182 fiy3 = _mm_setzero_pd();
183 fiz3 = _mm_setzero_pd();
184
185 /* Reset potential sums */
186 velecsum = _mm_setzero_pd();
187
188 /* Start inner kernel loop */
189 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
190 {
191
192 /* Get j neighbor index, and coordinate index */
193 jnrA = jjnr[jidx];
194 jnrB = jjnr[jidx+1];
195 j_coord_offsetA = DIM*jnrA;
196 j_coord_offsetB = DIM*jnrB;
197
198 /* load j atom coordinates */
199 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
200 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
201
202 /* Calculate displacement vector */
203 dx11 = _mm_sub_pd(ix1,jx1);
204 dy11 = _mm_sub_pd(iy1,jy1);
205 dz11 = _mm_sub_pd(iz1,jz1);
206 dx12 = _mm_sub_pd(ix1,jx2);
207 dy12 = _mm_sub_pd(iy1,jy2);
208 dz12 = _mm_sub_pd(iz1,jz2);
209 dx13 = _mm_sub_pd(ix1,jx3);
210 dy13 = _mm_sub_pd(iy1,jy3);
211 dz13 = _mm_sub_pd(iz1,jz3);
212 dx21 = _mm_sub_pd(ix2,jx1);
213 dy21 = _mm_sub_pd(iy2,jy1);
214 dz21 = _mm_sub_pd(iz2,jz1);
215 dx22 = _mm_sub_pd(ix2,jx2);
216 dy22 = _mm_sub_pd(iy2,jy2);
217 dz22 = _mm_sub_pd(iz2,jz2);
218 dx23 = _mm_sub_pd(ix2,jx3);
219 dy23 = _mm_sub_pd(iy2,jy3);
220 dz23 = _mm_sub_pd(iz2,jz3);
221 dx31 = _mm_sub_pd(ix3,jx1);
222 dy31 = _mm_sub_pd(iy3,jy1);
223 dz31 = _mm_sub_pd(iz3,jz1);
224 dx32 = _mm_sub_pd(ix3,jx2);
225 dy32 = _mm_sub_pd(iy3,jy2);
226 dz32 = _mm_sub_pd(iz3,jz2);
227 dx33 = _mm_sub_pd(ix3,jx3);
228 dy33 = _mm_sub_pd(iy3,jy3);
229 dz33 = _mm_sub_pd(iz3,jz3);
230
231 /* Calculate squared distance and things based on it */
232 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
233 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
234 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
235 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
236 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
237 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
238 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
239 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
240 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
241
242 rinv11 = gmx_mm_invsqrt_pd(rsq11);
243 rinv12 = gmx_mm_invsqrt_pd(rsq12);
244 rinv13 = gmx_mm_invsqrt_pd(rsq13);
245 rinv21 = gmx_mm_invsqrt_pd(rsq21);
246 rinv22 = gmx_mm_invsqrt_pd(rsq22);
247 rinv23 = gmx_mm_invsqrt_pd(rsq23);
248 rinv31 = gmx_mm_invsqrt_pd(rsq31);
249 rinv32 = gmx_mm_invsqrt_pd(rsq32);
250 rinv33 = gmx_mm_invsqrt_pd(rsq33);
251
252 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
253 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
254 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
255 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
256 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
257 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
258 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
259 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
260 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
261
262 fjx1 = _mm_setzero_pd();
263 fjy1 = _mm_setzero_pd();
264 fjz1 = _mm_setzero_pd();
265 fjx2 = _mm_setzero_pd();
266 fjy2 = _mm_setzero_pd();
267 fjz2 = _mm_setzero_pd();
268 fjx3 = _mm_setzero_pd();
269 fjy3 = _mm_setzero_pd();
270 fjz3 = _mm_setzero_pd();
271
272 /**************************
273 * CALCULATE INTERACTIONS *
274 **************************/
275
276 if (gmx_mm_any_lt(rsq11,rcutoff2))
277 {
278
279 /* REACTION-FIELD ELECTROSTATICS */
280 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
281 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
282
283 cutoff_mask = _mm_cmplt_pd(rsq11,rcutoff2);
284
285 /* Update potential sum for this i atom from the interaction with this j atom. */
286 velec = _mm_and_pd(velec,cutoff_mask);
287 velecsum = _mm_add_pd(velecsum,velec);
288
289 fscal = felec;
290
291 fscal = _mm_and_pd(fscal,cutoff_mask);
292
293 /* Calculate temporary vectorial force */
294 tx = _mm_mul_pd(fscal,dx11);
295 ty = _mm_mul_pd(fscal,dy11);
296 tz = _mm_mul_pd(fscal,dz11);
297
298 /* Update vectorial force */
299 fix1 = _mm_add_pd(fix1,tx);
300 fiy1 = _mm_add_pd(fiy1,ty);
301 fiz1 = _mm_add_pd(fiz1,tz);
302
303 fjx1 = _mm_add_pd(fjx1,tx);
304 fjy1 = _mm_add_pd(fjy1,ty);
305 fjz1 = _mm_add_pd(fjz1,tz);
306
307 }
308
309 /**************************
310 * CALCULATE INTERACTIONS *
311 **************************/
312
313 if (gmx_mm_any_lt(rsq12,rcutoff2))
314 {
315
316 /* REACTION-FIELD ELECTROSTATICS */
317 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
318 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
319
320 cutoff_mask = _mm_cmplt_pd(rsq12,rcutoff2);
321
322 /* Update potential sum for this i atom from the interaction with this j atom. */
323 velec = _mm_and_pd(velec,cutoff_mask);
324 velecsum = _mm_add_pd(velecsum,velec);
325
326 fscal = felec;
327
328 fscal = _mm_and_pd(fscal,cutoff_mask);
329
330 /* Calculate temporary vectorial force */
331 tx = _mm_mul_pd(fscal,dx12);
332 ty = _mm_mul_pd(fscal,dy12);
333 tz = _mm_mul_pd(fscal,dz12);
334
335 /* Update vectorial force */
336 fix1 = _mm_add_pd(fix1,tx);
337 fiy1 = _mm_add_pd(fiy1,ty);
338 fiz1 = _mm_add_pd(fiz1,tz);
339
340 fjx2 = _mm_add_pd(fjx2,tx);
341 fjy2 = _mm_add_pd(fjy2,ty);
342 fjz2 = _mm_add_pd(fjz2,tz);
343
344 }
345
346 /**************************
347 * CALCULATE INTERACTIONS *
348 **************************/
349
350 if (gmx_mm_any_lt(rsq13,rcutoff2))
351 {
352
353 /* REACTION-FIELD ELECTROSTATICS */
354 velec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
355 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
356
357 cutoff_mask = _mm_cmplt_pd(rsq13,rcutoff2);
358
359 /* Update potential sum for this i atom from the interaction with this j atom. */
360 velec = _mm_and_pd(velec,cutoff_mask);
361 velecsum = _mm_add_pd(velecsum,velec);
362
363 fscal = felec;
364
365 fscal = _mm_and_pd(fscal,cutoff_mask);
366
367 /* Calculate temporary vectorial force */
368 tx = _mm_mul_pd(fscal,dx13);
369 ty = _mm_mul_pd(fscal,dy13);
370 tz = _mm_mul_pd(fscal,dz13);
371
372 /* Update vectorial force */
373 fix1 = _mm_add_pd(fix1,tx);
374 fiy1 = _mm_add_pd(fiy1,ty);
375 fiz1 = _mm_add_pd(fiz1,tz);
376
377 fjx3 = _mm_add_pd(fjx3,tx);
378 fjy3 = _mm_add_pd(fjy3,ty);
379 fjz3 = _mm_add_pd(fjz3,tz);
380
381 }
382
383 /**************************
384 * CALCULATE INTERACTIONS *
385 **************************/
386
387 if (gmx_mm_any_lt(rsq21,rcutoff2))
388 {
389
390 /* REACTION-FIELD ELECTROSTATICS */
391 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
392 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
393
394 cutoff_mask = _mm_cmplt_pd(rsq21,rcutoff2);
395
396 /* Update potential sum for this i atom from the interaction with this j atom. */
397 velec = _mm_and_pd(velec,cutoff_mask);
398 velecsum = _mm_add_pd(velecsum,velec);
399
400 fscal = felec;
401
402 fscal = _mm_and_pd(fscal,cutoff_mask);
403
404 /* Calculate temporary vectorial force */
405 tx = _mm_mul_pd(fscal,dx21);
406 ty = _mm_mul_pd(fscal,dy21);
407 tz = _mm_mul_pd(fscal,dz21);
408
409 /* Update vectorial force */
410 fix2 = _mm_add_pd(fix2,tx);
411 fiy2 = _mm_add_pd(fiy2,ty);
412 fiz2 = _mm_add_pd(fiz2,tz);
413
414 fjx1 = _mm_add_pd(fjx1,tx);
415 fjy1 = _mm_add_pd(fjy1,ty);
416 fjz1 = _mm_add_pd(fjz1,tz);
417
418 }
419
420 /**************************
421 * CALCULATE INTERACTIONS *
422 **************************/
423
424 if (gmx_mm_any_lt(rsq22,rcutoff2))
425 {
426
427 /* REACTION-FIELD ELECTROSTATICS */
428 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
429 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
430
431 cutoff_mask = _mm_cmplt_pd(rsq22,rcutoff2);
432
433 /* Update potential sum for this i atom from the interaction with this j atom. */
434 velec = _mm_and_pd(velec,cutoff_mask);
435 velecsum = _mm_add_pd(velecsum,velec);
436
437 fscal = felec;
438
439 fscal = _mm_and_pd(fscal,cutoff_mask);
440
441 /* Calculate temporary vectorial force */
442 tx = _mm_mul_pd(fscal,dx22);
443 ty = _mm_mul_pd(fscal,dy22);
444 tz = _mm_mul_pd(fscal,dz22);
445
446 /* Update vectorial force */
447 fix2 = _mm_add_pd(fix2,tx);
448 fiy2 = _mm_add_pd(fiy2,ty);
449 fiz2 = _mm_add_pd(fiz2,tz);
450
451 fjx2 = _mm_add_pd(fjx2,tx);
452 fjy2 = _mm_add_pd(fjy2,ty);
453 fjz2 = _mm_add_pd(fjz2,tz);
454
455 }
456
457 /**************************
458 * CALCULATE INTERACTIONS *
459 **************************/
460
461 if (gmx_mm_any_lt(rsq23,rcutoff2))
462 {
463
464 /* REACTION-FIELD ELECTROSTATICS */
465 velec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
466 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
467
468 cutoff_mask = _mm_cmplt_pd(rsq23,rcutoff2);
469
470 /* Update potential sum for this i atom from the interaction with this j atom. */
471 velec = _mm_and_pd(velec,cutoff_mask);
472 velecsum = _mm_add_pd(velecsum,velec);
473
474 fscal = felec;
475
476 fscal = _mm_and_pd(fscal,cutoff_mask);
477
478 /* Calculate temporary vectorial force */
479 tx = _mm_mul_pd(fscal,dx23);
480 ty = _mm_mul_pd(fscal,dy23);
481 tz = _mm_mul_pd(fscal,dz23);
482
483 /* Update vectorial force */
484 fix2 = _mm_add_pd(fix2,tx);
485 fiy2 = _mm_add_pd(fiy2,ty);
486 fiz2 = _mm_add_pd(fiz2,tz);
487
488 fjx3 = _mm_add_pd(fjx3,tx);
489 fjy3 = _mm_add_pd(fjy3,ty);
490 fjz3 = _mm_add_pd(fjz3,tz);
491
492 }
493
494 /**************************
495 * CALCULATE INTERACTIONS *
496 **************************/
497
498 if (gmx_mm_any_lt(rsq31,rcutoff2))
499 {
500
501 /* REACTION-FIELD ELECTROSTATICS */
502 velec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
503 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
504
505 cutoff_mask = _mm_cmplt_pd(rsq31,rcutoff2);
506
507 /* Update potential sum for this i atom from the interaction with this j atom. */
508 velec = _mm_and_pd(velec,cutoff_mask);
509 velecsum = _mm_add_pd(velecsum,velec);
510
511 fscal = felec;
512
513 fscal = _mm_and_pd(fscal,cutoff_mask);
514
515 /* Calculate temporary vectorial force */
516 tx = _mm_mul_pd(fscal,dx31);
517 ty = _mm_mul_pd(fscal,dy31);
518 tz = _mm_mul_pd(fscal,dz31);
519
520 /* Update vectorial force */
521 fix3 = _mm_add_pd(fix3,tx);
522 fiy3 = _mm_add_pd(fiy3,ty);
523 fiz3 = _mm_add_pd(fiz3,tz);
524
525 fjx1 = _mm_add_pd(fjx1,tx);
526 fjy1 = _mm_add_pd(fjy1,ty);
527 fjz1 = _mm_add_pd(fjz1,tz);
528
529 }
530
531 /**************************
532 * CALCULATE INTERACTIONS *
533 **************************/
534
535 if (gmx_mm_any_lt(rsq32,rcutoff2))
536 {
537
538 /* REACTION-FIELD ELECTROSTATICS */
539 velec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
540 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
541
542 cutoff_mask = _mm_cmplt_pd(rsq32,rcutoff2);
543
544 /* Update potential sum for this i atom from the interaction with this j atom. */
545 velec = _mm_and_pd(velec,cutoff_mask);
546 velecsum = _mm_add_pd(velecsum,velec);
547
548 fscal = felec;
549
550 fscal = _mm_and_pd(fscal,cutoff_mask);
551
552 /* Calculate temporary vectorial force */
553 tx = _mm_mul_pd(fscal,dx32);
554 ty = _mm_mul_pd(fscal,dy32);
555 tz = _mm_mul_pd(fscal,dz32);
556
557 /* Update vectorial force */
558 fix3 = _mm_add_pd(fix3,tx);
559 fiy3 = _mm_add_pd(fiy3,ty);
560 fiz3 = _mm_add_pd(fiz3,tz);
561
562 fjx2 = _mm_add_pd(fjx2,tx);
563 fjy2 = _mm_add_pd(fjy2,ty);
564 fjz2 = _mm_add_pd(fjz2,tz);
565
566 }
567
568 /**************************
569 * CALCULATE INTERACTIONS *
570 **************************/
571
572 if (gmx_mm_any_lt(rsq33,rcutoff2))
573 {
574
575 /* REACTION-FIELD ELECTROSTATICS */
576 velec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
577 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
578
579 cutoff_mask = _mm_cmplt_pd(rsq33,rcutoff2);
580
581 /* Update potential sum for this i atom from the interaction with this j atom. */
582 velec = _mm_and_pd(velec,cutoff_mask);
583 velecsum = _mm_add_pd(velecsum,velec);
584
585 fscal = felec;
586
587 fscal = _mm_and_pd(fscal,cutoff_mask);
588
589 /* Calculate temporary vectorial force */
590 tx = _mm_mul_pd(fscal,dx33);
591 ty = _mm_mul_pd(fscal,dy33);
592 tz = _mm_mul_pd(fscal,dz33);
593
594 /* Update vectorial force */
595 fix3 = _mm_add_pd(fix3,tx);
596 fiy3 = _mm_add_pd(fiy3,ty);
597 fiz3 = _mm_add_pd(fiz3,tz);
598
599 fjx3 = _mm_add_pd(fjx3,tx);
600 fjy3 = _mm_add_pd(fjy3,ty);
601 fjz3 = _mm_add_pd(fjz3,tz);
602
603 }
604
605 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);
606
607 /* Inner loop uses 324 flops */
608 }
609
610 if(jidx<j_index_end)
611 {
612
613 jnrA = jjnr[jidx];
614 j_coord_offsetA = DIM*jnrA;
615
616 /* load j atom coordinates */
617 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
618 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
619
620 /* Calculate displacement vector */
621 dx11 = _mm_sub_pd(ix1,jx1);
622 dy11 = _mm_sub_pd(iy1,jy1);
623 dz11 = _mm_sub_pd(iz1,jz1);
624 dx12 = _mm_sub_pd(ix1,jx2);
625 dy12 = _mm_sub_pd(iy1,jy2);
626 dz12 = _mm_sub_pd(iz1,jz2);
627 dx13 = _mm_sub_pd(ix1,jx3);
628 dy13 = _mm_sub_pd(iy1,jy3);
629 dz13 = _mm_sub_pd(iz1,jz3);
630 dx21 = _mm_sub_pd(ix2,jx1);
631 dy21 = _mm_sub_pd(iy2,jy1);
632 dz21 = _mm_sub_pd(iz2,jz1);
633 dx22 = _mm_sub_pd(ix2,jx2);
634 dy22 = _mm_sub_pd(iy2,jy2);
635 dz22 = _mm_sub_pd(iz2,jz2);
636 dx23 = _mm_sub_pd(ix2,jx3);
637 dy23 = _mm_sub_pd(iy2,jy3);
638 dz23 = _mm_sub_pd(iz2,jz3);
639 dx31 = _mm_sub_pd(ix3,jx1);
640 dy31 = _mm_sub_pd(iy3,jy1);
641 dz31 = _mm_sub_pd(iz3,jz1);
642 dx32 = _mm_sub_pd(ix3,jx2);
643 dy32 = _mm_sub_pd(iy3,jy2);
644 dz32 = _mm_sub_pd(iz3,jz2);
645 dx33 = _mm_sub_pd(ix3,jx3);
646 dy33 = _mm_sub_pd(iy3,jy3);
647 dz33 = _mm_sub_pd(iz3,jz3);
648
649 /* Calculate squared distance and things based on it */
650 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
651 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
652 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
653 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
654 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
655 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
656 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
657 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
658 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
659
660 rinv11 = gmx_mm_invsqrt_pd(rsq11);
661 rinv12 = gmx_mm_invsqrt_pd(rsq12);
662 rinv13 = gmx_mm_invsqrt_pd(rsq13);
663 rinv21 = gmx_mm_invsqrt_pd(rsq21);
664 rinv22 = gmx_mm_invsqrt_pd(rsq22);
665 rinv23 = gmx_mm_invsqrt_pd(rsq23);
666 rinv31 = gmx_mm_invsqrt_pd(rsq31);
667 rinv32 = gmx_mm_invsqrt_pd(rsq32);
668 rinv33 = gmx_mm_invsqrt_pd(rsq33);
669
670 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
671 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
672 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
673 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
674 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
675 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
676 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
677 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
678 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
679
680 fjx1 = _mm_setzero_pd();
681 fjy1 = _mm_setzero_pd();
682 fjz1 = _mm_setzero_pd();
683 fjx2 = _mm_setzero_pd();
684 fjy2 = _mm_setzero_pd();
685 fjz2 = _mm_setzero_pd();
686 fjx3 = _mm_setzero_pd();
687 fjy3 = _mm_setzero_pd();
688 fjz3 = _mm_setzero_pd();
689
690 /**************************
691 * CALCULATE INTERACTIONS *
692 **************************/
693
694 if (gmx_mm_any_lt(rsq11,rcutoff2))
695 {
696
697 /* REACTION-FIELD ELECTROSTATICS */
698 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
699 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
700
701 cutoff_mask = _mm_cmplt_pd(rsq11,rcutoff2);
702
703 /* Update potential sum for this i atom from the interaction with this j atom. */
704 velec = _mm_and_pd(velec,cutoff_mask);
705 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
706 velecsum = _mm_add_pd(velecsum,velec);
707
708 fscal = felec;
709
710 fscal = _mm_and_pd(fscal,cutoff_mask);
711
712 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
713
714 /* Calculate temporary vectorial force */
715 tx = _mm_mul_pd(fscal,dx11);
716 ty = _mm_mul_pd(fscal,dy11);
717 tz = _mm_mul_pd(fscal,dz11);
718
719 /* Update vectorial force */
720 fix1 = _mm_add_pd(fix1,tx);
721 fiy1 = _mm_add_pd(fiy1,ty);
722 fiz1 = _mm_add_pd(fiz1,tz);
723
724 fjx1 = _mm_add_pd(fjx1,tx);
725 fjy1 = _mm_add_pd(fjy1,ty);
726 fjz1 = _mm_add_pd(fjz1,tz);
727
728 }
729
730 /**************************
731 * CALCULATE INTERACTIONS *
732 **************************/
733
734 if (gmx_mm_any_lt(rsq12,rcutoff2))
735 {
736
737 /* REACTION-FIELD ELECTROSTATICS */
738 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
739 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
740
741 cutoff_mask = _mm_cmplt_pd(rsq12,rcutoff2);
742
743 /* Update potential sum for this i atom from the interaction with this j atom. */
744 velec = _mm_and_pd(velec,cutoff_mask);
745 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
746 velecsum = _mm_add_pd(velecsum,velec);
747
748 fscal = felec;
749
750 fscal = _mm_and_pd(fscal,cutoff_mask);
751
752 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
753
754 /* Calculate temporary vectorial force */
755 tx = _mm_mul_pd(fscal,dx12);
756 ty = _mm_mul_pd(fscal,dy12);
757 tz = _mm_mul_pd(fscal,dz12);
758
759 /* Update vectorial force */
760 fix1 = _mm_add_pd(fix1,tx);
761 fiy1 = _mm_add_pd(fiy1,ty);
762 fiz1 = _mm_add_pd(fiz1,tz);
763
764 fjx2 = _mm_add_pd(fjx2,tx);
765 fjy2 = _mm_add_pd(fjy2,ty);
766 fjz2 = _mm_add_pd(fjz2,tz);
767
768 }
769
770 /**************************
771 * CALCULATE INTERACTIONS *
772 **************************/
773
774 if (gmx_mm_any_lt(rsq13,rcutoff2))
775 {
776
777 /* REACTION-FIELD ELECTROSTATICS */
778 velec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
779 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
780
781 cutoff_mask = _mm_cmplt_pd(rsq13,rcutoff2);
782
783 /* Update potential sum for this i atom from the interaction with this j atom. */
784 velec = _mm_and_pd(velec,cutoff_mask);
785 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
786 velecsum = _mm_add_pd(velecsum,velec);
787
788 fscal = felec;
789
790 fscal = _mm_and_pd(fscal,cutoff_mask);
791
792 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
793
794 /* Calculate temporary vectorial force */
795 tx = _mm_mul_pd(fscal,dx13);
796 ty = _mm_mul_pd(fscal,dy13);
797 tz = _mm_mul_pd(fscal,dz13);
798
799 /* Update vectorial force */
800 fix1 = _mm_add_pd(fix1,tx);
801 fiy1 = _mm_add_pd(fiy1,ty);
802 fiz1 = _mm_add_pd(fiz1,tz);
803
804 fjx3 = _mm_add_pd(fjx3,tx);
805 fjy3 = _mm_add_pd(fjy3,ty);
806 fjz3 = _mm_add_pd(fjz3,tz);
807
808 }
809
810 /**************************
811 * CALCULATE INTERACTIONS *
812 **************************/
813
814 if (gmx_mm_any_lt(rsq21,rcutoff2))
815 {
816
817 /* REACTION-FIELD ELECTROSTATICS */
818 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
819 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
820
821 cutoff_mask = _mm_cmplt_pd(rsq21,rcutoff2);
822
823 /* Update potential sum for this i atom from the interaction with this j atom. */
824 velec = _mm_and_pd(velec,cutoff_mask);
825 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
826 velecsum = _mm_add_pd(velecsum,velec);
827
828 fscal = felec;
829
830 fscal = _mm_and_pd(fscal,cutoff_mask);
831
832 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
833
834 /* Calculate temporary vectorial force */
835 tx = _mm_mul_pd(fscal,dx21);
836 ty = _mm_mul_pd(fscal,dy21);
837 tz = _mm_mul_pd(fscal,dz21);
838
839 /* Update vectorial force */
840 fix2 = _mm_add_pd(fix2,tx);
841 fiy2 = _mm_add_pd(fiy2,ty);
842 fiz2 = _mm_add_pd(fiz2,tz);
843
844 fjx1 = _mm_add_pd(fjx1,tx);
845 fjy1 = _mm_add_pd(fjy1,ty);
846 fjz1 = _mm_add_pd(fjz1,tz);
847
848 }
849
850 /**************************
851 * CALCULATE INTERACTIONS *
852 **************************/
853
854 if (gmx_mm_any_lt(rsq22,rcutoff2))
855 {
856
857 /* REACTION-FIELD ELECTROSTATICS */
858 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
859 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
860
861 cutoff_mask = _mm_cmplt_pd(rsq22,rcutoff2);
862
863 /* Update potential sum for this i atom from the interaction with this j atom. */
864 velec = _mm_and_pd(velec,cutoff_mask);
865 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
866 velecsum = _mm_add_pd(velecsum,velec);
867
868 fscal = felec;
869
870 fscal = _mm_and_pd(fscal,cutoff_mask);
871
872 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
873
874 /* Calculate temporary vectorial force */
875 tx = _mm_mul_pd(fscal,dx22);
876 ty = _mm_mul_pd(fscal,dy22);
877 tz = _mm_mul_pd(fscal,dz22);
878
879 /* Update vectorial force */
880 fix2 = _mm_add_pd(fix2,tx);
881 fiy2 = _mm_add_pd(fiy2,ty);
882 fiz2 = _mm_add_pd(fiz2,tz);
883
884 fjx2 = _mm_add_pd(fjx2,tx);
885 fjy2 = _mm_add_pd(fjy2,ty);
886 fjz2 = _mm_add_pd(fjz2,tz);
887
888 }
889
890 /**************************
891 * CALCULATE INTERACTIONS *
892 **************************/
893
894 if (gmx_mm_any_lt(rsq23,rcutoff2))
895 {
896
897 /* REACTION-FIELD ELECTROSTATICS */
898 velec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
899 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
900
901 cutoff_mask = _mm_cmplt_pd(rsq23,rcutoff2);
902
903 /* Update potential sum for this i atom from the interaction with this j atom. */
904 velec = _mm_and_pd(velec,cutoff_mask);
905 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
906 velecsum = _mm_add_pd(velecsum,velec);
907
908 fscal = felec;
909
910 fscal = _mm_and_pd(fscal,cutoff_mask);
911
912 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
913
914 /* Calculate temporary vectorial force */
915 tx = _mm_mul_pd(fscal,dx23);
916 ty = _mm_mul_pd(fscal,dy23);
917 tz = _mm_mul_pd(fscal,dz23);
918
919 /* Update vectorial force */
920 fix2 = _mm_add_pd(fix2,tx);
921 fiy2 = _mm_add_pd(fiy2,ty);
922 fiz2 = _mm_add_pd(fiz2,tz);
923
924 fjx3 = _mm_add_pd(fjx3,tx);
925 fjy3 = _mm_add_pd(fjy3,ty);
926 fjz3 = _mm_add_pd(fjz3,tz);
927
928 }
929
930 /**************************
931 * CALCULATE INTERACTIONS *
932 **************************/
933
934 if (gmx_mm_any_lt(rsq31,rcutoff2))
935 {
936
937 /* REACTION-FIELD ELECTROSTATICS */
938 velec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
939 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
940
941 cutoff_mask = _mm_cmplt_pd(rsq31,rcutoff2);
942
943 /* Update potential sum for this i atom from the interaction with this j atom. */
944 velec = _mm_and_pd(velec,cutoff_mask);
945 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
946 velecsum = _mm_add_pd(velecsum,velec);
947
948 fscal = felec;
949
950 fscal = _mm_and_pd(fscal,cutoff_mask);
951
952 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
953
954 /* Calculate temporary vectorial force */
955 tx = _mm_mul_pd(fscal,dx31);
956 ty = _mm_mul_pd(fscal,dy31);
957 tz = _mm_mul_pd(fscal,dz31);
958
959 /* Update vectorial force */
960 fix3 = _mm_add_pd(fix3,tx);
961 fiy3 = _mm_add_pd(fiy3,ty);
962 fiz3 = _mm_add_pd(fiz3,tz);
963
964 fjx1 = _mm_add_pd(fjx1,tx);
965 fjy1 = _mm_add_pd(fjy1,ty);
966 fjz1 = _mm_add_pd(fjz1,tz);
967
968 }
969
970 /**************************
971 * CALCULATE INTERACTIONS *
972 **************************/
973
974 if (gmx_mm_any_lt(rsq32,rcutoff2))
975 {
976
977 /* REACTION-FIELD ELECTROSTATICS */
978 velec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
979 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
980
981 cutoff_mask = _mm_cmplt_pd(rsq32,rcutoff2);
982
983 /* Update potential sum for this i atom from the interaction with this j atom. */
984 velec = _mm_and_pd(velec,cutoff_mask);
985 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
986 velecsum = _mm_add_pd(velecsum,velec);
987
988 fscal = felec;
989
990 fscal = _mm_and_pd(fscal,cutoff_mask);
991
992 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
993
994 /* Calculate temporary vectorial force */
995 tx = _mm_mul_pd(fscal,dx32);
996 ty = _mm_mul_pd(fscal,dy32);
997 tz = _mm_mul_pd(fscal,dz32);
998
999 /* Update vectorial force */
1000 fix3 = _mm_add_pd(fix3,tx);
1001 fiy3 = _mm_add_pd(fiy3,ty);
1002 fiz3 = _mm_add_pd(fiz3,tz);
1003
1004 fjx2 = _mm_add_pd(fjx2,tx);
1005 fjy2 = _mm_add_pd(fjy2,ty);
1006 fjz2 = _mm_add_pd(fjz2,tz);
1007
1008 }
1009
1010 /**************************
1011 * CALCULATE INTERACTIONS *
1012 **************************/
1013
1014 if (gmx_mm_any_lt(rsq33,rcutoff2))
1015 {
1016
1017 /* REACTION-FIELD ELECTROSTATICS */
1018 velec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1019 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1020
1021 cutoff_mask = _mm_cmplt_pd(rsq33,rcutoff2);
1022
1023 /* Update potential sum for this i atom from the interaction with this j atom. */
1024 velec = _mm_and_pd(velec,cutoff_mask);
1025 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1026 velecsum = _mm_add_pd(velecsum,velec);
1027
1028 fscal = felec;
1029
1030 fscal = _mm_and_pd(fscal,cutoff_mask);
1031
1032 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1033
1034 /* Calculate temporary vectorial force */
1035 tx = _mm_mul_pd(fscal,dx33);
1036 ty = _mm_mul_pd(fscal,dy33);
1037 tz = _mm_mul_pd(fscal,dz33);
1038
1039 /* Update vectorial force */
1040 fix3 = _mm_add_pd(fix3,tx);
1041 fiy3 = _mm_add_pd(fiy3,ty);
1042 fiz3 = _mm_add_pd(fiz3,tz);
1043
1044 fjx3 = _mm_add_pd(fjx3,tx);
1045 fjy3 = _mm_add_pd(fjy3,ty);
1046 fjz3 = _mm_add_pd(fjz3,tz);
1047
1048 }
1049
1050 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1051
1052 /* Inner loop uses 324 flops */
1053 }
1054
1055 /* End of innermost loop */
1056
1057 gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1058 f+i_coord_offset+DIM,fshift+i_shift_offset);
1059
1060 ggid = gid[iidx];
1061 /* Update potential energies */
1062 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1063
1064 /* Increment number of inner iterations */
1065 inneriter += j_index_end - j_index_start;
1066
1067 /* Outer loop uses 19 flops */
1068 }
1069
1070 /* Increment number of outer iterations */
1071 outeriter += nri;
1072
1073 /* Update outer/inner flops */
1074
1075 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*324);
1076 }
1077 /*
1078 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_double
1079 * Electrostatics interaction: ReactionField
1080 * VdW interaction: None
1081 * Geometry: Water4-Water4
1082 * Calculate force/pot: Force
1083 */
1084 void
1085 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_double
1086 (t_nblist * gmx_restrict nlist,
1087 rvec * gmx_restrict xx,
1088 rvec * gmx_restrict ff,
1089 t_forcerec * gmx_restrict fr,
1090 t_mdatoms * gmx_restrict mdatoms,
1091 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1092 t_nrnb * gmx_restrict nrnb)
1093 {
1094 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1095 * just 0 for non-waters.
1096 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1097 * jnr indices corresponding to data put in the four positions in the SIMD register.
1098 */
1099 int i_shift_offset,i_coord_offset,outeriter,inneriter;
1100 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1101 int jnrA,jnrB;
1102 int j_coord_offsetA,j_coord_offsetB;
1103 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
1104 real rcutoff_scalar;
1105 real *shiftvec,*fshift,*x,*f;
1106 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1107 int vdwioffset1;
1108 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1109 int vdwioffset2;
1110 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1111 int vdwioffset3;
1112 __m128d ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1113 int vdwjidx1A,vdwjidx1B;
1114 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1115 int vdwjidx2A,vdwjidx2B;
1116 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1117 int vdwjidx3A,vdwjidx3B;
1118 __m128d jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1119 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1120 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1121 __m128d dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1122 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1123 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1124 __m128d dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1125 __m128d dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1126 __m128d dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1127 __m128d dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1128 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
1129 real *charge;
1130 __m128d dummy_mask,cutoff_mask;
1131 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1132 __m128d one = _mm_set1_pd(1.0);
1133 __m128d two = _mm_set1_pd(2.0);
1134 x = xx[0];
1135 f = ff[0];
1136
1137 nri = nlist->nri;
1138 iinr = nlist->iinr;
1139 jindex = nlist->jindex;
1140 jjnr = nlist->jjnr;
1141 shiftidx = nlist->shift;
1142 gid = nlist->gid;
1143 shiftvec = fr->shift_vec[0];
1144 fshift = fr->fshift[0];
1145 facel = _mm_set1_pd(fr->epsfac);
1146 charge = mdatoms->chargeA;
1147 krf = _mm_set1_pd(fr->ic->k_rf);
1148 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
1149 crf = _mm_set1_pd(fr->ic->c_rf);
1150
1151 /* Setup water-specific parameters */
1152 inr = nlist->iinr[0];
1153 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1154 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1155 iq3 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1156
1157 jq1 = _mm_set1_pd(charge[inr+1]);
1158 jq2 = _mm_set1_pd(charge[inr+2]);
1159 jq3 = _mm_set1_pd(charge[inr+3]);
1160 qq11 = _mm_mul_pd(iq1,jq1);
1161 qq12 = _mm_mul_pd(iq1,jq2);
1162 qq13 = _mm_mul_pd(iq1,jq3);
1163 qq21 = _mm_mul_pd(iq2,jq1);
1164 qq22 = _mm_mul_pd(iq2,jq2);
1165 qq23 = _mm_mul_pd(iq2,jq3);
1166 qq31 = _mm_mul_pd(iq3,jq1);
1167 qq32 = _mm_mul_pd(iq3,jq2);
1168 qq33 = _mm_mul_pd(iq3,jq3);
1169
1170 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1171 rcutoff_scalar = fr->rcoulomb;
1172 rcutoff = _mm_set1_pd(rcutoff_scalar);
1173 rcutoff2 = _mm_mul_pd(rcutoff,rcutoff);
1174
1175 /* Avoid stupid compiler warnings */
1176 jnrA = jnrB = 0;
1177 j_coord_offsetA = 0;
1178 j_coord_offsetB = 0;
1179
1180 outeriter = 0;
1181 inneriter = 0;
1182
1183 /* Start outer loop over neighborlists */
1184 for(iidx=0; iidx<nri; iidx++)
1185 {
1186 /* Load shift vector for this list */
1187 i_shift_offset = DIM*shiftidx[iidx];
1188
1189 /* Load limits for loop over neighbors */
1190 j_index_start = jindex[iidx];
1191 j_index_end = jindex[iidx+1];
1192
1193 /* Get outer coordinate index */
1194 inr = iinr[iidx];
1195 i_coord_offset = DIM*inr;
1196
1197 /* Load i particle coords and add shift vector */
1198 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1199 &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1200
1201 fix1 = _mm_setzero_pd();
1202 fiy1 = _mm_setzero_pd();
1203 fiz1 = _mm_setzero_pd();
1204 fix2 = _mm_setzero_pd();
1205 fiy2 = _mm_setzero_pd();
1206 fiz2 = _mm_setzero_pd();
1207 fix3 = _mm_setzero_pd();
1208 fiy3 = _mm_setzero_pd();
1209 fiz3 = _mm_setzero_pd();
1210
1211 /* Start inner kernel loop */
1212 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1213 {
1214
1215 /* Get j neighbor index, and coordinate index */
1216 jnrA = jjnr[jidx];
1217 jnrB = jjnr[jidx+1];
1218 j_coord_offsetA = DIM*jnrA;
1219 j_coord_offsetB = DIM*jnrB;
1220
1221 /* load j atom coordinates */
1222 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1223 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1224
1225 /* Calculate displacement vector */
1226 dx11 = _mm_sub_pd(ix1,jx1);
1227 dy11 = _mm_sub_pd(iy1,jy1);
1228 dz11 = _mm_sub_pd(iz1,jz1);
1229 dx12 = _mm_sub_pd(ix1,jx2);
1230 dy12 = _mm_sub_pd(iy1,jy2);
1231 dz12 = _mm_sub_pd(iz1,jz2);
1232 dx13 = _mm_sub_pd(ix1,jx3);
1233 dy13 = _mm_sub_pd(iy1,jy3);
1234 dz13 = _mm_sub_pd(iz1,jz3);
1235 dx21 = _mm_sub_pd(ix2,jx1);
1236 dy21 = _mm_sub_pd(iy2,jy1);
1237 dz21 = _mm_sub_pd(iz2,jz1);
1238 dx22 = _mm_sub_pd(ix2,jx2);
1239 dy22 = _mm_sub_pd(iy2,jy2);
1240 dz22 = _mm_sub_pd(iz2,jz2);
1241 dx23 = _mm_sub_pd(ix2,jx3);
1242 dy23 = _mm_sub_pd(iy2,jy3);
1243 dz23 = _mm_sub_pd(iz2,jz3);
1244 dx31 = _mm_sub_pd(ix3,jx1);
1245 dy31 = _mm_sub_pd(iy3,jy1);
1246 dz31 = _mm_sub_pd(iz3,jz1);
1247 dx32 = _mm_sub_pd(ix3,jx2);
1248 dy32 = _mm_sub_pd(iy3,jy2);
1249 dz32 = _mm_sub_pd(iz3,jz2);
1250 dx33 = _mm_sub_pd(ix3,jx3);
1251 dy33 = _mm_sub_pd(iy3,jy3);
1252 dz33 = _mm_sub_pd(iz3,jz3);
1253
1254 /* Calculate squared distance and things based on it */
1255 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1256 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1257 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1258 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1259 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1260 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1261 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1262 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1263 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1264
1265 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1266 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1267 rinv13 = gmx_mm_invsqrt_pd(rsq13);
1268 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1269 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1270 rinv23 = gmx_mm_invsqrt_pd(rsq23);
1271 rinv31 = gmx_mm_invsqrt_pd(rsq31);
1272 rinv32 = gmx_mm_invsqrt_pd(rsq32);
1273 rinv33 = gmx_mm_invsqrt_pd(rsq33);
1274
1275 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1276 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1277 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
1278 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1279 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1280 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
1281 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
1282 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
1283 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
1284
1285 fjx1 = _mm_setzero_pd();
1286 fjy1 = _mm_setzero_pd();
1287 fjz1 = _mm_setzero_pd();
1288 fjx2 = _mm_setzero_pd();
1289 fjy2 = _mm_setzero_pd();
1290 fjz2 = _mm_setzero_pd();
1291 fjx3 = _mm_setzero_pd();
1292 fjy3 = _mm_setzero_pd();
1293 fjz3 = _mm_setzero_pd();
1294
1295 /**************************
1296 * CALCULATE INTERACTIONS *
1297 **************************/
1298
1299 if (gmx_mm_any_lt(rsq11,rcutoff2))
1300 {
1301
1302 /* REACTION-FIELD ELECTROSTATICS */
1303 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1304
1305 cutoff_mask = _mm_cmplt_pd(rsq11,rcutoff2);
1306
1307 fscal = felec;
1308
1309 fscal = _mm_and_pd(fscal,cutoff_mask);
1310
1311 /* Calculate temporary vectorial force */
1312 tx = _mm_mul_pd(fscal,dx11);
1313 ty = _mm_mul_pd(fscal,dy11);
1314 tz = _mm_mul_pd(fscal,dz11);
1315
1316 /* Update vectorial force */
1317 fix1 = _mm_add_pd(fix1,tx);
1318 fiy1 = _mm_add_pd(fiy1,ty);
1319 fiz1 = _mm_add_pd(fiz1,tz);
1320
1321 fjx1 = _mm_add_pd(fjx1,tx);
1322 fjy1 = _mm_add_pd(fjy1,ty);
1323 fjz1 = _mm_add_pd(fjz1,tz);
1324
1325 }
1326
1327 /**************************
1328 * CALCULATE INTERACTIONS *
1329 **************************/
1330
1331 if (gmx_mm_any_lt(rsq12,rcutoff2))
1332 {
1333
1334 /* REACTION-FIELD ELECTROSTATICS */
1335 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1336
1337 cutoff_mask = _mm_cmplt_pd(rsq12,rcutoff2);
1338
1339 fscal = felec;
1340
1341 fscal = _mm_and_pd(fscal,cutoff_mask);
1342
1343 /* Calculate temporary vectorial force */
1344 tx = _mm_mul_pd(fscal,dx12);
1345 ty = _mm_mul_pd(fscal,dy12);
1346 tz = _mm_mul_pd(fscal,dz12);
1347
1348 /* Update vectorial force */
1349 fix1 = _mm_add_pd(fix1,tx);
1350 fiy1 = _mm_add_pd(fiy1,ty);
1351 fiz1 = _mm_add_pd(fiz1,tz);
1352
1353 fjx2 = _mm_add_pd(fjx2,tx);
1354 fjy2 = _mm_add_pd(fjy2,ty);
1355 fjz2 = _mm_add_pd(fjz2,tz);
1356
1357 }
1358
1359 /**************************
1360 * CALCULATE INTERACTIONS *
1361 **************************/
1362
1363 if (gmx_mm_any_lt(rsq13,rcutoff2))
1364 {
1365
1366 /* REACTION-FIELD ELECTROSTATICS */
1367 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1368
1369 cutoff_mask = _mm_cmplt_pd(rsq13,rcutoff2);
1370
1371 fscal = felec;
1372
1373 fscal = _mm_and_pd(fscal,cutoff_mask);
1374
1375 /* Calculate temporary vectorial force */
1376 tx = _mm_mul_pd(fscal,dx13);
1377 ty = _mm_mul_pd(fscal,dy13);
1378 tz = _mm_mul_pd(fscal,dz13);
1379
1380 /* Update vectorial force */
1381 fix1 = _mm_add_pd(fix1,tx);
1382 fiy1 = _mm_add_pd(fiy1,ty);
1383 fiz1 = _mm_add_pd(fiz1,tz);
1384
1385 fjx3 = _mm_add_pd(fjx3,tx);
1386 fjy3 = _mm_add_pd(fjy3,ty);
1387 fjz3 = _mm_add_pd(fjz3,tz);
1388
1389 }
1390
1391 /**************************
1392 * CALCULATE INTERACTIONS *
1393 **************************/
1394
1395 if (gmx_mm_any_lt(rsq21,rcutoff2))
1396 {
1397
1398 /* REACTION-FIELD ELECTROSTATICS */
1399 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1400
1401 cutoff_mask = _mm_cmplt_pd(rsq21,rcutoff2);
1402
1403 fscal = felec;
1404
1405 fscal = _mm_and_pd(fscal,cutoff_mask);
1406
1407 /* Calculate temporary vectorial force */
1408 tx = _mm_mul_pd(fscal,dx21);
1409 ty = _mm_mul_pd(fscal,dy21);
1410 tz = _mm_mul_pd(fscal,dz21);
1411
1412 /* Update vectorial force */
1413 fix2 = _mm_add_pd(fix2,tx);
1414 fiy2 = _mm_add_pd(fiy2,ty);
1415 fiz2 = _mm_add_pd(fiz2,tz);
1416
1417 fjx1 = _mm_add_pd(fjx1,tx);
1418 fjy1 = _mm_add_pd(fjy1,ty);
1419 fjz1 = _mm_add_pd(fjz1,tz);
1420
1421 }
1422
1423 /**************************
1424 * CALCULATE INTERACTIONS *
1425 **************************/
1426
1427 if (gmx_mm_any_lt(rsq22,rcutoff2))
1428 {
1429
1430 /* REACTION-FIELD ELECTROSTATICS */
1431 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1432
1433 cutoff_mask = _mm_cmplt_pd(rsq22,rcutoff2);
1434
1435 fscal = felec;
1436
1437 fscal = _mm_and_pd(fscal,cutoff_mask);
1438
1439 /* Calculate temporary vectorial force */
1440 tx = _mm_mul_pd(fscal,dx22);
1441 ty = _mm_mul_pd(fscal,dy22);
1442 tz = _mm_mul_pd(fscal,dz22);
1443
1444 /* Update vectorial force */
1445 fix2 = _mm_add_pd(fix2,tx);
1446 fiy2 = _mm_add_pd(fiy2,ty);
1447 fiz2 = _mm_add_pd(fiz2,tz);
1448
1449 fjx2 = _mm_add_pd(fjx2,tx);
1450 fjy2 = _mm_add_pd(fjy2,ty);
1451 fjz2 = _mm_add_pd(fjz2,tz);
1452
1453 }
1454
1455 /**************************
1456 * CALCULATE INTERACTIONS *
1457 **************************/
1458
1459 if (gmx_mm_any_lt(rsq23,rcutoff2))
1460 {
1461
1462 /* REACTION-FIELD ELECTROSTATICS */
1463 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1464
1465 cutoff_mask = _mm_cmplt_pd(rsq23,rcutoff2);
1466
1467 fscal = felec;
1468
1469 fscal = _mm_and_pd(fscal,cutoff_mask);
1470
1471 /* Calculate temporary vectorial force */
1472 tx = _mm_mul_pd(fscal,dx23);
1473 ty = _mm_mul_pd(fscal,dy23);
1474 tz = _mm_mul_pd(fscal,dz23);
1475
1476 /* Update vectorial force */
1477 fix2 = _mm_add_pd(fix2,tx);
1478 fiy2 = _mm_add_pd(fiy2,ty);
1479 fiz2 = _mm_add_pd(fiz2,tz);
1480
1481 fjx3 = _mm_add_pd(fjx3,tx);
1482 fjy3 = _mm_add_pd(fjy3,ty);
1483 fjz3 = _mm_add_pd(fjz3,tz);
1484
1485 }
1486
1487 /**************************
1488 * CALCULATE INTERACTIONS *
1489 **************************/
1490
1491 if (gmx_mm_any_lt(rsq31,rcutoff2))
1492 {
1493
1494 /* REACTION-FIELD ELECTROSTATICS */
1495 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1496
1497 cutoff_mask = _mm_cmplt_pd(rsq31,rcutoff2);
1498
1499 fscal = felec;
1500
1501 fscal = _mm_and_pd(fscal,cutoff_mask);
1502
1503 /* Calculate temporary vectorial force */
1504 tx = _mm_mul_pd(fscal,dx31);
1505 ty = _mm_mul_pd(fscal,dy31);
1506 tz = _mm_mul_pd(fscal,dz31);
1507
1508 /* Update vectorial force */
1509 fix3 = _mm_add_pd(fix3,tx);
1510 fiy3 = _mm_add_pd(fiy3,ty);
1511 fiz3 = _mm_add_pd(fiz3,tz);
1512
1513 fjx1 = _mm_add_pd(fjx1,tx);
1514 fjy1 = _mm_add_pd(fjy1,ty);
1515 fjz1 = _mm_add_pd(fjz1,tz);
1516
1517 }
1518
1519 /**************************
1520 * CALCULATE INTERACTIONS *
1521 **************************/
1522
1523 if (gmx_mm_any_lt(rsq32,rcutoff2))
1524 {
1525
1526 /* REACTION-FIELD ELECTROSTATICS */
1527 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1528
1529 cutoff_mask = _mm_cmplt_pd(rsq32,rcutoff2);
1530
1531 fscal = felec;
1532
1533 fscal = _mm_and_pd(fscal,cutoff_mask);
1534
1535 /* Calculate temporary vectorial force */
1536 tx = _mm_mul_pd(fscal,dx32);
1537 ty = _mm_mul_pd(fscal,dy32);
1538 tz = _mm_mul_pd(fscal,dz32);
1539
1540 /* Update vectorial force */
1541 fix3 = _mm_add_pd(fix3,tx);
1542 fiy3 = _mm_add_pd(fiy3,ty);
1543 fiz3 = _mm_add_pd(fiz3,tz);
1544
1545 fjx2 = _mm_add_pd(fjx2,tx);
1546 fjy2 = _mm_add_pd(fjy2,ty);
1547 fjz2 = _mm_add_pd(fjz2,tz);
1548
1549 }
1550
1551 /**************************
1552 * CALCULATE INTERACTIONS *
1553 **************************/
1554
1555 if (gmx_mm_any_lt(rsq33,rcutoff2))
1556 {
1557
1558 /* REACTION-FIELD ELECTROSTATICS */
1559 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1560
1561 cutoff_mask = _mm_cmplt_pd(rsq33,rcutoff2);
1562
1563 fscal = felec;
1564
1565 fscal = _mm_and_pd(fscal,cutoff_mask);
1566
1567 /* Calculate temporary vectorial force */
1568 tx = _mm_mul_pd(fscal,dx33);
1569 ty = _mm_mul_pd(fscal,dy33);
1570 tz = _mm_mul_pd(fscal,dz33);
1571
1572 /* Update vectorial force */
1573 fix3 = _mm_add_pd(fix3,tx);
1574 fiy3 = _mm_add_pd(fiy3,ty);
1575 fiz3 = _mm_add_pd(fiz3,tz);
1576
1577 fjx3 = _mm_add_pd(fjx3,tx);
1578 fjy3 = _mm_add_pd(fjy3,ty);
1579 fjz3 = _mm_add_pd(fjz3,tz);
1580
1581 }
1582
1583 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);
1584
1585 /* Inner loop uses 270 flops */
1586 }
1587
1588 if(jidx<j_index_end)
1589 {
1590
1591 jnrA = jjnr[jidx];
1592 j_coord_offsetA = DIM*jnrA;
1593
1594 /* load j atom coordinates */
1595 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1596 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1597
1598 /* Calculate displacement vector */
1599 dx11 = _mm_sub_pd(ix1,jx1);
1600 dy11 = _mm_sub_pd(iy1,jy1);
1601 dz11 = _mm_sub_pd(iz1,jz1);
1602 dx12 = _mm_sub_pd(ix1,jx2);
1603 dy12 = _mm_sub_pd(iy1,jy2);
1604 dz12 = _mm_sub_pd(iz1,jz2);
1605 dx13 = _mm_sub_pd(ix1,jx3);
1606 dy13 = _mm_sub_pd(iy1,jy3);
1607 dz13 = _mm_sub_pd(iz1,jz3);
1608 dx21 = _mm_sub_pd(ix2,jx1);
1609 dy21 = _mm_sub_pd(iy2,jy1);
1610 dz21 = _mm_sub_pd(iz2,jz1);
1611 dx22 = _mm_sub_pd(ix2,jx2);
1612 dy22 = _mm_sub_pd(iy2,jy2);
1613 dz22 = _mm_sub_pd(iz2,jz2);
1614 dx23 = _mm_sub_pd(ix2,jx3);
1615 dy23 = _mm_sub_pd(iy2,jy3);
1616 dz23 = _mm_sub_pd(iz2,jz3);
1617 dx31 = _mm_sub_pd(ix3,jx1);
1618 dy31 = _mm_sub_pd(iy3,jy1);
1619 dz31 = _mm_sub_pd(iz3,jz1);
1620 dx32 = _mm_sub_pd(ix3,jx2);
1621 dy32 = _mm_sub_pd(iy3,jy2);
1622 dz32 = _mm_sub_pd(iz3,jz2);
1623 dx33 = _mm_sub_pd(ix3,jx3);
1624 dy33 = _mm_sub_pd(iy3,jy3);
1625 dz33 = _mm_sub_pd(iz3,jz3);
1626
1627 /* Calculate squared distance and things based on it */
1628 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1629 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1630 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1631 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1632 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1633 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1634 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1635 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1636 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1637
1638 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1639 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1640 rinv13 = gmx_mm_invsqrt_pd(rsq13);
1641 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1642 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1643 rinv23 = gmx_mm_invsqrt_pd(rsq23);
1644 rinv31 = gmx_mm_invsqrt_pd(rsq31);
1645 rinv32 = gmx_mm_invsqrt_pd(rsq32);
1646 rinv33 = gmx_mm_invsqrt_pd(rsq33);
1647
1648 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1649 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1650 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
1651 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1652 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1653 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
1654 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
1655 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
1656 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
1657
1658 fjx1 = _mm_setzero_pd();
1659 fjy1 = _mm_setzero_pd();
1660 fjz1 = _mm_setzero_pd();
1661 fjx2 = _mm_setzero_pd();
1662 fjy2 = _mm_setzero_pd();
1663 fjz2 = _mm_setzero_pd();
1664 fjx3 = _mm_setzero_pd();
1665 fjy3 = _mm_setzero_pd();
1666 fjz3 = _mm_setzero_pd();
1667
1668 /**************************
1669 * CALCULATE INTERACTIONS *
1670 **************************/
1671
1672 if (gmx_mm_any_lt(rsq11,rcutoff2))
1673 {
1674
1675 /* REACTION-FIELD ELECTROSTATICS */
1676 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1677
1678 cutoff_mask = _mm_cmplt_pd(rsq11,rcutoff2);
1679
1680 fscal = felec;
1681
1682 fscal = _mm_and_pd(fscal,cutoff_mask);
1683
1684 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1685
1686 /* Calculate temporary vectorial force */
1687 tx = _mm_mul_pd(fscal,dx11);
1688 ty = _mm_mul_pd(fscal,dy11);
1689 tz = _mm_mul_pd(fscal,dz11);
1690
1691 /* Update vectorial force */
1692 fix1 = _mm_add_pd(fix1,tx);
1693 fiy1 = _mm_add_pd(fiy1,ty);
1694 fiz1 = _mm_add_pd(fiz1,tz);
1695
1696 fjx1 = _mm_add_pd(fjx1,tx);
1697 fjy1 = _mm_add_pd(fjy1,ty);
1698 fjz1 = _mm_add_pd(fjz1,tz);
1699
1700 }
1701
1702 /**************************
1703 * CALCULATE INTERACTIONS *
1704 **************************/
1705
1706 if (gmx_mm_any_lt(rsq12,rcutoff2))
1707 {
1708
1709 /* REACTION-FIELD ELECTROSTATICS */
1710 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1711
1712 cutoff_mask = _mm_cmplt_pd(rsq12,rcutoff2);
1713
1714 fscal = felec;
1715
1716 fscal = _mm_and_pd(fscal,cutoff_mask);
1717
1718 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1719
1720 /* Calculate temporary vectorial force */
1721 tx = _mm_mul_pd(fscal,dx12);
1722 ty = _mm_mul_pd(fscal,dy12);
1723 tz = _mm_mul_pd(fscal,dz12);
1724
1725 /* Update vectorial force */
1726 fix1 = _mm_add_pd(fix1,tx);
1727 fiy1 = _mm_add_pd(fiy1,ty);
1728 fiz1 = _mm_add_pd(fiz1,tz);
1729
1730 fjx2 = _mm_add_pd(fjx2,tx);
1731 fjy2 = _mm_add_pd(fjy2,ty);
1732 fjz2 = _mm_add_pd(fjz2,tz);
1733
1734 }
1735
1736 /**************************
1737 * CALCULATE INTERACTIONS *
1738 **************************/
1739
1740 if (gmx_mm_any_lt(rsq13,rcutoff2))
1741 {
1742
1743 /* REACTION-FIELD ELECTROSTATICS */
1744 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1745
1746 cutoff_mask = _mm_cmplt_pd(rsq13,rcutoff2);
1747
1748 fscal = felec;
1749
1750 fscal = _mm_and_pd(fscal,cutoff_mask);
1751
1752 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1753
1754 /* Calculate temporary vectorial force */
1755 tx = _mm_mul_pd(fscal,dx13);
1756 ty = _mm_mul_pd(fscal,dy13);
1757 tz = _mm_mul_pd(fscal,dz13);
1758
1759 /* Update vectorial force */
1760 fix1 = _mm_add_pd(fix1,tx);
1761 fiy1 = _mm_add_pd(fiy1,ty);
1762 fiz1 = _mm_add_pd(fiz1,tz);
1763
1764 fjx3 = _mm_add_pd(fjx3,tx);
1765 fjy3 = _mm_add_pd(fjy3,ty);
1766 fjz3 = _mm_add_pd(fjz3,tz);
1767
1768 }
1769
1770 /**************************
1771 * CALCULATE INTERACTIONS *
1772 **************************/
1773
1774 if (gmx_mm_any_lt(rsq21,rcutoff2))
1775 {
1776
1777 /* REACTION-FIELD ELECTROSTATICS */
1778 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1779
1780 cutoff_mask = _mm_cmplt_pd(rsq21,rcutoff2);
1781
1782 fscal = felec;
1783
1784 fscal = _mm_and_pd(fscal,cutoff_mask);
1785
1786 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1787
1788 /* Calculate temporary vectorial force */
1789 tx = _mm_mul_pd(fscal,dx21);
1790 ty = _mm_mul_pd(fscal,dy21);
1791 tz = _mm_mul_pd(fscal,dz21);
1792
1793 /* Update vectorial force */
1794 fix2 = _mm_add_pd(fix2,tx);
1795 fiy2 = _mm_add_pd(fiy2,ty);
1796 fiz2 = _mm_add_pd(fiz2,tz);
1797
1798 fjx1 = _mm_add_pd(fjx1,tx);
1799 fjy1 = _mm_add_pd(fjy1,ty);
1800 fjz1 = _mm_add_pd(fjz1,tz);
1801
1802 }
1803
1804 /**************************
1805 * CALCULATE INTERACTIONS *
1806 **************************/
1807
1808 if (gmx_mm_any_lt(rsq22,rcutoff2))
1809 {
1810
1811 /* REACTION-FIELD ELECTROSTATICS */
1812 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1813
1814 cutoff_mask = _mm_cmplt_pd(rsq22,rcutoff2);
1815
1816 fscal = felec;
1817
1818 fscal = _mm_and_pd(fscal,cutoff_mask);
1819
1820 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1821
1822 /* Calculate temporary vectorial force */
1823 tx = _mm_mul_pd(fscal,dx22);
1824 ty = _mm_mul_pd(fscal,dy22);
1825 tz = _mm_mul_pd(fscal,dz22);
1826
1827 /* Update vectorial force */
1828 fix2 = _mm_add_pd(fix2,tx);
1829 fiy2 = _mm_add_pd(fiy2,ty);
1830 fiz2 = _mm_add_pd(fiz2,tz);
1831
1832 fjx2 = _mm_add_pd(fjx2,tx);
1833 fjy2 = _mm_add_pd(fjy2,ty);
1834 fjz2 = _mm_add_pd(fjz2,tz);
1835
1836 }
1837
1838 /**************************
1839 * CALCULATE INTERACTIONS *
1840 **************************/
1841
1842 if (gmx_mm_any_lt(rsq23,rcutoff2))
1843 {
1844
1845 /* REACTION-FIELD ELECTROSTATICS */
1846 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1847
1848 cutoff_mask = _mm_cmplt_pd(rsq23,rcutoff2);
1849
1850 fscal = felec;
1851
1852 fscal = _mm_and_pd(fscal,cutoff_mask);
1853
1854 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1855
1856 /* Calculate temporary vectorial force */
1857 tx = _mm_mul_pd(fscal,dx23);
1858 ty = _mm_mul_pd(fscal,dy23);
1859 tz = _mm_mul_pd(fscal,dz23);
1860
1861 /* Update vectorial force */
1862 fix2 = _mm_add_pd(fix2,tx);
1863 fiy2 = _mm_add_pd(fiy2,ty);
1864 fiz2 = _mm_add_pd(fiz2,tz);
1865
1866 fjx3 = _mm_add_pd(fjx3,tx);
1867 fjy3 = _mm_add_pd(fjy3,ty);
1868 fjz3 = _mm_add_pd(fjz3,tz);
1869
1870 }
1871
1872 /**************************
1873 * CALCULATE INTERACTIONS *
1874 **************************/
1875
1876 if (gmx_mm_any_lt(rsq31,rcutoff2))
1877 {
1878
1879 /* REACTION-FIELD ELECTROSTATICS */
1880 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1881
1882 cutoff_mask = _mm_cmplt_pd(rsq31,rcutoff2);
1883
1884 fscal = felec;
1885
1886 fscal = _mm_and_pd(fscal,cutoff_mask);
1887
1888 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1889
1890 /* Calculate temporary vectorial force */
1891 tx = _mm_mul_pd(fscal,dx31);
1892 ty = _mm_mul_pd(fscal,dy31);
1893 tz = _mm_mul_pd(fscal,dz31);
1894
1895 /* Update vectorial force */
1896 fix3 = _mm_add_pd(fix3,tx);
1897 fiy3 = _mm_add_pd(fiy3,ty);
1898 fiz3 = _mm_add_pd(fiz3,tz);
1899
1900 fjx1 = _mm_add_pd(fjx1,tx);
1901 fjy1 = _mm_add_pd(fjy1,ty);
1902 fjz1 = _mm_add_pd(fjz1,tz);
1903
1904 }
1905
1906 /**************************
1907 * CALCULATE INTERACTIONS *
1908 **************************/
1909
1910 if (gmx_mm_any_lt(rsq32,rcutoff2))
1911 {
1912
1913 /* REACTION-FIELD ELECTROSTATICS */
1914 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1915
1916 cutoff_mask = _mm_cmplt_pd(rsq32,rcutoff2);
1917
1918 fscal = felec;
1919
1920 fscal = _mm_and_pd(fscal,cutoff_mask);
1921
1922 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1923
1924 /* Calculate temporary vectorial force */
1925 tx = _mm_mul_pd(fscal,dx32);
1926 ty = _mm_mul_pd(fscal,dy32);
1927 tz = _mm_mul_pd(fscal,dz32);
1928
1929 /* Update vectorial force */
1930 fix3 = _mm_add_pd(fix3,tx);
1931 fiy3 = _mm_add_pd(fiy3,ty);
1932 fiz3 = _mm_add_pd(fiz3,tz);
1933
1934 fjx2 = _mm_add_pd(fjx2,tx);
1935 fjy2 = _mm_add_pd(fjy2,ty);
1936 fjz2 = _mm_add_pd(fjz2,tz);
1937
1938 }
1939
1940 /**************************
1941 * CALCULATE INTERACTIONS *
1942 **************************/
1943
1944 if (gmx_mm_any_lt(rsq33,rcutoff2))
1945 {
1946
1947 /* REACTION-FIELD ELECTROSTATICS */
1948 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1949
1950 cutoff_mask = _mm_cmplt_pd(rsq33,rcutoff2);
1951
1952 fscal = felec;
1953
1954 fscal = _mm_and_pd(fscal,cutoff_mask);
1955
1956 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1957
1958 /* Calculate temporary vectorial force */
1959 tx = _mm_mul_pd(fscal,dx33);
1960 ty = _mm_mul_pd(fscal,dy33);
1961 tz = _mm_mul_pd(fscal,dz33);
1962
1963 /* Update vectorial force */
1964 fix3 = _mm_add_pd(fix3,tx);
1965 fiy3 = _mm_add_pd(fiy3,ty);
1966 fiz3 = _mm_add_pd(fiz3,tz);
1967
1968 fjx3 = _mm_add_pd(fjx3,tx);
1969 fjy3 = _mm_add_pd(fjy3,ty);
1970 fjz3 = _mm_add_pd(fjz3,tz);
1971
1972 }
1973
1974 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1975
1976 /* Inner loop uses 270 flops */
1977 }
1978
1979 /* End of innermost loop */
1980
1981 gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1982 f+i_coord_offset+DIM,fshift+i_shift_offset);
1983
1984 /* Increment number of inner iterations */
1985 inneriter += j_index_end - j_index_start;
1986
1987 /* Outer loop uses 18 flops */
1988 }
1989
1990 /* Increment number of outer iterations */
1991 outeriter += nri;
1992
1993 /* Update outer/inner flops */
1994
1995 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1996 }
The ultimate molecular dynamics simulation package