src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRFCut_VdwNone_GeomW4W4_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_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_sse4_1_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset1;
  70     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  71     int              vdwioffset2;
  72     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  73     int              vdwioffset3;
  74     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  75     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  76     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  77     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  78     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  79     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  80     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  81     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  82     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  83     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  84     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  85     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  86     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  87     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  88     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  89     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128           dummy_mask,cutoff_mask;
  93     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  94     __m128           one     = _mm_set1_ps(1.0);
  95     __m128           two     = _mm_set1_ps(2.0);
  96     x                = xx[0];
  97     f                = ff[0];
  98
  99     nri              = nlist->nri;
 100     iinr             = nlist->iinr;
 101     jindex           = nlist->jindex;
 102     jjnr             = nlist->jjnr;
 103     shiftidx         = nlist->shift;
 104     gid              = nlist->gid;
 105     shiftvec         = fr->shift_vec[0];
 106     fshift           = fr->fshift[0];
 107     facel            = _mm_set1_ps(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109     krf              = _mm_set1_ps(fr->ic->k_rf);
 110     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 111     crf              = _mm_set1_ps(fr->ic->c_rf);
 112
 113     /* Setup water-specific parameters */
 114     inr              = nlist->iinr[0];
 115     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 116     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 117     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 118
 119     jq1              = _mm_set1_ps(charge[inr+1]);
 120     jq2              = _mm_set1_ps(charge[inr+2]);
 121     jq3              = _mm_set1_ps(charge[inr+3]);
 122     qq11             = _mm_mul_ps(iq1,jq1);
 123     qq12             = _mm_mul_ps(iq1,jq2);
 124     qq13             = _mm_mul_ps(iq1,jq3);
 125     qq21             = _mm_mul_ps(iq2,jq1);
 126     qq22             = _mm_mul_ps(iq2,jq2);
 127     qq23             = _mm_mul_ps(iq2,jq3);
 128     qq31             = _mm_mul_ps(iq3,jq1);
 129     qq32             = _mm_mul_ps(iq3,jq2);
 130     qq33             = _mm_mul_ps(iq3,jq3);
 131
 132     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 133     rcutoff_scalar   = fr->rcoulomb;
 134     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 135     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 136
 137     /* Avoid stupid compiler warnings */
 138     jnrA = jnrB = jnrC = jnrD = 0;
 139     j_coord_offsetA = 0;
 140     j_coord_offsetB = 0;
 141     j_coord_offsetC = 0;
 142     j_coord_offsetD = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     for(iidx=0;iidx<4*DIM;iidx++)
 148     {
 149         scratch[iidx] = 0.0;
 150     }
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157
 158         /* Load limits for loop over neighbors */
 159         j_index_start    = jindex[iidx];
 160         j_index_end      = jindex[iidx+1];
 161
 162         /* Get outer coordinate index */
 163         inr              = iinr[iidx];
 164         i_coord_offset   = DIM*inr;
 165
 166         /* Load i particle coords and add shift vector */
 167         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 168                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 169
 170         fix1             = _mm_setzero_ps();
 171         fiy1             = _mm_setzero_ps();
 172         fiz1             = _mm_setzero_ps();
 173         fix2             = _mm_setzero_ps();
 174         fiy2             = _mm_setzero_ps();
 175         fiz2             = _mm_setzero_ps();
 176         fix3             = _mm_setzero_ps();
 177         fiy3             = _mm_setzero_ps();
 178         fiz3             = _mm_setzero_ps();
 179
 180         /* Reset potential sums */
 181         velecsum         = _mm_setzero_ps();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             jnrC             = jjnr[jidx+2];
 191             jnrD             = jjnr[jidx+3];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194             j_coord_offsetC  = DIM*jnrC;
 195             j_coord_offsetD  = DIM*jnrD;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 199                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 200                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 201
 202             /* Calculate displacement vector */
 203             dx11             = _mm_sub_ps(ix1,jx1);
 204             dy11             = _mm_sub_ps(iy1,jy1);
 205             dz11             = _mm_sub_ps(iz1,jz1);
 206             dx12             = _mm_sub_ps(ix1,jx2);
 207             dy12             = _mm_sub_ps(iy1,jy2);
 208             dz12             = _mm_sub_ps(iz1,jz2);
 209             dx13             = _mm_sub_ps(ix1,jx3);
 210             dy13             = _mm_sub_ps(iy1,jy3);
 211             dz13             = _mm_sub_ps(iz1,jz3);
 212             dx21             = _mm_sub_ps(ix2,jx1);
 213             dy21             = _mm_sub_ps(iy2,jy1);
 214             dz21             = _mm_sub_ps(iz2,jz1);
 215             dx22             = _mm_sub_ps(ix2,jx2);
 216             dy22             = _mm_sub_ps(iy2,jy2);
 217             dz22             = _mm_sub_ps(iz2,jz2);
 218             dx23             = _mm_sub_ps(ix2,jx3);
 219             dy23             = _mm_sub_ps(iy2,jy3);
 220             dz23             = _mm_sub_ps(iz2,jz3);
 221             dx31             = _mm_sub_ps(ix3,jx1);
 222             dy31             = _mm_sub_ps(iy3,jy1);
 223             dz31             = _mm_sub_ps(iz3,jz1);
 224             dx32             = _mm_sub_ps(ix3,jx2);
 225             dy32             = _mm_sub_ps(iy3,jy2);
 226             dz32             = _mm_sub_ps(iz3,jz2);
 227             dx33             = _mm_sub_ps(ix3,jx3);
 228             dy33             = _mm_sub_ps(iy3,jy3);
 229             dz33             = _mm_sub_ps(iz3,jz3);
 230
 231             /* Calculate squared distance and things based on it */
 232             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 233             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 234             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 235             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 236             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 237             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 238             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 239             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 240             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 241
 242             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 243             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 244             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 245             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 246             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 247             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 248             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 249             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 250             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 251
 252             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 253             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 254             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 255             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 256             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 257             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 258             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 259             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 260             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 261
 262             fjx1             = _mm_setzero_ps();
 263             fjy1             = _mm_setzero_ps();
 264             fjz1             = _mm_setzero_ps();
 265             fjx2             = _mm_setzero_ps();
 266             fjy2             = _mm_setzero_ps();
 267             fjz2             = _mm_setzero_ps();
 268             fjx3             = _mm_setzero_ps();
 269             fjy3             = _mm_setzero_ps();
 270             fjz3             = _mm_setzero_ps();
 271
 272             /**************************
 273              * CALCULATE INTERACTIONS *
 274              **************************/
 275
 276             if (gmx_mm_any_lt(rsq11,rcutoff2))
 277             {
 278
 279             /* REACTION-FIELD ELECTROSTATICS */
 280             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 281             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 282
 283             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 284
 285             /* Update potential sum for this i atom from the interaction with this j atom. */
 286             velec            = _mm_and_ps(velec,cutoff_mask);
 287             velecsum         = _mm_add_ps(velecsum,velec);
 288
 289             fscal            = felec;
 290
 291             fscal            = _mm_and_ps(fscal,cutoff_mask);
 292
 293             /* Calculate temporary vectorial force */
 294             tx               = _mm_mul_ps(fscal,dx11);
 295             ty               = _mm_mul_ps(fscal,dy11);
 296             tz               = _mm_mul_ps(fscal,dz11);
 297
 298             /* Update vectorial force */
 299             fix1             = _mm_add_ps(fix1,tx);
 300             fiy1             = _mm_add_ps(fiy1,ty);
 301             fiz1             = _mm_add_ps(fiz1,tz);
 302
 303             fjx1             = _mm_add_ps(fjx1,tx);
 304             fjy1             = _mm_add_ps(fjy1,ty);
 305             fjz1             = _mm_add_ps(fjz1,tz);
 306
 307             }
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             if (gmx_mm_any_lt(rsq12,rcutoff2))
 314             {
 315
 316             /* REACTION-FIELD ELECTROSTATICS */
 317             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 318             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 319
 320             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 321
 322             /* Update potential sum for this i atom from the interaction with this j atom. */
 323             velec            = _mm_and_ps(velec,cutoff_mask);
 324             velecsum         = _mm_add_ps(velecsum,velec);
 325
 326             fscal            = felec;
 327
 328             fscal            = _mm_and_ps(fscal,cutoff_mask);
 329
 330             /* Calculate temporary vectorial force */
 331             tx               = _mm_mul_ps(fscal,dx12);
 332             ty               = _mm_mul_ps(fscal,dy12);
 333             tz               = _mm_mul_ps(fscal,dz12);
 334
 335             /* Update vectorial force */
 336             fix1             = _mm_add_ps(fix1,tx);
 337             fiy1             = _mm_add_ps(fiy1,ty);
 338             fiz1             = _mm_add_ps(fiz1,tz);
 339
 340             fjx2             = _mm_add_ps(fjx2,tx);
 341             fjy2             = _mm_add_ps(fjy2,ty);
 342             fjz2             = _mm_add_ps(fjz2,tz);
 343
 344             }
 345
 346             /**************************
 347              * CALCULATE INTERACTIONS *
 348              **************************/
 349
 350             if (gmx_mm_any_lt(rsq13,rcutoff2))
 351             {
 352
 353             /* REACTION-FIELD ELECTROSTATICS */
 354             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 355             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 356
 357             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velec            = _mm_and_ps(velec,cutoff_mask);
 361             velecsum         = _mm_add_ps(velecsum,velec);
 362
 363             fscal            = felec;
 364
 365             fscal            = _mm_and_ps(fscal,cutoff_mask);
 366
 367             /* Calculate temporary vectorial force */
 368             tx               = _mm_mul_ps(fscal,dx13);
 369             ty               = _mm_mul_ps(fscal,dy13);
 370             tz               = _mm_mul_ps(fscal,dz13);
 371
 372             /* Update vectorial force */
 373             fix1             = _mm_add_ps(fix1,tx);
 374             fiy1             = _mm_add_ps(fiy1,ty);
 375             fiz1             = _mm_add_ps(fiz1,tz);
 376
 377             fjx3             = _mm_add_ps(fjx3,tx);
 378             fjy3             = _mm_add_ps(fjy3,ty);
 379             fjz3             = _mm_add_ps(fjz3,tz);
 380
 381             }
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             if (gmx_mm_any_lt(rsq21,rcutoff2))
 388             {
 389
 390             /* REACTION-FIELD ELECTROSTATICS */
 391             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 392             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 393
 394             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velec            = _mm_and_ps(velec,cutoff_mask);
 398             velecsum         = _mm_add_ps(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402             fscal            = _mm_and_ps(fscal,cutoff_mask);
 403
 404             /* Calculate temporary vectorial force */
 405             tx               = _mm_mul_ps(fscal,dx21);
 406             ty               = _mm_mul_ps(fscal,dy21);
 407             tz               = _mm_mul_ps(fscal,dz21);
 408
 409             /* Update vectorial force */
 410             fix2             = _mm_add_ps(fix2,tx);
 411             fiy2             = _mm_add_ps(fiy2,ty);
 412             fiz2             = _mm_add_ps(fiz2,tz);
 413
 414             fjx1             = _mm_add_ps(fjx1,tx);
 415             fjy1             = _mm_add_ps(fjy1,ty);
 416             fjz1             = _mm_add_ps(fjz1,tz);
 417
 418             }
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             if (gmx_mm_any_lt(rsq22,rcutoff2))
 425             {
 426
 427             /* REACTION-FIELD ELECTROSTATICS */
 428             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 429             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 430
 431             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velec            = _mm_and_ps(velec,cutoff_mask);
 435             velecsum         = _mm_add_ps(velecsum,velec);
 436
 437             fscal            = felec;
 438
 439             fscal            = _mm_and_ps(fscal,cutoff_mask);
 440
 441             /* Calculate temporary vectorial force */
 442             tx               = _mm_mul_ps(fscal,dx22);
 443             ty               = _mm_mul_ps(fscal,dy22);
 444             tz               = _mm_mul_ps(fscal,dz22);
 445
 446             /* Update vectorial force */
 447             fix2             = _mm_add_ps(fix2,tx);
 448             fiy2             = _mm_add_ps(fiy2,ty);
 449             fiz2             = _mm_add_ps(fiz2,tz);
 450
 451             fjx2             = _mm_add_ps(fjx2,tx);
 452             fjy2             = _mm_add_ps(fjy2,ty);
 453             fjz2             = _mm_add_ps(fjz2,tz);
 454
 455             }
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             if (gmx_mm_any_lt(rsq23,rcutoff2))
 462             {
 463
 464             /* REACTION-FIELD ELECTROSTATICS */
 465             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 466             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 467
 468             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velec            = _mm_and_ps(velec,cutoff_mask);
 472             velecsum         = _mm_add_ps(velecsum,velec);
 473
 474             fscal            = felec;
 475
 476             fscal            = _mm_and_ps(fscal,cutoff_mask);
 477
 478             /* Calculate temporary vectorial force */
 479             tx               = _mm_mul_ps(fscal,dx23);
 480             ty               = _mm_mul_ps(fscal,dy23);
 481             tz               = _mm_mul_ps(fscal,dz23);
 482
 483             /* Update vectorial force */
 484             fix2             = _mm_add_ps(fix2,tx);
 485             fiy2             = _mm_add_ps(fiy2,ty);
 486             fiz2             = _mm_add_ps(fiz2,tz);
 487
 488             fjx3             = _mm_add_ps(fjx3,tx);
 489             fjy3             = _mm_add_ps(fjy3,ty);
 490             fjz3             = _mm_add_ps(fjz3,tz);
 491
 492             }
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             if (gmx_mm_any_lt(rsq31,rcutoff2))
 499             {
 500
 501             /* REACTION-FIELD ELECTROSTATICS */
 502             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 503             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 504
 505             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velec            = _mm_and_ps(velec,cutoff_mask);
 509             velecsum         = _mm_add_ps(velecsum,velec);
 510
 511             fscal            = felec;
 512
 513             fscal            = _mm_and_ps(fscal,cutoff_mask);
 514
 515             /* Calculate temporary vectorial force */
 516             tx               = _mm_mul_ps(fscal,dx31);
 517             ty               = _mm_mul_ps(fscal,dy31);
 518             tz               = _mm_mul_ps(fscal,dz31);
 519
 520             /* Update vectorial force */
 521             fix3             = _mm_add_ps(fix3,tx);
 522             fiy3             = _mm_add_ps(fiy3,ty);
 523             fiz3             = _mm_add_ps(fiz3,tz);
 524
 525             fjx1             = _mm_add_ps(fjx1,tx);
 526             fjy1             = _mm_add_ps(fjy1,ty);
 527             fjz1             = _mm_add_ps(fjz1,tz);
 528
 529             }
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             if (gmx_mm_any_lt(rsq32,rcutoff2))
 536             {
 537
 538             /* REACTION-FIELD ELECTROSTATICS */
 539             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 540             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 541
 542             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 543
 544             /* Update potential sum for this i atom from the interaction with this j atom. */
 545             velec            = _mm_and_ps(velec,cutoff_mask);
 546             velecsum         = _mm_add_ps(velecsum,velec);
 547
 548             fscal            = felec;
 549
 550             fscal            = _mm_and_ps(fscal,cutoff_mask);
 551
 552             /* Calculate temporary vectorial force */
 553             tx               = _mm_mul_ps(fscal,dx32);
 554             ty               = _mm_mul_ps(fscal,dy32);
 555             tz               = _mm_mul_ps(fscal,dz32);
 556
 557             /* Update vectorial force */
 558             fix3             = _mm_add_ps(fix3,tx);
 559             fiy3             = _mm_add_ps(fiy3,ty);
 560             fiz3             = _mm_add_ps(fiz3,tz);
 561
 562             fjx2             = _mm_add_ps(fjx2,tx);
 563             fjy2             = _mm_add_ps(fjy2,ty);
 564             fjz2             = _mm_add_ps(fjz2,tz);
 565
 566             }
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             if (gmx_mm_any_lt(rsq33,rcutoff2))
 573             {
 574
 575             /* REACTION-FIELD ELECTROSTATICS */
 576             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 577             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 578
 579             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 580
 581             /* Update potential sum for this i atom from the interaction with this j atom. */
 582             velec            = _mm_and_ps(velec,cutoff_mask);
 583             velecsum         = _mm_add_ps(velecsum,velec);
 584
 585             fscal            = felec;
 586
 587             fscal            = _mm_and_ps(fscal,cutoff_mask);
 588
 589             /* Calculate temporary vectorial force */
 590             tx               = _mm_mul_ps(fscal,dx33);
 591             ty               = _mm_mul_ps(fscal,dy33);
 592             tz               = _mm_mul_ps(fscal,dz33);
 593
 594             /* Update vectorial force */
 595             fix3             = _mm_add_ps(fix3,tx);
 596             fiy3             = _mm_add_ps(fiy3,ty);
 597             fiz3             = _mm_add_ps(fiz3,tz);
 598
 599             fjx3             = _mm_add_ps(fjx3,tx);
 600             fjy3             = _mm_add_ps(fjy3,ty);
 601             fjz3             = _mm_add_ps(fjz3,tz);
 602
 603             }
 604
 605             fjptrA             = f+j_coord_offsetA;
 606             fjptrB             = f+j_coord_offsetB;
 607             fjptrC             = f+j_coord_offsetC;
 608             fjptrD             = f+j_coord_offsetD;
 609
 610             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 611                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 612
 613             /* Inner loop uses 324 flops */
 614         }
 615
 616         if(jidx<j_index_end)
 617         {
 618
 619             /* Get j neighbor index, and coordinate index */
 620             jnrlistA         = jjnr[jidx];
 621             jnrlistB         = jjnr[jidx+1];
 622             jnrlistC         = jjnr[jidx+2];
 623             jnrlistD         = jjnr[jidx+3];
 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       = (jnrlistA>=0) ? jnrlistA : 0;
 630             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 631             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 632             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 633             j_coord_offsetA  = DIM*jnrA;
 634             j_coord_offsetB  = DIM*jnrB;
 635             j_coord_offsetC  = DIM*jnrC;
 636             j_coord_offsetD  = DIM*jnrD;
 637
 638             /* load j atom coordinates */
 639             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 640                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 641                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 642
 643             /* Calculate displacement vector */
 644             dx11             = _mm_sub_ps(ix1,jx1);
 645             dy11             = _mm_sub_ps(iy1,jy1);
 646             dz11             = _mm_sub_ps(iz1,jz1);
 647             dx12             = _mm_sub_ps(ix1,jx2);
 648             dy12             = _mm_sub_ps(iy1,jy2);
 649             dz12             = _mm_sub_ps(iz1,jz2);
 650             dx13             = _mm_sub_ps(ix1,jx3);
 651             dy13             = _mm_sub_ps(iy1,jy3);
 652             dz13             = _mm_sub_ps(iz1,jz3);
 653             dx21             = _mm_sub_ps(ix2,jx1);
 654             dy21             = _mm_sub_ps(iy2,jy1);
 655             dz21             = _mm_sub_ps(iz2,jz1);
 656             dx22             = _mm_sub_ps(ix2,jx2);
 657             dy22             = _mm_sub_ps(iy2,jy2);
 658             dz22             = _mm_sub_ps(iz2,jz2);
 659             dx23             = _mm_sub_ps(ix2,jx3);
 660             dy23             = _mm_sub_ps(iy2,jy3);
 661             dz23             = _mm_sub_ps(iz2,jz3);
 662             dx31             = _mm_sub_ps(ix3,jx1);
 663             dy31             = _mm_sub_ps(iy3,jy1);
 664             dz31             = _mm_sub_ps(iz3,jz1);
 665             dx32             = _mm_sub_ps(ix3,jx2);
 666             dy32             = _mm_sub_ps(iy3,jy2);
 667             dz32             = _mm_sub_ps(iz3,jz2);
 668             dx33             = _mm_sub_ps(ix3,jx3);
 669             dy33             = _mm_sub_ps(iy3,jy3);
 670             dz33             = _mm_sub_ps(iz3,jz3);
 671
 672             /* Calculate squared distance and things based on it */
 673             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 674             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 675             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 676             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 677             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 678             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 679             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 680             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 681             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 682
 683             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 684             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 685             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 686             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 687             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 688             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 689             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 690             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 691             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 692
 693             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 694             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 695             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 696             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 697             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 698             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 699             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 700             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 701             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 702
 703             fjx1             = _mm_setzero_ps();
 704             fjy1             = _mm_setzero_ps();
 705             fjz1             = _mm_setzero_ps();
 706             fjx2             = _mm_setzero_ps();
 707             fjy2             = _mm_setzero_ps();
 708             fjz2             = _mm_setzero_ps();
 709             fjx3             = _mm_setzero_ps();
 710             fjy3             = _mm_setzero_ps();
 711             fjz3             = _mm_setzero_ps();
 712
 713             /**************************
 714              * CALCULATE INTERACTIONS *
 715              **************************/
 716
 717             if (gmx_mm_any_lt(rsq11,rcutoff2))
 718             {
 719
 720             /* REACTION-FIELD ELECTROSTATICS */
 721             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 722             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 723
 724             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 725
 726             /* Update potential sum for this i atom from the interaction with this j atom. */
 727             velec            = _mm_and_ps(velec,cutoff_mask);
 728             velec            = _mm_andnot_ps(dummy_mask,velec);
 729             velecsum         = _mm_add_ps(velecsum,velec);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_and_ps(fscal,cutoff_mask);
 734
 735             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 736
 737             /* Calculate temporary vectorial force */
 738             tx               = _mm_mul_ps(fscal,dx11);
 739             ty               = _mm_mul_ps(fscal,dy11);
 740             tz               = _mm_mul_ps(fscal,dz11);
 741
 742             /* Update vectorial force */
 743             fix1             = _mm_add_ps(fix1,tx);
 744             fiy1             = _mm_add_ps(fiy1,ty);
 745             fiz1             = _mm_add_ps(fiz1,tz);
 746
 747             fjx1             = _mm_add_ps(fjx1,tx);
 748             fjy1             = _mm_add_ps(fjy1,ty);
 749             fjz1             = _mm_add_ps(fjz1,tz);
 750
 751             }
 752
 753             /**************************
 754              * CALCULATE INTERACTIONS *
 755              **************************/
 756
 757             if (gmx_mm_any_lt(rsq12,rcutoff2))
 758             {
 759
 760             /* REACTION-FIELD ELECTROSTATICS */
 761             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 762             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 763
 764             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 765
 766             /* Update potential sum for this i atom from the interaction with this j atom. */
 767             velec            = _mm_and_ps(velec,cutoff_mask);
 768             velec            = _mm_andnot_ps(dummy_mask,velec);
 769             velecsum         = _mm_add_ps(velecsum,velec);
 770
 771             fscal            = felec;
 772
 773             fscal            = _mm_and_ps(fscal,cutoff_mask);
 774
 775             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 776
 777             /* Calculate temporary vectorial force */
 778             tx               = _mm_mul_ps(fscal,dx12);
 779             ty               = _mm_mul_ps(fscal,dy12);
 780             tz               = _mm_mul_ps(fscal,dz12);
 781
 782             /* Update vectorial force */
 783             fix1             = _mm_add_ps(fix1,tx);
 784             fiy1             = _mm_add_ps(fiy1,ty);
 785             fiz1             = _mm_add_ps(fiz1,tz);
 786
 787             fjx2             = _mm_add_ps(fjx2,tx);
 788             fjy2             = _mm_add_ps(fjy2,ty);
 789             fjz2             = _mm_add_ps(fjz2,tz);
 790
 791             }
 792
 793             /**************************
 794              * CALCULATE INTERACTIONS *
 795              **************************/
 796
 797             if (gmx_mm_any_lt(rsq13,rcutoff2))
 798             {
 799
 800             /* REACTION-FIELD ELECTROSTATICS */
 801             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 802             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 803
 804             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 805
 806             /* Update potential sum for this i atom from the interaction with this j atom. */
 807             velec            = _mm_and_ps(velec,cutoff_mask);
 808             velec            = _mm_andnot_ps(dummy_mask,velec);
 809             velecsum         = _mm_add_ps(velecsum,velec);
 810
 811             fscal            = felec;
 812
 813             fscal            = _mm_and_ps(fscal,cutoff_mask);
 814
 815             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 816
 817             /* Calculate temporary vectorial force */
 818             tx               = _mm_mul_ps(fscal,dx13);
 819             ty               = _mm_mul_ps(fscal,dy13);
 820             tz               = _mm_mul_ps(fscal,dz13);
 821
 822             /* Update vectorial force */
 823             fix1             = _mm_add_ps(fix1,tx);
 824             fiy1             = _mm_add_ps(fiy1,ty);
 825             fiz1             = _mm_add_ps(fiz1,tz);
 826
 827             fjx3             = _mm_add_ps(fjx3,tx);
 828             fjy3             = _mm_add_ps(fjy3,ty);
 829             fjz3             = _mm_add_ps(fjz3,tz);
 830
 831             }
 832
 833             /**************************
 834              * CALCULATE INTERACTIONS *
 835              **************************/
 836
 837             if (gmx_mm_any_lt(rsq21,rcutoff2))
 838             {
 839
 840             /* REACTION-FIELD ELECTROSTATICS */
 841             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 842             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 843
 844             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 845
 846             /* Update potential sum for this i atom from the interaction with this j atom. */
 847             velec            = _mm_and_ps(velec,cutoff_mask);
 848             velec            = _mm_andnot_ps(dummy_mask,velec);
 849             velecsum         = _mm_add_ps(velecsum,velec);
 850
 851             fscal            = felec;
 852
 853             fscal            = _mm_and_ps(fscal,cutoff_mask);
 854
 855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 856
 857             /* Calculate temporary vectorial force */
 858             tx               = _mm_mul_ps(fscal,dx21);
 859             ty               = _mm_mul_ps(fscal,dy21);
 860             tz               = _mm_mul_ps(fscal,dz21);
 861
 862             /* Update vectorial force */
 863             fix2             = _mm_add_ps(fix2,tx);
 864             fiy2             = _mm_add_ps(fiy2,ty);
 865             fiz2             = _mm_add_ps(fiz2,tz);
 866
 867             fjx1             = _mm_add_ps(fjx1,tx);
 868             fjy1             = _mm_add_ps(fjy1,ty);
 869             fjz1             = _mm_add_ps(fjz1,tz);
 870
 871             }
 872
 873             /**************************
 874              * CALCULATE INTERACTIONS *
 875              **************************/
 876
 877             if (gmx_mm_any_lt(rsq22,rcutoff2))
 878             {
 879
 880             /* REACTION-FIELD ELECTROSTATICS */
 881             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 882             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 883
 884             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 885
 886             /* Update potential sum for this i atom from the interaction with this j atom. */
 887             velec            = _mm_and_ps(velec,cutoff_mask);
 888             velec            = _mm_andnot_ps(dummy_mask,velec);
 889             velecsum         = _mm_add_ps(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _mm_and_ps(fscal,cutoff_mask);
 894
 895             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 896
 897             /* Calculate temporary vectorial force */
 898             tx               = _mm_mul_ps(fscal,dx22);
 899             ty               = _mm_mul_ps(fscal,dy22);
 900             tz               = _mm_mul_ps(fscal,dz22);
 901
 902             /* Update vectorial force */
 903             fix2             = _mm_add_ps(fix2,tx);
 904             fiy2             = _mm_add_ps(fiy2,ty);
 905             fiz2             = _mm_add_ps(fiz2,tz);
 906
 907             fjx2             = _mm_add_ps(fjx2,tx);
 908             fjy2             = _mm_add_ps(fjy2,ty);
 909             fjz2             = _mm_add_ps(fjz2,tz);
 910
 911             }
 912
 913             /**************************
 914              * CALCULATE INTERACTIONS *
 915              **************************/
 916
 917             if (gmx_mm_any_lt(rsq23,rcutoff2))
 918             {
 919
 920             /* REACTION-FIELD ELECTROSTATICS */
 921             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 922             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 923
 924             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 925
 926             /* Update potential sum for this i atom from the interaction with this j atom. */
 927             velec            = _mm_and_ps(velec,cutoff_mask);
 928             velec            = _mm_andnot_ps(dummy_mask,velec);
 929             velecsum         = _mm_add_ps(velecsum,velec);
 930
 931             fscal            = felec;
 932
 933             fscal            = _mm_and_ps(fscal,cutoff_mask);
 934
 935             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 936
 937             /* Calculate temporary vectorial force */
 938             tx               = _mm_mul_ps(fscal,dx23);
 939             ty               = _mm_mul_ps(fscal,dy23);
 940             tz               = _mm_mul_ps(fscal,dz23);
 941
 942             /* Update vectorial force */
 943             fix2             = _mm_add_ps(fix2,tx);
 944             fiy2             = _mm_add_ps(fiy2,ty);
 945             fiz2             = _mm_add_ps(fiz2,tz);
 946
 947             fjx3             = _mm_add_ps(fjx3,tx);
 948             fjy3             = _mm_add_ps(fjy3,ty);
 949             fjz3             = _mm_add_ps(fjz3,tz);
 950
 951             }
 952
 953             /**************************
 954              * CALCULATE INTERACTIONS *
 955              **************************/
 956
 957             if (gmx_mm_any_lt(rsq31,rcutoff2))
 958             {
 959
 960             /* REACTION-FIELD ELECTROSTATICS */
 961             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 962             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 963
 964             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 965
 966             /* Update potential sum for this i atom from the interaction with this j atom. */
 967             velec            = _mm_and_ps(velec,cutoff_mask);
 968             velec            = _mm_andnot_ps(dummy_mask,velec);
 969             velecsum         = _mm_add_ps(velecsum,velec);
 970
 971             fscal            = felec;
 972
 973             fscal            = _mm_and_ps(fscal,cutoff_mask);
 974
 975             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 976
 977             /* Calculate temporary vectorial force */
 978             tx               = _mm_mul_ps(fscal,dx31);
 979             ty               = _mm_mul_ps(fscal,dy31);
 980             tz               = _mm_mul_ps(fscal,dz31);
 981
 982             /* Update vectorial force */
 983             fix3             = _mm_add_ps(fix3,tx);
 984             fiy3             = _mm_add_ps(fiy3,ty);
 985             fiz3             = _mm_add_ps(fiz3,tz);
 986
 987             fjx1             = _mm_add_ps(fjx1,tx);
 988             fjy1             = _mm_add_ps(fjy1,ty);
 989             fjz1             = _mm_add_ps(fjz1,tz);
 990
 991             }
 992
 993             /**************************
 994              * CALCULATE INTERACTIONS *
 995              **************************/
 996
 997             if (gmx_mm_any_lt(rsq32,rcutoff2))
 998             {
 999
1000             /* REACTION-FIELD ELECTROSTATICS */
1001             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1002             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1003
1004             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1005
1006             /* Update potential sum for this i atom from the interaction with this j atom. */
1007             velec            = _mm_and_ps(velec,cutoff_mask);
1008             velec            = _mm_andnot_ps(dummy_mask,velec);
1009             velecsum         = _mm_add_ps(velecsum,velec);
1010
1011             fscal            = felec;
1012
1013             fscal            = _mm_and_ps(fscal,cutoff_mask);
1014
1015             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1016
1017             /* Calculate temporary vectorial force */
1018             tx               = _mm_mul_ps(fscal,dx32);
1019             ty               = _mm_mul_ps(fscal,dy32);
1020             tz               = _mm_mul_ps(fscal,dz32);
1021
1022             /* Update vectorial force */
1023             fix3             = _mm_add_ps(fix3,tx);
1024             fiy3             = _mm_add_ps(fiy3,ty);
1025             fiz3             = _mm_add_ps(fiz3,tz);
1026
1027             fjx2             = _mm_add_ps(fjx2,tx);
1028             fjy2             = _mm_add_ps(fjy2,ty);
1029             fjz2             = _mm_add_ps(fjz2,tz);
1030
1031             }
1032
1033             /**************************
1034              * CALCULATE INTERACTIONS *
1035              **************************/
1036
1037             if (gmx_mm_any_lt(rsq33,rcutoff2))
1038             {
1039
1040             /* REACTION-FIELD ELECTROSTATICS */
1041             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1042             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1043
1044             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1045
1046             /* Update potential sum for this i atom from the interaction with this j atom. */
1047             velec            = _mm_and_ps(velec,cutoff_mask);
1048             velec            = _mm_andnot_ps(dummy_mask,velec);
1049             velecsum         = _mm_add_ps(velecsum,velec);
1050
1051             fscal            = felec;
1052
1053             fscal            = _mm_and_ps(fscal,cutoff_mask);
1054
1055             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1056
1057             /* Calculate temporary vectorial force */
1058             tx               = _mm_mul_ps(fscal,dx33);
1059             ty               = _mm_mul_ps(fscal,dy33);
1060             tz               = _mm_mul_ps(fscal,dz33);
1061
1062             /* Update vectorial force */
1063             fix3             = _mm_add_ps(fix3,tx);
1064             fiy3             = _mm_add_ps(fiy3,ty);
1065             fiz3             = _mm_add_ps(fiz3,tz);
1066
1067             fjx3             = _mm_add_ps(fjx3,tx);
1068             fjy3             = _mm_add_ps(fjy3,ty);
1069             fjz3             = _mm_add_ps(fjz3,tz);
1070
1071             }
1072
1073             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1074             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1075             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1076             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1077
1078             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1079                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1080
1081             /* Inner loop uses 324 flops */
1082         }
1083
1084         /* End of innermost loop */
1085
1086         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1087                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1088
1089         ggid                        = gid[iidx];
1090         /* Update potential energies */
1091         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1092
1093         /* Increment number of inner iterations */
1094         inneriter                  += j_index_end - j_index_start;
1095
1096         /* Outer loop uses 19 flops */
1097     }
1098
1099     /* Increment number of outer iterations */
1100     outeriter        += nri;
1101
1102     /* Update outer/inner flops */
1103
1104     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*324);
1105 }
1106 /*
1107  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_single
1108  * Electrostatics interaction: ReactionField
1109  * VdW interaction:            None
1110  * Geometry:                   Water4-Water4
1111  * Calculate force/pot:        Force
1112  */
1113 void
1114 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_single
1115                     (t_nblist * gmx_restrict                nlist,
1116                      rvec * gmx_restrict                    xx,
1117                      rvec * gmx_restrict                    ff,
1118                      t_forcerec * gmx_restrict              fr,
1119                      t_mdatoms * gmx_restrict               mdatoms,
1120                      nb_kernel_data_t * gmx_restrict        kernel_data,
1121                      t_nrnb * gmx_restrict                  nrnb)
1122 {
1123     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1124      * just 0 for non-waters.
1125      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1126      * jnr indices corresponding to data put in the four positions in the SIMD register.
1127      */
1128     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1129     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1130     int              jnrA,jnrB,jnrC,jnrD;
1131     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1132     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1133     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1134     real             rcutoff_scalar;
1135     real             *shiftvec,*fshift,*x,*f;
1136     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1137     real             scratch[4*DIM];
1138     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1139     int              vdwioffset1;
1140     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1141     int              vdwioffset2;
1142     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1143     int              vdwioffset3;
1144     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1145     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1146     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1147     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1148     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1149     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1150     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1151     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1152     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1153     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1154     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1155     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1156     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1157     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1158     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1159     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1160     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1161     real             *charge;
1162     __m128           dummy_mask,cutoff_mask;
1163     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1164     __m128           one     = _mm_set1_ps(1.0);
1165     __m128           two     = _mm_set1_ps(2.0);
1166     x                = xx[0];
1167     f                = ff[0];
1168
1169     nri              = nlist->nri;
1170     iinr             = nlist->iinr;
1171     jindex           = nlist->jindex;
1172     jjnr             = nlist->jjnr;
1173     shiftidx         = nlist->shift;
1174     gid              = nlist->gid;
1175     shiftvec         = fr->shift_vec[0];
1176     fshift           = fr->fshift[0];
1177     facel            = _mm_set1_ps(fr->epsfac);
1178     charge           = mdatoms->chargeA;
1179     krf              = _mm_set1_ps(fr->ic->k_rf);
1180     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1181     crf              = _mm_set1_ps(fr->ic->c_rf);
1182
1183     /* Setup water-specific parameters */
1184     inr              = nlist->iinr[0];
1185     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1186     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1187     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1188
1189     jq1              = _mm_set1_ps(charge[inr+1]);
1190     jq2              = _mm_set1_ps(charge[inr+2]);
1191     jq3              = _mm_set1_ps(charge[inr+3]);
1192     qq11             = _mm_mul_ps(iq1,jq1);
1193     qq12             = _mm_mul_ps(iq1,jq2);
1194     qq13             = _mm_mul_ps(iq1,jq3);
1195     qq21             = _mm_mul_ps(iq2,jq1);
1196     qq22             = _mm_mul_ps(iq2,jq2);
1197     qq23             = _mm_mul_ps(iq2,jq3);
1198     qq31             = _mm_mul_ps(iq3,jq1);
1199     qq32             = _mm_mul_ps(iq3,jq2);
1200     qq33             = _mm_mul_ps(iq3,jq3);
1201
1202     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1203     rcutoff_scalar   = fr->rcoulomb;
1204     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1205     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1206
1207     /* Avoid stupid compiler warnings */
1208     jnrA = jnrB = jnrC = jnrD = 0;
1209     j_coord_offsetA = 0;
1210     j_coord_offsetB = 0;
1211     j_coord_offsetC = 0;
1212     j_coord_offsetD = 0;
1213
1214     outeriter        = 0;
1215     inneriter        = 0;
1216
1217     for(iidx=0;iidx<4*DIM;iidx++)
1218     {
1219         scratch[iidx] = 0.0;
1220     }
1221
1222     /* Start outer loop over neighborlists */
1223     for(iidx=0; iidx<nri; iidx++)
1224     {
1225         /* Load shift vector for this list */
1226         i_shift_offset   = DIM*shiftidx[iidx];
1227
1228         /* Load limits for loop over neighbors */
1229         j_index_start    = jindex[iidx];
1230         j_index_end      = jindex[iidx+1];
1231
1232         /* Get outer coordinate index */
1233         inr              = iinr[iidx];
1234         i_coord_offset   = DIM*inr;
1235
1236         /* Load i particle coords and add shift vector */
1237         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1238                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1239
1240         fix1             = _mm_setzero_ps();
1241         fiy1             = _mm_setzero_ps();
1242         fiz1             = _mm_setzero_ps();
1243         fix2             = _mm_setzero_ps();
1244         fiy2             = _mm_setzero_ps();
1245         fiz2             = _mm_setzero_ps();
1246         fix3             = _mm_setzero_ps();
1247         fiy3             = _mm_setzero_ps();
1248         fiz3             = _mm_setzero_ps();
1249
1250         /* Start inner kernel loop */
1251         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1252         {
1253
1254             /* Get j neighbor index, and coordinate index */
1255             jnrA             = jjnr[jidx];
1256             jnrB             = jjnr[jidx+1];
1257             jnrC             = jjnr[jidx+2];
1258             jnrD             = jjnr[jidx+3];
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             fjptrA             = f+j_coord_offsetA;
1628             fjptrB             = f+j_coord_offsetB;
1629             fjptrC             = f+j_coord_offsetC;
1630             fjptrD             = f+j_coord_offsetD;
1631
1632             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1633                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1634
1635             /* Inner loop uses 270 flops */
1636         }
1637
1638         if(jidx<j_index_end)
1639         {
1640
1641             /* Get j neighbor index, and coordinate index */
1642             jnrlistA         = jjnr[jidx];
1643             jnrlistB         = jjnr[jidx+1];
1644             jnrlistC         = jjnr[jidx+2];
1645             jnrlistD         = jjnr[jidx+3];
1646             /* Sign of each element will be negative for non-real atoms.
1647              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1648              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1649              */
1650             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1651             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1652             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1653             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1654             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1655             j_coord_offsetA  = DIM*jnrA;
1656             j_coord_offsetB  = DIM*jnrB;
1657             j_coord_offsetC  = DIM*jnrC;
1658             j_coord_offsetD  = DIM*jnrD;
1659
1660             /* load j atom coordinates */
1661             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1662                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1663                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1664
1665             /* Calculate displacement vector */
1666             dx11             = _mm_sub_ps(ix1,jx1);
1667             dy11             = _mm_sub_ps(iy1,jy1);
1668             dz11             = _mm_sub_ps(iz1,jz1);
1669             dx12             = _mm_sub_ps(ix1,jx2);
1670             dy12             = _mm_sub_ps(iy1,jy2);
1671             dz12             = _mm_sub_ps(iz1,jz2);
1672             dx13             = _mm_sub_ps(ix1,jx3);
1673             dy13             = _mm_sub_ps(iy1,jy3);
1674             dz13             = _mm_sub_ps(iz1,jz3);
1675             dx21             = _mm_sub_ps(ix2,jx1);
1676             dy21             = _mm_sub_ps(iy2,jy1);
1677             dz21             = _mm_sub_ps(iz2,jz1);
1678             dx22             = _mm_sub_ps(ix2,jx2);
1679             dy22             = _mm_sub_ps(iy2,jy2);
1680             dz22             = _mm_sub_ps(iz2,jz2);
1681             dx23             = _mm_sub_ps(ix2,jx3);
1682             dy23             = _mm_sub_ps(iy2,jy3);
1683             dz23             = _mm_sub_ps(iz2,jz3);
1684             dx31             = _mm_sub_ps(ix3,jx1);
1685             dy31             = _mm_sub_ps(iy3,jy1);
1686             dz31             = _mm_sub_ps(iz3,jz1);
1687             dx32             = _mm_sub_ps(ix3,jx2);
1688             dy32             = _mm_sub_ps(iy3,jy2);
1689             dz32             = _mm_sub_ps(iz3,jz2);
1690             dx33             = _mm_sub_ps(ix3,jx3);
1691             dy33             = _mm_sub_ps(iy3,jy3);
1692             dz33             = _mm_sub_ps(iz3,jz3);
1693
1694             /* Calculate squared distance and things based on it */
1695             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1696             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1697             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1698             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1699             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1700             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1701             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1702             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1703             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1704
1705             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1706             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1707             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1708             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1709             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1710             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1711             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1712             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1713             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1714
1715             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1716             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1717             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1718             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1719             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1720             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1721             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1722             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1723             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1724
1725             fjx1             = _mm_setzero_ps();
1726             fjy1             = _mm_setzero_ps();
1727             fjz1             = _mm_setzero_ps();
1728             fjx2             = _mm_setzero_ps();
1729             fjy2             = _mm_setzero_ps();
1730             fjz2             = _mm_setzero_ps();
1731             fjx3             = _mm_setzero_ps();
1732             fjy3             = _mm_setzero_ps();
1733             fjz3             = _mm_setzero_ps();
1734
1735             /**************************
1736              * CALCULATE INTERACTIONS *
1737              **************************/
1738
1739             if (gmx_mm_any_lt(rsq11,rcutoff2))
1740             {
1741
1742             /* REACTION-FIELD ELECTROSTATICS */
1743             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1744
1745             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1746
1747             fscal            = felec;
1748
1749             fscal            = _mm_and_ps(fscal,cutoff_mask);
1750
1751             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1752
1753             /* Calculate temporary vectorial force */
1754             tx               = _mm_mul_ps(fscal,dx11);
1755             ty               = _mm_mul_ps(fscal,dy11);
1756             tz               = _mm_mul_ps(fscal,dz11);
1757
1758             /* Update vectorial force */
1759             fix1             = _mm_add_ps(fix1,tx);
1760             fiy1             = _mm_add_ps(fiy1,ty);
1761             fiz1             = _mm_add_ps(fiz1,tz);
1762
1763             fjx1             = _mm_add_ps(fjx1,tx);
1764             fjy1             = _mm_add_ps(fjy1,ty);
1765             fjz1             = _mm_add_ps(fjz1,tz);
1766
1767             }
1768
1769             /**************************
1770              * CALCULATE INTERACTIONS *
1771              **************************/
1772
1773             if (gmx_mm_any_lt(rsq12,rcutoff2))
1774             {
1775
1776             /* REACTION-FIELD ELECTROSTATICS */
1777             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1778
1779             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1780
1781             fscal            = felec;
1782
1783             fscal            = _mm_and_ps(fscal,cutoff_mask);
1784
1785             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1786
1787             /* Calculate temporary vectorial force */
1788             tx               = _mm_mul_ps(fscal,dx12);
1789             ty               = _mm_mul_ps(fscal,dy12);
1790             tz               = _mm_mul_ps(fscal,dz12);
1791
1792             /* Update vectorial force */
1793             fix1             = _mm_add_ps(fix1,tx);
1794             fiy1             = _mm_add_ps(fiy1,ty);
1795             fiz1             = _mm_add_ps(fiz1,tz);
1796
1797             fjx2             = _mm_add_ps(fjx2,tx);
1798             fjy2             = _mm_add_ps(fjy2,ty);
1799             fjz2             = _mm_add_ps(fjz2,tz);
1800
1801             }
1802
1803             /**************************
1804              * CALCULATE INTERACTIONS *
1805              **************************/
1806
1807             if (gmx_mm_any_lt(rsq13,rcutoff2))
1808             {
1809
1810             /* REACTION-FIELD ELECTROSTATICS */
1811             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1812
1813             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1814
1815             fscal            = felec;
1816
1817             fscal            = _mm_and_ps(fscal,cutoff_mask);
1818
1819             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1820
1821             /* Calculate temporary vectorial force */
1822             tx               = _mm_mul_ps(fscal,dx13);
1823             ty               = _mm_mul_ps(fscal,dy13);
1824             tz               = _mm_mul_ps(fscal,dz13);
1825
1826             /* Update vectorial force */
1827             fix1             = _mm_add_ps(fix1,tx);
1828             fiy1             = _mm_add_ps(fiy1,ty);
1829             fiz1             = _mm_add_ps(fiz1,tz);
1830
1831             fjx3             = _mm_add_ps(fjx3,tx);
1832             fjy3             = _mm_add_ps(fjy3,ty);
1833             fjz3             = _mm_add_ps(fjz3,tz);
1834
1835             }
1836
1837             /**************************
1838              * CALCULATE INTERACTIONS *
1839              **************************/
1840
1841             if (gmx_mm_any_lt(rsq21,rcutoff2))
1842             {
1843
1844             /* REACTION-FIELD ELECTROSTATICS */
1845             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1846
1847             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1848
1849             fscal            = felec;
1850
1851             fscal            = _mm_and_ps(fscal,cutoff_mask);
1852
1853             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1854
1855             /* Calculate temporary vectorial force */
1856             tx               = _mm_mul_ps(fscal,dx21);
1857             ty               = _mm_mul_ps(fscal,dy21);
1858             tz               = _mm_mul_ps(fscal,dz21);
1859
1860             /* Update vectorial force */
1861             fix2             = _mm_add_ps(fix2,tx);
1862             fiy2             = _mm_add_ps(fiy2,ty);
1863             fiz2             = _mm_add_ps(fiz2,tz);
1864
1865             fjx1             = _mm_add_ps(fjx1,tx);
1866             fjy1             = _mm_add_ps(fjy1,ty);
1867             fjz1             = _mm_add_ps(fjz1,tz);
1868
1869             }
1870
1871             /**************************
1872              * CALCULATE INTERACTIONS *
1873              **************************/
1874
1875             if (gmx_mm_any_lt(rsq22,rcutoff2))
1876             {
1877
1878             /* REACTION-FIELD ELECTROSTATICS */
1879             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1880
1881             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1882
1883             fscal            = felec;
1884
1885             fscal            = _mm_and_ps(fscal,cutoff_mask);
1886
1887             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1888
1889             /* Calculate temporary vectorial force */
1890             tx               = _mm_mul_ps(fscal,dx22);
1891             ty               = _mm_mul_ps(fscal,dy22);
1892             tz               = _mm_mul_ps(fscal,dz22);
1893
1894             /* Update vectorial force */
1895             fix2             = _mm_add_ps(fix2,tx);
1896             fiy2             = _mm_add_ps(fiy2,ty);
1897             fiz2             = _mm_add_ps(fiz2,tz);
1898
1899             fjx2             = _mm_add_ps(fjx2,tx);
1900             fjy2             = _mm_add_ps(fjy2,ty);
1901             fjz2             = _mm_add_ps(fjz2,tz);
1902
1903             }
1904
1905             /**************************
1906              * CALCULATE INTERACTIONS *
1907              **************************/
1908
1909             if (gmx_mm_any_lt(rsq23,rcutoff2))
1910             {
1911
1912             /* REACTION-FIELD ELECTROSTATICS */
1913             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1914
1915             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1916
1917             fscal            = felec;
1918
1919             fscal            = _mm_and_ps(fscal,cutoff_mask);
1920
1921             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1922
1923             /* Calculate temporary vectorial force */
1924             tx               = _mm_mul_ps(fscal,dx23);
1925             ty               = _mm_mul_ps(fscal,dy23);
1926             tz               = _mm_mul_ps(fscal,dz23);
1927
1928             /* Update vectorial force */
1929             fix2             = _mm_add_ps(fix2,tx);
1930             fiy2             = _mm_add_ps(fiy2,ty);
1931             fiz2             = _mm_add_ps(fiz2,tz);
1932
1933             fjx3             = _mm_add_ps(fjx3,tx);
1934             fjy3             = _mm_add_ps(fjy3,ty);
1935             fjz3             = _mm_add_ps(fjz3,tz);
1936
1937             }
1938
1939             /**************************
1940              * CALCULATE INTERACTIONS *
1941              **************************/
1942
1943             if (gmx_mm_any_lt(rsq31,rcutoff2))
1944             {
1945
1946             /* REACTION-FIELD ELECTROSTATICS */
1947             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1948
1949             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1950
1951             fscal            = felec;
1952
1953             fscal            = _mm_and_ps(fscal,cutoff_mask);
1954
1955             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1956
1957             /* Calculate temporary vectorial force */
1958             tx               = _mm_mul_ps(fscal,dx31);
1959             ty               = _mm_mul_ps(fscal,dy31);
1960             tz               = _mm_mul_ps(fscal,dz31);
1961
1962             /* Update vectorial force */
1963             fix3             = _mm_add_ps(fix3,tx);
1964             fiy3             = _mm_add_ps(fiy3,ty);
1965             fiz3             = _mm_add_ps(fiz3,tz);
1966
1967             fjx1             = _mm_add_ps(fjx1,tx);
1968             fjy1             = _mm_add_ps(fjy1,ty);
1969             fjz1             = _mm_add_ps(fjz1,tz);
1970
1971             }
1972
1973             /**************************
1974              * CALCULATE INTERACTIONS *
1975              **************************/
1976
1977             if (gmx_mm_any_lt(rsq32,rcutoff2))
1978             {
1979
1980             /* REACTION-FIELD ELECTROSTATICS */
1981             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1982
1983             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1984
1985             fscal            = felec;
1986
1987             fscal            = _mm_and_ps(fscal,cutoff_mask);
1988
1989             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1990
1991             /* Calculate temporary vectorial force */
1992             tx               = _mm_mul_ps(fscal,dx32);
1993             ty               = _mm_mul_ps(fscal,dy32);
1994             tz               = _mm_mul_ps(fscal,dz32);
1995
1996             /* Update vectorial force */
1997             fix3             = _mm_add_ps(fix3,tx);
1998             fiy3             = _mm_add_ps(fiy3,ty);
1999             fiz3             = _mm_add_ps(fiz3,tz);
2000
2001             fjx2             = _mm_add_ps(fjx2,tx);
2002             fjy2             = _mm_add_ps(fjy2,ty);
2003             fjz2             = _mm_add_ps(fjz2,tz);
2004
2005             }
2006
2007             /**************************
2008              * CALCULATE INTERACTIONS *
2009              **************************/
2010
2011             if (gmx_mm_any_lt(rsq33,rcutoff2))
2012             {
2013
2014             /* REACTION-FIELD ELECTROSTATICS */
2015             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2016
2017             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2018
2019             fscal            = felec;
2020
2021             fscal            = _mm_and_ps(fscal,cutoff_mask);
2022
2023             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2024
2025             /* Calculate temporary vectorial force */
2026             tx               = _mm_mul_ps(fscal,dx33);
2027             ty               = _mm_mul_ps(fscal,dy33);
2028             tz               = _mm_mul_ps(fscal,dz33);
2029
2030             /* Update vectorial force */
2031             fix3             = _mm_add_ps(fix3,tx);
2032             fiy3             = _mm_add_ps(fiy3,ty);
2033             fiz3             = _mm_add_ps(fiz3,tz);
2034
2035             fjx3             = _mm_add_ps(fjx3,tx);
2036             fjy3             = _mm_add_ps(fjy3,ty);
2037             fjz3             = _mm_add_ps(fjz3,tz);
2038
2039             }
2040
2041             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2042             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2043             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2044             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2045
2046             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
2047                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2048
2049             /* Inner loop uses 270 flops */
2050         }
2051
2052         /* End of innermost loop */
2053
2054         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2055                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2056
2057         /* Increment number of inner iterations */
2058         inneriter                  += j_index_end - j_index_start;
2059
2060         /* Outer loop uses 18 flops */
2061     }
2062
2063     /* Increment number of outer iterations */
2064     outeriter        += nri;
2065
2066     /* Update outer/inner flops */
2067
2068     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
2069 }