src/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_sse4_1_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_sse4_1_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm_set1_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115     krf              = _mm_set1_ps(fr->ic->k_rf);
 116     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 117     crf              = _mm_set1_ps(fr->ic->c_rf);
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     /* Setup water-specific parameters */
 123     inr              = nlist->iinr[0];
 124     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 125     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 126     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 127     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 128
 129     jq0              = _mm_set1_ps(charge[inr+0]);
 130     jq1              = _mm_set1_ps(charge[inr+1]);
 131     jq2              = _mm_set1_ps(charge[inr+2]);
 132     vdwjidx0A        = 2*vdwtype[inr+0];
 133     qq00             = _mm_mul_ps(iq0,jq0);
 134     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 135     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 136     qq01             = _mm_mul_ps(iq0,jq1);
 137     qq02             = _mm_mul_ps(iq0,jq2);
 138     qq10             = _mm_mul_ps(iq1,jq0);
 139     qq11             = _mm_mul_ps(iq1,jq1);
 140     qq12             = _mm_mul_ps(iq1,jq2);
 141     qq20             = _mm_mul_ps(iq2,jq0);
 142     qq21             = _mm_mul_ps(iq2,jq1);
 143     qq22             = _mm_mul_ps(iq2,jq2);
 144
 145     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 146     rcutoff_scalar   = fr->rcoulomb;
 147     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 148     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 149
 150     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 151     rvdw             = _mm_set1_ps(fr->rvdw);
 152
 153     /* Avoid stupid compiler warnings */
 154     jnrA = jnrB = jnrC = jnrD = 0;
 155     j_coord_offsetA = 0;
 156     j_coord_offsetB = 0;
 157     j_coord_offsetC = 0;
 158     j_coord_offsetD = 0;
 159
 160     outeriter        = 0;
 161     inneriter        = 0;
 162
 163     for(iidx=0;iidx<4*DIM;iidx++)
 164     {
 165         scratch[iidx] = 0.0;
 166     }
 167
 168     /* Start outer loop over neighborlists */
 169     for(iidx=0; iidx<nri; iidx++)
 170     {
 171         /* Load shift vector for this list */
 172         i_shift_offset   = DIM*shiftidx[iidx];
 173
 174         /* Load limits for loop over neighbors */
 175         j_index_start    = jindex[iidx];
 176         j_index_end      = jindex[iidx+1];
 177
 178         /* Get outer coordinate index */
 179         inr              = iinr[iidx];
 180         i_coord_offset   = DIM*inr;
 181
 182         /* Load i particle coords and add shift vector */
 183         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 184                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 185
 186         fix0             = _mm_setzero_ps();
 187         fiy0             = _mm_setzero_ps();
 188         fiz0             = _mm_setzero_ps();
 189         fix1             = _mm_setzero_ps();
 190         fiy1             = _mm_setzero_ps();
 191         fiz1             = _mm_setzero_ps();
 192         fix2             = _mm_setzero_ps();
 193         fiy2             = _mm_setzero_ps();
 194         fiz2             = _mm_setzero_ps();
 195
 196         /* Reset potential sums */
 197         velecsum         = _mm_setzero_ps();
 198         vvdwsum          = _mm_setzero_ps();
 199
 200         /* Start inner kernel loop */
 201         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 202         {
 203
 204             /* Get j neighbor index, and coordinate index */
 205             jnrA             = jjnr[jidx];
 206             jnrB             = jjnr[jidx+1];
 207             jnrC             = jjnr[jidx+2];
 208             jnrD             = jjnr[jidx+3];
 209             j_coord_offsetA  = DIM*jnrA;
 210             j_coord_offsetB  = DIM*jnrB;
 211             j_coord_offsetC  = DIM*jnrC;
 212             j_coord_offsetD  = DIM*jnrD;
 213
 214             /* load j atom coordinates */
 215             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 216                                               x+j_coord_offsetC,x+j_coord_offsetD,
 217                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 218
 219             /* Calculate displacement vector */
 220             dx00             = _mm_sub_ps(ix0,jx0);
 221             dy00             = _mm_sub_ps(iy0,jy0);
 222             dz00             = _mm_sub_ps(iz0,jz0);
 223             dx01             = _mm_sub_ps(ix0,jx1);
 224             dy01             = _mm_sub_ps(iy0,jy1);
 225             dz01             = _mm_sub_ps(iz0,jz1);
 226             dx02             = _mm_sub_ps(ix0,jx2);
 227             dy02             = _mm_sub_ps(iy0,jy2);
 228             dz02             = _mm_sub_ps(iz0,jz2);
 229             dx10             = _mm_sub_ps(ix1,jx0);
 230             dy10             = _mm_sub_ps(iy1,jy0);
 231             dz10             = _mm_sub_ps(iz1,jz0);
 232             dx11             = _mm_sub_ps(ix1,jx1);
 233             dy11             = _mm_sub_ps(iy1,jy1);
 234             dz11             = _mm_sub_ps(iz1,jz1);
 235             dx12             = _mm_sub_ps(ix1,jx2);
 236             dy12             = _mm_sub_ps(iy1,jy2);
 237             dz12             = _mm_sub_ps(iz1,jz2);
 238             dx20             = _mm_sub_ps(ix2,jx0);
 239             dy20             = _mm_sub_ps(iy2,jy0);
 240             dz20             = _mm_sub_ps(iz2,jz0);
 241             dx21             = _mm_sub_ps(ix2,jx1);
 242             dy21             = _mm_sub_ps(iy2,jy1);
 243             dz21             = _mm_sub_ps(iz2,jz1);
 244             dx22             = _mm_sub_ps(ix2,jx2);
 245             dy22             = _mm_sub_ps(iy2,jy2);
 246             dz22             = _mm_sub_ps(iz2,jz2);
 247
 248             /* Calculate squared distance and things based on it */
 249             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 250             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 251             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 252             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 253             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 254             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 255             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 256             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 257             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 258
 259             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 260             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 261             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 262             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 263             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 264             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 265             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 266             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 267             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 268
 269             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 270             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 271             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 272             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 273             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 274             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 275             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 276             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 277             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 278
 279             fjx0             = _mm_setzero_ps();
 280             fjy0             = _mm_setzero_ps();
 281             fjz0             = _mm_setzero_ps();
 282             fjx1             = _mm_setzero_ps();
 283             fjy1             = _mm_setzero_ps();
 284             fjz1             = _mm_setzero_ps();
 285             fjx2             = _mm_setzero_ps();
 286             fjy2             = _mm_setzero_ps();
 287             fjz2             = _mm_setzero_ps();
 288
 289             /**************************
 290              * CALCULATE INTERACTIONS *
 291              **************************/
 292
 293             if (gmx_mm_any_lt(rsq00,rcutoff2))
 294             {
 295
 296             /* REACTION-FIELD ELECTROSTATICS */
 297             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 298             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 299
 300             /* LENNARD-JONES DISPERSION/REPULSION */
 301
 302             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 303             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 304             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 305             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 306                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 307             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 308
 309             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 310
 311             /* Update potential sum for this i atom from the interaction with this j atom. */
 312             velec            = _mm_and_ps(velec,cutoff_mask);
 313             velecsum         = _mm_add_ps(velecsum,velec);
 314             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 315             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 316
 317             fscal            = _mm_add_ps(felec,fvdw);
 318
 319             fscal            = _mm_and_ps(fscal,cutoff_mask);
 320
 321             /* Calculate temporary vectorial force */
 322             tx               = _mm_mul_ps(fscal,dx00);
 323             ty               = _mm_mul_ps(fscal,dy00);
 324             tz               = _mm_mul_ps(fscal,dz00);
 325
 326             /* Update vectorial force */
 327             fix0             = _mm_add_ps(fix0,tx);
 328             fiy0             = _mm_add_ps(fiy0,ty);
 329             fiz0             = _mm_add_ps(fiz0,tz);
 330
 331             fjx0             = _mm_add_ps(fjx0,tx);
 332             fjy0             = _mm_add_ps(fjy0,ty);
 333             fjz0             = _mm_add_ps(fjz0,tz);
 334
 335             }
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             if (gmx_mm_any_lt(rsq01,rcutoff2))
 342             {
 343
 344             /* REACTION-FIELD ELECTROSTATICS */
 345             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 346             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 347
 348             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 349
 350             /* Update potential sum for this i atom from the interaction with this j atom. */
 351             velec            = _mm_and_ps(velec,cutoff_mask);
 352             velecsum         = _mm_add_ps(velecsum,velec);
 353
 354             fscal            = felec;
 355
 356             fscal            = _mm_and_ps(fscal,cutoff_mask);
 357
 358             /* Calculate temporary vectorial force */
 359             tx               = _mm_mul_ps(fscal,dx01);
 360             ty               = _mm_mul_ps(fscal,dy01);
 361             tz               = _mm_mul_ps(fscal,dz01);
 362
 363             /* Update vectorial force */
 364             fix0             = _mm_add_ps(fix0,tx);
 365             fiy0             = _mm_add_ps(fiy0,ty);
 366             fiz0             = _mm_add_ps(fiz0,tz);
 367
 368             fjx1             = _mm_add_ps(fjx1,tx);
 369             fjy1             = _mm_add_ps(fjy1,ty);
 370             fjz1             = _mm_add_ps(fjz1,tz);
 371
 372             }
 373
 374             /**************************
 375              * CALCULATE INTERACTIONS *
 376              **************************/
 377
 378             if (gmx_mm_any_lt(rsq02,rcutoff2))
 379             {
 380
 381             /* REACTION-FIELD ELECTROSTATICS */
 382             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 383             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 384
 385             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velec            = _mm_and_ps(velec,cutoff_mask);
 389             velecsum         = _mm_add_ps(velecsum,velec);
 390
 391             fscal            = felec;
 392
 393             fscal            = _mm_and_ps(fscal,cutoff_mask);
 394
 395             /* Calculate temporary vectorial force */
 396             tx               = _mm_mul_ps(fscal,dx02);
 397             ty               = _mm_mul_ps(fscal,dy02);
 398             tz               = _mm_mul_ps(fscal,dz02);
 399
 400             /* Update vectorial force */
 401             fix0             = _mm_add_ps(fix0,tx);
 402             fiy0             = _mm_add_ps(fiy0,ty);
 403             fiz0             = _mm_add_ps(fiz0,tz);
 404
 405             fjx2             = _mm_add_ps(fjx2,tx);
 406             fjy2             = _mm_add_ps(fjy2,ty);
 407             fjz2             = _mm_add_ps(fjz2,tz);
 408
 409             }
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             if (gmx_mm_any_lt(rsq10,rcutoff2))
 416             {
 417
 418             /* REACTION-FIELD ELECTROSTATICS */
 419             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 420             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 421
 422             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 423
 424             /* Update potential sum for this i atom from the interaction with this j atom. */
 425             velec            = _mm_and_ps(velec,cutoff_mask);
 426             velecsum         = _mm_add_ps(velecsum,velec);
 427
 428             fscal            = felec;
 429
 430             fscal            = _mm_and_ps(fscal,cutoff_mask);
 431
 432             /* Calculate temporary vectorial force */
 433             tx               = _mm_mul_ps(fscal,dx10);
 434             ty               = _mm_mul_ps(fscal,dy10);
 435             tz               = _mm_mul_ps(fscal,dz10);
 436
 437             /* Update vectorial force */
 438             fix1             = _mm_add_ps(fix1,tx);
 439             fiy1             = _mm_add_ps(fiy1,ty);
 440             fiz1             = _mm_add_ps(fiz1,tz);
 441
 442             fjx0             = _mm_add_ps(fjx0,tx);
 443             fjy0             = _mm_add_ps(fjy0,ty);
 444             fjz0             = _mm_add_ps(fjz0,tz);
 445
 446             }
 447
 448             /**************************
 449              * CALCULATE INTERACTIONS *
 450              **************************/
 451
 452             if (gmx_mm_any_lt(rsq11,rcutoff2))
 453             {
 454
 455             /* REACTION-FIELD ELECTROSTATICS */
 456             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 457             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 458
 459             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             velec            = _mm_and_ps(velec,cutoff_mask);
 463             velecsum         = _mm_add_ps(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467             fscal            = _mm_and_ps(fscal,cutoff_mask);
 468
 469             /* Calculate temporary vectorial force */
 470             tx               = _mm_mul_ps(fscal,dx11);
 471             ty               = _mm_mul_ps(fscal,dy11);
 472             tz               = _mm_mul_ps(fscal,dz11);
 473
 474             /* Update vectorial force */
 475             fix1             = _mm_add_ps(fix1,tx);
 476             fiy1             = _mm_add_ps(fiy1,ty);
 477             fiz1             = _mm_add_ps(fiz1,tz);
 478
 479             fjx1             = _mm_add_ps(fjx1,tx);
 480             fjy1             = _mm_add_ps(fjy1,ty);
 481             fjz1             = _mm_add_ps(fjz1,tz);
 482
 483             }
 484
 485             /**************************
 486              * CALCULATE INTERACTIONS *
 487              **************************/
 488
 489             if (gmx_mm_any_lt(rsq12,rcutoff2))
 490             {
 491
 492             /* REACTION-FIELD ELECTROSTATICS */
 493             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 494             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 495
 496             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 497
 498             /* Update potential sum for this i atom from the interaction with this j atom. */
 499             velec            = _mm_and_ps(velec,cutoff_mask);
 500             velecsum         = _mm_add_ps(velecsum,velec);
 501
 502             fscal            = felec;
 503
 504             fscal            = _mm_and_ps(fscal,cutoff_mask);
 505
 506             /* Calculate temporary vectorial force */
 507             tx               = _mm_mul_ps(fscal,dx12);
 508             ty               = _mm_mul_ps(fscal,dy12);
 509             tz               = _mm_mul_ps(fscal,dz12);
 510
 511             /* Update vectorial force */
 512             fix1             = _mm_add_ps(fix1,tx);
 513             fiy1             = _mm_add_ps(fiy1,ty);
 514             fiz1             = _mm_add_ps(fiz1,tz);
 515
 516             fjx2             = _mm_add_ps(fjx2,tx);
 517             fjy2             = _mm_add_ps(fjy2,ty);
 518             fjz2             = _mm_add_ps(fjz2,tz);
 519
 520             }
 521
 522             /**************************
 523              * CALCULATE INTERACTIONS *
 524              **************************/
 525
 526             if (gmx_mm_any_lt(rsq20,rcutoff2))
 527             {
 528
 529             /* REACTION-FIELD ELECTROSTATICS */
 530             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 531             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 532
 533             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 534
 535             /* Update potential sum for this i atom from the interaction with this j atom. */
 536             velec            = _mm_and_ps(velec,cutoff_mask);
 537             velecsum         = _mm_add_ps(velecsum,velec);
 538
 539             fscal            = felec;
 540
 541             fscal            = _mm_and_ps(fscal,cutoff_mask);
 542
 543             /* Calculate temporary vectorial force */
 544             tx               = _mm_mul_ps(fscal,dx20);
 545             ty               = _mm_mul_ps(fscal,dy20);
 546             tz               = _mm_mul_ps(fscal,dz20);
 547
 548             /* Update vectorial force */
 549             fix2             = _mm_add_ps(fix2,tx);
 550             fiy2             = _mm_add_ps(fiy2,ty);
 551             fiz2             = _mm_add_ps(fiz2,tz);
 552
 553             fjx0             = _mm_add_ps(fjx0,tx);
 554             fjy0             = _mm_add_ps(fjy0,ty);
 555             fjz0             = _mm_add_ps(fjz0,tz);
 556
 557             }
 558
 559             /**************************
 560              * CALCULATE INTERACTIONS *
 561              **************************/
 562
 563             if (gmx_mm_any_lt(rsq21,rcutoff2))
 564             {
 565
 566             /* REACTION-FIELD ELECTROSTATICS */
 567             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 568             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 569
 570             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 571
 572             /* Update potential sum for this i atom from the interaction with this j atom. */
 573             velec            = _mm_and_ps(velec,cutoff_mask);
 574             velecsum         = _mm_add_ps(velecsum,velec);
 575
 576             fscal            = felec;
 577
 578             fscal            = _mm_and_ps(fscal,cutoff_mask);
 579
 580             /* Calculate temporary vectorial force */
 581             tx               = _mm_mul_ps(fscal,dx21);
 582             ty               = _mm_mul_ps(fscal,dy21);
 583             tz               = _mm_mul_ps(fscal,dz21);
 584
 585             /* Update vectorial force */
 586             fix2             = _mm_add_ps(fix2,tx);
 587             fiy2             = _mm_add_ps(fiy2,ty);
 588             fiz2             = _mm_add_ps(fiz2,tz);
 589
 590             fjx1             = _mm_add_ps(fjx1,tx);
 591             fjy1             = _mm_add_ps(fjy1,ty);
 592             fjz1             = _mm_add_ps(fjz1,tz);
 593
 594             }
 595
 596             /**************************
 597              * CALCULATE INTERACTIONS *
 598              **************************/
 599
 600             if (gmx_mm_any_lt(rsq22,rcutoff2))
 601             {
 602
 603             /* REACTION-FIELD ELECTROSTATICS */
 604             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 605             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 606
 607             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 608
 609             /* Update potential sum for this i atom from the interaction with this j atom. */
 610             velec            = _mm_and_ps(velec,cutoff_mask);
 611             velecsum         = _mm_add_ps(velecsum,velec);
 612
 613             fscal            = felec;
 614
 615             fscal            = _mm_and_ps(fscal,cutoff_mask);
 616
 617             /* Calculate temporary vectorial force */
 618             tx               = _mm_mul_ps(fscal,dx22);
 619             ty               = _mm_mul_ps(fscal,dy22);
 620             tz               = _mm_mul_ps(fscal,dz22);
 621
 622             /* Update vectorial force */
 623             fix2             = _mm_add_ps(fix2,tx);
 624             fiy2             = _mm_add_ps(fiy2,ty);
 625             fiz2             = _mm_add_ps(fiz2,tz);
 626
 627             fjx2             = _mm_add_ps(fjx2,tx);
 628             fjy2             = _mm_add_ps(fjy2,ty);
 629             fjz2             = _mm_add_ps(fjz2,tz);
 630
 631             }
 632
 633             fjptrA             = f+j_coord_offsetA;
 634             fjptrB             = f+j_coord_offsetB;
 635             fjptrC             = f+j_coord_offsetC;
 636             fjptrD             = f+j_coord_offsetD;
 637
 638             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 639                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 640
 641             /* Inner loop uses 342 flops */
 642         }
 643
 644         if(jidx<j_index_end)
 645         {
 646
 647             /* Get j neighbor index, and coordinate index */
 648             jnrlistA         = jjnr[jidx];
 649             jnrlistB         = jjnr[jidx+1];
 650             jnrlistC         = jjnr[jidx+2];
 651             jnrlistD         = jjnr[jidx+3];
 652             /* Sign of each element will be negative for non-real atoms.
 653              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 654              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 655              */
 656             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 657             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 658             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 659             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 660             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 661             j_coord_offsetA  = DIM*jnrA;
 662             j_coord_offsetB  = DIM*jnrB;
 663             j_coord_offsetC  = DIM*jnrC;
 664             j_coord_offsetD  = DIM*jnrD;
 665
 666             /* load j atom coordinates */
 667             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 668                                               x+j_coord_offsetC,x+j_coord_offsetD,
 669                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 670
 671             /* Calculate displacement vector */
 672             dx00             = _mm_sub_ps(ix0,jx0);
 673             dy00             = _mm_sub_ps(iy0,jy0);
 674             dz00             = _mm_sub_ps(iz0,jz0);
 675             dx01             = _mm_sub_ps(ix0,jx1);
 676             dy01             = _mm_sub_ps(iy0,jy1);
 677             dz01             = _mm_sub_ps(iz0,jz1);
 678             dx02             = _mm_sub_ps(ix0,jx2);
 679             dy02             = _mm_sub_ps(iy0,jy2);
 680             dz02             = _mm_sub_ps(iz0,jz2);
 681             dx10             = _mm_sub_ps(ix1,jx0);
 682             dy10             = _mm_sub_ps(iy1,jy0);
 683             dz10             = _mm_sub_ps(iz1,jz0);
 684             dx11             = _mm_sub_ps(ix1,jx1);
 685             dy11             = _mm_sub_ps(iy1,jy1);
 686             dz11             = _mm_sub_ps(iz1,jz1);
 687             dx12             = _mm_sub_ps(ix1,jx2);
 688             dy12             = _mm_sub_ps(iy1,jy2);
 689             dz12             = _mm_sub_ps(iz1,jz2);
 690             dx20             = _mm_sub_ps(ix2,jx0);
 691             dy20             = _mm_sub_ps(iy2,jy0);
 692             dz20             = _mm_sub_ps(iz2,jz0);
 693             dx21             = _mm_sub_ps(ix2,jx1);
 694             dy21             = _mm_sub_ps(iy2,jy1);
 695             dz21             = _mm_sub_ps(iz2,jz1);
 696             dx22             = _mm_sub_ps(ix2,jx2);
 697             dy22             = _mm_sub_ps(iy2,jy2);
 698             dz22             = _mm_sub_ps(iz2,jz2);
 699
 700             /* Calculate squared distance and things based on it */
 701             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 702             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 703             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 704             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 705             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 706             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 707             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 708             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 709             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 710
 711             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 712             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 713             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 714             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 715             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 716             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 717             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 718             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 719             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 720
 721             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 722             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 723             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 724             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 725             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 726             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 727             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 728             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 729             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 730
 731             fjx0             = _mm_setzero_ps();
 732             fjy0             = _mm_setzero_ps();
 733             fjz0             = _mm_setzero_ps();
 734             fjx1             = _mm_setzero_ps();
 735             fjy1             = _mm_setzero_ps();
 736             fjz1             = _mm_setzero_ps();
 737             fjx2             = _mm_setzero_ps();
 738             fjy2             = _mm_setzero_ps();
 739             fjz2             = _mm_setzero_ps();
 740
 741             /**************************
 742              * CALCULATE INTERACTIONS *
 743              **************************/
 744
 745             if (gmx_mm_any_lt(rsq00,rcutoff2))
 746             {
 747
 748             /* REACTION-FIELD ELECTROSTATICS */
 749             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 750             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 751
 752             /* LENNARD-JONES DISPERSION/REPULSION */
 753
 754             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 755             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 756             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 757             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 758                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 759             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 760
 761             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _mm_and_ps(velec,cutoff_mask);
 765             velec            = _mm_andnot_ps(dummy_mask,velec);
 766             velecsum         = _mm_add_ps(velecsum,velec);
 767             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 768             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 769             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 770
 771             fscal            = _mm_add_ps(felec,fvdw);
 772
 773             fscal            = _mm_and_ps(fscal,cutoff_mask);
 774
 775             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 776
 777             /* Calculate temporary vectorial force */
 778             tx               = _mm_mul_ps(fscal,dx00);
 779             ty               = _mm_mul_ps(fscal,dy00);
 780             tz               = _mm_mul_ps(fscal,dz00);
 781
 782             /* Update vectorial force */
 783             fix0             = _mm_add_ps(fix0,tx);
 784             fiy0             = _mm_add_ps(fiy0,ty);
 785             fiz0             = _mm_add_ps(fiz0,tz);
 786
 787             fjx0             = _mm_add_ps(fjx0,tx);
 788             fjy0             = _mm_add_ps(fjy0,ty);
 789             fjz0             = _mm_add_ps(fjz0,tz);
 790
 791             }
 792
 793             /**************************
 794              * CALCULATE INTERACTIONS *
 795              **************************/
 796
 797             if (gmx_mm_any_lt(rsq01,rcutoff2))
 798             {
 799
 800             /* REACTION-FIELD ELECTROSTATICS */
 801             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 802             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 803
 804             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 805
 806             /* Update potential sum for this i atom from the interaction with this j atom. */
 807             velec            = _mm_and_ps(velec,cutoff_mask);
 808             velec            = _mm_andnot_ps(dummy_mask,velec);
 809             velecsum         = _mm_add_ps(velecsum,velec);
 810
 811             fscal            = felec;
 812
 813             fscal            = _mm_and_ps(fscal,cutoff_mask);
 814
 815             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 816
 817             /* Calculate temporary vectorial force */
 818             tx               = _mm_mul_ps(fscal,dx01);
 819             ty               = _mm_mul_ps(fscal,dy01);
 820             tz               = _mm_mul_ps(fscal,dz01);
 821
 822             /* Update vectorial force */
 823             fix0             = _mm_add_ps(fix0,tx);
 824             fiy0             = _mm_add_ps(fiy0,ty);
 825             fiz0             = _mm_add_ps(fiz0,tz);
 826
 827             fjx1             = _mm_add_ps(fjx1,tx);
 828             fjy1             = _mm_add_ps(fjy1,ty);
 829             fjz1             = _mm_add_ps(fjz1,tz);
 830
 831             }
 832
 833             /**************************
 834              * CALCULATE INTERACTIONS *
 835              **************************/
 836
 837             if (gmx_mm_any_lt(rsq02,rcutoff2))
 838             {
 839
 840             /* REACTION-FIELD ELECTROSTATICS */
 841             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 842             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 843
 844             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 845
 846             /* Update potential sum for this i atom from the interaction with this j atom. */
 847             velec            = _mm_and_ps(velec,cutoff_mask);
 848             velec            = _mm_andnot_ps(dummy_mask,velec);
 849             velecsum         = _mm_add_ps(velecsum,velec);
 850
 851             fscal            = felec;
 852
 853             fscal            = _mm_and_ps(fscal,cutoff_mask);
 854
 855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 856
 857             /* Calculate temporary vectorial force */
 858             tx               = _mm_mul_ps(fscal,dx02);
 859             ty               = _mm_mul_ps(fscal,dy02);
 860             tz               = _mm_mul_ps(fscal,dz02);
 861
 862             /* Update vectorial force */
 863             fix0             = _mm_add_ps(fix0,tx);
 864             fiy0             = _mm_add_ps(fiy0,ty);
 865             fiz0             = _mm_add_ps(fiz0,tz);
 866
 867             fjx2             = _mm_add_ps(fjx2,tx);
 868             fjy2             = _mm_add_ps(fjy2,ty);
 869             fjz2             = _mm_add_ps(fjz2,tz);
 870
 871             }
 872
 873             /**************************
 874              * CALCULATE INTERACTIONS *
 875              **************************/
 876
 877             if (gmx_mm_any_lt(rsq10,rcutoff2))
 878             {
 879
 880             /* REACTION-FIELD ELECTROSTATICS */
 881             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 882             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 883
 884             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 885
 886             /* Update potential sum for this i atom from the interaction with this j atom. */
 887             velec            = _mm_and_ps(velec,cutoff_mask);
 888             velec            = _mm_andnot_ps(dummy_mask,velec);
 889             velecsum         = _mm_add_ps(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _mm_and_ps(fscal,cutoff_mask);
 894
 895             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 896
 897             /* Calculate temporary vectorial force */
 898             tx               = _mm_mul_ps(fscal,dx10);
 899             ty               = _mm_mul_ps(fscal,dy10);
 900             tz               = _mm_mul_ps(fscal,dz10);
 901
 902             /* Update vectorial force */
 903             fix1             = _mm_add_ps(fix1,tx);
 904             fiy1             = _mm_add_ps(fiy1,ty);
 905             fiz1             = _mm_add_ps(fiz1,tz);
 906
 907             fjx0             = _mm_add_ps(fjx0,tx);
 908             fjy0             = _mm_add_ps(fjy0,ty);
 909             fjz0             = _mm_add_ps(fjz0,tz);
 910
 911             }
 912
 913             /**************************
 914              * CALCULATE INTERACTIONS *
 915              **************************/
 916
 917             if (gmx_mm_any_lt(rsq11,rcutoff2))
 918             {
 919
 920             /* REACTION-FIELD ELECTROSTATICS */
 921             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 922             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 923
 924             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 925
 926             /* Update potential sum for this i atom from the interaction with this j atom. */
 927             velec            = _mm_and_ps(velec,cutoff_mask);
 928             velec            = _mm_andnot_ps(dummy_mask,velec);
 929             velecsum         = _mm_add_ps(velecsum,velec);
 930
 931             fscal            = felec;
 932
 933             fscal            = _mm_and_ps(fscal,cutoff_mask);
 934
 935             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 936
 937             /* Calculate temporary vectorial force */
 938             tx               = _mm_mul_ps(fscal,dx11);
 939             ty               = _mm_mul_ps(fscal,dy11);
 940             tz               = _mm_mul_ps(fscal,dz11);
 941
 942             /* Update vectorial force */
 943             fix1             = _mm_add_ps(fix1,tx);
 944             fiy1             = _mm_add_ps(fiy1,ty);
 945             fiz1             = _mm_add_ps(fiz1,tz);
 946
 947             fjx1             = _mm_add_ps(fjx1,tx);
 948             fjy1             = _mm_add_ps(fjy1,ty);
 949             fjz1             = _mm_add_ps(fjz1,tz);
 950
 951             }
 952
 953             /**************************
 954              * CALCULATE INTERACTIONS *
 955              **************************/
 956
 957             if (gmx_mm_any_lt(rsq12,rcutoff2))
 958             {
 959
 960             /* REACTION-FIELD ELECTROSTATICS */
 961             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 962             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 963
 964             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 965
 966             /* Update potential sum for this i atom from the interaction with this j atom. */
 967             velec            = _mm_and_ps(velec,cutoff_mask);
 968             velec            = _mm_andnot_ps(dummy_mask,velec);
 969             velecsum         = _mm_add_ps(velecsum,velec);
 970
 971             fscal            = felec;
 972
 973             fscal            = _mm_and_ps(fscal,cutoff_mask);
 974
 975             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 976
 977             /* Calculate temporary vectorial force */
 978             tx               = _mm_mul_ps(fscal,dx12);
 979             ty               = _mm_mul_ps(fscal,dy12);
 980             tz               = _mm_mul_ps(fscal,dz12);
 981
 982             /* Update vectorial force */
 983             fix1             = _mm_add_ps(fix1,tx);
 984             fiy1             = _mm_add_ps(fiy1,ty);
 985             fiz1             = _mm_add_ps(fiz1,tz);
 986
 987             fjx2             = _mm_add_ps(fjx2,tx);
 988             fjy2             = _mm_add_ps(fjy2,ty);
 989             fjz2             = _mm_add_ps(fjz2,tz);
 990
 991             }
 992
 993             /**************************
 994              * CALCULATE INTERACTIONS *
 995              **************************/
 996
 997             if (gmx_mm_any_lt(rsq20,rcutoff2))
 998             {
 999
1000             /* REACTION-FIELD ELECTROSTATICS */
1001             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
1002             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1003
1004             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1005
1006             /* Update potential sum for this i atom from the interaction with this j atom. */
1007             velec            = _mm_and_ps(velec,cutoff_mask);
1008             velec            = _mm_andnot_ps(dummy_mask,velec);
1009             velecsum         = _mm_add_ps(velecsum,velec);
1010
1011             fscal            = felec;
1012
1013             fscal            = _mm_and_ps(fscal,cutoff_mask);
1014
1015             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1016
1017             /* Calculate temporary vectorial force */
1018             tx               = _mm_mul_ps(fscal,dx20);
1019             ty               = _mm_mul_ps(fscal,dy20);
1020             tz               = _mm_mul_ps(fscal,dz20);
1021
1022             /* Update vectorial force */
1023             fix2             = _mm_add_ps(fix2,tx);
1024             fiy2             = _mm_add_ps(fiy2,ty);
1025             fiz2             = _mm_add_ps(fiz2,tz);
1026
1027             fjx0             = _mm_add_ps(fjx0,tx);
1028             fjy0             = _mm_add_ps(fjy0,ty);
1029             fjz0             = _mm_add_ps(fjz0,tz);
1030
1031             }
1032
1033             /**************************
1034              * CALCULATE INTERACTIONS *
1035              **************************/
1036
1037             if (gmx_mm_any_lt(rsq21,rcutoff2))
1038             {
1039
1040             /* REACTION-FIELD ELECTROSTATICS */
1041             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
1042             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1043
1044             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1045
1046             /* Update potential sum for this i atom from the interaction with this j atom. */
1047             velec            = _mm_and_ps(velec,cutoff_mask);
1048             velec            = _mm_andnot_ps(dummy_mask,velec);
1049             velecsum         = _mm_add_ps(velecsum,velec);
1050
1051             fscal            = felec;
1052
1053             fscal            = _mm_and_ps(fscal,cutoff_mask);
1054
1055             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1056
1057             /* Calculate temporary vectorial force */
1058             tx               = _mm_mul_ps(fscal,dx21);
1059             ty               = _mm_mul_ps(fscal,dy21);
1060             tz               = _mm_mul_ps(fscal,dz21);
1061
1062             /* Update vectorial force */
1063             fix2             = _mm_add_ps(fix2,tx);
1064             fiy2             = _mm_add_ps(fiy2,ty);
1065             fiz2             = _mm_add_ps(fiz2,tz);
1066
1067             fjx1             = _mm_add_ps(fjx1,tx);
1068             fjy1             = _mm_add_ps(fjy1,ty);
1069             fjz1             = _mm_add_ps(fjz1,tz);
1070
1071             }
1072
1073             /**************************
1074              * CALCULATE INTERACTIONS *
1075              **************************/
1076
1077             if (gmx_mm_any_lt(rsq22,rcutoff2))
1078             {
1079
1080             /* REACTION-FIELD ELECTROSTATICS */
1081             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1082             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1083
1084             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1085
1086             /* Update potential sum for this i atom from the interaction with this j atom. */
1087             velec            = _mm_and_ps(velec,cutoff_mask);
1088             velec            = _mm_andnot_ps(dummy_mask,velec);
1089             velecsum         = _mm_add_ps(velecsum,velec);
1090
1091             fscal            = felec;
1092
1093             fscal            = _mm_and_ps(fscal,cutoff_mask);
1094
1095             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1096
1097             /* Calculate temporary vectorial force */
1098             tx               = _mm_mul_ps(fscal,dx22);
1099             ty               = _mm_mul_ps(fscal,dy22);
1100             tz               = _mm_mul_ps(fscal,dz22);
1101
1102             /* Update vectorial force */
1103             fix2             = _mm_add_ps(fix2,tx);
1104             fiy2             = _mm_add_ps(fiy2,ty);
1105             fiz2             = _mm_add_ps(fiz2,tz);
1106
1107             fjx2             = _mm_add_ps(fjx2,tx);
1108             fjy2             = _mm_add_ps(fjy2,ty);
1109             fjz2             = _mm_add_ps(fjz2,tz);
1110
1111             }
1112
1113             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1114             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1115             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1116             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1117
1118             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1119                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1120
1121             /* Inner loop uses 342 flops */
1122         }
1123
1124         /* End of innermost loop */
1125
1126         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1127                                               f+i_coord_offset,fshift+i_shift_offset);
1128
1129         ggid                        = gid[iidx];
1130         /* Update potential energies */
1131         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1132         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1133
1134         /* Increment number of inner iterations */
1135         inneriter                  += j_index_end - j_index_start;
1136
1137         /* Outer loop uses 20 flops */
1138     }
1139
1140     /* Increment number of outer iterations */
1141     outeriter        += nri;
1142
1143     /* Update outer/inner flops */
1144
1145     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*342);
1146 }
1147 /*
1148  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sse4_1_single
1149  * Electrostatics interaction: ReactionField
1150  * VdW interaction:            LennardJones
1151  * Geometry:                   Water3-Water3
1152  * Calculate force/pot:        Force
1153  */
1154 void
1155 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sse4_1_single
1156                     (t_nblist * gmx_restrict                nlist,
1157                      rvec * gmx_restrict                    xx,
1158                      rvec * gmx_restrict                    ff,
1159                      t_forcerec * gmx_restrict              fr,
1160                      t_mdatoms * gmx_restrict               mdatoms,
1161                      nb_kernel_data_t * gmx_restrict        kernel_data,
1162                      t_nrnb * gmx_restrict                  nrnb)
1163 {
1164     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1165      * just 0 for non-waters.
1166      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1167      * jnr indices corresponding to data put in the four positions in the SIMD register.
1168      */
1169     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1170     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1171     int              jnrA,jnrB,jnrC,jnrD;
1172     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1173     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1174     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1175     real             rcutoff_scalar;
1176     real             *shiftvec,*fshift,*x,*f;
1177     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1178     real             scratch[4*DIM];
1179     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1180     int              vdwioffset0;
1181     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1182     int              vdwioffset1;
1183     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1184     int              vdwioffset2;
1185     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1186     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1187     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1188     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1189     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1190     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1191     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1192     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1193     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1194     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1195     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1196     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1197     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1198     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1199     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1200     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1201     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1202     real             *charge;
1203     int              nvdwtype;
1204     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1205     int              *vdwtype;
1206     real             *vdwparam;
1207     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1208     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1209     __m128           dummy_mask,cutoff_mask;
1210     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1211     __m128           one     = _mm_set1_ps(1.0);
1212     __m128           two     = _mm_set1_ps(2.0);
1213     x                = xx[0];
1214     f                = ff[0];
1215
1216     nri              = nlist->nri;
1217     iinr             = nlist->iinr;
1218     jindex           = nlist->jindex;
1219     jjnr             = nlist->jjnr;
1220     shiftidx         = nlist->shift;
1221     gid              = nlist->gid;
1222     shiftvec         = fr->shift_vec[0];
1223     fshift           = fr->fshift[0];
1224     facel            = _mm_set1_ps(fr->epsfac);
1225     charge           = mdatoms->chargeA;
1226     krf              = _mm_set1_ps(fr->ic->k_rf);
1227     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1228     crf              = _mm_set1_ps(fr->ic->c_rf);
1229     nvdwtype         = fr->ntype;
1230     vdwparam         = fr->nbfp;
1231     vdwtype          = mdatoms->typeA;
1232
1233     /* Setup water-specific parameters */
1234     inr              = nlist->iinr[0];
1235     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1236     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1237     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1238     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1239
1240     jq0              = _mm_set1_ps(charge[inr+0]);
1241     jq1              = _mm_set1_ps(charge[inr+1]);
1242     jq2              = _mm_set1_ps(charge[inr+2]);
1243     vdwjidx0A        = 2*vdwtype[inr+0];
1244     qq00             = _mm_mul_ps(iq0,jq0);
1245     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1246     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1247     qq01             = _mm_mul_ps(iq0,jq1);
1248     qq02             = _mm_mul_ps(iq0,jq2);
1249     qq10             = _mm_mul_ps(iq1,jq0);
1250     qq11             = _mm_mul_ps(iq1,jq1);
1251     qq12             = _mm_mul_ps(iq1,jq2);
1252     qq20             = _mm_mul_ps(iq2,jq0);
1253     qq21             = _mm_mul_ps(iq2,jq1);
1254     qq22             = _mm_mul_ps(iq2,jq2);
1255
1256     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1257     rcutoff_scalar   = fr->rcoulomb;
1258     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1259     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1260
1261     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1262     rvdw             = _mm_set1_ps(fr->rvdw);
1263
1264     /* Avoid stupid compiler warnings */
1265     jnrA = jnrB = jnrC = jnrD = 0;
1266     j_coord_offsetA = 0;
1267     j_coord_offsetB = 0;
1268     j_coord_offsetC = 0;
1269     j_coord_offsetD = 0;
1270
1271     outeriter        = 0;
1272     inneriter        = 0;
1273
1274     for(iidx=0;iidx<4*DIM;iidx++)
1275     {
1276         scratch[iidx] = 0.0;
1277     }
1278
1279     /* Start outer loop over neighborlists */
1280     for(iidx=0; iidx<nri; iidx++)
1281     {
1282         /* Load shift vector for this list */
1283         i_shift_offset   = DIM*shiftidx[iidx];
1284
1285         /* Load limits for loop over neighbors */
1286         j_index_start    = jindex[iidx];
1287         j_index_end      = jindex[iidx+1];
1288
1289         /* Get outer coordinate index */
1290         inr              = iinr[iidx];
1291         i_coord_offset   = DIM*inr;
1292
1293         /* Load i particle coords and add shift vector */
1294         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1295                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1296
1297         fix0             = _mm_setzero_ps();
1298         fiy0             = _mm_setzero_ps();
1299         fiz0             = _mm_setzero_ps();
1300         fix1             = _mm_setzero_ps();
1301         fiy1             = _mm_setzero_ps();
1302         fiz1             = _mm_setzero_ps();
1303         fix2             = _mm_setzero_ps();
1304         fiy2             = _mm_setzero_ps();
1305         fiz2             = _mm_setzero_ps();
1306
1307         /* Start inner kernel loop */
1308         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1309         {
1310
1311             /* Get j neighbor index, and coordinate index */
1312             jnrA             = jjnr[jidx];
1313             jnrB             = jjnr[jidx+1];
1314             jnrC             = jjnr[jidx+2];
1315             jnrD             = jjnr[jidx+3];
1316             j_coord_offsetA  = DIM*jnrA;
1317             j_coord_offsetB  = DIM*jnrB;
1318             j_coord_offsetC  = DIM*jnrC;
1319             j_coord_offsetD  = DIM*jnrD;
1320
1321             /* load j atom coordinates */
1322             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1323                                               x+j_coord_offsetC,x+j_coord_offsetD,
1324                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1325
1326             /* Calculate displacement vector */
1327             dx00             = _mm_sub_ps(ix0,jx0);
1328             dy00             = _mm_sub_ps(iy0,jy0);
1329             dz00             = _mm_sub_ps(iz0,jz0);
1330             dx01             = _mm_sub_ps(ix0,jx1);
1331             dy01             = _mm_sub_ps(iy0,jy1);
1332             dz01             = _mm_sub_ps(iz0,jz1);
1333             dx02             = _mm_sub_ps(ix0,jx2);
1334             dy02             = _mm_sub_ps(iy0,jy2);
1335             dz02             = _mm_sub_ps(iz0,jz2);
1336             dx10             = _mm_sub_ps(ix1,jx0);
1337             dy10             = _mm_sub_ps(iy1,jy0);
1338             dz10             = _mm_sub_ps(iz1,jz0);
1339             dx11             = _mm_sub_ps(ix1,jx1);
1340             dy11             = _mm_sub_ps(iy1,jy1);
1341             dz11             = _mm_sub_ps(iz1,jz1);
1342             dx12             = _mm_sub_ps(ix1,jx2);
1343             dy12             = _mm_sub_ps(iy1,jy2);
1344             dz12             = _mm_sub_ps(iz1,jz2);
1345             dx20             = _mm_sub_ps(ix2,jx0);
1346             dy20             = _mm_sub_ps(iy2,jy0);
1347             dz20             = _mm_sub_ps(iz2,jz0);
1348             dx21             = _mm_sub_ps(ix2,jx1);
1349             dy21             = _mm_sub_ps(iy2,jy1);
1350             dz21             = _mm_sub_ps(iz2,jz1);
1351             dx22             = _mm_sub_ps(ix2,jx2);
1352             dy22             = _mm_sub_ps(iy2,jy2);
1353             dz22             = _mm_sub_ps(iz2,jz2);
1354
1355             /* Calculate squared distance and things based on it */
1356             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1357             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1358             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1359             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1360             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1361             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1362             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1363             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1364             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1365
1366             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1367             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1368             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1369             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1370             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1371             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1372             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1373             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1374             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1375
1376             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1377             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1378             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1379             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1380             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1381             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1382             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1383             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1384             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1385
1386             fjx0             = _mm_setzero_ps();
1387             fjy0             = _mm_setzero_ps();
1388             fjz0             = _mm_setzero_ps();
1389             fjx1             = _mm_setzero_ps();
1390             fjy1             = _mm_setzero_ps();
1391             fjz1             = _mm_setzero_ps();
1392             fjx2             = _mm_setzero_ps();
1393             fjy2             = _mm_setzero_ps();
1394             fjz2             = _mm_setzero_ps();
1395
1396             /**************************
1397              * CALCULATE INTERACTIONS *
1398              **************************/
1399
1400             if (gmx_mm_any_lt(rsq00,rcutoff2))
1401             {
1402
1403             /* REACTION-FIELD ELECTROSTATICS */
1404             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1405
1406             /* LENNARD-JONES DISPERSION/REPULSION */
1407
1408             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1409             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1410
1411             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1412
1413             fscal            = _mm_add_ps(felec,fvdw);
1414
1415             fscal            = _mm_and_ps(fscal,cutoff_mask);
1416
1417             /* Calculate temporary vectorial force */
1418             tx               = _mm_mul_ps(fscal,dx00);
1419             ty               = _mm_mul_ps(fscal,dy00);
1420             tz               = _mm_mul_ps(fscal,dz00);
1421
1422             /* Update vectorial force */
1423             fix0             = _mm_add_ps(fix0,tx);
1424             fiy0             = _mm_add_ps(fiy0,ty);
1425             fiz0             = _mm_add_ps(fiz0,tz);
1426
1427             fjx0             = _mm_add_ps(fjx0,tx);
1428             fjy0             = _mm_add_ps(fjy0,ty);
1429             fjz0             = _mm_add_ps(fjz0,tz);
1430
1431             }
1432
1433             /**************************
1434              * CALCULATE INTERACTIONS *
1435              **************************/
1436
1437             if (gmx_mm_any_lt(rsq01,rcutoff2))
1438             {
1439
1440             /* REACTION-FIELD ELECTROSTATICS */
1441             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1442
1443             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1444
1445             fscal            = felec;
1446
1447             fscal            = _mm_and_ps(fscal,cutoff_mask);
1448
1449             /* Calculate temporary vectorial force */
1450             tx               = _mm_mul_ps(fscal,dx01);
1451             ty               = _mm_mul_ps(fscal,dy01);
1452             tz               = _mm_mul_ps(fscal,dz01);
1453
1454             /* Update vectorial force */
1455             fix0             = _mm_add_ps(fix0,tx);
1456             fiy0             = _mm_add_ps(fiy0,ty);
1457             fiz0             = _mm_add_ps(fiz0,tz);
1458
1459             fjx1             = _mm_add_ps(fjx1,tx);
1460             fjy1             = _mm_add_ps(fjy1,ty);
1461             fjz1             = _mm_add_ps(fjz1,tz);
1462
1463             }
1464
1465             /**************************
1466              * CALCULATE INTERACTIONS *
1467              **************************/
1468
1469             if (gmx_mm_any_lt(rsq02,rcutoff2))
1470             {
1471
1472             /* REACTION-FIELD ELECTROSTATICS */
1473             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1474
1475             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1476
1477             fscal            = felec;
1478
1479             fscal            = _mm_and_ps(fscal,cutoff_mask);
1480
1481             /* Calculate temporary vectorial force */
1482             tx               = _mm_mul_ps(fscal,dx02);
1483             ty               = _mm_mul_ps(fscal,dy02);
1484             tz               = _mm_mul_ps(fscal,dz02);
1485
1486             /* Update vectorial force */
1487             fix0             = _mm_add_ps(fix0,tx);
1488             fiy0             = _mm_add_ps(fiy0,ty);
1489             fiz0             = _mm_add_ps(fiz0,tz);
1490
1491             fjx2             = _mm_add_ps(fjx2,tx);
1492             fjy2             = _mm_add_ps(fjy2,ty);
1493             fjz2             = _mm_add_ps(fjz2,tz);
1494
1495             }
1496
1497             /**************************
1498              * CALCULATE INTERACTIONS *
1499              **************************/
1500
1501             if (gmx_mm_any_lt(rsq10,rcutoff2))
1502             {
1503
1504             /* REACTION-FIELD ELECTROSTATICS */
1505             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1506
1507             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1508
1509             fscal            = felec;
1510
1511             fscal            = _mm_and_ps(fscal,cutoff_mask);
1512
1513             /* Calculate temporary vectorial force */
1514             tx               = _mm_mul_ps(fscal,dx10);
1515             ty               = _mm_mul_ps(fscal,dy10);
1516             tz               = _mm_mul_ps(fscal,dz10);
1517
1518             /* Update vectorial force */
1519             fix1             = _mm_add_ps(fix1,tx);
1520             fiy1             = _mm_add_ps(fiy1,ty);
1521             fiz1             = _mm_add_ps(fiz1,tz);
1522
1523             fjx0             = _mm_add_ps(fjx0,tx);
1524             fjy0             = _mm_add_ps(fjy0,ty);
1525             fjz0             = _mm_add_ps(fjz0,tz);
1526
1527             }
1528
1529             /**************************
1530              * CALCULATE INTERACTIONS *
1531              **************************/
1532
1533             if (gmx_mm_any_lt(rsq11,rcutoff2))
1534             {
1535
1536             /* REACTION-FIELD ELECTROSTATICS */
1537             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1538
1539             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1540
1541             fscal            = felec;
1542
1543             fscal            = _mm_and_ps(fscal,cutoff_mask);
1544
1545             /* Calculate temporary vectorial force */
1546             tx               = _mm_mul_ps(fscal,dx11);
1547             ty               = _mm_mul_ps(fscal,dy11);
1548             tz               = _mm_mul_ps(fscal,dz11);
1549
1550             /* Update vectorial force */
1551             fix1             = _mm_add_ps(fix1,tx);
1552             fiy1             = _mm_add_ps(fiy1,ty);
1553             fiz1             = _mm_add_ps(fiz1,tz);
1554
1555             fjx1             = _mm_add_ps(fjx1,tx);
1556             fjy1             = _mm_add_ps(fjy1,ty);
1557             fjz1             = _mm_add_ps(fjz1,tz);
1558
1559             }
1560
1561             /**************************
1562              * CALCULATE INTERACTIONS *
1563              **************************/
1564
1565             if (gmx_mm_any_lt(rsq12,rcutoff2))
1566             {
1567
1568             /* REACTION-FIELD ELECTROSTATICS */
1569             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1570
1571             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1572
1573             fscal            = felec;
1574
1575             fscal            = _mm_and_ps(fscal,cutoff_mask);
1576
1577             /* Calculate temporary vectorial force */
1578             tx               = _mm_mul_ps(fscal,dx12);
1579             ty               = _mm_mul_ps(fscal,dy12);
1580             tz               = _mm_mul_ps(fscal,dz12);
1581
1582             /* Update vectorial force */
1583             fix1             = _mm_add_ps(fix1,tx);
1584             fiy1             = _mm_add_ps(fiy1,ty);
1585             fiz1             = _mm_add_ps(fiz1,tz);
1586
1587             fjx2             = _mm_add_ps(fjx2,tx);
1588             fjy2             = _mm_add_ps(fjy2,ty);
1589             fjz2             = _mm_add_ps(fjz2,tz);
1590
1591             }
1592
1593             /**************************
1594              * CALCULATE INTERACTIONS *
1595              **************************/
1596
1597             if (gmx_mm_any_lt(rsq20,rcutoff2))
1598             {
1599
1600             /* REACTION-FIELD ELECTROSTATICS */
1601             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1602
1603             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1604
1605             fscal            = felec;
1606
1607             fscal            = _mm_and_ps(fscal,cutoff_mask);
1608
1609             /* Calculate temporary vectorial force */
1610             tx               = _mm_mul_ps(fscal,dx20);
1611             ty               = _mm_mul_ps(fscal,dy20);
1612             tz               = _mm_mul_ps(fscal,dz20);
1613
1614             /* Update vectorial force */
1615             fix2             = _mm_add_ps(fix2,tx);
1616             fiy2             = _mm_add_ps(fiy2,ty);
1617             fiz2             = _mm_add_ps(fiz2,tz);
1618
1619             fjx0             = _mm_add_ps(fjx0,tx);
1620             fjy0             = _mm_add_ps(fjy0,ty);
1621             fjz0             = _mm_add_ps(fjz0,tz);
1622
1623             }
1624
1625             /**************************
1626              * CALCULATE INTERACTIONS *
1627              **************************/
1628
1629             if (gmx_mm_any_lt(rsq21,rcutoff2))
1630             {
1631
1632             /* REACTION-FIELD ELECTROSTATICS */
1633             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1634
1635             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1636
1637             fscal            = felec;
1638
1639             fscal            = _mm_and_ps(fscal,cutoff_mask);
1640
1641             /* Calculate temporary vectorial force */
1642             tx               = _mm_mul_ps(fscal,dx21);
1643             ty               = _mm_mul_ps(fscal,dy21);
1644             tz               = _mm_mul_ps(fscal,dz21);
1645
1646             /* Update vectorial force */
1647             fix2             = _mm_add_ps(fix2,tx);
1648             fiy2             = _mm_add_ps(fiy2,ty);
1649             fiz2             = _mm_add_ps(fiz2,tz);
1650
1651             fjx1             = _mm_add_ps(fjx1,tx);
1652             fjy1             = _mm_add_ps(fjy1,ty);
1653             fjz1             = _mm_add_ps(fjz1,tz);
1654
1655             }
1656
1657             /**************************
1658              * CALCULATE INTERACTIONS *
1659              **************************/
1660
1661             if (gmx_mm_any_lt(rsq22,rcutoff2))
1662             {
1663
1664             /* REACTION-FIELD ELECTROSTATICS */
1665             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1666
1667             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1668
1669             fscal            = felec;
1670
1671             fscal            = _mm_and_ps(fscal,cutoff_mask);
1672
1673             /* Calculate temporary vectorial force */
1674             tx               = _mm_mul_ps(fscal,dx22);
1675             ty               = _mm_mul_ps(fscal,dy22);
1676             tz               = _mm_mul_ps(fscal,dz22);
1677
1678             /* Update vectorial force */
1679             fix2             = _mm_add_ps(fix2,tx);
1680             fiy2             = _mm_add_ps(fiy2,ty);
1681             fiz2             = _mm_add_ps(fiz2,tz);
1682
1683             fjx2             = _mm_add_ps(fjx2,tx);
1684             fjy2             = _mm_add_ps(fjy2,ty);
1685             fjz2             = _mm_add_ps(fjz2,tz);
1686
1687             }
1688
1689             fjptrA             = f+j_coord_offsetA;
1690             fjptrB             = f+j_coord_offsetB;
1691             fjptrC             = f+j_coord_offsetC;
1692             fjptrD             = f+j_coord_offsetD;
1693
1694             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1695                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1696
1697             /* Inner loop uses 277 flops */
1698         }
1699
1700         if(jidx<j_index_end)
1701         {
1702
1703             /* Get j neighbor index, and coordinate index */
1704             jnrlistA         = jjnr[jidx];
1705             jnrlistB         = jjnr[jidx+1];
1706             jnrlistC         = jjnr[jidx+2];
1707             jnrlistD         = jjnr[jidx+3];
1708             /* Sign of each element will be negative for non-real atoms.
1709              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1710              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1711              */
1712             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1713             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1714             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1715             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1716             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1717             j_coord_offsetA  = DIM*jnrA;
1718             j_coord_offsetB  = DIM*jnrB;
1719             j_coord_offsetC  = DIM*jnrC;
1720             j_coord_offsetD  = DIM*jnrD;
1721
1722             /* load j atom coordinates */
1723             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1724                                               x+j_coord_offsetC,x+j_coord_offsetD,
1725                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1726
1727             /* Calculate displacement vector */
1728             dx00             = _mm_sub_ps(ix0,jx0);
1729             dy00             = _mm_sub_ps(iy0,jy0);
1730             dz00             = _mm_sub_ps(iz0,jz0);
1731             dx01             = _mm_sub_ps(ix0,jx1);
1732             dy01             = _mm_sub_ps(iy0,jy1);
1733             dz01             = _mm_sub_ps(iz0,jz1);
1734             dx02             = _mm_sub_ps(ix0,jx2);
1735             dy02             = _mm_sub_ps(iy0,jy2);
1736             dz02             = _mm_sub_ps(iz0,jz2);
1737             dx10             = _mm_sub_ps(ix1,jx0);
1738             dy10             = _mm_sub_ps(iy1,jy0);
1739             dz10             = _mm_sub_ps(iz1,jz0);
1740             dx11             = _mm_sub_ps(ix1,jx1);
1741             dy11             = _mm_sub_ps(iy1,jy1);
1742             dz11             = _mm_sub_ps(iz1,jz1);
1743             dx12             = _mm_sub_ps(ix1,jx2);
1744             dy12             = _mm_sub_ps(iy1,jy2);
1745             dz12             = _mm_sub_ps(iz1,jz2);
1746             dx20             = _mm_sub_ps(ix2,jx0);
1747             dy20             = _mm_sub_ps(iy2,jy0);
1748             dz20             = _mm_sub_ps(iz2,jz0);
1749             dx21             = _mm_sub_ps(ix2,jx1);
1750             dy21             = _mm_sub_ps(iy2,jy1);
1751             dz21             = _mm_sub_ps(iz2,jz1);
1752             dx22             = _mm_sub_ps(ix2,jx2);
1753             dy22             = _mm_sub_ps(iy2,jy2);
1754             dz22             = _mm_sub_ps(iz2,jz2);
1755
1756             /* Calculate squared distance and things based on it */
1757             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1758             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1759             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1760             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1761             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1762             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1763             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1764             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1765             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1766
1767             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1768             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1769             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1770             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1771             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1772             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1773             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1774             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1775             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1776
1777             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1778             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1779             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1780             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1781             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1782             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1783             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1784             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1785             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1786
1787             fjx0             = _mm_setzero_ps();
1788             fjy0             = _mm_setzero_ps();
1789             fjz0             = _mm_setzero_ps();
1790             fjx1             = _mm_setzero_ps();
1791             fjy1             = _mm_setzero_ps();
1792             fjz1             = _mm_setzero_ps();
1793             fjx2             = _mm_setzero_ps();
1794             fjy2             = _mm_setzero_ps();
1795             fjz2             = _mm_setzero_ps();
1796
1797             /**************************
1798              * CALCULATE INTERACTIONS *
1799              **************************/
1800
1801             if (gmx_mm_any_lt(rsq00,rcutoff2))
1802             {
1803
1804             /* REACTION-FIELD ELECTROSTATICS */
1805             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1806
1807             /* LENNARD-JONES DISPERSION/REPULSION */
1808
1809             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1810             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1811
1812             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1813
1814             fscal            = _mm_add_ps(felec,fvdw);
1815
1816             fscal            = _mm_and_ps(fscal,cutoff_mask);
1817
1818             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1819
1820             /* Calculate temporary vectorial force */
1821             tx               = _mm_mul_ps(fscal,dx00);
1822             ty               = _mm_mul_ps(fscal,dy00);
1823             tz               = _mm_mul_ps(fscal,dz00);
1824
1825             /* Update vectorial force */
1826             fix0             = _mm_add_ps(fix0,tx);
1827             fiy0             = _mm_add_ps(fiy0,ty);
1828             fiz0             = _mm_add_ps(fiz0,tz);
1829
1830             fjx0             = _mm_add_ps(fjx0,tx);
1831             fjy0             = _mm_add_ps(fjy0,ty);
1832             fjz0             = _mm_add_ps(fjz0,tz);
1833
1834             }
1835
1836             /**************************
1837              * CALCULATE INTERACTIONS *
1838              **************************/
1839
1840             if (gmx_mm_any_lt(rsq01,rcutoff2))
1841             {
1842
1843             /* REACTION-FIELD ELECTROSTATICS */
1844             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1845
1846             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1847
1848             fscal            = felec;
1849
1850             fscal            = _mm_and_ps(fscal,cutoff_mask);
1851
1852             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1853
1854             /* Calculate temporary vectorial force */
1855             tx               = _mm_mul_ps(fscal,dx01);
1856             ty               = _mm_mul_ps(fscal,dy01);
1857             tz               = _mm_mul_ps(fscal,dz01);
1858
1859             /* Update vectorial force */
1860             fix0             = _mm_add_ps(fix0,tx);
1861             fiy0             = _mm_add_ps(fiy0,ty);
1862             fiz0             = _mm_add_ps(fiz0,tz);
1863
1864             fjx1             = _mm_add_ps(fjx1,tx);
1865             fjy1             = _mm_add_ps(fjy1,ty);
1866             fjz1             = _mm_add_ps(fjz1,tz);
1867
1868             }
1869
1870             /**************************
1871              * CALCULATE INTERACTIONS *
1872              **************************/
1873
1874             if (gmx_mm_any_lt(rsq02,rcutoff2))
1875             {
1876
1877             /* REACTION-FIELD ELECTROSTATICS */
1878             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1879
1880             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1881
1882             fscal            = felec;
1883
1884             fscal            = _mm_and_ps(fscal,cutoff_mask);
1885
1886             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1887
1888             /* Calculate temporary vectorial force */
1889             tx               = _mm_mul_ps(fscal,dx02);
1890             ty               = _mm_mul_ps(fscal,dy02);
1891             tz               = _mm_mul_ps(fscal,dz02);
1892
1893             /* Update vectorial force */
1894             fix0             = _mm_add_ps(fix0,tx);
1895             fiy0             = _mm_add_ps(fiy0,ty);
1896             fiz0             = _mm_add_ps(fiz0,tz);
1897
1898             fjx2             = _mm_add_ps(fjx2,tx);
1899             fjy2             = _mm_add_ps(fjy2,ty);
1900             fjz2             = _mm_add_ps(fjz2,tz);
1901
1902             }
1903
1904             /**************************
1905              * CALCULATE INTERACTIONS *
1906              **************************/
1907
1908             if (gmx_mm_any_lt(rsq10,rcutoff2))
1909             {
1910
1911             /* REACTION-FIELD ELECTROSTATICS */
1912             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1913
1914             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1915
1916             fscal            = felec;
1917
1918             fscal            = _mm_and_ps(fscal,cutoff_mask);
1919
1920             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1921
1922             /* Calculate temporary vectorial force */
1923             tx               = _mm_mul_ps(fscal,dx10);
1924             ty               = _mm_mul_ps(fscal,dy10);
1925             tz               = _mm_mul_ps(fscal,dz10);
1926
1927             /* Update vectorial force */
1928             fix1             = _mm_add_ps(fix1,tx);
1929             fiy1             = _mm_add_ps(fiy1,ty);
1930             fiz1             = _mm_add_ps(fiz1,tz);
1931
1932             fjx0             = _mm_add_ps(fjx0,tx);
1933             fjy0             = _mm_add_ps(fjy0,ty);
1934             fjz0             = _mm_add_ps(fjz0,tz);
1935
1936             }
1937
1938             /**************************
1939              * CALCULATE INTERACTIONS *
1940              **************************/
1941
1942             if (gmx_mm_any_lt(rsq11,rcutoff2))
1943             {
1944
1945             /* REACTION-FIELD ELECTROSTATICS */
1946             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1947
1948             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1949
1950             fscal            = felec;
1951
1952             fscal            = _mm_and_ps(fscal,cutoff_mask);
1953
1954             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1955
1956             /* Calculate temporary vectorial force */
1957             tx               = _mm_mul_ps(fscal,dx11);
1958             ty               = _mm_mul_ps(fscal,dy11);
1959             tz               = _mm_mul_ps(fscal,dz11);
1960
1961             /* Update vectorial force */
1962             fix1             = _mm_add_ps(fix1,tx);
1963             fiy1             = _mm_add_ps(fiy1,ty);
1964             fiz1             = _mm_add_ps(fiz1,tz);
1965
1966             fjx1             = _mm_add_ps(fjx1,tx);
1967             fjy1             = _mm_add_ps(fjy1,ty);
1968             fjz1             = _mm_add_ps(fjz1,tz);
1969
1970             }
1971
1972             /**************************
1973              * CALCULATE INTERACTIONS *
1974              **************************/
1975
1976             if (gmx_mm_any_lt(rsq12,rcutoff2))
1977             {
1978
1979             /* REACTION-FIELD ELECTROSTATICS */
1980             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1981
1982             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1983
1984             fscal            = felec;
1985
1986             fscal            = _mm_and_ps(fscal,cutoff_mask);
1987
1988             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1989
1990             /* Calculate temporary vectorial force */
1991             tx               = _mm_mul_ps(fscal,dx12);
1992             ty               = _mm_mul_ps(fscal,dy12);
1993             tz               = _mm_mul_ps(fscal,dz12);
1994
1995             /* Update vectorial force */
1996             fix1             = _mm_add_ps(fix1,tx);
1997             fiy1             = _mm_add_ps(fiy1,ty);
1998             fiz1             = _mm_add_ps(fiz1,tz);
1999
2000             fjx2             = _mm_add_ps(fjx2,tx);
2001             fjy2             = _mm_add_ps(fjy2,ty);
2002             fjz2             = _mm_add_ps(fjz2,tz);
2003
2004             }
2005
2006             /**************************
2007              * CALCULATE INTERACTIONS *
2008              **************************/
2009
2010             if (gmx_mm_any_lt(rsq20,rcutoff2))
2011             {
2012
2013             /* REACTION-FIELD ELECTROSTATICS */
2014             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
2015
2016             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2017
2018             fscal            = felec;
2019
2020             fscal            = _mm_and_ps(fscal,cutoff_mask);
2021
2022             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2023
2024             /* Calculate temporary vectorial force */
2025             tx               = _mm_mul_ps(fscal,dx20);
2026             ty               = _mm_mul_ps(fscal,dy20);
2027             tz               = _mm_mul_ps(fscal,dz20);
2028
2029             /* Update vectorial force */
2030             fix2             = _mm_add_ps(fix2,tx);
2031             fiy2             = _mm_add_ps(fiy2,ty);
2032             fiz2             = _mm_add_ps(fiz2,tz);
2033
2034             fjx0             = _mm_add_ps(fjx0,tx);
2035             fjy0             = _mm_add_ps(fjy0,ty);
2036             fjz0             = _mm_add_ps(fjz0,tz);
2037
2038             }
2039
2040             /**************************
2041              * CALCULATE INTERACTIONS *
2042              **************************/
2043
2044             if (gmx_mm_any_lt(rsq21,rcutoff2))
2045             {
2046
2047             /* REACTION-FIELD ELECTROSTATICS */
2048             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
2049
2050             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2051
2052             fscal            = felec;
2053
2054             fscal            = _mm_and_ps(fscal,cutoff_mask);
2055
2056             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2057
2058             /* Calculate temporary vectorial force */
2059             tx               = _mm_mul_ps(fscal,dx21);
2060             ty               = _mm_mul_ps(fscal,dy21);
2061             tz               = _mm_mul_ps(fscal,dz21);
2062
2063             /* Update vectorial force */
2064             fix2             = _mm_add_ps(fix2,tx);
2065             fiy2             = _mm_add_ps(fiy2,ty);
2066             fiz2             = _mm_add_ps(fiz2,tz);
2067
2068             fjx1             = _mm_add_ps(fjx1,tx);
2069             fjy1             = _mm_add_ps(fjy1,ty);
2070             fjz1             = _mm_add_ps(fjz1,tz);
2071
2072             }
2073
2074             /**************************
2075              * CALCULATE INTERACTIONS *
2076              **************************/
2077
2078             if (gmx_mm_any_lt(rsq22,rcutoff2))
2079             {
2080
2081             /* REACTION-FIELD ELECTROSTATICS */
2082             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2083
2084             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2085
2086             fscal            = felec;
2087
2088             fscal            = _mm_and_ps(fscal,cutoff_mask);
2089
2090             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2091
2092             /* Calculate temporary vectorial force */
2093             tx               = _mm_mul_ps(fscal,dx22);
2094             ty               = _mm_mul_ps(fscal,dy22);
2095             tz               = _mm_mul_ps(fscal,dz22);
2096
2097             /* Update vectorial force */
2098             fix2             = _mm_add_ps(fix2,tx);
2099             fiy2             = _mm_add_ps(fiy2,ty);
2100             fiz2             = _mm_add_ps(fiz2,tz);
2101
2102             fjx2             = _mm_add_ps(fjx2,tx);
2103             fjy2             = _mm_add_ps(fjy2,ty);
2104             fjz2             = _mm_add_ps(fjz2,tz);
2105
2106             }
2107
2108             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2109             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2110             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2111             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2112
2113             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2114                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2115
2116             /* Inner loop uses 277 flops */
2117         }
2118
2119         /* End of innermost loop */
2120
2121         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2122                                               f+i_coord_offset,fshift+i_shift_offset);
2123
2124         /* Increment number of inner iterations */
2125         inneriter                  += j_index_end - j_index_start;
2126
2127         /* Outer loop uses 18 flops */
2128     }
2129
2130     /* Increment number of outer iterations */
2131     outeriter        += nri;
2132
2133     /* Update outer/inner flops */
2134
2135     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
2136 }