src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRFCut_VdwNone_GeomW4W4_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_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_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse2_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              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             shX,shY,shZ,rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset1;
  67     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  68     int              vdwioffset2;
  69     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  70     int              vdwioffset3;
  71     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  72     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  73     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  74     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  75     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  76     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  77     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  78     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  79     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  80     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  81     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  82     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  83     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  84     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  85     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  86     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128           dummy_mask,cutoff_mask;
  90     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  91     __m128           one     = _mm_set1_ps(1.0);
  92     __m128           two     = _mm_set1_ps(2.0);
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = _mm_set1_ps(fr->epsfac);
 105     charge           = mdatoms->chargeA;
 106     krf              = _mm_set1_ps(fr->ic->k_rf);
 107     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 108     crf              = _mm_set1_ps(fr->ic->c_rf);
 109
 110     /* Setup water-specific parameters */
 111     inr              = nlist->iinr[0];
 112     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 113     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 114     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 115
 116     jq1              = _mm_set1_ps(charge[inr+1]);
 117     jq2              = _mm_set1_ps(charge[inr+2]);
 118     jq3              = _mm_set1_ps(charge[inr+3]);
 119     qq11             = _mm_mul_ps(iq1,jq1);
 120     qq12             = _mm_mul_ps(iq1,jq2);
 121     qq13             = _mm_mul_ps(iq1,jq3);
 122     qq21             = _mm_mul_ps(iq2,jq1);
 123     qq22             = _mm_mul_ps(iq2,jq2);
 124     qq23             = _mm_mul_ps(iq2,jq3);
 125     qq31             = _mm_mul_ps(iq3,jq1);
 126     qq32             = _mm_mul_ps(iq3,jq2);
 127     qq33             = _mm_mul_ps(iq3,jq3);
 128
 129     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 130     rcutoff_scalar   = fr->rcoulomb;
 131     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 132     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 133
 134     /* Avoid stupid compiler warnings */
 135     jnrA = jnrB = jnrC = jnrD = 0;
 136     j_coord_offsetA = 0;
 137     j_coord_offsetB = 0;
 138     j_coord_offsetC = 0;
 139     j_coord_offsetD = 0;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149         shX              = shiftvec[i_shift_offset+XX];
 150         shY              = shiftvec[i_shift_offset+YY];
 151         shZ              = shiftvec[i_shift_offset+ZZ];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 163         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 164         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 165         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 166         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 167         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 168         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 169         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 170         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 171
 172         fix1             = _mm_setzero_ps();
 173         fiy1             = _mm_setzero_ps();
 174         fiz1             = _mm_setzero_ps();
 175         fix2             = _mm_setzero_ps();
 176         fiy2             = _mm_setzero_ps();
 177         fiz2             = _mm_setzero_ps();
 178         fix3             = _mm_setzero_ps();
 179         fiy3             = _mm_setzero_ps();
 180         fiz3             = _mm_setzero_ps();
 181
 182         /* Reset potential sums */
 183         velecsum         = _mm_setzero_ps();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             jnrC             = jjnr[jidx+2];
 193             jnrD             = jjnr[jidx+3];
 194
 195             j_coord_offsetA  = DIM*jnrA;
 196             j_coord_offsetB  = DIM*jnrB;
 197             j_coord_offsetC  = DIM*jnrC;
 198             j_coord_offsetD  = DIM*jnrD;
 199
 200             /* load j atom coordinates */
 201             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 202                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 203                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 204
 205             /* Calculate displacement vector */
 206             dx11             = _mm_sub_ps(ix1,jx1);
 207             dy11             = _mm_sub_ps(iy1,jy1);
 208             dz11             = _mm_sub_ps(iz1,jz1);
 209             dx12             = _mm_sub_ps(ix1,jx2);
 210             dy12             = _mm_sub_ps(iy1,jy2);
 211             dz12             = _mm_sub_ps(iz1,jz2);
 212             dx13             = _mm_sub_ps(ix1,jx3);
 213             dy13             = _mm_sub_ps(iy1,jy3);
 214             dz13             = _mm_sub_ps(iz1,jz3);
 215             dx21             = _mm_sub_ps(ix2,jx1);
 216             dy21             = _mm_sub_ps(iy2,jy1);
 217             dz21             = _mm_sub_ps(iz2,jz1);
 218             dx22             = _mm_sub_ps(ix2,jx2);
 219             dy22             = _mm_sub_ps(iy2,jy2);
 220             dz22             = _mm_sub_ps(iz2,jz2);
 221             dx23             = _mm_sub_ps(ix2,jx3);
 222             dy23             = _mm_sub_ps(iy2,jy3);
 223             dz23             = _mm_sub_ps(iz2,jz3);
 224             dx31             = _mm_sub_ps(ix3,jx1);
 225             dy31             = _mm_sub_ps(iy3,jy1);
 226             dz31             = _mm_sub_ps(iz3,jz1);
 227             dx32             = _mm_sub_ps(ix3,jx2);
 228             dy32             = _mm_sub_ps(iy3,jy2);
 229             dz32             = _mm_sub_ps(iz3,jz2);
 230             dx33             = _mm_sub_ps(ix3,jx3);
 231             dy33             = _mm_sub_ps(iy3,jy3);
 232             dz33             = _mm_sub_ps(iz3,jz3);
 233
 234             /* Calculate squared distance and things based on it */
 235             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 236             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 237             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 238             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 239             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 240             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 241             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 242             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 243             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 244
 245             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 246             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 247             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 248             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 249             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 250             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 251             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 252             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 253             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 254
 255             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 256             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 257             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 258             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 259             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 260             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 261             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 262             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 263             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 264
 265             fjx1             = _mm_setzero_ps();
 266             fjy1             = _mm_setzero_ps();
 267             fjz1             = _mm_setzero_ps();
 268             fjx2             = _mm_setzero_ps();
 269             fjy2             = _mm_setzero_ps();
 270             fjz2             = _mm_setzero_ps();
 271             fjx3             = _mm_setzero_ps();
 272             fjy3             = _mm_setzero_ps();
 273             fjz3             = _mm_setzero_ps();
 274
 275             /**************************
 276              * CALCULATE INTERACTIONS *
 277              **************************/
 278
 279             if (gmx_mm_any_lt(rsq11,rcutoff2))
 280             {
 281
 282             /* REACTION-FIELD ELECTROSTATICS */
 283             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 284             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 285
 286             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 287
 288             /* Update potential sum for this i atom from the interaction with this j atom. */
 289             velec            = _mm_and_ps(velec,cutoff_mask);
 290             velecsum         = _mm_add_ps(velecsum,velec);
 291
 292             fscal            = felec;
 293
 294             fscal            = _mm_and_ps(fscal,cutoff_mask);
 295
 296             /* Calculate temporary vectorial force */
 297             tx               = _mm_mul_ps(fscal,dx11);
 298             ty               = _mm_mul_ps(fscal,dy11);
 299             tz               = _mm_mul_ps(fscal,dz11);
 300
 301             /* Update vectorial force */
 302             fix1             = _mm_add_ps(fix1,tx);
 303             fiy1             = _mm_add_ps(fiy1,ty);
 304             fiz1             = _mm_add_ps(fiz1,tz);
 305
 306             fjx1             = _mm_add_ps(fjx1,tx);
 307             fjy1             = _mm_add_ps(fjy1,ty);
 308             fjz1             = _mm_add_ps(fjz1,tz);
 309
 310             }
 311
 312             /**************************
 313              * CALCULATE INTERACTIONS *
 314              **************************/
 315
 316             if (gmx_mm_any_lt(rsq12,rcutoff2))
 317             {
 318
 319             /* REACTION-FIELD ELECTROSTATICS */
 320             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 321             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 322
 323             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 324
 325             /* Update potential sum for this i atom from the interaction with this j atom. */
 326             velec            = _mm_and_ps(velec,cutoff_mask);
 327             velecsum         = _mm_add_ps(velecsum,velec);
 328
 329             fscal            = felec;
 330
 331             fscal            = _mm_and_ps(fscal,cutoff_mask);
 332
 333             /* Calculate temporary vectorial force */
 334             tx               = _mm_mul_ps(fscal,dx12);
 335             ty               = _mm_mul_ps(fscal,dy12);
 336             tz               = _mm_mul_ps(fscal,dz12);
 337
 338             /* Update vectorial force */
 339             fix1             = _mm_add_ps(fix1,tx);
 340             fiy1             = _mm_add_ps(fiy1,ty);
 341             fiz1             = _mm_add_ps(fiz1,tz);
 342
 343             fjx2             = _mm_add_ps(fjx2,tx);
 344             fjy2             = _mm_add_ps(fjy2,ty);
 345             fjz2             = _mm_add_ps(fjz2,tz);
 346
 347             }
 348
 349             /**************************
 350              * CALCULATE INTERACTIONS *
 351              **************************/
 352
 353             if (gmx_mm_any_lt(rsq13,rcutoff2))
 354             {
 355
 356             /* REACTION-FIELD ELECTROSTATICS */
 357             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 358             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 359
 360             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velec            = _mm_and_ps(velec,cutoff_mask);
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368             fscal            = _mm_and_ps(fscal,cutoff_mask);
 369
 370             /* Calculate temporary vectorial force */
 371             tx               = _mm_mul_ps(fscal,dx13);
 372             ty               = _mm_mul_ps(fscal,dy13);
 373             tz               = _mm_mul_ps(fscal,dz13);
 374
 375             /* Update vectorial force */
 376             fix1             = _mm_add_ps(fix1,tx);
 377             fiy1             = _mm_add_ps(fiy1,ty);
 378             fiz1             = _mm_add_ps(fiz1,tz);
 379
 380             fjx3             = _mm_add_ps(fjx3,tx);
 381             fjy3             = _mm_add_ps(fjy3,ty);
 382             fjz3             = _mm_add_ps(fjz3,tz);
 383
 384             }
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             if (gmx_mm_any_lt(rsq21,rcutoff2))
 391             {
 392
 393             /* REACTION-FIELD ELECTROSTATICS */
 394             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 395             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 396
 397             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velec            = _mm_and_ps(velec,cutoff_mask);
 401             velecsum         = _mm_add_ps(velecsum,velec);
 402
 403             fscal            = felec;
 404
 405             fscal            = _mm_and_ps(fscal,cutoff_mask);
 406
 407             /* Calculate temporary vectorial force */
 408             tx               = _mm_mul_ps(fscal,dx21);
 409             ty               = _mm_mul_ps(fscal,dy21);
 410             tz               = _mm_mul_ps(fscal,dz21);
 411
 412             /* Update vectorial force */
 413             fix2             = _mm_add_ps(fix2,tx);
 414             fiy2             = _mm_add_ps(fiy2,ty);
 415             fiz2             = _mm_add_ps(fiz2,tz);
 416
 417             fjx1             = _mm_add_ps(fjx1,tx);
 418             fjy1             = _mm_add_ps(fjy1,ty);
 419             fjz1             = _mm_add_ps(fjz1,tz);
 420
 421             }
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             if (gmx_mm_any_lt(rsq22,rcutoff2))
 428             {
 429
 430             /* REACTION-FIELD ELECTROSTATICS */
 431             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 432             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 433
 434             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 435
 436             /* Update potential sum for this i atom from the interaction with this j atom. */
 437             velec            = _mm_and_ps(velec,cutoff_mask);
 438             velecsum         = _mm_add_ps(velecsum,velec);
 439
 440             fscal            = felec;
 441
 442             fscal            = _mm_and_ps(fscal,cutoff_mask);
 443
 444             /* Calculate temporary vectorial force */
 445             tx               = _mm_mul_ps(fscal,dx22);
 446             ty               = _mm_mul_ps(fscal,dy22);
 447             tz               = _mm_mul_ps(fscal,dz22);
 448
 449             /* Update vectorial force */
 450             fix2             = _mm_add_ps(fix2,tx);
 451             fiy2             = _mm_add_ps(fiy2,ty);
 452             fiz2             = _mm_add_ps(fiz2,tz);
 453
 454             fjx2             = _mm_add_ps(fjx2,tx);
 455             fjy2             = _mm_add_ps(fjy2,ty);
 456             fjz2             = _mm_add_ps(fjz2,tz);
 457
 458             }
 459
 460             /**************************
 461              * CALCULATE INTERACTIONS *
 462              **************************/
 463
 464             if (gmx_mm_any_lt(rsq23,rcutoff2))
 465             {
 466
 467             /* REACTION-FIELD ELECTROSTATICS */
 468             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 469             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 470
 471             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 472
 473             /* Update potential sum for this i atom from the interaction with this j atom. */
 474             velec            = _mm_and_ps(velec,cutoff_mask);
 475             velecsum         = _mm_add_ps(velecsum,velec);
 476
 477             fscal            = felec;
 478
 479             fscal            = _mm_and_ps(fscal,cutoff_mask);
 480
 481             /* Calculate temporary vectorial force */
 482             tx               = _mm_mul_ps(fscal,dx23);
 483             ty               = _mm_mul_ps(fscal,dy23);
 484             tz               = _mm_mul_ps(fscal,dz23);
 485
 486             /* Update vectorial force */
 487             fix2             = _mm_add_ps(fix2,tx);
 488             fiy2             = _mm_add_ps(fiy2,ty);
 489             fiz2             = _mm_add_ps(fiz2,tz);
 490
 491             fjx3             = _mm_add_ps(fjx3,tx);
 492             fjy3             = _mm_add_ps(fjy3,ty);
 493             fjz3             = _mm_add_ps(fjz3,tz);
 494
 495             }
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             if (gmx_mm_any_lt(rsq31,rcutoff2))
 502             {
 503
 504             /* REACTION-FIELD ELECTROSTATICS */
 505             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 506             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 507
 508             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 509
 510             /* Update potential sum for this i atom from the interaction with this j atom. */
 511             velec            = _mm_and_ps(velec,cutoff_mask);
 512             velecsum         = _mm_add_ps(velecsum,velec);
 513
 514             fscal            = felec;
 515
 516             fscal            = _mm_and_ps(fscal,cutoff_mask);
 517
 518             /* Calculate temporary vectorial force */
 519             tx               = _mm_mul_ps(fscal,dx31);
 520             ty               = _mm_mul_ps(fscal,dy31);
 521             tz               = _mm_mul_ps(fscal,dz31);
 522
 523             /* Update vectorial force */
 524             fix3             = _mm_add_ps(fix3,tx);
 525             fiy3             = _mm_add_ps(fiy3,ty);
 526             fiz3             = _mm_add_ps(fiz3,tz);
 527
 528             fjx1             = _mm_add_ps(fjx1,tx);
 529             fjy1             = _mm_add_ps(fjy1,ty);
 530             fjz1             = _mm_add_ps(fjz1,tz);
 531
 532             }
 533
 534             /**************************
 535              * CALCULATE INTERACTIONS *
 536              **************************/
 537
 538             if (gmx_mm_any_lt(rsq32,rcutoff2))
 539             {
 540
 541             /* REACTION-FIELD ELECTROSTATICS */
 542             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 543             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 544
 545             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 546
 547             /* Update potential sum for this i atom from the interaction with this j atom. */
 548             velec            = _mm_and_ps(velec,cutoff_mask);
 549             velecsum         = _mm_add_ps(velecsum,velec);
 550
 551             fscal            = felec;
 552
 553             fscal            = _mm_and_ps(fscal,cutoff_mask);
 554
 555             /* Calculate temporary vectorial force */
 556             tx               = _mm_mul_ps(fscal,dx32);
 557             ty               = _mm_mul_ps(fscal,dy32);
 558             tz               = _mm_mul_ps(fscal,dz32);
 559
 560             /* Update vectorial force */
 561             fix3             = _mm_add_ps(fix3,tx);
 562             fiy3             = _mm_add_ps(fiy3,ty);
 563             fiz3             = _mm_add_ps(fiz3,tz);
 564
 565             fjx2             = _mm_add_ps(fjx2,tx);
 566             fjy2             = _mm_add_ps(fjy2,ty);
 567             fjz2             = _mm_add_ps(fjz2,tz);
 568
 569             }
 570
 571             /**************************
 572              * CALCULATE INTERACTIONS *
 573              **************************/
 574
 575             if (gmx_mm_any_lt(rsq33,rcutoff2))
 576             {
 577
 578             /* REACTION-FIELD ELECTROSTATICS */
 579             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 580             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 581
 582             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 583
 584             /* Update potential sum for this i atom from the interaction with this j atom. */
 585             velec            = _mm_and_ps(velec,cutoff_mask);
 586             velecsum         = _mm_add_ps(velecsum,velec);
 587
 588             fscal            = felec;
 589
 590             fscal            = _mm_and_ps(fscal,cutoff_mask);
 591
 592             /* Calculate temporary vectorial force */
 593             tx               = _mm_mul_ps(fscal,dx33);
 594             ty               = _mm_mul_ps(fscal,dy33);
 595             tz               = _mm_mul_ps(fscal,dz33);
 596
 597             /* Update vectorial force */
 598             fix3             = _mm_add_ps(fix3,tx);
 599             fiy3             = _mm_add_ps(fiy3,ty);
 600             fiz3             = _mm_add_ps(fiz3,tz);
 601
 602             fjx3             = _mm_add_ps(fjx3,tx);
 603             fjy3             = _mm_add_ps(fjy3,ty);
 604             fjz3             = _mm_add_ps(fjz3,tz);
 605
 606             }
 607
 608             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
 609                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
 610                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 611
 612             /* Inner loop uses 324 flops */
 613         }
 614
 615         if(jidx<j_index_end)
 616         {
 617
 618             /* Get j neighbor index, and coordinate index */
 619             jnrA             = jjnr[jidx];
 620             jnrB             = jjnr[jidx+1];
 621             jnrC             = jjnr[jidx+2];
 622             jnrD             = jjnr[jidx+3];
 623
 624             /* Sign of each element will be negative for non-real atoms.
 625              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 626              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 627              */
 628             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 629             jnrA       = (jnrA>=0) ? jnrA : 0;
 630             jnrB       = (jnrB>=0) ? jnrB : 0;
 631             jnrC       = (jnrC>=0) ? jnrC : 0;
 632             jnrD       = (jnrD>=0) ? jnrD : 0;
 633
 634             j_coord_offsetA  = DIM*jnrA;
 635             j_coord_offsetB  = DIM*jnrB;
 636             j_coord_offsetC  = DIM*jnrC;
 637             j_coord_offsetD  = DIM*jnrD;
 638
 639             /* load j atom coordinates */
 640             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 641                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 642                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 643
 644             /* Calculate displacement vector */
 645             dx11             = _mm_sub_ps(ix1,jx1);
 646             dy11             = _mm_sub_ps(iy1,jy1);
 647             dz11             = _mm_sub_ps(iz1,jz1);
 648             dx12             = _mm_sub_ps(ix1,jx2);
 649             dy12             = _mm_sub_ps(iy1,jy2);
 650             dz12             = _mm_sub_ps(iz1,jz2);
 651             dx13             = _mm_sub_ps(ix1,jx3);
 652             dy13             = _mm_sub_ps(iy1,jy3);
 653             dz13             = _mm_sub_ps(iz1,jz3);
 654             dx21             = _mm_sub_ps(ix2,jx1);
 655             dy21             = _mm_sub_ps(iy2,jy1);
 656             dz21             = _mm_sub_ps(iz2,jz1);
 657             dx22             = _mm_sub_ps(ix2,jx2);
 658             dy22             = _mm_sub_ps(iy2,jy2);
 659             dz22             = _mm_sub_ps(iz2,jz2);
 660             dx23             = _mm_sub_ps(ix2,jx3);
 661             dy23             = _mm_sub_ps(iy2,jy3);
 662             dz23             = _mm_sub_ps(iz2,jz3);
 663             dx31             = _mm_sub_ps(ix3,jx1);
 664             dy31             = _mm_sub_ps(iy3,jy1);
 665             dz31             = _mm_sub_ps(iz3,jz1);
 666             dx32             = _mm_sub_ps(ix3,jx2);
 667             dy32             = _mm_sub_ps(iy3,jy2);
 668             dz32             = _mm_sub_ps(iz3,jz2);
 669             dx33             = _mm_sub_ps(ix3,jx3);
 670             dy33             = _mm_sub_ps(iy3,jy3);
 671             dz33             = _mm_sub_ps(iz3,jz3);
 672
 673             /* Calculate squared distance and things based on it */
 674             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 675             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 676             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 677             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 678             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 679             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 680             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 681             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 682             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 683
 684             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 685             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 686             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 687             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 688             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 689             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 690             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 691             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 692             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 693
 694             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 695             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 696             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 697             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 698             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 699             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 700             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 701             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 702             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 703
 704             fjx1             = _mm_setzero_ps();
 705             fjy1             = _mm_setzero_ps();
 706             fjz1             = _mm_setzero_ps();
 707             fjx2             = _mm_setzero_ps();
 708             fjy2             = _mm_setzero_ps();
 709             fjz2             = _mm_setzero_ps();
 710             fjx3             = _mm_setzero_ps();
 711             fjy3             = _mm_setzero_ps();
 712             fjz3             = _mm_setzero_ps();
 713
 714             /**************************
 715              * CALCULATE INTERACTIONS *
 716              **************************/
 717
 718             if (gmx_mm_any_lt(rsq11,rcutoff2))
 719             {
 720
 721             /* REACTION-FIELD ELECTROSTATICS */
 722             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 723             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 724
 725             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 726
 727             /* Update potential sum for this i atom from the interaction with this j atom. */
 728             velec            = _mm_and_ps(velec,cutoff_mask);
 729             velec            = _mm_andnot_ps(dummy_mask,velec);
 730             velecsum         = _mm_add_ps(velecsum,velec);
 731
 732             fscal            = felec;
 733
 734             fscal            = _mm_and_ps(fscal,cutoff_mask);
 735
 736             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 737
 738             /* Calculate temporary vectorial force */
 739             tx               = _mm_mul_ps(fscal,dx11);
 740             ty               = _mm_mul_ps(fscal,dy11);
 741             tz               = _mm_mul_ps(fscal,dz11);
 742
 743             /* Update vectorial force */
 744             fix1             = _mm_add_ps(fix1,tx);
 745             fiy1             = _mm_add_ps(fiy1,ty);
 746             fiz1             = _mm_add_ps(fiz1,tz);
 747
 748             fjx1             = _mm_add_ps(fjx1,tx);
 749             fjy1             = _mm_add_ps(fjy1,ty);
 750             fjz1             = _mm_add_ps(fjz1,tz);
 751
 752             }
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             if (gmx_mm_any_lt(rsq12,rcutoff2))
 759             {
 760
 761             /* REACTION-FIELD ELECTROSTATICS */
 762             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 763             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 764
 765             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 766
 767             /* Update potential sum for this i atom from the interaction with this j atom. */
 768             velec            = _mm_and_ps(velec,cutoff_mask);
 769             velec            = _mm_andnot_ps(dummy_mask,velec);
 770             velecsum         = _mm_add_ps(velecsum,velec);
 771
 772             fscal            = felec;
 773
 774             fscal            = _mm_and_ps(fscal,cutoff_mask);
 775
 776             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 777
 778             /* Calculate temporary vectorial force */
 779             tx               = _mm_mul_ps(fscal,dx12);
 780             ty               = _mm_mul_ps(fscal,dy12);
 781             tz               = _mm_mul_ps(fscal,dz12);
 782
 783             /* Update vectorial force */
 784             fix1             = _mm_add_ps(fix1,tx);
 785             fiy1             = _mm_add_ps(fiy1,ty);
 786             fiz1             = _mm_add_ps(fiz1,tz);
 787
 788             fjx2             = _mm_add_ps(fjx2,tx);
 789             fjy2             = _mm_add_ps(fjy2,ty);
 790             fjz2             = _mm_add_ps(fjz2,tz);
 791
 792             }
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             if (gmx_mm_any_lt(rsq13,rcutoff2))
 799             {
 800
 801             /* REACTION-FIELD ELECTROSTATICS */
 802             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 803             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 804
 805             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 806
 807             /* Update potential sum for this i atom from the interaction with this j atom. */
 808             velec            = _mm_and_ps(velec,cutoff_mask);
 809             velec            = _mm_andnot_ps(dummy_mask,velec);
 810             velecsum         = _mm_add_ps(velecsum,velec);
 811
 812             fscal            = felec;
 813
 814             fscal            = _mm_and_ps(fscal,cutoff_mask);
 815
 816             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 817
 818             /* Calculate temporary vectorial force */
 819             tx               = _mm_mul_ps(fscal,dx13);
 820             ty               = _mm_mul_ps(fscal,dy13);
 821             tz               = _mm_mul_ps(fscal,dz13);
 822
 823             /* Update vectorial force */
 824             fix1             = _mm_add_ps(fix1,tx);
 825             fiy1             = _mm_add_ps(fiy1,ty);
 826             fiz1             = _mm_add_ps(fiz1,tz);
 827
 828             fjx3             = _mm_add_ps(fjx3,tx);
 829             fjy3             = _mm_add_ps(fjy3,ty);
 830             fjz3             = _mm_add_ps(fjz3,tz);
 831
 832             }
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             if (gmx_mm_any_lt(rsq21,rcutoff2))
 839             {
 840
 841             /* REACTION-FIELD ELECTROSTATICS */
 842             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 843             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 844
 845             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 846
 847             /* Update potential sum for this i atom from the interaction with this j atom. */
 848             velec            = _mm_and_ps(velec,cutoff_mask);
 849             velec            = _mm_andnot_ps(dummy_mask,velec);
 850             velecsum         = _mm_add_ps(velecsum,velec);
 851
 852             fscal            = felec;
 853
 854             fscal            = _mm_and_ps(fscal,cutoff_mask);
 855
 856             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 857
 858             /* Calculate temporary vectorial force */
 859             tx               = _mm_mul_ps(fscal,dx21);
 860             ty               = _mm_mul_ps(fscal,dy21);
 861             tz               = _mm_mul_ps(fscal,dz21);
 862
 863             /* Update vectorial force */
 864             fix2             = _mm_add_ps(fix2,tx);
 865             fiy2             = _mm_add_ps(fiy2,ty);
 866             fiz2             = _mm_add_ps(fiz2,tz);
 867
 868             fjx1             = _mm_add_ps(fjx1,tx);
 869             fjy1             = _mm_add_ps(fjy1,ty);
 870             fjz1             = _mm_add_ps(fjz1,tz);
 871
 872             }
 873
 874             /**************************
 875              * CALCULATE INTERACTIONS *
 876              **************************/
 877
 878             if (gmx_mm_any_lt(rsq22,rcutoff2))
 879             {
 880
 881             /* REACTION-FIELD ELECTROSTATICS */
 882             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 883             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 884
 885             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 886
 887             /* Update potential sum for this i atom from the interaction with this j atom. */
 888             velec            = _mm_and_ps(velec,cutoff_mask);
 889             velec            = _mm_andnot_ps(dummy_mask,velec);
 890             velecsum         = _mm_add_ps(velecsum,velec);
 891
 892             fscal            = felec;
 893
 894             fscal            = _mm_and_ps(fscal,cutoff_mask);
 895
 896             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 897
 898             /* Calculate temporary vectorial force */
 899             tx               = _mm_mul_ps(fscal,dx22);
 900             ty               = _mm_mul_ps(fscal,dy22);
 901             tz               = _mm_mul_ps(fscal,dz22);
 902
 903             /* Update vectorial force */
 904             fix2             = _mm_add_ps(fix2,tx);
 905             fiy2             = _mm_add_ps(fiy2,ty);
 906             fiz2             = _mm_add_ps(fiz2,tz);
 907
 908             fjx2             = _mm_add_ps(fjx2,tx);
 909             fjy2             = _mm_add_ps(fjy2,ty);
 910             fjz2             = _mm_add_ps(fjz2,tz);
 911
 912             }
 913
 914             /**************************
 915              * CALCULATE INTERACTIONS *
 916              **************************/
 917
 918             if (gmx_mm_any_lt(rsq23,rcutoff2))
 919             {
 920
 921             /* REACTION-FIELD ELECTROSTATICS */
 922             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 923             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 924
 925             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 926
 927             /* Update potential sum for this i atom from the interaction with this j atom. */
 928             velec            = _mm_and_ps(velec,cutoff_mask);
 929             velec            = _mm_andnot_ps(dummy_mask,velec);
 930             velecsum         = _mm_add_ps(velecsum,velec);
 931
 932             fscal            = felec;
 933
 934             fscal            = _mm_and_ps(fscal,cutoff_mask);
 935
 936             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 937
 938             /* Calculate temporary vectorial force */
 939             tx               = _mm_mul_ps(fscal,dx23);
 940             ty               = _mm_mul_ps(fscal,dy23);
 941             tz               = _mm_mul_ps(fscal,dz23);
 942
 943             /* Update vectorial force */
 944             fix2             = _mm_add_ps(fix2,tx);
 945             fiy2             = _mm_add_ps(fiy2,ty);
 946             fiz2             = _mm_add_ps(fiz2,tz);
 947
 948             fjx3             = _mm_add_ps(fjx3,tx);
 949             fjy3             = _mm_add_ps(fjy3,ty);
 950             fjz3             = _mm_add_ps(fjz3,tz);
 951
 952             }
 953
 954             /**************************
 955              * CALCULATE INTERACTIONS *
 956              **************************/
 957
 958             if (gmx_mm_any_lt(rsq31,rcutoff2))
 959             {
 960
 961             /* REACTION-FIELD ELECTROSTATICS */
 962             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 963             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 964
 965             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 966
 967             /* Update potential sum for this i atom from the interaction with this j atom. */
 968             velec            = _mm_and_ps(velec,cutoff_mask);
 969             velec            = _mm_andnot_ps(dummy_mask,velec);
 970             velecsum         = _mm_add_ps(velecsum,velec);
 971
 972             fscal            = felec;
 973
 974             fscal            = _mm_and_ps(fscal,cutoff_mask);
 975
 976             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 977
 978             /* Calculate temporary vectorial force */
 979             tx               = _mm_mul_ps(fscal,dx31);
 980             ty               = _mm_mul_ps(fscal,dy31);
 981             tz               = _mm_mul_ps(fscal,dz31);
 982
 983             /* Update vectorial force */
 984             fix3             = _mm_add_ps(fix3,tx);
 985             fiy3             = _mm_add_ps(fiy3,ty);
 986             fiz3             = _mm_add_ps(fiz3,tz);
 987
 988             fjx1             = _mm_add_ps(fjx1,tx);
 989             fjy1             = _mm_add_ps(fjy1,ty);
 990             fjz1             = _mm_add_ps(fjz1,tz);
 991
 992             }
 993
 994             /**************************
 995              * CALCULATE INTERACTIONS *
 996              **************************/
 997
 998             if (gmx_mm_any_lt(rsq32,rcutoff2))
 999             {
1000
1001             /* REACTION-FIELD ELECTROSTATICS */
1002             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1003             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1004
1005             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1006
1007             /* Update potential sum for this i atom from the interaction with this j atom. */
1008             velec            = _mm_and_ps(velec,cutoff_mask);
1009             velec            = _mm_andnot_ps(dummy_mask,velec);
1010             velecsum         = _mm_add_ps(velecsum,velec);
1011
1012             fscal            = felec;
1013
1014             fscal            = _mm_and_ps(fscal,cutoff_mask);
1015
1016             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1017
1018             /* Calculate temporary vectorial force */
1019             tx               = _mm_mul_ps(fscal,dx32);
1020             ty               = _mm_mul_ps(fscal,dy32);
1021             tz               = _mm_mul_ps(fscal,dz32);
1022
1023             /* Update vectorial force */
1024             fix3             = _mm_add_ps(fix3,tx);
1025             fiy3             = _mm_add_ps(fiy3,ty);
1026             fiz3             = _mm_add_ps(fiz3,tz);
1027
1028             fjx2             = _mm_add_ps(fjx2,tx);
1029             fjy2             = _mm_add_ps(fjy2,ty);
1030             fjz2             = _mm_add_ps(fjz2,tz);
1031
1032             }
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             if (gmx_mm_any_lt(rsq33,rcutoff2))
1039             {
1040
1041             /* REACTION-FIELD ELECTROSTATICS */
1042             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1043             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1044
1045             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1046
1047             /* Update potential sum for this i atom from the interaction with this j atom. */
1048             velec            = _mm_and_ps(velec,cutoff_mask);
1049             velec            = _mm_andnot_ps(dummy_mask,velec);
1050             velecsum         = _mm_add_ps(velecsum,velec);
1051
1052             fscal            = felec;
1053
1054             fscal            = _mm_and_ps(fscal,cutoff_mask);
1055
1056             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1057
1058             /* Calculate temporary vectorial force */
1059             tx               = _mm_mul_ps(fscal,dx33);
1060             ty               = _mm_mul_ps(fscal,dy33);
1061             tz               = _mm_mul_ps(fscal,dz33);
1062
1063             /* Update vectorial force */
1064             fix3             = _mm_add_ps(fix3,tx);
1065             fiy3             = _mm_add_ps(fiy3,ty);
1066             fiz3             = _mm_add_ps(fiz3,tz);
1067
1068             fjx3             = _mm_add_ps(fjx3,tx);
1069             fjy3             = _mm_add_ps(fjy3,ty);
1070             fjz3             = _mm_add_ps(fjz3,tz);
1071
1072             }
1073
1074             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
1075                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
1076                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1077
1078             /* Inner loop uses 324 flops */
1079         }
1080
1081         /* End of innermost loop */
1082
1083         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1084                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1085
1086         ggid                        = gid[iidx];
1087         /* Update potential energies */
1088         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1089
1090         /* Increment number of inner iterations */
1091         inneriter                  += j_index_end - j_index_start;
1092
1093         /* Outer loop uses 28 flops */
1094     }
1095
1096     /* Increment number of outer iterations */
1097     outeriter        += nri;
1098
1099     /* Update outer/inner flops */
1100
1101     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*28 + inneriter*324);
1102 }
1103 /*
1104  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse2_single
1105  * Electrostatics interaction: ReactionField
1106  * VdW interaction:            None
1107  * Geometry:                   Water4-Water4
1108  * Calculate force/pot:        Force
1109  */
1110 void
1111 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse2_single
1112                     (t_nblist * gmx_restrict                nlist,
1113                      rvec * gmx_restrict                    xx,
1114                      rvec * gmx_restrict                    ff,
1115                      t_forcerec * gmx_restrict              fr,
1116                      t_mdatoms * gmx_restrict               mdatoms,
1117                      nb_kernel_data_t * gmx_restrict        kernel_data,
1118                      t_nrnb * gmx_restrict                  nrnb)
1119 {
1120     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1121      * just 0 for non-waters.
1122      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1123      * jnr indices corresponding to data put in the four positions in the SIMD register.
1124      */
1125     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1126     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1127     int              jnrA,jnrB,jnrC,jnrD;
1128     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1129     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1130     real             shX,shY,shZ,rcutoff_scalar;
1131     real             *shiftvec,*fshift,*x,*f;
1132     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1133     int              vdwioffset1;
1134     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1135     int              vdwioffset2;
1136     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1137     int              vdwioffset3;
1138     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1139     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1140     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1141     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1142     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1143     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1144     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1145     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1146     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1147     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1148     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1149     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1150     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1151     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1152     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1153     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1154     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1155     real             *charge;
1156     __m128           dummy_mask,cutoff_mask;
1157     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1158     __m128           one     = _mm_set1_ps(1.0);
1159     __m128           two     = _mm_set1_ps(2.0);
1160     x                = xx[0];
1161     f                = ff[0];
1162
1163     nri              = nlist->nri;
1164     iinr             = nlist->iinr;
1165     jindex           = nlist->jindex;
1166     jjnr             = nlist->jjnr;
1167     shiftidx         = nlist->shift;
1168     gid              = nlist->gid;
1169     shiftvec         = fr->shift_vec[0];
1170     fshift           = fr->fshift[0];
1171     facel            = _mm_set1_ps(fr->epsfac);
1172     charge           = mdatoms->chargeA;
1173     krf              = _mm_set1_ps(fr->ic->k_rf);
1174     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1175     crf              = _mm_set1_ps(fr->ic->c_rf);
1176
1177     /* Setup water-specific parameters */
1178     inr              = nlist->iinr[0];
1179     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1180     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1181     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1182
1183     jq1              = _mm_set1_ps(charge[inr+1]);
1184     jq2              = _mm_set1_ps(charge[inr+2]);
1185     jq3              = _mm_set1_ps(charge[inr+3]);
1186     qq11             = _mm_mul_ps(iq1,jq1);
1187     qq12             = _mm_mul_ps(iq1,jq2);
1188     qq13             = _mm_mul_ps(iq1,jq3);
1189     qq21             = _mm_mul_ps(iq2,jq1);
1190     qq22             = _mm_mul_ps(iq2,jq2);
1191     qq23             = _mm_mul_ps(iq2,jq3);
1192     qq31             = _mm_mul_ps(iq3,jq1);
1193     qq32             = _mm_mul_ps(iq3,jq2);
1194     qq33             = _mm_mul_ps(iq3,jq3);
1195
1196     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1197     rcutoff_scalar   = fr->rcoulomb;
1198     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1199     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1200
1201     /* Avoid stupid compiler warnings */
1202     jnrA = jnrB = jnrC = jnrD = 0;
1203     j_coord_offsetA = 0;
1204     j_coord_offsetB = 0;
1205     j_coord_offsetC = 0;
1206     j_coord_offsetD = 0;
1207
1208     outeriter        = 0;
1209     inneriter        = 0;
1210
1211     /* Start outer loop over neighborlists */
1212     for(iidx=0; iidx<nri; iidx++)
1213     {
1214         /* Load shift vector for this list */
1215         i_shift_offset   = DIM*shiftidx[iidx];
1216         shX              = shiftvec[i_shift_offset+XX];
1217         shY              = shiftvec[i_shift_offset+YY];
1218         shZ              = shiftvec[i_shift_offset+ZZ];
1219
1220         /* Load limits for loop over neighbors */
1221         j_index_start    = jindex[iidx];
1222         j_index_end      = jindex[iidx+1];
1223
1224         /* Get outer coordinate index */
1225         inr              = iinr[iidx];
1226         i_coord_offset   = DIM*inr;
1227
1228         /* Load i particle coords and add shift vector */
1229         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1230         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1231         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1232         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1233         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1234         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1235         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1236         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1237         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1238
1239         fix1             = _mm_setzero_ps();
1240         fiy1             = _mm_setzero_ps();
1241         fiz1             = _mm_setzero_ps();
1242         fix2             = _mm_setzero_ps();
1243         fiy2             = _mm_setzero_ps();
1244         fiz2             = _mm_setzero_ps();
1245         fix3             = _mm_setzero_ps();
1246         fiy3             = _mm_setzero_ps();
1247         fiz3             = _mm_setzero_ps();
1248
1249         /* Start inner kernel loop */
1250         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1251         {
1252
1253             /* Get j neighbor index, and coordinate index */
1254             jnrA             = jjnr[jidx];
1255             jnrB             = jjnr[jidx+1];
1256             jnrC             = jjnr[jidx+2];
1257             jnrD             = jjnr[jidx+3];
1258
1259             j_coord_offsetA  = DIM*jnrA;
1260             j_coord_offsetB  = DIM*jnrB;
1261             j_coord_offsetC  = DIM*jnrC;
1262             j_coord_offsetD  = DIM*jnrD;
1263
1264             /* load j atom coordinates */
1265             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1266                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1267                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1268
1269             /* Calculate displacement vector */
1270             dx11             = _mm_sub_ps(ix1,jx1);
1271             dy11             = _mm_sub_ps(iy1,jy1);
1272             dz11             = _mm_sub_ps(iz1,jz1);
1273             dx12             = _mm_sub_ps(ix1,jx2);
1274             dy12             = _mm_sub_ps(iy1,jy2);
1275             dz12             = _mm_sub_ps(iz1,jz2);
1276             dx13             = _mm_sub_ps(ix1,jx3);
1277             dy13             = _mm_sub_ps(iy1,jy3);
1278             dz13             = _mm_sub_ps(iz1,jz3);
1279             dx21             = _mm_sub_ps(ix2,jx1);
1280             dy21             = _mm_sub_ps(iy2,jy1);
1281             dz21             = _mm_sub_ps(iz2,jz1);
1282             dx22             = _mm_sub_ps(ix2,jx2);
1283             dy22             = _mm_sub_ps(iy2,jy2);
1284             dz22             = _mm_sub_ps(iz2,jz2);
1285             dx23             = _mm_sub_ps(ix2,jx3);
1286             dy23             = _mm_sub_ps(iy2,jy3);
1287             dz23             = _mm_sub_ps(iz2,jz3);
1288             dx31             = _mm_sub_ps(ix3,jx1);
1289             dy31             = _mm_sub_ps(iy3,jy1);
1290             dz31             = _mm_sub_ps(iz3,jz1);
1291             dx32             = _mm_sub_ps(ix3,jx2);
1292             dy32             = _mm_sub_ps(iy3,jy2);
1293             dz32             = _mm_sub_ps(iz3,jz2);
1294             dx33             = _mm_sub_ps(ix3,jx3);
1295             dy33             = _mm_sub_ps(iy3,jy3);
1296             dz33             = _mm_sub_ps(iz3,jz3);
1297
1298             /* Calculate squared distance and things based on it */
1299             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1300             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1301             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1302             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1303             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1304             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1305             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1306             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1307             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1308
1309             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1310             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1311             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1312             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1313             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1314             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1315             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1316             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1317             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1318
1319             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1320             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1321             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1322             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1323             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1324             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1325             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1326             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1327             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1328
1329             fjx1             = _mm_setzero_ps();
1330             fjy1             = _mm_setzero_ps();
1331             fjz1             = _mm_setzero_ps();
1332             fjx2             = _mm_setzero_ps();
1333             fjy2             = _mm_setzero_ps();
1334             fjz2             = _mm_setzero_ps();
1335             fjx3             = _mm_setzero_ps();
1336             fjy3             = _mm_setzero_ps();
1337             fjz3             = _mm_setzero_ps();
1338
1339             /**************************
1340              * CALCULATE INTERACTIONS *
1341              **************************/
1342
1343             if (gmx_mm_any_lt(rsq11,rcutoff2))
1344             {
1345
1346             /* REACTION-FIELD ELECTROSTATICS */
1347             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1348
1349             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1350
1351             fscal            = felec;
1352
1353             fscal            = _mm_and_ps(fscal,cutoff_mask);
1354
1355             /* Calculate temporary vectorial force */
1356             tx               = _mm_mul_ps(fscal,dx11);
1357             ty               = _mm_mul_ps(fscal,dy11);
1358             tz               = _mm_mul_ps(fscal,dz11);
1359
1360             /* Update vectorial force */
1361             fix1             = _mm_add_ps(fix1,tx);
1362             fiy1             = _mm_add_ps(fiy1,ty);
1363             fiz1             = _mm_add_ps(fiz1,tz);
1364
1365             fjx1             = _mm_add_ps(fjx1,tx);
1366             fjy1             = _mm_add_ps(fjy1,ty);
1367             fjz1             = _mm_add_ps(fjz1,tz);
1368
1369             }
1370
1371             /**************************
1372              * CALCULATE INTERACTIONS *
1373              **************************/
1374
1375             if (gmx_mm_any_lt(rsq12,rcutoff2))
1376             {
1377
1378             /* REACTION-FIELD ELECTROSTATICS */
1379             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1380
1381             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1382
1383             fscal            = felec;
1384
1385             fscal            = _mm_and_ps(fscal,cutoff_mask);
1386
1387             /* Calculate temporary vectorial force */
1388             tx               = _mm_mul_ps(fscal,dx12);
1389             ty               = _mm_mul_ps(fscal,dy12);
1390             tz               = _mm_mul_ps(fscal,dz12);
1391
1392             /* Update vectorial force */
1393             fix1             = _mm_add_ps(fix1,tx);
1394             fiy1             = _mm_add_ps(fiy1,ty);
1395             fiz1             = _mm_add_ps(fiz1,tz);
1396
1397             fjx2             = _mm_add_ps(fjx2,tx);
1398             fjy2             = _mm_add_ps(fjy2,ty);
1399             fjz2             = _mm_add_ps(fjz2,tz);
1400
1401             }
1402
1403             /**************************
1404              * CALCULATE INTERACTIONS *
1405              **************************/
1406
1407             if (gmx_mm_any_lt(rsq13,rcutoff2))
1408             {
1409
1410             /* REACTION-FIELD ELECTROSTATICS */
1411             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1412
1413             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1414
1415             fscal            = felec;
1416
1417             fscal            = _mm_and_ps(fscal,cutoff_mask);
1418
1419             /* Calculate temporary vectorial force */
1420             tx               = _mm_mul_ps(fscal,dx13);
1421             ty               = _mm_mul_ps(fscal,dy13);
1422             tz               = _mm_mul_ps(fscal,dz13);
1423
1424             /* Update vectorial force */
1425             fix1             = _mm_add_ps(fix1,tx);
1426             fiy1             = _mm_add_ps(fiy1,ty);
1427             fiz1             = _mm_add_ps(fiz1,tz);
1428
1429             fjx3             = _mm_add_ps(fjx3,tx);
1430             fjy3             = _mm_add_ps(fjy3,ty);
1431             fjz3             = _mm_add_ps(fjz3,tz);
1432
1433             }
1434
1435             /**************************
1436              * CALCULATE INTERACTIONS *
1437              **************************/
1438
1439             if (gmx_mm_any_lt(rsq21,rcutoff2))
1440             {
1441
1442             /* REACTION-FIELD ELECTROSTATICS */
1443             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1444
1445             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1446
1447             fscal            = felec;
1448
1449             fscal            = _mm_and_ps(fscal,cutoff_mask);
1450
1451             /* Calculate temporary vectorial force */
1452             tx               = _mm_mul_ps(fscal,dx21);
1453             ty               = _mm_mul_ps(fscal,dy21);
1454             tz               = _mm_mul_ps(fscal,dz21);
1455
1456             /* Update vectorial force */
1457             fix2             = _mm_add_ps(fix2,tx);
1458             fiy2             = _mm_add_ps(fiy2,ty);
1459             fiz2             = _mm_add_ps(fiz2,tz);
1460
1461             fjx1             = _mm_add_ps(fjx1,tx);
1462             fjy1             = _mm_add_ps(fjy1,ty);
1463             fjz1             = _mm_add_ps(fjz1,tz);
1464
1465             }
1466
1467             /**************************
1468              * CALCULATE INTERACTIONS *
1469              **************************/
1470
1471             if (gmx_mm_any_lt(rsq22,rcutoff2))
1472             {
1473
1474             /* REACTION-FIELD ELECTROSTATICS */
1475             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1476
1477             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1478
1479             fscal            = felec;
1480
1481             fscal            = _mm_and_ps(fscal,cutoff_mask);
1482
1483             /* Calculate temporary vectorial force */
1484             tx               = _mm_mul_ps(fscal,dx22);
1485             ty               = _mm_mul_ps(fscal,dy22);
1486             tz               = _mm_mul_ps(fscal,dz22);
1487
1488             /* Update vectorial force */
1489             fix2             = _mm_add_ps(fix2,tx);
1490             fiy2             = _mm_add_ps(fiy2,ty);
1491             fiz2             = _mm_add_ps(fiz2,tz);
1492
1493             fjx2             = _mm_add_ps(fjx2,tx);
1494             fjy2             = _mm_add_ps(fjy2,ty);
1495             fjz2             = _mm_add_ps(fjz2,tz);
1496
1497             }
1498
1499             /**************************
1500              * CALCULATE INTERACTIONS *
1501              **************************/
1502
1503             if (gmx_mm_any_lt(rsq23,rcutoff2))
1504             {
1505
1506             /* REACTION-FIELD ELECTROSTATICS */
1507             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1508
1509             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1510
1511             fscal            = felec;
1512
1513             fscal            = _mm_and_ps(fscal,cutoff_mask);
1514
1515             /* Calculate temporary vectorial force */
1516             tx               = _mm_mul_ps(fscal,dx23);
1517             ty               = _mm_mul_ps(fscal,dy23);
1518             tz               = _mm_mul_ps(fscal,dz23);
1519
1520             /* Update vectorial force */
1521             fix2             = _mm_add_ps(fix2,tx);
1522             fiy2             = _mm_add_ps(fiy2,ty);
1523             fiz2             = _mm_add_ps(fiz2,tz);
1524
1525             fjx3             = _mm_add_ps(fjx3,tx);
1526             fjy3             = _mm_add_ps(fjy3,ty);
1527             fjz3             = _mm_add_ps(fjz3,tz);
1528
1529             }
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             if (gmx_mm_any_lt(rsq31,rcutoff2))
1536             {
1537
1538             /* REACTION-FIELD ELECTROSTATICS */
1539             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1540
1541             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1542
1543             fscal            = felec;
1544
1545             fscal            = _mm_and_ps(fscal,cutoff_mask);
1546
1547             /* Calculate temporary vectorial force */
1548             tx               = _mm_mul_ps(fscal,dx31);
1549             ty               = _mm_mul_ps(fscal,dy31);
1550             tz               = _mm_mul_ps(fscal,dz31);
1551
1552             /* Update vectorial force */
1553             fix3             = _mm_add_ps(fix3,tx);
1554             fiy3             = _mm_add_ps(fiy3,ty);
1555             fiz3             = _mm_add_ps(fiz3,tz);
1556
1557             fjx1             = _mm_add_ps(fjx1,tx);
1558             fjy1             = _mm_add_ps(fjy1,ty);
1559             fjz1             = _mm_add_ps(fjz1,tz);
1560
1561             }
1562
1563             /**************************
1564              * CALCULATE INTERACTIONS *
1565              **************************/
1566
1567             if (gmx_mm_any_lt(rsq32,rcutoff2))
1568             {
1569
1570             /* REACTION-FIELD ELECTROSTATICS */
1571             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1572
1573             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1574
1575             fscal            = felec;
1576
1577             fscal            = _mm_and_ps(fscal,cutoff_mask);
1578
1579             /* Calculate temporary vectorial force */
1580             tx               = _mm_mul_ps(fscal,dx32);
1581             ty               = _mm_mul_ps(fscal,dy32);
1582             tz               = _mm_mul_ps(fscal,dz32);
1583
1584             /* Update vectorial force */
1585             fix3             = _mm_add_ps(fix3,tx);
1586             fiy3             = _mm_add_ps(fiy3,ty);
1587             fiz3             = _mm_add_ps(fiz3,tz);
1588
1589             fjx2             = _mm_add_ps(fjx2,tx);
1590             fjy2             = _mm_add_ps(fjy2,ty);
1591             fjz2             = _mm_add_ps(fjz2,tz);
1592
1593             }
1594
1595             /**************************
1596              * CALCULATE INTERACTIONS *
1597              **************************/
1598
1599             if (gmx_mm_any_lt(rsq33,rcutoff2))
1600             {
1601
1602             /* REACTION-FIELD ELECTROSTATICS */
1603             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1604
1605             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1606
1607             fscal            = felec;
1608
1609             fscal            = _mm_and_ps(fscal,cutoff_mask);
1610
1611             /* Calculate temporary vectorial force */
1612             tx               = _mm_mul_ps(fscal,dx33);
1613             ty               = _mm_mul_ps(fscal,dy33);
1614             tz               = _mm_mul_ps(fscal,dz33);
1615
1616             /* Update vectorial force */
1617             fix3             = _mm_add_ps(fix3,tx);
1618             fiy3             = _mm_add_ps(fiy3,ty);
1619             fiz3             = _mm_add_ps(fiz3,tz);
1620
1621             fjx3             = _mm_add_ps(fjx3,tx);
1622             fjy3             = _mm_add_ps(fjy3,ty);
1623             fjz3             = _mm_add_ps(fjz3,tz);
1624
1625             }
1626
1627             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
1628                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
1629                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1630
1631             /* Inner loop uses 270 flops */
1632         }
1633
1634         if(jidx<j_index_end)
1635         {
1636
1637             /* Get j neighbor index, and coordinate index */
1638             jnrA             = jjnr[jidx];
1639             jnrB             = jjnr[jidx+1];
1640             jnrC             = jjnr[jidx+2];
1641             jnrD             = jjnr[jidx+3];
1642
1643             /* Sign of each element will be negative for non-real atoms.
1644              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1645              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1646              */
1647             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1648             jnrA       = (jnrA>=0) ? jnrA : 0;
1649             jnrB       = (jnrB>=0) ? jnrB : 0;
1650             jnrC       = (jnrC>=0) ? jnrC : 0;
1651             jnrD       = (jnrD>=0) ? jnrD : 0;
1652
1653             j_coord_offsetA  = DIM*jnrA;
1654             j_coord_offsetB  = DIM*jnrB;
1655             j_coord_offsetC  = DIM*jnrC;
1656             j_coord_offsetD  = DIM*jnrD;
1657
1658             /* load j atom coordinates */
1659             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1660                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1661                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1662
1663             /* Calculate displacement vector */
1664             dx11             = _mm_sub_ps(ix1,jx1);
1665             dy11             = _mm_sub_ps(iy1,jy1);
1666             dz11             = _mm_sub_ps(iz1,jz1);
1667             dx12             = _mm_sub_ps(ix1,jx2);
1668             dy12             = _mm_sub_ps(iy1,jy2);
1669             dz12             = _mm_sub_ps(iz1,jz2);
1670             dx13             = _mm_sub_ps(ix1,jx3);
1671             dy13             = _mm_sub_ps(iy1,jy3);
1672             dz13             = _mm_sub_ps(iz1,jz3);
1673             dx21             = _mm_sub_ps(ix2,jx1);
1674             dy21             = _mm_sub_ps(iy2,jy1);
1675             dz21             = _mm_sub_ps(iz2,jz1);
1676             dx22             = _mm_sub_ps(ix2,jx2);
1677             dy22             = _mm_sub_ps(iy2,jy2);
1678             dz22             = _mm_sub_ps(iz2,jz2);
1679             dx23             = _mm_sub_ps(ix2,jx3);
1680             dy23             = _mm_sub_ps(iy2,jy3);
1681             dz23             = _mm_sub_ps(iz2,jz3);
1682             dx31             = _mm_sub_ps(ix3,jx1);
1683             dy31             = _mm_sub_ps(iy3,jy1);
1684             dz31             = _mm_sub_ps(iz3,jz1);
1685             dx32             = _mm_sub_ps(ix3,jx2);
1686             dy32             = _mm_sub_ps(iy3,jy2);
1687             dz32             = _mm_sub_ps(iz3,jz2);
1688             dx33             = _mm_sub_ps(ix3,jx3);
1689             dy33             = _mm_sub_ps(iy3,jy3);
1690             dz33             = _mm_sub_ps(iz3,jz3);
1691
1692             /* Calculate squared distance and things based on it */
1693             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1694             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1695             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1696             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1697             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1698             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1699             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1700             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1701             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1702
1703             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1704             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1705             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1706             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1707             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1708             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1709             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1710             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1711             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1712
1713             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1714             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1715             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1716             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1717             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1718             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1719             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1720             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1721             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1722
1723             fjx1             = _mm_setzero_ps();
1724             fjy1             = _mm_setzero_ps();
1725             fjz1             = _mm_setzero_ps();
1726             fjx2             = _mm_setzero_ps();
1727             fjy2             = _mm_setzero_ps();
1728             fjz2             = _mm_setzero_ps();
1729             fjx3             = _mm_setzero_ps();
1730             fjy3             = _mm_setzero_ps();
1731             fjz3             = _mm_setzero_ps();
1732
1733             /**************************
1734              * CALCULATE INTERACTIONS *
1735              **************************/
1736
1737             if (gmx_mm_any_lt(rsq11,rcutoff2))
1738             {
1739
1740             /* REACTION-FIELD ELECTROSTATICS */
1741             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1742
1743             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1744
1745             fscal            = felec;
1746
1747             fscal            = _mm_and_ps(fscal,cutoff_mask);
1748
1749             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1750
1751             /* Calculate temporary vectorial force */
1752             tx               = _mm_mul_ps(fscal,dx11);
1753             ty               = _mm_mul_ps(fscal,dy11);
1754             tz               = _mm_mul_ps(fscal,dz11);
1755
1756             /* Update vectorial force */
1757             fix1             = _mm_add_ps(fix1,tx);
1758             fiy1             = _mm_add_ps(fiy1,ty);
1759             fiz1             = _mm_add_ps(fiz1,tz);
1760
1761             fjx1             = _mm_add_ps(fjx1,tx);
1762             fjy1             = _mm_add_ps(fjy1,ty);
1763             fjz1             = _mm_add_ps(fjz1,tz);
1764
1765             }
1766
1767             /**************************
1768              * CALCULATE INTERACTIONS *
1769              **************************/
1770
1771             if (gmx_mm_any_lt(rsq12,rcutoff2))
1772             {
1773
1774             /* REACTION-FIELD ELECTROSTATICS */
1775             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1776
1777             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1778
1779             fscal            = felec;
1780
1781             fscal            = _mm_and_ps(fscal,cutoff_mask);
1782
1783             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1784
1785             /* Calculate temporary vectorial force */
1786             tx               = _mm_mul_ps(fscal,dx12);
1787             ty               = _mm_mul_ps(fscal,dy12);
1788             tz               = _mm_mul_ps(fscal,dz12);
1789
1790             /* Update vectorial force */
1791             fix1             = _mm_add_ps(fix1,tx);
1792             fiy1             = _mm_add_ps(fiy1,ty);
1793             fiz1             = _mm_add_ps(fiz1,tz);
1794
1795             fjx2             = _mm_add_ps(fjx2,tx);
1796             fjy2             = _mm_add_ps(fjy2,ty);
1797             fjz2             = _mm_add_ps(fjz2,tz);
1798
1799             }
1800
1801             /**************************
1802              * CALCULATE INTERACTIONS *
1803              **************************/
1804
1805             if (gmx_mm_any_lt(rsq13,rcutoff2))
1806             {
1807
1808             /* REACTION-FIELD ELECTROSTATICS */
1809             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1810
1811             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1812
1813             fscal            = felec;
1814
1815             fscal            = _mm_and_ps(fscal,cutoff_mask);
1816
1817             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1818
1819             /* Calculate temporary vectorial force */
1820             tx               = _mm_mul_ps(fscal,dx13);
1821             ty               = _mm_mul_ps(fscal,dy13);
1822             tz               = _mm_mul_ps(fscal,dz13);
1823
1824             /* Update vectorial force */
1825             fix1             = _mm_add_ps(fix1,tx);
1826             fiy1             = _mm_add_ps(fiy1,ty);
1827             fiz1             = _mm_add_ps(fiz1,tz);
1828
1829             fjx3             = _mm_add_ps(fjx3,tx);
1830             fjy3             = _mm_add_ps(fjy3,ty);
1831             fjz3             = _mm_add_ps(fjz3,tz);
1832
1833             }
1834
1835             /**************************
1836              * CALCULATE INTERACTIONS *
1837              **************************/
1838
1839             if (gmx_mm_any_lt(rsq21,rcutoff2))
1840             {
1841
1842             /* REACTION-FIELD ELECTROSTATICS */
1843             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1844
1845             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1846
1847             fscal            = felec;
1848
1849             fscal            = _mm_and_ps(fscal,cutoff_mask);
1850
1851             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1852
1853             /* Calculate temporary vectorial force */
1854             tx               = _mm_mul_ps(fscal,dx21);
1855             ty               = _mm_mul_ps(fscal,dy21);
1856             tz               = _mm_mul_ps(fscal,dz21);
1857
1858             /* Update vectorial force */
1859             fix2             = _mm_add_ps(fix2,tx);
1860             fiy2             = _mm_add_ps(fiy2,ty);
1861             fiz2             = _mm_add_ps(fiz2,tz);
1862
1863             fjx1             = _mm_add_ps(fjx1,tx);
1864             fjy1             = _mm_add_ps(fjy1,ty);
1865             fjz1             = _mm_add_ps(fjz1,tz);
1866
1867             }
1868
1869             /**************************
1870              * CALCULATE INTERACTIONS *
1871              **************************/
1872
1873             if (gmx_mm_any_lt(rsq22,rcutoff2))
1874             {
1875
1876             /* REACTION-FIELD ELECTROSTATICS */
1877             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1878
1879             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1880
1881             fscal            = felec;
1882
1883             fscal            = _mm_and_ps(fscal,cutoff_mask);
1884
1885             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1886
1887             /* Calculate temporary vectorial force */
1888             tx               = _mm_mul_ps(fscal,dx22);
1889             ty               = _mm_mul_ps(fscal,dy22);
1890             tz               = _mm_mul_ps(fscal,dz22);
1891
1892             /* Update vectorial force */
1893             fix2             = _mm_add_ps(fix2,tx);
1894             fiy2             = _mm_add_ps(fiy2,ty);
1895             fiz2             = _mm_add_ps(fiz2,tz);
1896
1897             fjx2             = _mm_add_ps(fjx2,tx);
1898             fjy2             = _mm_add_ps(fjy2,ty);
1899             fjz2             = _mm_add_ps(fjz2,tz);
1900
1901             }
1902
1903             /**************************
1904              * CALCULATE INTERACTIONS *
1905              **************************/
1906
1907             if (gmx_mm_any_lt(rsq23,rcutoff2))
1908             {
1909
1910             /* REACTION-FIELD ELECTROSTATICS */
1911             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1912
1913             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1914
1915             fscal            = felec;
1916
1917             fscal            = _mm_and_ps(fscal,cutoff_mask);
1918
1919             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1920
1921             /* Calculate temporary vectorial force */
1922             tx               = _mm_mul_ps(fscal,dx23);
1923             ty               = _mm_mul_ps(fscal,dy23);
1924             tz               = _mm_mul_ps(fscal,dz23);
1925
1926             /* Update vectorial force */
1927             fix2             = _mm_add_ps(fix2,tx);
1928             fiy2             = _mm_add_ps(fiy2,ty);
1929             fiz2             = _mm_add_ps(fiz2,tz);
1930
1931             fjx3             = _mm_add_ps(fjx3,tx);
1932             fjy3             = _mm_add_ps(fjy3,ty);
1933             fjz3             = _mm_add_ps(fjz3,tz);
1934
1935             }
1936
1937             /**************************
1938              * CALCULATE INTERACTIONS *
1939              **************************/
1940
1941             if (gmx_mm_any_lt(rsq31,rcutoff2))
1942             {
1943
1944             /* REACTION-FIELD ELECTROSTATICS */
1945             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1946
1947             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1948
1949             fscal            = felec;
1950
1951             fscal            = _mm_and_ps(fscal,cutoff_mask);
1952
1953             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1954
1955             /* Calculate temporary vectorial force */
1956             tx               = _mm_mul_ps(fscal,dx31);
1957             ty               = _mm_mul_ps(fscal,dy31);
1958             tz               = _mm_mul_ps(fscal,dz31);
1959
1960             /* Update vectorial force */
1961             fix3             = _mm_add_ps(fix3,tx);
1962             fiy3             = _mm_add_ps(fiy3,ty);
1963             fiz3             = _mm_add_ps(fiz3,tz);
1964
1965             fjx1             = _mm_add_ps(fjx1,tx);
1966             fjy1             = _mm_add_ps(fjy1,ty);
1967             fjz1             = _mm_add_ps(fjz1,tz);
1968
1969             }
1970
1971             /**************************
1972              * CALCULATE INTERACTIONS *
1973              **************************/
1974
1975             if (gmx_mm_any_lt(rsq32,rcutoff2))
1976             {
1977
1978             /* REACTION-FIELD ELECTROSTATICS */
1979             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1980
1981             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1982
1983             fscal            = felec;
1984
1985             fscal            = _mm_and_ps(fscal,cutoff_mask);
1986
1987             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1988
1989             /* Calculate temporary vectorial force */
1990             tx               = _mm_mul_ps(fscal,dx32);
1991             ty               = _mm_mul_ps(fscal,dy32);
1992             tz               = _mm_mul_ps(fscal,dz32);
1993
1994             /* Update vectorial force */
1995             fix3             = _mm_add_ps(fix3,tx);
1996             fiy3             = _mm_add_ps(fiy3,ty);
1997             fiz3             = _mm_add_ps(fiz3,tz);
1998
1999             fjx2             = _mm_add_ps(fjx2,tx);
2000             fjy2             = _mm_add_ps(fjy2,ty);
2001             fjz2             = _mm_add_ps(fjz2,tz);
2002
2003             }
2004
2005             /**************************
2006              * CALCULATE INTERACTIONS *
2007              **************************/
2008
2009             if (gmx_mm_any_lt(rsq33,rcutoff2))
2010             {
2011
2012             /* REACTION-FIELD ELECTROSTATICS */
2013             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2014
2015             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2016
2017             fscal            = felec;
2018
2019             fscal            = _mm_and_ps(fscal,cutoff_mask);
2020
2021             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2022
2023             /* Calculate temporary vectorial force */
2024             tx               = _mm_mul_ps(fscal,dx33);
2025             ty               = _mm_mul_ps(fscal,dy33);
2026             tz               = _mm_mul_ps(fscal,dz33);
2027
2028             /* Update vectorial force */
2029             fix3             = _mm_add_ps(fix3,tx);
2030             fiy3             = _mm_add_ps(fiy3,ty);
2031             fiz3             = _mm_add_ps(fiz3,tz);
2032
2033             fjx3             = _mm_add_ps(fjx3,tx);
2034             fjy3             = _mm_add_ps(fjy3,ty);
2035             fjz3             = _mm_add_ps(fjz3,tz);
2036
2037             }
2038
2039             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
2040                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
2041                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2042
2043             /* Inner loop uses 270 flops */
2044         }
2045
2046         /* End of innermost loop */
2047
2048         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2049                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2050
2051         /* Increment number of inner iterations */
2052         inneriter                  += j_index_end - j_index_start;
2053
2054         /* Outer loop uses 27 flops */
2055     }
2056
2057     /* Increment number of outer iterations */
2058     outeriter        += nri;
2059
2060     /* Update outer/inner flops */
2061
2062     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*27 + inneriter*270);
2063 }