f22ec32c9d010cd2298617950c9ef26a08d03d12
[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sparc64_hpc_ace_double / nb_kernel_ElecCoul_VdwLJ_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 "types/simple.h"
44 #include "gromacs/math/vec.h"
45 #include "nrnb.h"
46
47 #include "kernelutil_sparc64_hpc_ace_double.h"
48
49 /*
50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sparc64_hpc_ace_double
51  * Electrostatics interaction: Coulomb
52  * VdW interaction:            LennardJones
53  * Geometry:                   Water4-Water4
54  * Calculate force/pot:        PotentialAndForce
55  */
56 void
57 nb_kernel_ElecCoul_VdwLJ_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              vdwioffset0;
80     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
81     int              vdwioffset1;
82     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
83     int              vdwioffset2;
84     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
85     int              vdwioffset3;
86     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
87     int              vdwjidx0A,vdwjidx0B;
88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
89     int              vdwjidx1A,vdwjidx1B;
90     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
91     int              vdwjidx2A,vdwjidx2B;
92     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
93     int              vdwjidx3A,vdwjidx3B;
94     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
95     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
96     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
97     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
98     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
99     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
100     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
101     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
102     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
103     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
104     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
105     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
106     real             *charge;
107     int              nvdwtype;
108     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
109     int              *vdwtype;
110     real             *vdwparam;
111     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
112     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
113     _fjsp_v2r8       itab_tmp;
114     _fjsp_v2r8       dummy_mask,cutoff_mask;
115     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
116     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
117     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
118
119     x                = xx[0];
120     f                = ff[0];
121
122     nri              = nlist->nri;
123     iinr             = nlist->iinr;
124     jindex           = nlist->jindex;
125     jjnr             = nlist->jjnr;
126     shiftidx         = nlist->shift;
127     gid              = nlist->gid;
128     shiftvec         = fr->shift_vec[0];
129     fshift           = fr->fshift[0];
130     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
131     charge           = mdatoms->chargeA;
132     nvdwtype         = fr->ntype;
133     vdwparam         = fr->nbfp;
134     vdwtype          = mdatoms->typeA;
135
136     /* Setup water-specific parameters */
137     inr              = nlist->iinr[0];
138     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
139     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
140     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
141     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
142
143     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
144     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
145     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
146     vdwjidx0A        = 2*vdwtype[inr+0];
147     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
148     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
149     qq11             = _fjsp_mul_v2r8(iq1,jq1);
150     qq12             = _fjsp_mul_v2r8(iq1,jq2);
151     qq13             = _fjsp_mul_v2r8(iq1,jq3);
152     qq21             = _fjsp_mul_v2r8(iq2,jq1);
153     qq22             = _fjsp_mul_v2r8(iq2,jq2);
154     qq23             = _fjsp_mul_v2r8(iq2,jq3);
155     qq31             = _fjsp_mul_v2r8(iq3,jq1);
156     qq32             = _fjsp_mul_v2r8(iq3,jq2);
157     qq33             = _fjsp_mul_v2r8(iq3,jq3);
158
159     /* Avoid stupid compiler warnings */
160     jnrA = jnrB = 0;
161     j_coord_offsetA = 0;
162     j_coord_offsetB = 0;
163
164     outeriter        = 0;
165     inneriter        = 0;
166
167     /* Start outer loop over neighborlists */
168     for(iidx=0; iidx<nri; iidx++)
169     {
170         /* Load shift vector for this list */
171         i_shift_offset   = DIM*shiftidx[iidx];
172
173         /* Load limits for loop over neighbors */
174         j_index_start    = jindex[iidx];
175         j_index_end      = jindex[iidx+1];
176
177         /* Get outer coordinate index */
178         inr              = iinr[iidx];
179         i_coord_offset   = DIM*inr;
180
181         /* Load i particle coords and add shift vector */
182         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
183                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
184
185         fix0             = _fjsp_setzero_v2r8();
186         fiy0             = _fjsp_setzero_v2r8();
187         fiz0             = _fjsp_setzero_v2r8();
188         fix1             = _fjsp_setzero_v2r8();
189         fiy1             = _fjsp_setzero_v2r8();
190         fiz1             = _fjsp_setzero_v2r8();
191         fix2             = _fjsp_setzero_v2r8();
192         fiy2             = _fjsp_setzero_v2r8();
193         fiz2             = _fjsp_setzero_v2r8();
194         fix3             = _fjsp_setzero_v2r8();
195         fiy3             = _fjsp_setzero_v2r8();
196         fiz3             = _fjsp_setzero_v2r8();
197
198         /* Reset potential sums */
199         velecsum         = _fjsp_setzero_v2r8();
200         vvdwsum          = _fjsp_setzero_v2r8();
201
202         /* Start inner kernel loop */
203         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
204         {
205
206             /* Get j neighbor index, and coordinate index */
207             jnrA             = jjnr[jidx];
208             jnrB             = jjnr[jidx+1];
209             j_coord_offsetA  = DIM*jnrA;
210             j_coord_offsetB  = DIM*jnrB;
211
212             /* load j atom coordinates */
213             gmx_fjsp_load_4rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
214                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
215                                               &jy2,&jz2,&jx3,&jy3,&jz3);
216
217             /* Calculate displacement vector */
218             dx00             = _fjsp_sub_v2r8(ix0,jx0);
219             dy00             = _fjsp_sub_v2r8(iy0,jy0);
220             dz00             = _fjsp_sub_v2r8(iz0,jz0);
221             dx11             = _fjsp_sub_v2r8(ix1,jx1);
222             dy11             = _fjsp_sub_v2r8(iy1,jy1);
223             dz11             = _fjsp_sub_v2r8(iz1,jz1);
224             dx12             = _fjsp_sub_v2r8(ix1,jx2);
225             dy12             = _fjsp_sub_v2r8(iy1,jy2);
226             dz12             = _fjsp_sub_v2r8(iz1,jz2);
227             dx13             = _fjsp_sub_v2r8(ix1,jx3);
228             dy13             = _fjsp_sub_v2r8(iy1,jy3);
229             dz13             = _fjsp_sub_v2r8(iz1,jz3);
230             dx21             = _fjsp_sub_v2r8(ix2,jx1);
231             dy21             = _fjsp_sub_v2r8(iy2,jy1);
232             dz21             = _fjsp_sub_v2r8(iz2,jz1);
233             dx22             = _fjsp_sub_v2r8(ix2,jx2);
234             dy22             = _fjsp_sub_v2r8(iy2,jy2);
235             dz22             = _fjsp_sub_v2r8(iz2,jz2);
236             dx23             = _fjsp_sub_v2r8(ix2,jx3);
237             dy23             = _fjsp_sub_v2r8(iy2,jy3);
238             dz23             = _fjsp_sub_v2r8(iz2,jz3);
239             dx31             = _fjsp_sub_v2r8(ix3,jx1);
240             dy31             = _fjsp_sub_v2r8(iy3,jy1);
241             dz31             = _fjsp_sub_v2r8(iz3,jz1);
242             dx32             = _fjsp_sub_v2r8(ix3,jx2);
243             dy32             = _fjsp_sub_v2r8(iy3,jy2);
244             dz32             = _fjsp_sub_v2r8(iz3,jz2);
245             dx33             = _fjsp_sub_v2r8(ix3,jx3);
246             dy33             = _fjsp_sub_v2r8(iy3,jy3);
247             dz33             = _fjsp_sub_v2r8(iz3,jz3);
248
249             /* Calculate squared distance and things based on it */
250             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
251             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
252             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
253             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
254             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
255             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
256             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
257             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
258             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
259             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
260
261             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
262             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
263             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
264             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
265             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
266             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
267             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
268             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
269             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
270
271             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
272             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
273             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
274             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
275             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
276             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
277             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
278             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
279             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
280             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
281
282             fjx0             = _fjsp_setzero_v2r8();
283             fjy0             = _fjsp_setzero_v2r8();
284             fjz0             = _fjsp_setzero_v2r8();
285             fjx1             = _fjsp_setzero_v2r8();
286             fjy1             = _fjsp_setzero_v2r8();
287             fjz1             = _fjsp_setzero_v2r8();
288             fjx2             = _fjsp_setzero_v2r8();
289             fjy2             = _fjsp_setzero_v2r8();
290             fjz2             = _fjsp_setzero_v2r8();
291             fjx3             = _fjsp_setzero_v2r8();
292             fjy3             = _fjsp_setzero_v2r8();
293             fjz3             = _fjsp_setzero_v2r8();
294
295             /**************************
296              * CALCULATE INTERACTIONS *
297              **************************/
298
299             /* LENNARD-JONES DISPERSION/REPULSION */
300
301             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
302             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
303             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
304             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
305             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
306
307             /* Update potential sum for this i atom from the interaction with this j atom. */
308             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
309
310             fscal            = fvdw;
311
312             /* Update vectorial force */
313             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
314             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
315             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
316             
317             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
318             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
319             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
320
321             /**************************
322              * CALCULATE INTERACTIONS *
323              **************************/
324
325             /* COULOMB ELECTROSTATICS */
326             velec            = _fjsp_mul_v2r8(qq11,rinv11);
327             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
328
329             /* Update potential sum for this i atom from the interaction with this j atom. */
330             velecsum         = _fjsp_add_v2r8(velecsum,velec);
331
332             fscal            = felec;
333
334             /* Update vectorial force */
335             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
336             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
337             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
338             
339             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
340             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
341             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
342
343             /**************************
344              * CALCULATE INTERACTIONS *
345              **************************/
346
347             /* COULOMB ELECTROSTATICS */
348             velec            = _fjsp_mul_v2r8(qq12,rinv12);
349             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
350
351             /* Update potential sum for this i atom from the interaction with this j atom. */
352             velecsum         = _fjsp_add_v2r8(velecsum,velec);
353
354             fscal            = felec;
355
356             /* Update vectorial force */
357             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
358             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
359             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
360             
361             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
362             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
363             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
364
365             /**************************
366              * CALCULATE INTERACTIONS *
367              **************************/
368
369             /* COULOMB ELECTROSTATICS */
370             velec            = _fjsp_mul_v2r8(qq13,rinv13);
371             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
372
373             /* Update potential sum for this i atom from the interaction with this j atom. */
374             velecsum         = _fjsp_add_v2r8(velecsum,velec);
375
376             fscal            = felec;
377
378             /* Update vectorial force */
379             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
380             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
381             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
382             
383             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
384             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
385             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
386
387             /**************************
388              * CALCULATE INTERACTIONS *
389              **************************/
390
391             /* COULOMB ELECTROSTATICS */
392             velec            = _fjsp_mul_v2r8(qq21,rinv21);
393             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
394
395             /* Update potential sum for this i atom from the interaction with this j atom. */
396             velecsum         = _fjsp_add_v2r8(velecsum,velec);
397
398             fscal            = felec;
399
400             /* Update vectorial force */
401             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
402             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
403             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
404             
405             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
406             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
407             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
408
409             /**************************
410              * CALCULATE INTERACTIONS *
411              **************************/
412
413             /* COULOMB ELECTROSTATICS */
414             velec            = _fjsp_mul_v2r8(qq22,rinv22);
415             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
416
417             /* Update potential sum for this i atom from the interaction with this j atom. */
418             velecsum         = _fjsp_add_v2r8(velecsum,velec);
419
420             fscal            = felec;
421
422             /* Update vectorial force */
423             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
424             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
425             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
426             
427             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
428             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
429             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
430
431             /**************************
432              * CALCULATE INTERACTIONS *
433              **************************/
434
435             /* COULOMB ELECTROSTATICS */
436             velec            = _fjsp_mul_v2r8(qq23,rinv23);
437             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
438
439             /* Update potential sum for this i atom from the interaction with this j atom. */
440             velecsum         = _fjsp_add_v2r8(velecsum,velec);
441
442             fscal            = felec;
443
444             /* Update vectorial force */
445             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
446             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
447             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
448             
449             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
450             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
451             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
452
453             /**************************
454              * CALCULATE INTERACTIONS *
455              **************************/
456
457             /* COULOMB ELECTROSTATICS */
458             velec            = _fjsp_mul_v2r8(qq31,rinv31);
459             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
460
461             /* Update potential sum for this i atom from the interaction with this j atom. */
462             velecsum         = _fjsp_add_v2r8(velecsum,velec);
463
464             fscal            = felec;
465
466             /* Update vectorial force */
467             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
468             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
469             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
470             
471             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
472             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
473             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
474
475             /**************************
476              * CALCULATE INTERACTIONS *
477              **************************/
478
479             /* COULOMB ELECTROSTATICS */
480             velec            = _fjsp_mul_v2r8(qq32,rinv32);
481             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
482
483             /* Update potential sum for this i atom from the interaction with this j atom. */
484             velecsum         = _fjsp_add_v2r8(velecsum,velec);
485
486             fscal            = felec;
487
488             /* Update vectorial force */
489             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
490             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
491             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
492             
493             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
494             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
495             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
496
497             /**************************
498              * CALCULATE INTERACTIONS *
499              **************************/
500
501             /* COULOMB ELECTROSTATICS */
502             velec            = _fjsp_mul_v2r8(qq33,rinv33);
503             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
504
505             /* Update potential sum for this i atom from the interaction with this j atom. */
506             velecsum         = _fjsp_add_v2r8(velecsum,velec);
507
508             fscal            = felec;
509
510             /* Update vectorial force */
511             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
512             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
513             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
514             
515             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
516             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
517             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
518
519             gmx_fjsp_decrement_4rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
520
521             /* Inner loop uses 317 flops */
522         }
523
524         if(jidx<j_index_end)
525         {
526
527             jnrA             = jjnr[jidx];
528             j_coord_offsetA  = DIM*jnrA;
529
530             /* load j atom coordinates */
531             gmx_fjsp_load_4rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
532                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
533                                               &jy2,&jz2,&jx3,&jy3,&jz3);
534
535             /* Calculate displacement vector */
536             dx00             = _fjsp_sub_v2r8(ix0,jx0);
537             dy00             = _fjsp_sub_v2r8(iy0,jy0);
538             dz00             = _fjsp_sub_v2r8(iz0,jz0);
539             dx11             = _fjsp_sub_v2r8(ix1,jx1);
540             dy11             = _fjsp_sub_v2r8(iy1,jy1);
541             dz11             = _fjsp_sub_v2r8(iz1,jz1);
542             dx12             = _fjsp_sub_v2r8(ix1,jx2);
543             dy12             = _fjsp_sub_v2r8(iy1,jy2);
544             dz12             = _fjsp_sub_v2r8(iz1,jz2);
545             dx13             = _fjsp_sub_v2r8(ix1,jx3);
546             dy13             = _fjsp_sub_v2r8(iy1,jy3);
547             dz13             = _fjsp_sub_v2r8(iz1,jz3);
548             dx21             = _fjsp_sub_v2r8(ix2,jx1);
549             dy21             = _fjsp_sub_v2r8(iy2,jy1);
550             dz21             = _fjsp_sub_v2r8(iz2,jz1);
551             dx22             = _fjsp_sub_v2r8(ix2,jx2);
552             dy22             = _fjsp_sub_v2r8(iy2,jy2);
553             dz22             = _fjsp_sub_v2r8(iz2,jz2);
554             dx23             = _fjsp_sub_v2r8(ix2,jx3);
555             dy23             = _fjsp_sub_v2r8(iy2,jy3);
556             dz23             = _fjsp_sub_v2r8(iz2,jz3);
557             dx31             = _fjsp_sub_v2r8(ix3,jx1);
558             dy31             = _fjsp_sub_v2r8(iy3,jy1);
559             dz31             = _fjsp_sub_v2r8(iz3,jz1);
560             dx32             = _fjsp_sub_v2r8(ix3,jx2);
561             dy32             = _fjsp_sub_v2r8(iy3,jy2);
562             dz32             = _fjsp_sub_v2r8(iz3,jz2);
563             dx33             = _fjsp_sub_v2r8(ix3,jx3);
564             dy33             = _fjsp_sub_v2r8(iy3,jy3);
565             dz33             = _fjsp_sub_v2r8(iz3,jz3);
566
567             /* Calculate squared distance and things based on it */
568             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
569             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
570             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
571             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
572             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
573             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
574             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
575             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
576             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
577             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
578
579             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
580             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
581             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
582             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
583             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
584             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
585             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
586             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
587             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
588
589             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
590             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
591             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
592             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
593             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
594             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
595             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
596             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
597             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
598             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
599
600             fjx0             = _fjsp_setzero_v2r8();
601             fjy0             = _fjsp_setzero_v2r8();
602             fjz0             = _fjsp_setzero_v2r8();
603             fjx1             = _fjsp_setzero_v2r8();
604             fjy1             = _fjsp_setzero_v2r8();
605             fjz1             = _fjsp_setzero_v2r8();
606             fjx2             = _fjsp_setzero_v2r8();
607             fjy2             = _fjsp_setzero_v2r8();
608             fjz2             = _fjsp_setzero_v2r8();
609             fjx3             = _fjsp_setzero_v2r8();
610             fjy3             = _fjsp_setzero_v2r8();
611             fjz3             = _fjsp_setzero_v2r8();
612
613             /**************************
614              * CALCULATE INTERACTIONS *
615              **************************/
616
617             /* LENNARD-JONES DISPERSION/REPULSION */
618
619             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
620             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
621             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
622             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
623             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
624
625             /* Update potential sum for this i atom from the interaction with this j atom. */
626             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
627             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
628
629             fscal            = fvdw;
630
631             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
632
633             /* Update vectorial force */
634             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
635             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
636             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
637             
638             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
639             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
640             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
641
642             /**************************
643              * CALCULATE INTERACTIONS *
644              **************************/
645
646             /* COULOMB ELECTROSTATICS */
647             velec            = _fjsp_mul_v2r8(qq11,rinv11);
648             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
649
650             /* Update potential sum for this i atom from the interaction with this j atom. */
651             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
652             velecsum         = _fjsp_add_v2r8(velecsum,velec);
653
654             fscal            = felec;
655
656             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
657
658             /* Update vectorial force */
659             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
660             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
661             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
662             
663             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
664             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
665             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
666
667             /**************************
668              * CALCULATE INTERACTIONS *
669              **************************/
670
671             /* COULOMB ELECTROSTATICS */
672             velec            = _fjsp_mul_v2r8(qq12,rinv12);
673             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
674
675             /* Update potential sum for this i atom from the interaction with this j atom. */
676             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
677             velecsum         = _fjsp_add_v2r8(velecsum,velec);
678
679             fscal            = felec;
680
681             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
682
683             /* Update vectorial force */
684             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
685             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
686             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
687             
688             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
689             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
690             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
691
692             /**************************
693              * CALCULATE INTERACTIONS *
694              **************************/
695
696             /* COULOMB ELECTROSTATICS */
697             velec            = _fjsp_mul_v2r8(qq13,rinv13);
698             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
699
700             /* Update potential sum for this i atom from the interaction with this j atom. */
701             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
702             velecsum         = _fjsp_add_v2r8(velecsum,velec);
703
704             fscal            = felec;
705
706             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
707
708             /* Update vectorial force */
709             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
710             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
711             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
712             
713             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
714             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
715             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
716
717             /**************************
718              * CALCULATE INTERACTIONS *
719              **************************/
720
721             /* COULOMB ELECTROSTATICS */
722             velec            = _fjsp_mul_v2r8(qq21,rinv21);
723             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
724
725             /* Update potential sum for this i atom from the interaction with this j atom. */
726             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
727             velecsum         = _fjsp_add_v2r8(velecsum,velec);
728
729             fscal            = felec;
730
731             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
732
733             /* Update vectorial force */
734             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
735             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
736             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
737             
738             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
739             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
740             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
741
742             /**************************
743              * CALCULATE INTERACTIONS *
744              **************************/
745
746             /* COULOMB ELECTROSTATICS */
747             velec            = _fjsp_mul_v2r8(qq22,rinv22);
748             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
749
750             /* Update potential sum for this i atom from the interaction with this j atom. */
751             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
752             velecsum         = _fjsp_add_v2r8(velecsum,velec);
753
754             fscal            = felec;
755
756             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
757
758             /* Update vectorial force */
759             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
760             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
761             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
762             
763             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
764             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
765             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
766
767             /**************************
768              * CALCULATE INTERACTIONS *
769              **************************/
770
771             /* COULOMB ELECTROSTATICS */
772             velec            = _fjsp_mul_v2r8(qq23,rinv23);
773             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
774
775             /* Update potential sum for this i atom from the interaction with this j atom. */
776             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
777             velecsum         = _fjsp_add_v2r8(velecsum,velec);
778
779             fscal            = felec;
780
781             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
782
783             /* Update vectorial force */
784             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
785             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
786             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
787             
788             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
789             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
790             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
791
792             /**************************
793              * CALCULATE INTERACTIONS *
794              **************************/
795
796             /* COULOMB ELECTROSTATICS */
797             velec            = _fjsp_mul_v2r8(qq31,rinv31);
798             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
799
800             /* Update potential sum for this i atom from the interaction with this j atom. */
801             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
802             velecsum         = _fjsp_add_v2r8(velecsum,velec);
803
804             fscal            = felec;
805
806             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
807
808             /* Update vectorial force */
809             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
810             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
811             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
812             
813             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
814             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
815             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
816
817             /**************************
818              * CALCULATE INTERACTIONS *
819              **************************/
820
821             /* COULOMB ELECTROSTATICS */
822             velec            = _fjsp_mul_v2r8(qq32,rinv32);
823             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
824
825             /* Update potential sum for this i atom from the interaction with this j atom. */
826             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
827             velecsum         = _fjsp_add_v2r8(velecsum,velec);
828
829             fscal            = felec;
830
831             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
832
833             /* Update vectorial force */
834             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
835             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
836             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
837             
838             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
839             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
840             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
841
842             /**************************
843              * CALCULATE INTERACTIONS *
844              **************************/
845
846             /* COULOMB ELECTROSTATICS */
847             velec            = _fjsp_mul_v2r8(qq33,rinv33);
848             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
849
850             /* Update potential sum for this i atom from the interaction with this j atom. */
851             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
852             velecsum         = _fjsp_add_v2r8(velecsum,velec);
853
854             fscal            = felec;
855
856             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
857
858             /* Update vectorial force */
859             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
860             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
861             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
862             
863             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
864             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
865             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
866
867             gmx_fjsp_decrement_4rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
868
869             /* Inner loop uses 317 flops */
870         }
871
872         /* End of innermost loop */
873
874         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
875                                               f+i_coord_offset,fshift+i_shift_offset);
876
877         ggid                        = gid[iidx];
878         /* Update potential energies */
879         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
880         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
881
882         /* Increment number of inner iterations */
883         inneriter                  += j_index_end - j_index_start;
884
885         /* Outer loop uses 26 flops */
886     }
887
888     /* Increment number of outer iterations */
889     outeriter        += nri;
890
891     /* Update outer/inner flops */
892
893     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*317);
894 }
895 /*
896  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sparc64_hpc_ace_double
897  * Electrostatics interaction: Coulomb
898  * VdW interaction:            LennardJones
899  * Geometry:                   Water4-Water4
900  * Calculate force/pot:        Force
901  */
902 void
903 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sparc64_hpc_ace_double
904                     (t_nblist                    * gmx_restrict       nlist,
905                      rvec                        * gmx_restrict          xx,
906                      rvec                        * gmx_restrict          ff,
907                      t_forcerec                  * gmx_restrict          fr,
908                      t_mdatoms                   * gmx_restrict     mdatoms,
909                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
910                      t_nrnb                      * gmx_restrict        nrnb)
911 {
912     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
913      * just 0 for non-waters.
914      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
915      * jnr indices corresponding to data put in the four positions in the SIMD register.
916      */
917     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
918     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
919     int              jnrA,jnrB;
920     int              j_coord_offsetA,j_coord_offsetB;
921     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
922     real             rcutoff_scalar;
923     real             *shiftvec,*fshift,*x,*f;
924     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
925     int              vdwioffset0;
926     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
927     int              vdwioffset1;
928     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
929     int              vdwioffset2;
930     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
931     int              vdwioffset3;
932     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
933     int              vdwjidx0A,vdwjidx0B;
934     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
935     int              vdwjidx1A,vdwjidx1B;
936     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
937     int              vdwjidx2A,vdwjidx2B;
938     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
939     int              vdwjidx3A,vdwjidx3B;
940     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
941     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
942     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
943     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
944     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
945     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
946     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
947     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
948     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
949     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
950     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
951     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
952     real             *charge;
953     int              nvdwtype;
954     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
955     int              *vdwtype;
956     real             *vdwparam;
957     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
958     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
959     _fjsp_v2r8       itab_tmp;
960     _fjsp_v2r8       dummy_mask,cutoff_mask;
961     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
962     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
963     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
964
965     x                = xx[0];
966     f                = ff[0];
967
968     nri              = nlist->nri;
969     iinr             = nlist->iinr;
970     jindex           = nlist->jindex;
971     jjnr             = nlist->jjnr;
972     shiftidx         = nlist->shift;
973     gid              = nlist->gid;
974     shiftvec         = fr->shift_vec[0];
975     fshift           = fr->fshift[0];
976     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
977     charge           = mdatoms->chargeA;
978     nvdwtype         = fr->ntype;
979     vdwparam         = fr->nbfp;
980     vdwtype          = mdatoms->typeA;
981
982     /* Setup water-specific parameters */
983     inr              = nlist->iinr[0];
984     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
985     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
986     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
987     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
988
989     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
990     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
991     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
992     vdwjidx0A        = 2*vdwtype[inr+0];
993     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
994     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
995     qq11             = _fjsp_mul_v2r8(iq1,jq1);
996     qq12             = _fjsp_mul_v2r8(iq1,jq2);
997     qq13             = _fjsp_mul_v2r8(iq1,jq3);
998     qq21             = _fjsp_mul_v2r8(iq2,jq1);
999     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1000     qq23             = _fjsp_mul_v2r8(iq2,jq3);
1001     qq31             = _fjsp_mul_v2r8(iq3,jq1);
1002     qq32             = _fjsp_mul_v2r8(iq3,jq2);
1003     qq33             = _fjsp_mul_v2r8(iq3,jq3);
1004
1005     /* Avoid stupid compiler warnings */
1006     jnrA = jnrB = 0;
1007     j_coord_offsetA = 0;
1008     j_coord_offsetB = 0;
1009
1010     outeriter        = 0;
1011     inneriter        = 0;
1012
1013     /* Start outer loop over neighborlists */
1014     for(iidx=0; iidx<nri; iidx++)
1015     {
1016         /* Load shift vector for this list */
1017         i_shift_offset   = DIM*shiftidx[iidx];
1018
1019         /* Load limits for loop over neighbors */
1020         j_index_start    = jindex[iidx];
1021         j_index_end      = jindex[iidx+1];
1022
1023         /* Get outer coordinate index */
1024         inr              = iinr[iidx];
1025         i_coord_offset   = DIM*inr;
1026
1027         /* Load i particle coords and add shift vector */
1028         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1029                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1030
1031         fix0             = _fjsp_setzero_v2r8();
1032         fiy0             = _fjsp_setzero_v2r8();
1033         fiz0             = _fjsp_setzero_v2r8();
1034         fix1             = _fjsp_setzero_v2r8();
1035         fiy1             = _fjsp_setzero_v2r8();
1036         fiz1             = _fjsp_setzero_v2r8();
1037         fix2             = _fjsp_setzero_v2r8();
1038         fiy2             = _fjsp_setzero_v2r8();
1039         fiz2             = _fjsp_setzero_v2r8();
1040         fix3             = _fjsp_setzero_v2r8();
1041         fiy3             = _fjsp_setzero_v2r8();
1042         fiz3             = _fjsp_setzero_v2r8();
1043
1044         /* Start inner kernel loop */
1045         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1046         {
1047
1048             /* Get j neighbor index, and coordinate index */
1049             jnrA             = jjnr[jidx];
1050             jnrB             = jjnr[jidx+1];
1051             j_coord_offsetA  = DIM*jnrA;
1052             j_coord_offsetB  = DIM*jnrB;
1053
1054             /* load j atom coordinates */
1055             gmx_fjsp_load_4rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1056                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1057                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1058
1059             /* Calculate displacement vector */
1060             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1061             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1062             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1063             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1064             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1065             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1066             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1067             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1068             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1069             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1070             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1071             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1072             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1073             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1074             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1075             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1076             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1077             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1078             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1079             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1080             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1081             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1082             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1083             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1084             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1085             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1086             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1087             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1088             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1089             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1090
1091             /* Calculate squared distance and things based on it */
1092             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1093             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1094             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1095             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1096             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1097             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1098             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1099             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1100             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1101             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1102
1103             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1104             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1105             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1106             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1107             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1108             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1109             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1110             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1111             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1112
1113             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
1114             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1115             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1116             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1117             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1118             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1119             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1120             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1121             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1122             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1123
1124             fjx0             = _fjsp_setzero_v2r8();
1125             fjy0             = _fjsp_setzero_v2r8();
1126             fjz0             = _fjsp_setzero_v2r8();
1127             fjx1             = _fjsp_setzero_v2r8();
1128             fjy1             = _fjsp_setzero_v2r8();
1129             fjz1             = _fjsp_setzero_v2r8();
1130             fjx2             = _fjsp_setzero_v2r8();
1131             fjy2             = _fjsp_setzero_v2r8();
1132             fjz2             = _fjsp_setzero_v2r8();
1133             fjx3             = _fjsp_setzero_v2r8();
1134             fjy3             = _fjsp_setzero_v2r8();
1135             fjz3             = _fjsp_setzero_v2r8();
1136
1137             /**************************
1138              * CALCULATE INTERACTIONS *
1139              **************************/
1140
1141             /* LENNARD-JONES DISPERSION/REPULSION */
1142
1143             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1144             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1145
1146             fscal            = fvdw;
1147
1148             /* Update vectorial force */
1149             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1150             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1151             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1152             
1153             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1154             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1155             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1156
1157             /**************************
1158              * CALCULATE INTERACTIONS *
1159              **************************/
1160
1161             /* COULOMB ELECTROSTATICS */
1162             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1163             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1164
1165             fscal            = felec;
1166
1167             /* Update vectorial force */
1168             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1169             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1170             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1171             
1172             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1173             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1174             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1175
1176             /**************************
1177              * CALCULATE INTERACTIONS *
1178              **************************/
1179
1180             /* COULOMB ELECTROSTATICS */
1181             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1182             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1183
1184             fscal            = felec;
1185
1186             /* Update vectorial force */
1187             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1188             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1189             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1190             
1191             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1192             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1193             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1194
1195             /**************************
1196              * CALCULATE INTERACTIONS *
1197              **************************/
1198
1199             /* COULOMB ELECTROSTATICS */
1200             velec            = _fjsp_mul_v2r8(qq13,rinv13);
1201             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
1202
1203             fscal            = felec;
1204
1205             /* Update vectorial force */
1206             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1207             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1208             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1209             
1210             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1211             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1212             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1213
1214             /**************************
1215              * CALCULATE INTERACTIONS *
1216              **************************/
1217
1218             /* COULOMB ELECTROSTATICS */
1219             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1220             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1221
1222             fscal            = felec;
1223
1224             /* Update vectorial force */
1225             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1226             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1227             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1228             
1229             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1230             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1231             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1232
1233             /**************************
1234              * CALCULATE INTERACTIONS *
1235              **************************/
1236
1237             /* COULOMB ELECTROSTATICS */
1238             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1239             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1240
1241             fscal            = felec;
1242
1243             /* Update vectorial force */
1244             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1245             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1246             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1247             
1248             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1249             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1250             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1251
1252             /**************************
1253              * CALCULATE INTERACTIONS *
1254              **************************/
1255
1256             /* COULOMB ELECTROSTATICS */
1257             velec            = _fjsp_mul_v2r8(qq23,rinv23);
1258             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
1259
1260             fscal            = felec;
1261
1262             /* Update vectorial force */
1263             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1264             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1265             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1266             
1267             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1268             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1269             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1270
1271             /**************************
1272              * CALCULATE INTERACTIONS *
1273              **************************/
1274
1275             /* COULOMB ELECTROSTATICS */
1276             velec            = _fjsp_mul_v2r8(qq31,rinv31);
1277             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
1278
1279             fscal            = felec;
1280
1281             /* Update vectorial force */
1282             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1283             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1284             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1285             
1286             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1287             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1288             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             /* COULOMB ELECTROSTATICS */
1295             velec            = _fjsp_mul_v2r8(qq32,rinv32);
1296             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
1297
1298             fscal            = felec;
1299
1300             /* Update vectorial force */
1301             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1302             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1303             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1304             
1305             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1306             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1307             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1308
1309             /**************************
1310              * CALCULATE INTERACTIONS *
1311              **************************/
1312
1313             /* COULOMB ELECTROSTATICS */
1314             velec            = _fjsp_mul_v2r8(qq33,rinv33);
1315             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
1316
1317             fscal            = felec;
1318
1319             /* Update vectorial force */
1320             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1321             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1322             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1323             
1324             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1325             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1326             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1327
1328             gmx_fjsp_decrement_4rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1329
1330             /* Inner loop uses 303 flops */
1331         }
1332
1333         if(jidx<j_index_end)
1334         {
1335
1336             jnrA             = jjnr[jidx];
1337             j_coord_offsetA  = DIM*jnrA;
1338
1339             /* load j atom coordinates */
1340             gmx_fjsp_load_4rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1341                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1342                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1343
1344             /* Calculate displacement vector */
1345             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1346             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1347             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1348             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1349             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1350             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1351             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1352             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1353             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1354             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1355             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1356             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1357             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1358             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1359             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1360             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1361             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1362             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1363             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1364             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1365             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1366             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1367             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1368             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1369             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1370             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1371             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1372             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1373             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1374             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1375
1376             /* Calculate squared distance and things based on it */
1377             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1378             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1379             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1380             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1381             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1382             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1383             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1384             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1385             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1386             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1387
1388             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1389             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1390             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1391             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1392             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1393             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1394             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1395             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1396             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1397
1398             rinvsq00         = gmx_fjsp_inv_v2r8(rsq00);
1399             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1400             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1401             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1402             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1403             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1404             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1405             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1406             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1407             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1408
1409             fjx0             = _fjsp_setzero_v2r8();
1410             fjy0             = _fjsp_setzero_v2r8();
1411             fjz0             = _fjsp_setzero_v2r8();
1412             fjx1             = _fjsp_setzero_v2r8();
1413             fjy1             = _fjsp_setzero_v2r8();
1414             fjz1             = _fjsp_setzero_v2r8();
1415             fjx2             = _fjsp_setzero_v2r8();
1416             fjy2             = _fjsp_setzero_v2r8();
1417             fjz2             = _fjsp_setzero_v2r8();
1418             fjx3             = _fjsp_setzero_v2r8();
1419             fjy3             = _fjsp_setzero_v2r8();
1420             fjz3             = _fjsp_setzero_v2r8();
1421
1422             /**************************
1423              * CALCULATE INTERACTIONS *
1424              **************************/
1425
1426             /* LENNARD-JONES DISPERSION/REPULSION */
1427
1428             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1429             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
1430
1431             fscal            = fvdw;
1432
1433             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1434
1435             /* Update vectorial force */
1436             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1437             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1438             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1439             
1440             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1441             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1442             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1443
1444             /**************************
1445              * CALCULATE INTERACTIONS *
1446              **************************/
1447
1448             /* COULOMB ELECTROSTATICS */
1449             velec            = _fjsp_mul_v2r8(qq11,rinv11);
1450             felec            = _fjsp_mul_v2r8(velec,rinvsq11);
1451
1452             fscal            = felec;
1453
1454             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1455
1456             /* Update vectorial force */
1457             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1458             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1459             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1460             
1461             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1462             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1463             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1464
1465             /**************************
1466              * CALCULATE INTERACTIONS *
1467              **************************/
1468
1469             /* COULOMB ELECTROSTATICS */
1470             velec            = _fjsp_mul_v2r8(qq12,rinv12);
1471             felec            = _fjsp_mul_v2r8(velec,rinvsq12);
1472
1473             fscal            = felec;
1474
1475             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1476
1477             /* Update vectorial force */
1478             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1479             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1480             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1481             
1482             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1483             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1484             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1485
1486             /**************************
1487              * CALCULATE INTERACTIONS *
1488              **************************/
1489
1490             /* COULOMB ELECTROSTATICS */
1491             velec            = _fjsp_mul_v2r8(qq13,rinv13);
1492             felec            = _fjsp_mul_v2r8(velec,rinvsq13);
1493
1494             fscal            = felec;
1495
1496             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1497
1498             /* Update vectorial force */
1499             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1500             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1501             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1502             
1503             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1504             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1505             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1506
1507             /**************************
1508              * CALCULATE INTERACTIONS *
1509              **************************/
1510
1511             /* COULOMB ELECTROSTATICS */
1512             velec            = _fjsp_mul_v2r8(qq21,rinv21);
1513             felec            = _fjsp_mul_v2r8(velec,rinvsq21);
1514
1515             fscal            = felec;
1516
1517             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1518
1519             /* Update vectorial force */
1520             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1521             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1522             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1523             
1524             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1525             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1526             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1527
1528             /**************************
1529              * CALCULATE INTERACTIONS *
1530              **************************/
1531
1532             /* COULOMB ELECTROSTATICS */
1533             velec            = _fjsp_mul_v2r8(qq22,rinv22);
1534             felec            = _fjsp_mul_v2r8(velec,rinvsq22);
1535
1536             fscal            = felec;
1537
1538             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1539
1540             /* Update vectorial force */
1541             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1542             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1543             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1544             
1545             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1546             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1547             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1548
1549             /**************************
1550              * CALCULATE INTERACTIONS *
1551              **************************/
1552
1553             /* COULOMB ELECTROSTATICS */
1554             velec            = _fjsp_mul_v2r8(qq23,rinv23);
1555             felec            = _fjsp_mul_v2r8(velec,rinvsq23);
1556
1557             fscal            = felec;
1558
1559             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1560
1561             /* Update vectorial force */
1562             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1563             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1564             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1565             
1566             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1567             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1568             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1569
1570             /**************************
1571              * CALCULATE INTERACTIONS *
1572              **************************/
1573
1574             /* COULOMB ELECTROSTATICS */
1575             velec            = _fjsp_mul_v2r8(qq31,rinv31);
1576             felec            = _fjsp_mul_v2r8(velec,rinvsq31);
1577
1578             fscal            = felec;
1579
1580             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1581
1582             /* Update vectorial force */
1583             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1584             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1585             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1586             
1587             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1588             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1589             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1590
1591             /**************************
1592              * CALCULATE INTERACTIONS *
1593              **************************/
1594
1595             /* COULOMB ELECTROSTATICS */
1596             velec            = _fjsp_mul_v2r8(qq32,rinv32);
1597             felec            = _fjsp_mul_v2r8(velec,rinvsq32);
1598
1599             fscal            = felec;
1600
1601             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1602
1603             /* Update vectorial force */
1604             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1605             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1606             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1607             
1608             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1609             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1610             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1611
1612             /**************************
1613              * CALCULATE INTERACTIONS *
1614              **************************/
1615
1616             /* COULOMB ELECTROSTATICS */
1617             velec            = _fjsp_mul_v2r8(qq33,rinv33);
1618             felec            = _fjsp_mul_v2r8(velec,rinvsq33);
1619
1620             fscal            = felec;
1621
1622             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1623
1624             /* Update vectorial force */
1625             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1626             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1627             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1628             
1629             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1630             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1631             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1632
1633             gmx_fjsp_decrement_4rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1634
1635             /* Inner loop uses 303 flops */
1636         }
1637
1638         /* End of innermost loop */
1639
1640         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1641                                               f+i_coord_offset,fshift+i_shift_offset);
1642
1643         /* Increment number of inner iterations */
1644         inneriter                  += j_index_end - j_index_start;
1645
1646         /* Outer loop uses 24 flops */
1647     }
1648
1649     /* Increment number of outer iterations */
1650     outeriter        += nri;
1651
1652     /* Update outer/inner flops */
1653
1654     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1655 }