e0faa0da4b010b33096c22d1a307691b0d9ea351
[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sparc64_hpc_ace_double / nb_kernel_ElecRFCut_VdwLJSw_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_ElecRFCut_VdwLJSw_GeomW4W4_VF_sparc64_hpc_ace_double
51  * Electrostatics interaction: ReactionField
52  * VdW interaction:            LennardJones
53  * Geometry:                   Water4-Water4
54  * Calculate force/pot:        PotentialAndForce
55  */
56 void
57 nb_kernel_ElecRFCut_VdwLJSw_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       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
114     real             rswitch_scalar,d_scalar;
115     _fjsp_v2r8       itab_tmp;
116     _fjsp_v2r8       dummy_mask,cutoff_mask;
117     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
118     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
119     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
120
121     x                = xx[0];
122     f                = ff[0];
123
124     nri              = nlist->nri;
125     iinr             = nlist->iinr;
126     jindex           = nlist->jindex;
127     jjnr             = nlist->jjnr;
128     shiftidx         = nlist->shift;
129     gid              = nlist->gid;
130     shiftvec         = fr->shift_vec[0];
131     fshift           = fr->fshift[0];
132     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
133     charge           = mdatoms->chargeA;
134     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
135     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
136     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
137     nvdwtype         = fr->ntype;
138     vdwparam         = fr->nbfp;
139     vdwtype          = mdatoms->typeA;
140
141     /* Setup water-specific parameters */
142     inr              = nlist->iinr[0];
143     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
144     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
145     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
146     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
147
148     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
149     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
150     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
151     vdwjidx0A        = 2*vdwtype[inr+0];
152     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
153     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
154     qq11             = _fjsp_mul_v2r8(iq1,jq1);
155     qq12             = _fjsp_mul_v2r8(iq1,jq2);
156     qq13             = _fjsp_mul_v2r8(iq1,jq3);
157     qq21             = _fjsp_mul_v2r8(iq2,jq1);
158     qq22             = _fjsp_mul_v2r8(iq2,jq2);
159     qq23             = _fjsp_mul_v2r8(iq2,jq3);
160     qq31             = _fjsp_mul_v2r8(iq3,jq1);
161     qq32             = _fjsp_mul_v2r8(iq3,jq2);
162     qq33             = _fjsp_mul_v2r8(iq3,jq3);
163
164     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
165     rcutoff_scalar   = fr->rcoulomb;
166     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
167     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
168
169     rswitch_scalar   = fr->rvdw_switch;
170     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
171     /* Setup switch parameters */
172     d_scalar         = rcutoff_scalar-rswitch_scalar;
173     d                = gmx_fjsp_set1_v2r8(d_scalar);
174     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
175     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
176     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
177     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
178     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
179     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
180
181     /* Avoid stupid compiler warnings */
182     jnrA = jnrB = 0;
183     j_coord_offsetA = 0;
184     j_coord_offsetB = 0;
185
186     outeriter        = 0;
187     inneriter        = 0;
188
189     /* Start outer loop over neighborlists */
190     for(iidx=0; iidx<nri; iidx++)
191     {
192         /* Load shift vector for this list */
193         i_shift_offset   = DIM*shiftidx[iidx];
194
195         /* Load limits for loop over neighbors */
196         j_index_start    = jindex[iidx];
197         j_index_end      = jindex[iidx+1];
198
199         /* Get outer coordinate index */
200         inr              = iinr[iidx];
201         i_coord_offset   = DIM*inr;
202
203         /* Load i particle coords and add shift vector */
204         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
205                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
206
207         fix0             = _fjsp_setzero_v2r8();
208         fiy0             = _fjsp_setzero_v2r8();
209         fiz0             = _fjsp_setzero_v2r8();
210         fix1             = _fjsp_setzero_v2r8();
211         fiy1             = _fjsp_setzero_v2r8();
212         fiz1             = _fjsp_setzero_v2r8();
213         fix2             = _fjsp_setzero_v2r8();
214         fiy2             = _fjsp_setzero_v2r8();
215         fiz2             = _fjsp_setzero_v2r8();
216         fix3             = _fjsp_setzero_v2r8();
217         fiy3             = _fjsp_setzero_v2r8();
218         fiz3             = _fjsp_setzero_v2r8();
219
220         /* Reset potential sums */
221         velecsum         = _fjsp_setzero_v2r8();
222         vvdwsum          = _fjsp_setzero_v2r8();
223
224         /* Start inner kernel loop */
225         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
226         {
227
228             /* Get j neighbor index, and coordinate index */
229             jnrA             = jjnr[jidx];
230             jnrB             = jjnr[jidx+1];
231             j_coord_offsetA  = DIM*jnrA;
232             j_coord_offsetB  = DIM*jnrB;
233
234             /* load j atom coordinates */
235             gmx_fjsp_load_4rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
236                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
237                                               &jy2,&jz2,&jx3,&jy3,&jz3);
238
239             /* Calculate displacement vector */
240             dx00             = _fjsp_sub_v2r8(ix0,jx0);
241             dy00             = _fjsp_sub_v2r8(iy0,jy0);
242             dz00             = _fjsp_sub_v2r8(iz0,jz0);
243             dx11             = _fjsp_sub_v2r8(ix1,jx1);
244             dy11             = _fjsp_sub_v2r8(iy1,jy1);
245             dz11             = _fjsp_sub_v2r8(iz1,jz1);
246             dx12             = _fjsp_sub_v2r8(ix1,jx2);
247             dy12             = _fjsp_sub_v2r8(iy1,jy2);
248             dz12             = _fjsp_sub_v2r8(iz1,jz2);
249             dx13             = _fjsp_sub_v2r8(ix1,jx3);
250             dy13             = _fjsp_sub_v2r8(iy1,jy3);
251             dz13             = _fjsp_sub_v2r8(iz1,jz3);
252             dx21             = _fjsp_sub_v2r8(ix2,jx1);
253             dy21             = _fjsp_sub_v2r8(iy2,jy1);
254             dz21             = _fjsp_sub_v2r8(iz2,jz1);
255             dx22             = _fjsp_sub_v2r8(ix2,jx2);
256             dy22             = _fjsp_sub_v2r8(iy2,jy2);
257             dz22             = _fjsp_sub_v2r8(iz2,jz2);
258             dx23             = _fjsp_sub_v2r8(ix2,jx3);
259             dy23             = _fjsp_sub_v2r8(iy2,jy3);
260             dz23             = _fjsp_sub_v2r8(iz2,jz3);
261             dx31             = _fjsp_sub_v2r8(ix3,jx1);
262             dy31             = _fjsp_sub_v2r8(iy3,jy1);
263             dz31             = _fjsp_sub_v2r8(iz3,jz1);
264             dx32             = _fjsp_sub_v2r8(ix3,jx2);
265             dy32             = _fjsp_sub_v2r8(iy3,jy2);
266             dz32             = _fjsp_sub_v2r8(iz3,jz2);
267             dx33             = _fjsp_sub_v2r8(ix3,jx3);
268             dy33             = _fjsp_sub_v2r8(iy3,jy3);
269             dz33             = _fjsp_sub_v2r8(iz3,jz3);
270
271             /* Calculate squared distance and things based on it */
272             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
273             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
274             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
275             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
276             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
277             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
278             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
279             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
280             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
281             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
282
283             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
284             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
285             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
286             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
287             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
288             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
289             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
290             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
291             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
292             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
293
294             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
295             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
296             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
297             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
298             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
299             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
300             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
301             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
302             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
303             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
304
305             fjx0             = _fjsp_setzero_v2r8();
306             fjy0             = _fjsp_setzero_v2r8();
307             fjz0             = _fjsp_setzero_v2r8();
308             fjx1             = _fjsp_setzero_v2r8();
309             fjy1             = _fjsp_setzero_v2r8();
310             fjz1             = _fjsp_setzero_v2r8();
311             fjx2             = _fjsp_setzero_v2r8();
312             fjy2             = _fjsp_setzero_v2r8();
313             fjz2             = _fjsp_setzero_v2r8();
314             fjx3             = _fjsp_setzero_v2r8();
315             fjy3             = _fjsp_setzero_v2r8();
316             fjz3             = _fjsp_setzero_v2r8();
317
318             /**************************
319              * CALCULATE INTERACTIONS *
320              **************************/
321
322             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
323             {
324
325             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
326
327             /* LENNARD-JONES DISPERSION/REPULSION */
328
329             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
330             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
331             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
332             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
333             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
334
335             d                = _fjsp_sub_v2r8(r00,rswitch);
336             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
337             d2               = _fjsp_mul_v2r8(d,d);
338             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
339
340             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
341
342             /* Evaluate switch function */
343             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
344             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
345             vvdw             = _fjsp_mul_v2r8(vvdw,sw);
346             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
347
348             /* Update potential sum for this i atom from the interaction with this j atom. */
349             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
350             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
351
352             fscal            = fvdw;
353
354             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
355
356             /* Update vectorial force */
357             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
358             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
359             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
360             
361             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
362             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
363             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
364
365             }
366
367             /**************************
368              * CALCULATE INTERACTIONS *
369              **************************/
370
371             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
372             {
373
374             /* REACTION-FIELD ELECTROSTATICS */
375             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
376             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
377
378             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
379
380             /* Update potential sum for this i atom from the interaction with this j atom. */
381             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
382             velecsum         = _fjsp_add_v2r8(velecsum,velec);
383
384             fscal            = felec;
385
386             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
387
388             /* Update vectorial force */
389             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
390             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
391             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
392             
393             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
394             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
395             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
396
397             }
398
399             /**************************
400              * CALCULATE INTERACTIONS *
401              **************************/
402
403             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
404             {
405
406             /* REACTION-FIELD ELECTROSTATICS */
407             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
408             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
409
410             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
411
412             /* Update potential sum for this i atom from the interaction with this j atom. */
413             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
414             velecsum         = _fjsp_add_v2r8(velecsum,velec);
415
416             fscal            = felec;
417
418             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
419
420             /* Update vectorial force */
421             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
422             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
423             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
424             
425             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
426             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
427             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
428
429             }
430
431             /**************************
432              * CALCULATE INTERACTIONS *
433              **************************/
434
435             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
436             {
437
438             /* REACTION-FIELD ELECTROSTATICS */
439             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq13,rinv13),crf));
440             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
441
442             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
443
444             /* Update potential sum for this i atom from the interaction with this j atom. */
445             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
446             velecsum         = _fjsp_add_v2r8(velecsum,velec);
447
448             fscal            = felec;
449
450             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
451
452             /* Update vectorial force */
453             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
454             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
455             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
456             
457             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
458             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
459             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
460
461             }
462
463             /**************************
464              * CALCULATE INTERACTIONS *
465              **************************/
466
467             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
468             {
469
470             /* REACTION-FIELD ELECTROSTATICS */
471             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
472             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
473
474             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
475
476             /* Update potential sum for this i atom from the interaction with this j atom. */
477             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
478             velecsum         = _fjsp_add_v2r8(velecsum,velec);
479
480             fscal            = felec;
481
482             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
483
484             /* Update vectorial force */
485             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
486             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
487             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
488             
489             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
490             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
491             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
492
493             }
494
495             /**************************
496              * CALCULATE INTERACTIONS *
497              **************************/
498
499             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
500             {
501
502             /* REACTION-FIELD ELECTROSTATICS */
503             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
504             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
505
506             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
507
508             /* Update potential sum for this i atom from the interaction with this j atom. */
509             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
510             velecsum         = _fjsp_add_v2r8(velecsum,velec);
511
512             fscal            = felec;
513
514             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
515
516             /* Update vectorial force */
517             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
518             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
519             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
520             
521             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
522             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
523             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
524
525             }
526
527             /**************************
528              * CALCULATE INTERACTIONS *
529              **************************/
530
531             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
532             {
533
534             /* REACTION-FIELD ELECTROSTATICS */
535             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq23,rinv23),crf));
536             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
537
538             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
539
540             /* Update potential sum for this i atom from the interaction with this j atom. */
541             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
542             velecsum         = _fjsp_add_v2r8(velecsum,velec);
543
544             fscal            = felec;
545
546             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
547
548             /* Update vectorial force */
549             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
550             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
551             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
552             
553             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
554             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
555             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
556
557             }
558
559             /**************************
560              * CALCULATE INTERACTIONS *
561              **************************/
562
563             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
564             {
565
566             /* REACTION-FIELD ELECTROSTATICS */
567             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq31,rinv31),crf));
568             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
569
570             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
571
572             /* Update potential sum for this i atom from the interaction with this j atom. */
573             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
574             velecsum         = _fjsp_add_v2r8(velecsum,velec);
575
576             fscal            = felec;
577
578             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
579
580             /* Update vectorial force */
581             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
582             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
583             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
584             
585             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
586             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
587             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
588
589             }
590
591             /**************************
592              * CALCULATE INTERACTIONS *
593              **************************/
594
595             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
596             {
597
598             /* REACTION-FIELD ELECTROSTATICS */
599             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq32,rinv32),crf));
600             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
601
602             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
603
604             /* Update potential sum for this i atom from the interaction with this j atom. */
605             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
606             velecsum         = _fjsp_add_v2r8(velecsum,velec);
607
608             fscal            = felec;
609
610             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
611
612             /* Update vectorial force */
613             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
614             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
615             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
616             
617             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
618             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
619             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
620
621             }
622
623             /**************************
624              * CALCULATE INTERACTIONS *
625              **************************/
626
627             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
628             {
629
630             /* REACTION-FIELD ELECTROSTATICS */
631             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq33,rinv33),crf));
632             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
633
634             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
635
636             /* Update potential sum for this i atom from the interaction with this j atom. */
637             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
638             velecsum         = _fjsp_add_v2r8(velecsum,velec);
639
640             fscal            = felec;
641
642             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
643
644             /* Update vectorial force */
645             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
646             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
647             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
648             
649             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
650             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
651             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
652
653             }
654
655             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);
656
657             /* Inner loop uses 416 flops */
658         }
659
660         if(jidx<j_index_end)
661         {
662
663             jnrA             = jjnr[jidx];
664             j_coord_offsetA  = DIM*jnrA;
665
666             /* load j atom coordinates */
667             gmx_fjsp_load_4rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
668                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
669                                               &jy2,&jz2,&jx3,&jy3,&jz3);
670
671             /* Calculate displacement vector */
672             dx00             = _fjsp_sub_v2r8(ix0,jx0);
673             dy00             = _fjsp_sub_v2r8(iy0,jy0);
674             dz00             = _fjsp_sub_v2r8(iz0,jz0);
675             dx11             = _fjsp_sub_v2r8(ix1,jx1);
676             dy11             = _fjsp_sub_v2r8(iy1,jy1);
677             dz11             = _fjsp_sub_v2r8(iz1,jz1);
678             dx12             = _fjsp_sub_v2r8(ix1,jx2);
679             dy12             = _fjsp_sub_v2r8(iy1,jy2);
680             dz12             = _fjsp_sub_v2r8(iz1,jz2);
681             dx13             = _fjsp_sub_v2r8(ix1,jx3);
682             dy13             = _fjsp_sub_v2r8(iy1,jy3);
683             dz13             = _fjsp_sub_v2r8(iz1,jz3);
684             dx21             = _fjsp_sub_v2r8(ix2,jx1);
685             dy21             = _fjsp_sub_v2r8(iy2,jy1);
686             dz21             = _fjsp_sub_v2r8(iz2,jz1);
687             dx22             = _fjsp_sub_v2r8(ix2,jx2);
688             dy22             = _fjsp_sub_v2r8(iy2,jy2);
689             dz22             = _fjsp_sub_v2r8(iz2,jz2);
690             dx23             = _fjsp_sub_v2r8(ix2,jx3);
691             dy23             = _fjsp_sub_v2r8(iy2,jy3);
692             dz23             = _fjsp_sub_v2r8(iz2,jz3);
693             dx31             = _fjsp_sub_v2r8(ix3,jx1);
694             dy31             = _fjsp_sub_v2r8(iy3,jy1);
695             dz31             = _fjsp_sub_v2r8(iz3,jz1);
696             dx32             = _fjsp_sub_v2r8(ix3,jx2);
697             dy32             = _fjsp_sub_v2r8(iy3,jy2);
698             dz32             = _fjsp_sub_v2r8(iz3,jz2);
699             dx33             = _fjsp_sub_v2r8(ix3,jx3);
700             dy33             = _fjsp_sub_v2r8(iy3,jy3);
701             dz33             = _fjsp_sub_v2r8(iz3,jz3);
702
703             /* Calculate squared distance and things based on it */
704             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
705             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
706             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
707             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
708             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
709             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
710             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
711             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
712             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
713             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
714
715             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
716             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
717             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
718             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
719             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
720             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
721             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
722             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
723             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
724             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
725
726             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
727             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
728             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
729             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
730             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
731             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
732             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
733             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
734             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
735             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
736
737             fjx0             = _fjsp_setzero_v2r8();
738             fjy0             = _fjsp_setzero_v2r8();
739             fjz0             = _fjsp_setzero_v2r8();
740             fjx1             = _fjsp_setzero_v2r8();
741             fjy1             = _fjsp_setzero_v2r8();
742             fjz1             = _fjsp_setzero_v2r8();
743             fjx2             = _fjsp_setzero_v2r8();
744             fjy2             = _fjsp_setzero_v2r8();
745             fjz2             = _fjsp_setzero_v2r8();
746             fjx3             = _fjsp_setzero_v2r8();
747             fjy3             = _fjsp_setzero_v2r8();
748             fjz3             = _fjsp_setzero_v2r8();
749
750             /**************************
751              * CALCULATE INTERACTIONS *
752              **************************/
753
754             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
755             {
756
757             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
758
759             /* LENNARD-JONES DISPERSION/REPULSION */
760
761             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
762             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
763             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
764             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
765             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
766
767             d                = _fjsp_sub_v2r8(r00,rswitch);
768             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
769             d2               = _fjsp_mul_v2r8(d,d);
770             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
771
772             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
773
774             /* Evaluate switch function */
775             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
776             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
777             vvdw             = _fjsp_mul_v2r8(vvdw,sw);
778             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
779
780             /* Update potential sum for this i atom from the interaction with this j atom. */
781             vvdw             = _fjsp_and_v2r8(vvdw,cutoff_mask);
782             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
783             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
784
785             fscal            = fvdw;
786
787             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
788
789             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
790
791             /* Update vectorial force */
792             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
793             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
794             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
795             
796             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
797             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
798             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
799
800             }
801
802             /**************************
803              * CALCULATE INTERACTIONS *
804              **************************/
805
806             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
807             {
808
809             /* REACTION-FIELD ELECTROSTATICS */
810             velec            = _fjsp_mul_v2r8(qq11,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq11,rinv11),crf));
811             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
812
813             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
814
815             /* Update potential sum for this i atom from the interaction with this j atom. */
816             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
817             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
818             velecsum         = _fjsp_add_v2r8(velecsum,velec);
819
820             fscal            = felec;
821
822             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
823
824             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
825
826             /* Update vectorial force */
827             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
828             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
829             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
830             
831             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
832             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
833             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
834
835             }
836
837             /**************************
838              * CALCULATE INTERACTIONS *
839              **************************/
840
841             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
842             {
843
844             /* REACTION-FIELD ELECTROSTATICS */
845             velec            = _fjsp_mul_v2r8(qq12,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq12,rinv12),crf));
846             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
847
848             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
849
850             /* Update potential sum for this i atom from the interaction with this j atom. */
851             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
852             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
853             velecsum         = _fjsp_add_v2r8(velecsum,velec);
854
855             fscal            = felec;
856
857             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
858
859             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
860
861             /* Update vectorial force */
862             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
863             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
864             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
865             
866             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
867             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
868             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
869
870             }
871
872             /**************************
873              * CALCULATE INTERACTIONS *
874              **************************/
875
876             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
877             {
878
879             /* REACTION-FIELD ELECTROSTATICS */
880             velec            = _fjsp_mul_v2r8(qq13,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq13,rinv13),crf));
881             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
882
883             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
884
885             /* Update potential sum for this i atom from the interaction with this j atom. */
886             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
887             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
888             velecsum         = _fjsp_add_v2r8(velecsum,velec);
889
890             fscal            = felec;
891
892             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
893
894             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
895
896             /* Update vectorial force */
897             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
898             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
899             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
900             
901             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
902             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
903             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
904
905             }
906
907             /**************************
908              * CALCULATE INTERACTIONS *
909              **************************/
910
911             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
912             {
913
914             /* REACTION-FIELD ELECTROSTATICS */
915             velec            = _fjsp_mul_v2r8(qq21,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq21,rinv21),crf));
916             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
917
918             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
919
920             /* Update potential sum for this i atom from the interaction with this j atom. */
921             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
922             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
923             velecsum         = _fjsp_add_v2r8(velecsum,velec);
924
925             fscal            = felec;
926
927             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
928
929             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
930
931             /* Update vectorial force */
932             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
933             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
934             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
935             
936             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
937             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
938             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
939
940             }
941
942             /**************************
943              * CALCULATE INTERACTIONS *
944              **************************/
945
946             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
947             {
948
949             /* REACTION-FIELD ELECTROSTATICS */
950             velec            = _fjsp_mul_v2r8(qq22,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq22,rinv22),crf));
951             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
952
953             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
954
955             /* Update potential sum for this i atom from the interaction with this j atom. */
956             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
957             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
958             velecsum         = _fjsp_add_v2r8(velecsum,velec);
959
960             fscal            = felec;
961
962             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
963
964             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
965
966             /* Update vectorial force */
967             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
968             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
969             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
970             
971             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
972             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
973             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
974
975             }
976
977             /**************************
978              * CALCULATE INTERACTIONS *
979              **************************/
980
981             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
982             {
983
984             /* REACTION-FIELD ELECTROSTATICS */
985             velec            = _fjsp_mul_v2r8(qq23,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq23,rinv23),crf));
986             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
987
988             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
989
990             /* Update potential sum for this i atom from the interaction with this j atom. */
991             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
992             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
993             velecsum         = _fjsp_add_v2r8(velecsum,velec);
994
995             fscal            = felec;
996
997             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
998
999             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1000
1001             /* Update vectorial force */
1002             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1003             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1004             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1005             
1006             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1007             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1008             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1009
1010             }
1011
1012             /**************************
1013              * CALCULATE INTERACTIONS *
1014              **************************/
1015
1016             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
1017             {
1018
1019             /* REACTION-FIELD ELECTROSTATICS */
1020             velec            = _fjsp_mul_v2r8(qq31,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq31,rinv31),crf));
1021             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
1022
1023             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
1024
1025             /* Update potential sum for this i atom from the interaction with this j atom. */
1026             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1027             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1028             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1029
1030             fscal            = felec;
1031
1032             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1033
1034             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1035
1036             /* Update vectorial force */
1037             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1038             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1039             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1040             
1041             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1042             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1043             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1044
1045             }
1046
1047             /**************************
1048              * CALCULATE INTERACTIONS *
1049              **************************/
1050
1051             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
1052             {
1053
1054             /* REACTION-FIELD ELECTROSTATICS */
1055             velec            = _fjsp_mul_v2r8(qq32,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq32,rinv32),crf));
1056             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
1057
1058             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
1059
1060             /* Update potential sum for this i atom from the interaction with this j atom. */
1061             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1062             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1063             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1064
1065             fscal            = felec;
1066
1067             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1068
1069             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1070
1071             /* Update vectorial force */
1072             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1073             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1074             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1075             
1076             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1077             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1078             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1079
1080             }
1081
1082             /**************************
1083              * CALCULATE INTERACTIONS *
1084              **************************/
1085
1086             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
1087             {
1088
1089             /* REACTION-FIELD ELECTROSTATICS */
1090             velec            = _fjsp_mul_v2r8(qq33,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq33,rinv33),crf));
1091             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
1092
1093             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
1094
1095             /* Update potential sum for this i atom from the interaction with this j atom. */
1096             velec            = _fjsp_and_v2r8(velec,cutoff_mask);
1097             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
1098             velecsum         = _fjsp_add_v2r8(velecsum,velec);
1099
1100             fscal            = felec;
1101
1102             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1103
1104             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1105
1106             /* Update vectorial force */
1107             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1108             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1109             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1110             
1111             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1112             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1113             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1114
1115             }
1116
1117             gmx_fjsp_decrement_4rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1118
1119             /* Inner loop uses 416 flops */
1120         }
1121
1122         /* End of innermost loop */
1123
1124         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1125                                               f+i_coord_offset,fshift+i_shift_offset);
1126
1127         ggid                        = gid[iidx];
1128         /* Update potential energies */
1129         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
1130         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
1131
1132         /* Increment number of inner iterations */
1133         inneriter                  += j_index_end - j_index_start;
1134
1135         /* Outer loop uses 26 flops */
1136     }
1137
1138     /* Increment number of outer iterations */
1139     outeriter        += nri;
1140
1141     /* Update outer/inner flops */
1142
1143     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*416);
1144 }
1145 /*
1146  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sparc64_hpc_ace_double
1147  * Electrostatics interaction: ReactionField
1148  * VdW interaction:            LennardJones
1149  * Geometry:                   Water4-Water4
1150  * Calculate force/pot:        Force
1151  */
1152 void
1153 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sparc64_hpc_ace_double
1154                     (t_nblist                    * gmx_restrict       nlist,
1155                      rvec                        * gmx_restrict          xx,
1156                      rvec                        * gmx_restrict          ff,
1157                      t_forcerec                  * gmx_restrict          fr,
1158                      t_mdatoms                   * gmx_restrict     mdatoms,
1159                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1160                      t_nrnb                      * gmx_restrict        nrnb)
1161 {
1162     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1163      * just 0 for non-waters.
1164      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
1165      * jnr indices corresponding to data put in the four positions in the SIMD register.
1166      */
1167     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1168     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1169     int              jnrA,jnrB;
1170     int              j_coord_offsetA,j_coord_offsetB;
1171     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1172     real             rcutoff_scalar;
1173     real             *shiftvec,*fshift,*x,*f;
1174     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1175     int              vdwioffset0;
1176     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1177     int              vdwioffset1;
1178     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1179     int              vdwioffset2;
1180     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1181     int              vdwioffset3;
1182     _fjsp_v2r8       ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1183     int              vdwjidx0A,vdwjidx0B;
1184     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1185     int              vdwjidx1A,vdwjidx1B;
1186     _fjsp_v2r8       jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1187     int              vdwjidx2A,vdwjidx2B;
1188     _fjsp_v2r8       jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1189     int              vdwjidx3A,vdwjidx3B;
1190     _fjsp_v2r8       jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1191     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1192     _fjsp_v2r8       dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1193     _fjsp_v2r8       dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1194     _fjsp_v2r8       dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1195     _fjsp_v2r8       dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1196     _fjsp_v2r8       dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1197     _fjsp_v2r8       dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1198     _fjsp_v2r8       dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1199     _fjsp_v2r8       dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1200     _fjsp_v2r8       dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1201     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
1202     real             *charge;
1203     int              nvdwtype;
1204     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1205     int              *vdwtype;
1206     real             *vdwparam;
1207     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
1208     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
1209     _fjsp_v2r8       rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1210     real             rswitch_scalar,d_scalar;
1211     _fjsp_v2r8       itab_tmp;
1212     _fjsp_v2r8       dummy_mask,cutoff_mask;
1213     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
1214     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
1215     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
1216
1217     x                = xx[0];
1218     f                = ff[0];
1219
1220     nri              = nlist->nri;
1221     iinr             = nlist->iinr;
1222     jindex           = nlist->jindex;
1223     jjnr             = nlist->jjnr;
1224     shiftidx         = nlist->shift;
1225     gid              = nlist->gid;
1226     shiftvec         = fr->shift_vec[0];
1227     fshift           = fr->fshift[0];
1228     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
1229     charge           = mdatoms->chargeA;
1230     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
1231     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
1232     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
1233     nvdwtype         = fr->ntype;
1234     vdwparam         = fr->nbfp;
1235     vdwtype          = mdatoms->typeA;
1236
1237     /* Setup water-specific parameters */
1238     inr              = nlist->iinr[0];
1239     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
1240     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
1241     iq3              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+3]));
1242     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1243
1244     jq1              = gmx_fjsp_set1_v2r8(charge[inr+1]);
1245     jq2              = gmx_fjsp_set1_v2r8(charge[inr+2]);
1246     jq3              = gmx_fjsp_set1_v2r8(charge[inr+3]);
1247     vdwjidx0A        = 2*vdwtype[inr+0];
1248     c6_00            = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A]);
1249     c12_00           = gmx_fjsp_set1_v2r8(vdwparam[vdwioffset0+vdwjidx0A+1]);
1250     qq11             = _fjsp_mul_v2r8(iq1,jq1);
1251     qq12             = _fjsp_mul_v2r8(iq1,jq2);
1252     qq13             = _fjsp_mul_v2r8(iq1,jq3);
1253     qq21             = _fjsp_mul_v2r8(iq2,jq1);
1254     qq22             = _fjsp_mul_v2r8(iq2,jq2);
1255     qq23             = _fjsp_mul_v2r8(iq2,jq3);
1256     qq31             = _fjsp_mul_v2r8(iq3,jq1);
1257     qq32             = _fjsp_mul_v2r8(iq3,jq2);
1258     qq33             = _fjsp_mul_v2r8(iq3,jq3);
1259
1260     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1261     rcutoff_scalar   = fr->rcoulomb;
1262     rcutoff          = gmx_fjsp_set1_v2r8(rcutoff_scalar);
1263     rcutoff2         = _fjsp_mul_v2r8(rcutoff,rcutoff);
1264
1265     rswitch_scalar   = fr->rvdw_switch;
1266     rswitch          = gmx_fjsp_set1_v2r8(rswitch_scalar);
1267     /* Setup switch parameters */
1268     d_scalar         = rcutoff_scalar-rswitch_scalar;
1269     d                = gmx_fjsp_set1_v2r8(d_scalar);
1270     swV3             = gmx_fjsp_set1_v2r8(-10.0/(d_scalar*d_scalar*d_scalar));
1271     swV4             = gmx_fjsp_set1_v2r8( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1272     swV5             = gmx_fjsp_set1_v2r8( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1273     swF2             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar));
1274     swF3             = gmx_fjsp_set1_v2r8( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1275     swF4             = gmx_fjsp_set1_v2r8(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1276
1277     /* Avoid stupid compiler warnings */
1278     jnrA = jnrB = 0;
1279     j_coord_offsetA = 0;
1280     j_coord_offsetB = 0;
1281
1282     outeriter        = 0;
1283     inneriter        = 0;
1284
1285     /* Start outer loop over neighborlists */
1286     for(iidx=0; iidx<nri; iidx++)
1287     {
1288         /* Load shift vector for this list */
1289         i_shift_offset   = DIM*shiftidx[iidx];
1290
1291         /* Load limits for loop over neighbors */
1292         j_index_start    = jindex[iidx];
1293         j_index_end      = jindex[iidx+1];
1294
1295         /* Get outer coordinate index */
1296         inr              = iinr[iidx];
1297         i_coord_offset   = DIM*inr;
1298
1299         /* Load i particle coords and add shift vector */
1300         gmx_fjsp_load_shift_and_4rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
1301                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1302
1303         fix0             = _fjsp_setzero_v2r8();
1304         fiy0             = _fjsp_setzero_v2r8();
1305         fiz0             = _fjsp_setzero_v2r8();
1306         fix1             = _fjsp_setzero_v2r8();
1307         fiy1             = _fjsp_setzero_v2r8();
1308         fiz1             = _fjsp_setzero_v2r8();
1309         fix2             = _fjsp_setzero_v2r8();
1310         fiy2             = _fjsp_setzero_v2r8();
1311         fiz2             = _fjsp_setzero_v2r8();
1312         fix3             = _fjsp_setzero_v2r8();
1313         fiy3             = _fjsp_setzero_v2r8();
1314         fiz3             = _fjsp_setzero_v2r8();
1315
1316         /* Start inner kernel loop */
1317         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1318         {
1319
1320             /* Get j neighbor index, and coordinate index */
1321             jnrA             = jjnr[jidx];
1322             jnrB             = jjnr[jidx+1];
1323             j_coord_offsetA  = DIM*jnrA;
1324             j_coord_offsetB  = DIM*jnrB;
1325
1326             /* load j atom coordinates */
1327             gmx_fjsp_load_4rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
1328                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1329                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1330
1331             /* Calculate displacement vector */
1332             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1333             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1334             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1335             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1336             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1337             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1338             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1339             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1340             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1341             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1342             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1343             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1344             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1345             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1346             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1347             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1348             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1349             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1350             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1351             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1352             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1353             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1354             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1355             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1356             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1357             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1358             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1359             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1360             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1361             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1362
1363             /* Calculate squared distance and things based on it */
1364             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1365             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1366             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1367             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1368             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1369             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1370             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1371             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1372             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1373             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1374
1375             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1376             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1377             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1378             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1379             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1380             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1381             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1382             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1383             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1384             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1385
1386             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1387             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1388             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1389             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1390             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1391             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1392             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1393             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1394             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1395             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1396
1397             fjx0             = _fjsp_setzero_v2r8();
1398             fjy0             = _fjsp_setzero_v2r8();
1399             fjz0             = _fjsp_setzero_v2r8();
1400             fjx1             = _fjsp_setzero_v2r8();
1401             fjy1             = _fjsp_setzero_v2r8();
1402             fjz1             = _fjsp_setzero_v2r8();
1403             fjx2             = _fjsp_setzero_v2r8();
1404             fjy2             = _fjsp_setzero_v2r8();
1405             fjz2             = _fjsp_setzero_v2r8();
1406             fjx3             = _fjsp_setzero_v2r8();
1407             fjy3             = _fjsp_setzero_v2r8();
1408             fjz3             = _fjsp_setzero_v2r8();
1409
1410             /**************************
1411              * CALCULATE INTERACTIONS *
1412              **************************/
1413
1414             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1415             {
1416
1417             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1418
1419             /* LENNARD-JONES DISPERSION/REPULSION */
1420
1421             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1422             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
1423             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
1424             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
1425             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
1426
1427             d                = _fjsp_sub_v2r8(r00,rswitch);
1428             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1429             d2               = _fjsp_mul_v2r8(d,d);
1430             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1431
1432             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1433
1434             /* Evaluate switch function */
1435             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1436             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
1437             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1438
1439             fscal            = fvdw;
1440
1441             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1442
1443             /* Update vectorial force */
1444             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1445             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1446             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1447             
1448             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1449             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1450             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1451
1452             }
1453
1454             /**************************
1455              * CALCULATE INTERACTIONS *
1456              **************************/
1457
1458             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1459             {
1460
1461             /* REACTION-FIELD ELECTROSTATICS */
1462             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1463
1464             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1465
1466             fscal            = felec;
1467
1468             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1469
1470             /* Update vectorial force */
1471             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1472             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1473             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1474             
1475             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1476             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1477             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1478
1479             }
1480
1481             /**************************
1482              * CALCULATE INTERACTIONS *
1483              **************************/
1484
1485             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1486             {
1487
1488             /* REACTION-FIELD ELECTROSTATICS */
1489             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1490
1491             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1492
1493             fscal            = felec;
1494
1495             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1496
1497             /* Update vectorial force */
1498             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1499             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1500             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1501             
1502             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1503             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1504             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1505
1506             }
1507
1508             /**************************
1509              * CALCULATE INTERACTIONS *
1510              **************************/
1511
1512             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
1513             {
1514
1515             /* REACTION-FIELD ELECTROSTATICS */
1516             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
1517
1518             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
1519
1520             fscal            = felec;
1521
1522             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1523
1524             /* Update vectorial force */
1525             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1526             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1527             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1528             
1529             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1530             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1531             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1532
1533             }
1534
1535             /**************************
1536              * CALCULATE INTERACTIONS *
1537              **************************/
1538
1539             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1540             {
1541
1542             /* REACTION-FIELD ELECTROSTATICS */
1543             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1544
1545             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1546
1547             fscal            = felec;
1548
1549             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1550
1551             /* Update vectorial force */
1552             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1553             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1554             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1555             
1556             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1557             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1558             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1559
1560             }
1561
1562             /**************************
1563              * CALCULATE INTERACTIONS *
1564              **************************/
1565
1566             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1567             {
1568
1569             /* REACTION-FIELD ELECTROSTATICS */
1570             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1571
1572             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1573
1574             fscal            = felec;
1575
1576             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1577
1578             /* Update vectorial force */
1579             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1580             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1581             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1582             
1583             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1584             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1585             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1586
1587             }
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
1594             {
1595
1596             /* REACTION-FIELD ELECTROSTATICS */
1597             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
1598
1599             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
1600
1601             fscal            = felec;
1602
1603             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1604
1605             /* Update vectorial force */
1606             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
1607             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
1608             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
1609             
1610             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
1611             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
1612             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
1613
1614             }
1615
1616             /**************************
1617              * CALCULATE INTERACTIONS *
1618              **************************/
1619
1620             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
1621             {
1622
1623             /* REACTION-FIELD ELECTROSTATICS */
1624             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
1625
1626             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
1627
1628             fscal            = felec;
1629
1630             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1631
1632             /* Update vectorial force */
1633             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
1634             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
1635             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
1636             
1637             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
1638             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
1639             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
1640
1641             }
1642
1643             /**************************
1644              * CALCULATE INTERACTIONS *
1645              **************************/
1646
1647             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
1648             {
1649
1650             /* REACTION-FIELD ELECTROSTATICS */
1651             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
1652
1653             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
1654
1655             fscal            = felec;
1656
1657             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1658
1659             /* Update vectorial force */
1660             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
1661             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
1662             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
1663             
1664             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
1665             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
1666             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
1667
1668             }
1669
1670             /**************************
1671              * CALCULATE INTERACTIONS *
1672              **************************/
1673
1674             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
1675             {
1676
1677             /* REACTION-FIELD ELECTROSTATICS */
1678             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
1679
1680             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
1681
1682             fscal            = felec;
1683
1684             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1685
1686             /* Update vectorial force */
1687             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
1688             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
1689             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
1690             
1691             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
1692             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
1693             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
1694
1695             }
1696
1697             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);
1698
1699             /* Inner loop uses 359 flops */
1700         }
1701
1702         if(jidx<j_index_end)
1703         {
1704
1705             jnrA             = jjnr[jidx];
1706             j_coord_offsetA  = DIM*jnrA;
1707
1708             /* load j atom coordinates */
1709             gmx_fjsp_load_4rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
1710                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1711                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1712
1713             /* Calculate displacement vector */
1714             dx00             = _fjsp_sub_v2r8(ix0,jx0);
1715             dy00             = _fjsp_sub_v2r8(iy0,jy0);
1716             dz00             = _fjsp_sub_v2r8(iz0,jz0);
1717             dx11             = _fjsp_sub_v2r8(ix1,jx1);
1718             dy11             = _fjsp_sub_v2r8(iy1,jy1);
1719             dz11             = _fjsp_sub_v2r8(iz1,jz1);
1720             dx12             = _fjsp_sub_v2r8(ix1,jx2);
1721             dy12             = _fjsp_sub_v2r8(iy1,jy2);
1722             dz12             = _fjsp_sub_v2r8(iz1,jz2);
1723             dx13             = _fjsp_sub_v2r8(ix1,jx3);
1724             dy13             = _fjsp_sub_v2r8(iy1,jy3);
1725             dz13             = _fjsp_sub_v2r8(iz1,jz3);
1726             dx21             = _fjsp_sub_v2r8(ix2,jx1);
1727             dy21             = _fjsp_sub_v2r8(iy2,jy1);
1728             dz21             = _fjsp_sub_v2r8(iz2,jz1);
1729             dx22             = _fjsp_sub_v2r8(ix2,jx2);
1730             dy22             = _fjsp_sub_v2r8(iy2,jy2);
1731             dz22             = _fjsp_sub_v2r8(iz2,jz2);
1732             dx23             = _fjsp_sub_v2r8(ix2,jx3);
1733             dy23             = _fjsp_sub_v2r8(iy2,jy3);
1734             dz23             = _fjsp_sub_v2r8(iz2,jz3);
1735             dx31             = _fjsp_sub_v2r8(ix3,jx1);
1736             dy31             = _fjsp_sub_v2r8(iy3,jy1);
1737             dz31             = _fjsp_sub_v2r8(iz3,jz1);
1738             dx32             = _fjsp_sub_v2r8(ix3,jx2);
1739             dy32             = _fjsp_sub_v2r8(iy3,jy2);
1740             dz32             = _fjsp_sub_v2r8(iz3,jz2);
1741             dx33             = _fjsp_sub_v2r8(ix3,jx3);
1742             dy33             = _fjsp_sub_v2r8(iy3,jy3);
1743             dz33             = _fjsp_sub_v2r8(iz3,jz3);
1744
1745             /* Calculate squared distance and things based on it */
1746             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
1747             rsq11            = gmx_fjsp_calc_rsq_v2r8(dx11,dy11,dz11);
1748             rsq12            = gmx_fjsp_calc_rsq_v2r8(dx12,dy12,dz12);
1749             rsq13            = gmx_fjsp_calc_rsq_v2r8(dx13,dy13,dz13);
1750             rsq21            = gmx_fjsp_calc_rsq_v2r8(dx21,dy21,dz21);
1751             rsq22            = gmx_fjsp_calc_rsq_v2r8(dx22,dy22,dz22);
1752             rsq23            = gmx_fjsp_calc_rsq_v2r8(dx23,dy23,dz23);
1753             rsq31            = gmx_fjsp_calc_rsq_v2r8(dx31,dy31,dz31);
1754             rsq32            = gmx_fjsp_calc_rsq_v2r8(dx32,dy32,dz32);
1755             rsq33            = gmx_fjsp_calc_rsq_v2r8(dx33,dy33,dz33);
1756
1757             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
1758             rinv11           = gmx_fjsp_invsqrt_v2r8(rsq11);
1759             rinv12           = gmx_fjsp_invsqrt_v2r8(rsq12);
1760             rinv13           = gmx_fjsp_invsqrt_v2r8(rsq13);
1761             rinv21           = gmx_fjsp_invsqrt_v2r8(rsq21);
1762             rinv22           = gmx_fjsp_invsqrt_v2r8(rsq22);
1763             rinv23           = gmx_fjsp_invsqrt_v2r8(rsq23);
1764             rinv31           = gmx_fjsp_invsqrt_v2r8(rsq31);
1765             rinv32           = gmx_fjsp_invsqrt_v2r8(rsq32);
1766             rinv33           = gmx_fjsp_invsqrt_v2r8(rsq33);
1767
1768             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
1769             rinvsq11         = _fjsp_mul_v2r8(rinv11,rinv11);
1770             rinvsq12         = _fjsp_mul_v2r8(rinv12,rinv12);
1771             rinvsq13         = _fjsp_mul_v2r8(rinv13,rinv13);
1772             rinvsq21         = _fjsp_mul_v2r8(rinv21,rinv21);
1773             rinvsq22         = _fjsp_mul_v2r8(rinv22,rinv22);
1774             rinvsq23         = _fjsp_mul_v2r8(rinv23,rinv23);
1775             rinvsq31         = _fjsp_mul_v2r8(rinv31,rinv31);
1776             rinvsq32         = _fjsp_mul_v2r8(rinv32,rinv32);
1777             rinvsq33         = _fjsp_mul_v2r8(rinv33,rinv33);
1778
1779             fjx0             = _fjsp_setzero_v2r8();
1780             fjy0             = _fjsp_setzero_v2r8();
1781             fjz0             = _fjsp_setzero_v2r8();
1782             fjx1             = _fjsp_setzero_v2r8();
1783             fjy1             = _fjsp_setzero_v2r8();
1784             fjz1             = _fjsp_setzero_v2r8();
1785             fjx2             = _fjsp_setzero_v2r8();
1786             fjy2             = _fjsp_setzero_v2r8();
1787             fjz2             = _fjsp_setzero_v2r8();
1788             fjx3             = _fjsp_setzero_v2r8();
1789             fjy3             = _fjsp_setzero_v2r8();
1790             fjz3             = _fjsp_setzero_v2r8();
1791
1792             /**************************
1793              * CALCULATE INTERACTIONS *
1794              **************************/
1795
1796             if (gmx_fjsp_any_lt_v2r8(rsq00,rcutoff2))
1797             {
1798
1799             r00              = _fjsp_mul_v2r8(rsq00,rinv00);
1800
1801             /* LENNARD-JONES DISPERSION/REPULSION */
1802
1803             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
1804             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
1805             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
1806             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
1807             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
1808
1809             d                = _fjsp_sub_v2r8(r00,rswitch);
1810             d                = _fjsp_max_v2r8(d,_fjsp_setzero_v2r8());
1811             d2               = _fjsp_mul_v2r8(d,d);
1812             sw               = _fjsp_add_v2r8(one,_fjsp_mul_v2r8(d2,_fjsp_mul_v2r8(d,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swV5,swV4),swV3))));
1813
1814             dsw              = _fjsp_mul_v2r8(d2,_fjsp_madd_v2r8(d,_fjsp_madd_v2r8(d,swF4,swF3),swF2));
1815
1816             /* Evaluate switch function */
1817             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1818             fvdw             = _fjsp_msub_v2r8( fvdw,sw , _fjsp_mul_v2r8(rinv00,_fjsp_mul_v2r8(vvdw,dsw)) );
1819             cutoff_mask      = _fjsp_cmplt_v2r8(rsq00,rcutoff2);
1820
1821             fscal            = fvdw;
1822
1823             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1824
1825             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1826
1827             /* Update vectorial force */
1828             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
1829             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
1830             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
1831             
1832             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
1833             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
1834             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
1835
1836             }
1837
1838             /**************************
1839              * CALCULATE INTERACTIONS *
1840              **************************/
1841
1842             if (gmx_fjsp_any_lt_v2r8(rsq11,rcutoff2))
1843             {
1844
1845             /* REACTION-FIELD ELECTROSTATICS */
1846             felec            = _fjsp_mul_v2r8(qq11,_fjsp_msub_v2r8(rinv11,rinvsq11,krf2));
1847
1848             cutoff_mask      = _fjsp_cmplt_v2r8(rsq11,rcutoff2);
1849
1850             fscal            = felec;
1851
1852             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1853
1854             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1855
1856             /* Update vectorial force */
1857             fix1             = _fjsp_madd_v2r8(dx11,fscal,fix1);
1858             fiy1             = _fjsp_madd_v2r8(dy11,fscal,fiy1);
1859             fiz1             = _fjsp_madd_v2r8(dz11,fscal,fiz1);
1860             
1861             fjx1             = _fjsp_madd_v2r8(dx11,fscal,fjx1);
1862             fjy1             = _fjsp_madd_v2r8(dy11,fscal,fjy1);
1863             fjz1             = _fjsp_madd_v2r8(dz11,fscal,fjz1);
1864
1865             }
1866
1867             /**************************
1868              * CALCULATE INTERACTIONS *
1869              **************************/
1870
1871             if (gmx_fjsp_any_lt_v2r8(rsq12,rcutoff2))
1872             {
1873
1874             /* REACTION-FIELD ELECTROSTATICS */
1875             felec            = _fjsp_mul_v2r8(qq12,_fjsp_msub_v2r8(rinv12,rinvsq12,krf2));
1876
1877             cutoff_mask      = _fjsp_cmplt_v2r8(rsq12,rcutoff2);
1878
1879             fscal            = felec;
1880
1881             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1882
1883             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1884
1885             /* Update vectorial force */
1886             fix1             = _fjsp_madd_v2r8(dx12,fscal,fix1);
1887             fiy1             = _fjsp_madd_v2r8(dy12,fscal,fiy1);
1888             fiz1             = _fjsp_madd_v2r8(dz12,fscal,fiz1);
1889             
1890             fjx2             = _fjsp_madd_v2r8(dx12,fscal,fjx2);
1891             fjy2             = _fjsp_madd_v2r8(dy12,fscal,fjy2);
1892             fjz2             = _fjsp_madd_v2r8(dz12,fscal,fjz2);
1893
1894             }
1895
1896             /**************************
1897              * CALCULATE INTERACTIONS *
1898              **************************/
1899
1900             if (gmx_fjsp_any_lt_v2r8(rsq13,rcutoff2))
1901             {
1902
1903             /* REACTION-FIELD ELECTROSTATICS */
1904             felec            = _fjsp_mul_v2r8(qq13,_fjsp_msub_v2r8(rinv13,rinvsq13,krf2));
1905
1906             cutoff_mask      = _fjsp_cmplt_v2r8(rsq13,rcutoff2);
1907
1908             fscal            = felec;
1909
1910             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1911
1912             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1913
1914             /* Update vectorial force */
1915             fix1             = _fjsp_madd_v2r8(dx13,fscal,fix1);
1916             fiy1             = _fjsp_madd_v2r8(dy13,fscal,fiy1);
1917             fiz1             = _fjsp_madd_v2r8(dz13,fscal,fiz1);
1918             
1919             fjx3             = _fjsp_madd_v2r8(dx13,fscal,fjx3);
1920             fjy3             = _fjsp_madd_v2r8(dy13,fscal,fjy3);
1921             fjz3             = _fjsp_madd_v2r8(dz13,fscal,fjz3);
1922
1923             }
1924
1925             /**************************
1926              * CALCULATE INTERACTIONS *
1927              **************************/
1928
1929             if (gmx_fjsp_any_lt_v2r8(rsq21,rcutoff2))
1930             {
1931
1932             /* REACTION-FIELD ELECTROSTATICS */
1933             felec            = _fjsp_mul_v2r8(qq21,_fjsp_msub_v2r8(rinv21,rinvsq21,krf2));
1934
1935             cutoff_mask      = _fjsp_cmplt_v2r8(rsq21,rcutoff2);
1936
1937             fscal            = felec;
1938
1939             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1940
1941             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1942
1943             /* Update vectorial force */
1944             fix2             = _fjsp_madd_v2r8(dx21,fscal,fix2);
1945             fiy2             = _fjsp_madd_v2r8(dy21,fscal,fiy2);
1946             fiz2             = _fjsp_madd_v2r8(dz21,fscal,fiz2);
1947             
1948             fjx1             = _fjsp_madd_v2r8(dx21,fscal,fjx1);
1949             fjy1             = _fjsp_madd_v2r8(dy21,fscal,fjy1);
1950             fjz1             = _fjsp_madd_v2r8(dz21,fscal,fjz1);
1951
1952             }
1953
1954             /**************************
1955              * CALCULATE INTERACTIONS *
1956              **************************/
1957
1958             if (gmx_fjsp_any_lt_v2r8(rsq22,rcutoff2))
1959             {
1960
1961             /* REACTION-FIELD ELECTROSTATICS */
1962             felec            = _fjsp_mul_v2r8(qq22,_fjsp_msub_v2r8(rinv22,rinvsq22,krf2));
1963
1964             cutoff_mask      = _fjsp_cmplt_v2r8(rsq22,rcutoff2);
1965
1966             fscal            = felec;
1967
1968             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1969
1970             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
1971
1972             /* Update vectorial force */
1973             fix2             = _fjsp_madd_v2r8(dx22,fscal,fix2);
1974             fiy2             = _fjsp_madd_v2r8(dy22,fscal,fiy2);
1975             fiz2             = _fjsp_madd_v2r8(dz22,fscal,fiz2);
1976             
1977             fjx2             = _fjsp_madd_v2r8(dx22,fscal,fjx2);
1978             fjy2             = _fjsp_madd_v2r8(dy22,fscal,fjy2);
1979             fjz2             = _fjsp_madd_v2r8(dz22,fscal,fjz2);
1980
1981             }
1982
1983             /**************************
1984              * CALCULATE INTERACTIONS *
1985              **************************/
1986
1987             if (gmx_fjsp_any_lt_v2r8(rsq23,rcutoff2))
1988             {
1989
1990             /* REACTION-FIELD ELECTROSTATICS */
1991             felec            = _fjsp_mul_v2r8(qq23,_fjsp_msub_v2r8(rinv23,rinvsq23,krf2));
1992
1993             cutoff_mask      = _fjsp_cmplt_v2r8(rsq23,rcutoff2);
1994
1995             fscal            = felec;
1996
1997             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
1998
1999             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2000
2001             /* Update vectorial force */
2002             fix2             = _fjsp_madd_v2r8(dx23,fscal,fix2);
2003             fiy2             = _fjsp_madd_v2r8(dy23,fscal,fiy2);
2004             fiz2             = _fjsp_madd_v2r8(dz23,fscal,fiz2);
2005             
2006             fjx3             = _fjsp_madd_v2r8(dx23,fscal,fjx3);
2007             fjy3             = _fjsp_madd_v2r8(dy23,fscal,fjy3);
2008             fjz3             = _fjsp_madd_v2r8(dz23,fscal,fjz3);
2009
2010             }
2011
2012             /**************************
2013              * CALCULATE INTERACTIONS *
2014              **************************/
2015
2016             if (gmx_fjsp_any_lt_v2r8(rsq31,rcutoff2))
2017             {
2018
2019             /* REACTION-FIELD ELECTROSTATICS */
2020             felec            = _fjsp_mul_v2r8(qq31,_fjsp_msub_v2r8(rinv31,rinvsq31,krf2));
2021
2022             cutoff_mask      = _fjsp_cmplt_v2r8(rsq31,rcutoff2);
2023
2024             fscal            = felec;
2025
2026             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2027
2028             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2029
2030             /* Update vectorial force */
2031             fix3             = _fjsp_madd_v2r8(dx31,fscal,fix3);
2032             fiy3             = _fjsp_madd_v2r8(dy31,fscal,fiy3);
2033             fiz3             = _fjsp_madd_v2r8(dz31,fscal,fiz3);
2034             
2035             fjx1             = _fjsp_madd_v2r8(dx31,fscal,fjx1);
2036             fjy1             = _fjsp_madd_v2r8(dy31,fscal,fjy1);
2037             fjz1             = _fjsp_madd_v2r8(dz31,fscal,fjz1);
2038
2039             }
2040
2041             /**************************
2042              * CALCULATE INTERACTIONS *
2043              **************************/
2044
2045             if (gmx_fjsp_any_lt_v2r8(rsq32,rcutoff2))
2046             {
2047
2048             /* REACTION-FIELD ELECTROSTATICS */
2049             felec            = _fjsp_mul_v2r8(qq32,_fjsp_msub_v2r8(rinv32,rinvsq32,krf2));
2050
2051             cutoff_mask      = _fjsp_cmplt_v2r8(rsq32,rcutoff2);
2052
2053             fscal            = felec;
2054
2055             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2056
2057             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2058
2059             /* Update vectorial force */
2060             fix3             = _fjsp_madd_v2r8(dx32,fscal,fix3);
2061             fiy3             = _fjsp_madd_v2r8(dy32,fscal,fiy3);
2062             fiz3             = _fjsp_madd_v2r8(dz32,fscal,fiz3);
2063             
2064             fjx2             = _fjsp_madd_v2r8(dx32,fscal,fjx2);
2065             fjy2             = _fjsp_madd_v2r8(dy32,fscal,fjy2);
2066             fjz2             = _fjsp_madd_v2r8(dz32,fscal,fjz2);
2067
2068             }
2069
2070             /**************************
2071              * CALCULATE INTERACTIONS *
2072              **************************/
2073
2074             if (gmx_fjsp_any_lt_v2r8(rsq33,rcutoff2))
2075             {
2076
2077             /* REACTION-FIELD ELECTROSTATICS */
2078             felec            = _fjsp_mul_v2r8(qq33,_fjsp_msub_v2r8(rinv33,rinvsq33,krf2));
2079
2080             cutoff_mask      = _fjsp_cmplt_v2r8(rsq33,rcutoff2);
2081
2082             fscal            = felec;
2083
2084             fscal            = _fjsp_and_v2r8(fscal,cutoff_mask);
2085
2086             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
2087
2088             /* Update vectorial force */
2089             fix3             = _fjsp_madd_v2r8(dx33,fscal,fix3);
2090             fiy3             = _fjsp_madd_v2r8(dy33,fscal,fiy3);
2091             fiz3             = _fjsp_madd_v2r8(dz33,fscal,fiz3);
2092             
2093             fjx3             = _fjsp_madd_v2r8(dx33,fscal,fjx3);
2094             fjy3             = _fjsp_madd_v2r8(dy33,fscal,fjy3);
2095             fjz3             = _fjsp_madd_v2r8(dz33,fscal,fjz3);
2096
2097             }
2098
2099             gmx_fjsp_decrement_4rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2100
2101             /* Inner loop uses 359 flops */
2102         }
2103
2104         /* End of innermost loop */
2105
2106         gmx_fjsp_update_iforce_4atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2107                                               f+i_coord_offset,fshift+i_shift_offset);
2108
2109         /* Increment number of inner iterations */
2110         inneriter                  += j_index_end - j_index_start;
2111
2112         /* Outer loop uses 24 flops */
2113     }
2114
2115     /* Increment number of outer iterations */
2116     outeriter        += nri;
2117
2118     /* Update outer/inner flops */
2119
2120     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*359);
2121 }