?  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_single / nb_kernel_ElecCoul_VdwNone_GeomW4W4_sse2_single.c
blob4f1a443067b813a92581f6dc84fa8aaac52e06ce
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_single 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_single.h"
49 #include "kernelutil_x86_sse2_single.h"
50
51 /*
52 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse2_single
53 * Electrostatics interaction: Coulomb
54 * VdW interaction: None
55 * Geometry: Water4-Water4
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse2_single
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,C,D refer to j loop unrolling done with SSE, e.g. for the four 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,jnrC,jnrD;
76 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
77 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
78 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
79 real rcutoff_scalar;
80 real *shiftvec,*fshift,*x,*f;
81 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
82 real scratch[4*DIM];
83 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
84 int vdwioffset1;
85 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
86 int vdwioffset2;
87 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
88 int vdwioffset3;
89 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
90 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
91 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
92 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
93 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
94 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
95 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
96 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
97 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
98 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
99 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
100 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
101 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
102 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
103 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
104 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
105 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
106 real *charge;
107 __m128 dummy_mask,cutoff_mask;
108 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
109 __m128 one = _mm_set1_ps(1.0);
110 __m128 two = _mm_set1_ps(2.0);
111 x = xx[0];
112 f = ff[0];
113
114 nri = nlist->nri;
115 iinr = nlist->iinr;
116 jindex = nlist->jindex;
117 jjnr = nlist->jjnr;
118 shiftidx = nlist->shift;
119 gid = nlist->gid;
120 shiftvec = fr->shift_vec[0];
121 fshift = fr->fshift[0];
122 facel = _mm_set1_ps(fr->epsfac);
123 charge = mdatoms->chargeA;
124
125 /* Setup water-specific parameters */
126 inr = nlist->iinr[0];
127 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
128 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
129 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
130
131 jq1 = _mm_set1_ps(charge[inr+1]);
132 jq2 = _mm_set1_ps(charge[inr+2]);
133 jq3 = _mm_set1_ps(charge[inr+3]);
134 qq11 = _mm_mul_ps(iq1,jq1);
135 qq12 = _mm_mul_ps(iq1,jq2);
136 qq13 = _mm_mul_ps(iq1,jq3);
137 qq21 = _mm_mul_ps(iq2,jq1);
138 qq22 = _mm_mul_ps(iq2,jq2);
139 qq23 = _mm_mul_ps(iq2,jq3);
140 qq31 = _mm_mul_ps(iq3,jq1);
141 qq32 = _mm_mul_ps(iq3,jq2);
142 qq33 = _mm_mul_ps(iq3,jq3);
143
144 /* Avoid stupid compiler warnings */
145 jnrA = jnrB = jnrC = jnrD = 0;
146 j_coord_offsetA = 0;
147 j_coord_offsetB = 0;
148 j_coord_offsetC = 0;
149 j_coord_offsetD = 0;
150
151 outeriter = 0;
152 inneriter = 0;
153
154 for(iidx=0;iidx<4*DIM;iidx++)
155 {
156 scratch[iidx] = 0.0;
157 }
158
159 /* Start outer loop over neighborlists */
160 for(iidx=0; iidx<nri; iidx++)
161 {
162 /* Load shift vector for this list */
163 i_shift_offset = DIM*shiftidx[iidx];
164
165 /* Load limits for loop over neighbors */
166 j_index_start = jindex[iidx];
167 j_index_end = jindex[iidx+1];
168
169 /* Get outer coordinate index */
170 inr = iinr[iidx];
171 i_coord_offset = DIM*inr;
172
173 /* Load i particle coords and add shift vector */
174 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
175 &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
176
177 fix1 = _mm_setzero_ps();
178 fiy1 = _mm_setzero_ps();
179 fiz1 = _mm_setzero_ps();
180 fix2 = _mm_setzero_ps();
181 fiy2 = _mm_setzero_ps();
182 fiz2 = _mm_setzero_ps();
183 fix3 = _mm_setzero_ps();
184 fiy3 = _mm_setzero_ps();
185 fiz3 = _mm_setzero_ps();
186
187 /* Reset potential sums */
188 velecsum = _mm_setzero_ps();
189
190 /* Start inner kernel loop */
191 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
192 {
193
194 /* Get j neighbor index, and coordinate index */
195 jnrA = jjnr[jidx];
196 jnrB = jjnr[jidx+1];
197 jnrC = jjnr[jidx+2];
198 jnrD = jjnr[jidx+3];
199 j_coord_offsetA = DIM*jnrA;
200 j_coord_offsetB = DIM*jnrB;
201 j_coord_offsetC = DIM*jnrC;
202 j_coord_offsetD = DIM*jnrD;
203
204 /* load j atom coordinates */
205 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
206 x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
207 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
208
209 /* Calculate displacement vector */
210 dx11 = _mm_sub_ps(ix1,jx1);
211 dy11 = _mm_sub_ps(iy1,jy1);
212 dz11 = _mm_sub_ps(iz1,jz1);
213 dx12 = _mm_sub_ps(ix1,jx2);
214 dy12 = _mm_sub_ps(iy1,jy2);
215 dz12 = _mm_sub_ps(iz1,jz2);
216 dx13 = _mm_sub_ps(ix1,jx3);
217 dy13 = _mm_sub_ps(iy1,jy3);
218 dz13 = _mm_sub_ps(iz1,jz3);
219 dx21 = _mm_sub_ps(ix2,jx1);
220 dy21 = _mm_sub_ps(iy2,jy1);
221 dz21 = _mm_sub_ps(iz2,jz1);
222 dx22 = _mm_sub_ps(ix2,jx2);
223 dy22 = _mm_sub_ps(iy2,jy2);
224 dz22 = _mm_sub_ps(iz2,jz2);
225 dx23 = _mm_sub_ps(ix2,jx3);
226 dy23 = _mm_sub_ps(iy2,jy3);
227 dz23 = _mm_sub_ps(iz2,jz3);
228 dx31 = _mm_sub_ps(ix3,jx1);
229 dy31 = _mm_sub_ps(iy3,jy1);
230 dz31 = _mm_sub_ps(iz3,jz1);
231 dx32 = _mm_sub_ps(ix3,jx2);
232 dy32 = _mm_sub_ps(iy3,jy2);
233 dz32 = _mm_sub_ps(iz3,jz2);
234 dx33 = _mm_sub_ps(ix3,jx3);
235 dy33 = _mm_sub_ps(iy3,jy3);
236 dz33 = _mm_sub_ps(iz3,jz3);
237
238 /* Calculate squared distance and things based on it */
239 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
240 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
241 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
242 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
243 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
244 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
245 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
246 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
247 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
248
249 rinv11 = gmx_mm_invsqrt_ps(rsq11);
250 rinv12 = gmx_mm_invsqrt_ps(rsq12);
251 rinv13 = gmx_mm_invsqrt_ps(rsq13);
252 rinv21 = gmx_mm_invsqrt_ps(rsq21);
253 rinv22 = gmx_mm_invsqrt_ps(rsq22);
254 rinv23 = gmx_mm_invsqrt_ps(rsq23);
255 rinv31 = gmx_mm_invsqrt_ps(rsq31);
256 rinv32 = gmx_mm_invsqrt_ps(rsq32);
257 rinv33 = gmx_mm_invsqrt_ps(rsq33);
258
259 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
260 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
261 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
262 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
263 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
264 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
265 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
266 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
267 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
268
269 fjx1 = _mm_setzero_ps();
270 fjy1 = _mm_setzero_ps();
271 fjz1 = _mm_setzero_ps();
272 fjx2 = _mm_setzero_ps();
273 fjy2 = _mm_setzero_ps();
274 fjz2 = _mm_setzero_ps();
275 fjx3 = _mm_setzero_ps();
276 fjy3 = _mm_setzero_ps();
277 fjz3 = _mm_setzero_ps();
278
279 /**************************
280 * CALCULATE INTERACTIONS *
281 **************************/
282
283 /* COULOMB ELECTROSTATICS */
284 velec = _mm_mul_ps(qq11,rinv11);
285 felec = _mm_mul_ps(velec,rinvsq11);
286
287 /* Update potential sum for this i atom from the interaction with this j atom. */
288 velecsum = _mm_add_ps(velecsum,velec);
289
290 fscal = felec;
291
292 /* Calculate temporary vectorial force */
293 tx = _mm_mul_ps(fscal,dx11);
294 ty = _mm_mul_ps(fscal,dy11);
295 tz = _mm_mul_ps(fscal,dz11);
296
297 /* Update vectorial force */
298 fix1 = _mm_add_ps(fix1,tx);
299 fiy1 = _mm_add_ps(fiy1,ty);
300 fiz1 = _mm_add_ps(fiz1,tz);
301
302 fjx1 = _mm_add_ps(fjx1,tx);
303 fjy1 = _mm_add_ps(fjy1,ty);
304 fjz1 = _mm_add_ps(fjz1,tz);
305
306 /**************************
307 * CALCULATE INTERACTIONS *
308 **************************/
309
310 /* COULOMB ELECTROSTATICS */
311 velec = _mm_mul_ps(qq12,rinv12);
312 felec = _mm_mul_ps(velec,rinvsq12);
313
314 /* Update potential sum for this i atom from the interaction with this j atom. */
315 velecsum = _mm_add_ps(velecsum,velec);
316
317 fscal = felec;
318
319 /* Calculate temporary vectorial force */
320 tx = _mm_mul_ps(fscal,dx12);
321 ty = _mm_mul_ps(fscal,dy12);
322 tz = _mm_mul_ps(fscal,dz12);
323
324 /* Update vectorial force */
325 fix1 = _mm_add_ps(fix1,tx);
326 fiy1 = _mm_add_ps(fiy1,ty);
327 fiz1 = _mm_add_ps(fiz1,tz);
328
329 fjx2 = _mm_add_ps(fjx2,tx);
330 fjy2 = _mm_add_ps(fjy2,ty);
331 fjz2 = _mm_add_ps(fjz2,tz);
332
333 /**************************
334 * CALCULATE INTERACTIONS *
335 **************************/
336
337 /* COULOMB ELECTROSTATICS */
338 velec = _mm_mul_ps(qq13,rinv13);
339 felec = _mm_mul_ps(velec,rinvsq13);
340
341 /* Update potential sum for this i atom from the interaction with this j atom. */
342 velecsum = _mm_add_ps(velecsum,velec);
343
344 fscal = felec;
345
346 /* Calculate temporary vectorial force */
347 tx = _mm_mul_ps(fscal,dx13);
348 ty = _mm_mul_ps(fscal,dy13);
349 tz = _mm_mul_ps(fscal,dz13);
350
351 /* Update vectorial force */
352 fix1 = _mm_add_ps(fix1,tx);
353 fiy1 = _mm_add_ps(fiy1,ty);
354 fiz1 = _mm_add_ps(fiz1,tz);
355
356 fjx3 = _mm_add_ps(fjx3,tx);
357 fjy3 = _mm_add_ps(fjy3,ty);
358 fjz3 = _mm_add_ps(fjz3,tz);
359
360 /**************************
361 * CALCULATE INTERACTIONS *
362 **************************/
363
364 /* COULOMB ELECTROSTATICS */
365 velec = _mm_mul_ps(qq21,rinv21);
366 felec = _mm_mul_ps(velec,rinvsq21);
367
368 /* Update potential sum for this i atom from the interaction with this j atom. */
369 velecsum = _mm_add_ps(velecsum,velec);
370
371 fscal = felec;
372
373 /* Calculate temporary vectorial force */
374 tx = _mm_mul_ps(fscal,dx21);
375 ty = _mm_mul_ps(fscal,dy21);
376 tz = _mm_mul_ps(fscal,dz21);
377
378 /* Update vectorial force */
379 fix2 = _mm_add_ps(fix2,tx);
380 fiy2 = _mm_add_ps(fiy2,ty);
381 fiz2 = _mm_add_ps(fiz2,tz);
382
383 fjx1 = _mm_add_ps(fjx1,tx);
384 fjy1 = _mm_add_ps(fjy1,ty);
385 fjz1 = _mm_add_ps(fjz1,tz);
386
387 /**************************
388 * CALCULATE INTERACTIONS *
389 **************************/
390
391 /* COULOMB ELECTROSTATICS */
392 velec = _mm_mul_ps(qq22,rinv22);
393 felec = _mm_mul_ps(velec,rinvsq22);
394
395 /* Update potential sum for this i atom from the interaction with this j atom. */
396 velecsum = _mm_add_ps(velecsum,velec);
397
398 fscal = felec;
399
400 /* Calculate temporary vectorial force */
401 tx = _mm_mul_ps(fscal,dx22);
402 ty = _mm_mul_ps(fscal,dy22);
403 tz = _mm_mul_ps(fscal,dz22);
404
405 /* Update vectorial force */
406 fix2 = _mm_add_ps(fix2,tx);
407 fiy2 = _mm_add_ps(fiy2,ty);
408 fiz2 = _mm_add_ps(fiz2,tz);
409
410 fjx2 = _mm_add_ps(fjx2,tx);
411 fjy2 = _mm_add_ps(fjy2,ty);
412 fjz2 = _mm_add_ps(fjz2,tz);
413
414 /**************************
415 * CALCULATE INTERACTIONS *
416 **************************/
417
418 /* COULOMB ELECTROSTATICS */
419 velec = _mm_mul_ps(qq23,rinv23);
420 felec = _mm_mul_ps(velec,rinvsq23);
421
422 /* Update potential sum for this i atom from the interaction with this j atom. */
423 velecsum = _mm_add_ps(velecsum,velec);
424
425 fscal = felec;
426
427 /* Calculate temporary vectorial force */
428 tx = _mm_mul_ps(fscal,dx23);
429 ty = _mm_mul_ps(fscal,dy23);
430 tz = _mm_mul_ps(fscal,dz23);
431
432 /* Update vectorial force */
433 fix2 = _mm_add_ps(fix2,tx);
434 fiy2 = _mm_add_ps(fiy2,ty);
435 fiz2 = _mm_add_ps(fiz2,tz);
436
437 fjx3 = _mm_add_ps(fjx3,tx);
438 fjy3 = _mm_add_ps(fjy3,ty);
439 fjz3 = _mm_add_ps(fjz3,tz);
440
441 /**************************
442 * CALCULATE INTERACTIONS *
443 **************************/
444
445 /* COULOMB ELECTROSTATICS */
446 velec = _mm_mul_ps(qq31,rinv31);
447 felec = _mm_mul_ps(velec,rinvsq31);
448
449 /* Update potential sum for this i atom from the interaction with this j atom. */
450 velecsum = _mm_add_ps(velecsum,velec);
451
452 fscal = felec;
453
454 /* Calculate temporary vectorial force */
455 tx = _mm_mul_ps(fscal,dx31);
456 ty = _mm_mul_ps(fscal,dy31);
457 tz = _mm_mul_ps(fscal,dz31);
458
459 /* Update vectorial force */
460 fix3 = _mm_add_ps(fix3,tx);
461 fiy3 = _mm_add_ps(fiy3,ty);
462 fiz3 = _mm_add_ps(fiz3,tz);
463
464 fjx1 = _mm_add_ps(fjx1,tx);
465 fjy1 = _mm_add_ps(fjy1,ty);
466 fjz1 = _mm_add_ps(fjz1,tz);
467
468 /**************************
469 * CALCULATE INTERACTIONS *
470 **************************/
471
472 /* COULOMB ELECTROSTATICS */
473 velec = _mm_mul_ps(qq32,rinv32);
474 felec = _mm_mul_ps(velec,rinvsq32);
475
476 /* Update potential sum for this i atom from the interaction with this j atom. */
477 velecsum = _mm_add_ps(velecsum,velec);
478
479 fscal = felec;
480
481 /* Calculate temporary vectorial force */
482 tx = _mm_mul_ps(fscal,dx32);
483 ty = _mm_mul_ps(fscal,dy32);
484 tz = _mm_mul_ps(fscal,dz32);
485
486 /* Update vectorial force */
487 fix3 = _mm_add_ps(fix3,tx);
488 fiy3 = _mm_add_ps(fiy3,ty);
489 fiz3 = _mm_add_ps(fiz3,tz);
490
491 fjx2 = _mm_add_ps(fjx2,tx);
492 fjy2 = _mm_add_ps(fjy2,ty);
493 fjz2 = _mm_add_ps(fjz2,tz);
494
495 /**************************
496 * CALCULATE INTERACTIONS *
497 **************************/
498
499 /* COULOMB ELECTROSTATICS */
500 velec = _mm_mul_ps(qq33,rinv33);
501 felec = _mm_mul_ps(velec,rinvsq33);
502
503 /* Update potential sum for this i atom from the interaction with this j atom. */
504 velecsum = _mm_add_ps(velecsum,velec);
505
506 fscal = felec;
507
508 /* Calculate temporary vectorial force */
509 tx = _mm_mul_ps(fscal,dx33);
510 ty = _mm_mul_ps(fscal,dy33);
511 tz = _mm_mul_ps(fscal,dz33);
512
513 /* Update vectorial force */
514 fix3 = _mm_add_ps(fix3,tx);
515 fiy3 = _mm_add_ps(fiy3,ty);
516 fiz3 = _mm_add_ps(fiz3,tz);
517
518 fjx3 = _mm_add_ps(fjx3,tx);
519 fjy3 = _mm_add_ps(fjy3,ty);
520 fjz3 = _mm_add_ps(fjz3,tz);
521
522 fjptrA = f+j_coord_offsetA;
523 fjptrB = f+j_coord_offsetB;
524 fjptrC = f+j_coord_offsetC;
525 fjptrD = f+j_coord_offsetD;
526
527 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
528 fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
529
530 /* Inner loop uses 252 flops */
531 }
532
533 if(jidx<j_index_end)
534 {
535
536 /* Get j neighbor index, and coordinate index */
537 jnrlistA = jjnr[jidx];
538 jnrlistB = jjnr[jidx+1];
539 jnrlistC = jjnr[jidx+2];
540 jnrlistD = jjnr[jidx+3];
541 /* Sign of each element will be negative for non-real atoms.
542 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
543 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
544 */
545 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
546 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
547 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
548 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
549 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
550 j_coord_offsetA = DIM*jnrA;
551 j_coord_offsetB = DIM*jnrB;
552 j_coord_offsetC = DIM*jnrC;
553 j_coord_offsetD = DIM*jnrD;
554
555 /* load j atom coordinates */
556 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
557 x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
558 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
559
560 /* Calculate displacement vector */
561 dx11 = _mm_sub_ps(ix1,jx1);
562 dy11 = _mm_sub_ps(iy1,jy1);
563 dz11 = _mm_sub_ps(iz1,jz1);
564 dx12 = _mm_sub_ps(ix1,jx2);
565 dy12 = _mm_sub_ps(iy1,jy2);
566 dz12 = _mm_sub_ps(iz1,jz2);
567 dx13 = _mm_sub_ps(ix1,jx3);
568 dy13 = _mm_sub_ps(iy1,jy3);
569 dz13 = _mm_sub_ps(iz1,jz3);
570 dx21 = _mm_sub_ps(ix2,jx1);
571 dy21 = _mm_sub_ps(iy2,jy1);
572 dz21 = _mm_sub_ps(iz2,jz1);
573 dx22 = _mm_sub_ps(ix2,jx2);
574 dy22 = _mm_sub_ps(iy2,jy2);
575 dz22 = _mm_sub_ps(iz2,jz2);
576 dx23 = _mm_sub_ps(ix2,jx3);
577 dy23 = _mm_sub_ps(iy2,jy3);
578 dz23 = _mm_sub_ps(iz2,jz3);
579 dx31 = _mm_sub_ps(ix3,jx1);
580 dy31 = _mm_sub_ps(iy3,jy1);
581 dz31 = _mm_sub_ps(iz3,jz1);
582 dx32 = _mm_sub_ps(ix3,jx2);
583 dy32 = _mm_sub_ps(iy3,jy2);
584 dz32 = _mm_sub_ps(iz3,jz2);
585 dx33 = _mm_sub_ps(ix3,jx3);
586 dy33 = _mm_sub_ps(iy3,jy3);
587 dz33 = _mm_sub_ps(iz3,jz3);
588
589 /* Calculate squared distance and things based on it */
590 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
591 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
592 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
593 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
594 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
595 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
596 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
597 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
598 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
599
600 rinv11 = gmx_mm_invsqrt_ps(rsq11);
601 rinv12 = gmx_mm_invsqrt_ps(rsq12);
602 rinv13 = gmx_mm_invsqrt_ps(rsq13);
603 rinv21 = gmx_mm_invsqrt_ps(rsq21);
604 rinv22 = gmx_mm_invsqrt_ps(rsq22);
605 rinv23 = gmx_mm_invsqrt_ps(rsq23);
606 rinv31 = gmx_mm_invsqrt_ps(rsq31);
607 rinv32 = gmx_mm_invsqrt_ps(rsq32);
608 rinv33 = gmx_mm_invsqrt_ps(rsq33);
609
610 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
611 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
612 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
613 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
614 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
615 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
616 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
617 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
618 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
619
620 fjx1 = _mm_setzero_ps();
621 fjy1 = _mm_setzero_ps();
622 fjz1 = _mm_setzero_ps();
623 fjx2 = _mm_setzero_ps();
624 fjy2 = _mm_setzero_ps();
625 fjz2 = _mm_setzero_ps();
626 fjx3 = _mm_setzero_ps();
627 fjy3 = _mm_setzero_ps();
628 fjz3 = _mm_setzero_ps();
629
630 /**************************
631 * CALCULATE INTERACTIONS *
632 **************************/
633
634 /* COULOMB ELECTROSTATICS */
635 velec = _mm_mul_ps(qq11,rinv11);
636 felec = _mm_mul_ps(velec,rinvsq11);
637
638 /* Update potential sum for this i atom from the interaction with this j atom. */
639 velec = _mm_andnot_ps(dummy_mask,velec);
640 velecsum = _mm_add_ps(velecsum,velec);
641
642 fscal = felec;
643
644 fscal = _mm_andnot_ps(dummy_mask,fscal);
645
646 /* Calculate temporary vectorial force */
647 tx = _mm_mul_ps(fscal,dx11);
648 ty = _mm_mul_ps(fscal,dy11);
649 tz = _mm_mul_ps(fscal,dz11);
650
651 /* Update vectorial force */
652 fix1 = _mm_add_ps(fix1,tx);
653 fiy1 = _mm_add_ps(fiy1,ty);
654 fiz1 = _mm_add_ps(fiz1,tz);
655
656 fjx1 = _mm_add_ps(fjx1,tx);
657 fjy1 = _mm_add_ps(fjy1,ty);
658 fjz1 = _mm_add_ps(fjz1,tz);
659
660 /**************************
661 * CALCULATE INTERACTIONS *
662 **************************/
663
664 /* COULOMB ELECTROSTATICS */
665 velec = _mm_mul_ps(qq12,rinv12);
666 felec = _mm_mul_ps(velec,rinvsq12);
667
668 /* Update potential sum for this i atom from the interaction with this j atom. */
669 velec = _mm_andnot_ps(dummy_mask,velec);
670 velecsum = _mm_add_ps(velecsum,velec);
671
672 fscal = felec;
673
674 fscal = _mm_andnot_ps(dummy_mask,fscal);
675
676 /* Calculate temporary vectorial force */
677 tx = _mm_mul_ps(fscal,dx12);
678 ty = _mm_mul_ps(fscal,dy12);
679 tz = _mm_mul_ps(fscal,dz12);
680
681 /* Update vectorial force */
682 fix1 = _mm_add_ps(fix1,tx);
683 fiy1 = _mm_add_ps(fiy1,ty);
684 fiz1 = _mm_add_ps(fiz1,tz);
685
686 fjx2 = _mm_add_ps(fjx2,tx);
687 fjy2 = _mm_add_ps(fjy2,ty);
688 fjz2 = _mm_add_ps(fjz2,tz);
689
690 /**************************
691 * CALCULATE INTERACTIONS *
692 **************************/
693
694 /* COULOMB ELECTROSTATICS */
695 velec = _mm_mul_ps(qq13,rinv13);
696 felec = _mm_mul_ps(velec,rinvsq13);
697
698 /* Update potential sum for this i atom from the interaction with this j atom. */
699 velec = _mm_andnot_ps(dummy_mask,velec);
700 velecsum = _mm_add_ps(velecsum,velec);
701
702 fscal = felec;
703
704 fscal = _mm_andnot_ps(dummy_mask,fscal);
705
706 /* Calculate temporary vectorial force */
707 tx = _mm_mul_ps(fscal,dx13);
708 ty = _mm_mul_ps(fscal,dy13);
709 tz = _mm_mul_ps(fscal,dz13);
710
711 /* Update vectorial force */
712 fix1 = _mm_add_ps(fix1,tx);
713 fiy1 = _mm_add_ps(fiy1,ty);
714 fiz1 = _mm_add_ps(fiz1,tz);
715
716 fjx3 = _mm_add_ps(fjx3,tx);
717 fjy3 = _mm_add_ps(fjy3,ty);
718 fjz3 = _mm_add_ps(fjz3,tz);
719
720 /**************************
721 * CALCULATE INTERACTIONS *
722 **************************/
723
724 /* COULOMB ELECTROSTATICS */
725 velec = _mm_mul_ps(qq21,rinv21);
726 felec = _mm_mul_ps(velec,rinvsq21);
727
728 /* Update potential sum for this i atom from the interaction with this j atom. */
729 velec = _mm_andnot_ps(dummy_mask,velec);
730 velecsum = _mm_add_ps(velecsum,velec);
731
732 fscal = felec;
733
734 fscal = _mm_andnot_ps(dummy_mask,fscal);
735
736 /* Calculate temporary vectorial force */
737 tx = _mm_mul_ps(fscal,dx21);
738 ty = _mm_mul_ps(fscal,dy21);
739 tz = _mm_mul_ps(fscal,dz21);
740
741 /* Update vectorial force */
742 fix2 = _mm_add_ps(fix2,tx);
743 fiy2 = _mm_add_ps(fiy2,ty);
744 fiz2 = _mm_add_ps(fiz2,tz);
745
746 fjx1 = _mm_add_ps(fjx1,tx);
747 fjy1 = _mm_add_ps(fjy1,ty);
748 fjz1 = _mm_add_ps(fjz1,tz);
749
750 /**************************
751 * CALCULATE INTERACTIONS *
752 **************************/
753
754 /* COULOMB ELECTROSTATICS */
755 velec = _mm_mul_ps(qq22,rinv22);
756 felec = _mm_mul_ps(velec,rinvsq22);
757
758 /* Update potential sum for this i atom from the interaction with this j atom. */
759 velec = _mm_andnot_ps(dummy_mask,velec);
760 velecsum = _mm_add_ps(velecsum,velec);
761
762 fscal = felec;
763
764 fscal = _mm_andnot_ps(dummy_mask,fscal);
765
766 /* Calculate temporary vectorial force */
767 tx = _mm_mul_ps(fscal,dx22);
768 ty = _mm_mul_ps(fscal,dy22);
769 tz = _mm_mul_ps(fscal,dz22);
770
771 /* Update vectorial force */
772 fix2 = _mm_add_ps(fix2,tx);
773 fiy2 = _mm_add_ps(fiy2,ty);
774 fiz2 = _mm_add_ps(fiz2,tz);
775
776 fjx2 = _mm_add_ps(fjx2,tx);
777 fjy2 = _mm_add_ps(fjy2,ty);
778 fjz2 = _mm_add_ps(fjz2,tz);
779
780 /**************************
781 * CALCULATE INTERACTIONS *
782 **************************/
783
784 /* COULOMB ELECTROSTATICS */
785 velec = _mm_mul_ps(qq23,rinv23);
786 felec = _mm_mul_ps(velec,rinvsq23);
787
788 /* Update potential sum for this i atom from the interaction with this j atom. */
789 velec = _mm_andnot_ps(dummy_mask,velec);
790 velecsum = _mm_add_ps(velecsum,velec);
791
792 fscal = felec;
793
794 fscal = _mm_andnot_ps(dummy_mask,fscal);
795
796 /* Calculate temporary vectorial force */
797 tx = _mm_mul_ps(fscal,dx23);
798 ty = _mm_mul_ps(fscal,dy23);
799 tz = _mm_mul_ps(fscal,dz23);
800
801 /* Update vectorial force */
802 fix2 = _mm_add_ps(fix2,tx);
803 fiy2 = _mm_add_ps(fiy2,ty);
804 fiz2 = _mm_add_ps(fiz2,tz);
805
806 fjx3 = _mm_add_ps(fjx3,tx);
807 fjy3 = _mm_add_ps(fjy3,ty);
808 fjz3 = _mm_add_ps(fjz3,tz);
809
810 /**************************
811 * CALCULATE INTERACTIONS *
812 **************************/
813
814 /* COULOMB ELECTROSTATICS */
815 velec = _mm_mul_ps(qq31,rinv31);
816 felec = _mm_mul_ps(velec,rinvsq31);
817
818 /* Update potential sum for this i atom from the interaction with this j atom. */
819 velec = _mm_andnot_ps(dummy_mask,velec);
820 velecsum = _mm_add_ps(velecsum,velec);
821
822 fscal = felec;
823
824 fscal = _mm_andnot_ps(dummy_mask,fscal);
825
826 /* Calculate temporary vectorial force */
827 tx = _mm_mul_ps(fscal,dx31);
828 ty = _mm_mul_ps(fscal,dy31);
829 tz = _mm_mul_ps(fscal,dz31);
830
831 /* Update vectorial force */
832 fix3 = _mm_add_ps(fix3,tx);
833 fiy3 = _mm_add_ps(fiy3,ty);
834 fiz3 = _mm_add_ps(fiz3,tz);
835
836 fjx1 = _mm_add_ps(fjx1,tx);
837 fjy1 = _mm_add_ps(fjy1,ty);
838 fjz1 = _mm_add_ps(fjz1,tz);
839
840 /**************************
841 * CALCULATE INTERACTIONS *
842 **************************/
843
844 /* COULOMB ELECTROSTATICS */
845 velec = _mm_mul_ps(qq32,rinv32);
846 felec = _mm_mul_ps(velec,rinvsq32);
847
848 /* Update potential sum for this i atom from the interaction with this j atom. */
849 velec = _mm_andnot_ps(dummy_mask,velec);
850 velecsum = _mm_add_ps(velecsum,velec);
851
852 fscal = felec;
853
854 fscal = _mm_andnot_ps(dummy_mask,fscal);
855
856 /* Calculate temporary vectorial force */
857 tx = _mm_mul_ps(fscal,dx32);
858 ty = _mm_mul_ps(fscal,dy32);
859 tz = _mm_mul_ps(fscal,dz32);
860
861 /* Update vectorial force */
862 fix3 = _mm_add_ps(fix3,tx);
863 fiy3 = _mm_add_ps(fiy3,ty);
864 fiz3 = _mm_add_ps(fiz3,tz);
865
866 fjx2 = _mm_add_ps(fjx2,tx);
867 fjy2 = _mm_add_ps(fjy2,ty);
868 fjz2 = _mm_add_ps(fjz2,tz);
869
870 /**************************
871 * CALCULATE INTERACTIONS *
872 **************************/
873
874 /* COULOMB ELECTROSTATICS */
875 velec = _mm_mul_ps(qq33,rinv33);
876 felec = _mm_mul_ps(velec,rinvsq33);
877
878 /* Update potential sum for this i atom from the interaction with this j atom. */
879 velec = _mm_andnot_ps(dummy_mask,velec);
880 velecsum = _mm_add_ps(velecsum,velec);
881
882 fscal = felec;
883
884 fscal = _mm_andnot_ps(dummy_mask,fscal);
885
886 /* Calculate temporary vectorial force */
887 tx = _mm_mul_ps(fscal,dx33);
888 ty = _mm_mul_ps(fscal,dy33);
889 tz = _mm_mul_ps(fscal,dz33);
890
891 /* Update vectorial force */
892 fix3 = _mm_add_ps(fix3,tx);
893 fiy3 = _mm_add_ps(fiy3,ty);
894 fiz3 = _mm_add_ps(fiz3,tz);
895
896 fjx3 = _mm_add_ps(fjx3,tx);
897 fjy3 = _mm_add_ps(fjy3,ty);
898 fjz3 = _mm_add_ps(fjz3,tz);
899
900 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
901 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
902 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
903 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
904
905 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
906 fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
907
908 /* Inner loop uses 252 flops */
909 }
910
911 /* End of innermost loop */
912
913 gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
914 f+i_coord_offset+DIM,fshift+i_shift_offset);
915
916 ggid = gid[iidx];
917 /* Update potential energies */
918 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
919
920 /* Increment number of inner iterations */
921 inneriter += j_index_end - j_index_start;
922
923 /* Outer loop uses 19 flops */
924 }
925
926 /* Increment number of outer iterations */
927 outeriter += nri;
928
929 /* Update outer/inner flops */
930
931 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*252);
932 }
933 /*
934 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse2_single
935 * Electrostatics interaction: Coulomb
936 * VdW interaction: None
937 * Geometry: Water4-Water4
938 * Calculate force/pot: Force
939 */
940 void
941 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse2_single
942 (t_nblist * gmx_restrict nlist,
943 rvec * gmx_restrict xx,
944 rvec * gmx_restrict ff,
945 t_forcerec * gmx_restrict fr,
946 t_mdatoms * gmx_restrict mdatoms,
947 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
948 t_nrnb * gmx_restrict nrnb)
949 {
950 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
951 * just 0 for non-waters.
952 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
953 * jnr indices corresponding to data put in the four positions in the SIMD register.
954 */
955 int i_shift_offset,i_coord_offset,outeriter,inneriter;
956 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
957 int jnrA,jnrB,jnrC,jnrD;
958 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
959 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
960 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
961 real rcutoff_scalar;
962 real *shiftvec,*fshift,*x,*f;
963 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
964 real scratch[4*DIM];
965 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
966 int vdwioffset1;
967 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
968 int vdwioffset2;
969 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
970 int vdwioffset3;
971 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
972 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
973 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
974 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
975 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
976 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
977 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
978 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
979 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
980 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
981 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
982 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
983 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
984 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
985 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
986 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
987 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
988 real *charge;
989 __m128 dummy_mask,cutoff_mask;
990 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
991 __m128 one = _mm_set1_ps(1.0);
992 __m128 two = _mm_set1_ps(2.0);
993 x = xx[0];
994 f = ff[0];
995
996 nri = nlist->nri;
997 iinr = nlist->iinr;
998 jindex = nlist->jindex;
999 jjnr = nlist->jjnr;
1000 shiftidx = nlist->shift;
1001 gid = nlist->gid;
1002 shiftvec = fr->shift_vec[0];
1003 fshift = fr->fshift[0];
1004 facel = _mm_set1_ps(fr->epsfac);
1005 charge = mdatoms->chargeA;
1006
1007 /* Setup water-specific parameters */
1008 inr = nlist->iinr[0];
1009 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1010 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1011 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1012
1013 jq1 = _mm_set1_ps(charge[inr+1]);
1014 jq2 = _mm_set1_ps(charge[inr+2]);
1015 jq3 = _mm_set1_ps(charge[inr+3]);
1016 qq11 = _mm_mul_ps(iq1,jq1);
1017 qq12 = _mm_mul_ps(iq1,jq2);
1018 qq13 = _mm_mul_ps(iq1,jq3);
1019 qq21 = _mm_mul_ps(iq2,jq1);
1020 qq22 = _mm_mul_ps(iq2,jq2);
1021 qq23 = _mm_mul_ps(iq2,jq3);
1022 qq31 = _mm_mul_ps(iq3,jq1);
1023 qq32 = _mm_mul_ps(iq3,jq2);
1024 qq33 = _mm_mul_ps(iq3,jq3);
1025
1026 /* Avoid stupid compiler warnings */
1027 jnrA = jnrB = jnrC = jnrD = 0;
1028 j_coord_offsetA = 0;
1029 j_coord_offsetB = 0;
1030 j_coord_offsetC = 0;
1031 j_coord_offsetD = 0;
1032
1033 outeriter = 0;
1034 inneriter = 0;
1035
1036 for(iidx=0;iidx<4*DIM;iidx++)
1037 {
1038 scratch[iidx] = 0.0;
1039 }
1040
1041 /* Start outer loop over neighborlists */
1042 for(iidx=0; iidx<nri; iidx++)
1043 {
1044 /* Load shift vector for this list */
1045 i_shift_offset = DIM*shiftidx[iidx];
1046
1047 /* Load limits for loop over neighbors */
1048 j_index_start = jindex[iidx];
1049 j_index_end = jindex[iidx+1];
1050
1051 /* Get outer coordinate index */
1052 inr = iinr[iidx];
1053 i_coord_offset = DIM*inr;
1054
1055 /* Load i particle coords and add shift vector */
1056 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1057 &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1058
1059 fix1 = _mm_setzero_ps();
1060 fiy1 = _mm_setzero_ps();
1061 fiz1 = _mm_setzero_ps();
1062 fix2 = _mm_setzero_ps();
1063 fiy2 = _mm_setzero_ps();
1064 fiz2 = _mm_setzero_ps();
1065 fix3 = _mm_setzero_ps();
1066 fiy3 = _mm_setzero_ps();
1067 fiz3 = _mm_setzero_ps();
1068
1069 /* Start inner kernel loop */
1070 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1071 {
1072
1073 /* Get j neighbor index, and coordinate index */
1074 jnrA = jjnr[jidx];
1075 jnrB = jjnr[jidx+1];
1076 jnrC = jjnr[jidx+2];
1077 jnrD = jjnr[jidx+3];
1078 j_coord_offsetA = DIM*jnrA;
1079 j_coord_offsetB = DIM*jnrB;
1080 j_coord_offsetC = DIM*jnrC;
1081 j_coord_offsetD = DIM*jnrD;
1082
1083 /* load j atom coordinates */
1084 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1085 x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1086 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1087
1088 /* Calculate displacement vector */
1089 dx11 = _mm_sub_ps(ix1,jx1);
1090 dy11 = _mm_sub_ps(iy1,jy1);
1091 dz11 = _mm_sub_ps(iz1,jz1);
1092 dx12 = _mm_sub_ps(ix1,jx2);
1093 dy12 = _mm_sub_ps(iy1,jy2);
1094 dz12 = _mm_sub_ps(iz1,jz2);
1095 dx13 = _mm_sub_ps(ix1,jx3);
1096 dy13 = _mm_sub_ps(iy1,jy3);
1097 dz13 = _mm_sub_ps(iz1,jz3);
1098 dx21 = _mm_sub_ps(ix2,jx1);
1099 dy21 = _mm_sub_ps(iy2,jy1);
1100 dz21 = _mm_sub_ps(iz2,jz1);
1101 dx22 = _mm_sub_ps(ix2,jx2);
1102 dy22 = _mm_sub_ps(iy2,jy2);
1103 dz22 = _mm_sub_ps(iz2,jz2);
1104 dx23 = _mm_sub_ps(ix2,jx3);
1105 dy23 = _mm_sub_ps(iy2,jy3);
1106 dz23 = _mm_sub_ps(iz2,jz3);
1107 dx31 = _mm_sub_ps(ix3,jx1);
1108 dy31 = _mm_sub_ps(iy3,jy1);
1109 dz31 = _mm_sub_ps(iz3,jz1);
1110 dx32 = _mm_sub_ps(ix3,jx2);
1111 dy32 = _mm_sub_ps(iy3,jy2);
1112 dz32 = _mm_sub_ps(iz3,jz2);
1113 dx33 = _mm_sub_ps(ix3,jx3);
1114 dy33 = _mm_sub_ps(iy3,jy3);
1115 dz33 = _mm_sub_ps(iz3,jz3);
1116
1117 /* Calculate squared distance and things based on it */
1118 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1119 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1120 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1121 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1122 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1123 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1124 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1125 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1126 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1127
1128 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1129 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1130 rinv13 = gmx_mm_invsqrt_ps(rsq13);
1131 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1132 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1133 rinv23 = gmx_mm_invsqrt_ps(rsq23);
1134 rinv31 = gmx_mm_invsqrt_ps(rsq31);
1135 rinv32 = gmx_mm_invsqrt_ps(rsq32);
1136 rinv33 = gmx_mm_invsqrt_ps(rsq33);
1137
1138 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1139 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1140 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1141 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1142 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1143 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1144 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1145 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1146 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1147
1148 fjx1 = _mm_setzero_ps();
1149 fjy1 = _mm_setzero_ps();
1150 fjz1 = _mm_setzero_ps();
1151 fjx2 = _mm_setzero_ps();
1152 fjy2 = _mm_setzero_ps();
1153 fjz2 = _mm_setzero_ps();
1154 fjx3 = _mm_setzero_ps();
1155 fjy3 = _mm_setzero_ps();
1156 fjz3 = _mm_setzero_ps();
1157
1158 /**************************
1159 * CALCULATE INTERACTIONS *
1160 **************************/
1161
1162 /* COULOMB ELECTROSTATICS */
1163 velec = _mm_mul_ps(qq11,rinv11);
1164 felec = _mm_mul_ps(velec,rinvsq11);
1165
1166 fscal = felec;
1167
1168 /* Calculate temporary vectorial force */
1169 tx = _mm_mul_ps(fscal,dx11);
1170 ty = _mm_mul_ps(fscal,dy11);
1171 tz = _mm_mul_ps(fscal,dz11);
1172
1173 /* Update vectorial force */
1174 fix1 = _mm_add_ps(fix1,tx);
1175 fiy1 = _mm_add_ps(fiy1,ty);
1176 fiz1 = _mm_add_ps(fiz1,tz);
1177
1178 fjx1 = _mm_add_ps(fjx1,tx);
1179 fjy1 = _mm_add_ps(fjy1,ty);
1180 fjz1 = _mm_add_ps(fjz1,tz);
1181
1182 /**************************
1183 * CALCULATE INTERACTIONS *
1184 **************************/
1185
1186 /* COULOMB ELECTROSTATICS */
1187 velec = _mm_mul_ps(qq12,rinv12);
1188 felec = _mm_mul_ps(velec,rinvsq12);
1189
1190 fscal = felec;
1191
1192 /* Calculate temporary vectorial force */
1193 tx = _mm_mul_ps(fscal,dx12);
1194 ty = _mm_mul_ps(fscal,dy12);
1195 tz = _mm_mul_ps(fscal,dz12);
1196
1197 /* Update vectorial force */
1198 fix1 = _mm_add_ps(fix1,tx);
1199 fiy1 = _mm_add_ps(fiy1,ty);
1200 fiz1 = _mm_add_ps(fiz1,tz);
1201
1202 fjx2 = _mm_add_ps(fjx2,tx);
1203 fjy2 = _mm_add_ps(fjy2,ty);
1204 fjz2 = _mm_add_ps(fjz2,tz);
1205
1206 /**************************
1207 * CALCULATE INTERACTIONS *
1208 **************************/
1209
1210 /* COULOMB ELECTROSTATICS */
1211 velec = _mm_mul_ps(qq13,rinv13);
1212 felec = _mm_mul_ps(velec,rinvsq13);
1213
1214 fscal = felec;
1215
1216 /* Calculate temporary vectorial force */
1217 tx = _mm_mul_ps(fscal,dx13);
1218 ty = _mm_mul_ps(fscal,dy13);
1219 tz = _mm_mul_ps(fscal,dz13);
1220
1221 /* Update vectorial force */
1222 fix1 = _mm_add_ps(fix1,tx);
1223 fiy1 = _mm_add_ps(fiy1,ty);
1224 fiz1 = _mm_add_ps(fiz1,tz);
1225
1226 fjx3 = _mm_add_ps(fjx3,tx);
1227 fjy3 = _mm_add_ps(fjy3,ty);
1228 fjz3 = _mm_add_ps(fjz3,tz);
1229
1230 /**************************
1231 * CALCULATE INTERACTIONS *
1232 **************************/
1233
1234 /* COULOMB ELECTROSTATICS */
1235 velec = _mm_mul_ps(qq21,rinv21);
1236 felec = _mm_mul_ps(velec,rinvsq21);
1237
1238 fscal = felec;
1239
1240 /* Calculate temporary vectorial force */
1241 tx = _mm_mul_ps(fscal,dx21);
1242 ty = _mm_mul_ps(fscal,dy21);
1243 tz = _mm_mul_ps(fscal,dz21);
1244
1245 /* Update vectorial force */
1246 fix2 = _mm_add_ps(fix2,tx);
1247 fiy2 = _mm_add_ps(fiy2,ty);
1248 fiz2 = _mm_add_ps(fiz2,tz);
1249
1250 fjx1 = _mm_add_ps(fjx1,tx);
1251 fjy1 = _mm_add_ps(fjy1,ty);
1252 fjz1 = _mm_add_ps(fjz1,tz);
1253
1254 /**************************
1255 * CALCULATE INTERACTIONS *
1256 **************************/
1257
1258 /* COULOMB ELECTROSTATICS */
1259 velec = _mm_mul_ps(qq22,rinv22);
1260 felec = _mm_mul_ps(velec,rinvsq22);
1261
1262 fscal = felec;
1263
1264 /* Calculate temporary vectorial force */
1265 tx = _mm_mul_ps(fscal,dx22);
1266 ty = _mm_mul_ps(fscal,dy22);
1267 tz = _mm_mul_ps(fscal,dz22);
1268
1269 /* Update vectorial force */
1270 fix2 = _mm_add_ps(fix2,tx);
1271 fiy2 = _mm_add_ps(fiy2,ty);
1272 fiz2 = _mm_add_ps(fiz2,tz);
1273
1274 fjx2 = _mm_add_ps(fjx2,tx);
1275 fjy2 = _mm_add_ps(fjy2,ty);
1276 fjz2 = _mm_add_ps(fjz2,tz);
1277
1278 /**************************
1279 * CALCULATE INTERACTIONS *
1280 **************************/
1281
1282 /* COULOMB ELECTROSTATICS */
1283 velec = _mm_mul_ps(qq23,rinv23);
1284 felec = _mm_mul_ps(velec,rinvsq23);
1285
1286 fscal = felec;
1287
1288 /* Calculate temporary vectorial force */
1289 tx = _mm_mul_ps(fscal,dx23);
1290 ty = _mm_mul_ps(fscal,dy23);
1291 tz = _mm_mul_ps(fscal,dz23);
1292
1293 /* Update vectorial force */
1294 fix2 = _mm_add_ps(fix2,tx);
1295 fiy2 = _mm_add_ps(fiy2,ty);
1296 fiz2 = _mm_add_ps(fiz2,tz);
1297
1298 fjx3 = _mm_add_ps(fjx3,tx);
1299 fjy3 = _mm_add_ps(fjy3,ty);
1300 fjz3 = _mm_add_ps(fjz3,tz);
1301
1302 /**************************
1303 * CALCULATE INTERACTIONS *
1304 **************************/
1305
1306 /* COULOMB ELECTROSTATICS */
1307 velec = _mm_mul_ps(qq31,rinv31);
1308 felec = _mm_mul_ps(velec,rinvsq31);
1309
1310 fscal = felec;
1311
1312 /* Calculate temporary vectorial force */
1313 tx = _mm_mul_ps(fscal,dx31);
1314 ty = _mm_mul_ps(fscal,dy31);
1315 tz = _mm_mul_ps(fscal,dz31);
1316
1317 /* Update vectorial force */
1318 fix3 = _mm_add_ps(fix3,tx);
1319 fiy3 = _mm_add_ps(fiy3,ty);
1320 fiz3 = _mm_add_ps(fiz3,tz);
1321
1322 fjx1 = _mm_add_ps(fjx1,tx);
1323 fjy1 = _mm_add_ps(fjy1,ty);
1324 fjz1 = _mm_add_ps(fjz1,tz);
1325
1326 /**************************
1327 * CALCULATE INTERACTIONS *
1328 **************************/
1329
1330 /* COULOMB ELECTROSTATICS */
1331 velec = _mm_mul_ps(qq32,rinv32);
1332 felec = _mm_mul_ps(velec,rinvsq32);
1333
1334 fscal = felec;
1335
1336 /* Calculate temporary vectorial force */
1337 tx = _mm_mul_ps(fscal,dx32);
1338 ty = _mm_mul_ps(fscal,dy32);
1339 tz = _mm_mul_ps(fscal,dz32);
1340
1341 /* Update vectorial force */
1342 fix3 = _mm_add_ps(fix3,tx);
1343 fiy3 = _mm_add_ps(fiy3,ty);
1344 fiz3 = _mm_add_ps(fiz3,tz);
1345
1346 fjx2 = _mm_add_ps(fjx2,tx);
1347 fjy2 = _mm_add_ps(fjy2,ty);
1348 fjz2 = _mm_add_ps(fjz2,tz);
1349
1350 /**************************
1351 * CALCULATE INTERACTIONS *
1352 **************************/
1353
1354 /* COULOMB ELECTROSTATICS */
1355 velec = _mm_mul_ps(qq33,rinv33);
1356 felec = _mm_mul_ps(velec,rinvsq33);
1357
1358 fscal = felec;
1359
1360 /* Calculate temporary vectorial force */
1361 tx = _mm_mul_ps(fscal,dx33);
1362 ty = _mm_mul_ps(fscal,dy33);
1363 tz = _mm_mul_ps(fscal,dz33);
1364
1365 /* Update vectorial force */
1366 fix3 = _mm_add_ps(fix3,tx);
1367 fiy3 = _mm_add_ps(fiy3,ty);
1368 fiz3 = _mm_add_ps(fiz3,tz);
1369
1370 fjx3 = _mm_add_ps(fjx3,tx);
1371 fjy3 = _mm_add_ps(fjy3,ty);
1372 fjz3 = _mm_add_ps(fjz3,tz);
1373
1374 fjptrA = f+j_coord_offsetA;
1375 fjptrB = f+j_coord_offsetB;
1376 fjptrC = f+j_coord_offsetC;
1377 fjptrD = f+j_coord_offsetD;
1378
1379 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1380 fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1381
1382 /* Inner loop uses 243 flops */
1383 }
1384
1385 if(jidx<j_index_end)
1386 {
1387
1388 /* Get j neighbor index, and coordinate index */
1389 jnrlistA = jjnr[jidx];
1390 jnrlistB = jjnr[jidx+1];
1391 jnrlistC = jjnr[jidx+2];
1392 jnrlistD = jjnr[jidx+3];
1393 /* Sign of each element will be negative for non-real atoms.
1394 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1395 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1396 */
1397 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1398 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1399 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1400 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1401 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1402 j_coord_offsetA = DIM*jnrA;
1403 j_coord_offsetB = DIM*jnrB;
1404 j_coord_offsetC = DIM*jnrC;
1405 j_coord_offsetD = DIM*jnrD;
1406
1407 /* load j atom coordinates */
1408 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1409 x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1410 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1411
1412 /* Calculate displacement vector */
1413 dx11 = _mm_sub_ps(ix1,jx1);
1414 dy11 = _mm_sub_ps(iy1,jy1);
1415 dz11 = _mm_sub_ps(iz1,jz1);
1416 dx12 = _mm_sub_ps(ix1,jx2);
1417 dy12 = _mm_sub_ps(iy1,jy2);
1418 dz12 = _mm_sub_ps(iz1,jz2);
1419 dx13 = _mm_sub_ps(ix1,jx3);
1420 dy13 = _mm_sub_ps(iy1,jy3);
1421 dz13 = _mm_sub_ps(iz1,jz3);
1422 dx21 = _mm_sub_ps(ix2,jx1);
1423 dy21 = _mm_sub_ps(iy2,jy1);
1424 dz21 = _mm_sub_ps(iz2,jz1);
1425 dx22 = _mm_sub_ps(ix2,jx2);
1426 dy22 = _mm_sub_ps(iy2,jy2);
1427 dz22 = _mm_sub_ps(iz2,jz2);
1428 dx23 = _mm_sub_ps(ix2,jx3);
1429 dy23 = _mm_sub_ps(iy2,jy3);
1430 dz23 = _mm_sub_ps(iz2,jz3);
1431 dx31 = _mm_sub_ps(ix3,jx1);
1432 dy31 = _mm_sub_ps(iy3,jy1);
1433 dz31 = _mm_sub_ps(iz3,jz1);
1434 dx32 = _mm_sub_ps(ix3,jx2);
1435 dy32 = _mm_sub_ps(iy3,jy2);
1436 dz32 = _mm_sub_ps(iz3,jz2);
1437 dx33 = _mm_sub_ps(ix3,jx3);
1438 dy33 = _mm_sub_ps(iy3,jy3);
1439 dz33 = _mm_sub_ps(iz3,jz3);
1440
1441 /* Calculate squared distance and things based on it */
1442 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1443 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1444 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1445 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1446 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1447 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1448 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1449 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1450 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1451
1452 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1453 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1454 rinv13 = gmx_mm_invsqrt_ps(rsq13);
1455 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1456 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1457 rinv23 = gmx_mm_invsqrt_ps(rsq23);
1458 rinv31 = gmx_mm_invsqrt_ps(rsq31);
1459 rinv32 = gmx_mm_invsqrt_ps(rsq32);
1460 rinv33 = gmx_mm_invsqrt_ps(rsq33);
1461
1462 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1463 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1464 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1465 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1466 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1467 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1468 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1469 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1470 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1471
1472 fjx1 = _mm_setzero_ps();
1473 fjy1 = _mm_setzero_ps();
1474 fjz1 = _mm_setzero_ps();
1475 fjx2 = _mm_setzero_ps();
1476 fjy2 = _mm_setzero_ps();
1477 fjz2 = _mm_setzero_ps();
1478 fjx3 = _mm_setzero_ps();
1479 fjy3 = _mm_setzero_ps();
1480 fjz3 = _mm_setzero_ps();
1481
1482 /**************************
1483 * CALCULATE INTERACTIONS *
1484 **************************/
1485
1486 /* COULOMB ELECTROSTATICS */
1487 velec = _mm_mul_ps(qq11,rinv11);
1488 felec = _mm_mul_ps(velec,rinvsq11);
1489
1490 fscal = felec;
1491
1492 fscal = _mm_andnot_ps(dummy_mask,fscal);
1493
1494 /* Calculate temporary vectorial force */
1495 tx = _mm_mul_ps(fscal,dx11);
1496 ty = _mm_mul_ps(fscal,dy11);
1497 tz = _mm_mul_ps(fscal,dz11);
1498
1499 /* Update vectorial force */
1500 fix1 = _mm_add_ps(fix1,tx);
1501 fiy1 = _mm_add_ps(fiy1,ty);
1502 fiz1 = _mm_add_ps(fiz1,tz);
1503
1504 fjx1 = _mm_add_ps(fjx1,tx);
1505 fjy1 = _mm_add_ps(fjy1,ty);
1506 fjz1 = _mm_add_ps(fjz1,tz);
1507
1508 /**************************
1509 * CALCULATE INTERACTIONS *
1510 **************************/
1511
1512 /* COULOMB ELECTROSTATICS */
1513 velec = _mm_mul_ps(qq12,rinv12);
1514 felec = _mm_mul_ps(velec,rinvsq12);
1515
1516 fscal = felec;
1517
1518 fscal = _mm_andnot_ps(dummy_mask,fscal);
1519
1520 /* Calculate temporary vectorial force */
1521 tx = _mm_mul_ps(fscal,dx12);
1522 ty = _mm_mul_ps(fscal,dy12);
1523 tz = _mm_mul_ps(fscal,dz12);
1524
1525 /* Update vectorial force */
1526 fix1 = _mm_add_ps(fix1,tx);
1527 fiy1 = _mm_add_ps(fiy1,ty);
1528 fiz1 = _mm_add_ps(fiz1,tz);
1529
1530 fjx2 = _mm_add_ps(fjx2,tx);
1531 fjy2 = _mm_add_ps(fjy2,ty);
1532 fjz2 = _mm_add_ps(fjz2,tz);
1533
1534 /**************************
1535 * CALCULATE INTERACTIONS *
1536 **************************/
1537
1538 /* COULOMB ELECTROSTATICS */
1539 velec = _mm_mul_ps(qq13,rinv13);
1540 felec = _mm_mul_ps(velec,rinvsq13);
1541
1542 fscal = felec;
1543
1544 fscal = _mm_andnot_ps(dummy_mask,fscal);
1545
1546 /* Calculate temporary vectorial force */
1547 tx = _mm_mul_ps(fscal,dx13);
1548 ty = _mm_mul_ps(fscal,dy13);
1549 tz = _mm_mul_ps(fscal,dz13);
1550
1551 /* Update vectorial force */
1552 fix1 = _mm_add_ps(fix1,tx);
1553 fiy1 = _mm_add_ps(fiy1,ty);
1554 fiz1 = _mm_add_ps(fiz1,tz);
1555
1556 fjx3 = _mm_add_ps(fjx3,tx);
1557 fjy3 = _mm_add_ps(fjy3,ty);
1558 fjz3 = _mm_add_ps(fjz3,tz);
1559
1560 /**************************
1561 * CALCULATE INTERACTIONS *
1562 **************************/
1563
1564 /* COULOMB ELECTROSTATICS */
1565 velec = _mm_mul_ps(qq21,rinv21);
1566 felec = _mm_mul_ps(velec,rinvsq21);
1567
1568 fscal = felec;
1569
1570 fscal = _mm_andnot_ps(dummy_mask,fscal);
1571
1572 /* Calculate temporary vectorial force */
1573 tx = _mm_mul_ps(fscal,dx21);
1574 ty = _mm_mul_ps(fscal,dy21);
1575 tz = _mm_mul_ps(fscal,dz21);
1576
1577 /* Update vectorial force */
1578 fix2 = _mm_add_ps(fix2,tx);
1579 fiy2 = _mm_add_ps(fiy2,ty);
1580 fiz2 = _mm_add_ps(fiz2,tz);
1581
1582 fjx1 = _mm_add_ps(fjx1,tx);
1583 fjy1 = _mm_add_ps(fjy1,ty);
1584 fjz1 = _mm_add_ps(fjz1,tz);
1585
1586 /**************************
1587 * CALCULATE INTERACTIONS *
1588 **************************/
1589
1590 /* COULOMB ELECTROSTATICS */
1591 velec = _mm_mul_ps(qq22,rinv22);
1592 felec = _mm_mul_ps(velec,rinvsq22);
1593
1594 fscal = felec;
1595
1596 fscal = _mm_andnot_ps(dummy_mask,fscal);
1597
1598 /* Calculate temporary vectorial force */
1599 tx = _mm_mul_ps(fscal,dx22);
1600 ty = _mm_mul_ps(fscal,dy22);
1601 tz = _mm_mul_ps(fscal,dz22);
1602
1603 /* Update vectorial force */
1604 fix2 = _mm_add_ps(fix2,tx);
1605 fiy2 = _mm_add_ps(fiy2,ty);
1606 fiz2 = _mm_add_ps(fiz2,tz);
1607
1608 fjx2 = _mm_add_ps(fjx2,tx);
1609 fjy2 = _mm_add_ps(fjy2,ty);
1610 fjz2 = _mm_add_ps(fjz2,tz);
1611
1612 /**************************
1613 * CALCULATE INTERACTIONS *
1614 **************************/
1615
1616 /* COULOMB ELECTROSTATICS */
1617 velec = _mm_mul_ps(qq23,rinv23);
1618 felec = _mm_mul_ps(velec,rinvsq23);
1619
1620 fscal = felec;
1621
1622 fscal = _mm_andnot_ps(dummy_mask,fscal);
1623
1624 /* Calculate temporary vectorial force */
1625 tx = _mm_mul_ps(fscal,dx23);
1626 ty = _mm_mul_ps(fscal,dy23);
1627 tz = _mm_mul_ps(fscal,dz23);
1628
1629 /* Update vectorial force */
1630 fix2 = _mm_add_ps(fix2,tx);
1631 fiy2 = _mm_add_ps(fiy2,ty);
1632 fiz2 = _mm_add_ps(fiz2,tz);
1633
1634 fjx3 = _mm_add_ps(fjx3,tx);
1635 fjy3 = _mm_add_ps(fjy3,ty);
1636 fjz3 = _mm_add_ps(fjz3,tz);
1637
1638 /**************************
1639 * CALCULATE INTERACTIONS *
1640 **************************/
1641
1642 /* COULOMB ELECTROSTATICS */
1643 velec = _mm_mul_ps(qq31,rinv31);
1644 felec = _mm_mul_ps(velec,rinvsq31);
1645
1646 fscal = felec;
1647
1648 fscal = _mm_andnot_ps(dummy_mask,fscal);
1649
1650 /* Calculate temporary vectorial force */
1651 tx = _mm_mul_ps(fscal,dx31);
1652 ty = _mm_mul_ps(fscal,dy31);
1653 tz = _mm_mul_ps(fscal,dz31);
1654
1655 /* Update vectorial force */
1656 fix3 = _mm_add_ps(fix3,tx);
1657 fiy3 = _mm_add_ps(fiy3,ty);
1658 fiz3 = _mm_add_ps(fiz3,tz);
1659
1660 fjx1 = _mm_add_ps(fjx1,tx);
1661 fjy1 = _mm_add_ps(fjy1,ty);
1662 fjz1 = _mm_add_ps(fjz1,tz);
1663
1664 /**************************
1665 * CALCULATE INTERACTIONS *
1666 **************************/
1667
1668 /* COULOMB ELECTROSTATICS */
1669 velec = _mm_mul_ps(qq32,rinv32);
1670 felec = _mm_mul_ps(velec,rinvsq32);
1671
1672 fscal = felec;
1673
1674 fscal = _mm_andnot_ps(dummy_mask,fscal);
1675
1676 /* Calculate temporary vectorial force */
1677 tx = _mm_mul_ps(fscal,dx32);
1678 ty = _mm_mul_ps(fscal,dy32);
1679 tz = _mm_mul_ps(fscal,dz32);
1680
1681 /* Update vectorial force */
1682 fix3 = _mm_add_ps(fix3,tx);
1683 fiy3 = _mm_add_ps(fiy3,ty);
1684 fiz3 = _mm_add_ps(fiz3,tz);
1685
1686 fjx2 = _mm_add_ps(fjx2,tx);
1687 fjy2 = _mm_add_ps(fjy2,ty);
1688 fjz2 = _mm_add_ps(fjz2,tz);
1689
1690 /**************************
1691 * CALCULATE INTERACTIONS *
1692 **************************/
1693
1694 /* COULOMB ELECTROSTATICS */
1695 velec = _mm_mul_ps(qq33,rinv33);
1696 felec = _mm_mul_ps(velec,rinvsq33);
1697
1698 fscal = felec;
1699
1700 fscal = _mm_andnot_ps(dummy_mask,fscal);
1701
1702 /* Calculate temporary vectorial force */
1703 tx = _mm_mul_ps(fscal,dx33);
1704 ty = _mm_mul_ps(fscal,dy33);
1705 tz = _mm_mul_ps(fscal,dz33);
1706
1707 /* Update vectorial force */
1708 fix3 = _mm_add_ps(fix3,tx);
1709 fiy3 = _mm_add_ps(fiy3,ty);
1710 fiz3 = _mm_add_ps(fiz3,tz);
1711
1712 fjx3 = _mm_add_ps(fjx3,tx);
1713 fjy3 = _mm_add_ps(fjy3,ty);
1714 fjz3 = _mm_add_ps(fjz3,tz);
1715
1716 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1717 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1718 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1719 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1720
1721 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1722 fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1723
1724 /* Inner loop uses 243 flops */
1725 }
1726
1727 /* End of innermost loop */
1728
1729 gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1730 f+i_coord_offset+DIM,fshift+i_shift_offset);
1731
1732 /* Increment number of inner iterations */
1733 inneriter += j_index_end - j_index_start;
1734
1735 /* Outer loop uses 18 flops */
1736 }
1737
1738 /* Increment number of outer iterations */
1739 outeriter += nri;
1740
1741 /* Update outer/inner flops */
1742
1743 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1744 }
The ultimate molecular dynamics simulation package