?  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_VdwLJ_GeomW4W4_sse2_double.c
blob7cd3a39ed0b8eccb8f26a9bf31a0cdad4709540a
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_VdwLJ_GeomW4W4_VF_sse2_double
53 * Electrostatics interaction: ReactionField
54 * VdW interaction: LennardJones
55 * Geometry: Water4-Water4
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecRF_VdwLJ_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 vdwioffset0;
82 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
83 int vdwioffset1;
84 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
85 int vdwioffset2;
86 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
87 int vdwioffset3;
88 __m128d ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
89 int vdwjidx0A,vdwjidx0B;
90 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
91 int vdwjidx1A,vdwjidx1B;
92 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
93 int vdwjidx2A,vdwjidx2B;
94 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
95 int vdwjidx3A,vdwjidx3B;
96 __m128d jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
97 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
98 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
99 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
100 __m128d dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
101 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
102 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
103 __m128d dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
104 __m128d dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
105 __m128d dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
106 __m128d dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
107 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
108 real *charge;
109 int nvdwtype;
110 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
111 int *vdwtype;
112 real *vdwparam;
113 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
114 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
115 __m128d dummy_mask,cutoff_mask;
116 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
117 __m128d one = _mm_set1_pd(1.0);
118 __m128d two = _mm_set1_pd(2.0);
119 x = xx[0];
120 f = ff[0];
121
122 nri = nlist->nri;
123 iinr = nlist->iinr;
124 jindex = nlist->jindex;
125 jjnr = nlist->jjnr;
126 shiftidx = nlist->shift;
127 gid = nlist->gid;
128 shiftvec = fr->shift_vec[0];
129 fshift = fr->fshift[0];
130 facel = _mm_set1_pd(fr->epsfac);
131 charge = mdatoms->chargeA;
132 krf = _mm_set1_pd(fr->ic->k_rf);
133 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
134 crf = _mm_set1_pd(fr->ic->c_rf);
135 nvdwtype = fr->ntype;
136 vdwparam = fr->nbfp;
137 vdwtype = mdatoms->typeA;
138
139 /* Setup water-specific parameters */
140 inr = nlist->iinr[0];
141 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
142 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
143 iq3 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
144 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
145
146 jq1 = _mm_set1_pd(charge[inr+1]);
147 jq2 = _mm_set1_pd(charge[inr+2]);
148 jq3 = _mm_set1_pd(charge[inr+3]);
149 vdwjidx0A = 2*vdwtype[inr+0];
150 c6_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
151 c12_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
152 qq11 = _mm_mul_pd(iq1,jq1);
153 qq12 = _mm_mul_pd(iq1,jq2);
154 qq13 = _mm_mul_pd(iq1,jq3);
155 qq21 = _mm_mul_pd(iq2,jq1);
156 qq22 = _mm_mul_pd(iq2,jq2);
157 qq23 = _mm_mul_pd(iq2,jq3);
158 qq31 = _mm_mul_pd(iq3,jq1);
159 qq32 = _mm_mul_pd(iq3,jq2);
160 qq33 = _mm_mul_pd(iq3,jq3);
161
162 /* Avoid stupid compiler warnings */
163 jnrA = jnrB = 0;
164 j_coord_offsetA = 0;
165 j_coord_offsetB = 0;
166
167 outeriter = 0;
168 inneriter = 0;
169
170 /* Start outer loop over neighborlists */
171 for(iidx=0; iidx<nri; iidx++)
172 {
173 /* Load shift vector for this list */
174 i_shift_offset = DIM*shiftidx[iidx];
175
176 /* Load limits for loop over neighbors */
177 j_index_start = jindex[iidx];
178 j_index_end = jindex[iidx+1];
179
180 /* Get outer coordinate index */
181 inr = iinr[iidx];
182 i_coord_offset = DIM*inr;
183
184 /* Load i particle coords and add shift vector */
185 gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
186 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
187
188 fix0 = _mm_setzero_pd();
189 fiy0 = _mm_setzero_pd();
190 fiz0 = _mm_setzero_pd();
191 fix1 = _mm_setzero_pd();
192 fiy1 = _mm_setzero_pd();
193 fiz1 = _mm_setzero_pd();
194 fix2 = _mm_setzero_pd();
195 fiy2 = _mm_setzero_pd();
196 fiz2 = _mm_setzero_pd();
197 fix3 = _mm_setzero_pd();
198 fiy3 = _mm_setzero_pd();
199 fiz3 = _mm_setzero_pd();
200
201 /* Reset potential sums */
202 velecsum = _mm_setzero_pd();
203 vvdwsum = _mm_setzero_pd();
204
205 /* Start inner kernel loop */
206 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
207 {
208
209 /* Get j neighbor index, and coordinate index */
210 jnrA = jjnr[jidx];
211 jnrB = jjnr[jidx+1];
212 j_coord_offsetA = DIM*jnrA;
213 j_coord_offsetB = DIM*jnrB;
214
215 /* load j atom coordinates */
216 gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
217 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
218 &jy2,&jz2,&jx3,&jy3,&jz3);
219
220 /* Calculate displacement vector */
221 dx00 = _mm_sub_pd(ix0,jx0);
222 dy00 = _mm_sub_pd(iy0,jy0);
223 dz00 = _mm_sub_pd(iz0,jz0);
224 dx11 = _mm_sub_pd(ix1,jx1);
225 dy11 = _mm_sub_pd(iy1,jy1);
226 dz11 = _mm_sub_pd(iz1,jz1);
227 dx12 = _mm_sub_pd(ix1,jx2);
228 dy12 = _mm_sub_pd(iy1,jy2);
229 dz12 = _mm_sub_pd(iz1,jz2);
230 dx13 = _mm_sub_pd(ix1,jx3);
231 dy13 = _mm_sub_pd(iy1,jy3);
232 dz13 = _mm_sub_pd(iz1,jz3);
233 dx21 = _mm_sub_pd(ix2,jx1);
234 dy21 = _mm_sub_pd(iy2,jy1);
235 dz21 = _mm_sub_pd(iz2,jz1);
236 dx22 = _mm_sub_pd(ix2,jx2);
237 dy22 = _mm_sub_pd(iy2,jy2);
238 dz22 = _mm_sub_pd(iz2,jz2);
239 dx23 = _mm_sub_pd(ix2,jx3);
240 dy23 = _mm_sub_pd(iy2,jy3);
241 dz23 = _mm_sub_pd(iz2,jz3);
242 dx31 = _mm_sub_pd(ix3,jx1);
243 dy31 = _mm_sub_pd(iy3,jy1);
244 dz31 = _mm_sub_pd(iz3,jz1);
245 dx32 = _mm_sub_pd(ix3,jx2);
246 dy32 = _mm_sub_pd(iy3,jy2);
247 dz32 = _mm_sub_pd(iz3,jz2);
248 dx33 = _mm_sub_pd(ix3,jx3);
249 dy33 = _mm_sub_pd(iy3,jy3);
250 dz33 = _mm_sub_pd(iz3,jz3);
251
252 /* Calculate squared distance and things based on it */
253 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
254 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
255 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
256 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
257 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
258 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
259 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
260 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
261 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
262 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
263
264 rinv11 = gmx_mm_invsqrt_pd(rsq11);
265 rinv12 = gmx_mm_invsqrt_pd(rsq12);
266 rinv13 = gmx_mm_invsqrt_pd(rsq13);
267 rinv21 = gmx_mm_invsqrt_pd(rsq21);
268 rinv22 = gmx_mm_invsqrt_pd(rsq22);
269 rinv23 = gmx_mm_invsqrt_pd(rsq23);
270 rinv31 = gmx_mm_invsqrt_pd(rsq31);
271 rinv32 = gmx_mm_invsqrt_pd(rsq32);
272 rinv33 = gmx_mm_invsqrt_pd(rsq33);
273
274 rinvsq00 = gmx_mm_inv_pd(rsq00);
275 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
276 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
277 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
278 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
279 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
280 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
281 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
282 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
283 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
284
285 fjx0 = _mm_setzero_pd();
286 fjy0 = _mm_setzero_pd();
287 fjz0 = _mm_setzero_pd();
288 fjx1 = _mm_setzero_pd();
289 fjy1 = _mm_setzero_pd();
290 fjz1 = _mm_setzero_pd();
291 fjx2 = _mm_setzero_pd();
292 fjy2 = _mm_setzero_pd();
293 fjz2 = _mm_setzero_pd();
294 fjx3 = _mm_setzero_pd();
295 fjy3 = _mm_setzero_pd();
296 fjz3 = _mm_setzero_pd();
297
298 /**************************
299 * CALCULATE INTERACTIONS *
300 **************************/
301
302 /* LENNARD-JONES DISPERSION/REPULSION */
303
304 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
305 vvdw6 = _mm_mul_pd(c6_00,rinvsix);
306 vvdw12 = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
307 vvdw = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
308 fvdw = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
309
310 /* Update potential sum for this i atom from the interaction with this j atom. */
311 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
312
313 fscal = fvdw;
314
315 /* Calculate temporary vectorial force */
316 tx = _mm_mul_pd(fscal,dx00);
317 ty = _mm_mul_pd(fscal,dy00);
318 tz = _mm_mul_pd(fscal,dz00);
319
320 /* Update vectorial force */
321 fix0 = _mm_add_pd(fix0,tx);
322 fiy0 = _mm_add_pd(fiy0,ty);
323 fiz0 = _mm_add_pd(fiz0,tz);
324
325 fjx0 = _mm_add_pd(fjx0,tx);
326 fjy0 = _mm_add_pd(fjy0,ty);
327 fjz0 = _mm_add_pd(fjz0,tz);
328
329 /**************************
330 * CALCULATE INTERACTIONS *
331 **************************/
332
333 /* REACTION-FIELD ELECTROSTATICS */
334 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
335 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
336
337 /* Update potential sum for this i atom from the interaction with this j atom. */
338 velecsum = _mm_add_pd(velecsum,velec);
339
340 fscal = felec;
341
342 /* Calculate temporary vectorial force */
343 tx = _mm_mul_pd(fscal,dx11);
344 ty = _mm_mul_pd(fscal,dy11);
345 tz = _mm_mul_pd(fscal,dz11);
346
347 /* Update vectorial force */
348 fix1 = _mm_add_pd(fix1,tx);
349 fiy1 = _mm_add_pd(fiy1,ty);
350 fiz1 = _mm_add_pd(fiz1,tz);
351
352 fjx1 = _mm_add_pd(fjx1,tx);
353 fjy1 = _mm_add_pd(fjy1,ty);
354 fjz1 = _mm_add_pd(fjz1,tz);
355
356 /**************************
357 * CALCULATE INTERACTIONS *
358 **************************/
359
360 /* REACTION-FIELD ELECTROSTATICS */
361 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
362 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
363
364 /* Update potential sum for this i atom from the interaction with this j atom. */
365 velecsum = _mm_add_pd(velecsum,velec);
366
367 fscal = felec;
368
369 /* Calculate temporary vectorial force */
370 tx = _mm_mul_pd(fscal,dx12);
371 ty = _mm_mul_pd(fscal,dy12);
372 tz = _mm_mul_pd(fscal,dz12);
373
374 /* Update vectorial force */
375 fix1 = _mm_add_pd(fix1,tx);
376 fiy1 = _mm_add_pd(fiy1,ty);
377 fiz1 = _mm_add_pd(fiz1,tz);
378
379 fjx2 = _mm_add_pd(fjx2,tx);
380 fjy2 = _mm_add_pd(fjy2,ty);
381 fjz2 = _mm_add_pd(fjz2,tz);
382
383 /**************************
384 * CALCULATE INTERACTIONS *
385 **************************/
386
387 /* REACTION-FIELD ELECTROSTATICS */
388 velec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
389 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
390
391 /* Update potential sum for this i atom from the interaction with this j atom. */
392 velecsum = _mm_add_pd(velecsum,velec);
393
394 fscal = felec;
395
396 /* Calculate temporary vectorial force */
397 tx = _mm_mul_pd(fscal,dx13);
398 ty = _mm_mul_pd(fscal,dy13);
399 tz = _mm_mul_pd(fscal,dz13);
400
401 /* Update vectorial force */
402 fix1 = _mm_add_pd(fix1,tx);
403 fiy1 = _mm_add_pd(fiy1,ty);
404 fiz1 = _mm_add_pd(fiz1,tz);
405
406 fjx3 = _mm_add_pd(fjx3,tx);
407 fjy3 = _mm_add_pd(fjy3,ty);
408 fjz3 = _mm_add_pd(fjz3,tz);
409
410 /**************************
411 * CALCULATE INTERACTIONS *
412 **************************/
413
414 /* REACTION-FIELD ELECTROSTATICS */
415 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
416 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
417
418 /* Update potential sum for this i atom from the interaction with this j atom. */
419 velecsum = _mm_add_pd(velecsum,velec);
420
421 fscal = felec;
422
423 /* Calculate temporary vectorial force */
424 tx = _mm_mul_pd(fscal,dx21);
425 ty = _mm_mul_pd(fscal,dy21);
426 tz = _mm_mul_pd(fscal,dz21);
427
428 /* Update vectorial force */
429 fix2 = _mm_add_pd(fix2,tx);
430 fiy2 = _mm_add_pd(fiy2,ty);
431 fiz2 = _mm_add_pd(fiz2,tz);
432
433 fjx1 = _mm_add_pd(fjx1,tx);
434 fjy1 = _mm_add_pd(fjy1,ty);
435 fjz1 = _mm_add_pd(fjz1,tz);
436
437 /**************************
438 * CALCULATE INTERACTIONS *
439 **************************/
440
441 /* REACTION-FIELD ELECTROSTATICS */
442 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
443 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
444
445 /* Update potential sum for this i atom from the interaction with this j atom. */
446 velecsum = _mm_add_pd(velecsum,velec);
447
448 fscal = felec;
449
450 /* Calculate temporary vectorial force */
451 tx = _mm_mul_pd(fscal,dx22);
452 ty = _mm_mul_pd(fscal,dy22);
453 tz = _mm_mul_pd(fscal,dz22);
454
455 /* Update vectorial force */
456 fix2 = _mm_add_pd(fix2,tx);
457 fiy2 = _mm_add_pd(fiy2,ty);
458 fiz2 = _mm_add_pd(fiz2,tz);
459
460 fjx2 = _mm_add_pd(fjx2,tx);
461 fjy2 = _mm_add_pd(fjy2,ty);
462 fjz2 = _mm_add_pd(fjz2,tz);
463
464 /**************************
465 * CALCULATE INTERACTIONS *
466 **************************/
467
468 /* REACTION-FIELD ELECTROSTATICS */
469 velec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
470 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
471
472 /* Update potential sum for this i atom from the interaction with this j atom. */
473 velecsum = _mm_add_pd(velecsum,velec);
474
475 fscal = felec;
476
477 /* Calculate temporary vectorial force */
478 tx = _mm_mul_pd(fscal,dx23);
479 ty = _mm_mul_pd(fscal,dy23);
480 tz = _mm_mul_pd(fscal,dz23);
481
482 /* Update vectorial force */
483 fix2 = _mm_add_pd(fix2,tx);
484 fiy2 = _mm_add_pd(fiy2,ty);
485 fiz2 = _mm_add_pd(fiz2,tz);
486
487 fjx3 = _mm_add_pd(fjx3,tx);
488 fjy3 = _mm_add_pd(fjy3,ty);
489 fjz3 = _mm_add_pd(fjz3,tz);
490
491 /**************************
492 * CALCULATE INTERACTIONS *
493 **************************/
494
495 /* REACTION-FIELD ELECTROSTATICS */
496 velec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
497 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
498
499 /* Update potential sum for this i atom from the interaction with this j atom. */
500 velecsum = _mm_add_pd(velecsum,velec);
501
502 fscal = felec;
503
504 /* Calculate temporary vectorial force */
505 tx = _mm_mul_pd(fscal,dx31);
506 ty = _mm_mul_pd(fscal,dy31);
507 tz = _mm_mul_pd(fscal,dz31);
508
509 /* Update vectorial force */
510 fix3 = _mm_add_pd(fix3,tx);
511 fiy3 = _mm_add_pd(fiy3,ty);
512 fiz3 = _mm_add_pd(fiz3,tz);
513
514 fjx1 = _mm_add_pd(fjx1,tx);
515 fjy1 = _mm_add_pd(fjy1,ty);
516 fjz1 = _mm_add_pd(fjz1,tz);
517
518 /**************************
519 * CALCULATE INTERACTIONS *
520 **************************/
521
522 /* REACTION-FIELD ELECTROSTATICS */
523 velec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
524 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
525
526 /* Update potential sum for this i atom from the interaction with this j atom. */
527 velecsum = _mm_add_pd(velecsum,velec);
528
529 fscal = felec;
530
531 /* Calculate temporary vectorial force */
532 tx = _mm_mul_pd(fscal,dx32);
533 ty = _mm_mul_pd(fscal,dy32);
534 tz = _mm_mul_pd(fscal,dz32);
535
536 /* Update vectorial force */
537 fix3 = _mm_add_pd(fix3,tx);
538 fiy3 = _mm_add_pd(fiy3,ty);
539 fiz3 = _mm_add_pd(fiz3,tz);
540
541 fjx2 = _mm_add_pd(fjx2,tx);
542 fjy2 = _mm_add_pd(fjy2,ty);
543 fjz2 = _mm_add_pd(fjz2,tz);
544
545 /**************************
546 * CALCULATE INTERACTIONS *
547 **************************/
548
549 /* REACTION-FIELD ELECTROSTATICS */
550 velec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
551 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
552
553 /* Update potential sum for this i atom from the interaction with this j atom. */
554 velecsum = _mm_add_pd(velecsum,velec);
555
556 fscal = felec;
557
558 /* Calculate temporary vectorial force */
559 tx = _mm_mul_pd(fscal,dx33);
560 ty = _mm_mul_pd(fscal,dy33);
561 tz = _mm_mul_pd(fscal,dz33);
562
563 /* Update vectorial force */
564 fix3 = _mm_add_pd(fix3,tx);
565 fiy3 = _mm_add_pd(fiy3,ty);
566 fiz3 = _mm_add_pd(fiz3,tz);
567
568 fjx3 = _mm_add_pd(fjx3,tx);
569 fjy3 = _mm_add_pd(fjy3,ty);
570 fjz3 = _mm_add_pd(fjz3,tz);
571
572 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
573
574 /* Inner loop uses 323 flops */
575 }
576
577 if(jidx<j_index_end)
578 {
579
580 jnrA = jjnr[jidx];
581 j_coord_offsetA = DIM*jnrA;
582
583 /* load j atom coordinates */
584 gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
585 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
586 &jy2,&jz2,&jx3,&jy3,&jz3);
587
588 /* Calculate displacement vector */
589 dx00 = _mm_sub_pd(ix0,jx0);
590 dy00 = _mm_sub_pd(iy0,jy0);
591 dz00 = _mm_sub_pd(iz0,jz0);
592 dx11 = _mm_sub_pd(ix1,jx1);
593 dy11 = _mm_sub_pd(iy1,jy1);
594 dz11 = _mm_sub_pd(iz1,jz1);
595 dx12 = _mm_sub_pd(ix1,jx2);
596 dy12 = _mm_sub_pd(iy1,jy2);
597 dz12 = _mm_sub_pd(iz1,jz2);
598 dx13 = _mm_sub_pd(ix1,jx3);
599 dy13 = _mm_sub_pd(iy1,jy3);
600 dz13 = _mm_sub_pd(iz1,jz3);
601 dx21 = _mm_sub_pd(ix2,jx1);
602 dy21 = _mm_sub_pd(iy2,jy1);
603 dz21 = _mm_sub_pd(iz2,jz1);
604 dx22 = _mm_sub_pd(ix2,jx2);
605 dy22 = _mm_sub_pd(iy2,jy2);
606 dz22 = _mm_sub_pd(iz2,jz2);
607 dx23 = _mm_sub_pd(ix2,jx3);
608 dy23 = _mm_sub_pd(iy2,jy3);
609 dz23 = _mm_sub_pd(iz2,jz3);
610 dx31 = _mm_sub_pd(ix3,jx1);
611 dy31 = _mm_sub_pd(iy3,jy1);
612 dz31 = _mm_sub_pd(iz3,jz1);
613 dx32 = _mm_sub_pd(ix3,jx2);
614 dy32 = _mm_sub_pd(iy3,jy2);
615 dz32 = _mm_sub_pd(iz3,jz2);
616 dx33 = _mm_sub_pd(ix3,jx3);
617 dy33 = _mm_sub_pd(iy3,jy3);
618 dz33 = _mm_sub_pd(iz3,jz3);
619
620 /* Calculate squared distance and things based on it */
621 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
622 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
623 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
624 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
625 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
626 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
627 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
628 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
629 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
630 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
631
632 rinv11 = gmx_mm_invsqrt_pd(rsq11);
633 rinv12 = gmx_mm_invsqrt_pd(rsq12);
634 rinv13 = gmx_mm_invsqrt_pd(rsq13);
635 rinv21 = gmx_mm_invsqrt_pd(rsq21);
636 rinv22 = gmx_mm_invsqrt_pd(rsq22);
637 rinv23 = gmx_mm_invsqrt_pd(rsq23);
638 rinv31 = gmx_mm_invsqrt_pd(rsq31);
639 rinv32 = gmx_mm_invsqrt_pd(rsq32);
640 rinv33 = gmx_mm_invsqrt_pd(rsq33);
641
642 rinvsq00 = gmx_mm_inv_pd(rsq00);
643 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
644 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
645 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
646 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
647 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
648 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
649 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
650 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
651 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
652
653 fjx0 = _mm_setzero_pd();
654 fjy0 = _mm_setzero_pd();
655 fjz0 = _mm_setzero_pd();
656 fjx1 = _mm_setzero_pd();
657 fjy1 = _mm_setzero_pd();
658 fjz1 = _mm_setzero_pd();
659 fjx2 = _mm_setzero_pd();
660 fjy2 = _mm_setzero_pd();
661 fjz2 = _mm_setzero_pd();
662 fjx3 = _mm_setzero_pd();
663 fjy3 = _mm_setzero_pd();
664 fjz3 = _mm_setzero_pd();
665
666 /**************************
667 * CALCULATE INTERACTIONS *
668 **************************/
669
670 /* LENNARD-JONES DISPERSION/REPULSION */
671
672 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
673 vvdw6 = _mm_mul_pd(c6_00,rinvsix);
674 vvdw12 = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
675 vvdw = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
676 fvdw = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
677
678 /* Update potential sum for this i atom from the interaction with this j atom. */
679 vvdw = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
680 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
681
682 fscal = fvdw;
683
684 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
685
686 /* Calculate temporary vectorial force */
687 tx = _mm_mul_pd(fscal,dx00);
688 ty = _mm_mul_pd(fscal,dy00);
689 tz = _mm_mul_pd(fscal,dz00);
690
691 /* Update vectorial force */
692 fix0 = _mm_add_pd(fix0,tx);
693 fiy0 = _mm_add_pd(fiy0,ty);
694 fiz0 = _mm_add_pd(fiz0,tz);
695
696 fjx0 = _mm_add_pd(fjx0,tx);
697 fjy0 = _mm_add_pd(fjy0,ty);
698 fjz0 = _mm_add_pd(fjz0,tz);
699
700 /**************************
701 * CALCULATE INTERACTIONS *
702 **************************/
703
704 /* REACTION-FIELD ELECTROSTATICS */
705 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
706 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
707
708 /* Update potential sum for this i atom from the interaction with this j atom. */
709 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
710 velecsum = _mm_add_pd(velecsum,velec);
711
712 fscal = felec;
713
714 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
715
716 /* Calculate temporary vectorial force */
717 tx = _mm_mul_pd(fscal,dx11);
718 ty = _mm_mul_pd(fscal,dy11);
719 tz = _mm_mul_pd(fscal,dz11);
720
721 /* Update vectorial force */
722 fix1 = _mm_add_pd(fix1,tx);
723 fiy1 = _mm_add_pd(fiy1,ty);
724 fiz1 = _mm_add_pd(fiz1,tz);
725
726 fjx1 = _mm_add_pd(fjx1,tx);
727 fjy1 = _mm_add_pd(fjy1,ty);
728 fjz1 = _mm_add_pd(fjz1,tz);
729
730 /**************************
731 * CALCULATE INTERACTIONS *
732 **************************/
733
734 /* REACTION-FIELD ELECTROSTATICS */
735 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
736 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
737
738 /* Update potential sum for this i atom from the interaction with this j atom. */
739 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
740 velecsum = _mm_add_pd(velecsum,velec);
741
742 fscal = felec;
743
744 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
745
746 /* Calculate temporary vectorial force */
747 tx = _mm_mul_pd(fscal,dx12);
748 ty = _mm_mul_pd(fscal,dy12);
749 tz = _mm_mul_pd(fscal,dz12);
750
751 /* Update vectorial force */
752 fix1 = _mm_add_pd(fix1,tx);
753 fiy1 = _mm_add_pd(fiy1,ty);
754 fiz1 = _mm_add_pd(fiz1,tz);
755
756 fjx2 = _mm_add_pd(fjx2,tx);
757 fjy2 = _mm_add_pd(fjy2,ty);
758 fjz2 = _mm_add_pd(fjz2,tz);
759
760 /**************************
761 * CALCULATE INTERACTIONS *
762 **************************/
763
764 /* REACTION-FIELD ELECTROSTATICS */
765 velec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
766 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
767
768 /* Update potential sum for this i atom from the interaction with this j atom. */
769 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
770 velecsum = _mm_add_pd(velecsum,velec);
771
772 fscal = felec;
773
774 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
775
776 /* Calculate temporary vectorial force */
777 tx = _mm_mul_pd(fscal,dx13);
778 ty = _mm_mul_pd(fscal,dy13);
779 tz = _mm_mul_pd(fscal,dz13);
780
781 /* Update vectorial force */
782 fix1 = _mm_add_pd(fix1,tx);
783 fiy1 = _mm_add_pd(fiy1,ty);
784 fiz1 = _mm_add_pd(fiz1,tz);
785
786 fjx3 = _mm_add_pd(fjx3,tx);
787 fjy3 = _mm_add_pd(fjy3,ty);
788 fjz3 = _mm_add_pd(fjz3,tz);
789
790 /**************************
791 * CALCULATE INTERACTIONS *
792 **************************/
793
794 /* REACTION-FIELD ELECTROSTATICS */
795 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
796 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
797
798 /* Update potential sum for this i atom from the interaction with this j atom. */
799 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
800 velecsum = _mm_add_pd(velecsum,velec);
801
802 fscal = felec;
803
804 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
805
806 /* Calculate temporary vectorial force */
807 tx = _mm_mul_pd(fscal,dx21);
808 ty = _mm_mul_pd(fscal,dy21);
809 tz = _mm_mul_pd(fscal,dz21);
810
811 /* Update vectorial force */
812 fix2 = _mm_add_pd(fix2,tx);
813 fiy2 = _mm_add_pd(fiy2,ty);
814 fiz2 = _mm_add_pd(fiz2,tz);
815
816 fjx1 = _mm_add_pd(fjx1,tx);
817 fjy1 = _mm_add_pd(fjy1,ty);
818 fjz1 = _mm_add_pd(fjz1,tz);
819
820 /**************************
821 * CALCULATE INTERACTIONS *
822 **************************/
823
824 /* REACTION-FIELD ELECTROSTATICS */
825 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
826 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
827
828 /* Update potential sum for this i atom from the interaction with this j atom. */
829 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
830 velecsum = _mm_add_pd(velecsum,velec);
831
832 fscal = felec;
833
834 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
835
836 /* Calculate temporary vectorial force */
837 tx = _mm_mul_pd(fscal,dx22);
838 ty = _mm_mul_pd(fscal,dy22);
839 tz = _mm_mul_pd(fscal,dz22);
840
841 /* Update vectorial force */
842 fix2 = _mm_add_pd(fix2,tx);
843 fiy2 = _mm_add_pd(fiy2,ty);
844 fiz2 = _mm_add_pd(fiz2,tz);
845
846 fjx2 = _mm_add_pd(fjx2,tx);
847 fjy2 = _mm_add_pd(fjy2,ty);
848 fjz2 = _mm_add_pd(fjz2,tz);
849
850 /**************************
851 * CALCULATE INTERACTIONS *
852 **************************/
853
854 /* REACTION-FIELD ELECTROSTATICS */
855 velec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
856 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
857
858 /* Update potential sum for this i atom from the interaction with this j atom. */
859 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
860 velecsum = _mm_add_pd(velecsum,velec);
861
862 fscal = felec;
863
864 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
865
866 /* Calculate temporary vectorial force */
867 tx = _mm_mul_pd(fscal,dx23);
868 ty = _mm_mul_pd(fscal,dy23);
869 tz = _mm_mul_pd(fscal,dz23);
870
871 /* Update vectorial force */
872 fix2 = _mm_add_pd(fix2,tx);
873 fiy2 = _mm_add_pd(fiy2,ty);
874 fiz2 = _mm_add_pd(fiz2,tz);
875
876 fjx3 = _mm_add_pd(fjx3,tx);
877 fjy3 = _mm_add_pd(fjy3,ty);
878 fjz3 = _mm_add_pd(fjz3,tz);
879
880 /**************************
881 * CALCULATE INTERACTIONS *
882 **************************/
883
884 /* REACTION-FIELD ELECTROSTATICS */
885 velec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
886 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
887
888 /* Update potential sum for this i atom from the interaction with this j atom. */
889 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
890 velecsum = _mm_add_pd(velecsum,velec);
891
892 fscal = felec;
893
894 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
895
896 /* Calculate temporary vectorial force */
897 tx = _mm_mul_pd(fscal,dx31);
898 ty = _mm_mul_pd(fscal,dy31);
899 tz = _mm_mul_pd(fscal,dz31);
900
901 /* Update vectorial force */
902 fix3 = _mm_add_pd(fix3,tx);
903 fiy3 = _mm_add_pd(fiy3,ty);
904 fiz3 = _mm_add_pd(fiz3,tz);
905
906 fjx1 = _mm_add_pd(fjx1,tx);
907 fjy1 = _mm_add_pd(fjy1,ty);
908 fjz1 = _mm_add_pd(fjz1,tz);
909
910 /**************************
911 * CALCULATE INTERACTIONS *
912 **************************/
913
914 /* REACTION-FIELD ELECTROSTATICS */
915 velec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
916 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
917
918 /* Update potential sum for this i atom from the interaction with this j atom. */
919 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
920 velecsum = _mm_add_pd(velecsum,velec);
921
922 fscal = felec;
923
924 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
925
926 /* Calculate temporary vectorial force */
927 tx = _mm_mul_pd(fscal,dx32);
928 ty = _mm_mul_pd(fscal,dy32);
929 tz = _mm_mul_pd(fscal,dz32);
930
931 /* Update vectorial force */
932 fix3 = _mm_add_pd(fix3,tx);
933 fiy3 = _mm_add_pd(fiy3,ty);
934 fiz3 = _mm_add_pd(fiz3,tz);
935
936 fjx2 = _mm_add_pd(fjx2,tx);
937 fjy2 = _mm_add_pd(fjy2,ty);
938 fjz2 = _mm_add_pd(fjz2,tz);
939
940 /**************************
941 * CALCULATE INTERACTIONS *
942 **************************/
943
944 /* REACTION-FIELD ELECTROSTATICS */
945 velec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
946 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
947
948 /* Update potential sum for this i atom from the interaction with this j atom. */
949 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
950 velecsum = _mm_add_pd(velecsum,velec);
951
952 fscal = felec;
953
954 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
955
956 /* Calculate temporary vectorial force */
957 tx = _mm_mul_pd(fscal,dx33);
958 ty = _mm_mul_pd(fscal,dy33);
959 tz = _mm_mul_pd(fscal,dz33);
960
961 /* Update vectorial force */
962 fix3 = _mm_add_pd(fix3,tx);
963 fiy3 = _mm_add_pd(fiy3,ty);
964 fiz3 = _mm_add_pd(fiz3,tz);
965
966 fjx3 = _mm_add_pd(fjx3,tx);
967 fjy3 = _mm_add_pd(fjy3,ty);
968 fjz3 = _mm_add_pd(fjz3,tz);
969
970 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
971
972 /* Inner loop uses 323 flops */
973 }
974
975 /* End of innermost loop */
976
977 gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
978 f+i_coord_offset,fshift+i_shift_offset);
979
980 ggid = gid[iidx];
981 /* Update potential energies */
982 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
983 gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
984
985 /* Increment number of inner iterations */
986 inneriter += j_index_end - j_index_start;
987
988 /* Outer loop uses 26 flops */
989 }
990
991 /* Increment number of outer iterations */
992 outeriter += nri;
993
994 /* Update outer/inner flops */
995
996 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*323);
997 }
998 /*
999 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_double
1000 * Electrostatics interaction: ReactionField
1001 * VdW interaction: LennardJones
1002 * Geometry: Water4-Water4
1003 * Calculate force/pot: Force
1004 */
1005 void
1006 nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_double
1007 (t_nblist * gmx_restrict nlist,
1008 rvec * gmx_restrict xx,
1009 rvec * gmx_restrict ff,
1010 t_forcerec * gmx_restrict fr,
1011 t_mdatoms * gmx_restrict mdatoms,
1012 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1013 t_nrnb * gmx_restrict nrnb)
1014 {
1015 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1016 * just 0 for non-waters.
1017 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1018 * jnr indices corresponding to data put in the four positions in the SIMD register.
1019 */
1020 int i_shift_offset,i_coord_offset,outeriter,inneriter;
1021 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1022 int jnrA,jnrB;
1023 int j_coord_offsetA,j_coord_offsetB;
1024 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
1025 real rcutoff_scalar;
1026 real *shiftvec,*fshift,*x,*f;
1027 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1028 int vdwioffset0;
1029 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1030 int vdwioffset1;
1031 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1032 int vdwioffset2;
1033 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1034 int vdwioffset3;
1035 __m128d ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1036 int vdwjidx0A,vdwjidx0B;
1037 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1038 int vdwjidx1A,vdwjidx1B;
1039 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1040 int vdwjidx2A,vdwjidx2B;
1041 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1042 int vdwjidx3A,vdwjidx3B;
1043 __m128d jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1044 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1045 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1046 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1047 __m128d dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1048 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1049 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1050 __m128d dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1051 __m128d dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1052 __m128d dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1053 __m128d dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1054 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
1055 real *charge;
1056 int nvdwtype;
1057 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1058 int *vdwtype;
1059 real *vdwparam;
1060 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
1061 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
1062 __m128d dummy_mask,cutoff_mask;
1063 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1064 __m128d one = _mm_set1_pd(1.0);
1065 __m128d two = _mm_set1_pd(2.0);
1066 x = xx[0];
1067 f = ff[0];
1068
1069 nri = nlist->nri;
1070 iinr = nlist->iinr;
1071 jindex = nlist->jindex;
1072 jjnr = nlist->jjnr;
1073 shiftidx = nlist->shift;
1074 gid = nlist->gid;
1075 shiftvec = fr->shift_vec[0];
1076 fshift = fr->fshift[0];
1077 facel = _mm_set1_pd(fr->epsfac);
1078 charge = mdatoms->chargeA;
1079 krf = _mm_set1_pd(fr->ic->k_rf);
1080 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
1081 crf = _mm_set1_pd(fr->ic->c_rf);
1082 nvdwtype = fr->ntype;
1083 vdwparam = fr->nbfp;
1084 vdwtype = mdatoms->typeA;
1085
1086 /* Setup water-specific parameters */
1087 inr = nlist->iinr[0];
1088 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1089 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1090 iq3 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1091 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1092
1093 jq1 = _mm_set1_pd(charge[inr+1]);
1094 jq2 = _mm_set1_pd(charge[inr+2]);
1095 jq3 = _mm_set1_pd(charge[inr+3]);
1096 vdwjidx0A = 2*vdwtype[inr+0];
1097 c6_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1098 c12_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1099 qq11 = _mm_mul_pd(iq1,jq1);
1100 qq12 = _mm_mul_pd(iq1,jq2);
1101 qq13 = _mm_mul_pd(iq1,jq3);
1102 qq21 = _mm_mul_pd(iq2,jq1);
1103 qq22 = _mm_mul_pd(iq2,jq2);
1104 qq23 = _mm_mul_pd(iq2,jq3);
1105 qq31 = _mm_mul_pd(iq3,jq1);
1106 qq32 = _mm_mul_pd(iq3,jq2);
1107 qq33 = _mm_mul_pd(iq3,jq3);
1108
1109 /* Avoid stupid compiler warnings */
1110 jnrA = jnrB = 0;
1111 j_coord_offsetA = 0;
1112 j_coord_offsetB = 0;
1113
1114 outeriter = 0;
1115 inneriter = 0;
1116
1117 /* Start outer loop over neighborlists */
1118 for(iidx=0; iidx<nri; iidx++)
1119 {
1120 /* Load shift vector for this list */
1121 i_shift_offset = DIM*shiftidx[iidx];
1122
1123 /* Load limits for loop over neighbors */
1124 j_index_start = jindex[iidx];
1125 j_index_end = jindex[iidx+1];
1126
1127 /* Get outer coordinate index */
1128 inr = iinr[iidx];
1129 i_coord_offset = DIM*inr;
1130
1131 /* Load i particle coords and add shift vector */
1132 gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1133 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1134
1135 fix0 = _mm_setzero_pd();
1136 fiy0 = _mm_setzero_pd();
1137 fiz0 = _mm_setzero_pd();
1138 fix1 = _mm_setzero_pd();
1139 fiy1 = _mm_setzero_pd();
1140 fiz1 = _mm_setzero_pd();
1141 fix2 = _mm_setzero_pd();
1142 fiy2 = _mm_setzero_pd();
1143 fiz2 = _mm_setzero_pd();
1144 fix3 = _mm_setzero_pd();
1145 fiy3 = _mm_setzero_pd();
1146 fiz3 = _mm_setzero_pd();
1147
1148 /* Start inner kernel loop */
1149 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1150 {
1151
1152 /* Get j neighbor index, and coordinate index */
1153 jnrA = jjnr[jidx];
1154 jnrB = jjnr[jidx+1];
1155 j_coord_offsetA = DIM*jnrA;
1156 j_coord_offsetB = DIM*jnrB;
1157
1158 /* load j atom coordinates */
1159 gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1160 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1161 &jy2,&jz2,&jx3,&jy3,&jz3);
1162
1163 /* Calculate displacement vector */
1164 dx00 = _mm_sub_pd(ix0,jx0);
1165 dy00 = _mm_sub_pd(iy0,jy0);
1166 dz00 = _mm_sub_pd(iz0,jz0);
1167 dx11 = _mm_sub_pd(ix1,jx1);
1168 dy11 = _mm_sub_pd(iy1,jy1);
1169 dz11 = _mm_sub_pd(iz1,jz1);
1170 dx12 = _mm_sub_pd(ix1,jx2);
1171 dy12 = _mm_sub_pd(iy1,jy2);
1172 dz12 = _mm_sub_pd(iz1,jz2);
1173 dx13 = _mm_sub_pd(ix1,jx3);
1174 dy13 = _mm_sub_pd(iy1,jy3);
1175 dz13 = _mm_sub_pd(iz1,jz3);
1176 dx21 = _mm_sub_pd(ix2,jx1);
1177 dy21 = _mm_sub_pd(iy2,jy1);
1178 dz21 = _mm_sub_pd(iz2,jz1);
1179 dx22 = _mm_sub_pd(ix2,jx2);
1180 dy22 = _mm_sub_pd(iy2,jy2);
1181 dz22 = _mm_sub_pd(iz2,jz2);
1182 dx23 = _mm_sub_pd(ix2,jx3);
1183 dy23 = _mm_sub_pd(iy2,jy3);
1184 dz23 = _mm_sub_pd(iz2,jz3);
1185 dx31 = _mm_sub_pd(ix3,jx1);
1186 dy31 = _mm_sub_pd(iy3,jy1);
1187 dz31 = _mm_sub_pd(iz3,jz1);
1188 dx32 = _mm_sub_pd(ix3,jx2);
1189 dy32 = _mm_sub_pd(iy3,jy2);
1190 dz32 = _mm_sub_pd(iz3,jz2);
1191 dx33 = _mm_sub_pd(ix3,jx3);
1192 dy33 = _mm_sub_pd(iy3,jy3);
1193 dz33 = _mm_sub_pd(iz3,jz3);
1194
1195 /* Calculate squared distance and things based on it */
1196 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1197 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1198 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1199 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1200 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1201 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1202 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1203 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1204 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1205 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1206
1207 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1208 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1209 rinv13 = gmx_mm_invsqrt_pd(rsq13);
1210 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1211 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1212 rinv23 = gmx_mm_invsqrt_pd(rsq23);
1213 rinv31 = gmx_mm_invsqrt_pd(rsq31);
1214 rinv32 = gmx_mm_invsqrt_pd(rsq32);
1215 rinv33 = gmx_mm_invsqrt_pd(rsq33);
1216
1217 rinvsq00 = gmx_mm_inv_pd(rsq00);
1218 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1219 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1220 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
1221 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1222 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1223 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
1224 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
1225 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
1226 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
1227
1228 fjx0 = _mm_setzero_pd();
1229 fjy0 = _mm_setzero_pd();
1230 fjz0 = _mm_setzero_pd();
1231 fjx1 = _mm_setzero_pd();
1232 fjy1 = _mm_setzero_pd();
1233 fjz1 = _mm_setzero_pd();
1234 fjx2 = _mm_setzero_pd();
1235 fjy2 = _mm_setzero_pd();
1236 fjz2 = _mm_setzero_pd();
1237 fjx3 = _mm_setzero_pd();
1238 fjy3 = _mm_setzero_pd();
1239 fjz3 = _mm_setzero_pd();
1240
1241 /**************************
1242 * CALCULATE INTERACTIONS *
1243 **************************/
1244
1245 /* LENNARD-JONES DISPERSION/REPULSION */
1246
1247 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1248 fvdw = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1249
1250 fscal = fvdw;
1251
1252 /* Calculate temporary vectorial force */
1253 tx = _mm_mul_pd(fscal,dx00);
1254 ty = _mm_mul_pd(fscal,dy00);
1255 tz = _mm_mul_pd(fscal,dz00);
1256
1257 /* Update vectorial force */
1258 fix0 = _mm_add_pd(fix0,tx);
1259 fiy0 = _mm_add_pd(fiy0,ty);
1260 fiz0 = _mm_add_pd(fiz0,tz);
1261
1262 fjx0 = _mm_add_pd(fjx0,tx);
1263 fjy0 = _mm_add_pd(fjy0,ty);
1264 fjz0 = _mm_add_pd(fjz0,tz);
1265
1266 /**************************
1267 * CALCULATE INTERACTIONS *
1268 **************************/
1269
1270 /* REACTION-FIELD ELECTROSTATICS */
1271 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1272
1273 fscal = felec;
1274
1275 /* Calculate temporary vectorial force */
1276 tx = _mm_mul_pd(fscal,dx11);
1277 ty = _mm_mul_pd(fscal,dy11);
1278 tz = _mm_mul_pd(fscal,dz11);
1279
1280 /* Update vectorial force */
1281 fix1 = _mm_add_pd(fix1,tx);
1282 fiy1 = _mm_add_pd(fiy1,ty);
1283 fiz1 = _mm_add_pd(fiz1,tz);
1284
1285 fjx1 = _mm_add_pd(fjx1,tx);
1286 fjy1 = _mm_add_pd(fjy1,ty);
1287 fjz1 = _mm_add_pd(fjz1,tz);
1288
1289 /**************************
1290 * CALCULATE INTERACTIONS *
1291 **************************/
1292
1293 /* REACTION-FIELD ELECTROSTATICS */
1294 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1295
1296 fscal = felec;
1297
1298 /* Calculate temporary vectorial force */
1299 tx = _mm_mul_pd(fscal,dx12);
1300 ty = _mm_mul_pd(fscal,dy12);
1301 tz = _mm_mul_pd(fscal,dz12);
1302
1303 /* Update vectorial force */
1304 fix1 = _mm_add_pd(fix1,tx);
1305 fiy1 = _mm_add_pd(fiy1,ty);
1306 fiz1 = _mm_add_pd(fiz1,tz);
1307
1308 fjx2 = _mm_add_pd(fjx2,tx);
1309 fjy2 = _mm_add_pd(fjy2,ty);
1310 fjz2 = _mm_add_pd(fjz2,tz);
1311
1312 /**************************
1313 * CALCULATE INTERACTIONS *
1314 **************************/
1315
1316 /* REACTION-FIELD ELECTROSTATICS */
1317 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1318
1319 fscal = felec;
1320
1321 /* Calculate temporary vectorial force */
1322 tx = _mm_mul_pd(fscal,dx13);
1323 ty = _mm_mul_pd(fscal,dy13);
1324 tz = _mm_mul_pd(fscal,dz13);
1325
1326 /* Update vectorial force */
1327 fix1 = _mm_add_pd(fix1,tx);
1328 fiy1 = _mm_add_pd(fiy1,ty);
1329 fiz1 = _mm_add_pd(fiz1,tz);
1330
1331 fjx3 = _mm_add_pd(fjx3,tx);
1332 fjy3 = _mm_add_pd(fjy3,ty);
1333 fjz3 = _mm_add_pd(fjz3,tz);
1334
1335 /**************************
1336 * CALCULATE INTERACTIONS *
1337 **************************/
1338
1339 /* REACTION-FIELD ELECTROSTATICS */
1340 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1341
1342 fscal = felec;
1343
1344 /* Calculate temporary vectorial force */
1345 tx = _mm_mul_pd(fscal,dx21);
1346 ty = _mm_mul_pd(fscal,dy21);
1347 tz = _mm_mul_pd(fscal,dz21);
1348
1349 /* Update vectorial force */
1350 fix2 = _mm_add_pd(fix2,tx);
1351 fiy2 = _mm_add_pd(fiy2,ty);
1352 fiz2 = _mm_add_pd(fiz2,tz);
1353
1354 fjx1 = _mm_add_pd(fjx1,tx);
1355 fjy1 = _mm_add_pd(fjy1,ty);
1356 fjz1 = _mm_add_pd(fjz1,tz);
1357
1358 /**************************
1359 * CALCULATE INTERACTIONS *
1360 **************************/
1361
1362 /* REACTION-FIELD ELECTROSTATICS */
1363 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1364
1365 fscal = felec;
1366
1367 /* Calculate temporary vectorial force */
1368 tx = _mm_mul_pd(fscal,dx22);
1369 ty = _mm_mul_pd(fscal,dy22);
1370 tz = _mm_mul_pd(fscal,dz22);
1371
1372 /* Update vectorial force */
1373 fix2 = _mm_add_pd(fix2,tx);
1374 fiy2 = _mm_add_pd(fiy2,ty);
1375 fiz2 = _mm_add_pd(fiz2,tz);
1376
1377 fjx2 = _mm_add_pd(fjx2,tx);
1378 fjy2 = _mm_add_pd(fjy2,ty);
1379 fjz2 = _mm_add_pd(fjz2,tz);
1380
1381 /**************************
1382 * CALCULATE INTERACTIONS *
1383 **************************/
1384
1385 /* REACTION-FIELD ELECTROSTATICS */
1386 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1387
1388 fscal = felec;
1389
1390 /* Calculate temporary vectorial force */
1391 tx = _mm_mul_pd(fscal,dx23);
1392 ty = _mm_mul_pd(fscal,dy23);
1393 tz = _mm_mul_pd(fscal,dz23);
1394
1395 /* Update vectorial force */
1396 fix2 = _mm_add_pd(fix2,tx);
1397 fiy2 = _mm_add_pd(fiy2,ty);
1398 fiz2 = _mm_add_pd(fiz2,tz);
1399
1400 fjx3 = _mm_add_pd(fjx3,tx);
1401 fjy3 = _mm_add_pd(fjy3,ty);
1402 fjz3 = _mm_add_pd(fjz3,tz);
1403
1404 /**************************
1405 * CALCULATE INTERACTIONS *
1406 **************************/
1407
1408 /* REACTION-FIELD ELECTROSTATICS */
1409 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1410
1411 fscal = felec;
1412
1413 /* Calculate temporary vectorial force */
1414 tx = _mm_mul_pd(fscal,dx31);
1415 ty = _mm_mul_pd(fscal,dy31);
1416 tz = _mm_mul_pd(fscal,dz31);
1417
1418 /* Update vectorial force */
1419 fix3 = _mm_add_pd(fix3,tx);
1420 fiy3 = _mm_add_pd(fiy3,ty);
1421 fiz3 = _mm_add_pd(fiz3,tz);
1422
1423 fjx1 = _mm_add_pd(fjx1,tx);
1424 fjy1 = _mm_add_pd(fjy1,ty);
1425 fjz1 = _mm_add_pd(fjz1,tz);
1426
1427 /**************************
1428 * CALCULATE INTERACTIONS *
1429 **************************/
1430
1431 /* REACTION-FIELD ELECTROSTATICS */
1432 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1433
1434 fscal = felec;
1435
1436 /* Calculate temporary vectorial force */
1437 tx = _mm_mul_pd(fscal,dx32);
1438 ty = _mm_mul_pd(fscal,dy32);
1439 tz = _mm_mul_pd(fscal,dz32);
1440
1441 /* Update vectorial force */
1442 fix3 = _mm_add_pd(fix3,tx);
1443 fiy3 = _mm_add_pd(fiy3,ty);
1444 fiz3 = _mm_add_pd(fiz3,tz);
1445
1446 fjx2 = _mm_add_pd(fjx2,tx);
1447 fjy2 = _mm_add_pd(fjy2,ty);
1448 fjz2 = _mm_add_pd(fjz2,tz);
1449
1450 /**************************
1451 * CALCULATE INTERACTIONS *
1452 **************************/
1453
1454 /* REACTION-FIELD ELECTROSTATICS */
1455 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1456
1457 fscal = felec;
1458
1459 /* Calculate temporary vectorial force */
1460 tx = _mm_mul_pd(fscal,dx33);
1461 ty = _mm_mul_pd(fscal,dy33);
1462 tz = _mm_mul_pd(fscal,dz33);
1463
1464 /* Update vectorial force */
1465 fix3 = _mm_add_pd(fix3,tx);
1466 fiy3 = _mm_add_pd(fiy3,ty);
1467 fiz3 = _mm_add_pd(fiz3,tz);
1468
1469 fjx3 = _mm_add_pd(fjx3,tx);
1470 fjy3 = _mm_add_pd(fjy3,ty);
1471 fjz3 = _mm_add_pd(fjz3,tz);
1472
1473 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1474
1475 /* Inner loop uses 273 flops */
1476 }
1477
1478 if(jidx<j_index_end)
1479 {
1480
1481 jnrA = jjnr[jidx];
1482 j_coord_offsetA = DIM*jnrA;
1483
1484 /* load j atom coordinates */
1485 gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1486 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1487 &jy2,&jz2,&jx3,&jy3,&jz3);
1488
1489 /* Calculate displacement vector */
1490 dx00 = _mm_sub_pd(ix0,jx0);
1491 dy00 = _mm_sub_pd(iy0,jy0);
1492 dz00 = _mm_sub_pd(iz0,jz0);
1493 dx11 = _mm_sub_pd(ix1,jx1);
1494 dy11 = _mm_sub_pd(iy1,jy1);
1495 dz11 = _mm_sub_pd(iz1,jz1);
1496 dx12 = _mm_sub_pd(ix1,jx2);
1497 dy12 = _mm_sub_pd(iy1,jy2);
1498 dz12 = _mm_sub_pd(iz1,jz2);
1499 dx13 = _mm_sub_pd(ix1,jx3);
1500 dy13 = _mm_sub_pd(iy1,jy3);
1501 dz13 = _mm_sub_pd(iz1,jz3);
1502 dx21 = _mm_sub_pd(ix2,jx1);
1503 dy21 = _mm_sub_pd(iy2,jy1);
1504 dz21 = _mm_sub_pd(iz2,jz1);
1505 dx22 = _mm_sub_pd(ix2,jx2);
1506 dy22 = _mm_sub_pd(iy2,jy2);
1507 dz22 = _mm_sub_pd(iz2,jz2);
1508 dx23 = _mm_sub_pd(ix2,jx3);
1509 dy23 = _mm_sub_pd(iy2,jy3);
1510 dz23 = _mm_sub_pd(iz2,jz3);
1511 dx31 = _mm_sub_pd(ix3,jx1);
1512 dy31 = _mm_sub_pd(iy3,jy1);
1513 dz31 = _mm_sub_pd(iz3,jz1);
1514 dx32 = _mm_sub_pd(ix3,jx2);
1515 dy32 = _mm_sub_pd(iy3,jy2);
1516 dz32 = _mm_sub_pd(iz3,jz2);
1517 dx33 = _mm_sub_pd(ix3,jx3);
1518 dy33 = _mm_sub_pd(iy3,jy3);
1519 dz33 = _mm_sub_pd(iz3,jz3);
1520
1521 /* Calculate squared distance and things based on it */
1522 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1523 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1524 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1525 rsq13 = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1526 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1527 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1528 rsq23 = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1529 rsq31 = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1530 rsq32 = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1531 rsq33 = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1532
1533 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1534 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1535 rinv13 = gmx_mm_invsqrt_pd(rsq13);
1536 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1537 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1538 rinv23 = gmx_mm_invsqrt_pd(rsq23);
1539 rinv31 = gmx_mm_invsqrt_pd(rsq31);
1540 rinv32 = gmx_mm_invsqrt_pd(rsq32);
1541 rinv33 = gmx_mm_invsqrt_pd(rsq33);
1542
1543 rinvsq00 = gmx_mm_inv_pd(rsq00);
1544 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1545 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1546 rinvsq13 = _mm_mul_pd(rinv13,rinv13);
1547 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1548 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1549 rinvsq23 = _mm_mul_pd(rinv23,rinv23);
1550 rinvsq31 = _mm_mul_pd(rinv31,rinv31);
1551 rinvsq32 = _mm_mul_pd(rinv32,rinv32);
1552 rinvsq33 = _mm_mul_pd(rinv33,rinv33);
1553
1554 fjx0 = _mm_setzero_pd();
1555 fjy0 = _mm_setzero_pd();
1556 fjz0 = _mm_setzero_pd();
1557 fjx1 = _mm_setzero_pd();
1558 fjy1 = _mm_setzero_pd();
1559 fjz1 = _mm_setzero_pd();
1560 fjx2 = _mm_setzero_pd();
1561 fjy2 = _mm_setzero_pd();
1562 fjz2 = _mm_setzero_pd();
1563 fjx3 = _mm_setzero_pd();
1564 fjy3 = _mm_setzero_pd();
1565 fjz3 = _mm_setzero_pd();
1566
1567 /**************************
1568 * CALCULATE INTERACTIONS *
1569 **************************/
1570
1571 /* LENNARD-JONES DISPERSION/REPULSION */
1572
1573 rinvsix = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1574 fvdw = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1575
1576 fscal = fvdw;
1577
1578 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1579
1580 /* Calculate temporary vectorial force */
1581 tx = _mm_mul_pd(fscal,dx00);
1582 ty = _mm_mul_pd(fscal,dy00);
1583 tz = _mm_mul_pd(fscal,dz00);
1584
1585 /* Update vectorial force */
1586 fix0 = _mm_add_pd(fix0,tx);
1587 fiy0 = _mm_add_pd(fiy0,ty);
1588 fiz0 = _mm_add_pd(fiz0,tz);
1589
1590 fjx0 = _mm_add_pd(fjx0,tx);
1591 fjy0 = _mm_add_pd(fjy0,ty);
1592 fjz0 = _mm_add_pd(fjz0,tz);
1593
1594 /**************************
1595 * CALCULATE INTERACTIONS *
1596 **************************/
1597
1598 /* REACTION-FIELD ELECTROSTATICS */
1599 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1600
1601 fscal = felec;
1602
1603 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1604
1605 /* Calculate temporary vectorial force */
1606 tx = _mm_mul_pd(fscal,dx11);
1607 ty = _mm_mul_pd(fscal,dy11);
1608 tz = _mm_mul_pd(fscal,dz11);
1609
1610 /* Update vectorial force */
1611 fix1 = _mm_add_pd(fix1,tx);
1612 fiy1 = _mm_add_pd(fiy1,ty);
1613 fiz1 = _mm_add_pd(fiz1,tz);
1614
1615 fjx1 = _mm_add_pd(fjx1,tx);
1616 fjy1 = _mm_add_pd(fjy1,ty);
1617 fjz1 = _mm_add_pd(fjz1,tz);
1618
1619 /**************************
1620 * CALCULATE INTERACTIONS *
1621 **************************/
1622
1623 /* REACTION-FIELD ELECTROSTATICS */
1624 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1625
1626 fscal = felec;
1627
1628 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1629
1630 /* Calculate temporary vectorial force */
1631 tx = _mm_mul_pd(fscal,dx12);
1632 ty = _mm_mul_pd(fscal,dy12);
1633 tz = _mm_mul_pd(fscal,dz12);
1634
1635 /* Update vectorial force */
1636 fix1 = _mm_add_pd(fix1,tx);
1637 fiy1 = _mm_add_pd(fiy1,ty);
1638 fiz1 = _mm_add_pd(fiz1,tz);
1639
1640 fjx2 = _mm_add_pd(fjx2,tx);
1641 fjy2 = _mm_add_pd(fjy2,ty);
1642 fjz2 = _mm_add_pd(fjz2,tz);
1643
1644 /**************************
1645 * CALCULATE INTERACTIONS *
1646 **************************/
1647
1648 /* REACTION-FIELD ELECTROSTATICS */
1649 felec = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1650
1651 fscal = felec;
1652
1653 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1654
1655 /* Calculate temporary vectorial force */
1656 tx = _mm_mul_pd(fscal,dx13);
1657 ty = _mm_mul_pd(fscal,dy13);
1658 tz = _mm_mul_pd(fscal,dz13);
1659
1660 /* Update vectorial force */
1661 fix1 = _mm_add_pd(fix1,tx);
1662 fiy1 = _mm_add_pd(fiy1,ty);
1663 fiz1 = _mm_add_pd(fiz1,tz);
1664
1665 fjx3 = _mm_add_pd(fjx3,tx);
1666 fjy3 = _mm_add_pd(fjy3,ty);
1667 fjz3 = _mm_add_pd(fjz3,tz);
1668
1669 /**************************
1670 * CALCULATE INTERACTIONS *
1671 **************************/
1672
1673 /* REACTION-FIELD ELECTROSTATICS */
1674 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1675
1676 fscal = felec;
1677
1678 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1679
1680 /* Calculate temporary vectorial force */
1681 tx = _mm_mul_pd(fscal,dx21);
1682 ty = _mm_mul_pd(fscal,dy21);
1683 tz = _mm_mul_pd(fscal,dz21);
1684
1685 /* Update vectorial force */
1686 fix2 = _mm_add_pd(fix2,tx);
1687 fiy2 = _mm_add_pd(fiy2,ty);
1688 fiz2 = _mm_add_pd(fiz2,tz);
1689
1690 fjx1 = _mm_add_pd(fjx1,tx);
1691 fjy1 = _mm_add_pd(fjy1,ty);
1692 fjz1 = _mm_add_pd(fjz1,tz);
1693
1694 /**************************
1695 * CALCULATE INTERACTIONS *
1696 **************************/
1697
1698 /* REACTION-FIELD ELECTROSTATICS */
1699 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1700
1701 fscal = felec;
1702
1703 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1704
1705 /* Calculate temporary vectorial force */
1706 tx = _mm_mul_pd(fscal,dx22);
1707 ty = _mm_mul_pd(fscal,dy22);
1708 tz = _mm_mul_pd(fscal,dz22);
1709
1710 /* Update vectorial force */
1711 fix2 = _mm_add_pd(fix2,tx);
1712 fiy2 = _mm_add_pd(fiy2,ty);
1713 fiz2 = _mm_add_pd(fiz2,tz);
1714
1715 fjx2 = _mm_add_pd(fjx2,tx);
1716 fjy2 = _mm_add_pd(fjy2,ty);
1717 fjz2 = _mm_add_pd(fjz2,tz);
1718
1719 /**************************
1720 * CALCULATE INTERACTIONS *
1721 **************************/
1722
1723 /* REACTION-FIELD ELECTROSTATICS */
1724 felec = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1725
1726 fscal = felec;
1727
1728 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1729
1730 /* Calculate temporary vectorial force */
1731 tx = _mm_mul_pd(fscal,dx23);
1732 ty = _mm_mul_pd(fscal,dy23);
1733 tz = _mm_mul_pd(fscal,dz23);
1734
1735 /* Update vectorial force */
1736 fix2 = _mm_add_pd(fix2,tx);
1737 fiy2 = _mm_add_pd(fiy2,ty);
1738 fiz2 = _mm_add_pd(fiz2,tz);
1739
1740 fjx3 = _mm_add_pd(fjx3,tx);
1741 fjy3 = _mm_add_pd(fjy3,ty);
1742 fjz3 = _mm_add_pd(fjz3,tz);
1743
1744 /**************************
1745 * CALCULATE INTERACTIONS *
1746 **************************/
1747
1748 /* REACTION-FIELD ELECTROSTATICS */
1749 felec = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1750
1751 fscal = felec;
1752
1753 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1754
1755 /* Calculate temporary vectorial force */
1756 tx = _mm_mul_pd(fscal,dx31);
1757 ty = _mm_mul_pd(fscal,dy31);
1758 tz = _mm_mul_pd(fscal,dz31);
1759
1760 /* Update vectorial force */
1761 fix3 = _mm_add_pd(fix3,tx);
1762 fiy3 = _mm_add_pd(fiy3,ty);
1763 fiz3 = _mm_add_pd(fiz3,tz);
1764
1765 fjx1 = _mm_add_pd(fjx1,tx);
1766 fjy1 = _mm_add_pd(fjy1,ty);
1767 fjz1 = _mm_add_pd(fjz1,tz);
1768
1769 /**************************
1770 * CALCULATE INTERACTIONS *
1771 **************************/
1772
1773 /* REACTION-FIELD ELECTROSTATICS */
1774 felec = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1775
1776 fscal = felec;
1777
1778 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1779
1780 /* Calculate temporary vectorial force */
1781 tx = _mm_mul_pd(fscal,dx32);
1782 ty = _mm_mul_pd(fscal,dy32);
1783 tz = _mm_mul_pd(fscal,dz32);
1784
1785 /* Update vectorial force */
1786 fix3 = _mm_add_pd(fix3,tx);
1787 fiy3 = _mm_add_pd(fiy3,ty);
1788 fiz3 = _mm_add_pd(fiz3,tz);
1789
1790 fjx2 = _mm_add_pd(fjx2,tx);
1791 fjy2 = _mm_add_pd(fjy2,ty);
1792 fjz2 = _mm_add_pd(fjz2,tz);
1793
1794 /**************************
1795 * CALCULATE INTERACTIONS *
1796 **************************/
1797
1798 /* REACTION-FIELD ELECTROSTATICS */
1799 felec = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1800
1801 fscal = felec;
1802
1803 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1804
1805 /* Calculate temporary vectorial force */
1806 tx = _mm_mul_pd(fscal,dx33);
1807 ty = _mm_mul_pd(fscal,dy33);
1808 tz = _mm_mul_pd(fscal,dz33);
1809
1810 /* Update vectorial force */
1811 fix3 = _mm_add_pd(fix3,tx);
1812 fiy3 = _mm_add_pd(fiy3,ty);
1813 fiz3 = _mm_add_pd(fiz3,tz);
1814
1815 fjx3 = _mm_add_pd(fjx3,tx);
1816 fjy3 = _mm_add_pd(fjy3,ty);
1817 fjz3 = _mm_add_pd(fjz3,tz);
1818
1819 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1820
1821 /* Inner loop uses 273 flops */
1822 }
1823
1824 /* End of innermost loop */
1825
1826 gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1827 f+i_coord_offset,fshift+i_shift_offset);
1828
1829 /* Increment number of inner iterations */
1830 inneriter += j_index_end - j_index_start;
1831
1832 /* Outer loop uses 24 flops */
1833 }
1834
1835 /* Increment number of outer iterations */
1836 outeriter += nri;
1837
1838 /* Update outer/inner flops */
1839
1840 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1841 }
The ultimate molecular dynamics simulation package