?  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_VdwLJ_GeomW3W3_sse2_single.c
blobf70c55a4f30c8f5c4f5e088043cf0a7907bea1b0
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_VdwLJ_GeomW3W3_VF_sse2_single
53 * Electrostatics interaction: Coulomb
54 * VdW interaction: LennardJones
55 * Geometry: Water3-Water3
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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 vdwioffset0;
85 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
86 int vdwioffset1;
87 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
88 int vdwioffset2;
89 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
90 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
91 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
92 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
93 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
94 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
95 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
96 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
97 __m128 dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
98 __m128 dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
99 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
100 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
101 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
102 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
103 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
104 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
105 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
106 real *charge;
107 int nvdwtype;
108 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
109 int *vdwtype;
110 real *vdwparam;
111 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
112 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
113 __m128 dummy_mask,cutoff_mask;
114 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
115 __m128 one = _mm_set1_ps(1.0);
116 __m128 two = _mm_set1_ps(2.0);
117 x = xx[0];
118 f = ff[0];
119
120 nri = nlist->nri;
121 iinr = nlist->iinr;
122 jindex = nlist->jindex;
123 jjnr = nlist->jjnr;
124 shiftidx = nlist->shift;
125 gid = nlist->gid;
126 shiftvec = fr->shift_vec[0];
127 fshift = fr->fshift[0];
128 facel = _mm_set1_ps(fr->epsfac);
129 charge = mdatoms->chargeA;
130 nvdwtype = fr->ntype;
131 vdwparam = fr->nbfp;
132 vdwtype = mdatoms->typeA;
133
134 /* Setup water-specific parameters */
135 inr = nlist->iinr[0];
136 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
137 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
138 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
139 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
140
141 jq0 = _mm_set1_ps(charge[inr+0]);
142 jq1 = _mm_set1_ps(charge[inr+1]);
143 jq2 = _mm_set1_ps(charge[inr+2]);
144 vdwjidx0A = 2*vdwtype[inr+0];
145 qq00 = _mm_mul_ps(iq0,jq0);
146 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
147 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
148 qq01 = _mm_mul_ps(iq0,jq1);
149 qq02 = _mm_mul_ps(iq0,jq2);
150 qq10 = _mm_mul_ps(iq1,jq0);
151 qq11 = _mm_mul_ps(iq1,jq1);
152 qq12 = _mm_mul_ps(iq1,jq2);
153 qq20 = _mm_mul_ps(iq2,jq0);
154 qq21 = _mm_mul_ps(iq2,jq1);
155 qq22 = _mm_mul_ps(iq2,jq2);
156
157 /* Avoid stupid compiler warnings */
158 jnrA = jnrB = jnrC = jnrD = 0;
159 j_coord_offsetA = 0;
160 j_coord_offsetB = 0;
161 j_coord_offsetC = 0;
162 j_coord_offsetD = 0;
163
164 outeriter = 0;
165 inneriter = 0;
166
167 for(iidx=0;iidx<4*DIM;iidx++)
168 {
169 scratch[iidx] = 0.0;
170 }
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_ps(shiftvec+i_shift_offset,x+i_coord_offset,
188 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
189
190 fix0 = _mm_setzero_ps();
191 fiy0 = _mm_setzero_ps();
192 fiz0 = _mm_setzero_ps();
193 fix1 = _mm_setzero_ps();
194 fiy1 = _mm_setzero_ps();
195 fiz1 = _mm_setzero_ps();
196 fix2 = _mm_setzero_ps();
197 fiy2 = _mm_setzero_ps();
198 fiz2 = _mm_setzero_ps();
199
200 /* Reset potential sums */
201 velecsum = _mm_setzero_ps();
202 vvdwsum = _mm_setzero_ps();
203
204 /* Start inner kernel loop */
205 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
206 {
207
208 /* Get j neighbor index, and coordinate index */
209 jnrA = jjnr[jidx];
210 jnrB = jjnr[jidx+1];
211 jnrC = jjnr[jidx+2];
212 jnrD = jjnr[jidx+3];
213 j_coord_offsetA = DIM*jnrA;
214 j_coord_offsetB = DIM*jnrB;
215 j_coord_offsetC = DIM*jnrC;
216 j_coord_offsetD = DIM*jnrD;
217
218 /* load j atom coordinates */
219 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
220 x+j_coord_offsetC,x+j_coord_offsetD,
221 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
222
223 /* Calculate displacement vector */
224 dx00 = _mm_sub_ps(ix0,jx0);
225 dy00 = _mm_sub_ps(iy0,jy0);
226 dz00 = _mm_sub_ps(iz0,jz0);
227 dx01 = _mm_sub_ps(ix0,jx1);
228 dy01 = _mm_sub_ps(iy0,jy1);
229 dz01 = _mm_sub_ps(iz0,jz1);
230 dx02 = _mm_sub_ps(ix0,jx2);
231 dy02 = _mm_sub_ps(iy0,jy2);
232 dz02 = _mm_sub_ps(iz0,jz2);
233 dx10 = _mm_sub_ps(ix1,jx0);
234 dy10 = _mm_sub_ps(iy1,jy0);
235 dz10 = _mm_sub_ps(iz1,jz0);
236 dx11 = _mm_sub_ps(ix1,jx1);
237 dy11 = _mm_sub_ps(iy1,jy1);
238 dz11 = _mm_sub_ps(iz1,jz1);
239 dx12 = _mm_sub_ps(ix1,jx2);
240 dy12 = _mm_sub_ps(iy1,jy2);
241 dz12 = _mm_sub_ps(iz1,jz2);
242 dx20 = _mm_sub_ps(ix2,jx0);
243 dy20 = _mm_sub_ps(iy2,jy0);
244 dz20 = _mm_sub_ps(iz2,jz0);
245 dx21 = _mm_sub_ps(ix2,jx1);
246 dy21 = _mm_sub_ps(iy2,jy1);
247 dz21 = _mm_sub_ps(iz2,jz1);
248 dx22 = _mm_sub_ps(ix2,jx2);
249 dy22 = _mm_sub_ps(iy2,jy2);
250 dz22 = _mm_sub_ps(iz2,jz2);
251
252 /* Calculate squared distance and things based on it */
253 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
254 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
255 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
256 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
257 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
258 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
259 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
260 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
261 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
262
263 rinv00 = gmx_mm_invsqrt_ps(rsq00);
264 rinv01 = gmx_mm_invsqrt_ps(rsq01);
265 rinv02 = gmx_mm_invsqrt_ps(rsq02);
266 rinv10 = gmx_mm_invsqrt_ps(rsq10);
267 rinv11 = gmx_mm_invsqrt_ps(rsq11);
268 rinv12 = gmx_mm_invsqrt_ps(rsq12);
269 rinv20 = gmx_mm_invsqrt_ps(rsq20);
270 rinv21 = gmx_mm_invsqrt_ps(rsq21);
271 rinv22 = gmx_mm_invsqrt_ps(rsq22);
272
273 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
274 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
275 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
276 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
277 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
278 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
279 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
280 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
281 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
282
283 fjx0 = _mm_setzero_ps();
284 fjy0 = _mm_setzero_ps();
285 fjz0 = _mm_setzero_ps();
286 fjx1 = _mm_setzero_ps();
287 fjy1 = _mm_setzero_ps();
288 fjz1 = _mm_setzero_ps();
289 fjx2 = _mm_setzero_ps();
290 fjy2 = _mm_setzero_ps();
291 fjz2 = _mm_setzero_ps();
292
293 /**************************
294 * CALCULATE INTERACTIONS *
295 **************************/
296
297 /* COULOMB ELECTROSTATICS */
298 velec = _mm_mul_ps(qq00,rinv00);
299 felec = _mm_mul_ps(velec,rinvsq00);
300
301 /* LENNARD-JONES DISPERSION/REPULSION */
302
303 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
304 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
305 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
306 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
307 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
308
309 /* Update potential sum for this i atom from the interaction with this j atom. */
310 velecsum = _mm_add_ps(velecsum,velec);
311 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
312
313 fscal = _mm_add_ps(felec,fvdw);
314
315 /* Calculate temporary vectorial force */
316 tx = _mm_mul_ps(fscal,dx00);
317 ty = _mm_mul_ps(fscal,dy00);
318 tz = _mm_mul_ps(fscal,dz00);
319
320 /* Update vectorial force */
321 fix0 = _mm_add_ps(fix0,tx);
322 fiy0 = _mm_add_ps(fiy0,ty);
323 fiz0 = _mm_add_ps(fiz0,tz);
324
325 fjx0 = _mm_add_ps(fjx0,tx);
326 fjy0 = _mm_add_ps(fjy0,ty);
327 fjz0 = _mm_add_ps(fjz0,tz);
328
329 /**************************
330 * CALCULATE INTERACTIONS *
331 **************************/
332
333 /* COULOMB ELECTROSTATICS */
334 velec = _mm_mul_ps(qq01,rinv01);
335 felec = _mm_mul_ps(velec,rinvsq01);
336
337 /* Update potential sum for this i atom from the interaction with this j atom. */
338 velecsum = _mm_add_ps(velecsum,velec);
339
340 fscal = felec;
341
342 /* Calculate temporary vectorial force */
343 tx = _mm_mul_ps(fscal,dx01);
344 ty = _mm_mul_ps(fscal,dy01);
345 tz = _mm_mul_ps(fscal,dz01);
346
347 /* Update vectorial force */
348 fix0 = _mm_add_ps(fix0,tx);
349 fiy0 = _mm_add_ps(fiy0,ty);
350 fiz0 = _mm_add_ps(fiz0,tz);
351
352 fjx1 = _mm_add_ps(fjx1,tx);
353 fjy1 = _mm_add_ps(fjy1,ty);
354 fjz1 = _mm_add_ps(fjz1,tz);
355
356 /**************************
357 * CALCULATE INTERACTIONS *
358 **************************/
359
360 /* COULOMB ELECTROSTATICS */
361 velec = _mm_mul_ps(qq02,rinv02);
362 felec = _mm_mul_ps(velec,rinvsq02);
363
364 /* Update potential sum for this i atom from the interaction with this j atom. */
365 velecsum = _mm_add_ps(velecsum,velec);
366
367 fscal = felec;
368
369 /* Calculate temporary vectorial force */
370 tx = _mm_mul_ps(fscal,dx02);
371 ty = _mm_mul_ps(fscal,dy02);
372 tz = _mm_mul_ps(fscal,dz02);
373
374 /* Update vectorial force */
375 fix0 = _mm_add_ps(fix0,tx);
376 fiy0 = _mm_add_ps(fiy0,ty);
377 fiz0 = _mm_add_ps(fiz0,tz);
378
379 fjx2 = _mm_add_ps(fjx2,tx);
380 fjy2 = _mm_add_ps(fjy2,ty);
381 fjz2 = _mm_add_ps(fjz2,tz);
382
383 /**************************
384 * CALCULATE INTERACTIONS *
385 **************************/
386
387 /* COULOMB ELECTROSTATICS */
388 velec = _mm_mul_ps(qq10,rinv10);
389 felec = _mm_mul_ps(velec,rinvsq10);
390
391 /* Update potential sum for this i atom from the interaction with this j atom. */
392 velecsum = _mm_add_ps(velecsum,velec);
393
394 fscal = felec;
395
396 /* Calculate temporary vectorial force */
397 tx = _mm_mul_ps(fscal,dx10);
398 ty = _mm_mul_ps(fscal,dy10);
399 tz = _mm_mul_ps(fscal,dz10);
400
401 /* Update vectorial force */
402 fix1 = _mm_add_ps(fix1,tx);
403 fiy1 = _mm_add_ps(fiy1,ty);
404 fiz1 = _mm_add_ps(fiz1,tz);
405
406 fjx0 = _mm_add_ps(fjx0,tx);
407 fjy0 = _mm_add_ps(fjy0,ty);
408 fjz0 = _mm_add_ps(fjz0,tz);
409
410 /**************************
411 * CALCULATE INTERACTIONS *
412 **************************/
413
414 /* COULOMB ELECTROSTATICS */
415 velec = _mm_mul_ps(qq11,rinv11);
416 felec = _mm_mul_ps(velec,rinvsq11);
417
418 /* Update potential sum for this i atom from the interaction with this j atom. */
419 velecsum = _mm_add_ps(velecsum,velec);
420
421 fscal = felec;
422
423 /* Calculate temporary vectorial force */
424 tx = _mm_mul_ps(fscal,dx11);
425 ty = _mm_mul_ps(fscal,dy11);
426 tz = _mm_mul_ps(fscal,dz11);
427
428 /* Update vectorial force */
429 fix1 = _mm_add_ps(fix1,tx);
430 fiy1 = _mm_add_ps(fiy1,ty);
431 fiz1 = _mm_add_ps(fiz1,tz);
432
433 fjx1 = _mm_add_ps(fjx1,tx);
434 fjy1 = _mm_add_ps(fjy1,ty);
435 fjz1 = _mm_add_ps(fjz1,tz);
436
437 /**************************
438 * CALCULATE INTERACTIONS *
439 **************************/
440
441 /* COULOMB ELECTROSTATICS */
442 velec = _mm_mul_ps(qq12,rinv12);
443 felec = _mm_mul_ps(velec,rinvsq12);
444
445 /* Update potential sum for this i atom from the interaction with this j atom. */
446 velecsum = _mm_add_ps(velecsum,velec);
447
448 fscal = felec;
449
450 /* Calculate temporary vectorial force */
451 tx = _mm_mul_ps(fscal,dx12);
452 ty = _mm_mul_ps(fscal,dy12);
453 tz = _mm_mul_ps(fscal,dz12);
454
455 /* Update vectorial force */
456 fix1 = _mm_add_ps(fix1,tx);
457 fiy1 = _mm_add_ps(fiy1,ty);
458 fiz1 = _mm_add_ps(fiz1,tz);
459
460 fjx2 = _mm_add_ps(fjx2,tx);
461 fjy2 = _mm_add_ps(fjy2,ty);
462 fjz2 = _mm_add_ps(fjz2,tz);
463
464 /**************************
465 * CALCULATE INTERACTIONS *
466 **************************/
467
468 /* COULOMB ELECTROSTATICS */
469 velec = _mm_mul_ps(qq20,rinv20);
470 felec = _mm_mul_ps(velec,rinvsq20);
471
472 /* Update potential sum for this i atom from the interaction with this j atom. */
473 velecsum = _mm_add_ps(velecsum,velec);
474
475 fscal = felec;
476
477 /* Calculate temporary vectorial force */
478 tx = _mm_mul_ps(fscal,dx20);
479 ty = _mm_mul_ps(fscal,dy20);
480 tz = _mm_mul_ps(fscal,dz20);
481
482 /* Update vectorial force */
483 fix2 = _mm_add_ps(fix2,tx);
484 fiy2 = _mm_add_ps(fiy2,ty);
485 fiz2 = _mm_add_ps(fiz2,tz);
486
487 fjx0 = _mm_add_ps(fjx0,tx);
488 fjy0 = _mm_add_ps(fjy0,ty);
489 fjz0 = _mm_add_ps(fjz0,tz);
490
491 /**************************
492 * CALCULATE INTERACTIONS *
493 **************************/
494
495 /* COULOMB ELECTROSTATICS */
496 velec = _mm_mul_ps(qq21,rinv21);
497 felec = _mm_mul_ps(velec,rinvsq21);
498
499 /* Update potential sum for this i atom from the interaction with this j atom. */
500 velecsum = _mm_add_ps(velecsum,velec);
501
502 fscal = felec;
503
504 /* Calculate temporary vectorial force */
505 tx = _mm_mul_ps(fscal,dx21);
506 ty = _mm_mul_ps(fscal,dy21);
507 tz = _mm_mul_ps(fscal,dz21);
508
509 /* Update vectorial force */
510 fix2 = _mm_add_ps(fix2,tx);
511 fiy2 = _mm_add_ps(fiy2,ty);
512 fiz2 = _mm_add_ps(fiz2,tz);
513
514 fjx1 = _mm_add_ps(fjx1,tx);
515 fjy1 = _mm_add_ps(fjy1,ty);
516 fjz1 = _mm_add_ps(fjz1,tz);
517
518 /**************************
519 * CALCULATE INTERACTIONS *
520 **************************/
521
522 /* COULOMB ELECTROSTATICS */
523 velec = _mm_mul_ps(qq22,rinv22);
524 felec = _mm_mul_ps(velec,rinvsq22);
525
526 /* Update potential sum for this i atom from the interaction with this j atom. */
527 velecsum = _mm_add_ps(velecsum,velec);
528
529 fscal = felec;
530
531 /* Calculate temporary vectorial force */
532 tx = _mm_mul_ps(fscal,dx22);
533 ty = _mm_mul_ps(fscal,dy22);
534 tz = _mm_mul_ps(fscal,dz22);
535
536 /* Update vectorial force */
537 fix2 = _mm_add_ps(fix2,tx);
538 fiy2 = _mm_add_ps(fiy2,ty);
539 fiz2 = _mm_add_ps(fiz2,tz);
540
541 fjx2 = _mm_add_ps(fjx2,tx);
542 fjy2 = _mm_add_ps(fjy2,ty);
543 fjz2 = _mm_add_ps(fjz2,tz);
544
545 fjptrA = f+j_coord_offsetA;
546 fjptrB = f+j_coord_offsetB;
547 fjptrC = f+j_coord_offsetC;
548 fjptrD = f+j_coord_offsetD;
549
550 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
551 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
552
553 /* Inner loop uses 264 flops */
554 }
555
556 if(jidx<j_index_end)
557 {
558
559 /* Get j neighbor index, and coordinate index */
560 jnrlistA = jjnr[jidx];
561 jnrlistB = jjnr[jidx+1];
562 jnrlistC = jjnr[jidx+2];
563 jnrlistD = jjnr[jidx+3];
564 /* Sign of each element will be negative for non-real atoms.
565 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
566 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
567 */
568 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
569 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
570 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
571 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
572 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
573 j_coord_offsetA = DIM*jnrA;
574 j_coord_offsetB = DIM*jnrB;
575 j_coord_offsetC = DIM*jnrC;
576 j_coord_offsetD = DIM*jnrD;
577
578 /* load j atom coordinates */
579 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
580 x+j_coord_offsetC,x+j_coord_offsetD,
581 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
582
583 /* Calculate displacement vector */
584 dx00 = _mm_sub_ps(ix0,jx0);
585 dy00 = _mm_sub_ps(iy0,jy0);
586 dz00 = _mm_sub_ps(iz0,jz0);
587 dx01 = _mm_sub_ps(ix0,jx1);
588 dy01 = _mm_sub_ps(iy0,jy1);
589 dz01 = _mm_sub_ps(iz0,jz1);
590 dx02 = _mm_sub_ps(ix0,jx2);
591 dy02 = _mm_sub_ps(iy0,jy2);
592 dz02 = _mm_sub_ps(iz0,jz2);
593 dx10 = _mm_sub_ps(ix1,jx0);
594 dy10 = _mm_sub_ps(iy1,jy0);
595 dz10 = _mm_sub_ps(iz1,jz0);
596 dx11 = _mm_sub_ps(ix1,jx1);
597 dy11 = _mm_sub_ps(iy1,jy1);
598 dz11 = _mm_sub_ps(iz1,jz1);
599 dx12 = _mm_sub_ps(ix1,jx2);
600 dy12 = _mm_sub_ps(iy1,jy2);
601 dz12 = _mm_sub_ps(iz1,jz2);
602 dx20 = _mm_sub_ps(ix2,jx0);
603 dy20 = _mm_sub_ps(iy2,jy0);
604 dz20 = _mm_sub_ps(iz2,jz0);
605 dx21 = _mm_sub_ps(ix2,jx1);
606 dy21 = _mm_sub_ps(iy2,jy1);
607 dz21 = _mm_sub_ps(iz2,jz1);
608 dx22 = _mm_sub_ps(ix2,jx2);
609 dy22 = _mm_sub_ps(iy2,jy2);
610 dz22 = _mm_sub_ps(iz2,jz2);
611
612 /* Calculate squared distance and things based on it */
613 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
614 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
615 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
616 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
617 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
618 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
619 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
620 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
621 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
622
623 rinv00 = gmx_mm_invsqrt_ps(rsq00);
624 rinv01 = gmx_mm_invsqrt_ps(rsq01);
625 rinv02 = gmx_mm_invsqrt_ps(rsq02);
626 rinv10 = gmx_mm_invsqrt_ps(rsq10);
627 rinv11 = gmx_mm_invsqrt_ps(rsq11);
628 rinv12 = gmx_mm_invsqrt_ps(rsq12);
629 rinv20 = gmx_mm_invsqrt_ps(rsq20);
630 rinv21 = gmx_mm_invsqrt_ps(rsq21);
631 rinv22 = gmx_mm_invsqrt_ps(rsq22);
632
633 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
634 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
635 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
636 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
637 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
638 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
639 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
640 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
641 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
642
643 fjx0 = _mm_setzero_ps();
644 fjy0 = _mm_setzero_ps();
645 fjz0 = _mm_setzero_ps();
646 fjx1 = _mm_setzero_ps();
647 fjy1 = _mm_setzero_ps();
648 fjz1 = _mm_setzero_ps();
649 fjx2 = _mm_setzero_ps();
650 fjy2 = _mm_setzero_ps();
651 fjz2 = _mm_setzero_ps();
652
653 /**************************
654 * CALCULATE INTERACTIONS *
655 **************************/
656
657 /* COULOMB ELECTROSTATICS */
658 velec = _mm_mul_ps(qq00,rinv00);
659 felec = _mm_mul_ps(velec,rinvsq00);
660
661 /* LENNARD-JONES DISPERSION/REPULSION */
662
663 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
664 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
665 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
666 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
667 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
668
669 /* Update potential sum for this i atom from the interaction with this j atom. */
670 velec = _mm_andnot_ps(dummy_mask,velec);
671 velecsum = _mm_add_ps(velecsum,velec);
672 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
673 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
674
675 fscal = _mm_add_ps(felec,fvdw);
676
677 fscal = _mm_andnot_ps(dummy_mask,fscal);
678
679 /* Calculate temporary vectorial force */
680 tx = _mm_mul_ps(fscal,dx00);
681 ty = _mm_mul_ps(fscal,dy00);
682 tz = _mm_mul_ps(fscal,dz00);
683
684 /* Update vectorial force */
685 fix0 = _mm_add_ps(fix0,tx);
686 fiy0 = _mm_add_ps(fiy0,ty);
687 fiz0 = _mm_add_ps(fiz0,tz);
688
689 fjx0 = _mm_add_ps(fjx0,tx);
690 fjy0 = _mm_add_ps(fjy0,ty);
691 fjz0 = _mm_add_ps(fjz0,tz);
692
693 /**************************
694 * CALCULATE INTERACTIONS *
695 **************************/
696
697 /* COULOMB ELECTROSTATICS */
698 velec = _mm_mul_ps(qq01,rinv01);
699 felec = _mm_mul_ps(velec,rinvsq01);
700
701 /* Update potential sum for this i atom from the interaction with this j atom. */
702 velec = _mm_andnot_ps(dummy_mask,velec);
703 velecsum = _mm_add_ps(velecsum,velec);
704
705 fscal = felec;
706
707 fscal = _mm_andnot_ps(dummy_mask,fscal);
708
709 /* Calculate temporary vectorial force */
710 tx = _mm_mul_ps(fscal,dx01);
711 ty = _mm_mul_ps(fscal,dy01);
712 tz = _mm_mul_ps(fscal,dz01);
713
714 /* Update vectorial force */
715 fix0 = _mm_add_ps(fix0,tx);
716 fiy0 = _mm_add_ps(fiy0,ty);
717 fiz0 = _mm_add_ps(fiz0,tz);
718
719 fjx1 = _mm_add_ps(fjx1,tx);
720 fjy1 = _mm_add_ps(fjy1,ty);
721 fjz1 = _mm_add_ps(fjz1,tz);
722
723 /**************************
724 * CALCULATE INTERACTIONS *
725 **************************/
726
727 /* COULOMB ELECTROSTATICS */
728 velec = _mm_mul_ps(qq02,rinv02);
729 felec = _mm_mul_ps(velec,rinvsq02);
730
731 /* Update potential sum for this i atom from the interaction with this j atom. */
732 velec = _mm_andnot_ps(dummy_mask,velec);
733 velecsum = _mm_add_ps(velecsum,velec);
734
735 fscal = felec;
736
737 fscal = _mm_andnot_ps(dummy_mask,fscal);
738
739 /* Calculate temporary vectorial force */
740 tx = _mm_mul_ps(fscal,dx02);
741 ty = _mm_mul_ps(fscal,dy02);
742 tz = _mm_mul_ps(fscal,dz02);
743
744 /* Update vectorial force */
745 fix0 = _mm_add_ps(fix0,tx);
746 fiy0 = _mm_add_ps(fiy0,ty);
747 fiz0 = _mm_add_ps(fiz0,tz);
748
749 fjx2 = _mm_add_ps(fjx2,tx);
750 fjy2 = _mm_add_ps(fjy2,ty);
751 fjz2 = _mm_add_ps(fjz2,tz);
752
753 /**************************
754 * CALCULATE INTERACTIONS *
755 **************************/
756
757 /* COULOMB ELECTROSTATICS */
758 velec = _mm_mul_ps(qq10,rinv10);
759 felec = _mm_mul_ps(velec,rinvsq10);
760
761 /* Update potential sum for this i atom from the interaction with this j atom. */
762 velec = _mm_andnot_ps(dummy_mask,velec);
763 velecsum = _mm_add_ps(velecsum,velec);
764
765 fscal = felec;
766
767 fscal = _mm_andnot_ps(dummy_mask,fscal);
768
769 /* Calculate temporary vectorial force */
770 tx = _mm_mul_ps(fscal,dx10);
771 ty = _mm_mul_ps(fscal,dy10);
772 tz = _mm_mul_ps(fscal,dz10);
773
774 /* Update vectorial force */
775 fix1 = _mm_add_ps(fix1,tx);
776 fiy1 = _mm_add_ps(fiy1,ty);
777 fiz1 = _mm_add_ps(fiz1,tz);
778
779 fjx0 = _mm_add_ps(fjx0,tx);
780 fjy0 = _mm_add_ps(fjy0,ty);
781 fjz0 = _mm_add_ps(fjz0,tz);
782
783 /**************************
784 * CALCULATE INTERACTIONS *
785 **************************/
786
787 /* COULOMB ELECTROSTATICS */
788 velec = _mm_mul_ps(qq11,rinv11);
789 felec = _mm_mul_ps(velec,rinvsq11);
790
791 /* Update potential sum for this i atom from the interaction with this j atom. */
792 velec = _mm_andnot_ps(dummy_mask,velec);
793 velecsum = _mm_add_ps(velecsum,velec);
794
795 fscal = felec;
796
797 fscal = _mm_andnot_ps(dummy_mask,fscal);
798
799 /* Calculate temporary vectorial force */
800 tx = _mm_mul_ps(fscal,dx11);
801 ty = _mm_mul_ps(fscal,dy11);
802 tz = _mm_mul_ps(fscal,dz11);
803
804 /* Update vectorial force */
805 fix1 = _mm_add_ps(fix1,tx);
806 fiy1 = _mm_add_ps(fiy1,ty);
807 fiz1 = _mm_add_ps(fiz1,tz);
808
809 fjx1 = _mm_add_ps(fjx1,tx);
810 fjy1 = _mm_add_ps(fjy1,ty);
811 fjz1 = _mm_add_ps(fjz1,tz);
812
813 /**************************
814 * CALCULATE INTERACTIONS *
815 **************************/
816
817 /* COULOMB ELECTROSTATICS */
818 velec = _mm_mul_ps(qq12,rinv12);
819 felec = _mm_mul_ps(velec,rinvsq12);
820
821 /* Update potential sum for this i atom from the interaction with this j atom. */
822 velec = _mm_andnot_ps(dummy_mask,velec);
823 velecsum = _mm_add_ps(velecsum,velec);
824
825 fscal = felec;
826
827 fscal = _mm_andnot_ps(dummy_mask,fscal);
828
829 /* Calculate temporary vectorial force */
830 tx = _mm_mul_ps(fscal,dx12);
831 ty = _mm_mul_ps(fscal,dy12);
832 tz = _mm_mul_ps(fscal,dz12);
833
834 /* Update vectorial force */
835 fix1 = _mm_add_ps(fix1,tx);
836 fiy1 = _mm_add_ps(fiy1,ty);
837 fiz1 = _mm_add_ps(fiz1,tz);
838
839 fjx2 = _mm_add_ps(fjx2,tx);
840 fjy2 = _mm_add_ps(fjy2,ty);
841 fjz2 = _mm_add_ps(fjz2,tz);
842
843 /**************************
844 * CALCULATE INTERACTIONS *
845 **************************/
846
847 /* COULOMB ELECTROSTATICS */
848 velec = _mm_mul_ps(qq20,rinv20);
849 felec = _mm_mul_ps(velec,rinvsq20);
850
851 /* Update potential sum for this i atom from the interaction with this j atom. */
852 velec = _mm_andnot_ps(dummy_mask,velec);
853 velecsum = _mm_add_ps(velecsum,velec);
854
855 fscal = felec;
856
857 fscal = _mm_andnot_ps(dummy_mask,fscal);
858
859 /* Calculate temporary vectorial force */
860 tx = _mm_mul_ps(fscal,dx20);
861 ty = _mm_mul_ps(fscal,dy20);
862 tz = _mm_mul_ps(fscal,dz20);
863
864 /* Update vectorial force */
865 fix2 = _mm_add_ps(fix2,tx);
866 fiy2 = _mm_add_ps(fiy2,ty);
867 fiz2 = _mm_add_ps(fiz2,tz);
868
869 fjx0 = _mm_add_ps(fjx0,tx);
870 fjy0 = _mm_add_ps(fjy0,ty);
871 fjz0 = _mm_add_ps(fjz0,tz);
872
873 /**************************
874 * CALCULATE INTERACTIONS *
875 **************************/
876
877 /* COULOMB ELECTROSTATICS */
878 velec = _mm_mul_ps(qq21,rinv21);
879 felec = _mm_mul_ps(velec,rinvsq21);
880
881 /* Update potential sum for this i atom from the interaction with this j atom. */
882 velec = _mm_andnot_ps(dummy_mask,velec);
883 velecsum = _mm_add_ps(velecsum,velec);
884
885 fscal = felec;
886
887 fscal = _mm_andnot_ps(dummy_mask,fscal);
888
889 /* Calculate temporary vectorial force */
890 tx = _mm_mul_ps(fscal,dx21);
891 ty = _mm_mul_ps(fscal,dy21);
892 tz = _mm_mul_ps(fscal,dz21);
893
894 /* Update vectorial force */
895 fix2 = _mm_add_ps(fix2,tx);
896 fiy2 = _mm_add_ps(fiy2,ty);
897 fiz2 = _mm_add_ps(fiz2,tz);
898
899 fjx1 = _mm_add_ps(fjx1,tx);
900 fjy1 = _mm_add_ps(fjy1,ty);
901 fjz1 = _mm_add_ps(fjz1,tz);
902
903 /**************************
904 * CALCULATE INTERACTIONS *
905 **************************/
906
907 /* COULOMB ELECTROSTATICS */
908 velec = _mm_mul_ps(qq22,rinv22);
909 felec = _mm_mul_ps(velec,rinvsq22);
910
911 /* Update potential sum for this i atom from the interaction with this j atom. */
912 velec = _mm_andnot_ps(dummy_mask,velec);
913 velecsum = _mm_add_ps(velecsum,velec);
914
915 fscal = felec;
916
917 fscal = _mm_andnot_ps(dummy_mask,fscal);
918
919 /* Calculate temporary vectorial force */
920 tx = _mm_mul_ps(fscal,dx22);
921 ty = _mm_mul_ps(fscal,dy22);
922 tz = _mm_mul_ps(fscal,dz22);
923
924 /* Update vectorial force */
925 fix2 = _mm_add_ps(fix2,tx);
926 fiy2 = _mm_add_ps(fiy2,ty);
927 fiz2 = _mm_add_ps(fiz2,tz);
928
929 fjx2 = _mm_add_ps(fjx2,tx);
930 fjy2 = _mm_add_ps(fjy2,ty);
931 fjz2 = _mm_add_ps(fjz2,tz);
932
933 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
934 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
935 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
936 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
937
938 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
939 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
940
941 /* Inner loop uses 264 flops */
942 }
943
944 /* End of innermost loop */
945
946 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
947 f+i_coord_offset,fshift+i_shift_offset);
948
949 ggid = gid[iidx];
950 /* Update potential energies */
951 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
952 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
953
954 /* Increment number of inner iterations */
955 inneriter += j_index_end - j_index_start;
956
957 /* Outer loop uses 20 flops */
958 }
959
960 /* Increment number of outer iterations */
961 outeriter += nri;
962
963 /* Update outer/inner flops */
964
965 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*264);
966 }
967 /*
968 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
969 * Electrostatics interaction: Coulomb
970 * VdW interaction: LennardJones
971 * Geometry: Water3-Water3
972 * Calculate force/pot: Force
973 */
974 void
975 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
976 (t_nblist * gmx_restrict nlist,
977 rvec * gmx_restrict xx,
978 rvec * gmx_restrict ff,
979 t_forcerec * gmx_restrict fr,
980 t_mdatoms * gmx_restrict mdatoms,
981 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
982 t_nrnb * gmx_restrict nrnb)
983 {
984 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
985 * just 0 for non-waters.
986 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
987 * jnr indices corresponding to data put in the four positions in the SIMD register.
988 */
989 int i_shift_offset,i_coord_offset,outeriter,inneriter;
990 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
991 int jnrA,jnrB,jnrC,jnrD;
992 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
993 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
994 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
995 real rcutoff_scalar;
996 real *shiftvec,*fshift,*x,*f;
997 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
998 real scratch[4*DIM];
999 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1000 int vdwioffset0;
1001 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1002 int vdwioffset1;
1003 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1004 int vdwioffset2;
1005 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1006 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1007 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1008 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1009 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1010 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1011 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1012 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1013 __m128 dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1014 __m128 dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1015 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1016 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1017 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1018 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1019 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1020 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1021 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
1022 real *charge;
1023 int nvdwtype;
1024 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1025 int *vdwtype;
1026 real *vdwparam;
1027 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
1028 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
1029 __m128 dummy_mask,cutoff_mask;
1030 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1031 __m128 one = _mm_set1_ps(1.0);
1032 __m128 two = _mm_set1_ps(2.0);
1033 x = xx[0];
1034 f = ff[0];
1035
1036 nri = nlist->nri;
1037 iinr = nlist->iinr;
1038 jindex = nlist->jindex;
1039 jjnr = nlist->jjnr;
1040 shiftidx = nlist->shift;
1041 gid = nlist->gid;
1042 shiftvec = fr->shift_vec[0];
1043 fshift = fr->fshift[0];
1044 facel = _mm_set1_ps(fr->epsfac);
1045 charge = mdatoms->chargeA;
1046 nvdwtype = fr->ntype;
1047 vdwparam = fr->nbfp;
1048 vdwtype = mdatoms->typeA;
1049
1050 /* Setup water-specific parameters */
1051 inr = nlist->iinr[0];
1052 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1053 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1054 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1055 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1056
1057 jq0 = _mm_set1_ps(charge[inr+0]);
1058 jq1 = _mm_set1_ps(charge[inr+1]);
1059 jq2 = _mm_set1_ps(charge[inr+2]);
1060 vdwjidx0A = 2*vdwtype[inr+0];
1061 qq00 = _mm_mul_ps(iq0,jq0);
1062 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1063 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1064 qq01 = _mm_mul_ps(iq0,jq1);
1065 qq02 = _mm_mul_ps(iq0,jq2);
1066 qq10 = _mm_mul_ps(iq1,jq0);
1067 qq11 = _mm_mul_ps(iq1,jq1);
1068 qq12 = _mm_mul_ps(iq1,jq2);
1069 qq20 = _mm_mul_ps(iq2,jq0);
1070 qq21 = _mm_mul_ps(iq2,jq1);
1071 qq22 = _mm_mul_ps(iq2,jq2);
1072
1073 /* Avoid stupid compiler warnings */
1074 jnrA = jnrB = jnrC = jnrD = 0;
1075 j_coord_offsetA = 0;
1076 j_coord_offsetB = 0;
1077 j_coord_offsetC = 0;
1078 j_coord_offsetD = 0;
1079
1080 outeriter = 0;
1081 inneriter = 0;
1082
1083 for(iidx=0;iidx<4*DIM;iidx++)
1084 {
1085 scratch[iidx] = 0.0;
1086 }
1087
1088 /* Start outer loop over neighborlists */
1089 for(iidx=0; iidx<nri; iidx++)
1090 {
1091 /* Load shift vector for this list */
1092 i_shift_offset = DIM*shiftidx[iidx];
1093
1094 /* Load limits for loop over neighbors */
1095 j_index_start = jindex[iidx];
1096 j_index_end = jindex[iidx+1];
1097
1098 /* Get outer coordinate index */
1099 inr = iinr[iidx];
1100 i_coord_offset = DIM*inr;
1101
1102 /* Load i particle coords and add shift vector */
1103 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1104 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1105
1106 fix0 = _mm_setzero_ps();
1107 fiy0 = _mm_setzero_ps();
1108 fiz0 = _mm_setzero_ps();
1109 fix1 = _mm_setzero_ps();
1110 fiy1 = _mm_setzero_ps();
1111 fiz1 = _mm_setzero_ps();
1112 fix2 = _mm_setzero_ps();
1113 fiy2 = _mm_setzero_ps();
1114 fiz2 = _mm_setzero_ps();
1115
1116 /* Start inner kernel loop */
1117 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1118 {
1119
1120 /* Get j neighbor index, and coordinate index */
1121 jnrA = jjnr[jidx];
1122 jnrB = jjnr[jidx+1];
1123 jnrC = jjnr[jidx+2];
1124 jnrD = jjnr[jidx+3];
1125 j_coord_offsetA = DIM*jnrA;
1126 j_coord_offsetB = DIM*jnrB;
1127 j_coord_offsetC = DIM*jnrC;
1128 j_coord_offsetD = DIM*jnrD;
1129
1130 /* load j atom coordinates */
1131 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1132 x+j_coord_offsetC,x+j_coord_offsetD,
1133 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1134
1135 /* Calculate displacement vector */
1136 dx00 = _mm_sub_ps(ix0,jx0);
1137 dy00 = _mm_sub_ps(iy0,jy0);
1138 dz00 = _mm_sub_ps(iz0,jz0);
1139 dx01 = _mm_sub_ps(ix0,jx1);
1140 dy01 = _mm_sub_ps(iy0,jy1);
1141 dz01 = _mm_sub_ps(iz0,jz1);
1142 dx02 = _mm_sub_ps(ix0,jx2);
1143 dy02 = _mm_sub_ps(iy0,jy2);
1144 dz02 = _mm_sub_ps(iz0,jz2);
1145 dx10 = _mm_sub_ps(ix1,jx0);
1146 dy10 = _mm_sub_ps(iy1,jy0);
1147 dz10 = _mm_sub_ps(iz1,jz0);
1148 dx11 = _mm_sub_ps(ix1,jx1);
1149 dy11 = _mm_sub_ps(iy1,jy1);
1150 dz11 = _mm_sub_ps(iz1,jz1);
1151 dx12 = _mm_sub_ps(ix1,jx2);
1152 dy12 = _mm_sub_ps(iy1,jy2);
1153 dz12 = _mm_sub_ps(iz1,jz2);
1154 dx20 = _mm_sub_ps(ix2,jx0);
1155 dy20 = _mm_sub_ps(iy2,jy0);
1156 dz20 = _mm_sub_ps(iz2,jz0);
1157 dx21 = _mm_sub_ps(ix2,jx1);
1158 dy21 = _mm_sub_ps(iy2,jy1);
1159 dz21 = _mm_sub_ps(iz2,jz1);
1160 dx22 = _mm_sub_ps(ix2,jx2);
1161 dy22 = _mm_sub_ps(iy2,jy2);
1162 dz22 = _mm_sub_ps(iz2,jz2);
1163
1164 /* Calculate squared distance and things based on it */
1165 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1166 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1167 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1168 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1169 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1170 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1171 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1172 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1173 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1174
1175 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1176 rinv01 = gmx_mm_invsqrt_ps(rsq01);
1177 rinv02 = gmx_mm_invsqrt_ps(rsq02);
1178 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1179 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1180 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1181 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1182 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1183 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1184
1185 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1186 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
1187 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
1188 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1189 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1190 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1191 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1192 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1193 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1194
1195 fjx0 = _mm_setzero_ps();
1196 fjy0 = _mm_setzero_ps();
1197 fjz0 = _mm_setzero_ps();
1198 fjx1 = _mm_setzero_ps();
1199 fjy1 = _mm_setzero_ps();
1200 fjz1 = _mm_setzero_ps();
1201 fjx2 = _mm_setzero_ps();
1202 fjy2 = _mm_setzero_ps();
1203 fjz2 = _mm_setzero_ps();
1204
1205 /**************************
1206 * CALCULATE INTERACTIONS *
1207 **************************/
1208
1209 /* COULOMB ELECTROSTATICS */
1210 velec = _mm_mul_ps(qq00,rinv00);
1211 felec = _mm_mul_ps(velec,rinvsq00);
1212
1213 /* LENNARD-JONES DISPERSION/REPULSION */
1214
1215 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1216 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1217
1218 fscal = _mm_add_ps(felec,fvdw);
1219
1220 /* Calculate temporary vectorial force */
1221 tx = _mm_mul_ps(fscal,dx00);
1222 ty = _mm_mul_ps(fscal,dy00);
1223 tz = _mm_mul_ps(fscal,dz00);
1224
1225 /* Update vectorial force */
1226 fix0 = _mm_add_ps(fix0,tx);
1227 fiy0 = _mm_add_ps(fiy0,ty);
1228 fiz0 = _mm_add_ps(fiz0,tz);
1229
1230 fjx0 = _mm_add_ps(fjx0,tx);
1231 fjy0 = _mm_add_ps(fjy0,ty);
1232 fjz0 = _mm_add_ps(fjz0,tz);
1233
1234 /**************************
1235 * CALCULATE INTERACTIONS *
1236 **************************/
1237
1238 /* COULOMB ELECTROSTATICS */
1239 velec = _mm_mul_ps(qq01,rinv01);
1240 felec = _mm_mul_ps(velec,rinvsq01);
1241
1242 fscal = felec;
1243
1244 /* Calculate temporary vectorial force */
1245 tx = _mm_mul_ps(fscal,dx01);
1246 ty = _mm_mul_ps(fscal,dy01);
1247 tz = _mm_mul_ps(fscal,dz01);
1248
1249 /* Update vectorial force */
1250 fix0 = _mm_add_ps(fix0,tx);
1251 fiy0 = _mm_add_ps(fiy0,ty);
1252 fiz0 = _mm_add_ps(fiz0,tz);
1253
1254 fjx1 = _mm_add_ps(fjx1,tx);
1255 fjy1 = _mm_add_ps(fjy1,ty);
1256 fjz1 = _mm_add_ps(fjz1,tz);
1257
1258 /**************************
1259 * CALCULATE INTERACTIONS *
1260 **************************/
1261
1262 /* COULOMB ELECTROSTATICS */
1263 velec = _mm_mul_ps(qq02,rinv02);
1264 felec = _mm_mul_ps(velec,rinvsq02);
1265
1266 fscal = felec;
1267
1268 /* Calculate temporary vectorial force */
1269 tx = _mm_mul_ps(fscal,dx02);
1270 ty = _mm_mul_ps(fscal,dy02);
1271 tz = _mm_mul_ps(fscal,dz02);
1272
1273 /* Update vectorial force */
1274 fix0 = _mm_add_ps(fix0,tx);
1275 fiy0 = _mm_add_ps(fiy0,ty);
1276 fiz0 = _mm_add_ps(fiz0,tz);
1277
1278 fjx2 = _mm_add_ps(fjx2,tx);
1279 fjy2 = _mm_add_ps(fjy2,ty);
1280 fjz2 = _mm_add_ps(fjz2,tz);
1281
1282 /**************************
1283 * CALCULATE INTERACTIONS *
1284 **************************/
1285
1286 /* COULOMB ELECTROSTATICS */
1287 velec = _mm_mul_ps(qq10,rinv10);
1288 felec = _mm_mul_ps(velec,rinvsq10);
1289
1290 fscal = felec;
1291
1292 /* Calculate temporary vectorial force */
1293 tx = _mm_mul_ps(fscal,dx10);
1294 ty = _mm_mul_ps(fscal,dy10);
1295 tz = _mm_mul_ps(fscal,dz10);
1296
1297 /* Update vectorial force */
1298 fix1 = _mm_add_ps(fix1,tx);
1299 fiy1 = _mm_add_ps(fiy1,ty);
1300 fiz1 = _mm_add_ps(fiz1,tz);
1301
1302 fjx0 = _mm_add_ps(fjx0,tx);
1303 fjy0 = _mm_add_ps(fjy0,ty);
1304 fjz0 = _mm_add_ps(fjz0,tz);
1305
1306 /**************************
1307 * CALCULATE INTERACTIONS *
1308 **************************/
1309
1310 /* COULOMB ELECTROSTATICS */
1311 velec = _mm_mul_ps(qq11,rinv11);
1312 felec = _mm_mul_ps(velec,rinvsq11);
1313
1314 fscal = felec;
1315
1316 /* Calculate temporary vectorial force */
1317 tx = _mm_mul_ps(fscal,dx11);
1318 ty = _mm_mul_ps(fscal,dy11);
1319 tz = _mm_mul_ps(fscal,dz11);
1320
1321 /* Update vectorial force */
1322 fix1 = _mm_add_ps(fix1,tx);
1323 fiy1 = _mm_add_ps(fiy1,ty);
1324 fiz1 = _mm_add_ps(fiz1,tz);
1325
1326 fjx1 = _mm_add_ps(fjx1,tx);
1327 fjy1 = _mm_add_ps(fjy1,ty);
1328 fjz1 = _mm_add_ps(fjz1,tz);
1329
1330 /**************************
1331 * CALCULATE INTERACTIONS *
1332 **************************/
1333
1334 /* COULOMB ELECTROSTATICS */
1335 velec = _mm_mul_ps(qq12,rinv12);
1336 felec = _mm_mul_ps(velec,rinvsq12);
1337
1338 fscal = felec;
1339
1340 /* Calculate temporary vectorial force */
1341 tx = _mm_mul_ps(fscal,dx12);
1342 ty = _mm_mul_ps(fscal,dy12);
1343 tz = _mm_mul_ps(fscal,dz12);
1344
1345 /* Update vectorial force */
1346 fix1 = _mm_add_ps(fix1,tx);
1347 fiy1 = _mm_add_ps(fiy1,ty);
1348 fiz1 = _mm_add_ps(fiz1,tz);
1349
1350 fjx2 = _mm_add_ps(fjx2,tx);
1351 fjy2 = _mm_add_ps(fjy2,ty);
1352 fjz2 = _mm_add_ps(fjz2,tz);
1353
1354 /**************************
1355 * CALCULATE INTERACTIONS *
1356 **************************/
1357
1358 /* COULOMB ELECTROSTATICS */
1359 velec = _mm_mul_ps(qq20,rinv20);
1360 felec = _mm_mul_ps(velec,rinvsq20);
1361
1362 fscal = felec;
1363
1364 /* Calculate temporary vectorial force */
1365 tx = _mm_mul_ps(fscal,dx20);
1366 ty = _mm_mul_ps(fscal,dy20);
1367 tz = _mm_mul_ps(fscal,dz20);
1368
1369 /* Update vectorial force */
1370 fix2 = _mm_add_ps(fix2,tx);
1371 fiy2 = _mm_add_ps(fiy2,ty);
1372 fiz2 = _mm_add_ps(fiz2,tz);
1373
1374 fjx0 = _mm_add_ps(fjx0,tx);
1375 fjy0 = _mm_add_ps(fjy0,ty);
1376 fjz0 = _mm_add_ps(fjz0,tz);
1377
1378 /**************************
1379 * CALCULATE INTERACTIONS *
1380 **************************/
1381
1382 /* COULOMB ELECTROSTATICS */
1383 velec = _mm_mul_ps(qq21,rinv21);
1384 felec = _mm_mul_ps(velec,rinvsq21);
1385
1386 fscal = felec;
1387
1388 /* Calculate temporary vectorial force */
1389 tx = _mm_mul_ps(fscal,dx21);
1390 ty = _mm_mul_ps(fscal,dy21);
1391 tz = _mm_mul_ps(fscal,dz21);
1392
1393 /* Update vectorial force */
1394 fix2 = _mm_add_ps(fix2,tx);
1395 fiy2 = _mm_add_ps(fiy2,ty);
1396 fiz2 = _mm_add_ps(fiz2,tz);
1397
1398 fjx1 = _mm_add_ps(fjx1,tx);
1399 fjy1 = _mm_add_ps(fjy1,ty);
1400 fjz1 = _mm_add_ps(fjz1,tz);
1401
1402 /**************************
1403 * CALCULATE INTERACTIONS *
1404 **************************/
1405
1406 /* COULOMB ELECTROSTATICS */
1407 velec = _mm_mul_ps(qq22,rinv22);
1408 felec = _mm_mul_ps(velec,rinvsq22);
1409
1410 fscal = felec;
1411
1412 /* Calculate temporary vectorial force */
1413 tx = _mm_mul_ps(fscal,dx22);
1414 ty = _mm_mul_ps(fscal,dy22);
1415 tz = _mm_mul_ps(fscal,dz22);
1416
1417 /* Update vectorial force */
1418 fix2 = _mm_add_ps(fix2,tx);
1419 fiy2 = _mm_add_ps(fiy2,ty);
1420 fiz2 = _mm_add_ps(fiz2,tz);
1421
1422 fjx2 = _mm_add_ps(fjx2,tx);
1423 fjy2 = _mm_add_ps(fjy2,ty);
1424 fjz2 = _mm_add_ps(fjz2,tz);
1425
1426 fjptrA = f+j_coord_offsetA;
1427 fjptrB = f+j_coord_offsetB;
1428 fjptrC = f+j_coord_offsetC;
1429 fjptrD = f+j_coord_offsetD;
1430
1431 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1432 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1433
1434 /* Inner loop uses 250 flops */
1435 }
1436
1437 if(jidx<j_index_end)
1438 {
1439
1440 /* Get j neighbor index, and coordinate index */
1441 jnrlistA = jjnr[jidx];
1442 jnrlistB = jjnr[jidx+1];
1443 jnrlistC = jjnr[jidx+2];
1444 jnrlistD = jjnr[jidx+3];
1445 /* Sign of each element will be negative for non-real atoms.
1446 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1447 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1448 */
1449 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1450 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1451 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1452 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1453 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1454 j_coord_offsetA = DIM*jnrA;
1455 j_coord_offsetB = DIM*jnrB;
1456 j_coord_offsetC = DIM*jnrC;
1457 j_coord_offsetD = DIM*jnrD;
1458
1459 /* load j atom coordinates */
1460 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1461 x+j_coord_offsetC,x+j_coord_offsetD,
1462 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1463
1464 /* Calculate displacement vector */
1465 dx00 = _mm_sub_ps(ix0,jx0);
1466 dy00 = _mm_sub_ps(iy0,jy0);
1467 dz00 = _mm_sub_ps(iz0,jz0);
1468 dx01 = _mm_sub_ps(ix0,jx1);
1469 dy01 = _mm_sub_ps(iy0,jy1);
1470 dz01 = _mm_sub_ps(iz0,jz1);
1471 dx02 = _mm_sub_ps(ix0,jx2);
1472 dy02 = _mm_sub_ps(iy0,jy2);
1473 dz02 = _mm_sub_ps(iz0,jz2);
1474 dx10 = _mm_sub_ps(ix1,jx0);
1475 dy10 = _mm_sub_ps(iy1,jy0);
1476 dz10 = _mm_sub_ps(iz1,jz0);
1477 dx11 = _mm_sub_ps(ix1,jx1);
1478 dy11 = _mm_sub_ps(iy1,jy1);
1479 dz11 = _mm_sub_ps(iz1,jz1);
1480 dx12 = _mm_sub_ps(ix1,jx2);
1481 dy12 = _mm_sub_ps(iy1,jy2);
1482 dz12 = _mm_sub_ps(iz1,jz2);
1483 dx20 = _mm_sub_ps(ix2,jx0);
1484 dy20 = _mm_sub_ps(iy2,jy0);
1485 dz20 = _mm_sub_ps(iz2,jz0);
1486 dx21 = _mm_sub_ps(ix2,jx1);
1487 dy21 = _mm_sub_ps(iy2,jy1);
1488 dz21 = _mm_sub_ps(iz2,jz1);
1489 dx22 = _mm_sub_ps(ix2,jx2);
1490 dy22 = _mm_sub_ps(iy2,jy2);
1491 dz22 = _mm_sub_ps(iz2,jz2);
1492
1493 /* Calculate squared distance and things based on it */
1494 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1495 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1496 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1497 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1498 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1499 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1500 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1501 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1502 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1503
1504 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1505 rinv01 = gmx_mm_invsqrt_ps(rsq01);
1506 rinv02 = gmx_mm_invsqrt_ps(rsq02);
1507 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1508 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1509 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1510 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1511 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1512 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1513
1514 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1515 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
1516 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
1517 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1518 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1519 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1520 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1521 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1522 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1523
1524 fjx0 = _mm_setzero_ps();
1525 fjy0 = _mm_setzero_ps();
1526 fjz0 = _mm_setzero_ps();
1527 fjx1 = _mm_setzero_ps();
1528 fjy1 = _mm_setzero_ps();
1529 fjz1 = _mm_setzero_ps();
1530 fjx2 = _mm_setzero_ps();
1531 fjy2 = _mm_setzero_ps();
1532 fjz2 = _mm_setzero_ps();
1533
1534 /**************************
1535 * CALCULATE INTERACTIONS *
1536 **************************/
1537
1538 /* COULOMB ELECTROSTATICS */
1539 velec = _mm_mul_ps(qq00,rinv00);
1540 felec = _mm_mul_ps(velec,rinvsq00);
1541
1542 /* LENNARD-JONES DISPERSION/REPULSION */
1543
1544 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1545 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1546
1547 fscal = _mm_add_ps(felec,fvdw);
1548
1549 fscal = _mm_andnot_ps(dummy_mask,fscal);
1550
1551 /* Calculate temporary vectorial force */
1552 tx = _mm_mul_ps(fscal,dx00);
1553 ty = _mm_mul_ps(fscal,dy00);
1554 tz = _mm_mul_ps(fscal,dz00);
1555
1556 /* Update vectorial force */
1557 fix0 = _mm_add_ps(fix0,tx);
1558 fiy0 = _mm_add_ps(fiy0,ty);
1559 fiz0 = _mm_add_ps(fiz0,tz);
1560
1561 fjx0 = _mm_add_ps(fjx0,tx);
1562 fjy0 = _mm_add_ps(fjy0,ty);
1563 fjz0 = _mm_add_ps(fjz0,tz);
1564
1565 /**************************
1566 * CALCULATE INTERACTIONS *
1567 **************************/
1568
1569 /* COULOMB ELECTROSTATICS */
1570 velec = _mm_mul_ps(qq01,rinv01);
1571 felec = _mm_mul_ps(velec,rinvsq01);
1572
1573 fscal = felec;
1574
1575 fscal = _mm_andnot_ps(dummy_mask,fscal);
1576
1577 /* Calculate temporary vectorial force */
1578 tx = _mm_mul_ps(fscal,dx01);
1579 ty = _mm_mul_ps(fscal,dy01);
1580 tz = _mm_mul_ps(fscal,dz01);
1581
1582 /* Update vectorial force */
1583 fix0 = _mm_add_ps(fix0,tx);
1584 fiy0 = _mm_add_ps(fiy0,ty);
1585 fiz0 = _mm_add_ps(fiz0,tz);
1586
1587 fjx1 = _mm_add_ps(fjx1,tx);
1588 fjy1 = _mm_add_ps(fjy1,ty);
1589 fjz1 = _mm_add_ps(fjz1,tz);
1590
1591 /**************************
1592 * CALCULATE INTERACTIONS *
1593 **************************/
1594
1595 /* COULOMB ELECTROSTATICS */
1596 velec = _mm_mul_ps(qq02,rinv02);
1597 felec = _mm_mul_ps(velec,rinvsq02);
1598
1599 fscal = felec;
1600
1601 fscal = _mm_andnot_ps(dummy_mask,fscal);
1602
1603 /* Calculate temporary vectorial force */
1604 tx = _mm_mul_ps(fscal,dx02);
1605 ty = _mm_mul_ps(fscal,dy02);
1606 tz = _mm_mul_ps(fscal,dz02);
1607
1608 /* Update vectorial force */
1609 fix0 = _mm_add_ps(fix0,tx);
1610 fiy0 = _mm_add_ps(fiy0,ty);
1611 fiz0 = _mm_add_ps(fiz0,tz);
1612
1613 fjx2 = _mm_add_ps(fjx2,tx);
1614 fjy2 = _mm_add_ps(fjy2,ty);
1615 fjz2 = _mm_add_ps(fjz2,tz);
1616
1617 /**************************
1618 * CALCULATE INTERACTIONS *
1619 **************************/
1620
1621 /* COULOMB ELECTROSTATICS */
1622 velec = _mm_mul_ps(qq10,rinv10);
1623 felec = _mm_mul_ps(velec,rinvsq10);
1624
1625 fscal = felec;
1626
1627 fscal = _mm_andnot_ps(dummy_mask,fscal);
1628
1629 /* Calculate temporary vectorial force */
1630 tx = _mm_mul_ps(fscal,dx10);
1631 ty = _mm_mul_ps(fscal,dy10);
1632 tz = _mm_mul_ps(fscal,dz10);
1633
1634 /* Update vectorial force */
1635 fix1 = _mm_add_ps(fix1,tx);
1636 fiy1 = _mm_add_ps(fiy1,ty);
1637 fiz1 = _mm_add_ps(fiz1,tz);
1638
1639 fjx0 = _mm_add_ps(fjx0,tx);
1640 fjy0 = _mm_add_ps(fjy0,ty);
1641 fjz0 = _mm_add_ps(fjz0,tz);
1642
1643 /**************************
1644 * CALCULATE INTERACTIONS *
1645 **************************/
1646
1647 /* COULOMB ELECTROSTATICS */
1648 velec = _mm_mul_ps(qq11,rinv11);
1649 felec = _mm_mul_ps(velec,rinvsq11);
1650
1651 fscal = felec;
1652
1653 fscal = _mm_andnot_ps(dummy_mask,fscal);
1654
1655 /* Calculate temporary vectorial force */
1656 tx = _mm_mul_ps(fscal,dx11);
1657 ty = _mm_mul_ps(fscal,dy11);
1658 tz = _mm_mul_ps(fscal,dz11);
1659
1660 /* Update vectorial force */
1661 fix1 = _mm_add_ps(fix1,tx);
1662 fiy1 = _mm_add_ps(fiy1,ty);
1663 fiz1 = _mm_add_ps(fiz1,tz);
1664
1665 fjx1 = _mm_add_ps(fjx1,tx);
1666 fjy1 = _mm_add_ps(fjy1,ty);
1667 fjz1 = _mm_add_ps(fjz1,tz);
1668
1669 /**************************
1670 * CALCULATE INTERACTIONS *
1671 **************************/
1672
1673 /* COULOMB ELECTROSTATICS */
1674 velec = _mm_mul_ps(qq12,rinv12);
1675 felec = _mm_mul_ps(velec,rinvsq12);
1676
1677 fscal = felec;
1678
1679 fscal = _mm_andnot_ps(dummy_mask,fscal);
1680
1681 /* Calculate temporary vectorial force */
1682 tx = _mm_mul_ps(fscal,dx12);
1683 ty = _mm_mul_ps(fscal,dy12);
1684 tz = _mm_mul_ps(fscal,dz12);
1685
1686 /* Update vectorial force */
1687 fix1 = _mm_add_ps(fix1,tx);
1688 fiy1 = _mm_add_ps(fiy1,ty);
1689 fiz1 = _mm_add_ps(fiz1,tz);
1690
1691 fjx2 = _mm_add_ps(fjx2,tx);
1692 fjy2 = _mm_add_ps(fjy2,ty);
1693 fjz2 = _mm_add_ps(fjz2,tz);
1694
1695 /**************************
1696 * CALCULATE INTERACTIONS *
1697 **************************/
1698
1699 /* COULOMB ELECTROSTATICS */
1700 velec = _mm_mul_ps(qq20,rinv20);
1701 felec = _mm_mul_ps(velec,rinvsq20);
1702
1703 fscal = felec;
1704
1705 fscal = _mm_andnot_ps(dummy_mask,fscal);
1706
1707 /* Calculate temporary vectorial force */
1708 tx = _mm_mul_ps(fscal,dx20);
1709 ty = _mm_mul_ps(fscal,dy20);
1710 tz = _mm_mul_ps(fscal,dz20);
1711
1712 /* Update vectorial force */
1713 fix2 = _mm_add_ps(fix2,tx);
1714 fiy2 = _mm_add_ps(fiy2,ty);
1715 fiz2 = _mm_add_ps(fiz2,tz);
1716
1717 fjx0 = _mm_add_ps(fjx0,tx);
1718 fjy0 = _mm_add_ps(fjy0,ty);
1719 fjz0 = _mm_add_ps(fjz0,tz);
1720
1721 /**************************
1722 * CALCULATE INTERACTIONS *
1723 **************************/
1724
1725 /* COULOMB ELECTROSTATICS */
1726 velec = _mm_mul_ps(qq21,rinv21);
1727 felec = _mm_mul_ps(velec,rinvsq21);
1728
1729 fscal = felec;
1730
1731 fscal = _mm_andnot_ps(dummy_mask,fscal);
1732
1733 /* Calculate temporary vectorial force */
1734 tx = _mm_mul_ps(fscal,dx21);
1735 ty = _mm_mul_ps(fscal,dy21);
1736 tz = _mm_mul_ps(fscal,dz21);
1737
1738 /* Update vectorial force */
1739 fix2 = _mm_add_ps(fix2,tx);
1740 fiy2 = _mm_add_ps(fiy2,ty);
1741 fiz2 = _mm_add_ps(fiz2,tz);
1742
1743 fjx1 = _mm_add_ps(fjx1,tx);
1744 fjy1 = _mm_add_ps(fjy1,ty);
1745 fjz1 = _mm_add_ps(fjz1,tz);
1746
1747 /**************************
1748 * CALCULATE INTERACTIONS *
1749 **************************/
1750
1751 /* COULOMB ELECTROSTATICS */
1752 velec = _mm_mul_ps(qq22,rinv22);
1753 felec = _mm_mul_ps(velec,rinvsq22);
1754
1755 fscal = felec;
1756
1757 fscal = _mm_andnot_ps(dummy_mask,fscal);
1758
1759 /* Calculate temporary vectorial force */
1760 tx = _mm_mul_ps(fscal,dx22);
1761 ty = _mm_mul_ps(fscal,dy22);
1762 tz = _mm_mul_ps(fscal,dz22);
1763
1764 /* Update vectorial force */
1765 fix2 = _mm_add_ps(fix2,tx);
1766 fiy2 = _mm_add_ps(fiy2,ty);
1767 fiz2 = _mm_add_ps(fiz2,tz);
1768
1769 fjx2 = _mm_add_ps(fjx2,tx);
1770 fjy2 = _mm_add_ps(fjy2,ty);
1771 fjz2 = _mm_add_ps(fjz2,tz);
1772
1773 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1774 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1775 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1776 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1777
1778 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1779 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1780
1781 /* Inner loop uses 250 flops */
1782 }
1783
1784 /* End of innermost loop */
1785
1786 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1787 f+i_coord_offset,fshift+i_shift_offset);
1788
1789 /* Increment number of inner iterations */
1790 inneriter += j_index_end - j_index_start;
1791
1792 /* Outer loop uses 18 flops */
1793 }
1794
1795 /* Increment number of outer iterations */
1796 outeriter += nri;
1797
1798 /* Update outer/inner flops */
1799
1800 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1801 }
The ultimate molecular dynamics simulation package