?  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_GeomW4W4_sse2_single.c
blobe9f3cd75b508b4b416fde7cbd26724a9d8064600
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_GeomW4W4_VF_sse2_single
53 * Electrostatics interaction: Coulomb
54 * VdW interaction: LennardJones
55 * Geometry: Water4-Water4
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecCoul_VdwLJ_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 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 vdwioffset3;
91 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
92 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
93 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
94 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
95 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
96 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
97 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
98 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
99 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
100 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
101 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
102 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
103 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
104 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
105 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
106 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
107 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
108 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
109 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
110 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
111 real *charge;
112 int nvdwtype;
113 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
114 int *vdwtype;
115 real *vdwparam;
116 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
117 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
118 __m128 dummy_mask,cutoff_mask;
119 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
120 __m128 one = _mm_set1_ps(1.0);
121 __m128 two = _mm_set1_ps(2.0);
122 x = xx[0];
123 f = ff[0];
124
125 nri = nlist->nri;
126 iinr = nlist->iinr;
127 jindex = nlist->jindex;
128 jjnr = nlist->jjnr;
129 shiftidx = nlist->shift;
130 gid = nlist->gid;
131 shiftvec = fr->shift_vec[0];
132 fshift = fr->fshift[0];
133 facel = _mm_set1_ps(fr->epsfac);
134 charge = mdatoms->chargeA;
135 nvdwtype = fr->ntype;
136 vdwparam = fr->nbfp;
137 vdwtype = mdatoms->typeA;
138
139 /* Setup water-specific parameters */
140 inr = nlist->iinr[0];
141 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
142 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
143 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
144 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
145
146 jq1 = _mm_set1_ps(charge[inr+1]);
147 jq2 = _mm_set1_ps(charge[inr+2]);
148 jq3 = _mm_set1_ps(charge[inr+3]);
149 vdwjidx0A = 2*vdwtype[inr+0];
150 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
151 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
152 qq11 = _mm_mul_ps(iq1,jq1);
153 qq12 = _mm_mul_ps(iq1,jq2);
154 qq13 = _mm_mul_ps(iq1,jq3);
155 qq21 = _mm_mul_ps(iq2,jq1);
156 qq22 = _mm_mul_ps(iq2,jq2);
157 qq23 = _mm_mul_ps(iq2,jq3);
158 qq31 = _mm_mul_ps(iq3,jq1);
159 qq32 = _mm_mul_ps(iq3,jq2);
160 qq33 = _mm_mul_ps(iq3,jq3);
161
162 /* Avoid stupid compiler warnings */
163 jnrA = jnrB = jnrC = jnrD = 0;
164 j_coord_offsetA = 0;
165 j_coord_offsetB = 0;
166 j_coord_offsetC = 0;
167 j_coord_offsetD = 0;
168
169 outeriter = 0;
170 inneriter = 0;
171
172 for(iidx=0;iidx<4*DIM;iidx++)
173 {
174 scratch[iidx] = 0.0;
175 }
176
177 /* Start outer loop over neighborlists */
178 for(iidx=0; iidx<nri; iidx++)
179 {
180 /* Load shift vector for this list */
181 i_shift_offset = DIM*shiftidx[iidx];
182
183 /* Load limits for loop over neighbors */
184 j_index_start = jindex[iidx];
185 j_index_end = jindex[iidx+1];
186
187 /* Get outer coordinate index */
188 inr = iinr[iidx];
189 i_coord_offset = DIM*inr;
190
191 /* Load i particle coords and add shift vector */
192 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
193 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
194
195 fix0 = _mm_setzero_ps();
196 fiy0 = _mm_setzero_ps();
197 fiz0 = _mm_setzero_ps();
198 fix1 = _mm_setzero_ps();
199 fiy1 = _mm_setzero_ps();
200 fiz1 = _mm_setzero_ps();
201 fix2 = _mm_setzero_ps();
202 fiy2 = _mm_setzero_ps();
203 fiz2 = _mm_setzero_ps();
204 fix3 = _mm_setzero_ps();
205 fiy3 = _mm_setzero_ps();
206 fiz3 = _mm_setzero_ps();
207
208 /* Reset potential sums */
209 velecsum = _mm_setzero_ps();
210 vvdwsum = _mm_setzero_ps();
211
212 /* Start inner kernel loop */
213 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
214 {
215
216 /* Get j neighbor index, and coordinate index */
217 jnrA = jjnr[jidx];
218 jnrB = jjnr[jidx+1];
219 jnrC = jjnr[jidx+2];
220 jnrD = jjnr[jidx+3];
221 j_coord_offsetA = DIM*jnrA;
222 j_coord_offsetB = DIM*jnrB;
223 j_coord_offsetC = DIM*jnrC;
224 j_coord_offsetD = DIM*jnrD;
225
226 /* load j atom coordinates */
227 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
228 x+j_coord_offsetC,x+j_coord_offsetD,
229 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
230 &jy2,&jz2,&jx3,&jy3,&jz3);
231
232 /* Calculate displacement vector */
233 dx00 = _mm_sub_ps(ix0,jx0);
234 dy00 = _mm_sub_ps(iy0,jy0);
235 dz00 = _mm_sub_ps(iz0,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 dx13 = _mm_sub_ps(ix1,jx3);
243 dy13 = _mm_sub_ps(iy1,jy3);
244 dz13 = _mm_sub_ps(iz1,jz3);
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 dx23 = _mm_sub_ps(ix2,jx3);
252 dy23 = _mm_sub_ps(iy2,jy3);
253 dz23 = _mm_sub_ps(iz2,jz3);
254 dx31 = _mm_sub_ps(ix3,jx1);
255 dy31 = _mm_sub_ps(iy3,jy1);
256 dz31 = _mm_sub_ps(iz3,jz1);
257 dx32 = _mm_sub_ps(ix3,jx2);
258 dy32 = _mm_sub_ps(iy3,jy2);
259 dz32 = _mm_sub_ps(iz3,jz2);
260 dx33 = _mm_sub_ps(ix3,jx3);
261 dy33 = _mm_sub_ps(iy3,jy3);
262 dz33 = _mm_sub_ps(iz3,jz3);
263
264 /* Calculate squared distance and things based on it */
265 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
266 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
267 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
268 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
269 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
270 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
271 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
272 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
273 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
274 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
275
276 rinv11 = gmx_mm_invsqrt_ps(rsq11);
277 rinv12 = gmx_mm_invsqrt_ps(rsq12);
278 rinv13 = gmx_mm_invsqrt_ps(rsq13);
279 rinv21 = gmx_mm_invsqrt_ps(rsq21);
280 rinv22 = gmx_mm_invsqrt_ps(rsq22);
281 rinv23 = gmx_mm_invsqrt_ps(rsq23);
282 rinv31 = gmx_mm_invsqrt_ps(rsq31);
283 rinv32 = gmx_mm_invsqrt_ps(rsq32);
284 rinv33 = gmx_mm_invsqrt_ps(rsq33);
285
286 rinvsq00 = gmx_mm_inv_ps(rsq00);
287 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
288 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
289 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
290 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
291 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
292 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
293 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
294 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
295 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
296
297 fjx0 = _mm_setzero_ps();
298 fjy0 = _mm_setzero_ps();
299 fjz0 = _mm_setzero_ps();
300 fjx1 = _mm_setzero_ps();
301 fjy1 = _mm_setzero_ps();
302 fjz1 = _mm_setzero_ps();
303 fjx2 = _mm_setzero_ps();
304 fjy2 = _mm_setzero_ps();
305 fjz2 = _mm_setzero_ps();
306 fjx3 = _mm_setzero_ps();
307 fjy3 = _mm_setzero_ps();
308 fjz3 = _mm_setzero_ps();
309
310 /**************************
311 * CALCULATE INTERACTIONS *
312 **************************/
313
314 /* LENNARD-JONES DISPERSION/REPULSION */
315
316 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
317 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
318 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
319 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
320 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
321
322 /* Update potential sum for this i atom from the interaction with this j atom. */
323 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
324
325 fscal = fvdw;
326
327 /* Calculate temporary vectorial force */
328 tx = _mm_mul_ps(fscal,dx00);
329 ty = _mm_mul_ps(fscal,dy00);
330 tz = _mm_mul_ps(fscal,dz00);
331
332 /* Update vectorial force */
333 fix0 = _mm_add_ps(fix0,tx);
334 fiy0 = _mm_add_ps(fiy0,ty);
335 fiz0 = _mm_add_ps(fiz0,tz);
336
337 fjx0 = _mm_add_ps(fjx0,tx);
338 fjy0 = _mm_add_ps(fjy0,ty);
339 fjz0 = _mm_add_ps(fjz0,tz);
340
341 /**************************
342 * CALCULATE INTERACTIONS *
343 **************************/
344
345 /* COULOMB ELECTROSTATICS */
346 velec = _mm_mul_ps(qq11,rinv11);
347 felec = _mm_mul_ps(velec,rinvsq11);
348
349 /* Update potential sum for this i atom from the interaction with this j atom. */
350 velecsum = _mm_add_ps(velecsum,velec);
351
352 fscal = felec;
353
354 /* Calculate temporary vectorial force */
355 tx = _mm_mul_ps(fscal,dx11);
356 ty = _mm_mul_ps(fscal,dy11);
357 tz = _mm_mul_ps(fscal,dz11);
358
359 /* Update vectorial force */
360 fix1 = _mm_add_ps(fix1,tx);
361 fiy1 = _mm_add_ps(fiy1,ty);
362 fiz1 = _mm_add_ps(fiz1,tz);
363
364 fjx1 = _mm_add_ps(fjx1,tx);
365 fjy1 = _mm_add_ps(fjy1,ty);
366 fjz1 = _mm_add_ps(fjz1,tz);
367
368 /**************************
369 * CALCULATE INTERACTIONS *
370 **************************/
371
372 /* COULOMB ELECTROSTATICS */
373 velec = _mm_mul_ps(qq12,rinv12);
374 felec = _mm_mul_ps(velec,rinvsq12);
375
376 /* Update potential sum for this i atom from the interaction with this j atom. */
377 velecsum = _mm_add_ps(velecsum,velec);
378
379 fscal = felec;
380
381 /* Calculate temporary vectorial force */
382 tx = _mm_mul_ps(fscal,dx12);
383 ty = _mm_mul_ps(fscal,dy12);
384 tz = _mm_mul_ps(fscal,dz12);
385
386 /* Update vectorial force */
387 fix1 = _mm_add_ps(fix1,tx);
388 fiy1 = _mm_add_ps(fiy1,ty);
389 fiz1 = _mm_add_ps(fiz1,tz);
390
391 fjx2 = _mm_add_ps(fjx2,tx);
392 fjy2 = _mm_add_ps(fjy2,ty);
393 fjz2 = _mm_add_ps(fjz2,tz);
394
395 /**************************
396 * CALCULATE INTERACTIONS *
397 **************************/
398
399 /* COULOMB ELECTROSTATICS */
400 velec = _mm_mul_ps(qq13,rinv13);
401 felec = _mm_mul_ps(velec,rinvsq13);
402
403 /* Update potential sum for this i atom from the interaction with this j atom. */
404 velecsum = _mm_add_ps(velecsum,velec);
405
406 fscal = felec;
407
408 /* Calculate temporary vectorial force */
409 tx = _mm_mul_ps(fscal,dx13);
410 ty = _mm_mul_ps(fscal,dy13);
411 tz = _mm_mul_ps(fscal,dz13);
412
413 /* Update vectorial force */
414 fix1 = _mm_add_ps(fix1,tx);
415 fiy1 = _mm_add_ps(fiy1,ty);
416 fiz1 = _mm_add_ps(fiz1,tz);
417
418 fjx3 = _mm_add_ps(fjx3,tx);
419 fjy3 = _mm_add_ps(fjy3,ty);
420 fjz3 = _mm_add_ps(fjz3,tz);
421
422 /**************************
423 * CALCULATE INTERACTIONS *
424 **************************/
425
426 /* COULOMB ELECTROSTATICS */
427 velec = _mm_mul_ps(qq21,rinv21);
428 felec = _mm_mul_ps(velec,rinvsq21);
429
430 /* Update potential sum for this i atom from the interaction with this j atom. */
431 velecsum = _mm_add_ps(velecsum,velec);
432
433 fscal = felec;
434
435 /* Calculate temporary vectorial force */
436 tx = _mm_mul_ps(fscal,dx21);
437 ty = _mm_mul_ps(fscal,dy21);
438 tz = _mm_mul_ps(fscal,dz21);
439
440 /* Update vectorial force */
441 fix2 = _mm_add_ps(fix2,tx);
442 fiy2 = _mm_add_ps(fiy2,ty);
443 fiz2 = _mm_add_ps(fiz2,tz);
444
445 fjx1 = _mm_add_ps(fjx1,tx);
446 fjy1 = _mm_add_ps(fjy1,ty);
447 fjz1 = _mm_add_ps(fjz1,tz);
448
449 /**************************
450 * CALCULATE INTERACTIONS *
451 **************************/
452
453 /* COULOMB ELECTROSTATICS */
454 velec = _mm_mul_ps(qq22,rinv22);
455 felec = _mm_mul_ps(velec,rinvsq22);
456
457 /* Update potential sum for this i atom from the interaction with this j atom. */
458 velecsum = _mm_add_ps(velecsum,velec);
459
460 fscal = felec;
461
462 /* Calculate temporary vectorial force */
463 tx = _mm_mul_ps(fscal,dx22);
464 ty = _mm_mul_ps(fscal,dy22);
465 tz = _mm_mul_ps(fscal,dz22);
466
467 /* Update vectorial force */
468 fix2 = _mm_add_ps(fix2,tx);
469 fiy2 = _mm_add_ps(fiy2,ty);
470 fiz2 = _mm_add_ps(fiz2,tz);
471
472 fjx2 = _mm_add_ps(fjx2,tx);
473 fjy2 = _mm_add_ps(fjy2,ty);
474 fjz2 = _mm_add_ps(fjz2,tz);
475
476 /**************************
477 * CALCULATE INTERACTIONS *
478 **************************/
479
480 /* COULOMB ELECTROSTATICS */
481 velec = _mm_mul_ps(qq23,rinv23);
482 felec = _mm_mul_ps(velec,rinvsq23);
483
484 /* Update potential sum for this i atom from the interaction with this j atom. */
485 velecsum = _mm_add_ps(velecsum,velec);
486
487 fscal = felec;
488
489 /* Calculate temporary vectorial force */
490 tx = _mm_mul_ps(fscal,dx23);
491 ty = _mm_mul_ps(fscal,dy23);
492 tz = _mm_mul_ps(fscal,dz23);
493
494 /* Update vectorial force */
495 fix2 = _mm_add_ps(fix2,tx);
496 fiy2 = _mm_add_ps(fiy2,ty);
497 fiz2 = _mm_add_ps(fiz2,tz);
498
499 fjx3 = _mm_add_ps(fjx3,tx);
500 fjy3 = _mm_add_ps(fjy3,ty);
501 fjz3 = _mm_add_ps(fjz3,tz);
502
503 /**************************
504 * CALCULATE INTERACTIONS *
505 **************************/
506
507 /* COULOMB ELECTROSTATICS */
508 velec = _mm_mul_ps(qq31,rinv31);
509 felec = _mm_mul_ps(velec,rinvsq31);
510
511 /* Update potential sum for this i atom from the interaction with this j atom. */
512 velecsum = _mm_add_ps(velecsum,velec);
513
514 fscal = felec;
515
516 /* Calculate temporary vectorial force */
517 tx = _mm_mul_ps(fscal,dx31);
518 ty = _mm_mul_ps(fscal,dy31);
519 tz = _mm_mul_ps(fscal,dz31);
520
521 /* Update vectorial force */
522 fix3 = _mm_add_ps(fix3,tx);
523 fiy3 = _mm_add_ps(fiy3,ty);
524 fiz3 = _mm_add_ps(fiz3,tz);
525
526 fjx1 = _mm_add_ps(fjx1,tx);
527 fjy1 = _mm_add_ps(fjy1,ty);
528 fjz1 = _mm_add_ps(fjz1,tz);
529
530 /**************************
531 * CALCULATE INTERACTIONS *
532 **************************/
533
534 /* COULOMB ELECTROSTATICS */
535 velec = _mm_mul_ps(qq32,rinv32);
536 felec = _mm_mul_ps(velec,rinvsq32);
537
538 /* Update potential sum for this i atom from the interaction with this j atom. */
539 velecsum = _mm_add_ps(velecsum,velec);
540
541 fscal = felec;
542
543 /* Calculate temporary vectorial force */
544 tx = _mm_mul_ps(fscal,dx32);
545 ty = _mm_mul_ps(fscal,dy32);
546 tz = _mm_mul_ps(fscal,dz32);
547
548 /* Update vectorial force */
549 fix3 = _mm_add_ps(fix3,tx);
550 fiy3 = _mm_add_ps(fiy3,ty);
551 fiz3 = _mm_add_ps(fiz3,tz);
552
553 fjx2 = _mm_add_ps(fjx2,tx);
554 fjy2 = _mm_add_ps(fjy2,ty);
555 fjz2 = _mm_add_ps(fjz2,tz);
556
557 /**************************
558 * CALCULATE INTERACTIONS *
559 **************************/
560
561 /* COULOMB ELECTROSTATICS */
562 velec = _mm_mul_ps(qq33,rinv33);
563 felec = _mm_mul_ps(velec,rinvsq33);
564
565 /* Update potential sum for this i atom from the interaction with this j atom. */
566 velecsum = _mm_add_ps(velecsum,velec);
567
568 fscal = felec;
569
570 /* Calculate temporary vectorial force */
571 tx = _mm_mul_ps(fscal,dx33);
572 ty = _mm_mul_ps(fscal,dy33);
573 tz = _mm_mul_ps(fscal,dz33);
574
575 /* Update vectorial force */
576 fix3 = _mm_add_ps(fix3,tx);
577 fiy3 = _mm_add_ps(fiy3,ty);
578 fiz3 = _mm_add_ps(fiz3,tz);
579
580 fjx3 = _mm_add_ps(fjx3,tx);
581 fjy3 = _mm_add_ps(fjy3,ty);
582 fjz3 = _mm_add_ps(fjz3,tz);
583
584 fjptrA = f+j_coord_offsetA;
585 fjptrB = f+j_coord_offsetB;
586 fjptrC = f+j_coord_offsetC;
587 fjptrD = f+j_coord_offsetD;
588
589 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
590 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
591 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
592
593 /* Inner loop uses 287 flops */
594 }
595
596 if(jidx<j_index_end)
597 {
598
599 /* Get j neighbor index, and coordinate index */
600 jnrlistA = jjnr[jidx];
601 jnrlistB = jjnr[jidx+1];
602 jnrlistC = jjnr[jidx+2];
603 jnrlistD = jjnr[jidx+3];
604 /* Sign of each element will be negative for non-real atoms.
605 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
606 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
607 */
608 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
609 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
610 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
611 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
612 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
613 j_coord_offsetA = DIM*jnrA;
614 j_coord_offsetB = DIM*jnrB;
615 j_coord_offsetC = DIM*jnrC;
616 j_coord_offsetD = DIM*jnrD;
617
618 /* load j atom coordinates */
619 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
620 x+j_coord_offsetC,x+j_coord_offsetD,
621 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
622 &jy2,&jz2,&jx3,&jy3,&jz3);
623
624 /* Calculate displacement vector */
625 dx00 = _mm_sub_ps(ix0,jx0);
626 dy00 = _mm_sub_ps(iy0,jy0);
627 dz00 = _mm_sub_ps(iz0,jz0);
628 dx11 = _mm_sub_ps(ix1,jx1);
629 dy11 = _mm_sub_ps(iy1,jy1);
630 dz11 = _mm_sub_ps(iz1,jz1);
631 dx12 = _mm_sub_ps(ix1,jx2);
632 dy12 = _mm_sub_ps(iy1,jy2);
633 dz12 = _mm_sub_ps(iz1,jz2);
634 dx13 = _mm_sub_ps(ix1,jx3);
635 dy13 = _mm_sub_ps(iy1,jy3);
636 dz13 = _mm_sub_ps(iz1,jz3);
637 dx21 = _mm_sub_ps(ix2,jx1);
638 dy21 = _mm_sub_ps(iy2,jy1);
639 dz21 = _mm_sub_ps(iz2,jz1);
640 dx22 = _mm_sub_ps(ix2,jx2);
641 dy22 = _mm_sub_ps(iy2,jy2);
642 dz22 = _mm_sub_ps(iz2,jz2);
643 dx23 = _mm_sub_ps(ix2,jx3);
644 dy23 = _mm_sub_ps(iy2,jy3);
645 dz23 = _mm_sub_ps(iz2,jz3);
646 dx31 = _mm_sub_ps(ix3,jx1);
647 dy31 = _mm_sub_ps(iy3,jy1);
648 dz31 = _mm_sub_ps(iz3,jz1);
649 dx32 = _mm_sub_ps(ix3,jx2);
650 dy32 = _mm_sub_ps(iy3,jy2);
651 dz32 = _mm_sub_ps(iz3,jz2);
652 dx33 = _mm_sub_ps(ix3,jx3);
653 dy33 = _mm_sub_ps(iy3,jy3);
654 dz33 = _mm_sub_ps(iz3,jz3);
655
656 /* Calculate squared distance and things based on it */
657 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
658 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
659 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
660 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
661 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
662 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
663 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
664 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
665 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
666 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
667
668 rinv11 = gmx_mm_invsqrt_ps(rsq11);
669 rinv12 = gmx_mm_invsqrt_ps(rsq12);
670 rinv13 = gmx_mm_invsqrt_ps(rsq13);
671 rinv21 = gmx_mm_invsqrt_ps(rsq21);
672 rinv22 = gmx_mm_invsqrt_ps(rsq22);
673 rinv23 = gmx_mm_invsqrt_ps(rsq23);
674 rinv31 = gmx_mm_invsqrt_ps(rsq31);
675 rinv32 = gmx_mm_invsqrt_ps(rsq32);
676 rinv33 = gmx_mm_invsqrt_ps(rsq33);
677
678 rinvsq00 = gmx_mm_inv_ps(rsq00);
679 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
680 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
681 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
682 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
683 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
684 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
685 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
686 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
687 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
688
689 fjx0 = _mm_setzero_ps();
690 fjy0 = _mm_setzero_ps();
691 fjz0 = _mm_setzero_ps();
692 fjx1 = _mm_setzero_ps();
693 fjy1 = _mm_setzero_ps();
694 fjz1 = _mm_setzero_ps();
695 fjx2 = _mm_setzero_ps();
696 fjy2 = _mm_setzero_ps();
697 fjz2 = _mm_setzero_ps();
698 fjx3 = _mm_setzero_ps();
699 fjy3 = _mm_setzero_ps();
700 fjz3 = _mm_setzero_ps();
701
702 /**************************
703 * CALCULATE INTERACTIONS *
704 **************************/
705
706 /* LENNARD-JONES DISPERSION/REPULSION */
707
708 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
709 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
710 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
711 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
712 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
713
714 /* Update potential sum for this i atom from the interaction with this j atom. */
715 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
716 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
717
718 fscal = fvdw;
719
720 fscal = _mm_andnot_ps(dummy_mask,fscal);
721
722 /* Calculate temporary vectorial force */
723 tx = _mm_mul_ps(fscal,dx00);
724 ty = _mm_mul_ps(fscal,dy00);
725 tz = _mm_mul_ps(fscal,dz00);
726
727 /* Update vectorial force */
728 fix0 = _mm_add_ps(fix0,tx);
729 fiy0 = _mm_add_ps(fiy0,ty);
730 fiz0 = _mm_add_ps(fiz0,tz);
731
732 fjx0 = _mm_add_ps(fjx0,tx);
733 fjy0 = _mm_add_ps(fjy0,ty);
734 fjz0 = _mm_add_ps(fjz0,tz);
735
736 /**************************
737 * CALCULATE INTERACTIONS *
738 **************************/
739
740 /* COULOMB ELECTROSTATICS */
741 velec = _mm_mul_ps(qq11,rinv11);
742 felec = _mm_mul_ps(velec,rinvsq11);
743
744 /* Update potential sum for this i atom from the interaction with this j atom. */
745 velec = _mm_andnot_ps(dummy_mask,velec);
746 velecsum = _mm_add_ps(velecsum,velec);
747
748 fscal = felec;
749
750 fscal = _mm_andnot_ps(dummy_mask,fscal);
751
752 /* Calculate temporary vectorial force */
753 tx = _mm_mul_ps(fscal,dx11);
754 ty = _mm_mul_ps(fscal,dy11);
755 tz = _mm_mul_ps(fscal,dz11);
756
757 /* Update vectorial force */
758 fix1 = _mm_add_ps(fix1,tx);
759 fiy1 = _mm_add_ps(fiy1,ty);
760 fiz1 = _mm_add_ps(fiz1,tz);
761
762 fjx1 = _mm_add_ps(fjx1,tx);
763 fjy1 = _mm_add_ps(fjy1,ty);
764 fjz1 = _mm_add_ps(fjz1,tz);
765
766 /**************************
767 * CALCULATE INTERACTIONS *
768 **************************/
769
770 /* COULOMB ELECTROSTATICS */
771 velec = _mm_mul_ps(qq12,rinv12);
772 felec = _mm_mul_ps(velec,rinvsq12);
773
774 /* Update potential sum for this i atom from the interaction with this j atom. */
775 velec = _mm_andnot_ps(dummy_mask,velec);
776 velecsum = _mm_add_ps(velecsum,velec);
777
778 fscal = felec;
779
780 fscal = _mm_andnot_ps(dummy_mask,fscal);
781
782 /* Calculate temporary vectorial force */
783 tx = _mm_mul_ps(fscal,dx12);
784 ty = _mm_mul_ps(fscal,dy12);
785 tz = _mm_mul_ps(fscal,dz12);
786
787 /* Update vectorial force */
788 fix1 = _mm_add_ps(fix1,tx);
789 fiy1 = _mm_add_ps(fiy1,ty);
790 fiz1 = _mm_add_ps(fiz1,tz);
791
792 fjx2 = _mm_add_ps(fjx2,tx);
793 fjy2 = _mm_add_ps(fjy2,ty);
794 fjz2 = _mm_add_ps(fjz2,tz);
795
796 /**************************
797 * CALCULATE INTERACTIONS *
798 **************************/
799
800 /* COULOMB ELECTROSTATICS */
801 velec = _mm_mul_ps(qq13,rinv13);
802 felec = _mm_mul_ps(velec,rinvsq13);
803
804 /* Update potential sum for this i atom from the interaction with this j atom. */
805 velec = _mm_andnot_ps(dummy_mask,velec);
806 velecsum = _mm_add_ps(velecsum,velec);
807
808 fscal = felec;
809
810 fscal = _mm_andnot_ps(dummy_mask,fscal);
811
812 /* Calculate temporary vectorial force */
813 tx = _mm_mul_ps(fscal,dx13);
814 ty = _mm_mul_ps(fscal,dy13);
815 tz = _mm_mul_ps(fscal,dz13);
816
817 /* Update vectorial force */
818 fix1 = _mm_add_ps(fix1,tx);
819 fiy1 = _mm_add_ps(fiy1,ty);
820 fiz1 = _mm_add_ps(fiz1,tz);
821
822 fjx3 = _mm_add_ps(fjx3,tx);
823 fjy3 = _mm_add_ps(fjy3,ty);
824 fjz3 = _mm_add_ps(fjz3,tz);
825
826 /**************************
827 * CALCULATE INTERACTIONS *
828 **************************/
829
830 /* COULOMB ELECTROSTATICS */
831 velec = _mm_mul_ps(qq21,rinv21);
832 felec = _mm_mul_ps(velec,rinvsq21);
833
834 /* Update potential sum for this i atom from the interaction with this j atom. */
835 velec = _mm_andnot_ps(dummy_mask,velec);
836 velecsum = _mm_add_ps(velecsum,velec);
837
838 fscal = felec;
839
840 fscal = _mm_andnot_ps(dummy_mask,fscal);
841
842 /* Calculate temporary vectorial force */
843 tx = _mm_mul_ps(fscal,dx21);
844 ty = _mm_mul_ps(fscal,dy21);
845 tz = _mm_mul_ps(fscal,dz21);
846
847 /* Update vectorial force */
848 fix2 = _mm_add_ps(fix2,tx);
849 fiy2 = _mm_add_ps(fiy2,ty);
850 fiz2 = _mm_add_ps(fiz2,tz);
851
852 fjx1 = _mm_add_ps(fjx1,tx);
853 fjy1 = _mm_add_ps(fjy1,ty);
854 fjz1 = _mm_add_ps(fjz1,tz);
855
856 /**************************
857 * CALCULATE INTERACTIONS *
858 **************************/
859
860 /* COULOMB ELECTROSTATICS */
861 velec = _mm_mul_ps(qq22,rinv22);
862 felec = _mm_mul_ps(velec,rinvsq22);
863
864 /* Update potential sum for this i atom from the interaction with this j atom. */
865 velec = _mm_andnot_ps(dummy_mask,velec);
866 velecsum = _mm_add_ps(velecsum,velec);
867
868 fscal = felec;
869
870 fscal = _mm_andnot_ps(dummy_mask,fscal);
871
872 /* Calculate temporary vectorial force */
873 tx = _mm_mul_ps(fscal,dx22);
874 ty = _mm_mul_ps(fscal,dy22);
875 tz = _mm_mul_ps(fscal,dz22);
876
877 /* Update vectorial force */
878 fix2 = _mm_add_ps(fix2,tx);
879 fiy2 = _mm_add_ps(fiy2,ty);
880 fiz2 = _mm_add_ps(fiz2,tz);
881
882 fjx2 = _mm_add_ps(fjx2,tx);
883 fjy2 = _mm_add_ps(fjy2,ty);
884 fjz2 = _mm_add_ps(fjz2,tz);
885
886 /**************************
887 * CALCULATE INTERACTIONS *
888 **************************/
889
890 /* COULOMB ELECTROSTATICS */
891 velec = _mm_mul_ps(qq23,rinv23);
892 felec = _mm_mul_ps(velec,rinvsq23);
893
894 /* Update potential sum for this i atom from the interaction with this j atom. */
895 velec = _mm_andnot_ps(dummy_mask,velec);
896 velecsum = _mm_add_ps(velecsum,velec);
897
898 fscal = felec;
899
900 fscal = _mm_andnot_ps(dummy_mask,fscal);
901
902 /* Calculate temporary vectorial force */
903 tx = _mm_mul_ps(fscal,dx23);
904 ty = _mm_mul_ps(fscal,dy23);
905 tz = _mm_mul_ps(fscal,dz23);
906
907 /* Update vectorial force */
908 fix2 = _mm_add_ps(fix2,tx);
909 fiy2 = _mm_add_ps(fiy2,ty);
910 fiz2 = _mm_add_ps(fiz2,tz);
911
912 fjx3 = _mm_add_ps(fjx3,tx);
913 fjy3 = _mm_add_ps(fjy3,ty);
914 fjz3 = _mm_add_ps(fjz3,tz);
915
916 /**************************
917 * CALCULATE INTERACTIONS *
918 **************************/
919
920 /* COULOMB ELECTROSTATICS */
921 velec = _mm_mul_ps(qq31,rinv31);
922 felec = _mm_mul_ps(velec,rinvsq31);
923
924 /* Update potential sum for this i atom from the interaction with this j atom. */
925 velec = _mm_andnot_ps(dummy_mask,velec);
926 velecsum = _mm_add_ps(velecsum,velec);
927
928 fscal = felec;
929
930 fscal = _mm_andnot_ps(dummy_mask,fscal);
931
932 /* Calculate temporary vectorial force */
933 tx = _mm_mul_ps(fscal,dx31);
934 ty = _mm_mul_ps(fscal,dy31);
935 tz = _mm_mul_ps(fscal,dz31);
936
937 /* Update vectorial force */
938 fix3 = _mm_add_ps(fix3,tx);
939 fiy3 = _mm_add_ps(fiy3,ty);
940 fiz3 = _mm_add_ps(fiz3,tz);
941
942 fjx1 = _mm_add_ps(fjx1,tx);
943 fjy1 = _mm_add_ps(fjy1,ty);
944 fjz1 = _mm_add_ps(fjz1,tz);
945
946 /**************************
947 * CALCULATE INTERACTIONS *
948 **************************/
949
950 /* COULOMB ELECTROSTATICS */
951 velec = _mm_mul_ps(qq32,rinv32);
952 felec = _mm_mul_ps(velec,rinvsq32);
953
954 /* Update potential sum for this i atom from the interaction with this j atom. */
955 velec = _mm_andnot_ps(dummy_mask,velec);
956 velecsum = _mm_add_ps(velecsum,velec);
957
958 fscal = felec;
959
960 fscal = _mm_andnot_ps(dummy_mask,fscal);
961
962 /* Calculate temporary vectorial force */
963 tx = _mm_mul_ps(fscal,dx32);
964 ty = _mm_mul_ps(fscal,dy32);
965 tz = _mm_mul_ps(fscal,dz32);
966
967 /* Update vectorial force */
968 fix3 = _mm_add_ps(fix3,tx);
969 fiy3 = _mm_add_ps(fiy3,ty);
970 fiz3 = _mm_add_ps(fiz3,tz);
971
972 fjx2 = _mm_add_ps(fjx2,tx);
973 fjy2 = _mm_add_ps(fjy2,ty);
974 fjz2 = _mm_add_ps(fjz2,tz);
975
976 /**************************
977 * CALCULATE INTERACTIONS *
978 **************************/
979
980 /* COULOMB ELECTROSTATICS */
981 velec = _mm_mul_ps(qq33,rinv33);
982 felec = _mm_mul_ps(velec,rinvsq33);
983
984 /* Update potential sum for this i atom from the interaction with this j atom. */
985 velec = _mm_andnot_ps(dummy_mask,velec);
986 velecsum = _mm_add_ps(velecsum,velec);
987
988 fscal = felec;
989
990 fscal = _mm_andnot_ps(dummy_mask,fscal);
991
992 /* Calculate temporary vectorial force */
993 tx = _mm_mul_ps(fscal,dx33);
994 ty = _mm_mul_ps(fscal,dy33);
995 tz = _mm_mul_ps(fscal,dz33);
996
997 /* Update vectorial force */
998 fix3 = _mm_add_ps(fix3,tx);
999 fiy3 = _mm_add_ps(fiy3,ty);
1000 fiz3 = _mm_add_ps(fiz3,tz);
1001
1002 fjx3 = _mm_add_ps(fjx3,tx);
1003 fjy3 = _mm_add_ps(fjy3,ty);
1004 fjz3 = _mm_add_ps(fjz3,tz);
1005
1006 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1007 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1008 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1009 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1010
1011 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1012 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1013 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1014
1015 /* Inner loop uses 287 flops */
1016 }
1017
1018 /* End of innermost loop */
1019
1020 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1021 f+i_coord_offset,fshift+i_shift_offset);
1022
1023 ggid = gid[iidx];
1024 /* Update potential energies */
1025 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1026 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1027
1028 /* Increment number of inner iterations */
1029 inneriter += j_index_end - j_index_start;
1030
1031 /* Outer loop uses 26 flops */
1032 }
1033
1034 /* Increment number of outer iterations */
1035 outeriter += nri;
1036
1037 /* Update outer/inner flops */
1038
1039 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*287);
1040 }
1041 /*
1042 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse2_single
1043 * Electrostatics interaction: Coulomb
1044 * VdW interaction: LennardJones
1045 * Geometry: Water4-Water4
1046 * Calculate force/pot: Force
1047 */
1048 void
1049 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse2_single
1050 (t_nblist * gmx_restrict nlist,
1051 rvec * gmx_restrict xx,
1052 rvec * gmx_restrict ff,
1053 t_forcerec * gmx_restrict fr,
1054 t_mdatoms * gmx_restrict mdatoms,
1055 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1056 t_nrnb * gmx_restrict nrnb)
1057 {
1058 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1059 * just 0 for non-waters.
1060 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1061 * jnr indices corresponding to data put in the four positions in the SIMD register.
1062 */
1063 int i_shift_offset,i_coord_offset,outeriter,inneriter;
1064 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1065 int jnrA,jnrB,jnrC,jnrD;
1066 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1067 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1068 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
1069 real rcutoff_scalar;
1070 real *shiftvec,*fshift,*x,*f;
1071 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1072 real scratch[4*DIM];
1073 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1074 int vdwioffset0;
1075 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1076 int vdwioffset1;
1077 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1078 int vdwioffset2;
1079 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1080 int vdwioffset3;
1081 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1082 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1083 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1084 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1085 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1086 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1087 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1088 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1089 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1090 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1091 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1092 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1093 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1094 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1095 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1096 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1097 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1098 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1099 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1100 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
1101 real *charge;
1102 int nvdwtype;
1103 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1104 int *vdwtype;
1105 real *vdwparam;
1106 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
1107 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
1108 __m128 dummy_mask,cutoff_mask;
1109 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1110 __m128 one = _mm_set1_ps(1.0);
1111 __m128 two = _mm_set1_ps(2.0);
1112 x = xx[0];
1113 f = ff[0];
1114
1115 nri = nlist->nri;
1116 iinr = nlist->iinr;
1117 jindex = nlist->jindex;
1118 jjnr = nlist->jjnr;
1119 shiftidx = nlist->shift;
1120 gid = nlist->gid;
1121 shiftvec = fr->shift_vec[0];
1122 fshift = fr->fshift[0];
1123 facel = _mm_set1_ps(fr->epsfac);
1124 charge = mdatoms->chargeA;
1125 nvdwtype = fr->ntype;
1126 vdwparam = fr->nbfp;
1127 vdwtype = mdatoms->typeA;
1128
1129 /* Setup water-specific parameters */
1130 inr = nlist->iinr[0];
1131 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1132 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1133 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1134 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1135
1136 jq1 = _mm_set1_ps(charge[inr+1]);
1137 jq2 = _mm_set1_ps(charge[inr+2]);
1138 jq3 = _mm_set1_ps(charge[inr+3]);
1139 vdwjidx0A = 2*vdwtype[inr+0];
1140 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1141 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1142 qq11 = _mm_mul_ps(iq1,jq1);
1143 qq12 = _mm_mul_ps(iq1,jq2);
1144 qq13 = _mm_mul_ps(iq1,jq3);
1145 qq21 = _mm_mul_ps(iq2,jq1);
1146 qq22 = _mm_mul_ps(iq2,jq2);
1147 qq23 = _mm_mul_ps(iq2,jq3);
1148 qq31 = _mm_mul_ps(iq3,jq1);
1149 qq32 = _mm_mul_ps(iq3,jq2);
1150 qq33 = _mm_mul_ps(iq3,jq3);
1151
1152 /* Avoid stupid compiler warnings */
1153 jnrA = jnrB = jnrC = jnrD = 0;
1154 j_coord_offsetA = 0;
1155 j_coord_offsetB = 0;
1156 j_coord_offsetC = 0;
1157 j_coord_offsetD = 0;
1158
1159 outeriter = 0;
1160 inneriter = 0;
1161
1162 for(iidx=0;iidx<4*DIM;iidx++)
1163 {
1164 scratch[iidx] = 0.0;
1165 }
1166
1167 /* Start outer loop over neighborlists */
1168 for(iidx=0; iidx<nri; iidx++)
1169 {
1170 /* Load shift vector for this list */
1171 i_shift_offset = DIM*shiftidx[iidx];
1172
1173 /* Load limits for loop over neighbors */
1174 j_index_start = jindex[iidx];
1175 j_index_end = jindex[iidx+1];
1176
1177 /* Get outer coordinate index */
1178 inr = iinr[iidx];
1179 i_coord_offset = DIM*inr;
1180
1181 /* Load i particle coords and add shift vector */
1182 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1183 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1184
1185 fix0 = _mm_setzero_ps();
1186 fiy0 = _mm_setzero_ps();
1187 fiz0 = _mm_setzero_ps();
1188 fix1 = _mm_setzero_ps();
1189 fiy1 = _mm_setzero_ps();
1190 fiz1 = _mm_setzero_ps();
1191 fix2 = _mm_setzero_ps();
1192 fiy2 = _mm_setzero_ps();
1193 fiz2 = _mm_setzero_ps();
1194 fix3 = _mm_setzero_ps();
1195 fiy3 = _mm_setzero_ps();
1196 fiz3 = _mm_setzero_ps();
1197
1198 /* Start inner kernel loop */
1199 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1200 {
1201
1202 /* Get j neighbor index, and coordinate index */
1203 jnrA = jjnr[jidx];
1204 jnrB = jjnr[jidx+1];
1205 jnrC = jjnr[jidx+2];
1206 jnrD = jjnr[jidx+3];
1207 j_coord_offsetA = DIM*jnrA;
1208 j_coord_offsetB = DIM*jnrB;
1209 j_coord_offsetC = DIM*jnrC;
1210 j_coord_offsetD = DIM*jnrD;
1211
1212 /* load j atom coordinates */
1213 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1214 x+j_coord_offsetC,x+j_coord_offsetD,
1215 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1216 &jy2,&jz2,&jx3,&jy3,&jz3);
1217
1218 /* Calculate displacement vector */
1219 dx00 = _mm_sub_ps(ix0,jx0);
1220 dy00 = _mm_sub_ps(iy0,jy0);
1221 dz00 = _mm_sub_ps(iz0,jz0);
1222 dx11 = _mm_sub_ps(ix1,jx1);
1223 dy11 = _mm_sub_ps(iy1,jy1);
1224 dz11 = _mm_sub_ps(iz1,jz1);
1225 dx12 = _mm_sub_ps(ix1,jx2);
1226 dy12 = _mm_sub_ps(iy1,jy2);
1227 dz12 = _mm_sub_ps(iz1,jz2);
1228 dx13 = _mm_sub_ps(ix1,jx3);
1229 dy13 = _mm_sub_ps(iy1,jy3);
1230 dz13 = _mm_sub_ps(iz1,jz3);
1231 dx21 = _mm_sub_ps(ix2,jx1);
1232 dy21 = _mm_sub_ps(iy2,jy1);
1233 dz21 = _mm_sub_ps(iz2,jz1);
1234 dx22 = _mm_sub_ps(ix2,jx2);
1235 dy22 = _mm_sub_ps(iy2,jy2);
1236 dz22 = _mm_sub_ps(iz2,jz2);
1237 dx23 = _mm_sub_ps(ix2,jx3);
1238 dy23 = _mm_sub_ps(iy2,jy3);
1239 dz23 = _mm_sub_ps(iz2,jz3);
1240 dx31 = _mm_sub_ps(ix3,jx1);
1241 dy31 = _mm_sub_ps(iy3,jy1);
1242 dz31 = _mm_sub_ps(iz3,jz1);
1243 dx32 = _mm_sub_ps(ix3,jx2);
1244 dy32 = _mm_sub_ps(iy3,jy2);
1245 dz32 = _mm_sub_ps(iz3,jz2);
1246 dx33 = _mm_sub_ps(ix3,jx3);
1247 dy33 = _mm_sub_ps(iy3,jy3);
1248 dz33 = _mm_sub_ps(iz3,jz3);
1249
1250 /* Calculate squared distance and things based on it */
1251 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1252 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1253 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1254 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1255 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1256 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1257 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1258 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1259 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1260 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1261
1262 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1263 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1264 rinv13 = gmx_mm_invsqrt_ps(rsq13);
1265 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1266 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1267 rinv23 = gmx_mm_invsqrt_ps(rsq23);
1268 rinv31 = gmx_mm_invsqrt_ps(rsq31);
1269 rinv32 = gmx_mm_invsqrt_ps(rsq32);
1270 rinv33 = gmx_mm_invsqrt_ps(rsq33);
1271
1272 rinvsq00 = gmx_mm_inv_ps(rsq00);
1273 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1274 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1275 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1276 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1277 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1278 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1279 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1280 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1281 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1282
1283 fjx0 = _mm_setzero_ps();
1284 fjy0 = _mm_setzero_ps();
1285 fjz0 = _mm_setzero_ps();
1286 fjx1 = _mm_setzero_ps();
1287 fjy1 = _mm_setzero_ps();
1288 fjz1 = _mm_setzero_ps();
1289 fjx2 = _mm_setzero_ps();
1290 fjy2 = _mm_setzero_ps();
1291 fjz2 = _mm_setzero_ps();
1292 fjx3 = _mm_setzero_ps();
1293 fjy3 = _mm_setzero_ps();
1294 fjz3 = _mm_setzero_ps();
1295
1296 /**************************
1297 * CALCULATE INTERACTIONS *
1298 **************************/
1299
1300 /* LENNARD-JONES DISPERSION/REPULSION */
1301
1302 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1303 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1304
1305 fscal = fvdw;
1306
1307 /* Calculate temporary vectorial force */
1308 tx = _mm_mul_ps(fscal,dx00);
1309 ty = _mm_mul_ps(fscal,dy00);
1310 tz = _mm_mul_ps(fscal,dz00);
1311
1312 /* Update vectorial force */
1313 fix0 = _mm_add_ps(fix0,tx);
1314 fiy0 = _mm_add_ps(fiy0,ty);
1315 fiz0 = _mm_add_ps(fiz0,tz);
1316
1317 fjx0 = _mm_add_ps(fjx0,tx);
1318 fjy0 = _mm_add_ps(fjy0,ty);
1319 fjz0 = _mm_add_ps(fjz0,tz);
1320
1321 /**************************
1322 * CALCULATE INTERACTIONS *
1323 **************************/
1324
1325 /* COULOMB ELECTROSTATICS */
1326 velec = _mm_mul_ps(qq11,rinv11);
1327 felec = _mm_mul_ps(velec,rinvsq11);
1328
1329 fscal = felec;
1330
1331 /* Calculate temporary vectorial force */
1332 tx = _mm_mul_ps(fscal,dx11);
1333 ty = _mm_mul_ps(fscal,dy11);
1334 tz = _mm_mul_ps(fscal,dz11);
1335
1336 /* Update vectorial force */
1337 fix1 = _mm_add_ps(fix1,tx);
1338 fiy1 = _mm_add_ps(fiy1,ty);
1339 fiz1 = _mm_add_ps(fiz1,tz);
1340
1341 fjx1 = _mm_add_ps(fjx1,tx);
1342 fjy1 = _mm_add_ps(fjy1,ty);
1343 fjz1 = _mm_add_ps(fjz1,tz);
1344
1345 /**************************
1346 * CALCULATE INTERACTIONS *
1347 **************************/
1348
1349 /* COULOMB ELECTROSTATICS */
1350 velec = _mm_mul_ps(qq12,rinv12);
1351 felec = _mm_mul_ps(velec,rinvsq12);
1352
1353 fscal = felec;
1354
1355 /* Calculate temporary vectorial force */
1356 tx = _mm_mul_ps(fscal,dx12);
1357 ty = _mm_mul_ps(fscal,dy12);
1358 tz = _mm_mul_ps(fscal,dz12);
1359
1360 /* Update vectorial force */
1361 fix1 = _mm_add_ps(fix1,tx);
1362 fiy1 = _mm_add_ps(fiy1,ty);
1363 fiz1 = _mm_add_ps(fiz1,tz);
1364
1365 fjx2 = _mm_add_ps(fjx2,tx);
1366 fjy2 = _mm_add_ps(fjy2,ty);
1367 fjz2 = _mm_add_ps(fjz2,tz);
1368
1369 /**************************
1370 * CALCULATE INTERACTIONS *
1371 **************************/
1372
1373 /* COULOMB ELECTROSTATICS */
1374 velec = _mm_mul_ps(qq13,rinv13);
1375 felec = _mm_mul_ps(velec,rinvsq13);
1376
1377 fscal = felec;
1378
1379 /* Calculate temporary vectorial force */
1380 tx = _mm_mul_ps(fscal,dx13);
1381 ty = _mm_mul_ps(fscal,dy13);
1382 tz = _mm_mul_ps(fscal,dz13);
1383
1384 /* Update vectorial force */
1385 fix1 = _mm_add_ps(fix1,tx);
1386 fiy1 = _mm_add_ps(fiy1,ty);
1387 fiz1 = _mm_add_ps(fiz1,tz);
1388
1389 fjx3 = _mm_add_ps(fjx3,tx);
1390 fjy3 = _mm_add_ps(fjy3,ty);
1391 fjz3 = _mm_add_ps(fjz3,tz);
1392
1393 /**************************
1394 * CALCULATE INTERACTIONS *
1395 **************************/
1396
1397 /* COULOMB ELECTROSTATICS */
1398 velec = _mm_mul_ps(qq21,rinv21);
1399 felec = _mm_mul_ps(velec,rinvsq21);
1400
1401 fscal = felec;
1402
1403 /* Calculate temporary vectorial force */
1404 tx = _mm_mul_ps(fscal,dx21);
1405 ty = _mm_mul_ps(fscal,dy21);
1406 tz = _mm_mul_ps(fscal,dz21);
1407
1408 /* Update vectorial force */
1409 fix2 = _mm_add_ps(fix2,tx);
1410 fiy2 = _mm_add_ps(fiy2,ty);
1411 fiz2 = _mm_add_ps(fiz2,tz);
1412
1413 fjx1 = _mm_add_ps(fjx1,tx);
1414 fjy1 = _mm_add_ps(fjy1,ty);
1415 fjz1 = _mm_add_ps(fjz1,tz);
1416
1417 /**************************
1418 * CALCULATE INTERACTIONS *
1419 **************************/
1420
1421 /* COULOMB ELECTROSTATICS */
1422 velec = _mm_mul_ps(qq22,rinv22);
1423 felec = _mm_mul_ps(velec,rinvsq22);
1424
1425 fscal = felec;
1426
1427 /* Calculate temporary vectorial force */
1428 tx = _mm_mul_ps(fscal,dx22);
1429 ty = _mm_mul_ps(fscal,dy22);
1430 tz = _mm_mul_ps(fscal,dz22);
1431
1432 /* Update vectorial force */
1433 fix2 = _mm_add_ps(fix2,tx);
1434 fiy2 = _mm_add_ps(fiy2,ty);
1435 fiz2 = _mm_add_ps(fiz2,tz);
1436
1437 fjx2 = _mm_add_ps(fjx2,tx);
1438 fjy2 = _mm_add_ps(fjy2,ty);
1439 fjz2 = _mm_add_ps(fjz2,tz);
1440
1441 /**************************
1442 * CALCULATE INTERACTIONS *
1443 **************************/
1444
1445 /* COULOMB ELECTROSTATICS */
1446 velec = _mm_mul_ps(qq23,rinv23);
1447 felec = _mm_mul_ps(velec,rinvsq23);
1448
1449 fscal = felec;
1450
1451 /* Calculate temporary vectorial force */
1452 tx = _mm_mul_ps(fscal,dx23);
1453 ty = _mm_mul_ps(fscal,dy23);
1454 tz = _mm_mul_ps(fscal,dz23);
1455
1456 /* Update vectorial force */
1457 fix2 = _mm_add_ps(fix2,tx);
1458 fiy2 = _mm_add_ps(fiy2,ty);
1459 fiz2 = _mm_add_ps(fiz2,tz);
1460
1461 fjx3 = _mm_add_ps(fjx3,tx);
1462 fjy3 = _mm_add_ps(fjy3,ty);
1463 fjz3 = _mm_add_ps(fjz3,tz);
1464
1465 /**************************
1466 * CALCULATE INTERACTIONS *
1467 **************************/
1468
1469 /* COULOMB ELECTROSTATICS */
1470 velec = _mm_mul_ps(qq31,rinv31);
1471 felec = _mm_mul_ps(velec,rinvsq31);
1472
1473 fscal = felec;
1474
1475 /* Calculate temporary vectorial force */
1476 tx = _mm_mul_ps(fscal,dx31);
1477 ty = _mm_mul_ps(fscal,dy31);
1478 tz = _mm_mul_ps(fscal,dz31);
1479
1480 /* Update vectorial force */
1481 fix3 = _mm_add_ps(fix3,tx);
1482 fiy3 = _mm_add_ps(fiy3,ty);
1483 fiz3 = _mm_add_ps(fiz3,tz);
1484
1485 fjx1 = _mm_add_ps(fjx1,tx);
1486 fjy1 = _mm_add_ps(fjy1,ty);
1487 fjz1 = _mm_add_ps(fjz1,tz);
1488
1489 /**************************
1490 * CALCULATE INTERACTIONS *
1491 **************************/
1492
1493 /* COULOMB ELECTROSTATICS */
1494 velec = _mm_mul_ps(qq32,rinv32);
1495 felec = _mm_mul_ps(velec,rinvsq32);
1496
1497 fscal = felec;
1498
1499 /* Calculate temporary vectorial force */
1500 tx = _mm_mul_ps(fscal,dx32);
1501 ty = _mm_mul_ps(fscal,dy32);
1502 tz = _mm_mul_ps(fscal,dz32);
1503
1504 /* Update vectorial force */
1505 fix3 = _mm_add_ps(fix3,tx);
1506 fiy3 = _mm_add_ps(fiy3,ty);
1507 fiz3 = _mm_add_ps(fiz3,tz);
1508
1509 fjx2 = _mm_add_ps(fjx2,tx);
1510 fjy2 = _mm_add_ps(fjy2,ty);
1511 fjz2 = _mm_add_ps(fjz2,tz);
1512
1513 /**************************
1514 * CALCULATE INTERACTIONS *
1515 **************************/
1516
1517 /* COULOMB ELECTROSTATICS */
1518 velec = _mm_mul_ps(qq33,rinv33);
1519 felec = _mm_mul_ps(velec,rinvsq33);
1520
1521 fscal = felec;
1522
1523 /* Calculate temporary vectorial force */
1524 tx = _mm_mul_ps(fscal,dx33);
1525 ty = _mm_mul_ps(fscal,dy33);
1526 tz = _mm_mul_ps(fscal,dz33);
1527
1528 /* Update vectorial force */
1529 fix3 = _mm_add_ps(fix3,tx);
1530 fiy3 = _mm_add_ps(fiy3,ty);
1531 fiz3 = _mm_add_ps(fiz3,tz);
1532
1533 fjx3 = _mm_add_ps(fjx3,tx);
1534 fjy3 = _mm_add_ps(fjy3,ty);
1535 fjz3 = _mm_add_ps(fjz3,tz);
1536
1537 fjptrA = f+j_coord_offsetA;
1538 fjptrB = f+j_coord_offsetB;
1539 fjptrC = f+j_coord_offsetC;
1540 fjptrD = f+j_coord_offsetD;
1541
1542 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1543 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1544 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1545
1546 /* Inner loop uses 273 flops */
1547 }
1548
1549 if(jidx<j_index_end)
1550 {
1551
1552 /* Get j neighbor index, and coordinate index */
1553 jnrlistA = jjnr[jidx];
1554 jnrlistB = jjnr[jidx+1];
1555 jnrlistC = jjnr[jidx+2];
1556 jnrlistD = jjnr[jidx+3];
1557 /* Sign of each element will be negative for non-real atoms.
1558 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1559 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1560 */
1561 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1562 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1563 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1564 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1565 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1566 j_coord_offsetA = DIM*jnrA;
1567 j_coord_offsetB = DIM*jnrB;
1568 j_coord_offsetC = DIM*jnrC;
1569 j_coord_offsetD = DIM*jnrD;
1570
1571 /* load j atom coordinates */
1572 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1573 x+j_coord_offsetC,x+j_coord_offsetD,
1574 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1575 &jy2,&jz2,&jx3,&jy3,&jz3);
1576
1577 /* Calculate displacement vector */
1578 dx00 = _mm_sub_ps(ix0,jx0);
1579 dy00 = _mm_sub_ps(iy0,jy0);
1580 dz00 = _mm_sub_ps(iz0,jz0);
1581 dx11 = _mm_sub_ps(ix1,jx1);
1582 dy11 = _mm_sub_ps(iy1,jy1);
1583 dz11 = _mm_sub_ps(iz1,jz1);
1584 dx12 = _mm_sub_ps(ix1,jx2);
1585 dy12 = _mm_sub_ps(iy1,jy2);
1586 dz12 = _mm_sub_ps(iz1,jz2);
1587 dx13 = _mm_sub_ps(ix1,jx3);
1588 dy13 = _mm_sub_ps(iy1,jy3);
1589 dz13 = _mm_sub_ps(iz1,jz3);
1590 dx21 = _mm_sub_ps(ix2,jx1);
1591 dy21 = _mm_sub_ps(iy2,jy1);
1592 dz21 = _mm_sub_ps(iz2,jz1);
1593 dx22 = _mm_sub_ps(ix2,jx2);
1594 dy22 = _mm_sub_ps(iy2,jy2);
1595 dz22 = _mm_sub_ps(iz2,jz2);
1596 dx23 = _mm_sub_ps(ix2,jx3);
1597 dy23 = _mm_sub_ps(iy2,jy3);
1598 dz23 = _mm_sub_ps(iz2,jz3);
1599 dx31 = _mm_sub_ps(ix3,jx1);
1600 dy31 = _mm_sub_ps(iy3,jy1);
1601 dz31 = _mm_sub_ps(iz3,jz1);
1602 dx32 = _mm_sub_ps(ix3,jx2);
1603 dy32 = _mm_sub_ps(iy3,jy2);
1604 dz32 = _mm_sub_ps(iz3,jz2);
1605 dx33 = _mm_sub_ps(ix3,jx3);
1606 dy33 = _mm_sub_ps(iy3,jy3);
1607 dz33 = _mm_sub_ps(iz3,jz3);
1608
1609 /* Calculate squared distance and things based on it */
1610 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1611 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1612 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1613 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1614 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1615 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1616 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1617 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1618 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1619 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1620
1621 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1622 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1623 rinv13 = gmx_mm_invsqrt_ps(rsq13);
1624 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1625 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1626 rinv23 = gmx_mm_invsqrt_ps(rsq23);
1627 rinv31 = gmx_mm_invsqrt_ps(rsq31);
1628 rinv32 = gmx_mm_invsqrt_ps(rsq32);
1629 rinv33 = gmx_mm_invsqrt_ps(rsq33);
1630
1631 rinvsq00 = gmx_mm_inv_ps(rsq00);
1632 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1633 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1634 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1635 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1636 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1637 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1638 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1639 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1640 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1641
1642 fjx0 = _mm_setzero_ps();
1643 fjy0 = _mm_setzero_ps();
1644 fjz0 = _mm_setzero_ps();
1645 fjx1 = _mm_setzero_ps();
1646 fjy1 = _mm_setzero_ps();
1647 fjz1 = _mm_setzero_ps();
1648 fjx2 = _mm_setzero_ps();
1649 fjy2 = _mm_setzero_ps();
1650 fjz2 = _mm_setzero_ps();
1651 fjx3 = _mm_setzero_ps();
1652 fjy3 = _mm_setzero_ps();
1653 fjz3 = _mm_setzero_ps();
1654
1655 /**************************
1656 * CALCULATE INTERACTIONS *
1657 **************************/
1658
1659 /* LENNARD-JONES DISPERSION/REPULSION */
1660
1661 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1662 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1663
1664 fscal = fvdw;
1665
1666 fscal = _mm_andnot_ps(dummy_mask,fscal);
1667
1668 /* Calculate temporary vectorial force */
1669 tx = _mm_mul_ps(fscal,dx00);
1670 ty = _mm_mul_ps(fscal,dy00);
1671 tz = _mm_mul_ps(fscal,dz00);
1672
1673 /* Update vectorial force */
1674 fix0 = _mm_add_ps(fix0,tx);
1675 fiy0 = _mm_add_ps(fiy0,ty);
1676 fiz0 = _mm_add_ps(fiz0,tz);
1677
1678 fjx0 = _mm_add_ps(fjx0,tx);
1679 fjy0 = _mm_add_ps(fjy0,ty);
1680 fjz0 = _mm_add_ps(fjz0,tz);
1681
1682 /**************************
1683 * CALCULATE INTERACTIONS *
1684 **************************/
1685
1686 /* COULOMB ELECTROSTATICS */
1687 velec = _mm_mul_ps(qq11,rinv11);
1688 felec = _mm_mul_ps(velec,rinvsq11);
1689
1690 fscal = felec;
1691
1692 fscal = _mm_andnot_ps(dummy_mask,fscal);
1693
1694 /* Calculate temporary vectorial force */
1695 tx = _mm_mul_ps(fscal,dx11);
1696 ty = _mm_mul_ps(fscal,dy11);
1697 tz = _mm_mul_ps(fscal,dz11);
1698
1699 /* Update vectorial force */
1700 fix1 = _mm_add_ps(fix1,tx);
1701 fiy1 = _mm_add_ps(fiy1,ty);
1702 fiz1 = _mm_add_ps(fiz1,tz);
1703
1704 fjx1 = _mm_add_ps(fjx1,tx);
1705 fjy1 = _mm_add_ps(fjy1,ty);
1706 fjz1 = _mm_add_ps(fjz1,tz);
1707
1708 /**************************
1709 * CALCULATE INTERACTIONS *
1710 **************************/
1711
1712 /* COULOMB ELECTROSTATICS */
1713 velec = _mm_mul_ps(qq12,rinv12);
1714 felec = _mm_mul_ps(velec,rinvsq12);
1715
1716 fscal = felec;
1717
1718 fscal = _mm_andnot_ps(dummy_mask,fscal);
1719
1720 /* Calculate temporary vectorial force */
1721 tx = _mm_mul_ps(fscal,dx12);
1722 ty = _mm_mul_ps(fscal,dy12);
1723 tz = _mm_mul_ps(fscal,dz12);
1724
1725 /* Update vectorial force */
1726 fix1 = _mm_add_ps(fix1,tx);
1727 fiy1 = _mm_add_ps(fiy1,ty);
1728 fiz1 = _mm_add_ps(fiz1,tz);
1729
1730 fjx2 = _mm_add_ps(fjx2,tx);
1731 fjy2 = _mm_add_ps(fjy2,ty);
1732 fjz2 = _mm_add_ps(fjz2,tz);
1733
1734 /**************************
1735 * CALCULATE INTERACTIONS *
1736 **************************/
1737
1738 /* COULOMB ELECTROSTATICS */
1739 velec = _mm_mul_ps(qq13,rinv13);
1740 felec = _mm_mul_ps(velec,rinvsq13);
1741
1742 fscal = felec;
1743
1744 fscal = _mm_andnot_ps(dummy_mask,fscal);
1745
1746 /* Calculate temporary vectorial force */
1747 tx = _mm_mul_ps(fscal,dx13);
1748 ty = _mm_mul_ps(fscal,dy13);
1749 tz = _mm_mul_ps(fscal,dz13);
1750
1751 /* Update vectorial force */
1752 fix1 = _mm_add_ps(fix1,tx);
1753 fiy1 = _mm_add_ps(fiy1,ty);
1754 fiz1 = _mm_add_ps(fiz1,tz);
1755
1756 fjx3 = _mm_add_ps(fjx3,tx);
1757 fjy3 = _mm_add_ps(fjy3,ty);
1758 fjz3 = _mm_add_ps(fjz3,tz);
1759
1760 /**************************
1761 * CALCULATE INTERACTIONS *
1762 **************************/
1763
1764 /* COULOMB ELECTROSTATICS */
1765 velec = _mm_mul_ps(qq21,rinv21);
1766 felec = _mm_mul_ps(velec,rinvsq21);
1767
1768 fscal = felec;
1769
1770 fscal = _mm_andnot_ps(dummy_mask,fscal);
1771
1772 /* Calculate temporary vectorial force */
1773 tx = _mm_mul_ps(fscal,dx21);
1774 ty = _mm_mul_ps(fscal,dy21);
1775 tz = _mm_mul_ps(fscal,dz21);
1776
1777 /* Update vectorial force */
1778 fix2 = _mm_add_ps(fix2,tx);
1779 fiy2 = _mm_add_ps(fiy2,ty);
1780 fiz2 = _mm_add_ps(fiz2,tz);
1781
1782 fjx1 = _mm_add_ps(fjx1,tx);
1783 fjy1 = _mm_add_ps(fjy1,ty);
1784 fjz1 = _mm_add_ps(fjz1,tz);
1785
1786 /**************************
1787 * CALCULATE INTERACTIONS *
1788 **************************/
1789
1790 /* COULOMB ELECTROSTATICS */
1791 velec = _mm_mul_ps(qq22,rinv22);
1792 felec = _mm_mul_ps(velec,rinvsq22);
1793
1794 fscal = felec;
1795
1796 fscal = _mm_andnot_ps(dummy_mask,fscal);
1797
1798 /* Calculate temporary vectorial force */
1799 tx = _mm_mul_ps(fscal,dx22);
1800 ty = _mm_mul_ps(fscal,dy22);
1801 tz = _mm_mul_ps(fscal,dz22);
1802
1803 /* Update vectorial force */
1804 fix2 = _mm_add_ps(fix2,tx);
1805 fiy2 = _mm_add_ps(fiy2,ty);
1806 fiz2 = _mm_add_ps(fiz2,tz);
1807
1808 fjx2 = _mm_add_ps(fjx2,tx);
1809 fjy2 = _mm_add_ps(fjy2,ty);
1810 fjz2 = _mm_add_ps(fjz2,tz);
1811
1812 /**************************
1813 * CALCULATE INTERACTIONS *
1814 **************************/
1815
1816 /* COULOMB ELECTROSTATICS */
1817 velec = _mm_mul_ps(qq23,rinv23);
1818 felec = _mm_mul_ps(velec,rinvsq23);
1819
1820 fscal = felec;
1821
1822 fscal = _mm_andnot_ps(dummy_mask,fscal);
1823
1824 /* Calculate temporary vectorial force */
1825 tx = _mm_mul_ps(fscal,dx23);
1826 ty = _mm_mul_ps(fscal,dy23);
1827 tz = _mm_mul_ps(fscal,dz23);
1828
1829 /* Update vectorial force */
1830 fix2 = _mm_add_ps(fix2,tx);
1831 fiy2 = _mm_add_ps(fiy2,ty);
1832 fiz2 = _mm_add_ps(fiz2,tz);
1833
1834 fjx3 = _mm_add_ps(fjx3,tx);
1835 fjy3 = _mm_add_ps(fjy3,ty);
1836 fjz3 = _mm_add_ps(fjz3,tz);
1837
1838 /**************************
1839 * CALCULATE INTERACTIONS *
1840 **************************/
1841
1842 /* COULOMB ELECTROSTATICS */
1843 velec = _mm_mul_ps(qq31,rinv31);
1844 felec = _mm_mul_ps(velec,rinvsq31);
1845
1846 fscal = felec;
1847
1848 fscal = _mm_andnot_ps(dummy_mask,fscal);
1849
1850 /* Calculate temporary vectorial force */
1851 tx = _mm_mul_ps(fscal,dx31);
1852 ty = _mm_mul_ps(fscal,dy31);
1853 tz = _mm_mul_ps(fscal,dz31);
1854
1855 /* Update vectorial force */
1856 fix3 = _mm_add_ps(fix3,tx);
1857 fiy3 = _mm_add_ps(fiy3,ty);
1858 fiz3 = _mm_add_ps(fiz3,tz);
1859
1860 fjx1 = _mm_add_ps(fjx1,tx);
1861 fjy1 = _mm_add_ps(fjy1,ty);
1862 fjz1 = _mm_add_ps(fjz1,tz);
1863
1864 /**************************
1865 * CALCULATE INTERACTIONS *
1866 **************************/
1867
1868 /* COULOMB ELECTROSTATICS */
1869 velec = _mm_mul_ps(qq32,rinv32);
1870 felec = _mm_mul_ps(velec,rinvsq32);
1871
1872 fscal = felec;
1873
1874 fscal = _mm_andnot_ps(dummy_mask,fscal);
1875
1876 /* Calculate temporary vectorial force */
1877 tx = _mm_mul_ps(fscal,dx32);
1878 ty = _mm_mul_ps(fscal,dy32);
1879 tz = _mm_mul_ps(fscal,dz32);
1880
1881 /* Update vectorial force */
1882 fix3 = _mm_add_ps(fix3,tx);
1883 fiy3 = _mm_add_ps(fiy3,ty);
1884 fiz3 = _mm_add_ps(fiz3,tz);
1885
1886 fjx2 = _mm_add_ps(fjx2,tx);
1887 fjy2 = _mm_add_ps(fjy2,ty);
1888 fjz2 = _mm_add_ps(fjz2,tz);
1889
1890 /**************************
1891 * CALCULATE INTERACTIONS *
1892 **************************/
1893
1894 /* COULOMB ELECTROSTATICS */
1895 velec = _mm_mul_ps(qq33,rinv33);
1896 felec = _mm_mul_ps(velec,rinvsq33);
1897
1898 fscal = felec;
1899
1900 fscal = _mm_andnot_ps(dummy_mask,fscal);
1901
1902 /* Calculate temporary vectorial force */
1903 tx = _mm_mul_ps(fscal,dx33);
1904 ty = _mm_mul_ps(fscal,dy33);
1905 tz = _mm_mul_ps(fscal,dz33);
1906
1907 /* Update vectorial force */
1908 fix3 = _mm_add_ps(fix3,tx);
1909 fiy3 = _mm_add_ps(fiy3,ty);
1910 fiz3 = _mm_add_ps(fiz3,tz);
1911
1912 fjx3 = _mm_add_ps(fjx3,tx);
1913 fjy3 = _mm_add_ps(fjy3,ty);
1914 fjz3 = _mm_add_ps(fjz3,tz);
1915
1916 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1917 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1918 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1919 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1920
1921 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1922 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1923 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1924
1925 /* Inner loop uses 273 flops */
1926 }
1927
1928 /* End of innermost loop */
1929
1930 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1931 f+i_coord_offset,fshift+i_shift_offset);
1932
1933 /* Increment number of inner iterations */
1934 inneriter += j_index_end - j_index_start;
1935
1936 /* Outer loop uses 24 flops */
1937 }
1938
1939 /* Increment number of outer iterations */
1940 outeriter += nri;
1941
1942 /* Update outer/inner flops */
1943
1944 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1945 }
The ultimate molecular dynamics simulation package