Use full path for legacyheaders
[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sparc64_hpc_ace_double / nb_kernel_ElecCoul_VdwNone_GeomW4W4_sparc64_hpc_ace_double.c
1 /*
2  * This file is part of the GROMACS molecular simulation package.
3  *
4  * Copyright (c) 2012,2013,2014, 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 sparc64_hpc_ace_double kernel generator.
37  */
38 #include "config.h"
39
40 #include <math.h>
41
42 #include "../nb_kernel.h"
43 #include "gromacs/legacyheaders/types/simple.h"
44 #include "gromacs/math/vec.h"
45 #include "gromacs/legacyheaders/nrnb.h"
46
47 #include "kernelutil_sparc64_hpc_ace_double.h"
48
49 /*
50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
51  * Electrostatics interaction: Coulomb
52  * VdW interaction:            None
53  * Geometry:                   Water4-Water4
54  * Calculate force/pot:        PotentialAndForce
55  */
56 void
57 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sparc64_hpc_ace_double
58                     (t_nblist                    * gmx_restrict       nlist,
59                      rvec                        * gmx_restrict          xx,
60                      rvec                        * gmx_restrict          ff,
61                      t_forcerec                  * gmx_restrict          fr,
62                      t_mdatoms                   * gmx_restrict     mdatoms,
63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
64                      t_nrnb                      * gmx_restrict        nrnb)
65 {
66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
67      * just 0 for non-waters.
68      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
69      * jnr indices corresponding to data put in the four positions in the SIMD register.
70      */
71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
73     int              jnrA,jnrB;
74     int              j_coord_offsetA,j_coord_offsetB;
75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
76     real             rcutoff_scalar;
77     real             *shiftvec,*fshift,*x,*f;
78     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
79     int              vdwioffset1;
80     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
81     int              vdwioffset2;
82     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
83     int              vdwioffset3;
84     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
85     int              vdwjidx1A,vdwjidx1B;
86     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
87     int              vdwjidx2A,vdwjidx2B;
88     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
89     int              vdwjidx3A,vdwjidx3B;
90     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
91     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
92     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
93     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
94     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
95     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
96     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
97     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
98     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
99     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
100     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
101     real             *charge;
102     _fjsp_v2r8       itab_tmp;
103     _fjsp_v2r8       dummy_mask,cutoff_mask;
104     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
105     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
106     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
107
108     x                = xx[0];
109     f                = ff[0];
110
111     nri              = nlist->nri;
112     iinr             = nlist->iinr;
113     jindex           = nlist->jindex;
114     jjnr             = nlist->jjnr;
115     shiftidx         = nlist->shift;
116     gid              = nlist->gid;
117     shiftvec         = fr->shift_vec[0];
118     fshift           = fr->fshift[0];
119     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
120     charge           = mdatoms->chargeA;
121
122     /* Setup water-specific parameters */
123     inr              = nlist->iinr[0];
124     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
125     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
126     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
127
128     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
129     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
130     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
131     qq11             = _fjsp_mul_v2r8(iq1,jq1);
132     qq12             = _fjsp_mul_v2r8(iq1,jq2);
133     qq13             = _fjsp_mul_v2r8(iq1,jq3);
134     qq21             = _fjsp_mul_v2r8(iq2,jq1);
135     qq22             = _fjsp_mul_v2r8(iq2,jq2);
136     qq23             = _fjsp_mul_v2r8(iq2,jq3);
137     qq31             = _fjsp_mul_v2r8(iq3,jq1);
138     qq32             = _fjsp_mul_v2r8(iq3,jq2);
139     qq33             = _fjsp_mul_v2r8(iq3,jq3);
140
141     /* Avoid stupid compiler warnings */
142     jnrA = jnrB = 0;
143     j_coord_offsetA = 0;
144     j_coord_offsetB = 0;
145
146     outeriter        = 0;
147     inneriter        = 0;
148
149     /* Start outer loop over neighborlists */
150     for(iidx=0; iidx<nri; iidx++)
151     {
152         /* Load shift vector for this list */
153         i_shift_offset   = DIM*shiftidx[iidx];
154
155         /* Load limits for loop over neighbors */
156         j_index_start    = jindex[iidx];
157         j_index_end      = jindex[iidx+1];
158
159         /* Get outer coordinate index */
160         inr              = iinr[iidx];
161         i_coord_offset   = DIM*inr;
162
163         /* Load i particle coords and add shift vector */
164         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
165                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
166
167         fix1             = _fjsp_setzero_v2r8();
168         fiy1             = _fjsp_setzero_v2r8();
169         fiz1             = _fjsp_setzero_v2r8();
170         fix2             = _fjsp_setzero_v2r8();
171         fiy2             = _fjsp_setzero_v2r8();
172         fiz2             = _fjsp_setzero_v2r8();
173         fix3             = _fjsp_setzero_v2r8();
174         fiy3             = _fjsp_setzero_v2r8();
175         fiz3             = _fjsp_setzero_v2r8();
176
177         /* Reset potential sums */
178         velecsum         = _fjsp_setzero_v2r8();
179
180         /* Start inner kernel loop */
181         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
182         {
183
184             /* Get j neighbor index, and coordinate index */
185             jnrA             = jjnr[jidx];
186             jnrB             = jjnr[jidx+1];
187             j_coord_offsetA  = DIM*jnrA;
188             j_coord_offsetB  = DIM*jnrB;
189
190             /* load j atom coordinates */
191             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
192                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
193
194             /* Calculate displacement vector */
195             dx11             = _fjsp_sub_v2r8(ix1,jx1);
196             dy11             = _fjsp_sub_v2r8(iy1,jy1);
197             dz11             = _fjsp_sub_v2r8(iz1,jz1);
198             dx12             = _fjsp_sub_v2r8(ix1,jx2);
199             dy12             = _fjsp_sub_v2r8(iy1,jy2);
200             dz12             = _fjsp_sub_v2r8(iz1,jz2);
201             dx13             = _fjsp_sub_v2r8(ix1,jx3);
202             dy13             = _fjsp_sub_v2r8(iy1,jy3);
203             dz13             = _fjsp_sub_v2r8(iz1,jz3);
204             dx21             = _fjsp_sub_v2r8(ix2,jx1);
205             dy21             = _fjsp_sub_v2r8(iy2,jy1);
206             dz21             = _fjsp_sub_v2r8(iz2,jz1);
207             dx22             = _fjsp_sub_v2r8(ix2,jx2);
208             dy22             = _fjsp_sub_v2r8(iy2,jy2);
209             dz22             = _fjsp_sub_v2r8(iz2,jz2);
210             dx23             = _fjsp_sub_v2r8(ix2,jx3);
211             dy23             = _fjsp_sub_v2r8(iy2,jy3);
212             dz23             = _fjsp_sub_v2r8(iz2,jz3);
213             dx31             = _fjsp_sub_v2r8(ix3,jx1);
214             dy31             = _fjsp_sub_v2r8(iy3,jy1);
215             dz31             = _fjsp_sub_v2r8(iz3,jz1);
216             dx32             = _fjsp_sub_v2r8(ix3,jx2);
217             dy32             = _fjsp_sub_v2r8(iy3,jy2);
218             dz32             = _fjsp_sub_v2r8(iz3,jz2);
219             dx33             = _fjsp_sub_v2r8(ix3,jx3);
220             dy33             = _fjsp_sub_v2r8(iy3,jy3);
221             dz33             = _fjsp_sub_v2r8(iz3,jz3);
222
223             /* Calculate squared distance and things based on it */
224             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
225             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
226             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
227             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
228             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
229             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
230             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
231             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
232             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
233
234             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
235             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
236             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
237             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
238             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
239             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
240             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
241             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
242             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
243
244             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
245             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
246             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
247             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
248             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
249             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
250             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
251             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
252             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
253
254             fjx1             = _fjsp_setzero_v2r8();
255             fjy1             = _fjsp_setzero_v2r8();
256             fjz1             = _fjsp_setzero_v2r8();
257             fjx2             = _fjsp_setzero_v2r8();
258             fjy2             = _fjsp_setzero_v2r8();
259             fjz2             = _fjsp_setzero_v2r8();
260             fjx3             = _fjsp_setzero_v2r8();
261             fjy3             = _fjsp_setzero_v2r8();
262             fjz3             = _fjsp_setzero_v2r8();
263
264             /**************************
265              * CALCULATE INTERACTIONS *
266              **************************/
267
268             /* COULOMB ELECTROSTATICS */
269             velec            = _fjsp_mul_v2r8(qq11,rinv11);
270             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
271
272             /* Update potential sum for this i atom from the interaction with this j atom. */
273             velecsum         = _fjsp_add_v2r8(velecsum,velec);
274
275             fscal            = felec;
276
277             /* Update vectorial force */
278             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
279             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
280             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
281             
282             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
283             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
284             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
285
286             /**************************
287              * CALCULATE INTERACTIONS *
288              **************************/
289
290             /* COULOMB ELECTROSTATICS */
291             velec            = _fjsp_mul_v2r8(qq12,rinv12);
292             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
293
294             /* Update potential sum for this i atom from the interaction with this j atom. */
295             velecsum         = _fjsp_add_v2r8(velecsum,velec);
296
297             fscal            = felec;
298
299             /* Update vectorial force */
300             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
301             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
302             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
303             
304             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
305             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
306             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
307
308             /**************************
309              * CALCULATE INTERACTIONS *
310              **************************/
311
312             /* COULOMB ELECTROSTATICS */
313             velec            = _fjsp_mul_v2r8(qq13,rinv13);
314             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
315
316             /* Update potential sum for this i atom from the interaction with this j atom. */
317             velecsum         = _fjsp_add_v2r8(velecsum,velec);
318
319             fscal            = felec;
320
321             /* Update vectorial force */
322             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
323             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
324             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
325             
326             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
327             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
328             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
329
330             /**************************
331              * CALCULATE INTERACTIONS *
332              **************************/
333
334             /* COULOMB ELECTROSTATICS */
335             velec            = _fjsp_mul_v2r8(qq21,rinv21);
336             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
337
338             /* Update potential sum for this i atom from the interaction with this j atom. */
339             velecsum         = _fjsp_add_v2r8(velecsum,velec);
340
341             fscal            = felec;
342
343             /* Update vectorial force */
344             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
345             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
346             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
347             
348             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
349             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
350             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
351
352             /**************************
353              * CALCULATE INTERACTIONS *
354              **************************/
355
356             /* COULOMB ELECTROSTATICS */
357             velec            = _fjsp_mul_v2r8(qq22,rinv22);
358             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
359
360             /* Update potential sum for this i atom from the interaction with this j atom. */
361             velecsum         = _fjsp_add_v2r8(velecsum,velec);
362
363             fscal            = felec;
364
365             /* Update vectorial force */
366             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
367             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
368             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
369             
370             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
371             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
372             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
373
374             /**************************
375              * CALCULATE INTERACTIONS *
376              **************************/
377
378             /* COULOMB ELECTROSTATICS */
379             velec            = _fjsp_mul_v2r8(qq23,rinv23);
380             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
381
382             /* Update potential sum for this i atom from the interaction with this j atom. */
383             velecsum         = _fjsp_add_v2r8(velecsum,velec);
384
385             fscal            = felec;
386
387             /* Update vectorial force */
388             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
389             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
390             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
391             
392             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
393             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
394             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
395
396             /**************************
397              * CALCULATE INTERACTIONS *
398              **************************/
399
400             /* COULOMB ELECTROSTATICS */
401             velec            = _fjsp_mul_v2r8(qq31,rinv31);
402             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
403
404             /* Update potential sum for this i atom from the interaction with this j atom. */
405             velecsum         = _fjsp_add_v2r8(velecsum,velec);
406
407             fscal            = felec;
408
409             /* Update vectorial force */
410             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
411             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
412             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
413             
414             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
415             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
416             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
417
418             /**************************
419              * CALCULATE INTERACTIONS *
420              **************************/
421
422             /* COULOMB ELECTROSTATICS */
423             velec            = _fjsp_mul_v2r8(qq32,rinv32);
424             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
425
426             /* Update potential sum for this i atom from the interaction with this j atom. */
427             velecsum         = _fjsp_add_v2r8(velecsum,velec);
428
429             fscal            = felec;
430
431             /* Update vectorial force */
432             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
433             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
434             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
435             
436             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
437             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
438             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
439
440             /**************************
441              * CALCULATE INTERACTIONS *
442              **************************/
443
444             /* COULOMB ELECTROSTATICS */
445             velec            = _fjsp_mul_v2r8(qq33,rinv33);
446             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
447
448             /* Update potential sum for this i atom from the interaction with this j atom. */
449             velecsum         = _fjsp_add_v2r8(velecsum,velec);
450
451             fscal            = felec;
452
453             /* Update vectorial force */
454             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
455             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
456             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
457             
458             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
459             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
460             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
461
462             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
463
464             /* Inner loop uses 279 flops */
465         }
466
467         if(jidx<j_index_end)
468         {
469
470             jnrA             = jjnr[jidx];
471             j_coord_offsetA  = DIM*jnrA;
472
473             /* load j atom coordinates */
474             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
475                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
476
477             /* Calculate displacement vector */
478             dx11             = _fjsp_sub_v2r8(ix1,jx1);
479             dy11             = _fjsp_sub_v2r8(iy1,jy1);
480             dz11             = _fjsp_sub_v2r8(iz1,jz1);
481             dx12             = _fjsp_sub_v2r8(ix1,jx2);
482             dy12             = _fjsp_sub_v2r8(iy1,jy2);
483             dz12             = _fjsp_sub_v2r8(iz1,jz2);
484             dx13             = _fjsp_sub_v2r8(ix1,jx3);
485             dy13             = _fjsp_sub_v2r8(iy1,jy3);
486             dz13             = _fjsp_sub_v2r8(iz1,jz3);
487             dx21             = _fjsp_sub_v2r8(ix2,jx1);
488             dy21             = _fjsp_sub_v2r8(iy2,jy1);
489             dz21             = _fjsp_sub_v2r8(iz2,jz1);
490             dx22             = _fjsp_sub_v2r8(ix2,jx2);
491             dy22             = _fjsp_sub_v2r8(iy2,jy2);
492             dz22             = _fjsp_sub_v2r8(iz2,jz2);
493             dx23             = _fjsp_sub_v2r8(ix2,jx3);
494             dy23             = _fjsp_sub_v2r8(iy2,jy3);
495             dz23             = _fjsp_sub_v2r8(iz2,jz3);
496             dx31             = _fjsp_sub_v2r8(ix3,jx1);
497             dy31             = _fjsp_sub_v2r8(iy3,jy1);
498             dz31             = _fjsp_sub_v2r8(iz3,jz1);
499             dx32             = _fjsp_sub_v2r8(ix3,jx2);
500             dy32             = _fjsp_sub_v2r8(iy3,jy2);
501             dz32             = _fjsp_sub_v2r8(iz3,jz2);
502             dx33             = _fjsp_sub_v2r8(ix3,jx3);
503             dy33             = _fjsp_sub_v2r8(iy3,jy3);
504             dz33             = _fjsp_sub_v2r8(iz3,jz3);
505
506             /* Calculate squared distance and things based on it */
507             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
508             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
509             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
510             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
511             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
512             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
513             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
514             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
515             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
516
517             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
518             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
519             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
520             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
521             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
522             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
523             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
524             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
525             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
526
527             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
528             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
529             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
530             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
531             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
532             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
533             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
534             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
535             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
536
537             fjx1             = _fjsp_setzero_v2r8();
538             fjy1             = _fjsp_setzero_v2r8();
539             fjz1             = _fjsp_setzero_v2r8();
540             fjx2             = _fjsp_setzero_v2r8();
541             fjy2             = _fjsp_setzero_v2r8();
542             fjz2             = _fjsp_setzero_v2r8();
543             fjx3             = _fjsp_setzero_v2r8();
544             fjy3             = _fjsp_setzero_v2r8();
545             fjz3             = _fjsp_setzero_v2r8();
546
547             /**************************
548              * CALCULATE INTERACTIONS *
549              **************************/
550
551             /* COULOMB ELECTROSTATICS */
552             velec            = _fjsp_mul_v2r8(qq11,rinv11);
553             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
554
555             /* Update potential sum for this i atom from the interaction with this j atom. */
556             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
557             velecsum         = _fjsp_add_v2r8(velecsum,velec);
558
559             fscal            = felec;
560
561             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
562
563             /* Update vectorial force */
564             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
565             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
566             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
567             
568             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
569             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
570             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
571
572             /**************************
573              * CALCULATE INTERACTIONS *
574              **************************/
575
576             /* COULOMB ELECTROSTATICS */
577             velec            = _fjsp_mul_v2r8(qq12,rinv12);
578             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
579
580             /* Update potential sum for this i atom from the interaction with this j atom. */
581             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
582             velecsum         = _fjsp_add_v2r8(velecsum,velec);
583
584             fscal            = felec;
585
586             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
587
588             /* Update vectorial force */
589             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
590             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
591             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
592             
593             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
594             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
595             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
596
597             /**************************
598              * CALCULATE INTERACTIONS *
599              **************************/
600
601             /* COULOMB ELECTROSTATICS */
602             velec            = _fjsp_mul_v2r8(qq13,rinv13);
603             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
604
605             /* Update potential sum for this i atom from the interaction with this j atom. */
606             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
607             velecsum         = _fjsp_add_v2r8(velecsum,velec);
608
609             fscal            = felec;
610
611             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
612
613             /* Update vectorial force */
614             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
615             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
616             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
617             
618             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
619             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
620             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
621
622             /**************************
623              * CALCULATE INTERACTIONS *
624              **************************/
625
626             /* COULOMB ELECTROSTATICS */
627             velec            = _fjsp_mul_v2r8(qq21,rinv21);
628             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
629
630             /* Update potential sum for this i atom from the interaction with this j atom. */
631             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
632             velecsum         = _fjsp_add_v2r8(velecsum,velec);
633
634             fscal            = felec;
635
636             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
637
638             /* Update vectorial force */
639             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
640             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
641             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
642             
643             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
644             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
645             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
646
647             /**************************
648              * CALCULATE INTERACTIONS *
649              **************************/
650
651             /* COULOMB ELECTROSTATICS */
652             velec            = _fjsp_mul_v2r8(qq22,rinv22);
653             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
654
655             /* Update potential sum for this i atom from the interaction with this j atom. */
656             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
657             velecsum         = _fjsp_add_v2r8(velecsum,velec);
658
659             fscal            = felec;
660
661             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
662
663             /* Update vectorial force */
664             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
665             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
666             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
667             
668             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
669             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
670             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
671
672             /**************************
673              * CALCULATE INTERACTIONS *
674              **************************/
675
676             /* COULOMB ELECTROSTATICS */
677             velec            = _fjsp_mul_v2r8(qq23,rinv23);
678             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
679
680             /* Update potential sum for this i atom from the interaction with this j atom. */
681             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
682             velecsum         = _fjsp_add_v2r8(velecsum,velec);
683
684             fscal            = felec;
685
686             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
687
688             /* Update vectorial force */
689             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
690             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
691             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
692             
693             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
694             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
695             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
696
697             /**************************
698              * CALCULATE INTERACTIONS *
699              **************************/
700
701             /* COULOMB ELECTROSTATICS */
702             velec            = _fjsp_mul_v2r8(qq31,rinv31);
703             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
704
705             /* Update potential sum for this i atom from the interaction with this j atom. */
706             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
707             velecsum         = _fjsp_add_v2r8(velecsum,velec);
708
709             fscal            = felec;
710
711             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
712
713             /* Update vectorial force */
714             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
715             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
716             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
717             
718             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
719             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
720             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
721
722             /**************************
723              * CALCULATE INTERACTIONS *
724              **************************/
725
726             /* COULOMB ELECTROSTATICS */
727             velec            = _fjsp_mul_v2r8(qq32,rinv32);
728             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
729
730             /* Update potential sum for this i atom from the interaction with this j atom. */
731             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
732             velecsum         = _fjsp_add_v2r8(velecsum,velec);
733
734             fscal            = felec;
735
736             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
737
738             /* Update vectorial force */
739             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
740             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
741             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
742             
743             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
744             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
745             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
746
747             /**************************
748              * CALCULATE INTERACTIONS *
749              **************************/
750
751             /* COULOMB ELECTROSTATICS */
752             velec            = _fjsp_mul_v2r8(qq33,rinv33);
753             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
754
755             /* Update potential sum for this i atom from the interaction with this j atom. */
756             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
757             velecsum         = _fjsp_add_v2r8(velecsum,velec);
758
759             fscal            = felec;
760
761             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
762
763             /* Update vectorial force */
764             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
765             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
766             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
767             
768             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
769             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
770             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
771
772             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
773
774             /* Inner loop uses 279 flops */
775         }
776
777         /* End of innermost loop */
778
779         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
780                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
781
782         ggid                        = gid[iidx];
783         /* Update potential energies */
784         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
785
786         /* Increment number of inner iterations */
787         inneriter                  += j_index_end - j_index_start;
788
789         /* Outer loop uses 19 flops */
790     }
791
792     /* Increment number of outer iterations */
793     outeriter        += nri;
794
795     /* Update outer/inner flops */
796
797     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*279);
798 }
799 /*
800  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
801  * Electrostatics interaction: Coulomb
802  * VdW interaction:            None
803  * Geometry:                   Water4-Water4
804  * Calculate force/pot:        Force
805  */
806 void
807 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sparc64_hpc_ace_double
808                     (t_nblist                    * gmx_restrict       nlist,
809                      rvec                        * gmx_restrict          xx,
810                      rvec                        * gmx_restrict          ff,
811                      t_forcerec                  * gmx_restrict          fr,
812                      t_mdatoms                   * gmx_restrict     mdatoms,
813                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
814                      t_nrnb                      * gmx_restrict        nrnb)
815 {
816     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
817      * just 0 for non-waters.
818      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
819      * jnr indices corresponding to data put in the four positions in the SIMD register.
820      */
821     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
822     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
823     int              jnrA,jnrB;
824     int              j_coord_offsetA,j_coord_offsetB;
825     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
826     real             rcutoff_scalar;
827     real             *shiftvec,*fshift,*x,*f;
828     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
829     int              vdwioffset1;
830     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
831     int              vdwioffset2;
832     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
833     int              vdwioffset3;
834     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
835     int              vdwjidx1A,vdwjidx1B;
836     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
837     int              vdwjidx2A,vdwjidx2B;
838     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
839     int              vdwjidx3A,vdwjidx3B;
840     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
841     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
842     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
843     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
844     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
845     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
846     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
847     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
848     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
849     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
850     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
851     real             *charge;
852     _fjsp_v2r8       itab_tmp;
853     _fjsp_v2r8       dummy_mask,cutoff_mask;
854     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
855     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
856     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
857
858     x                = xx[0];
859     f                = ff[0];
860
861     nri              = nlist->nri;
862     iinr             = nlist->iinr;
863     jindex           = nlist->jindex;
864     jjnr             = nlist->jjnr;
865     shiftidx         = nlist->shift;
866     gid              = nlist->gid;
867     shiftvec         = fr->shift_vec[0];
868     fshift           = fr->fshift[0];
869     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
870     charge           = mdatoms->chargeA;
871
872     /* Setup water-specific parameters */
873     inr              = nlist->iinr[0];
874     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
875     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
876     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
877
878     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
879     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
880     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
881     qq11             = _fjsp_mul_v2r8(iq1,jq1);
882     qq12             = _fjsp_mul_v2r8(iq1,jq2);
883     qq13             = _fjsp_mul_v2r8(iq1,jq3);
884     qq21             = _fjsp_mul_v2r8(iq2,jq1);
885     qq22             = _fjsp_mul_v2r8(iq2,jq2);
886     qq23             = _fjsp_mul_v2r8(iq2,jq3);
887     qq31             = _fjsp_mul_v2r8(iq3,jq1);
888     qq32             = _fjsp_mul_v2r8(iq3,jq2);
889     qq33             = _fjsp_mul_v2r8(iq3,jq3);
890
891     /* Avoid stupid compiler warnings */
892     jnrA = jnrB = 0;
893     j_coord_offsetA = 0;
894     j_coord_offsetB = 0;
895
896     outeriter        = 0;
897     inneriter        = 0;
898
899     /* Start outer loop over neighborlists */
900     for(iidx=0; iidx<nri; iidx++)
901     {
902         /* Load shift vector for this list */
903         i_shift_offset   = DIM*shiftidx[iidx];
904
905         /* Load limits for loop over neighbors */
906         j_index_start    = jindex[iidx];
907         j_index_end      = jindex[iidx+1];
908
909         /* Get outer coordinate index */
910         inr              = iinr[iidx];
911         i_coord_offset   = DIM*inr;
912
913         /* Load i particle coords and add shift vector */
914         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
915                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
916
917         fix1             = _fjsp_setzero_v2r8();
918         fiy1             = _fjsp_setzero_v2r8();
919         fiz1             = _fjsp_setzero_v2r8();
920         fix2             = _fjsp_setzero_v2r8();
921         fiy2             = _fjsp_setzero_v2r8();
922         fiz2             = _fjsp_setzero_v2r8();
923         fix3             = _fjsp_setzero_v2r8();
924         fiy3             = _fjsp_setzero_v2r8();
925         fiz3             = _fjsp_setzero_v2r8();
926
927         /* Start inner kernel loop */
928         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
929         {
930
931             /* Get j neighbor index, and coordinate index */
932             jnrA             = jjnr[jidx];
933             jnrB             = jjnr[jidx+1];
934             j_coord_offsetA  = DIM*jnrA;
935             j_coord_offsetB  = DIM*jnrB;
936
937             /* load j atom coordinates */
938             gmx_fjsp_load_3rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
939                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
940
941             /* Calculate displacement vector */
942             dx11             = _fjsp_sub_v2r8(ix1,jx1);
943             dy11             = _fjsp_sub_v2r8(iy1,jy1);
944             dz11             = _fjsp_sub_v2r8(iz1,jz1);
945             dx12             = _fjsp_sub_v2r8(ix1,jx2);
946             dy12             = _fjsp_sub_v2r8(iy1,jy2);
947             dz12             = _fjsp_sub_v2r8(iz1,jz2);
948             dx13             = _fjsp_sub_v2r8(ix1,jx3);
949             dy13             = _fjsp_sub_v2r8(iy1,jy3);
950             dz13             = _fjsp_sub_v2r8(iz1,jz3);
951             dx21             = _fjsp_sub_v2r8(ix2,jx1);
952             dy21             = _fjsp_sub_v2r8(iy2,jy1);
953             dz21             = _fjsp_sub_v2r8(iz2,jz1);
954             dx22             = _fjsp_sub_v2r8(ix2,jx2);
955             dy22             = _fjsp_sub_v2r8(iy2,jy2);
956             dz22             = _fjsp_sub_v2r8(iz2,jz2);
957             dx23             = _fjsp_sub_v2r8(ix2,jx3);
958             dy23             = _fjsp_sub_v2r8(iy2,jy3);
959             dz23             = _fjsp_sub_v2r8(iz2,jz3);
960             dx31             = _fjsp_sub_v2r8(ix3,jx1);
961             dy31             = _fjsp_sub_v2r8(iy3,jy1);
962             dz31             = _fjsp_sub_v2r8(iz3,jz1);
963             dx32             = _fjsp_sub_v2r8(ix3,jx2);
964             dy32             = _fjsp_sub_v2r8(iy3,jy2);
965             dz32             = _fjsp_sub_v2r8(iz3,jz2);
966             dx33             = _fjsp_sub_v2r8(ix3,jx3);
967             dy33             = _fjsp_sub_v2r8(iy3,jy3);
968             dz33             = _fjsp_sub_v2r8(iz3,jz3);
969
970             /* Calculate squared distance and things based on it */
971             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
972             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
973             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
974             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
975             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
976             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
977             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
978             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
979             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
980
981             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
982             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
983             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
984             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
985             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
986             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
987             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
988             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
989             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
990
991             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
992             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
993             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
994             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
995             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
996             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
997             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
998             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
999             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1000
1001             fjx1             = _fjsp_setzero_v2r8();
1002             fjy1             = _fjsp_setzero_v2r8();
1003             fjz1             = _fjsp_setzero_v2r8();
1004             fjx2             = _fjsp_setzero_v2r8();
1005             fjy2             = _fjsp_setzero_v2r8();
1006             fjz2             = _fjsp_setzero_v2r8();
1007             fjx3             = _fjsp_setzero_v2r8();
1008             fjy3             = _fjsp_setzero_v2r8();
1009             fjz3             = _fjsp_setzero_v2r8();
1010
1011             /**************************
1012              * CALCULATE INTERACTIONS *
1013              **************************/
1014
1015             /* COULOMB ELECTROSTATICS */
1016             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1017             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1018
1019             fscal            = felec;
1020
1021             /* Update vectorial force */
1022             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1023             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1024             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1025             
1026             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1027             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1028             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1029
1030             /**************************
1031              * CALCULATE INTERACTIONS *
1032              **************************/
1033
1034             /* COULOMB ELECTROSTATICS */
1035             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1036             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1037
1038             fscal            = felec;
1039
1040             /* Update vectorial force */
1041             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1042             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1043             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1044             
1045             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1046             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1047             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1048
1049             /**************************
1050              * CALCULATE INTERACTIONS *
1051              **************************/
1052
1053             /* COULOMB ELECTROSTATICS */
1054             velec            = _fjsp_mul_v2r8(qq13,rinv13);
1055             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
1056
1057             fscal            = felec;
1058
1059             /* Update vectorial force */
1060             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1061             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1062             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1063             
1064             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1065             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1066             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1067
1068             /**************************
1069              * CALCULATE INTERACTIONS *
1070              **************************/
1071
1072             /* COULOMB ELECTROSTATICS */
1073             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1074             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1075
1076             fscal            = felec;
1077
1078             /* Update vectorial force */
1079             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1080             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1081             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1082             
1083             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1084             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1085             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1086
1087             /**************************
1088              * CALCULATE INTERACTIONS *
1089              **************************/
1090
1091             /* COULOMB ELECTROSTATICS */
1092             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1093             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1094
1095             fscal            = felec;
1096
1097             /* Update vectorial force */
1098             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1099             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1100             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1101             
1102             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1103             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1104             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1105
1106             /**************************
1107              * CALCULATE INTERACTIONS *
1108              **************************/
1109
1110             /* COULOMB ELECTROSTATICS */
1111             velec            = _fjsp_mul_v2r8(qq23,rinv23);
1112             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
1113
1114             fscal            = felec;
1115
1116             /* Update vectorial force */
1117             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1118             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1119             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1120             
1121             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1122             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1123             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1124
1125             /**************************
1126              * CALCULATE INTERACTIONS *
1127              **************************/
1128
1129             /* COULOMB ELECTROSTATICS */
1130             velec            = _fjsp_mul_v2r8(qq31,rinv31);
1131             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
1132
1133             fscal            = felec;
1134
1135             /* Update vectorial force */
1136             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1137             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1138             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1139             
1140             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1141             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1142             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1143
1144             /**************************
1145              * CALCULATE INTERACTIONS *
1146              **************************/
1147
1148             /* COULOMB ELECTROSTATICS */
1149             velec            = _fjsp_mul_v2r8(qq32,rinv32);
1150             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
1151
1152             fscal            = felec;
1153
1154             /* Update vectorial force */
1155             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1156             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1157             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1158             
1159             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1160             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1161             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1162
1163             /**************************
1164              * CALCULATE INTERACTIONS *
1165              **************************/
1166
1167             /* COULOMB ELECTROSTATICS */
1168             velec            = _fjsp_mul_v2r8(qq33,rinv33);
1169             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
1170
1171             fscal            = felec;
1172
1173             /* Update vectorial force */
1174             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1175             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1176             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1177             
1178             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1179             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1180             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1181
1182             gmx_fjsp_decrement_3rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1183
1184             /* Inner loop uses 270 flops */
1185         }
1186
1187         if(jidx<j_index_end)
1188         {
1189
1190             jnrA             = jjnr[jidx];
1191             j_coord_offsetA  = DIM*jnrA;
1192
1193             /* load j atom coordinates */
1194             gmx_fjsp_load_3rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA+DIM,
1195                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1196
1197             /* Calculate displacement vector */
1198             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1199             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1200             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1201             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1202             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1203             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1204             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1205             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1206             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1207             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1208             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1209             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1210             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1211             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1212             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1213             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1214             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1215             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1216             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1217             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1218             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1219             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1220             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1221             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1222             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1223             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1224             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1225
1226             /* Calculate squared distance and things based on it */
1227             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1228             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1229             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1230             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1231             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1232             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1233             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1234             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1235             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1236
1237             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1238             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1239             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1240             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1241             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1242             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1243             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1244             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1245             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1246
1247             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1248             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1249             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1250             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1251             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1252             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1253             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1254             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1255             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1256
1257             fjx1             = _fjsp_setzero_v2r8();
1258             fjy1             = _fjsp_setzero_v2r8();
1259             fjz1             = _fjsp_setzero_v2r8();
1260             fjx2             = _fjsp_setzero_v2r8();
1261             fjy2             = _fjsp_setzero_v2r8();
1262             fjz2             = _fjsp_setzero_v2r8();
1263             fjx3             = _fjsp_setzero_v2r8();
1264             fjy3             = _fjsp_setzero_v2r8();
1265             fjz3             = _fjsp_setzero_v2r8();
1266
1267             /**************************
1268              * CALCULATE INTERACTIONS *
1269              **************************/
1270
1271             /* COULOMB ELECTROSTATICS */
1272             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1273             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1274
1275             fscal            = felec;
1276
1277             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1278
1279             /* Update vectorial force */
1280             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1281             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1282             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1283             
1284             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1285             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1286             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1287
1288             /**************************
1289              * CALCULATE INTERACTIONS *
1290              **************************/
1291
1292             /* COULOMB ELECTROSTATICS */
1293             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1294             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1295
1296             fscal            = felec;
1297
1298             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1299
1300             /* Update vectorial force */
1301             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1302             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1303             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1304             
1305             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1306             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1307             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1308
1309             /**************************
1310              * CALCULATE INTERACTIONS *
1311              **************************/
1312
1313             /* COULOMB ELECTROSTATICS */
1314             velec            = _fjsp_mul_v2r8(qq13,rinv13);
1315             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
1316
1317             fscal            = felec;
1318
1319             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1320
1321             /* Update vectorial force */
1322             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1323             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1324             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1325             
1326             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1327             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1328             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1329
1330             /**************************
1331              * CALCULATE INTERACTIONS *
1332              **************************/
1333
1334             /* COULOMB ELECTROSTATICS */
1335             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1336             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1337
1338             fscal            = felec;
1339
1340             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1341
1342             /* Update vectorial force */
1343             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1344             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1345             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1346             
1347             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1348             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1349             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1350
1351             /**************************
1352              * CALCULATE INTERACTIONS *
1353              **************************/
1354
1355             /* COULOMB ELECTROSTATICS */
1356             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1357             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1358
1359             fscal            = felec;
1360
1361             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1362
1363             /* Update vectorial force */
1364             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1365             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1366             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1367             
1368             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1369             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1370             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1371
1372             /**************************
1373              * CALCULATE INTERACTIONS *
1374              **************************/
1375
1376             /* COULOMB ELECTROSTATICS */
1377             velec            = _fjsp_mul_v2r8(qq23,rinv23);
1378             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
1379
1380             fscal            = felec;
1381
1382             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1383
1384             /* Update vectorial force */
1385             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1386             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1387             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1388             
1389             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1390             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1391             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1392
1393             /**************************
1394              * CALCULATE INTERACTIONS *
1395              **************************/
1396
1397             /* COULOMB ELECTROSTATICS */
1398             velec            = _fjsp_mul_v2r8(qq31,rinv31);
1399             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
1400
1401             fscal            = felec;
1402
1403             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1404
1405             /* Update vectorial force */
1406             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1407             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1408             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1409             
1410             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1411             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1412             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1413
1414             /**************************
1415              * CALCULATE INTERACTIONS *
1416              **************************/
1417
1418             /* COULOMB ELECTROSTATICS */
1419             velec            = _fjsp_mul_v2r8(qq32,rinv32);
1420             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
1421
1422             fscal            = felec;
1423
1424             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1425
1426             /* Update vectorial force */
1427             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1428             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1429             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1430             
1431             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1432             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1433             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1434
1435             /**************************
1436              * CALCULATE INTERACTIONS *
1437              **************************/
1438
1439             /* COULOMB ELECTROSTATICS */
1440             velec            = _fjsp_mul_v2r8(qq33,rinv33);
1441             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
1442
1443             fscal            = felec;
1444
1445             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1446
1447             /* Update vectorial force */
1448             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1449             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1450             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1451             
1452             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1453             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1454             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1455
1456             gmx_fjsp_decrement_3rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1457
1458             /* Inner loop uses 270 flops */
1459         }
1460
1461         /* End of innermost loop */
1462
1463         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1464                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1465
1466         /* Increment number of inner iterations */
1467         inneriter                  += j_index_end - j_index_start;
1468
1469         /* Outer loop uses 18 flops */
1470     }
1471
1472     /* Increment number of outer iterations */
1473     outeriter        += nri;
1474
1475     /* Update outer/inner flops */
1476
1477     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1478 }