?  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_ElecRFCut_VdwLJSh_GeomW4W4_sse2_single.c
blobb903c622ce4d80a33087a0448849c32930da61f7
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_ElecRFCut_VdwLJSh_GeomW4W4_VF_sse2_single
53 * Electrostatics interaction: ReactionField
54 * VdW interaction: LennardJones
55 * Geometry: Water4-Water4
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecRFCut_VdwLJSh_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 krf = _mm_set1_ps(fr->ic->k_rf);
136 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
137 crf = _mm_set1_ps(fr->ic->c_rf);
138 nvdwtype = fr->ntype;
139 vdwparam = fr->nbfp;
140 vdwtype = mdatoms->typeA;
141
142 /* Setup water-specific parameters */
143 inr = nlist->iinr[0];
144 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
145 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
146 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
147 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
148
149 jq1 = _mm_set1_ps(charge[inr+1]);
150 jq2 = _mm_set1_ps(charge[inr+2]);
151 jq3 = _mm_set1_ps(charge[inr+3]);
152 vdwjidx0A = 2*vdwtype[inr+0];
153 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
154 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
155 qq11 = _mm_mul_ps(iq1,jq1);
156 qq12 = _mm_mul_ps(iq1,jq2);
157 qq13 = _mm_mul_ps(iq1,jq3);
158 qq21 = _mm_mul_ps(iq2,jq1);
159 qq22 = _mm_mul_ps(iq2,jq2);
160 qq23 = _mm_mul_ps(iq2,jq3);
161 qq31 = _mm_mul_ps(iq3,jq1);
162 qq32 = _mm_mul_ps(iq3,jq2);
163 qq33 = _mm_mul_ps(iq3,jq3);
164
165 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
166 rcutoff_scalar = fr->rcoulomb;
167 rcutoff = _mm_set1_ps(rcutoff_scalar);
168 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
169
170 sh_vdw_invrcut6 = _mm_set1_ps(fr->ic->sh_invrc6);
171 rvdw = _mm_set1_ps(fr->rvdw);
172
173 /* Avoid stupid compiler warnings */
174 jnrA = jnrB = jnrC = jnrD = 0;
175 j_coord_offsetA = 0;
176 j_coord_offsetB = 0;
177 j_coord_offsetC = 0;
178 j_coord_offsetD = 0;
179
180 outeriter = 0;
181 inneriter = 0;
182
183 for(iidx=0;iidx<4*DIM;iidx++)
184 {
185 scratch[iidx] = 0.0;
186 }
187
188 /* Start outer loop over neighborlists */
189 for(iidx=0; iidx<nri; iidx++)
190 {
191 /* Load shift vector for this list */
192 i_shift_offset = DIM*shiftidx[iidx];
193
194 /* Load limits for loop over neighbors */
195 j_index_start = jindex[iidx];
196 j_index_end = jindex[iidx+1];
197
198 /* Get outer coordinate index */
199 inr = iinr[iidx];
200 i_coord_offset = DIM*inr;
201
202 /* Load i particle coords and add shift vector */
203 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
204 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
205
206 fix0 = _mm_setzero_ps();
207 fiy0 = _mm_setzero_ps();
208 fiz0 = _mm_setzero_ps();
209 fix1 = _mm_setzero_ps();
210 fiy1 = _mm_setzero_ps();
211 fiz1 = _mm_setzero_ps();
212 fix2 = _mm_setzero_ps();
213 fiy2 = _mm_setzero_ps();
214 fiz2 = _mm_setzero_ps();
215 fix3 = _mm_setzero_ps();
216 fiy3 = _mm_setzero_ps();
217 fiz3 = _mm_setzero_ps();
218
219 /* Reset potential sums */
220 velecsum = _mm_setzero_ps();
221 vvdwsum = _mm_setzero_ps();
222
223 /* Start inner kernel loop */
224 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
225 {
226
227 /* Get j neighbor index, and coordinate index */
228 jnrA = jjnr[jidx];
229 jnrB = jjnr[jidx+1];
230 jnrC = jjnr[jidx+2];
231 jnrD = jjnr[jidx+3];
232 j_coord_offsetA = DIM*jnrA;
233 j_coord_offsetB = DIM*jnrB;
234 j_coord_offsetC = DIM*jnrC;
235 j_coord_offsetD = DIM*jnrD;
236
237 /* load j atom coordinates */
238 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
239 x+j_coord_offsetC,x+j_coord_offsetD,
240 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
241 &jy2,&jz2,&jx3,&jy3,&jz3);
242
243 /* Calculate displacement vector */
244 dx00 = _mm_sub_ps(ix0,jx0);
245 dy00 = _mm_sub_ps(iy0,jy0);
246 dz00 = _mm_sub_ps(iz0,jz0);
247 dx11 = _mm_sub_ps(ix1,jx1);
248 dy11 = _mm_sub_ps(iy1,jy1);
249 dz11 = _mm_sub_ps(iz1,jz1);
250 dx12 = _mm_sub_ps(ix1,jx2);
251 dy12 = _mm_sub_ps(iy1,jy2);
252 dz12 = _mm_sub_ps(iz1,jz2);
253 dx13 = _mm_sub_ps(ix1,jx3);
254 dy13 = _mm_sub_ps(iy1,jy3);
255 dz13 = _mm_sub_ps(iz1,jz3);
256 dx21 = _mm_sub_ps(ix2,jx1);
257 dy21 = _mm_sub_ps(iy2,jy1);
258 dz21 = _mm_sub_ps(iz2,jz1);
259 dx22 = _mm_sub_ps(ix2,jx2);
260 dy22 = _mm_sub_ps(iy2,jy2);
261 dz22 = _mm_sub_ps(iz2,jz2);
262 dx23 = _mm_sub_ps(ix2,jx3);
263 dy23 = _mm_sub_ps(iy2,jy3);
264 dz23 = _mm_sub_ps(iz2,jz3);
265 dx31 = _mm_sub_ps(ix3,jx1);
266 dy31 = _mm_sub_ps(iy3,jy1);
267 dz31 = _mm_sub_ps(iz3,jz1);
268 dx32 = _mm_sub_ps(ix3,jx2);
269 dy32 = _mm_sub_ps(iy3,jy2);
270 dz32 = _mm_sub_ps(iz3,jz2);
271 dx33 = _mm_sub_ps(ix3,jx3);
272 dy33 = _mm_sub_ps(iy3,jy3);
273 dz33 = _mm_sub_ps(iz3,jz3);
274
275 /* Calculate squared distance and things based on it */
276 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
277 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
278 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
279 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
280 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
281 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
282 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
283 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
284 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
285 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
286
287 rinv11 = gmx_mm_invsqrt_ps(rsq11);
288 rinv12 = gmx_mm_invsqrt_ps(rsq12);
289 rinv13 = gmx_mm_invsqrt_ps(rsq13);
290 rinv21 = gmx_mm_invsqrt_ps(rsq21);
291 rinv22 = gmx_mm_invsqrt_ps(rsq22);
292 rinv23 = gmx_mm_invsqrt_ps(rsq23);
293 rinv31 = gmx_mm_invsqrt_ps(rsq31);
294 rinv32 = gmx_mm_invsqrt_ps(rsq32);
295 rinv33 = gmx_mm_invsqrt_ps(rsq33);
296
297 rinvsq00 = gmx_mm_inv_ps(rsq00);
298 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
299 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
300 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
301 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
302 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
303 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
304 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
305 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
306 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
307
308 fjx0 = _mm_setzero_ps();
309 fjy0 = _mm_setzero_ps();
310 fjz0 = _mm_setzero_ps();
311 fjx1 = _mm_setzero_ps();
312 fjy1 = _mm_setzero_ps();
313 fjz1 = _mm_setzero_ps();
314 fjx2 = _mm_setzero_ps();
315 fjy2 = _mm_setzero_ps();
316 fjz2 = _mm_setzero_ps();
317 fjx3 = _mm_setzero_ps();
318 fjy3 = _mm_setzero_ps();
319 fjz3 = _mm_setzero_ps();
320
321 /**************************
322 * CALCULATE INTERACTIONS *
323 **************************/
324
325 if (gmx_mm_any_lt(rsq00,rcutoff2))
326 {
327
328 /* LENNARD-JONES DISPERSION/REPULSION */
329
330 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
331 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
332 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
333 vvdw = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
334 _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
335 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
336
337 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
338
339 /* Update potential sum for this i atom from the interaction with this j atom. */
340 vvdw = _mm_and_ps(vvdw,cutoff_mask);
341 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
342
343 fscal = fvdw;
344
345 fscal = _mm_and_ps(fscal,cutoff_mask);
346
347 /* Calculate temporary vectorial force */
348 tx = _mm_mul_ps(fscal,dx00);
349 ty = _mm_mul_ps(fscal,dy00);
350 tz = _mm_mul_ps(fscal,dz00);
351
352 /* Update vectorial force */
353 fix0 = _mm_add_ps(fix0,tx);
354 fiy0 = _mm_add_ps(fiy0,ty);
355 fiz0 = _mm_add_ps(fiz0,tz);
356
357 fjx0 = _mm_add_ps(fjx0,tx);
358 fjy0 = _mm_add_ps(fjy0,ty);
359 fjz0 = _mm_add_ps(fjz0,tz);
360
361 }
362
363 /**************************
364 * CALCULATE INTERACTIONS *
365 **************************/
366
367 if (gmx_mm_any_lt(rsq11,rcutoff2))
368 {
369
370 /* REACTION-FIELD ELECTROSTATICS */
371 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
372 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
373
374 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
375
376 /* Update potential sum for this i atom from the interaction with this j atom. */
377 velec = _mm_and_ps(velec,cutoff_mask);
378 velecsum = _mm_add_ps(velecsum,velec);
379
380 fscal = felec;
381
382 fscal = _mm_and_ps(fscal,cutoff_mask);
383
384 /* Calculate temporary vectorial force */
385 tx = _mm_mul_ps(fscal,dx11);
386 ty = _mm_mul_ps(fscal,dy11);
387 tz = _mm_mul_ps(fscal,dz11);
388
389 /* Update vectorial force */
390 fix1 = _mm_add_ps(fix1,tx);
391 fiy1 = _mm_add_ps(fiy1,ty);
392 fiz1 = _mm_add_ps(fiz1,tz);
393
394 fjx1 = _mm_add_ps(fjx1,tx);
395 fjy1 = _mm_add_ps(fjy1,ty);
396 fjz1 = _mm_add_ps(fjz1,tz);
397
398 }
399
400 /**************************
401 * CALCULATE INTERACTIONS *
402 **************************/
403
404 if (gmx_mm_any_lt(rsq12,rcutoff2))
405 {
406
407 /* REACTION-FIELD ELECTROSTATICS */
408 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
409 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
410
411 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
412
413 /* Update potential sum for this i atom from the interaction with this j atom. */
414 velec = _mm_and_ps(velec,cutoff_mask);
415 velecsum = _mm_add_ps(velecsum,velec);
416
417 fscal = felec;
418
419 fscal = _mm_and_ps(fscal,cutoff_mask);
420
421 /* Calculate temporary vectorial force */
422 tx = _mm_mul_ps(fscal,dx12);
423 ty = _mm_mul_ps(fscal,dy12);
424 tz = _mm_mul_ps(fscal,dz12);
425
426 /* Update vectorial force */
427 fix1 = _mm_add_ps(fix1,tx);
428 fiy1 = _mm_add_ps(fiy1,ty);
429 fiz1 = _mm_add_ps(fiz1,tz);
430
431 fjx2 = _mm_add_ps(fjx2,tx);
432 fjy2 = _mm_add_ps(fjy2,ty);
433 fjz2 = _mm_add_ps(fjz2,tz);
434
435 }
436
437 /**************************
438 * CALCULATE INTERACTIONS *
439 **************************/
440
441 if (gmx_mm_any_lt(rsq13,rcutoff2))
442 {
443
444 /* REACTION-FIELD ELECTROSTATICS */
445 velec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
446 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
447
448 cutoff_mask = _mm_cmplt_ps(rsq13,rcutoff2);
449
450 /* Update potential sum for this i atom from the interaction with this j atom. */
451 velec = _mm_and_ps(velec,cutoff_mask);
452 velecsum = _mm_add_ps(velecsum,velec);
453
454 fscal = felec;
455
456 fscal = _mm_and_ps(fscal,cutoff_mask);
457
458 /* Calculate temporary vectorial force */
459 tx = _mm_mul_ps(fscal,dx13);
460 ty = _mm_mul_ps(fscal,dy13);
461 tz = _mm_mul_ps(fscal,dz13);
462
463 /* Update vectorial force */
464 fix1 = _mm_add_ps(fix1,tx);
465 fiy1 = _mm_add_ps(fiy1,ty);
466 fiz1 = _mm_add_ps(fiz1,tz);
467
468 fjx3 = _mm_add_ps(fjx3,tx);
469 fjy3 = _mm_add_ps(fjy3,ty);
470 fjz3 = _mm_add_ps(fjz3,tz);
471
472 }
473
474 /**************************
475 * CALCULATE INTERACTIONS *
476 **************************/
477
478 if (gmx_mm_any_lt(rsq21,rcutoff2))
479 {
480
481 /* REACTION-FIELD ELECTROSTATICS */
482 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
483 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
484
485 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
486
487 /* Update potential sum for this i atom from the interaction with this j atom. */
488 velec = _mm_and_ps(velec,cutoff_mask);
489 velecsum = _mm_add_ps(velecsum,velec);
490
491 fscal = felec;
492
493 fscal = _mm_and_ps(fscal,cutoff_mask);
494
495 /* Calculate temporary vectorial force */
496 tx = _mm_mul_ps(fscal,dx21);
497 ty = _mm_mul_ps(fscal,dy21);
498 tz = _mm_mul_ps(fscal,dz21);
499
500 /* Update vectorial force */
501 fix2 = _mm_add_ps(fix2,tx);
502 fiy2 = _mm_add_ps(fiy2,ty);
503 fiz2 = _mm_add_ps(fiz2,tz);
504
505 fjx1 = _mm_add_ps(fjx1,tx);
506 fjy1 = _mm_add_ps(fjy1,ty);
507 fjz1 = _mm_add_ps(fjz1,tz);
508
509 }
510
511 /**************************
512 * CALCULATE INTERACTIONS *
513 **************************/
514
515 if (gmx_mm_any_lt(rsq22,rcutoff2))
516 {
517
518 /* REACTION-FIELD ELECTROSTATICS */
519 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
520 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
521
522 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
523
524 /* Update potential sum for this i atom from the interaction with this j atom. */
525 velec = _mm_and_ps(velec,cutoff_mask);
526 velecsum = _mm_add_ps(velecsum,velec);
527
528 fscal = felec;
529
530 fscal = _mm_and_ps(fscal,cutoff_mask);
531
532 /* Calculate temporary vectorial force */
533 tx = _mm_mul_ps(fscal,dx22);
534 ty = _mm_mul_ps(fscal,dy22);
535 tz = _mm_mul_ps(fscal,dz22);
536
537 /* Update vectorial force */
538 fix2 = _mm_add_ps(fix2,tx);
539 fiy2 = _mm_add_ps(fiy2,ty);
540 fiz2 = _mm_add_ps(fiz2,tz);
541
542 fjx2 = _mm_add_ps(fjx2,tx);
543 fjy2 = _mm_add_ps(fjy2,ty);
544 fjz2 = _mm_add_ps(fjz2,tz);
545
546 }
547
548 /**************************
549 * CALCULATE INTERACTIONS *
550 **************************/
551
552 if (gmx_mm_any_lt(rsq23,rcutoff2))
553 {
554
555 /* REACTION-FIELD ELECTROSTATICS */
556 velec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
557 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
558
559 cutoff_mask = _mm_cmplt_ps(rsq23,rcutoff2);
560
561 /* Update potential sum for this i atom from the interaction with this j atom. */
562 velec = _mm_and_ps(velec,cutoff_mask);
563 velecsum = _mm_add_ps(velecsum,velec);
564
565 fscal = felec;
566
567 fscal = _mm_and_ps(fscal,cutoff_mask);
568
569 /* Calculate temporary vectorial force */
570 tx = _mm_mul_ps(fscal,dx23);
571 ty = _mm_mul_ps(fscal,dy23);
572 tz = _mm_mul_ps(fscal,dz23);
573
574 /* Update vectorial force */
575 fix2 = _mm_add_ps(fix2,tx);
576 fiy2 = _mm_add_ps(fiy2,ty);
577 fiz2 = _mm_add_ps(fiz2,tz);
578
579 fjx3 = _mm_add_ps(fjx3,tx);
580 fjy3 = _mm_add_ps(fjy3,ty);
581 fjz3 = _mm_add_ps(fjz3,tz);
582
583 }
584
585 /**************************
586 * CALCULATE INTERACTIONS *
587 **************************/
588
589 if (gmx_mm_any_lt(rsq31,rcutoff2))
590 {
591
592 /* REACTION-FIELD ELECTROSTATICS */
593 velec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
594 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
595
596 cutoff_mask = _mm_cmplt_ps(rsq31,rcutoff2);
597
598 /* Update potential sum for this i atom from the interaction with this j atom. */
599 velec = _mm_and_ps(velec,cutoff_mask);
600 velecsum = _mm_add_ps(velecsum,velec);
601
602 fscal = felec;
603
604 fscal = _mm_and_ps(fscal,cutoff_mask);
605
606 /* Calculate temporary vectorial force */
607 tx = _mm_mul_ps(fscal,dx31);
608 ty = _mm_mul_ps(fscal,dy31);
609 tz = _mm_mul_ps(fscal,dz31);
610
611 /* Update vectorial force */
612 fix3 = _mm_add_ps(fix3,tx);
613 fiy3 = _mm_add_ps(fiy3,ty);
614 fiz3 = _mm_add_ps(fiz3,tz);
615
616 fjx1 = _mm_add_ps(fjx1,tx);
617 fjy1 = _mm_add_ps(fjy1,ty);
618 fjz1 = _mm_add_ps(fjz1,tz);
619
620 }
621
622 /**************************
623 * CALCULATE INTERACTIONS *
624 **************************/
625
626 if (gmx_mm_any_lt(rsq32,rcutoff2))
627 {
628
629 /* REACTION-FIELD ELECTROSTATICS */
630 velec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
631 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
632
633 cutoff_mask = _mm_cmplt_ps(rsq32,rcutoff2);
634
635 /* Update potential sum for this i atom from the interaction with this j atom. */
636 velec = _mm_and_ps(velec,cutoff_mask);
637 velecsum = _mm_add_ps(velecsum,velec);
638
639 fscal = felec;
640
641 fscal = _mm_and_ps(fscal,cutoff_mask);
642
643 /* Calculate temporary vectorial force */
644 tx = _mm_mul_ps(fscal,dx32);
645 ty = _mm_mul_ps(fscal,dy32);
646 tz = _mm_mul_ps(fscal,dz32);
647
648 /* Update vectorial force */
649 fix3 = _mm_add_ps(fix3,tx);
650 fiy3 = _mm_add_ps(fiy3,ty);
651 fiz3 = _mm_add_ps(fiz3,tz);
652
653 fjx2 = _mm_add_ps(fjx2,tx);
654 fjy2 = _mm_add_ps(fjy2,ty);
655 fjz2 = _mm_add_ps(fjz2,tz);
656
657 }
658
659 /**************************
660 * CALCULATE INTERACTIONS *
661 **************************/
662
663 if (gmx_mm_any_lt(rsq33,rcutoff2))
664 {
665
666 /* REACTION-FIELD ELECTROSTATICS */
667 velec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
668 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
669
670 cutoff_mask = _mm_cmplt_ps(rsq33,rcutoff2);
671
672 /* Update potential sum for this i atom from the interaction with this j atom. */
673 velec = _mm_and_ps(velec,cutoff_mask);
674 velecsum = _mm_add_ps(velecsum,velec);
675
676 fscal = felec;
677
678 fscal = _mm_and_ps(fscal,cutoff_mask);
679
680 /* Calculate temporary vectorial force */
681 tx = _mm_mul_ps(fscal,dx33);
682 ty = _mm_mul_ps(fscal,dy33);
683 tz = _mm_mul_ps(fscal,dz33);
684
685 /* Update vectorial force */
686 fix3 = _mm_add_ps(fix3,tx);
687 fiy3 = _mm_add_ps(fiy3,ty);
688 fiz3 = _mm_add_ps(fiz3,tz);
689
690 fjx3 = _mm_add_ps(fjx3,tx);
691 fjy3 = _mm_add_ps(fjy3,ty);
692 fjz3 = _mm_add_ps(fjz3,tz);
693
694 }
695
696 fjptrA = f+j_coord_offsetA;
697 fjptrB = f+j_coord_offsetB;
698 fjptrC = f+j_coord_offsetC;
699 fjptrD = f+j_coord_offsetD;
700
701 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
702 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
703 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
704
705 /* Inner loop uses 368 flops */
706 }
707
708 if(jidx<j_index_end)
709 {
710
711 /* Get j neighbor index, and coordinate index */
712 jnrlistA = jjnr[jidx];
713 jnrlistB = jjnr[jidx+1];
714 jnrlistC = jjnr[jidx+2];
715 jnrlistD = jjnr[jidx+3];
716 /* Sign of each element will be negative for non-real atoms.
717 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
718 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
719 */
720 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
721 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
722 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
723 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
724 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
725 j_coord_offsetA = DIM*jnrA;
726 j_coord_offsetB = DIM*jnrB;
727 j_coord_offsetC = DIM*jnrC;
728 j_coord_offsetD = DIM*jnrD;
729
730 /* load j atom coordinates */
731 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
732 x+j_coord_offsetC,x+j_coord_offsetD,
733 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
734 &jy2,&jz2,&jx3,&jy3,&jz3);
735
736 /* Calculate displacement vector */
737 dx00 = _mm_sub_ps(ix0,jx0);
738 dy00 = _mm_sub_ps(iy0,jy0);
739 dz00 = _mm_sub_ps(iz0,jz0);
740 dx11 = _mm_sub_ps(ix1,jx1);
741 dy11 = _mm_sub_ps(iy1,jy1);
742 dz11 = _mm_sub_ps(iz1,jz1);
743 dx12 = _mm_sub_ps(ix1,jx2);
744 dy12 = _mm_sub_ps(iy1,jy2);
745 dz12 = _mm_sub_ps(iz1,jz2);
746 dx13 = _mm_sub_ps(ix1,jx3);
747 dy13 = _mm_sub_ps(iy1,jy3);
748 dz13 = _mm_sub_ps(iz1,jz3);
749 dx21 = _mm_sub_ps(ix2,jx1);
750 dy21 = _mm_sub_ps(iy2,jy1);
751 dz21 = _mm_sub_ps(iz2,jz1);
752 dx22 = _mm_sub_ps(ix2,jx2);
753 dy22 = _mm_sub_ps(iy2,jy2);
754 dz22 = _mm_sub_ps(iz2,jz2);
755 dx23 = _mm_sub_ps(ix2,jx3);
756 dy23 = _mm_sub_ps(iy2,jy3);
757 dz23 = _mm_sub_ps(iz2,jz3);
758 dx31 = _mm_sub_ps(ix3,jx1);
759 dy31 = _mm_sub_ps(iy3,jy1);
760 dz31 = _mm_sub_ps(iz3,jz1);
761 dx32 = _mm_sub_ps(ix3,jx2);
762 dy32 = _mm_sub_ps(iy3,jy2);
763 dz32 = _mm_sub_ps(iz3,jz2);
764 dx33 = _mm_sub_ps(ix3,jx3);
765 dy33 = _mm_sub_ps(iy3,jy3);
766 dz33 = _mm_sub_ps(iz3,jz3);
767
768 /* Calculate squared distance and things based on it */
769 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
770 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
771 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
772 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
773 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
774 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
775 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
776 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
777 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
778 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
779
780 rinv11 = gmx_mm_invsqrt_ps(rsq11);
781 rinv12 = gmx_mm_invsqrt_ps(rsq12);
782 rinv13 = gmx_mm_invsqrt_ps(rsq13);
783 rinv21 = gmx_mm_invsqrt_ps(rsq21);
784 rinv22 = gmx_mm_invsqrt_ps(rsq22);
785 rinv23 = gmx_mm_invsqrt_ps(rsq23);
786 rinv31 = gmx_mm_invsqrt_ps(rsq31);
787 rinv32 = gmx_mm_invsqrt_ps(rsq32);
788 rinv33 = gmx_mm_invsqrt_ps(rsq33);
789
790 rinvsq00 = gmx_mm_inv_ps(rsq00);
791 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
792 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
793 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
794 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
795 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
796 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
797 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
798 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
799 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
800
801 fjx0 = _mm_setzero_ps();
802 fjy0 = _mm_setzero_ps();
803 fjz0 = _mm_setzero_ps();
804 fjx1 = _mm_setzero_ps();
805 fjy1 = _mm_setzero_ps();
806 fjz1 = _mm_setzero_ps();
807 fjx2 = _mm_setzero_ps();
808 fjy2 = _mm_setzero_ps();
809 fjz2 = _mm_setzero_ps();
810 fjx3 = _mm_setzero_ps();
811 fjy3 = _mm_setzero_ps();
812 fjz3 = _mm_setzero_ps();
813
814 /**************************
815 * CALCULATE INTERACTIONS *
816 **************************/
817
818 if (gmx_mm_any_lt(rsq00,rcutoff2))
819 {
820
821 /* LENNARD-JONES DISPERSION/REPULSION */
822
823 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
824 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
825 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
826 vvdw = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
827 _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
828 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
829
830 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
831
832 /* Update potential sum for this i atom from the interaction with this j atom. */
833 vvdw = _mm_and_ps(vvdw,cutoff_mask);
834 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
835 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
836
837 fscal = fvdw;
838
839 fscal = _mm_and_ps(fscal,cutoff_mask);
840
841 fscal = _mm_andnot_ps(dummy_mask,fscal);
842
843 /* Calculate temporary vectorial force */
844 tx = _mm_mul_ps(fscal,dx00);
845 ty = _mm_mul_ps(fscal,dy00);
846 tz = _mm_mul_ps(fscal,dz00);
847
848 /* Update vectorial force */
849 fix0 = _mm_add_ps(fix0,tx);
850 fiy0 = _mm_add_ps(fiy0,ty);
851 fiz0 = _mm_add_ps(fiz0,tz);
852
853 fjx0 = _mm_add_ps(fjx0,tx);
854 fjy0 = _mm_add_ps(fjy0,ty);
855 fjz0 = _mm_add_ps(fjz0,tz);
856
857 }
858
859 /**************************
860 * CALCULATE INTERACTIONS *
861 **************************/
862
863 if (gmx_mm_any_lt(rsq11,rcutoff2))
864 {
865
866 /* REACTION-FIELD ELECTROSTATICS */
867 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
868 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
869
870 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
871
872 /* Update potential sum for this i atom from the interaction with this j atom. */
873 velec = _mm_and_ps(velec,cutoff_mask);
874 velec = _mm_andnot_ps(dummy_mask,velec);
875 velecsum = _mm_add_ps(velecsum,velec);
876
877 fscal = felec;
878
879 fscal = _mm_and_ps(fscal,cutoff_mask);
880
881 fscal = _mm_andnot_ps(dummy_mask,fscal);
882
883 /* Calculate temporary vectorial force */
884 tx = _mm_mul_ps(fscal,dx11);
885 ty = _mm_mul_ps(fscal,dy11);
886 tz = _mm_mul_ps(fscal,dz11);
887
888 /* Update vectorial force */
889 fix1 = _mm_add_ps(fix1,tx);
890 fiy1 = _mm_add_ps(fiy1,ty);
891 fiz1 = _mm_add_ps(fiz1,tz);
892
893 fjx1 = _mm_add_ps(fjx1,tx);
894 fjy1 = _mm_add_ps(fjy1,ty);
895 fjz1 = _mm_add_ps(fjz1,tz);
896
897 }
898
899 /**************************
900 * CALCULATE INTERACTIONS *
901 **************************/
902
903 if (gmx_mm_any_lt(rsq12,rcutoff2))
904 {
905
906 /* REACTION-FIELD ELECTROSTATICS */
907 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
908 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
909
910 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
911
912 /* Update potential sum for this i atom from the interaction with this j atom. */
913 velec = _mm_and_ps(velec,cutoff_mask);
914 velec = _mm_andnot_ps(dummy_mask,velec);
915 velecsum = _mm_add_ps(velecsum,velec);
916
917 fscal = felec;
918
919 fscal = _mm_and_ps(fscal,cutoff_mask);
920
921 fscal = _mm_andnot_ps(dummy_mask,fscal);
922
923 /* Calculate temporary vectorial force */
924 tx = _mm_mul_ps(fscal,dx12);
925 ty = _mm_mul_ps(fscal,dy12);
926 tz = _mm_mul_ps(fscal,dz12);
927
928 /* Update vectorial force */
929 fix1 = _mm_add_ps(fix1,tx);
930 fiy1 = _mm_add_ps(fiy1,ty);
931 fiz1 = _mm_add_ps(fiz1,tz);
932
933 fjx2 = _mm_add_ps(fjx2,tx);
934 fjy2 = _mm_add_ps(fjy2,ty);
935 fjz2 = _mm_add_ps(fjz2,tz);
936
937 }
938
939 /**************************
940 * CALCULATE INTERACTIONS *
941 **************************/
942
943 if (gmx_mm_any_lt(rsq13,rcutoff2))
944 {
945
946 /* REACTION-FIELD ELECTROSTATICS */
947 velec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
948 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
949
950 cutoff_mask = _mm_cmplt_ps(rsq13,rcutoff2);
951
952 /* Update potential sum for this i atom from the interaction with this j atom. */
953 velec = _mm_and_ps(velec,cutoff_mask);
954 velec = _mm_andnot_ps(dummy_mask,velec);
955 velecsum = _mm_add_ps(velecsum,velec);
956
957 fscal = felec;
958
959 fscal = _mm_and_ps(fscal,cutoff_mask);
960
961 fscal = _mm_andnot_ps(dummy_mask,fscal);
962
963 /* Calculate temporary vectorial force */
964 tx = _mm_mul_ps(fscal,dx13);
965 ty = _mm_mul_ps(fscal,dy13);
966 tz = _mm_mul_ps(fscal,dz13);
967
968 /* Update vectorial force */
969 fix1 = _mm_add_ps(fix1,tx);
970 fiy1 = _mm_add_ps(fiy1,ty);
971 fiz1 = _mm_add_ps(fiz1,tz);
972
973 fjx3 = _mm_add_ps(fjx3,tx);
974 fjy3 = _mm_add_ps(fjy3,ty);
975 fjz3 = _mm_add_ps(fjz3,tz);
976
977 }
978
979 /**************************
980 * CALCULATE INTERACTIONS *
981 **************************/
982
983 if (gmx_mm_any_lt(rsq21,rcutoff2))
984 {
985
986 /* REACTION-FIELD ELECTROSTATICS */
987 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
988 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
989
990 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
991
992 /* Update potential sum for this i atom from the interaction with this j atom. */
993 velec = _mm_and_ps(velec,cutoff_mask);
994 velec = _mm_andnot_ps(dummy_mask,velec);
995 velecsum = _mm_add_ps(velecsum,velec);
996
997 fscal = felec;
998
999 fscal = _mm_and_ps(fscal,cutoff_mask);
1000
1001 fscal = _mm_andnot_ps(dummy_mask,fscal);
1002
1003 /* Calculate temporary vectorial force */
1004 tx = _mm_mul_ps(fscal,dx21);
1005 ty = _mm_mul_ps(fscal,dy21);
1006 tz = _mm_mul_ps(fscal,dz21);
1007
1008 /* Update vectorial force */
1009 fix2 = _mm_add_ps(fix2,tx);
1010 fiy2 = _mm_add_ps(fiy2,ty);
1011 fiz2 = _mm_add_ps(fiz2,tz);
1012
1013 fjx1 = _mm_add_ps(fjx1,tx);
1014 fjy1 = _mm_add_ps(fjy1,ty);
1015 fjz1 = _mm_add_ps(fjz1,tz);
1016
1017 }
1018
1019 /**************************
1020 * CALCULATE INTERACTIONS *
1021 **************************/
1022
1023 if (gmx_mm_any_lt(rsq22,rcutoff2))
1024 {
1025
1026 /* REACTION-FIELD ELECTROSTATICS */
1027 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1028 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1029
1030 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
1031
1032 /* Update potential sum for this i atom from the interaction with this j atom. */
1033 velec = _mm_and_ps(velec,cutoff_mask);
1034 velec = _mm_andnot_ps(dummy_mask,velec);
1035 velecsum = _mm_add_ps(velecsum,velec);
1036
1037 fscal = felec;
1038
1039 fscal = _mm_and_ps(fscal,cutoff_mask);
1040
1041 fscal = _mm_andnot_ps(dummy_mask,fscal);
1042
1043 /* Calculate temporary vectorial force */
1044 tx = _mm_mul_ps(fscal,dx22);
1045 ty = _mm_mul_ps(fscal,dy22);
1046 tz = _mm_mul_ps(fscal,dz22);
1047
1048 /* Update vectorial force */
1049 fix2 = _mm_add_ps(fix2,tx);
1050 fiy2 = _mm_add_ps(fiy2,ty);
1051 fiz2 = _mm_add_ps(fiz2,tz);
1052
1053 fjx2 = _mm_add_ps(fjx2,tx);
1054 fjy2 = _mm_add_ps(fjy2,ty);
1055 fjz2 = _mm_add_ps(fjz2,tz);
1056
1057 }
1058
1059 /**************************
1060 * CALCULATE INTERACTIONS *
1061 **************************/
1062
1063 if (gmx_mm_any_lt(rsq23,rcutoff2))
1064 {
1065
1066 /* REACTION-FIELD ELECTROSTATICS */
1067 velec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
1068 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1069
1070 cutoff_mask = _mm_cmplt_ps(rsq23,rcutoff2);
1071
1072 /* Update potential sum for this i atom from the interaction with this j atom. */
1073 velec = _mm_and_ps(velec,cutoff_mask);
1074 velec = _mm_andnot_ps(dummy_mask,velec);
1075 velecsum = _mm_add_ps(velecsum,velec);
1076
1077 fscal = felec;
1078
1079 fscal = _mm_and_ps(fscal,cutoff_mask);
1080
1081 fscal = _mm_andnot_ps(dummy_mask,fscal);
1082
1083 /* Calculate temporary vectorial force */
1084 tx = _mm_mul_ps(fscal,dx23);
1085 ty = _mm_mul_ps(fscal,dy23);
1086 tz = _mm_mul_ps(fscal,dz23);
1087
1088 /* Update vectorial force */
1089 fix2 = _mm_add_ps(fix2,tx);
1090 fiy2 = _mm_add_ps(fiy2,ty);
1091 fiz2 = _mm_add_ps(fiz2,tz);
1092
1093 fjx3 = _mm_add_ps(fjx3,tx);
1094 fjy3 = _mm_add_ps(fjy3,ty);
1095 fjz3 = _mm_add_ps(fjz3,tz);
1096
1097 }
1098
1099 /**************************
1100 * CALCULATE INTERACTIONS *
1101 **************************/
1102
1103 if (gmx_mm_any_lt(rsq31,rcutoff2))
1104 {
1105
1106 /* REACTION-FIELD ELECTROSTATICS */
1107 velec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
1108 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1109
1110 cutoff_mask = _mm_cmplt_ps(rsq31,rcutoff2);
1111
1112 /* Update potential sum for this i atom from the interaction with this j atom. */
1113 velec = _mm_and_ps(velec,cutoff_mask);
1114 velec = _mm_andnot_ps(dummy_mask,velec);
1115 velecsum = _mm_add_ps(velecsum,velec);
1116
1117 fscal = felec;
1118
1119 fscal = _mm_and_ps(fscal,cutoff_mask);
1120
1121 fscal = _mm_andnot_ps(dummy_mask,fscal);
1122
1123 /* Calculate temporary vectorial force */
1124 tx = _mm_mul_ps(fscal,dx31);
1125 ty = _mm_mul_ps(fscal,dy31);
1126 tz = _mm_mul_ps(fscal,dz31);
1127
1128 /* Update vectorial force */
1129 fix3 = _mm_add_ps(fix3,tx);
1130 fiy3 = _mm_add_ps(fiy3,ty);
1131 fiz3 = _mm_add_ps(fiz3,tz);
1132
1133 fjx1 = _mm_add_ps(fjx1,tx);
1134 fjy1 = _mm_add_ps(fjy1,ty);
1135 fjz1 = _mm_add_ps(fjz1,tz);
1136
1137 }
1138
1139 /**************************
1140 * CALCULATE INTERACTIONS *
1141 **************************/
1142
1143 if (gmx_mm_any_lt(rsq32,rcutoff2))
1144 {
1145
1146 /* REACTION-FIELD ELECTROSTATICS */
1147 velec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1148 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1149
1150 cutoff_mask = _mm_cmplt_ps(rsq32,rcutoff2);
1151
1152 /* Update potential sum for this i atom from the interaction with this j atom. */
1153 velec = _mm_and_ps(velec,cutoff_mask);
1154 velec = _mm_andnot_ps(dummy_mask,velec);
1155 velecsum = _mm_add_ps(velecsum,velec);
1156
1157 fscal = felec;
1158
1159 fscal = _mm_and_ps(fscal,cutoff_mask);
1160
1161 fscal = _mm_andnot_ps(dummy_mask,fscal);
1162
1163 /* Calculate temporary vectorial force */
1164 tx = _mm_mul_ps(fscal,dx32);
1165 ty = _mm_mul_ps(fscal,dy32);
1166 tz = _mm_mul_ps(fscal,dz32);
1167
1168 /* Update vectorial force */
1169 fix3 = _mm_add_ps(fix3,tx);
1170 fiy3 = _mm_add_ps(fiy3,ty);
1171 fiz3 = _mm_add_ps(fiz3,tz);
1172
1173 fjx2 = _mm_add_ps(fjx2,tx);
1174 fjy2 = _mm_add_ps(fjy2,ty);
1175 fjz2 = _mm_add_ps(fjz2,tz);
1176
1177 }
1178
1179 /**************************
1180 * CALCULATE INTERACTIONS *
1181 **************************/
1182
1183 if (gmx_mm_any_lt(rsq33,rcutoff2))
1184 {
1185
1186 /* REACTION-FIELD ELECTROSTATICS */
1187 velec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1188 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1189
1190 cutoff_mask = _mm_cmplt_ps(rsq33,rcutoff2);
1191
1192 /* Update potential sum for this i atom from the interaction with this j atom. */
1193 velec = _mm_and_ps(velec,cutoff_mask);
1194 velec = _mm_andnot_ps(dummy_mask,velec);
1195 velecsum = _mm_add_ps(velecsum,velec);
1196
1197 fscal = felec;
1198
1199 fscal = _mm_and_ps(fscal,cutoff_mask);
1200
1201 fscal = _mm_andnot_ps(dummy_mask,fscal);
1202
1203 /* Calculate temporary vectorial force */
1204 tx = _mm_mul_ps(fscal,dx33);
1205 ty = _mm_mul_ps(fscal,dy33);
1206 tz = _mm_mul_ps(fscal,dz33);
1207
1208 /* Update vectorial force */
1209 fix3 = _mm_add_ps(fix3,tx);
1210 fiy3 = _mm_add_ps(fiy3,ty);
1211 fiz3 = _mm_add_ps(fiz3,tz);
1212
1213 fjx3 = _mm_add_ps(fjx3,tx);
1214 fjy3 = _mm_add_ps(fjy3,ty);
1215 fjz3 = _mm_add_ps(fjz3,tz);
1216
1217 }
1218
1219 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1220 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1221 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1222 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1223
1224 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1225 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1226 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1227
1228 /* Inner loop uses 368 flops */
1229 }
1230
1231 /* End of innermost loop */
1232
1233 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1234 f+i_coord_offset,fshift+i_shift_offset);
1235
1236 ggid = gid[iidx];
1237 /* Update potential energies */
1238 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1239 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1240
1241 /* Increment number of inner iterations */
1242 inneriter += j_index_end - j_index_start;
1243
1244 /* Outer loop uses 26 flops */
1245 }
1246
1247 /* Increment number of outer iterations */
1248 outeriter += nri;
1249
1250 /* Update outer/inner flops */
1251
1252 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*368);
1253 }
1254 /*
1255 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_sse2_single
1256 * Electrostatics interaction: ReactionField
1257 * VdW interaction: LennardJones
1258 * Geometry: Water4-Water4
1259 * Calculate force/pot: Force
1260 */
1261 void
1262 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_sse2_single
1263 (t_nblist * gmx_restrict nlist,
1264 rvec * gmx_restrict xx,
1265 rvec * gmx_restrict ff,
1266 t_forcerec * gmx_restrict fr,
1267 t_mdatoms * gmx_restrict mdatoms,
1268 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1269 t_nrnb * gmx_restrict nrnb)
1270 {
1271 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1272 * just 0 for non-waters.
1273 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1274 * jnr indices corresponding to data put in the four positions in the SIMD register.
1275 */
1276 int i_shift_offset,i_coord_offset,outeriter,inneriter;
1277 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1278 int jnrA,jnrB,jnrC,jnrD;
1279 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1280 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1281 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
1282 real rcutoff_scalar;
1283 real *shiftvec,*fshift,*x,*f;
1284 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1285 real scratch[4*DIM];
1286 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1287 int vdwioffset0;
1288 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1289 int vdwioffset1;
1290 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1291 int vdwioffset2;
1292 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1293 int vdwioffset3;
1294 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1295 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1296 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1297 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1298 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1299 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1300 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1301 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1302 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1303 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1304 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1305 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1306 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1307 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1308 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1309 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1310 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1311 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1312 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1313 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
1314 real *charge;
1315 int nvdwtype;
1316 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1317 int *vdwtype;
1318 real *vdwparam;
1319 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
1320 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
1321 __m128 dummy_mask,cutoff_mask;
1322 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1323 __m128 one = _mm_set1_ps(1.0);
1324 __m128 two = _mm_set1_ps(2.0);
1325 x = xx[0];
1326 f = ff[0];
1327
1328 nri = nlist->nri;
1329 iinr = nlist->iinr;
1330 jindex = nlist->jindex;
1331 jjnr = nlist->jjnr;
1332 shiftidx = nlist->shift;
1333 gid = nlist->gid;
1334 shiftvec = fr->shift_vec[0];
1335 fshift = fr->fshift[0];
1336 facel = _mm_set1_ps(fr->epsfac);
1337 charge = mdatoms->chargeA;
1338 krf = _mm_set1_ps(fr->ic->k_rf);
1339 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
1340 crf = _mm_set1_ps(fr->ic->c_rf);
1341 nvdwtype = fr->ntype;
1342 vdwparam = fr->nbfp;
1343 vdwtype = mdatoms->typeA;
1344
1345 /* Setup water-specific parameters */
1346 inr = nlist->iinr[0];
1347 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1348 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1349 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1350 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1351
1352 jq1 = _mm_set1_ps(charge[inr+1]);
1353 jq2 = _mm_set1_ps(charge[inr+2]);
1354 jq3 = _mm_set1_ps(charge[inr+3]);
1355 vdwjidx0A = 2*vdwtype[inr+0];
1356 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1357 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1358 qq11 = _mm_mul_ps(iq1,jq1);
1359 qq12 = _mm_mul_ps(iq1,jq2);
1360 qq13 = _mm_mul_ps(iq1,jq3);
1361 qq21 = _mm_mul_ps(iq2,jq1);
1362 qq22 = _mm_mul_ps(iq2,jq2);
1363 qq23 = _mm_mul_ps(iq2,jq3);
1364 qq31 = _mm_mul_ps(iq3,jq1);
1365 qq32 = _mm_mul_ps(iq3,jq2);
1366 qq33 = _mm_mul_ps(iq3,jq3);
1367
1368 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1369 rcutoff_scalar = fr->rcoulomb;
1370 rcutoff = _mm_set1_ps(rcutoff_scalar);
1371 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
1372
1373 sh_vdw_invrcut6 = _mm_set1_ps(fr->ic->sh_invrc6);
1374 rvdw = _mm_set1_ps(fr->rvdw);
1375
1376 /* Avoid stupid compiler warnings */
1377 jnrA = jnrB = jnrC = jnrD = 0;
1378 j_coord_offsetA = 0;
1379 j_coord_offsetB = 0;
1380 j_coord_offsetC = 0;
1381 j_coord_offsetD = 0;
1382
1383 outeriter = 0;
1384 inneriter = 0;
1385
1386 for(iidx=0;iidx<4*DIM;iidx++)
1387 {
1388 scratch[iidx] = 0.0;
1389 }
1390
1391 /* Start outer loop over neighborlists */
1392 for(iidx=0; iidx<nri; iidx++)
1393 {
1394 /* Load shift vector for this list */
1395 i_shift_offset = DIM*shiftidx[iidx];
1396
1397 /* Load limits for loop over neighbors */
1398 j_index_start = jindex[iidx];
1399 j_index_end = jindex[iidx+1];
1400
1401 /* Get outer coordinate index */
1402 inr = iinr[iidx];
1403 i_coord_offset = DIM*inr;
1404
1405 /* Load i particle coords and add shift vector */
1406 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1407 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1408
1409 fix0 = _mm_setzero_ps();
1410 fiy0 = _mm_setzero_ps();
1411 fiz0 = _mm_setzero_ps();
1412 fix1 = _mm_setzero_ps();
1413 fiy1 = _mm_setzero_ps();
1414 fiz1 = _mm_setzero_ps();
1415 fix2 = _mm_setzero_ps();
1416 fiy2 = _mm_setzero_ps();
1417 fiz2 = _mm_setzero_ps();
1418 fix3 = _mm_setzero_ps();
1419 fiy3 = _mm_setzero_ps();
1420 fiz3 = _mm_setzero_ps();
1421
1422 /* Start inner kernel loop */
1423 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1424 {
1425
1426 /* Get j neighbor index, and coordinate index */
1427 jnrA = jjnr[jidx];
1428 jnrB = jjnr[jidx+1];
1429 jnrC = jjnr[jidx+2];
1430 jnrD = jjnr[jidx+3];
1431 j_coord_offsetA = DIM*jnrA;
1432 j_coord_offsetB = DIM*jnrB;
1433 j_coord_offsetC = DIM*jnrC;
1434 j_coord_offsetD = DIM*jnrD;
1435
1436 /* load j atom coordinates */
1437 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1438 x+j_coord_offsetC,x+j_coord_offsetD,
1439 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1440 &jy2,&jz2,&jx3,&jy3,&jz3);
1441
1442 /* Calculate displacement vector */
1443 dx00 = _mm_sub_ps(ix0,jx0);
1444 dy00 = _mm_sub_ps(iy0,jy0);
1445 dz00 = _mm_sub_ps(iz0,jz0);
1446 dx11 = _mm_sub_ps(ix1,jx1);
1447 dy11 = _mm_sub_ps(iy1,jy1);
1448 dz11 = _mm_sub_ps(iz1,jz1);
1449 dx12 = _mm_sub_ps(ix1,jx2);
1450 dy12 = _mm_sub_ps(iy1,jy2);
1451 dz12 = _mm_sub_ps(iz1,jz2);
1452 dx13 = _mm_sub_ps(ix1,jx3);
1453 dy13 = _mm_sub_ps(iy1,jy3);
1454 dz13 = _mm_sub_ps(iz1,jz3);
1455 dx21 = _mm_sub_ps(ix2,jx1);
1456 dy21 = _mm_sub_ps(iy2,jy1);
1457 dz21 = _mm_sub_ps(iz2,jz1);
1458 dx22 = _mm_sub_ps(ix2,jx2);
1459 dy22 = _mm_sub_ps(iy2,jy2);
1460 dz22 = _mm_sub_ps(iz2,jz2);
1461 dx23 = _mm_sub_ps(ix2,jx3);
1462 dy23 = _mm_sub_ps(iy2,jy3);
1463 dz23 = _mm_sub_ps(iz2,jz3);
1464 dx31 = _mm_sub_ps(ix3,jx1);
1465 dy31 = _mm_sub_ps(iy3,jy1);
1466 dz31 = _mm_sub_ps(iz3,jz1);
1467 dx32 = _mm_sub_ps(ix3,jx2);
1468 dy32 = _mm_sub_ps(iy3,jy2);
1469 dz32 = _mm_sub_ps(iz3,jz2);
1470 dx33 = _mm_sub_ps(ix3,jx3);
1471 dy33 = _mm_sub_ps(iy3,jy3);
1472 dz33 = _mm_sub_ps(iz3,jz3);
1473
1474 /* Calculate squared distance and things based on it */
1475 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1476 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1477 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1478 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1479 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1480 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1481 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1482 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1483 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1484 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1485
1486 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1487 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1488 rinv13 = gmx_mm_invsqrt_ps(rsq13);
1489 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1490 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1491 rinv23 = gmx_mm_invsqrt_ps(rsq23);
1492 rinv31 = gmx_mm_invsqrt_ps(rsq31);
1493 rinv32 = gmx_mm_invsqrt_ps(rsq32);
1494 rinv33 = gmx_mm_invsqrt_ps(rsq33);
1495
1496 rinvsq00 = gmx_mm_inv_ps(rsq00);
1497 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1498 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1499 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1500 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1501 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1502 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1503 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1504 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1505 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1506
1507 fjx0 = _mm_setzero_ps();
1508 fjy0 = _mm_setzero_ps();
1509 fjz0 = _mm_setzero_ps();
1510 fjx1 = _mm_setzero_ps();
1511 fjy1 = _mm_setzero_ps();
1512 fjz1 = _mm_setzero_ps();
1513 fjx2 = _mm_setzero_ps();
1514 fjy2 = _mm_setzero_ps();
1515 fjz2 = _mm_setzero_ps();
1516 fjx3 = _mm_setzero_ps();
1517 fjy3 = _mm_setzero_ps();
1518 fjz3 = _mm_setzero_ps();
1519
1520 /**************************
1521 * CALCULATE INTERACTIONS *
1522 **************************/
1523
1524 if (gmx_mm_any_lt(rsq00,rcutoff2))
1525 {
1526
1527 /* LENNARD-JONES DISPERSION/REPULSION */
1528
1529 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1530 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1531
1532 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1533
1534 fscal = fvdw;
1535
1536 fscal = _mm_and_ps(fscal,cutoff_mask);
1537
1538 /* Calculate temporary vectorial force */
1539 tx = _mm_mul_ps(fscal,dx00);
1540 ty = _mm_mul_ps(fscal,dy00);
1541 tz = _mm_mul_ps(fscal,dz00);
1542
1543 /* Update vectorial force */
1544 fix0 = _mm_add_ps(fix0,tx);
1545 fiy0 = _mm_add_ps(fiy0,ty);
1546 fiz0 = _mm_add_ps(fiz0,tz);
1547
1548 fjx0 = _mm_add_ps(fjx0,tx);
1549 fjy0 = _mm_add_ps(fjy0,ty);
1550 fjz0 = _mm_add_ps(fjz0,tz);
1551
1552 }
1553
1554 /**************************
1555 * CALCULATE INTERACTIONS *
1556 **************************/
1557
1558 if (gmx_mm_any_lt(rsq11,rcutoff2))
1559 {
1560
1561 /* REACTION-FIELD ELECTROSTATICS */
1562 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1563
1564 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
1565
1566 fscal = felec;
1567
1568 fscal = _mm_and_ps(fscal,cutoff_mask);
1569
1570 /* Calculate temporary vectorial force */
1571 tx = _mm_mul_ps(fscal,dx11);
1572 ty = _mm_mul_ps(fscal,dy11);
1573 tz = _mm_mul_ps(fscal,dz11);
1574
1575 /* Update vectorial force */
1576 fix1 = _mm_add_ps(fix1,tx);
1577 fiy1 = _mm_add_ps(fiy1,ty);
1578 fiz1 = _mm_add_ps(fiz1,tz);
1579
1580 fjx1 = _mm_add_ps(fjx1,tx);
1581 fjy1 = _mm_add_ps(fjy1,ty);
1582 fjz1 = _mm_add_ps(fjz1,tz);
1583
1584 }
1585
1586 /**************************
1587 * CALCULATE INTERACTIONS *
1588 **************************/
1589
1590 if (gmx_mm_any_lt(rsq12,rcutoff2))
1591 {
1592
1593 /* REACTION-FIELD ELECTROSTATICS */
1594 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1595
1596 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
1597
1598 fscal = felec;
1599
1600 fscal = _mm_and_ps(fscal,cutoff_mask);
1601
1602 /* Calculate temporary vectorial force */
1603 tx = _mm_mul_ps(fscal,dx12);
1604 ty = _mm_mul_ps(fscal,dy12);
1605 tz = _mm_mul_ps(fscal,dz12);
1606
1607 /* Update vectorial force */
1608 fix1 = _mm_add_ps(fix1,tx);
1609 fiy1 = _mm_add_ps(fiy1,ty);
1610 fiz1 = _mm_add_ps(fiz1,tz);
1611
1612 fjx2 = _mm_add_ps(fjx2,tx);
1613 fjy2 = _mm_add_ps(fjy2,ty);
1614 fjz2 = _mm_add_ps(fjz2,tz);
1615
1616 }
1617
1618 /**************************
1619 * CALCULATE INTERACTIONS *
1620 **************************/
1621
1622 if (gmx_mm_any_lt(rsq13,rcutoff2))
1623 {
1624
1625 /* REACTION-FIELD ELECTROSTATICS */
1626 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1627
1628 cutoff_mask = _mm_cmplt_ps(rsq13,rcutoff2);
1629
1630 fscal = felec;
1631
1632 fscal = _mm_and_ps(fscal,cutoff_mask);
1633
1634 /* Calculate temporary vectorial force */
1635 tx = _mm_mul_ps(fscal,dx13);
1636 ty = _mm_mul_ps(fscal,dy13);
1637 tz = _mm_mul_ps(fscal,dz13);
1638
1639 /* Update vectorial force */
1640 fix1 = _mm_add_ps(fix1,tx);
1641 fiy1 = _mm_add_ps(fiy1,ty);
1642 fiz1 = _mm_add_ps(fiz1,tz);
1643
1644 fjx3 = _mm_add_ps(fjx3,tx);
1645 fjy3 = _mm_add_ps(fjy3,ty);
1646 fjz3 = _mm_add_ps(fjz3,tz);
1647
1648 }
1649
1650 /**************************
1651 * CALCULATE INTERACTIONS *
1652 **************************/
1653
1654 if (gmx_mm_any_lt(rsq21,rcutoff2))
1655 {
1656
1657 /* REACTION-FIELD ELECTROSTATICS */
1658 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1659
1660 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
1661
1662 fscal = felec;
1663
1664 fscal = _mm_and_ps(fscal,cutoff_mask);
1665
1666 /* Calculate temporary vectorial force */
1667 tx = _mm_mul_ps(fscal,dx21);
1668 ty = _mm_mul_ps(fscal,dy21);
1669 tz = _mm_mul_ps(fscal,dz21);
1670
1671 /* Update vectorial force */
1672 fix2 = _mm_add_ps(fix2,tx);
1673 fiy2 = _mm_add_ps(fiy2,ty);
1674 fiz2 = _mm_add_ps(fiz2,tz);
1675
1676 fjx1 = _mm_add_ps(fjx1,tx);
1677 fjy1 = _mm_add_ps(fjy1,ty);
1678 fjz1 = _mm_add_ps(fjz1,tz);
1679
1680 }
1681
1682 /**************************
1683 * CALCULATE INTERACTIONS *
1684 **************************/
1685
1686 if (gmx_mm_any_lt(rsq22,rcutoff2))
1687 {
1688
1689 /* REACTION-FIELD ELECTROSTATICS */
1690 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1691
1692 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
1693
1694 fscal = felec;
1695
1696 fscal = _mm_and_ps(fscal,cutoff_mask);
1697
1698 /* Calculate temporary vectorial force */
1699 tx = _mm_mul_ps(fscal,dx22);
1700 ty = _mm_mul_ps(fscal,dy22);
1701 tz = _mm_mul_ps(fscal,dz22);
1702
1703 /* Update vectorial force */
1704 fix2 = _mm_add_ps(fix2,tx);
1705 fiy2 = _mm_add_ps(fiy2,ty);
1706 fiz2 = _mm_add_ps(fiz2,tz);
1707
1708 fjx2 = _mm_add_ps(fjx2,tx);
1709 fjy2 = _mm_add_ps(fjy2,ty);
1710 fjz2 = _mm_add_ps(fjz2,tz);
1711
1712 }
1713
1714 /**************************
1715 * CALCULATE INTERACTIONS *
1716 **************************/
1717
1718 if (gmx_mm_any_lt(rsq23,rcutoff2))
1719 {
1720
1721 /* REACTION-FIELD ELECTROSTATICS */
1722 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1723
1724 cutoff_mask = _mm_cmplt_ps(rsq23,rcutoff2);
1725
1726 fscal = felec;
1727
1728 fscal = _mm_and_ps(fscal,cutoff_mask);
1729
1730 /* Calculate temporary vectorial force */
1731 tx = _mm_mul_ps(fscal,dx23);
1732 ty = _mm_mul_ps(fscal,dy23);
1733 tz = _mm_mul_ps(fscal,dz23);
1734
1735 /* Update vectorial force */
1736 fix2 = _mm_add_ps(fix2,tx);
1737 fiy2 = _mm_add_ps(fiy2,ty);
1738 fiz2 = _mm_add_ps(fiz2,tz);
1739
1740 fjx3 = _mm_add_ps(fjx3,tx);
1741 fjy3 = _mm_add_ps(fjy3,ty);
1742 fjz3 = _mm_add_ps(fjz3,tz);
1743
1744 }
1745
1746 /**************************
1747 * CALCULATE INTERACTIONS *
1748 **************************/
1749
1750 if (gmx_mm_any_lt(rsq31,rcutoff2))
1751 {
1752
1753 /* REACTION-FIELD ELECTROSTATICS */
1754 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1755
1756 cutoff_mask = _mm_cmplt_ps(rsq31,rcutoff2);
1757
1758 fscal = felec;
1759
1760 fscal = _mm_and_ps(fscal,cutoff_mask);
1761
1762 /* Calculate temporary vectorial force */
1763 tx = _mm_mul_ps(fscal,dx31);
1764 ty = _mm_mul_ps(fscal,dy31);
1765 tz = _mm_mul_ps(fscal,dz31);
1766
1767 /* Update vectorial force */
1768 fix3 = _mm_add_ps(fix3,tx);
1769 fiy3 = _mm_add_ps(fiy3,ty);
1770 fiz3 = _mm_add_ps(fiz3,tz);
1771
1772 fjx1 = _mm_add_ps(fjx1,tx);
1773 fjy1 = _mm_add_ps(fjy1,ty);
1774 fjz1 = _mm_add_ps(fjz1,tz);
1775
1776 }
1777
1778 /**************************
1779 * CALCULATE INTERACTIONS *
1780 **************************/
1781
1782 if (gmx_mm_any_lt(rsq32,rcutoff2))
1783 {
1784
1785 /* REACTION-FIELD ELECTROSTATICS */
1786 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1787
1788 cutoff_mask = _mm_cmplt_ps(rsq32,rcutoff2);
1789
1790 fscal = felec;
1791
1792 fscal = _mm_and_ps(fscal,cutoff_mask);
1793
1794 /* Calculate temporary vectorial force */
1795 tx = _mm_mul_ps(fscal,dx32);
1796 ty = _mm_mul_ps(fscal,dy32);
1797 tz = _mm_mul_ps(fscal,dz32);
1798
1799 /* Update vectorial force */
1800 fix3 = _mm_add_ps(fix3,tx);
1801 fiy3 = _mm_add_ps(fiy3,ty);
1802 fiz3 = _mm_add_ps(fiz3,tz);
1803
1804 fjx2 = _mm_add_ps(fjx2,tx);
1805 fjy2 = _mm_add_ps(fjy2,ty);
1806 fjz2 = _mm_add_ps(fjz2,tz);
1807
1808 }
1809
1810 /**************************
1811 * CALCULATE INTERACTIONS *
1812 **************************/
1813
1814 if (gmx_mm_any_lt(rsq33,rcutoff2))
1815 {
1816
1817 /* REACTION-FIELD ELECTROSTATICS */
1818 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1819
1820 cutoff_mask = _mm_cmplt_ps(rsq33,rcutoff2);
1821
1822 fscal = felec;
1823
1824 fscal = _mm_and_ps(fscal,cutoff_mask);
1825
1826 /* Calculate temporary vectorial force */
1827 tx = _mm_mul_ps(fscal,dx33);
1828 ty = _mm_mul_ps(fscal,dy33);
1829 tz = _mm_mul_ps(fscal,dz33);
1830
1831 /* Update vectorial force */
1832 fix3 = _mm_add_ps(fix3,tx);
1833 fiy3 = _mm_add_ps(fiy3,ty);
1834 fiz3 = _mm_add_ps(fiz3,tz);
1835
1836 fjx3 = _mm_add_ps(fjx3,tx);
1837 fjy3 = _mm_add_ps(fjy3,ty);
1838 fjz3 = _mm_add_ps(fjz3,tz);
1839
1840 }
1841
1842 fjptrA = f+j_coord_offsetA;
1843 fjptrB = f+j_coord_offsetB;
1844 fjptrC = f+j_coord_offsetC;
1845 fjptrD = f+j_coord_offsetD;
1846
1847 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1848 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1849 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1850
1851 /* Inner loop uses 303 flops */
1852 }
1853
1854 if(jidx<j_index_end)
1855 {
1856
1857 /* Get j neighbor index, and coordinate index */
1858 jnrlistA = jjnr[jidx];
1859 jnrlistB = jjnr[jidx+1];
1860 jnrlistC = jjnr[jidx+2];
1861 jnrlistD = jjnr[jidx+3];
1862 /* Sign of each element will be negative for non-real atoms.
1863 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1864 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1865 */
1866 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1867 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1868 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1869 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1870 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1871 j_coord_offsetA = DIM*jnrA;
1872 j_coord_offsetB = DIM*jnrB;
1873 j_coord_offsetC = DIM*jnrC;
1874 j_coord_offsetD = DIM*jnrD;
1875
1876 /* load j atom coordinates */
1877 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1878 x+j_coord_offsetC,x+j_coord_offsetD,
1879 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1880 &jy2,&jz2,&jx3,&jy3,&jz3);
1881
1882 /* Calculate displacement vector */
1883 dx00 = _mm_sub_ps(ix0,jx0);
1884 dy00 = _mm_sub_ps(iy0,jy0);
1885 dz00 = _mm_sub_ps(iz0,jz0);
1886 dx11 = _mm_sub_ps(ix1,jx1);
1887 dy11 = _mm_sub_ps(iy1,jy1);
1888 dz11 = _mm_sub_ps(iz1,jz1);
1889 dx12 = _mm_sub_ps(ix1,jx2);
1890 dy12 = _mm_sub_ps(iy1,jy2);
1891 dz12 = _mm_sub_ps(iz1,jz2);
1892 dx13 = _mm_sub_ps(ix1,jx3);
1893 dy13 = _mm_sub_ps(iy1,jy3);
1894 dz13 = _mm_sub_ps(iz1,jz3);
1895 dx21 = _mm_sub_ps(ix2,jx1);
1896 dy21 = _mm_sub_ps(iy2,jy1);
1897 dz21 = _mm_sub_ps(iz2,jz1);
1898 dx22 = _mm_sub_ps(ix2,jx2);
1899 dy22 = _mm_sub_ps(iy2,jy2);
1900 dz22 = _mm_sub_ps(iz2,jz2);
1901 dx23 = _mm_sub_ps(ix2,jx3);
1902 dy23 = _mm_sub_ps(iy2,jy3);
1903 dz23 = _mm_sub_ps(iz2,jz3);
1904 dx31 = _mm_sub_ps(ix3,jx1);
1905 dy31 = _mm_sub_ps(iy3,jy1);
1906 dz31 = _mm_sub_ps(iz3,jz1);
1907 dx32 = _mm_sub_ps(ix3,jx2);
1908 dy32 = _mm_sub_ps(iy3,jy2);
1909 dz32 = _mm_sub_ps(iz3,jz2);
1910 dx33 = _mm_sub_ps(ix3,jx3);
1911 dy33 = _mm_sub_ps(iy3,jy3);
1912 dz33 = _mm_sub_ps(iz3,jz3);
1913
1914 /* Calculate squared distance and things based on it */
1915 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1916 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1917 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1918 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1919 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1920 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1921 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1922 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1923 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1924 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1925
1926 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1927 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1928 rinv13 = gmx_mm_invsqrt_ps(rsq13);
1929 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1930 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1931 rinv23 = gmx_mm_invsqrt_ps(rsq23);
1932 rinv31 = gmx_mm_invsqrt_ps(rsq31);
1933 rinv32 = gmx_mm_invsqrt_ps(rsq32);
1934 rinv33 = gmx_mm_invsqrt_ps(rsq33);
1935
1936 rinvsq00 = gmx_mm_inv_ps(rsq00);
1937 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1938 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1939 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1940 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1941 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1942 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1943 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1944 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1945 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1946
1947 fjx0 = _mm_setzero_ps();
1948 fjy0 = _mm_setzero_ps();
1949 fjz0 = _mm_setzero_ps();
1950 fjx1 = _mm_setzero_ps();
1951 fjy1 = _mm_setzero_ps();
1952 fjz1 = _mm_setzero_ps();
1953 fjx2 = _mm_setzero_ps();
1954 fjy2 = _mm_setzero_ps();
1955 fjz2 = _mm_setzero_ps();
1956 fjx3 = _mm_setzero_ps();
1957 fjy3 = _mm_setzero_ps();
1958 fjz3 = _mm_setzero_ps();
1959
1960 /**************************
1961 * CALCULATE INTERACTIONS *
1962 **************************/
1963
1964 if (gmx_mm_any_lt(rsq00,rcutoff2))
1965 {
1966
1967 /* LENNARD-JONES DISPERSION/REPULSION */
1968
1969 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1970 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1971
1972 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1973
1974 fscal = fvdw;
1975
1976 fscal = _mm_and_ps(fscal,cutoff_mask);
1977
1978 fscal = _mm_andnot_ps(dummy_mask,fscal);
1979
1980 /* Calculate temporary vectorial force */
1981 tx = _mm_mul_ps(fscal,dx00);
1982 ty = _mm_mul_ps(fscal,dy00);
1983 tz = _mm_mul_ps(fscal,dz00);
1984
1985 /* Update vectorial force */
1986 fix0 = _mm_add_ps(fix0,tx);
1987 fiy0 = _mm_add_ps(fiy0,ty);
1988 fiz0 = _mm_add_ps(fiz0,tz);
1989
1990 fjx0 = _mm_add_ps(fjx0,tx);
1991 fjy0 = _mm_add_ps(fjy0,ty);
1992 fjz0 = _mm_add_ps(fjz0,tz);
1993
1994 }
1995
1996 /**************************
1997 * CALCULATE INTERACTIONS *
1998 **************************/
1999
2000 if (gmx_mm_any_lt(rsq11,rcutoff2))
2001 {
2002
2003 /* REACTION-FIELD ELECTROSTATICS */
2004 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
2005
2006 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
2007
2008 fscal = felec;
2009
2010 fscal = _mm_and_ps(fscal,cutoff_mask);
2011
2012 fscal = _mm_andnot_ps(dummy_mask,fscal);
2013
2014 /* Calculate temporary vectorial force */
2015 tx = _mm_mul_ps(fscal,dx11);
2016 ty = _mm_mul_ps(fscal,dy11);
2017 tz = _mm_mul_ps(fscal,dz11);
2018
2019 /* Update vectorial force */
2020 fix1 = _mm_add_ps(fix1,tx);
2021 fiy1 = _mm_add_ps(fiy1,ty);
2022 fiz1 = _mm_add_ps(fiz1,tz);
2023
2024 fjx1 = _mm_add_ps(fjx1,tx);
2025 fjy1 = _mm_add_ps(fjy1,ty);
2026 fjz1 = _mm_add_ps(fjz1,tz);
2027
2028 }
2029
2030 /**************************
2031 * CALCULATE INTERACTIONS *
2032 **************************/
2033
2034 if (gmx_mm_any_lt(rsq12,rcutoff2))
2035 {
2036
2037 /* REACTION-FIELD ELECTROSTATICS */
2038 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
2039
2040 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
2041
2042 fscal = felec;
2043
2044 fscal = _mm_and_ps(fscal,cutoff_mask);
2045
2046 fscal = _mm_andnot_ps(dummy_mask,fscal);
2047
2048 /* Calculate temporary vectorial force */
2049 tx = _mm_mul_ps(fscal,dx12);
2050 ty = _mm_mul_ps(fscal,dy12);
2051 tz = _mm_mul_ps(fscal,dz12);
2052
2053 /* Update vectorial force */
2054 fix1 = _mm_add_ps(fix1,tx);
2055 fiy1 = _mm_add_ps(fiy1,ty);
2056 fiz1 = _mm_add_ps(fiz1,tz);
2057
2058 fjx2 = _mm_add_ps(fjx2,tx);
2059 fjy2 = _mm_add_ps(fjy2,ty);
2060 fjz2 = _mm_add_ps(fjz2,tz);
2061
2062 }
2063
2064 /**************************
2065 * CALCULATE INTERACTIONS *
2066 **************************/
2067
2068 if (gmx_mm_any_lt(rsq13,rcutoff2))
2069 {
2070
2071 /* REACTION-FIELD ELECTROSTATICS */
2072 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
2073
2074 cutoff_mask = _mm_cmplt_ps(rsq13,rcutoff2);
2075
2076 fscal = felec;
2077
2078 fscal = _mm_and_ps(fscal,cutoff_mask);
2079
2080 fscal = _mm_andnot_ps(dummy_mask,fscal);
2081
2082 /* Calculate temporary vectorial force */
2083 tx = _mm_mul_ps(fscal,dx13);
2084 ty = _mm_mul_ps(fscal,dy13);
2085 tz = _mm_mul_ps(fscal,dz13);
2086
2087 /* Update vectorial force */
2088 fix1 = _mm_add_ps(fix1,tx);
2089 fiy1 = _mm_add_ps(fiy1,ty);
2090 fiz1 = _mm_add_ps(fiz1,tz);
2091
2092 fjx3 = _mm_add_ps(fjx3,tx);
2093 fjy3 = _mm_add_ps(fjy3,ty);
2094 fjz3 = _mm_add_ps(fjz3,tz);
2095
2096 }
2097
2098 /**************************
2099 * CALCULATE INTERACTIONS *
2100 **************************/
2101
2102 if (gmx_mm_any_lt(rsq21,rcutoff2))
2103 {
2104
2105 /* REACTION-FIELD ELECTROSTATICS */
2106 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
2107
2108 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
2109
2110 fscal = felec;
2111
2112 fscal = _mm_and_ps(fscal,cutoff_mask);
2113
2114 fscal = _mm_andnot_ps(dummy_mask,fscal);
2115
2116 /* Calculate temporary vectorial force */
2117 tx = _mm_mul_ps(fscal,dx21);
2118 ty = _mm_mul_ps(fscal,dy21);
2119 tz = _mm_mul_ps(fscal,dz21);
2120
2121 /* Update vectorial force */
2122 fix2 = _mm_add_ps(fix2,tx);
2123 fiy2 = _mm_add_ps(fiy2,ty);
2124 fiz2 = _mm_add_ps(fiz2,tz);
2125
2126 fjx1 = _mm_add_ps(fjx1,tx);
2127 fjy1 = _mm_add_ps(fjy1,ty);
2128 fjz1 = _mm_add_ps(fjz1,tz);
2129
2130 }
2131
2132 /**************************
2133 * CALCULATE INTERACTIONS *
2134 **************************/
2135
2136 if (gmx_mm_any_lt(rsq22,rcutoff2))
2137 {
2138
2139 /* REACTION-FIELD ELECTROSTATICS */
2140 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2141
2142 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
2143
2144 fscal = felec;
2145
2146 fscal = _mm_and_ps(fscal,cutoff_mask);
2147
2148 fscal = _mm_andnot_ps(dummy_mask,fscal);
2149
2150 /* Calculate temporary vectorial force */
2151 tx = _mm_mul_ps(fscal,dx22);
2152 ty = _mm_mul_ps(fscal,dy22);
2153 tz = _mm_mul_ps(fscal,dz22);
2154
2155 /* Update vectorial force */
2156 fix2 = _mm_add_ps(fix2,tx);
2157 fiy2 = _mm_add_ps(fiy2,ty);
2158 fiz2 = _mm_add_ps(fiz2,tz);
2159
2160 fjx2 = _mm_add_ps(fjx2,tx);
2161 fjy2 = _mm_add_ps(fjy2,ty);
2162 fjz2 = _mm_add_ps(fjz2,tz);
2163
2164 }
2165
2166 /**************************
2167 * CALCULATE INTERACTIONS *
2168 **************************/
2169
2170 if (gmx_mm_any_lt(rsq23,rcutoff2))
2171 {
2172
2173 /* REACTION-FIELD ELECTROSTATICS */
2174 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
2175
2176 cutoff_mask = _mm_cmplt_ps(rsq23,rcutoff2);
2177
2178 fscal = felec;
2179
2180 fscal = _mm_and_ps(fscal,cutoff_mask);
2181
2182 fscal = _mm_andnot_ps(dummy_mask,fscal);
2183
2184 /* Calculate temporary vectorial force */
2185 tx = _mm_mul_ps(fscal,dx23);
2186 ty = _mm_mul_ps(fscal,dy23);
2187 tz = _mm_mul_ps(fscal,dz23);
2188
2189 /* Update vectorial force */
2190 fix2 = _mm_add_ps(fix2,tx);
2191 fiy2 = _mm_add_ps(fiy2,ty);
2192 fiz2 = _mm_add_ps(fiz2,tz);
2193
2194 fjx3 = _mm_add_ps(fjx3,tx);
2195 fjy3 = _mm_add_ps(fjy3,ty);
2196 fjz3 = _mm_add_ps(fjz3,tz);
2197
2198 }
2199
2200 /**************************
2201 * CALCULATE INTERACTIONS *
2202 **************************/
2203
2204 if (gmx_mm_any_lt(rsq31,rcutoff2))
2205 {
2206
2207 /* REACTION-FIELD ELECTROSTATICS */
2208 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
2209
2210 cutoff_mask = _mm_cmplt_ps(rsq31,rcutoff2);
2211
2212 fscal = felec;
2213
2214 fscal = _mm_and_ps(fscal,cutoff_mask);
2215
2216 fscal = _mm_andnot_ps(dummy_mask,fscal);
2217
2218 /* Calculate temporary vectorial force */
2219 tx = _mm_mul_ps(fscal,dx31);
2220 ty = _mm_mul_ps(fscal,dy31);
2221 tz = _mm_mul_ps(fscal,dz31);
2222
2223 /* Update vectorial force */
2224 fix3 = _mm_add_ps(fix3,tx);
2225 fiy3 = _mm_add_ps(fiy3,ty);
2226 fiz3 = _mm_add_ps(fiz3,tz);
2227
2228 fjx1 = _mm_add_ps(fjx1,tx);
2229 fjy1 = _mm_add_ps(fjy1,ty);
2230 fjz1 = _mm_add_ps(fjz1,tz);
2231
2232 }
2233
2234 /**************************
2235 * CALCULATE INTERACTIONS *
2236 **************************/
2237
2238 if (gmx_mm_any_lt(rsq32,rcutoff2))
2239 {
2240
2241 /* REACTION-FIELD ELECTROSTATICS */
2242 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
2243
2244 cutoff_mask = _mm_cmplt_ps(rsq32,rcutoff2);
2245
2246 fscal = felec;
2247
2248 fscal = _mm_and_ps(fscal,cutoff_mask);
2249
2250 fscal = _mm_andnot_ps(dummy_mask,fscal);
2251
2252 /* Calculate temporary vectorial force */
2253 tx = _mm_mul_ps(fscal,dx32);
2254 ty = _mm_mul_ps(fscal,dy32);
2255 tz = _mm_mul_ps(fscal,dz32);
2256
2257 /* Update vectorial force */
2258 fix3 = _mm_add_ps(fix3,tx);
2259 fiy3 = _mm_add_ps(fiy3,ty);
2260 fiz3 = _mm_add_ps(fiz3,tz);
2261
2262 fjx2 = _mm_add_ps(fjx2,tx);
2263 fjy2 = _mm_add_ps(fjy2,ty);
2264 fjz2 = _mm_add_ps(fjz2,tz);
2265
2266 }
2267
2268 /**************************
2269 * CALCULATE INTERACTIONS *
2270 **************************/
2271
2272 if (gmx_mm_any_lt(rsq33,rcutoff2))
2273 {
2274
2275 /* REACTION-FIELD ELECTROSTATICS */
2276 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2277
2278 cutoff_mask = _mm_cmplt_ps(rsq33,rcutoff2);
2279
2280 fscal = felec;
2281
2282 fscal = _mm_and_ps(fscal,cutoff_mask);
2283
2284 fscal = _mm_andnot_ps(dummy_mask,fscal);
2285
2286 /* Calculate temporary vectorial force */
2287 tx = _mm_mul_ps(fscal,dx33);
2288 ty = _mm_mul_ps(fscal,dy33);
2289 tz = _mm_mul_ps(fscal,dz33);
2290
2291 /* Update vectorial force */
2292 fix3 = _mm_add_ps(fix3,tx);
2293 fiy3 = _mm_add_ps(fiy3,ty);
2294 fiz3 = _mm_add_ps(fiz3,tz);
2295
2296 fjx3 = _mm_add_ps(fjx3,tx);
2297 fjy3 = _mm_add_ps(fjy3,ty);
2298 fjz3 = _mm_add_ps(fjz3,tz);
2299
2300 }
2301
2302 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2303 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2304 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2305 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2306
2307 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2308 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2309 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2310
2311 /* Inner loop uses 303 flops */
2312 }
2313
2314 /* End of innermost loop */
2315
2316 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2317 f+i_coord_offset,fshift+i_shift_offset);
2318
2319 /* Increment number of inner iterations */
2320 inneriter += j_index_end - j_index_start;
2321
2322 /* Outer loop uses 24 flops */
2323 }
2324
2325 /* Increment number of outer iterations */
2326 outeriter += nri;
2327
2328 /* Update outer/inner flops */
2329
2330 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
2331 }
The ultimate molecular dynamics simulation package