?  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_VdwCSTab_GeomW3W3_sse2_double.c
blobad7680c616f2054b0caa496be0315faffd8edc89
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_VdwCSTab_GeomW3W3_VF_sse2_double
53 * Electrostatics interaction: ReactionField
54 * VdW interaction: CubicSplineTable
55 * Geometry: Water3-Water3
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_sse2_double
60 (t_nblist * gmx_restrict nlist,
61 rvec * gmx_restrict xx,
62 rvec * gmx_restrict ff,
63 t_forcerec * gmx_restrict fr,
64 t_mdatoms * gmx_restrict mdatoms,
65 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
66 t_nrnb * gmx_restrict nrnb)
67 {
68 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
69 * just 0 for non-waters.
70 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
71 * jnr indices corresponding to data put in the four positions in the SIMD register.
72 */
73 int i_shift_offset,i_coord_offset,outeriter,inneriter;
74 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
75 int jnrA,jnrB;
76 int j_coord_offsetA,j_coord_offsetB;
77 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
78 real rcutoff_scalar;
79 real *shiftvec,*fshift,*x,*f;
80 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
81 int vdwioffset0;
82 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
83 int vdwioffset1;
84 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
85 int vdwioffset2;
86 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
87 int vdwjidx0A,vdwjidx0B;
88 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
89 int vdwjidx1A,vdwjidx1B;
90 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
91 int vdwjidx2A,vdwjidx2B;
92 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
93 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
94 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
95 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
96 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
97 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
98 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
99 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
100 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
101 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
102 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
103 real *charge;
104 int nvdwtype;
105 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
106 int *vdwtype;
107 real *vdwparam;
108 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
109 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
110 __m128i vfitab;
111 __m128i ifour = _mm_set1_epi32(4);
112 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
113 real *vftab;
114 __m128d dummy_mask,cutoff_mask;
115 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
116 __m128d one = _mm_set1_pd(1.0);
117 __m128d two = _mm_set1_pd(2.0);
118 x = xx[0];
119 f = ff[0];
120
121 nri = nlist->nri;
122 iinr = nlist->iinr;
123 jindex = nlist->jindex;
124 jjnr = nlist->jjnr;
125 shiftidx = nlist->shift;
126 gid = nlist->gid;
127 shiftvec = fr->shift_vec[0];
128 fshift = fr->fshift[0];
129 facel = _mm_set1_pd(fr->epsfac);
130 charge = mdatoms->chargeA;
131 krf = _mm_set1_pd(fr->ic->k_rf);
132 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
133 crf = _mm_set1_pd(fr->ic->c_rf);
134 nvdwtype = fr->ntype;
135 vdwparam = fr->nbfp;
136 vdwtype = mdatoms->typeA;
137
138 vftab = kernel_data->table_vdw->data;
139 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
140
141 /* Setup water-specific parameters */
142 inr = nlist->iinr[0];
143 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
144 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
145 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
146 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
147
148 jq0 = _mm_set1_pd(charge[inr+0]);
149 jq1 = _mm_set1_pd(charge[inr+1]);
150 jq2 = _mm_set1_pd(charge[inr+2]);
151 vdwjidx0A = 2*vdwtype[inr+0];
152 qq00 = _mm_mul_pd(iq0,jq0);
153 c6_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
154 c12_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
155 qq01 = _mm_mul_pd(iq0,jq1);
156 qq02 = _mm_mul_pd(iq0,jq2);
157 qq10 = _mm_mul_pd(iq1,jq0);
158 qq11 = _mm_mul_pd(iq1,jq1);
159 qq12 = _mm_mul_pd(iq1,jq2);
160 qq20 = _mm_mul_pd(iq2,jq0);
161 qq21 = _mm_mul_pd(iq2,jq1);
162 qq22 = _mm_mul_pd(iq2,jq2);
163
164 /* Avoid stupid compiler warnings */
165 jnrA = jnrB = 0;
166 j_coord_offsetA = 0;
167 j_coord_offsetB = 0;
168
169 outeriter = 0;
170 inneriter = 0;
171
172 /* Start outer loop over neighborlists */
173 for(iidx=0; iidx<nri; iidx++)
174 {
175 /* Load shift vector for this list */
176 i_shift_offset = DIM*shiftidx[iidx];
177
178 /* Load limits for loop over neighbors */
179 j_index_start = jindex[iidx];
180 j_index_end = jindex[iidx+1];
181
182 /* Get outer coordinate index */
183 inr = iinr[iidx];
184 i_coord_offset = DIM*inr;
185
186 /* Load i particle coords and add shift vector */
187 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
188 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
189
190 fix0 = _mm_setzero_pd();
191 fiy0 = _mm_setzero_pd();
192 fiz0 = _mm_setzero_pd();
193 fix1 = _mm_setzero_pd();
194 fiy1 = _mm_setzero_pd();
195 fiz1 = _mm_setzero_pd();
196 fix2 = _mm_setzero_pd();
197 fiy2 = _mm_setzero_pd();
198 fiz2 = _mm_setzero_pd();
199
200 /* Reset potential sums */
201 velecsum = _mm_setzero_pd();
202 vvdwsum = _mm_setzero_pd();
203
204 /* Start inner kernel loop */
205 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
206 {
207
208 /* Get j neighbor index, and coordinate index */
209 jnrA = jjnr[jidx];
210 jnrB = jjnr[jidx+1];
211 j_coord_offsetA = DIM*jnrA;
212 j_coord_offsetB = DIM*jnrB;
213
214 /* load j atom coordinates */
215 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
216 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
217
218 /* Calculate displacement vector */
219 dx00 = _mm_sub_pd(ix0,jx0);
220 dy00 = _mm_sub_pd(iy0,jy0);
221 dz00 = _mm_sub_pd(iz0,jz0);
222 dx01 = _mm_sub_pd(ix0,jx1);
223 dy01 = _mm_sub_pd(iy0,jy1);
224 dz01 = _mm_sub_pd(iz0,jz1);
225 dx02 = _mm_sub_pd(ix0,jx2);
226 dy02 = _mm_sub_pd(iy0,jy2);
227 dz02 = _mm_sub_pd(iz0,jz2);
228 dx10 = _mm_sub_pd(ix1,jx0);
229 dy10 = _mm_sub_pd(iy1,jy0);
230 dz10 = _mm_sub_pd(iz1,jz0);
231 dx11 = _mm_sub_pd(ix1,jx1);
232 dy11 = _mm_sub_pd(iy1,jy1);
233 dz11 = _mm_sub_pd(iz1,jz1);
234 dx12 = _mm_sub_pd(ix1,jx2);
235 dy12 = _mm_sub_pd(iy1,jy2);
236 dz12 = _mm_sub_pd(iz1,jz2);
237 dx20 = _mm_sub_pd(ix2,jx0);
238 dy20 = _mm_sub_pd(iy2,jy0);
239 dz20 = _mm_sub_pd(iz2,jz0);
240 dx21 = _mm_sub_pd(ix2,jx1);
241 dy21 = _mm_sub_pd(iy2,jy1);
242 dz21 = _mm_sub_pd(iz2,jz1);
243 dx22 = _mm_sub_pd(ix2,jx2);
244 dy22 = _mm_sub_pd(iy2,jy2);
245 dz22 = _mm_sub_pd(iz2,jz2);
246
247 /* Calculate squared distance and things based on it */
248 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
249 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
250 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
251 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
252 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
253 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
254 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
255 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
256 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
257
258 rinv00 = gmx_mm_invsqrt_pd(rsq00);
259 rinv01 = gmx_mm_invsqrt_pd(rsq01);
260 rinv02 = gmx_mm_invsqrt_pd(rsq02);
261 rinv10 = gmx_mm_invsqrt_pd(rsq10);
262 rinv11 = gmx_mm_invsqrt_pd(rsq11);
263 rinv12 = gmx_mm_invsqrt_pd(rsq12);
264 rinv20 = gmx_mm_invsqrt_pd(rsq20);
265 rinv21 = gmx_mm_invsqrt_pd(rsq21);
266 rinv22 = gmx_mm_invsqrt_pd(rsq22);
267
268 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
269 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
270 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
271 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
272 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
273 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
274 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
275 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
276 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
277
278 fjx0 = _mm_setzero_pd();
279 fjy0 = _mm_setzero_pd();
280 fjz0 = _mm_setzero_pd();
281 fjx1 = _mm_setzero_pd();
282 fjy1 = _mm_setzero_pd();
283 fjz1 = _mm_setzero_pd();
284 fjx2 = _mm_setzero_pd();
285 fjy2 = _mm_setzero_pd();
286 fjz2 = _mm_setzero_pd();
287
288 /**************************
289 * CALCULATE INTERACTIONS *
290 **************************/
291
292 r00 = _mm_mul_pd(rsq00,rinv00);
293
294 /* Calculate table index by multiplying r with table scale and truncate to integer */
295 rt = _mm_mul_pd(r00,vftabscale);
296 vfitab = _mm_cvttpd_epi32(rt);
297 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
298 vfitab = _mm_slli_epi32(vfitab,3);
299
300 /* REACTION-FIELD ELECTROSTATICS */
301 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
302 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
303
304 /* CUBIC SPLINE TABLE DISPERSION */
305 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
306 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
307 GMX_MM_TRANSPOSE2_PD(Y,F);
308 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
309 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
310 GMX_MM_TRANSPOSE2_PD(G,H);
311 Heps = _mm_mul_pd(vfeps,H);
312 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
313 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
314 vvdw6 = _mm_mul_pd(c6_00,VV);
315 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
316 fvdw6 = _mm_mul_pd(c6_00,FF);
317
318 /* CUBIC SPLINE TABLE REPULSION */
319 vfitab = _mm_add_epi32(vfitab,ifour);
320 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
321 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
322 GMX_MM_TRANSPOSE2_PD(Y,F);
323 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
324 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
325 GMX_MM_TRANSPOSE2_PD(G,H);
326 Heps = _mm_mul_pd(vfeps,H);
327 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
328 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
329 vvdw12 = _mm_mul_pd(c12_00,VV);
330 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
331 fvdw12 = _mm_mul_pd(c12_00,FF);
332 vvdw = _mm_add_pd(vvdw12,vvdw6);
333 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
334
335 /* Update potential sum for this i atom from the interaction with this j atom. */
336 velecsum = _mm_add_pd(velecsum,velec);
337 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
338
339 fscal = _mm_add_pd(felec,fvdw);
340
341 /* Calculate temporary vectorial force */
342 tx = _mm_mul_pd(fscal,dx00);
343 ty = _mm_mul_pd(fscal,dy00);
344 tz = _mm_mul_pd(fscal,dz00);
345
346 /* Update vectorial force */
347 fix0 = _mm_add_pd(fix0,tx);
348 fiy0 = _mm_add_pd(fiy0,ty);
349 fiz0 = _mm_add_pd(fiz0,tz);
350
351 fjx0 = _mm_add_pd(fjx0,tx);
352 fjy0 = _mm_add_pd(fjy0,ty);
353 fjz0 = _mm_add_pd(fjz0,tz);
354
355 /**************************
356 * CALCULATE INTERACTIONS *
357 **************************/
358
359 /* REACTION-FIELD ELECTROSTATICS */
360 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
361 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
362
363 /* Update potential sum for this i atom from the interaction with this j atom. */
364 velecsum = _mm_add_pd(velecsum,velec);
365
366 fscal = felec;
367
368 /* Calculate temporary vectorial force */
369 tx = _mm_mul_pd(fscal,dx01);
370 ty = _mm_mul_pd(fscal,dy01);
371 tz = _mm_mul_pd(fscal,dz01);
372
373 /* Update vectorial force */
374 fix0 = _mm_add_pd(fix0,tx);
375 fiy0 = _mm_add_pd(fiy0,ty);
376 fiz0 = _mm_add_pd(fiz0,tz);
377
378 fjx1 = _mm_add_pd(fjx1,tx);
379 fjy1 = _mm_add_pd(fjy1,ty);
380 fjz1 = _mm_add_pd(fjz1,tz);
381
382 /**************************
383 * CALCULATE INTERACTIONS *
384 **************************/
385
386 /* REACTION-FIELD ELECTROSTATICS */
387 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
388 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
389
390 /* Update potential sum for this i atom from the interaction with this j atom. */
391 velecsum = _mm_add_pd(velecsum,velec);
392
393 fscal = felec;
394
395 /* Calculate temporary vectorial force */
396 tx = _mm_mul_pd(fscal,dx02);
397 ty = _mm_mul_pd(fscal,dy02);
398 tz = _mm_mul_pd(fscal,dz02);
399
400 /* Update vectorial force */
401 fix0 = _mm_add_pd(fix0,tx);
402 fiy0 = _mm_add_pd(fiy0,ty);
403 fiz0 = _mm_add_pd(fiz0,tz);
404
405 fjx2 = _mm_add_pd(fjx2,tx);
406 fjy2 = _mm_add_pd(fjy2,ty);
407 fjz2 = _mm_add_pd(fjz2,tz);
408
409 /**************************
410 * CALCULATE INTERACTIONS *
411 **************************/
412
413 /* REACTION-FIELD ELECTROSTATICS */
414 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
415 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
416
417 /* Update potential sum for this i atom from the interaction with this j atom. */
418 velecsum = _mm_add_pd(velecsum,velec);
419
420 fscal = felec;
421
422 /* Calculate temporary vectorial force */
423 tx = _mm_mul_pd(fscal,dx10);
424 ty = _mm_mul_pd(fscal,dy10);
425 tz = _mm_mul_pd(fscal,dz10);
426
427 /* Update vectorial force */
428 fix1 = _mm_add_pd(fix1,tx);
429 fiy1 = _mm_add_pd(fiy1,ty);
430 fiz1 = _mm_add_pd(fiz1,tz);
431
432 fjx0 = _mm_add_pd(fjx0,tx);
433 fjy0 = _mm_add_pd(fjy0,ty);
434 fjz0 = _mm_add_pd(fjz0,tz);
435
436 /**************************
437 * CALCULATE INTERACTIONS *
438 **************************/
439
440 /* REACTION-FIELD ELECTROSTATICS */
441 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
442 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
443
444 /* Update potential sum for this i atom from the interaction with this j atom. */
445 velecsum = _mm_add_pd(velecsum,velec);
446
447 fscal = felec;
448
449 /* Calculate temporary vectorial force */
450 tx = _mm_mul_pd(fscal,dx11);
451 ty = _mm_mul_pd(fscal,dy11);
452 tz = _mm_mul_pd(fscal,dz11);
453
454 /* Update vectorial force */
455 fix1 = _mm_add_pd(fix1,tx);
456 fiy1 = _mm_add_pd(fiy1,ty);
457 fiz1 = _mm_add_pd(fiz1,tz);
458
459 fjx1 = _mm_add_pd(fjx1,tx);
460 fjy1 = _mm_add_pd(fjy1,ty);
461 fjz1 = _mm_add_pd(fjz1,tz);
462
463 /**************************
464 * CALCULATE INTERACTIONS *
465 **************************/
466
467 /* REACTION-FIELD ELECTROSTATICS */
468 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
469 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
470
471 /* Update potential sum for this i atom from the interaction with this j atom. */
472 velecsum = _mm_add_pd(velecsum,velec);
473
474 fscal = felec;
475
476 /* Calculate temporary vectorial force */
477 tx = _mm_mul_pd(fscal,dx12);
478 ty = _mm_mul_pd(fscal,dy12);
479 tz = _mm_mul_pd(fscal,dz12);
480
481 /* Update vectorial force */
482 fix1 = _mm_add_pd(fix1,tx);
483 fiy1 = _mm_add_pd(fiy1,ty);
484 fiz1 = _mm_add_pd(fiz1,tz);
485
486 fjx2 = _mm_add_pd(fjx2,tx);
487 fjy2 = _mm_add_pd(fjy2,ty);
488 fjz2 = _mm_add_pd(fjz2,tz);
489
490 /**************************
491 * CALCULATE INTERACTIONS *
492 **************************/
493
494 /* REACTION-FIELD ELECTROSTATICS */
495 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
496 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
497
498 /* Update potential sum for this i atom from the interaction with this j atom. */
499 velecsum = _mm_add_pd(velecsum,velec);
500
501 fscal = felec;
502
503 /* Calculate temporary vectorial force */
504 tx = _mm_mul_pd(fscal,dx20);
505 ty = _mm_mul_pd(fscal,dy20);
506 tz = _mm_mul_pd(fscal,dz20);
507
508 /* Update vectorial force */
509 fix2 = _mm_add_pd(fix2,tx);
510 fiy2 = _mm_add_pd(fiy2,ty);
511 fiz2 = _mm_add_pd(fiz2,tz);
512
513 fjx0 = _mm_add_pd(fjx0,tx);
514 fjy0 = _mm_add_pd(fjy0,ty);
515 fjz0 = _mm_add_pd(fjz0,tz);
516
517 /**************************
518 * CALCULATE INTERACTIONS *
519 **************************/
520
521 /* REACTION-FIELD ELECTROSTATICS */
522 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
523 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
524
525 /* Update potential sum for this i atom from the interaction with this j atom. */
526 velecsum = _mm_add_pd(velecsum,velec);
527
528 fscal = felec;
529
530 /* Calculate temporary vectorial force */
531 tx = _mm_mul_pd(fscal,dx21);
532 ty = _mm_mul_pd(fscal,dy21);
533 tz = _mm_mul_pd(fscal,dz21);
534
535 /* Update vectorial force */
536 fix2 = _mm_add_pd(fix2,tx);
537 fiy2 = _mm_add_pd(fiy2,ty);
538 fiz2 = _mm_add_pd(fiz2,tz);
539
540 fjx1 = _mm_add_pd(fjx1,tx);
541 fjy1 = _mm_add_pd(fjy1,ty);
542 fjz1 = _mm_add_pd(fjz1,tz);
543
544 /**************************
545 * CALCULATE INTERACTIONS *
546 **************************/
547
548 /* REACTION-FIELD ELECTROSTATICS */
549 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
550 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
551
552 /* Update potential sum for this i atom from the interaction with this j atom. */
553 velecsum = _mm_add_pd(velecsum,velec);
554
555 fscal = felec;
556
557 /* Calculate temporary vectorial force */
558 tx = _mm_mul_pd(fscal,dx22);
559 ty = _mm_mul_pd(fscal,dy22);
560 tz = _mm_mul_pd(fscal,dz22);
561
562 /* Update vectorial force */
563 fix2 = _mm_add_pd(fix2,tx);
564 fiy2 = _mm_add_pd(fiy2,ty);
565 fiz2 = _mm_add_pd(fiz2,tz);
566
567 fjx2 = _mm_add_pd(fjx2,tx);
568 fjy2 = _mm_add_pd(fjy2,ty);
569 fjz2 = _mm_add_pd(fjz2,tz);
570
571 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
572
573 /* Inner loop uses 323 flops */
574 }
575
576 if(jidx<j_index_end)
577 {
578
579 jnrA = jjnr[jidx];
580 j_coord_offsetA = DIM*jnrA;
581
582 /* load j atom coordinates */
583 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
584 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
585
586 /* Calculate displacement vector */
587 dx00 = _mm_sub_pd(ix0,jx0);
588 dy00 = _mm_sub_pd(iy0,jy0);
589 dz00 = _mm_sub_pd(iz0,jz0);
590 dx01 = _mm_sub_pd(ix0,jx1);
591 dy01 = _mm_sub_pd(iy0,jy1);
592 dz01 = _mm_sub_pd(iz0,jz1);
593 dx02 = _mm_sub_pd(ix0,jx2);
594 dy02 = _mm_sub_pd(iy0,jy2);
595 dz02 = _mm_sub_pd(iz0,jz2);
596 dx10 = _mm_sub_pd(ix1,jx0);
597 dy10 = _mm_sub_pd(iy1,jy0);
598 dz10 = _mm_sub_pd(iz1,jz0);
599 dx11 = _mm_sub_pd(ix1,jx1);
600 dy11 = _mm_sub_pd(iy1,jy1);
601 dz11 = _mm_sub_pd(iz1,jz1);
602 dx12 = _mm_sub_pd(ix1,jx2);
603 dy12 = _mm_sub_pd(iy1,jy2);
604 dz12 = _mm_sub_pd(iz1,jz2);
605 dx20 = _mm_sub_pd(ix2,jx0);
606 dy20 = _mm_sub_pd(iy2,jy0);
607 dz20 = _mm_sub_pd(iz2,jz0);
608 dx21 = _mm_sub_pd(ix2,jx1);
609 dy21 = _mm_sub_pd(iy2,jy1);
610 dz21 = _mm_sub_pd(iz2,jz1);
611 dx22 = _mm_sub_pd(ix2,jx2);
612 dy22 = _mm_sub_pd(iy2,jy2);
613 dz22 = _mm_sub_pd(iz2,jz2);
614
615 /* Calculate squared distance and things based on it */
616 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
617 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
618 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
619 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
620 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
621 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
622 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
623 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
624 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
625
626 rinv00 = gmx_mm_invsqrt_pd(rsq00);
627 rinv01 = gmx_mm_invsqrt_pd(rsq01);
628 rinv02 = gmx_mm_invsqrt_pd(rsq02);
629 rinv10 = gmx_mm_invsqrt_pd(rsq10);
630 rinv11 = gmx_mm_invsqrt_pd(rsq11);
631 rinv12 = gmx_mm_invsqrt_pd(rsq12);
632 rinv20 = gmx_mm_invsqrt_pd(rsq20);
633 rinv21 = gmx_mm_invsqrt_pd(rsq21);
634 rinv22 = gmx_mm_invsqrt_pd(rsq22);
635
636 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
637 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
638 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
639 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
640 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
641 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
642 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
643 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
644 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
645
646 fjx0 = _mm_setzero_pd();
647 fjy0 = _mm_setzero_pd();
648 fjz0 = _mm_setzero_pd();
649 fjx1 = _mm_setzero_pd();
650 fjy1 = _mm_setzero_pd();
651 fjz1 = _mm_setzero_pd();
652 fjx2 = _mm_setzero_pd();
653 fjy2 = _mm_setzero_pd();
654 fjz2 = _mm_setzero_pd();
655
656 /**************************
657 * CALCULATE INTERACTIONS *
658 **************************/
659
660 r00 = _mm_mul_pd(rsq00,rinv00);
661
662 /* Calculate table index by multiplying r with table scale and truncate to integer */
663 rt = _mm_mul_pd(r00,vftabscale);
664 vfitab = _mm_cvttpd_epi32(rt);
665 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
666 vfitab = _mm_slli_epi32(vfitab,3);
667
668 /* REACTION-FIELD ELECTROSTATICS */
669 velec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
670 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
671
672 /* CUBIC SPLINE TABLE DISPERSION */
673 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
674 F = _mm_setzero_pd();
675 GMX_MM_TRANSPOSE2_PD(Y,F);
676 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
677 H = _mm_setzero_pd();
678 GMX_MM_TRANSPOSE2_PD(G,H);
679 Heps = _mm_mul_pd(vfeps,H);
680 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
681 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
682 vvdw6 = _mm_mul_pd(c6_00,VV);
683 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
684 fvdw6 = _mm_mul_pd(c6_00,FF);
685
686 /* CUBIC SPLINE TABLE REPULSION */
687 vfitab = _mm_add_epi32(vfitab,ifour);
688 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
689 F = _mm_setzero_pd();
690 GMX_MM_TRANSPOSE2_PD(Y,F);
691 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
692 H = _mm_setzero_pd();
693 GMX_MM_TRANSPOSE2_PD(G,H);
694 Heps = _mm_mul_pd(vfeps,H);
695 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
696 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
697 vvdw12 = _mm_mul_pd(c12_00,VV);
698 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
699 fvdw12 = _mm_mul_pd(c12_00,FF);
700 vvdw = _mm_add_pd(vvdw12,vvdw6);
701 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
702
703 /* Update potential sum for this i atom from the interaction with this j atom. */
704 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
705 velecsum = _mm_add_pd(velecsum,velec);
706 vvdw = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
707 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
708
709 fscal = _mm_add_pd(felec,fvdw);
710
711 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
712
713 /* Calculate temporary vectorial force */
714 tx = _mm_mul_pd(fscal,dx00);
715 ty = _mm_mul_pd(fscal,dy00);
716 tz = _mm_mul_pd(fscal,dz00);
717
718 /* Update vectorial force */
719 fix0 = _mm_add_pd(fix0,tx);
720 fiy0 = _mm_add_pd(fiy0,ty);
721 fiz0 = _mm_add_pd(fiz0,tz);
722
723 fjx0 = _mm_add_pd(fjx0,tx);
724 fjy0 = _mm_add_pd(fjy0,ty);
725 fjz0 = _mm_add_pd(fjz0,tz);
726
727 /**************************
728 * CALCULATE INTERACTIONS *
729 **************************/
730
731 /* REACTION-FIELD ELECTROSTATICS */
732 velec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
733 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
734
735 /* Update potential sum for this i atom from the interaction with this j atom. */
736 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
737 velecsum = _mm_add_pd(velecsum,velec);
738
739 fscal = felec;
740
741 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
742
743 /* Calculate temporary vectorial force */
744 tx = _mm_mul_pd(fscal,dx01);
745 ty = _mm_mul_pd(fscal,dy01);
746 tz = _mm_mul_pd(fscal,dz01);
747
748 /* Update vectorial force */
749 fix0 = _mm_add_pd(fix0,tx);
750 fiy0 = _mm_add_pd(fiy0,ty);
751 fiz0 = _mm_add_pd(fiz0,tz);
752
753 fjx1 = _mm_add_pd(fjx1,tx);
754 fjy1 = _mm_add_pd(fjy1,ty);
755 fjz1 = _mm_add_pd(fjz1,tz);
756
757 /**************************
758 * CALCULATE INTERACTIONS *
759 **************************/
760
761 /* REACTION-FIELD ELECTROSTATICS */
762 velec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
763 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
764
765 /* Update potential sum for this i atom from the interaction with this j atom. */
766 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
767 velecsum = _mm_add_pd(velecsum,velec);
768
769 fscal = felec;
770
771 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
772
773 /* Calculate temporary vectorial force */
774 tx = _mm_mul_pd(fscal,dx02);
775 ty = _mm_mul_pd(fscal,dy02);
776 tz = _mm_mul_pd(fscal,dz02);
777
778 /* Update vectorial force */
779 fix0 = _mm_add_pd(fix0,tx);
780 fiy0 = _mm_add_pd(fiy0,ty);
781 fiz0 = _mm_add_pd(fiz0,tz);
782
783 fjx2 = _mm_add_pd(fjx2,tx);
784 fjy2 = _mm_add_pd(fjy2,ty);
785 fjz2 = _mm_add_pd(fjz2,tz);
786
787 /**************************
788 * CALCULATE INTERACTIONS *
789 **************************/
790
791 /* REACTION-FIELD ELECTROSTATICS */
792 velec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
793 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
794
795 /* Update potential sum for this i atom from the interaction with this j atom. */
796 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
797 velecsum = _mm_add_pd(velecsum,velec);
798
799 fscal = felec;
800
801 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
802
803 /* Calculate temporary vectorial force */
804 tx = _mm_mul_pd(fscal,dx10);
805 ty = _mm_mul_pd(fscal,dy10);
806 tz = _mm_mul_pd(fscal,dz10);
807
808 /* Update vectorial force */
809 fix1 = _mm_add_pd(fix1,tx);
810 fiy1 = _mm_add_pd(fiy1,ty);
811 fiz1 = _mm_add_pd(fiz1,tz);
812
813 fjx0 = _mm_add_pd(fjx0,tx);
814 fjy0 = _mm_add_pd(fjy0,ty);
815 fjz0 = _mm_add_pd(fjz0,tz);
816
817 /**************************
818 * CALCULATE INTERACTIONS *
819 **************************/
820
821 /* REACTION-FIELD ELECTROSTATICS */
822 velec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
823 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
824
825 /* Update potential sum for this i atom from the interaction with this j atom. */
826 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
827 velecsum = _mm_add_pd(velecsum,velec);
828
829 fscal = felec;
830
831 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
832
833 /* Calculate temporary vectorial force */
834 tx = _mm_mul_pd(fscal,dx11);
835 ty = _mm_mul_pd(fscal,dy11);
836 tz = _mm_mul_pd(fscal,dz11);
837
838 /* Update vectorial force */
839 fix1 = _mm_add_pd(fix1,tx);
840 fiy1 = _mm_add_pd(fiy1,ty);
841 fiz1 = _mm_add_pd(fiz1,tz);
842
843 fjx1 = _mm_add_pd(fjx1,tx);
844 fjy1 = _mm_add_pd(fjy1,ty);
845 fjz1 = _mm_add_pd(fjz1,tz);
846
847 /**************************
848 * CALCULATE INTERACTIONS *
849 **************************/
850
851 /* REACTION-FIELD ELECTROSTATICS */
852 velec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
853 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
854
855 /* Update potential sum for this i atom from the interaction with this j atom. */
856 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
857 velecsum = _mm_add_pd(velecsum,velec);
858
859 fscal = felec;
860
861 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
862
863 /* Calculate temporary vectorial force */
864 tx = _mm_mul_pd(fscal,dx12);
865 ty = _mm_mul_pd(fscal,dy12);
866 tz = _mm_mul_pd(fscal,dz12);
867
868 /* Update vectorial force */
869 fix1 = _mm_add_pd(fix1,tx);
870 fiy1 = _mm_add_pd(fiy1,ty);
871 fiz1 = _mm_add_pd(fiz1,tz);
872
873 fjx2 = _mm_add_pd(fjx2,tx);
874 fjy2 = _mm_add_pd(fjy2,ty);
875 fjz2 = _mm_add_pd(fjz2,tz);
876
877 /**************************
878 * CALCULATE INTERACTIONS *
879 **************************/
880
881 /* REACTION-FIELD ELECTROSTATICS */
882 velec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
883 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
884
885 /* Update potential sum for this i atom from the interaction with this j atom. */
886 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
887 velecsum = _mm_add_pd(velecsum,velec);
888
889 fscal = felec;
890
891 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
892
893 /* Calculate temporary vectorial force */
894 tx = _mm_mul_pd(fscal,dx20);
895 ty = _mm_mul_pd(fscal,dy20);
896 tz = _mm_mul_pd(fscal,dz20);
897
898 /* Update vectorial force */
899 fix2 = _mm_add_pd(fix2,tx);
900 fiy2 = _mm_add_pd(fiy2,ty);
901 fiz2 = _mm_add_pd(fiz2,tz);
902
903 fjx0 = _mm_add_pd(fjx0,tx);
904 fjy0 = _mm_add_pd(fjy0,ty);
905 fjz0 = _mm_add_pd(fjz0,tz);
906
907 /**************************
908 * CALCULATE INTERACTIONS *
909 **************************/
910
911 /* REACTION-FIELD ELECTROSTATICS */
912 velec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
913 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
914
915 /* Update potential sum for this i atom from the interaction with this j atom. */
916 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
917 velecsum = _mm_add_pd(velecsum,velec);
918
919 fscal = felec;
920
921 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
922
923 /* Calculate temporary vectorial force */
924 tx = _mm_mul_pd(fscal,dx21);
925 ty = _mm_mul_pd(fscal,dy21);
926 tz = _mm_mul_pd(fscal,dz21);
927
928 /* Update vectorial force */
929 fix2 = _mm_add_pd(fix2,tx);
930 fiy2 = _mm_add_pd(fiy2,ty);
931 fiz2 = _mm_add_pd(fiz2,tz);
932
933 fjx1 = _mm_add_pd(fjx1,tx);
934 fjy1 = _mm_add_pd(fjy1,ty);
935 fjz1 = _mm_add_pd(fjz1,tz);
936
937 /**************************
938 * CALCULATE INTERACTIONS *
939 **************************/
940
941 /* REACTION-FIELD ELECTROSTATICS */
942 velec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
943 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
944
945 /* Update potential sum for this i atom from the interaction with this j atom. */
946 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
947 velecsum = _mm_add_pd(velecsum,velec);
948
949 fscal = felec;
950
951 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
952
953 /* Calculate temporary vectorial force */
954 tx = _mm_mul_pd(fscal,dx22);
955 ty = _mm_mul_pd(fscal,dy22);
956 tz = _mm_mul_pd(fscal,dz22);
957
958 /* Update vectorial force */
959 fix2 = _mm_add_pd(fix2,tx);
960 fiy2 = _mm_add_pd(fiy2,ty);
961 fiz2 = _mm_add_pd(fiz2,tz);
962
963 fjx2 = _mm_add_pd(fjx2,tx);
964 fjy2 = _mm_add_pd(fjy2,ty);
965 fjz2 = _mm_add_pd(fjz2,tz);
966
967 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
968
969 /* Inner loop uses 323 flops */
970 }
971
972 /* End of innermost loop */
973
974 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
975 f+i_coord_offset,fshift+i_shift_offset);
976
977 ggid = gid[iidx];
978 /* Update potential energies */
979 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
980 gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
981
982 /* Increment number of inner iterations */
983 inneriter += j_index_end - j_index_start;
984
985 /* Outer loop uses 20 flops */
986 }
987
988 /* Increment number of outer iterations */
989 outeriter += nri;
990
991 /* Update outer/inner flops */
992
993 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*323);
994 }
995 /*
996 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_sse2_double
997 * Electrostatics interaction: ReactionField
998 * VdW interaction: CubicSplineTable
999 * Geometry: Water3-Water3
1000 * Calculate force/pot: Force
1001 */
1002 void
1003 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_sse2_double
1004 (t_nblist * gmx_restrict nlist,
1005 rvec * gmx_restrict xx,
1006 rvec * gmx_restrict ff,
1007 t_forcerec * gmx_restrict fr,
1008 t_mdatoms * gmx_restrict mdatoms,
1009 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1010 t_nrnb * gmx_restrict nrnb)
1011 {
1012 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1013 * just 0 for non-waters.
1014 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1015 * jnr indices corresponding to data put in the four positions in the SIMD register.
1016 */
1017 int i_shift_offset,i_coord_offset,outeriter,inneriter;
1018 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1019 int jnrA,jnrB;
1020 int j_coord_offsetA,j_coord_offsetB;
1021 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
1022 real rcutoff_scalar;
1023 real *shiftvec,*fshift,*x,*f;
1024 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1025 int vdwioffset0;
1026 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1027 int vdwioffset1;
1028 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1029 int vdwioffset2;
1030 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1031 int vdwjidx0A,vdwjidx0B;
1032 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1033 int vdwjidx1A,vdwjidx1B;
1034 __m128d jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1035 int vdwjidx2A,vdwjidx2B;
1036 __m128d jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1037 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1038 __m128d dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1039 __m128d dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1040 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1041 __m128d dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1042 __m128d dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1043 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1044 __m128d dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1045 __m128d dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1046 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
1047 real *charge;
1048 int nvdwtype;
1049 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1050 int *vdwtype;
1051 real *vdwparam;
1052 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
1053 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
1054 __m128i vfitab;
1055 __m128i ifour = _mm_set1_epi32(4);
1056 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1057 real *vftab;
1058 __m128d dummy_mask,cutoff_mask;
1059 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1060 __m128d one = _mm_set1_pd(1.0);
1061 __m128d two = _mm_set1_pd(2.0);
1062 x = xx[0];
1063 f = ff[0];
1064
1065 nri = nlist->nri;
1066 iinr = nlist->iinr;
1067 jindex = nlist->jindex;
1068 jjnr = nlist->jjnr;
1069 shiftidx = nlist->shift;
1070 gid = nlist->gid;
1071 shiftvec = fr->shift_vec[0];
1072 fshift = fr->fshift[0];
1073 facel = _mm_set1_pd(fr->epsfac);
1074 charge = mdatoms->chargeA;
1075 krf = _mm_set1_pd(fr->ic->k_rf);
1076 krf2 = _mm_set1_pd(fr->ic->k_rf*2.0);
1077 crf = _mm_set1_pd(fr->ic->c_rf);
1078 nvdwtype = fr->ntype;
1079 vdwparam = fr->nbfp;
1080 vdwtype = mdatoms->typeA;
1081
1082 vftab = kernel_data->table_vdw->data;
1083 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
1084
1085 /* Setup water-specific parameters */
1086 inr = nlist->iinr[0];
1087 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+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 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1091
1092 jq0 = _mm_set1_pd(charge[inr+0]);
1093 jq1 = _mm_set1_pd(charge[inr+1]);
1094 jq2 = _mm_set1_pd(charge[inr+2]);
1095 vdwjidx0A = 2*vdwtype[inr+0];
1096 qq00 = _mm_mul_pd(iq0,jq0);
1097 c6_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1098 c12_00 = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1099 qq01 = _mm_mul_pd(iq0,jq1);
1100 qq02 = _mm_mul_pd(iq0,jq2);
1101 qq10 = _mm_mul_pd(iq1,jq0);
1102 qq11 = _mm_mul_pd(iq1,jq1);
1103 qq12 = _mm_mul_pd(iq1,jq2);
1104 qq20 = _mm_mul_pd(iq2,jq0);
1105 qq21 = _mm_mul_pd(iq2,jq1);
1106 qq22 = _mm_mul_pd(iq2,jq2);
1107
1108 /* Avoid stupid compiler warnings */
1109 jnrA = jnrB = 0;
1110 j_coord_offsetA = 0;
1111 j_coord_offsetB = 0;
1112
1113 outeriter = 0;
1114 inneriter = 0;
1115
1116 /* Start outer loop over neighborlists */
1117 for(iidx=0; iidx<nri; iidx++)
1118 {
1119 /* Load shift vector for this list */
1120 i_shift_offset = DIM*shiftidx[iidx];
1121
1122 /* Load limits for loop over neighbors */
1123 j_index_start = jindex[iidx];
1124 j_index_end = jindex[iidx+1];
1125
1126 /* Get outer coordinate index */
1127 inr = iinr[iidx];
1128 i_coord_offset = DIM*inr;
1129
1130 /* Load i particle coords and add shift vector */
1131 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1132 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1133
1134 fix0 = _mm_setzero_pd();
1135 fiy0 = _mm_setzero_pd();
1136 fiz0 = _mm_setzero_pd();
1137 fix1 = _mm_setzero_pd();
1138 fiy1 = _mm_setzero_pd();
1139 fiz1 = _mm_setzero_pd();
1140 fix2 = _mm_setzero_pd();
1141 fiy2 = _mm_setzero_pd();
1142 fiz2 = _mm_setzero_pd();
1143
1144 /* Start inner kernel loop */
1145 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1146 {
1147
1148 /* Get j neighbor index, and coordinate index */
1149 jnrA = jjnr[jidx];
1150 jnrB = jjnr[jidx+1];
1151 j_coord_offsetA = DIM*jnrA;
1152 j_coord_offsetB = DIM*jnrB;
1153
1154 /* load j atom coordinates */
1155 gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1156 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1157
1158 /* Calculate displacement vector */
1159 dx00 = _mm_sub_pd(ix0,jx0);
1160 dy00 = _mm_sub_pd(iy0,jy0);
1161 dz00 = _mm_sub_pd(iz0,jz0);
1162 dx01 = _mm_sub_pd(ix0,jx1);
1163 dy01 = _mm_sub_pd(iy0,jy1);
1164 dz01 = _mm_sub_pd(iz0,jz1);
1165 dx02 = _mm_sub_pd(ix0,jx2);
1166 dy02 = _mm_sub_pd(iy0,jy2);
1167 dz02 = _mm_sub_pd(iz0,jz2);
1168 dx10 = _mm_sub_pd(ix1,jx0);
1169 dy10 = _mm_sub_pd(iy1,jy0);
1170 dz10 = _mm_sub_pd(iz1,jz0);
1171 dx11 = _mm_sub_pd(ix1,jx1);
1172 dy11 = _mm_sub_pd(iy1,jy1);
1173 dz11 = _mm_sub_pd(iz1,jz1);
1174 dx12 = _mm_sub_pd(ix1,jx2);
1175 dy12 = _mm_sub_pd(iy1,jy2);
1176 dz12 = _mm_sub_pd(iz1,jz2);
1177 dx20 = _mm_sub_pd(ix2,jx0);
1178 dy20 = _mm_sub_pd(iy2,jy0);
1179 dz20 = _mm_sub_pd(iz2,jz0);
1180 dx21 = _mm_sub_pd(ix2,jx1);
1181 dy21 = _mm_sub_pd(iy2,jy1);
1182 dz21 = _mm_sub_pd(iz2,jz1);
1183 dx22 = _mm_sub_pd(ix2,jx2);
1184 dy22 = _mm_sub_pd(iy2,jy2);
1185 dz22 = _mm_sub_pd(iz2,jz2);
1186
1187 /* Calculate squared distance and things based on it */
1188 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1189 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1190 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1191 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1192 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1193 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1194 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1195 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1196 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1197
1198 rinv00 = gmx_mm_invsqrt_pd(rsq00);
1199 rinv01 = gmx_mm_invsqrt_pd(rsq01);
1200 rinv02 = gmx_mm_invsqrt_pd(rsq02);
1201 rinv10 = gmx_mm_invsqrt_pd(rsq10);
1202 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1203 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1204 rinv20 = gmx_mm_invsqrt_pd(rsq20);
1205 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1206 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1207
1208 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1209 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1210 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1211 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1212 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1213 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1214 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1215 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1216 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1217
1218 fjx0 = _mm_setzero_pd();
1219 fjy0 = _mm_setzero_pd();
1220 fjz0 = _mm_setzero_pd();
1221 fjx1 = _mm_setzero_pd();
1222 fjy1 = _mm_setzero_pd();
1223 fjz1 = _mm_setzero_pd();
1224 fjx2 = _mm_setzero_pd();
1225 fjy2 = _mm_setzero_pd();
1226 fjz2 = _mm_setzero_pd();
1227
1228 /**************************
1229 * CALCULATE INTERACTIONS *
1230 **************************/
1231
1232 r00 = _mm_mul_pd(rsq00,rinv00);
1233
1234 /* Calculate table index by multiplying r with table scale and truncate to integer */
1235 rt = _mm_mul_pd(r00,vftabscale);
1236 vfitab = _mm_cvttpd_epi32(rt);
1237 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1238 vfitab = _mm_slli_epi32(vfitab,3);
1239
1240 /* REACTION-FIELD ELECTROSTATICS */
1241 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1242
1243 /* CUBIC SPLINE TABLE DISPERSION */
1244 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1245 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1246 GMX_MM_TRANSPOSE2_PD(Y,F);
1247 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1248 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1249 GMX_MM_TRANSPOSE2_PD(G,H);
1250 Heps = _mm_mul_pd(vfeps,H);
1251 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1252 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1253 fvdw6 = _mm_mul_pd(c6_00,FF);
1254
1255 /* CUBIC SPLINE TABLE REPULSION */
1256 vfitab = _mm_add_epi32(vfitab,ifour);
1257 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1258 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1259 GMX_MM_TRANSPOSE2_PD(Y,F);
1260 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1261 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1262 GMX_MM_TRANSPOSE2_PD(G,H);
1263 Heps = _mm_mul_pd(vfeps,H);
1264 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1265 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1266 fvdw12 = _mm_mul_pd(c12_00,FF);
1267 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1268
1269 fscal = _mm_add_pd(felec,fvdw);
1270
1271 /* Calculate temporary vectorial force */
1272 tx = _mm_mul_pd(fscal,dx00);
1273 ty = _mm_mul_pd(fscal,dy00);
1274 tz = _mm_mul_pd(fscal,dz00);
1275
1276 /* Update vectorial force */
1277 fix0 = _mm_add_pd(fix0,tx);
1278 fiy0 = _mm_add_pd(fiy0,ty);
1279 fiz0 = _mm_add_pd(fiz0,tz);
1280
1281 fjx0 = _mm_add_pd(fjx0,tx);
1282 fjy0 = _mm_add_pd(fjy0,ty);
1283 fjz0 = _mm_add_pd(fjz0,tz);
1284
1285 /**************************
1286 * CALCULATE INTERACTIONS *
1287 **************************/
1288
1289 /* REACTION-FIELD ELECTROSTATICS */
1290 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1291
1292 fscal = felec;
1293
1294 /* Calculate temporary vectorial force */
1295 tx = _mm_mul_pd(fscal,dx01);
1296 ty = _mm_mul_pd(fscal,dy01);
1297 tz = _mm_mul_pd(fscal,dz01);
1298
1299 /* Update vectorial force */
1300 fix0 = _mm_add_pd(fix0,tx);
1301 fiy0 = _mm_add_pd(fiy0,ty);
1302 fiz0 = _mm_add_pd(fiz0,tz);
1303
1304 fjx1 = _mm_add_pd(fjx1,tx);
1305 fjy1 = _mm_add_pd(fjy1,ty);
1306 fjz1 = _mm_add_pd(fjz1,tz);
1307
1308 /**************************
1309 * CALCULATE INTERACTIONS *
1310 **************************/
1311
1312 /* REACTION-FIELD ELECTROSTATICS */
1313 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1314
1315 fscal = felec;
1316
1317 /* Calculate temporary vectorial force */
1318 tx = _mm_mul_pd(fscal,dx02);
1319 ty = _mm_mul_pd(fscal,dy02);
1320 tz = _mm_mul_pd(fscal,dz02);
1321
1322 /* Update vectorial force */
1323 fix0 = _mm_add_pd(fix0,tx);
1324 fiy0 = _mm_add_pd(fiy0,ty);
1325 fiz0 = _mm_add_pd(fiz0,tz);
1326
1327 fjx2 = _mm_add_pd(fjx2,tx);
1328 fjy2 = _mm_add_pd(fjy2,ty);
1329 fjz2 = _mm_add_pd(fjz2,tz);
1330
1331 /**************************
1332 * CALCULATE INTERACTIONS *
1333 **************************/
1334
1335 /* REACTION-FIELD ELECTROSTATICS */
1336 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1337
1338 fscal = felec;
1339
1340 /* Calculate temporary vectorial force */
1341 tx = _mm_mul_pd(fscal,dx10);
1342 ty = _mm_mul_pd(fscal,dy10);
1343 tz = _mm_mul_pd(fscal,dz10);
1344
1345 /* Update vectorial force */
1346 fix1 = _mm_add_pd(fix1,tx);
1347 fiy1 = _mm_add_pd(fiy1,ty);
1348 fiz1 = _mm_add_pd(fiz1,tz);
1349
1350 fjx0 = _mm_add_pd(fjx0,tx);
1351 fjy0 = _mm_add_pd(fjy0,ty);
1352 fjz0 = _mm_add_pd(fjz0,tz);
1353
1354 /**************************
1355 * CALCULATE INTERACTIONS *
1356 **************************/
1357
1358 /* REACTION-FIELD ELECTROSTATICS */
1359 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1360
1361 fscal = felec;
1362
1363 /* Calculate temporary vectorial force */
1364 tx = _mm_mul_pd(fscal,dx11);
1365 ty = _mm_mul_pd(fscal,dy11);
1366 tz = _mm_mul_pd(fscal,dz11);
1367
1368 /* Update vectorial force */
1369 fix1 = _mm_add_pd(fix1,tx);
1370 fiy1 = _mm_add_pd(fiy1,ty);
1371 fiz1 = _mm_add_pd(fiz1,tz);
1372
1373 fjx1 = _mm_add_pd(fjx1,tx);
1374 fjy1 = _mm_add_pd(fjy1,ty);
1375 fjz1 = _mm_add_pd(fjz1,tz);
1376
1377 /**************************
1378 * CALCULATE INTERACTIONS *
1379 **************************/
1380
1381 /* REACTION-FIELD ELECTROSTATICS */
1382 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1383
1384 fscal = felec;
1385
1386 /* Calculate temporary vectorial force */
1387 tx = _mm_mul_pd(fscal,dx12);
1388 ty = _mm_mul_pd(fscal,dy12);
1389 tz = _mm_mul_pd(fscal,dz12);
1390
1391 /* Update vectorial force */
1392 fix1 = _mm_add_pd(fix1,tx);
1393 fiy1 = _mm_add_pd(fiy1,ty);
1394 fiz1 = _mm_add_pd(fiz1,tz);
1395
1396 fjx2 = _mm_add_pd(fjx2,tx);
1397 fjy2 = _mm_add_pd(fjy2,ty);
1398 fjz2 = _mm_add_pd(fjz2,tz);
1399
1400 /**************************
1401 * CALCULATE INTERACTIONS *
1402 **************************/
1403
1404 /* REACTION-FIELD ELECTROSTATICS */
1405 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1406
1407 fscal = felec;
1408
1409 /* Calculate temporary vectorial force */
1410 tx = _mm_mul_pd(fscal,dx20);
1411 ty = _mm_mul_pd(fscal,dy20);
1412 tz = _mm_mul_pd(fscal,dz20);
1413
1414 /* Update vectorial force */
1415 fix2 = _mm_add_pd(fix2,tx);
1416 fiy2 = _mm_add_pd(fiy2,ty);
1417 fiz2 = _mm_add_pd(fiz2,tz);
1418
1419 fjx0 = _mm_add_pd(fjx0,tx);
1420 fjy0 = _mm_add_pd(fjy0,ty);
1421 fjz0 = _mm_add_pd(fjz0,tz);
1422
1423 /**************************
1424 * CALCULATE INTERACTIONS *
1425 **************************/
1426
1427 /* REACTION-FIELD ELECTROSTATICS */
1428 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1429
1430 fscal = felec;
1431
1432 /* Calculate temporary vectorial force */
1433 tx = _mm_mul_pd(fscal,dx21);
1434 ty = _mm_mul_pd(fscal,dy21);
1435 tz = _mm_mul_pd(fscal,dz21);
1436
1437 /* Update vectorial force */
1438 fix2 = _mm_add_pd(fix2,tx);
1439 fiy2 = _mm_add_pd(fiy2,ty);
1440 fiz2 = _mm_add_pd(fiz2,tz);
1441
1442 fjx1 = _mm_add_pd(fjx1,tx);
1443 fjy1 = _mm_add_pd(fjy1,ty);
1444 fjz1 = _mm_add_pd(fjz1,tz);
1445
1446 /**************************
1447 * CALCULATE INTERACTIONS *
1448 **************************/
1449
1450 /* REACTION-FIELD ELECTROSTATICS */
1451 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1452
1453 fscal = felec;
1454
1455 /* Calculate temporary vectorial force */
1456 tx = _mm_mul_pd(fscal,dx22);
1457 ty = _mm_mul_pd(fscal,dy22);
1458 tz = _mm_mul_pd(fscal,dz22);
1459
1460 /* Update vectorial force */
1461 fix2 = _mm_add_pd(fix2,tx);
1462 fiy2 = _mm_add_pd(fiy2,ty);
1463 fiz2 = _mm_add_pd(fiz2,tz);
1464
1465 fjx2 = _mm_add_pd(fjx2,tx);
1466 fjy2 = _mm_add_pd(fjy2,ty);
1467 fjz2 = _mm_add_pd(fjz2,tz);
1468
1469 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1470
1471 /* Inner loop uses 270 flops */
1472 }
1473
1474 if(jidx<j_index_end)
1475 {
1476
1477 jnrA = jjnr[jidx];
1478 j_coord_offsetA = DIM*jnrA;
1479
1480 /* load j atom coordinates */
1481 gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1482 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1483
1484 /* Calculate displacement vector */
1485 dx00 = _mm_sub_pd(ix0,jx0);
1486 dy00 = _mm_sub_pd(iy0,jy0);
1487 dz00 = _mm_sub_pd(iz0,jz0);
1488 dx01 = _mm_sub_pd(ix0,jx1);
1489 dy01 = _mm_sub_pd(iy0,jy1);
1490 dz01 = _mm_sub_pd(iz0,jz1);
1491 dx02 = _mm_sub_pd(ix0,jx2);
1492 dy02 = _mm_sub_pd(iy0,jy2);
1493 dz02 = _mm_sub_pd(iz0,jz2);
1494 dx10 = _mm_sub_pd(ix1,jx0);
1495 dy10 = _mm_sub_pd(iy1,jy0);
1496 dz10 = _mm_sub_pd(iz1,jz0);
1497 dx11 = _mm_sub_pd(ix1,jx1);
1498 dy11 = _mm_sub_pd(iy1,jy1);
1499 dz11 = _mm_sub_pd(iz1,jz1);
1500 dx12 = _mm_sub_pd(ix1,jx2);
1501 dy12 = _mm_sub_pd(iy1,jy2);
1502 dz12 = _mm_sub_pd(iz1,jz2);
1503 dx20 = _mm_sub_pd(ix2,jx0);
1504 dy20 = _mm_sub_pd(iy2,jy0);
1505 dz20 = _mm_sub_pd(iz2,jz0);
1506 dx21 = _mm_sub_pd(ix2,jx1);
1507 dy21 = _mm_sub_pd(iy2,jy1);
1508 dz21 = _mm_sub_pd(iz2,jz1);
1509 dx22 = _mm_sub_pd(ix2,jx2);
1510 dy22 = _mm_sub_pd(iy2,jy2);
1511 dz22 = _mm_sub_pd(iz2,jz2);
1512
1513 /* Calculate squared distance and things based on it */
1514 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1515 rsq01 = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1516 rsq02 = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1517 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1518 rsq11 = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1519 rsq12 = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1520 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1521 rsq21 = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1522 rsq22 = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1523
1524 rinv00 = gmx_mm_invsqrt_pd(rsq00);
1525 rinv01 = gmx_mm_invsqrt_pd(rsq01);
1526 rinv02 = gmx_mm_invsqrt_pd(rsq02);
1527 rinv10 = gmx_mm_invsqrt_pd(rsq10);
1528 rinv11 = gmx_mm_invsqrt_pd(rsq11);
1529 rinv12 = gmx_mm_invsqrt_pd(rsq12);
1530 rinv20 = gmx_mm_invsqrt_pd(rsq20);
1531 rinv21 = gmx_mm_invsqrt_pd(rsq21);
1532 rinv22 = gmx_mm_invsqrt_pd(rsq22);
1533
1534 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
1535 rinvsq01 = _mm_mul_pd(rinv01,rinv01);
1536 rinvsq02 = _mm_mul_pd(rinv02,rinv02);
1537 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
1538 rinvsq11 = _mm_mul_pd(rinv11,rinv11);
1539 rinvsq12 = _mm_mul_pd(rinv12,rinv12);
1540 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
1541 rinvsq21 = _mm_mul_pd(rinv21,rinv21);
1542 rinvsq22 = _mm_mul_pd(rinv22,rinv22);
1543
1544 fjx0 = _mm_setzero_pd();
1545 fjy0 = _mm_setzero_pd();
1546 fjz0 = _mm_setzero_pd();
1547 fjx1 = _mm_setzero_pd();
1548 fjy1 = _mm_setzero_pd();
1549 fjz1 = _mm_setzero_pd();
1550 fjx2 = _mm_setzero_pd();
1551 fjy2 = _mm_setzero_pd();
1552 fjz2 = _mm_setzero_pd();
1553
1554 /**************************
1555 * CALCULATE INTERACTIONS *
1556 **************************/
1557
1558 r00 = _mm_mul_pd(rsq00,rinv00);
1559
1560 /* Calculate table index by multiplying r with table scale and truncate to integer */
1561 rt = _mm_mul_pd(r00,vftabscale);
1562 vfitab = _mm_cvttpd_epi32(rt);
1563 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1564 vfitab = _mm_slli_epi32(vfitab,3);
1565
1566 /* REACTION-FIELD ELECTROSTATICS */
1567 felec = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1568
1569 /* CUBIC SPLINE TABLE DISPERSION */
1570 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1571 F = _mm_setzero_pd();
1572 GMX_MM_TRANSPOSE2_PD(Y,F);
1573 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1574 H = _mm_setzero_pd();
1575 GMX_MM_TRANSPOSE2_PD(G,H);
1576 Heps = _mm_mul_pd(vfeps,H);
1577 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1578 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1579 fvdw6 = _mm_mul_pd(c6_00,FF);
1580
1581 /* CUBIC SPLINE TABLE REPULSION */
1582 vfitab = _mm_add_epi32(vfitab,ifour);
1583 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1584 F = _mm_setzero_pd();
1585 GMX_MM_TRANSPOSE2_PD(Y,F);
1586 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1587 H = _mm_setzero_pd();
1588 GMX_MM_TRANSPOSE2_PD(G,H);
1589 Heps = _mm_mul_pd(vfeps,H);
1590 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1591 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1592 fvdw12 = _mm_mul_pd(c12_00,FF);
1593 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1594
1595 fscal = _mm_add_pd(felec,fvdw);
1596
1597 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1598
1599 /* Calculate temporary vectorial force */
1600 tx = _mm_mul_pd(fscal,dx00);
1601 ty = _mm_mul_pd(fscal,dy00);
1602 tz = _mm_mul_pd(fscal,dz00);
1603
1604 /* Update vectorial force */
1605 fix0 = _mm_add_pd(fix0,tx);
1606 fiy0 = _mm_add_pd(fiy0,ty);
1607 fiz0 = _mm_add_pd(fiz0,tz);
1608
1609 fjx0 = _mm_add_pd(fjx0,tx);
1610 fjy0 = _mm_add_pd(fjy0,ty);
1611 fjz0 = _mm_add_pd(fjz0,tz);
1612
1613 /**************************
1614 * CALCULATE INTERACTIONS *
1615 **************************/
1616
1617 /* REACTION-FIELD ELECTROSTATICS */
1618 felec = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1619
1620 fscal = felec;
1621
1622 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1623
1624 /* Calculate temporary vectorial force */
1625 tx = _mm_mul_pd(fscal,dx01);
1626 ty = _mm_mul_pd(fscal,dy01);
1627 tz = _mm_mul_pd(fscal,dz01);
1628
1629 /* Update vectorial force */
1630 fix0 = _mm_add_pd(fix0,tx);
1631 fiy0 = _mm_add_pd(fiy0,ty);
1632 fiz0 = _mm_add_pd(fiz0,tz);
1633
1634 fjx1 = _mm_add_pd(fjx1,tx);
1635 fjy1 = _mm_add_pd(fjy1,ty);
1636 fjz1 = _mm_add_pd(fjz1,tz);
1637
1638 /**************************
1639 * CALCULATE INTERACTIONS *
1640 **************************/
1641
1642 /* REACTION-FIELD ELECTROSTATICS */
1643 felec = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1644
1645 fscal = felec;
1646
1647 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1648
1649 /* Calculate temporary vectorial force */
1650 tx = _mm_mul_pd(fscal,dx02);
1651 ty = _mm_mul_pd(fscal,dy02);
1652 tz = _mm_mul_pd(fscal,dz02);
1653
1654 /* Update vectorial force */
1655 fix0 = _mm_add_pd(fix0,tx);
1656 fiy0 = _mm_add_pd(fiy0,ty);
1657 fiz0 = _mm_add_pd(fiz0,tz);
1658
1659 fjx2 = _mm_add_pd(fjx2,tx);
1660 fjy2 = _mm_add_pd(fjy2,ty);
1661 fjz2 = _mm_add_pd(fjz2,tz);
1662
1663 /**************************
1664 * CALCULATE INTERACTIONS *
1665 **************************/
1666
1667 /* REACTION-FIELD ELECTROSTATICS */
1668 felec = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1669
1670 fscal = felec;
1671
1672 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1673
1674 /* Calculate temporary vectorial force */
1675 tx = _mm_mul_pd(fscal,dx10);
1676 ty = _mm_mul_pd(fscal,dy10);
1677 tz = _mm_mul_pd(fscal,dz10);
1678
1679 /* Update vectorial force */
1680 fix1 = _mm_add_pd(fix1,tx);
1681 fiy1 = _mm_add_pd(fiy1,ty);
1682 fiz1 = _mm_add_pd(fiz1,tz);
1683
1684 fjx0 = _mm_add_pd(fjx0,tx);
1685 fjy0 = _mm_add_pd(fjy0,ty);
1686 fjz0 = _mm_add_pd(fjz0,tz);
1687
1688 /**************************
1689 * CALCULATE INTERACTIONS *
1690 **************************/
1691
1692 /* REACTION-FIELD ELECTROSTATICS */
1693 felec = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1694
1695 fscal = felec;
1696
1697 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1698
1699 /* Calculate temporary vectorial force */
1700 tx = _mm_mul_pd(fscal,dx11);
1701 ty = _mm_mul_pd(fscal,dy11);
1702 tz = _mm_mul_pd(fscal,dz11);
1703
1704 /* Update vectorial force */
1705 fix1 = _mm_add_pd(fix1,tx);
1706 fiy1 = _mm_add_pd(fiy1,ty);
1707 fiz1 = _mm_add_pd(fiz1,tz);
1708
1709 fjx1 = _mm_add_pd(fjx1,tx);
1710 fjy1 = _mm_add_pd(fjy1,ty);
1711 fjz1 = _mm_add_pd(fjz1,tz);
1712
1713 /**************************
1714 * CALCULATE INTERACTIONS *
1715 **************************/
1716
1717 /* REACTION-FIELD ELECTROSTATICS */
1718 felec = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1719
1720 fscal = felec;
1721
1722 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1723
1724 /* Calculate temporary vectorial force */
1725 tx = _mm_mul_pd(fscal,dx12);
1726 ty = _mm_mul_pd(fscal,dy12);
1727 tz = _mm_mul_pd(fscal,dz12);
1728
1729 /* Update vectorial force */
1730 fix1 = _mm_add_pd(fix1,tx);
1731 fiy1 = _mm_add_pd(fiy1,ty);
1732 fiz1 = _mm_add_pd(fiz1,tz);
1733
1734 fjx2 = _mm_add_pd(fjx2,tx);
1735 fjy2 = _mm_add_pd(fjy2,ty);
1736 fjz2 = _mm_add_pd(fjz2,tz);
1737
1738 /**************************
1739 * CALCULATE INTERACTIONS *
1740 **************************/
1741
1742 /* REACTION-FIELD ELECTROSTATICS */
1743 felec = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1744
1745 fscal = felec;
1746
1747 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1748
1749 /* Calculate temporary vectorial force */
1750 tx = _mm_mul_pd(fscal,dx20);
1751 ty = _mm_mul_pd(fscal,dy20);
1752 tz = _mm_mul_pd(fscal,dz20);
1753
1754 /* Update vectorial force */
1755 fix2 = _mm_add_pd(fix2,tx);
1756 fiy2 = _mm_add_pd(fiy2,ty);
1757 fiz2 = _mm_add_pd(fiz2,tz);
1758
1759 fjx0 = _mm_add_pd(fjx0,tx);
1760 fjy0 = _mm_add_pd(fjy0,ty);
1761 fjz0 = _mm_add_pd(fjz0,tz);
1762
1763 /**************************
1764 * CALCULATE INTERACTIONS *
1765 **************************/
1766
1767 /* REACTION-FIELD ELECTROSTATICS */
1768 felec = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1769
1770 fscal = felec;
1771
1772 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1773
1774 /* Calculate temporary vectorial force */
1775 tx = _mm_mul_pd(fscal,dx21);
1776 ty = _mm_mul_pd(fscal,dy21);
1777 tz = _mm_mul_pd(fscal,dz21);
1778
1779 /* Update vectorial force */
1780 fix2 = _mm_add_pd(fix2,tx);
1781 fiy2 = _mm_add_pd(fiy2,ty);
1782 fiz2 = _mm_add_pd(fiz2,tz);
1783
1784 fjx1 = _mm_add_pd(fjx1,tx);
1785 fjy1 = _mm_add_pd(fjy1,ty);
1786 fjz1 = _mm_add_pd(fjz1,tz);
1787
1788 /**************************
1789 * CALCULATE INTERACTIONS *
1790 **************************/
1791
1792 /* REACTION-FIELD ELECTROSTATICS */
1793 felec = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1794
1795 fscal = felec;
1796
1797 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1798
1799 /* Calculate temporary vectorial force */
1800 tx = _mm_mul_pd(fscal,dx22);
1801 ty = _mm_mul_pd(fscal,dy22);
1802 tz = _mm_mul_pd(fscal,dz22);
1803
1804 /* Update vectorial force */
1805 fix2 = _mm_add_pd(fix2,tx);
1806 fiy2 = _mm_add_pd(fiy2,ty);
1807 fiz2 = _mm_add_pd(fiz2,tz);
1808
1809 fjx2 = _mm_add_pd(fjx2,tx);
1810 fjy2 = _mm_add_pd(fjy2,ty);
1811 fjz2 = _mm_add_pd(fjz2,tz);
1812
1813 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1814
1815 /* Inner loop uses 270 flops */
1816 }
1817
1818 /* End of innermost loop */
1819
1820 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1821 f+i_coord_offset,fshift+i_shift_offset);
1822
1823 /* Increment number of inner iterations */
1824 inneriter += j_index_end - j_index_start;
1825
1826 /* Outer loop uses 18 flops */
1827 }
1828
1829 /* Increment number of outer iterations */
1830 outeriter += nri;
1831
1832 /* Update outer/inner flops */
1833
1834 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*270);
1835 }
The ultimate molecular dynamics simulation package