src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecEw_VdwNone_GeomW4W4_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_sse2_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              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     __m128i          ewitab;
  93     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  94     real             *ewtab;
  95     __m128           dummy_mask,cutoff_mask;
  96     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  97     __m128           one     = _mm_set1_ps(1.0);
  98     __m128           two     = _mm_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_ps(fr->epsfac);
 111     charge           = mdatoms->chargeA;
 112
 113     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 114     ewtab            = fr->ic->tabq_coul_FDV0;
 115     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 116     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 117
 118     /* Setup water-specific parameters */
 119     inr              = nlist->iinr[0];
 120     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 121     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 122     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 123
 124     jq1              = _mm_set1_ps(charge[inr+1]);
 125     jq2              = _mm_set1_ps(charge[inr+2]);
 126     jq3              = _mm_set1_ps(charge[inr+3]);
 127     qq11             = _mm_mul_ps(iq1,jq1);
 128     qq12             = _mm_mul_ps(iq1,jq2);
 129     qq13             = _mm_mul_ps(iq1,jq3);
 130     qq21             = _mm_mul_ps(iq2,jq1);
 131     qq22             = _mm_mul_ps(iq2,jq2);
 132     qq23             = _mm_mul_ps(iq2,jq3);
 133     qq31             = _mm_mul_ps(iq3,jq1);
 134     qq32             = _mm_mul_ps(iq3,jq2);
 135     qq33             = _mm_mul_ps(iq3,jq3);
 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             r11              = _mm_mul_ps(rsq11,rinv11);
 277
 278             /* EWALD ELECTROSTATICS */
 279
 280             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 281             ewrt             = _mm_mul_ps(r11,ewtabscale);
 282             ewitab           = _mm_cvttps_epi32(ewrt);
 283             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 284             ewitab           = _mm_slli_epi32(ewitab,2);
 285             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 286             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 287             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 288             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 289             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 290             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 291             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 292             velec            = _mm_mul_ps(qq11,_mm_sub_ps(rinv11,velec));
 293             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
 294
 295             /* Update potential sum for this i atom from the interaction with this j atom. */
 296             velecsum         = _mm_add_ps(velecsum,velec);
 297
 298             fscal            = felec;
 299
 300             /* Calculate temporary vectorial force */
 301             tx               = _mm_mul_ps(fscal,dx11);
 302             ty               = _mm_mul_ps(fscal,dy11);
 303             tz               = _mm_mul_ps(fscal,dz11);
 304
 305             /* Update vectorial force */
 306             fix1             = _mm_add_ps(fix1,tx);
 307             fiy1             = _mm_add_ps(fiy1,ty);
 308             fiz1             = _mm_add_ps(fiz1,tz);
 309
 310             fjx1             = _mm_add_ps(fjx1,tx);
 311             fjy1             = _mm_add_ps(fjy1,ty);
 312             fjz1             = _mm_add_ps(fjz1,tz);
 313
 314             /**************************
 315              * CALCULATE INTERACTIONS *
 316              **************************/
 317
 318             r12              = _mm_mul_ps(rsq12,rinv12);
 319
 320             /* EWALD ELECTROSTATICS */
 321
 322             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 323             ewrt             = _mm_mul_ps(r12,ewtabscale);
 324             ewitab           = _mm_cvttps_epi32(ewrt);
 325             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 326             ewitab           = _mm_slli_epi32(ewitab,2);
 327             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 328             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 329             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 330             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 331             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 332             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 333             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 334             velec            = _mm_mul_ps(qq12,_mm_sub_ps(rinv12,velec));
 335             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
 336
 337             /* Update potential sum for this i atom from the interaction with this j atom. */
 338             velecsum         = _mm_add_ps(velecsum,velec);
 339
 340             fscal            = felec;
 341
 342             /* Calculate temporary vectorial force */
 343             tx               = _mm_mul_ps(fscal,dx12);
 344             ty               = _mm_mul_ps(fscal,dy12);
 345             tz               = _mm_mul_ps(fscal,dz12);
 346
 347             /* Update vectorial force */
 348             fix1             = _mm_add_ps(fix1,tx);
 349             fiy1             = _mm_add_ps(fiy1,ty);
 350             fiz1             = _mm_add_ps(fiz1,tz);
 351
 352             fjx2             = _mm_add_ps(fjx2,tx);
 353             fjy2             = _mm_add_ps(fjy2,ty);
 354             fjz2             = _mm_add_ps(fjz2,tz);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             r13              = _mm_mul_ps(rsq13,rinv13);
 361
 362             /* EWALD ELECTROSTATICS */
 363
 364             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 365             ewrt             = _mm_mul_ps(r13,ewtabscale);
 366             ewitab           = _mm_cvttps_epi32(ewrt);
 367             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 368             ewitab           = _mm_slli_epi32(ewitab,2);
 369             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 370             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 371             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 372             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 373             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 374             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 375             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 376             velec            = _mm_mul_ps(qq13,_mm_sub_ps(rinv13,velec));
 377             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velecsum         = _mm_add_ps(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             /* Calculate temporary vectorial force */
 385             tx               = _mm_mul_ps(fscal,dx13);
 386             ty               = _mm_mul_ps(fscal,dy13);
 387             tz               = _mm_mul_ps(fscal,dz13);
 388
 389             /* Update vectorial force */
 390             fix1             = _mm_add_ps(fix1,tx);
 391             fiy1             = _mm_add_ps(fiy1,ty);
 392             fiz1             = _mm_add_ps(fiz1,tz);
 393
 394             fjx3             = _mm_add_ps(fjx3,tx);
 395             fjy3             = _mm_add_ps(fjy3,ty);
 396             fjz3             = _mm_add_ps(fjz3,tz);
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             r21              = _mm_mul_ps(rsq21,rinv21);
 403
 404             /* EWALD ELECTROSTATICS */
 405
 406             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 407             ewrt             = _mm_mul_ps(r21,ewtabscale);
 408             ewitab           = _mm_cvttps_epi32(ewrt);
 409             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 410             ewitab           = _mm_slli_epi32(ewitab,2);
 411             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 412             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 413             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 414             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 415             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 416             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 417             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 418             velec            = _mm_mul_ps(qq21,_mm_sub_ps(rinv21,velec));
 419             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
 420
 421             /* Update potential sum for this i atom from the interaction with this j atom. */
 422             velecsum         = _mm_add_ps(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             /* Calculate temporary vectorial force */
 427             tx               = _mm_mul_ps(fscal,dx21);
 428             ty               = _mm_mul_ps(fscal,dy21);
 429             tz               = _mm_mul_ps(fscal,dz21);
 430
 431             /* Update vectorial force */
 432             fix2             = _mm_add_ps(fix2,tx);
 433             fiy2             = _mm_add_ps(fiy2,ty);
 434             fiz2             = _mm_add_ps(fiz2,tz);
 435
 436             fjx1             = _mm_add_ps(fjx1,tx);
 437             fjy1             = _mm_add_ps(fjy1,ty);
 438             fjz1             = _mm_add_ps(fjz1,tz);
 439
 440             /**************************
 441              * CALCULATE INTERACTIONS *
 442              **************************/
 443
 444             r22              = _mm_mul_ps(rsq22,rinv22);
 445
 446             /* EWALD ELECTROSTATICS */
 447
 448             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 449             ewrt             = _mm_mul_ps(r22,ewtabscale);
 450             ewitab           = _mm_cvttps_epi32(ewrt);
 451             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 452             ewitab           = _mm_slli_epi32(ewitab,2);
 453             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 454             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 455             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 456             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 457             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 458             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 459             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 460             velec            = _mm_mul_ps(qq22,_mm_sub_ps(rinv22,velec));
 461             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
 462
 463             /* Update potential sum for this i atom from the interaction with this j atom. */
 464             velecsum         = _mm_add_ps(velecsum,velec);
 465
 466             fscal            = felec;
 467
 468             /* Calculate temporary vectorial force */
 469             tx               = _mm_mul_ps(fscal,dx22);
 470             ty               = _mm_mul_ps(fscal,dy22);
 471             tz               = _mm_mul_ps(fscal,dz22);
 472
 473             /* Update vectorial force */
 474             fix2             = _mm_add_ps(fix2,tx);
 475             fiy2             = _mm_add_ps(fiy2,ty);
 476             fiz2             = _mm_add_ps(fiz2,tz);
 477
 478             fjx2             = _mm_add_ps(fjx2,tx);
 479             fjy2             = _mm_add_ps(fjy2,ty);
 480             fjz2             = _mm_add_ps(fjz2,tz);
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             r23              = _mm_mul_ps(rsq23,rinv23);
 487
 488             /* EWALD ELECTROSTATICS */
 489
 490             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 491             ewrt             = _mm_mul_ps(r23,ewtabscale);
 492             ewitab           = _mm_cvttps_epi32(ewrt);
 493             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 494             ewitab           = _mm_slli_epi32(ewitab,2);
 495             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 496             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 497             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 498             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 499             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 500             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 501             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 502             velec            = _mm_mul_ps(qq23,_mm_sub_ps(rinv23,velec));
 503             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
 504
 505             /* Update potential sum for this i atom from the interaction with this j atom. */
 506             velecsum         = _mm_add_ps(velecsum,velec);
 507
 508             fscal            = felec;
 509
 510             /* Calculate temporary vectorial force */
 511             tx               = _mm_mul_ps(fscal,dx23);
 512             ty               = _mm_mul_ps(fscal,dy23);
 513             tz               = _mm_mul_ps(fscal,dz23);
 514
 515             /* Update vectorial force */
 516             fix2             = _mm_add_ps(fix2,tx);
 517             fiy2             = _mm_add_ps(fiy2,ty);
 518             fiz2             = _mm_add_ps(fiz2,tz);
 519
 520             fjx3             = _mm_add_ps(fjx3,tx);
 521             fjy3             = _mm_add_ps(fjy3,ty);
 522             fjz3             = _mm_add_ps(fjz3,tz);
 523
 524             /**************************
 525              * CALCULATE INTERACTIONS *
 526              **************************/
 527
 528             r31              = _mm_mul_ps(rsq31,rinv31);
 529
 530             /* EWALD ELECTROSTATICS */
 531
 532             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 533             ewrt             = _mm_mul_ps(r31,ewtabscale);
 534             ewitab           = _mm_cvttps_epi32(ewrt);
 535             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 536             ewitab           = _mm_slli_epi32(ewitab,2);
 537             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 538             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 539             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 540             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 541             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 542             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 543             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 544             velec            = _mm_mul_ps(qq31,_mm_sub_ps(rinv31,velec));
 545             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
 546
 547             /* Update potential sum for this i atom from the interaction with this j atom. */
 548             velecsum         = _mm_add_ps(velecsum,velec);
 549
 550             fscal            = felec;
 551
 552             /* Calculate temporary vectorial force */
 553             tx               = _mm_mul_ps(fscal,dx31);
 554             ty               = _mm_mul_ps(fscal,dy31);
 555             tz               = _mm_mul_ps(fscal,dz31);
 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             fjx1             = _mm_add_ps(fjx1,tx);
 563             fjy1             = _mm_add_ps(fjy1,ty);
 564             fjz1             = _mm_add_ps(fjz1,tz);
 565
 566             /**************************
 567              * CALCULATE INTERACTIONS *
 568              **************************/
 569
 570             r32              = _mm_mul_ps(rsq32,rinv32);
 571
 572             /* EWALD ELECTROSTATICS */
 573
 574             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 575             ewrt             = _mm_mul_ps(r32,ewtabscale);
 576             ewitab           = _mm_cvttps_epi32(ewrt);
 577             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 578             ewitab           = _mm_slli_epi32(ewitab,2);
 579             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 580             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 581             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 582             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 583             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 584             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 585             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 586             velec            = _mm_mul_ps(qq32,_mm_sub_ps(rinv32,velec));
 587             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
 588
 589             /* Update potential sum for this i atom from the interaction with this j atom. */
 590             velecsum         = _mm_add_ps(velecsum,velec);
 591
 592             fscal            = felec;
 593
 594             /* Calculate temporary vectorial force */
 595             tx               = _mm_mul_ps(fscal,dx32);
 596             ty               = _mm_mul_ps(fscal,dy32);
 597             tz               = _mm_mul_ps(fscal,dz32);
 598
 599             /* Update vectorial force */
 600             fix3             = _mm_add_ps(fix3,tx);
 601             fiy3             = _mm_add_ps(fiy3,ty);
 602             fiz3             = _mm_add_ps(fiz3,tz);
 603
 604             fjx2             = _mm_add_ps(fjx2,tx);
 605             fjy2             = _mm_add_ps(fjy2,ty);
 606             fjz2             = _mm_add_ps(fjz2,tz);
 607
 608             /**************************
 609              * CALCULATE INTERACTIONS *
 610              **************************/
 611
 612             r33              = _mm_mul_ps(rsq33,rinv33);
 613
 614             /* EWALD ELECTROSTATICS */
 615
 616             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 617             ewrt             = _mm_mul_ps(r33,ewtabscale);
 618             ewitab           = _mm_cvttps_epi32(ewrt);
 619             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 620             ewitab           = _mm_slli_epi32(ewitab,2);
 621             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 622             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 623             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 624             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 625             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 626             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 627             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 628             velec            = _mm_mul_ps(qq33,_mm_sub_ps(rinv33,velec));
 629             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
 630
 631             /* Update potential sum for this i atom from the interaction with this j atom. */
 632             velecsum         = _mm_add_ps(velecsum,velec);
 633
 634             fscal            = felec;
 635
 636             /* Calculate temporary vectorial force */
 637             tx               = _mm_mul_ps(fscal,dx33);
 638             ty               = _mm_mul_ps(fscal,dy33);
 639             tz               = _mm_mul_ps(fscal,dz33);
 640
 641             /* Update vectorial force */
 642             fix3             = _mm_add_ps(fix3,tx);
 643             fiy3             = _mm_add_ps(fiy3,ty);
 644             fiz3             = _mm_add_ps(fiz3,tz);
 645
 646             fjx3             = _mm_add_ps(fjx3,tx);
 647             fjy3             = _mm_add_ps(fjy3,ty);
 648             fjz3             = _mm_add_ps(fjz3,tz);
 649
 650             fjptrA             = f+j_coord_offsetA;
 651             fjptrB             = f+j_coord_offsetB;
 652             fjptrC             = f+j_coord_offsetC;
 653             fjptrD             = f+j_coord_offsetD;
 654
 655             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 656                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 657
 658             /* Inner loop uses 369 flops */
 659         }
 660
 661         if(jidx<j_index_end)
 662         {
 663
 664             /* Get j neighbor index, and coordinate index */
 665             jnrlistA         = jjnr[jidx];
 666             jnrlistB         = jjnr[jidx+1];
 667             jnrlistC         = jjnr[jidx+2];
 668             jnrlistD         = jjnr[jidx+3];
 669             /* Sign of each element will be negative for non-real atoms.
 670              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 671              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 672              */
 673             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 674             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 675             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 676             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 677             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 678             j_coord_offsetA  = DIM*jnrA;
 679             j_coord_offsetB  = DIM*jnrB;
 680             j_coord_offsetC  = DIM*jnrC;
 681             j_coord_offsetD  = DIM*jnrD;
 682
 683             /* load j atom coordinates */
 684             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 685                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 686                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 687
 688             /* Calculate displacement vector */
 689             dx11             = _mm_sub_ps(ix1,jx1);
 690             dy11             = _mm_sub_ps(iy1,jy1);
 691             dz11             = _mm_sub_ps(iz1,jz1);
 692             dx12             = _mm_sub_ps(ix1,jx2);
 693             dy12             = _mm_sub_ps(iy1,jy2);
 694             dz12             = _mm_sub_ps(iz1,jz2);
 695             dx13             = _mm_sub_ps(ix1,jx3);
 696             dy13             = _mm_sub_ps(iy1,jy3);
 697             dz13             = _mm_sub_ps(iz1,jz3);
 698             dx21             = _mm_sub_ps(ix2,jx1);
 699             dy21             = _mm_sub_ps(iy2,jy1);
 700             dz21             = _mm_sub_ps(iz2,jz1);
 701             dx22             = _mm_sub_ps(ix2,jx2);
 702             dy22             = _mm_sub_ps(iy2,jy2);
 703             dz22             = _mm_sub_ps(iz2,jz2);
 704             dx23             = _mm_sub_ps(ix2,jx3);
 705             dy23             = _mm_sub_ps(iy2,jy3);
 706             dz23             = _mm_sub_ps(iz2,jz3);
 707             dx31             = _mm_sub_ps(ix3,jx1);
 708             dy31             = _mm_sub_ps(iy3,jy1);
 709             dz31             = _mm_sub_ps(iz3,jz1);
 710             dx32             = _mm_sub_ps(ix3,jx2);
 711             dy32             = _mm_sub_ps(iy3,jy2);
 712             dz32             = _mm_sub_ps(iz3,jz2);
 713             dx33             = _mm_sub_ps(ix3,jx3);
 714             dy33             = _mm_sub_ps(iy3,jy3);
 715             dz33             = _mm_sub_ps(iz3,jz3);
 716
 717             /* Calculate squared distance and things based on it */
 718             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 719             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 720             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 721             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 722             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 723             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 724             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 725             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 726             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 727
 728             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 729             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 730             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 731             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 732             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 733             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 734             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 735             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 736             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 737
 738             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 739             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 740             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 741             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 742             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 743             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 744             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 745             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 746             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 747
 748             fjx1             = _mm_setzero_ps();
 749             fjy1             = _mm_setzero_ps();
 750             fjz1             = _mm_setzero_ps();
 751             fjx2             = _mm_setzero_ps();
 752             fjy2             = _mm_setzero_ps();
 753             fjz2             = _mm_setzero_ps();
 754             fjx3             = _mm_setzero_ps();
 755             fjy3             = _mm_setzero_ps();
 756             fjz3             = _mm_setzero_ps();
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             r11              = _mm_mul_ps(rsq11,rinv11);
 763             r11              = _mm_andnot_ps(dummy_mask,r11);
 764
 765             /* EWALD ELECTROSTATICS */
 766
 767             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 768             ewrt             = _mm_mul_ps(r11,ewtabscale);
 769             ewitab           = _mm_cvttps_epi32(ewrt);
 770             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 771             ewitab           = _mm_slli_epi32(ewitab,2);
 772             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 773             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 774             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 775             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 776             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 777             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 778             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 779             velec            = _mm_mul_ps(qq11,_mm_sub_ps(rinv11,velec));
 780             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
 781
 782             /* Update potential sum for this i atom from the interaction with this j atom. */
 783             velec            = _mm_andnot_ps(dummy_mask,velec);
 784             velecsum         = _mm_add_ps(velecsum,velec);
 785
 786             fscal            = felec;
 787
 788             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 789
 790             /* Calculate temporary vectorial force */
 791             tx               = _mm_mul_ps(fscal,dx11);
 792             ty               = _mm_mul_ps(fscal,dy11);
 793             tz               = _mm_mul_ps(fscal,dz11);
 794
 795             /* Update vectorial force */
 796             fix1             = _mm_add_ps(fix1,tx);
 797             fiy1             = _mm_add_ps(fiy1,ty);
 798             fiz1             = _mm_add_ps(fiz1,tz);
 799
 800             fjx1             = _mm_add_ps(fjx1,tx);
 801             fjy1             = _mm_add_ps(fjy1,ty);
 802             fjz1             = _mm_add_ps(fjz1,tz);
 803
 804             /**************************
 805              * CALCULATE INTERACTIONS *
 806              **************************/
 807
 808             r12              = _mm_mul_ps(rsq12,rinv12);
 809             r12              = _mm_andnot_ps(dummy_mask,r12);
 810
 811             /* EWALD ELECTROSTATICS */
 812
 813             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 814             ewrt             = _mm_mul_ps(r12,ewtabscale);
 815             ewitab           = _mm_cvttps_epi32(ewrt);
 816             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 817             ewitab           = _mm_slli_epi32(ewitab,2);
 818             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 819             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 820             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 821             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 822             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 823             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 824             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 825             velec            = _mm_mul_ps(qq12,_mm_sub_ps(rinv12,velec));
 826             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
 827
 828             /* Update potential sum for this i atom from the interaction with this j atom. */
 829             velec            = _mm_andnot_ps(dummy_mask,velec);
 830             velecsum         = _mm_add_ps(velecsum,velec);
 831
 832             fscal            = felec;
 833
 834             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 835
 836             /* Calculate temporary vectorial force */
 837             tx               = _mm_mul_ps(fscal,dx12);
 838             ty               = _mm_mul_ps(fscal,dy12);
 839             tz               = _mm_mul_ps(fscal,dz12);
 840
 841             /* Update vectorial force */
 842             fix1             = _mm_add_ps(fix1,tx);
 843             fiy1             = _mm_add_ps(fiy1,ty);
 844             fiz1             = _mm_add_ps(fiz1,tz);
 845
 846             fjx2             = _mm_add_ps(fjx2,tx);
 847             fjy2             = _mm_add_ps(fjy2,ty);
 848             fjz2             = _mm_add_ps(fjz2,tz);
 849
 850             /**************************
 851              * CALCULATE INTERACTIONS *
 852              **************************/
 853
 854             r13              = _mm_mul_ps(rsq13,rinv13);
 855             r13              = _mm_andnot_ps(dummy_mask,r13);
 856
 857             /* EWALD ELECTROSTATICS */
 858
 859             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 860             ewrt             = _mm_mul_ps(r13,ewtabscale);
 861             ewitab           = _mm_cvttps_epi32(ewrt);
 862             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 863             ewitab           = _mm_slli_epi32(ewitab,2);
 864             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 865             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 866             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 867             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 868             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 869             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 870             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 871             velec            = _mm_mul_ps(qq13,_mm_sub_ps(rinv13,velec));
 872             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
 873
 874             /* Update potential sum for this i atom from the interaction with this j atom. */
 875             velec            = _mm_andnot_ps(dummy_mask,velec);
 876             velecsum         = _mm_add_ps(velecsum,velec);
 877
 878             fscal            = felec;
 879
 880             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 881
 882             /* Calculate temporary vectorial force */
 883             tx               = _mm_mul_ps(fscal,dx13);
 884             ty               = _mm_mul_ps(fscal,dy13);
 885             tz               = _mm_mul_ps(fscal,dz13);
 886
 887             /* Update vectorial force */
 888             fix1             = _mm_add_ps(fix1,tx);
 889             fiy1             = _mm_add_ps(fiy1,ty);
 890             fiz1             = _mm_add_ps(fiz1,tz);
 891
 892             fjx3             = _mm_add_ps(fjx3,tx);
 893             fjy3             = _mm_add_ps(fjy3,ty);
 894             fjz3             = _mm_add_ps(fjz3,tz);
 895
 896             /**************************
 897              * CALCULATE INTERACTIONS *
 898              **************************/
 899
 900             r21              = _mm_mul_ps(rsq21,rinv21);
 901             r21              = _mm_andnot_ps(dummy_mask,r21);
 902
 903             /* EWALD ELECTROSTATICS */
 904
 905             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 906             ewrt             = _mm_mul_ps(r21,ewtabscale);
 907             ewitab           = _mm_cvttps_epi32(ewrt);
 908             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 909             ewitab           = _mm_slli_epi32(ewitab,2);
 910             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 911             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 912             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 913             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 914             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 915             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 916             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 917             velec            = _mm_mul_ps(qq21,_mm_sub_ps(rinv21,velec));
 918             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
 919
 920             /* Update potential sum for this i atom from the interaction with this j atom. */
 921             velec            = _mm_andnot_ps(dummy_mask,velec);
 922             velecsum         = _mm_add_ps(velecsum,velec);
 923
 924             fscal            = felec;
 925
 926             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 927
 928             /* Calculate temporary vectorial force */
 929             tx               = _mm_mul_ps(fscal,dx21);
 930             ty               = _mm_mul_ps(fscal,dy21);
 931             tz               = _mm_mul_ps(fscal,dz21);
 932
 933             /* Update vectorial force */
 934             fix2             = _mm_add_ps(fix2,tx);
 935             fiy2             = _mm_add_ps(fiy2,ty);
 936             fiz2             = _mm_add_ps(fiz2,tz);
 937
 938             fjx1             = _mm_add_ps(fjx1,tx);
 939             fjy1             = _mm_add_ps(fjy1,ty);
 940             fjz1             = _mm_add_ps(fjz1,tz);
 941
 942             /**************************
 943              * CALCULATE INTERACTIONS *
 944              **************************/
 945
 946             r22              = _mm_mul_ps(rsq22,rinv22);
 947             r22              = _mm_andnot_ps(dummy_mask,r22);
 948
 949             /* EWALD ELECTROSTATICS */
 950
 951             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 952             ewrt             = _mm_mul_ps(r22,ewtabscale);
 953             ewitab           = _mm_cvttps_epi32(ewrt);
 954             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 955             ewitab           = _mm_slli_epi32(ewitab,2);
 956             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 957             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 958             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 959             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 960             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 961             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 962             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 963             velec            = _mm_mul_ps(qq22,_mm_sub_ps(rinv22,velec));
 964             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
 965
 966             /* Update potential sum for this i atom from the interaction with this j atom. */
 967             velec            = _mm_andnot_ps(dummy_mask,velec);
 968             velecsum         = _mm_add_ps(velecsum,velec);
 969
 970             fscal            = felec;
 971
 972             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 973
 974             /* Calculate temporary vectorial force */
 975             tx               = _mm_mul_ps(fscal,dx22);
 976             ty               = _mm_mul_ps(fscal,dy22);
 977             tz               = _mm_mul_ps(fscal,dz22);
 978
 979             /* Update vectorial force */
 980             fix2             = _mm_add_ps(fix2,tx);
 981             fiy2             = _mm_add_ps(fiy2,ty);
 982             fiz2             = _mm_add_ps(fiz2,tz);
 983
 984             fjx2             = _mm_add_ps(fjx2,tx);
 985             fjy2             = _mm_add_ps(fjy2,ty);
 986             fjz2             = _mm_add_ps(fjz2,tz);
 987
 988             /**************************
 989              * CALCULATE INTERACTIONS *
 990              **************************/
 991
 992             r23              = _mm_mul_ps(rsq23,rinv23);
 993             r23              = _mm_andnot_ps(dummy_mask,r23);
 994
 995             /* EWALD ELECTROSTATICS */
 996
 997             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 998             ewrt             = _mm_mul_ps(r23,ewtabscale);
 999             ewitab           = _mm_cvttps_epi32(ewrt);
1000             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1001             ewitab           = _mm_slli_epi32(ewitab,2);
1002             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1003             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1004             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1005             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1006             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1007             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1008             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1009             velec            = _mm_mul_ps(qq23,_mm_sub_ps(rinv23,velec));
1010             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
1011
1012             /* Update potential sum for this i atom from the interaction with this j atom. */
1013             velec            = _mm_andnot_ps(dummy_mask,velec);
1014             velecsum         = _mm_add_ps(velecsum,velec);
1015
1016             fscal            = felec;
1017
1018             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1019
1020             /* Calculate temporary vectorial force */
1021             tx               = _mm_mul_ps(fscal,dx23);
1022             ty               = _mm_mul_ps(fscal,dy23);
1023             tz               = _mm_mul_ps(fscal,dz23);
1024
1025             /* Update vectorial force */
1026             fix2             = _mm_add_ps(fix2,tx);
1027             fiy2             = _mm_add_ps(fiy2,ty);
1028             fiz2             = _mm_add_ps(fiz2,tz);
1029
1030             fjx3             = _mm_add_ps(fjx3,tx);
1031             fjy3             = _mm_add_ps(fjy3,ty);
1032             fjz3             = _mm_add_ps(fjz3,tz);
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             r31              = _mm_mul_ps(rsq31,rinv31);
1039             r31              = _mm_andnot_ps(dummy_mask,r31);
1040
1041             /* EWALD ELECTROSTATICS */
1042
1043             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1044             ewrt             = _mm_mul_ps(r31,ewtabscale);
1045             ewitab           = _mm_cvttps_epi32(ewrt);
1046             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1047             ewitab           = _mm_slli_epi32(ewitab,2);
1048             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1049             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1050             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1051             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1052             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1053             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1054             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1055             velec            = _mm_mul_ps(qq31,_mm_sub_ps(rinv31,velec));
1056             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
1057
1058             /* Update potential sum for this i atom from the interaction with this j atom. */
1059             velec            = _mm_andnot_ps(dummy_mask,velec);
1060             velecsum         = _mm_add_ps(velecsum,velec);
1061
1062             fscal            = felec;
1063
1064             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1065
1066             /* Calculate temporary vectorial force */
1067             tx               = _mm_mul_ps(fscal,dx31);
1068             ty               = _mm_mul_ps(fscal,dy31);
1069             tz               = _mm_mul_ps(fscal,dz31);
1070
1071             /* Update vectorial force */
1072             fix3             = _mm_add_ps(fix3,tx);
1073             fiy3             = _mm_add_ps(fiy3,ty);
1074             fiz3             = _mm_add_ps(fiz3,tz);
1075
1076             fjx1             = _mm_add_ps(fjx1,tx);
1077             fjy1             = _mm_add_ps(fjy1,ty);
1078             fjz1             = _mm_add_ps(fjz1,tz);
1079
1080             /**************************
1081              * CALCULATE INTERACTIONS *
1082              **************************/
1083
1084             r32              = _mm_mul_ps(rsq32,rinv32);
1085             r32              = _mm_andnot_ps(dummy_mask,r32);
1086
1087             /* EWALD ELECTROSTATICS */
1088
1089             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1090             ewrt             = _mm_mul_ps(r32,ewtabscale);
1091             ewitab           = _mm_cvttps_epi32(ewrt);
1092             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1093             ewitab           = _mm_slli_epi32(ewitab,2);
1094             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1095             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1096             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1097             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1098             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1099             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1100             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1101             velec            = _mm_mul_ps(qq32,_mm_sub_ps(rinv32,velec));
1102             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
1103
1104             /* Update potential sum for this i atom from the interaction with this j atom. */
1105             velec            = _mm_andnot_ps(dummy_mask,velec);
1106             velecsum         = _mm_add_ps(velecsum,velec);
1107
1108             fscal            = felec;
1109
1110             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1111
1112             /* Calculate temporary vectorial force */
1113             tx               = _mm_mul_ps(fscal,dx32);
1114             ty               = _mm_mul_ps(fscal,dy32);
1115             tz               = _mm_mul_ps(fscal,dz32);
1116
1117             /* Update vectorial force */
1118             fix3             = _mm_add_ps(fix3,tx);
1119             fiy3             = _mm_add_ps(fiy3,ty);
1120             fiz3             = _mm_add_ps(fiz3,tz);
1121
1122             fjx2             = _mm_add_ps(fjx2,tx);
1123             fjy2             = _mm_add_ps(fjy2,ty);
1124             fjz2             = _mm_add_ps(fjz2,tz);
1125
1126             /**************************
1127              * CALCULATE INTERACTIONS *
1128              **************************/
1129
1130             r33              = _mm_mul_ps(rsq33,rinv33);
1131             r33              = _mm_andnot_ps(dummy_mask,r33);
1132
1133             /* EWALD ELECTROSTATICS */
1134
1135             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1136             ewrt             = _mm_mul_ps(r33,ewtabscale);
1137             ewitab           = _mm_cvttps_epi32(ewrt);
1138             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1139             ewitab           = _mm_slli_epi32(ewitab,2);
1140             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1141             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1142             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1143             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1144             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1145             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1146             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1147             velec            = _mm_mul_ps(qq33,_mm_sub_ps(rinv33,velec));
1148             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
1149
1150             /* Update potential sum for this i atom from the interaction with this j atom. */
1151             velec            = _mm_andnot_ps(dummy_mask,velec);
1152             velecsum         = _mm_add_ps(velecsum,velec);
1153
1154             fscal            = felec;
1155
1156             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1157
1158             /* Calculate temporary vectorial force */
1159             tx               = _mm_mul_ps(fscal,dx33);
1160             ty               = _mm_mul_ps(fscal,dy33);
1161             tz               = _mm_mul_ps(fscal,dz33);
1162
1163             /* Update vectorial force */
1164             fix3             = _mm_add_ps(fix3,tx);
1165             fiy3             = _mm_add_ps(fiy3,ty);
1166             fiz3             = _mm_add_ps(fiz3,tz);
1167
1168             fjx3             = _mm_add_ps(fjx3,tx);
1169             fjy3             = _mm_add_ps(fjy3,ty);
1170             fjz3             = _mm_add_ps(fjz3,tz);
1171
1172             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1173             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1174             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1175             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1176
1177             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1178                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1179
1180             /* Inner loop uses 378 flops */
1181         }
1182
1183         /* End of innermost loop */
1184
1185         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1186                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1187
1188         ggid                        = gid[iidx];
1189         /* Update potential energies */
1190         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1191
1192         /* Increment number of inner iterations */
1193         inneriter                  += j_index_end - j_index_start;
1194
1195         /* Outer loop uses 19 flops */
1196     }
1197
1198     /* Increment number of outer iterations */
1199     outeriter        += nri;
1200
1201     /* Update outer/inner flops */
1202
1203     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*378);
1204 }
1205 /*
1206  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_F_sse2_single
1207  * Electrostatics interaction: Ewald
1208  * VdW interaction:            None
1209  * Geometry:                   Water4-Water4
1210  * Calculate force/pot:        Force
1211  */
1212 void
1213 nb_kernel_ElecEw_VdwNone_GeomW4W4_F_sse2_single
1214                     (t_nblist * gmx_restrict                nlist,
1215                      rvec * gmx_restrict                    xx,
1216                      rvec * gmx_restrict                    ff,
1217                      t_forcerec * gmx_restrict              fr,
1218                      t_mdatoms * gmx_restrict               mdatoms,
1219                      nb_kernel_data_t * gmx_restrict        kernel_data,
1220                      t_nrnb * gmx_restrict                  nrnb)
1221 {
1222     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1223      * just 0 for non-waters.
1224      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1225      * jnr indices corresponding to data put in the four positions in the SIMD register.
1226      */
1227     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1228     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1229     int              jnrA,jnrB,jnrC,jnrD;
1230     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1231     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1232     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1233     real             rcutoff_scalar;
1234     real             *shiftvec,*fshift,*x,*f;
1235     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1236     real             scratch[4*DIM];
1237     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1238     int              vdwioffset1;
1239     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1240     int              vdwioffset2;
1241     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1242     int              vdwioffset3;
1243     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1244     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1245     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1246     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1247     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1248     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1249     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1250     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1251     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1252     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1253     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1254     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1255     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1256     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1257     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1258     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1259     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1260     real             *charge;
1261     __m128i          ewitab;
1262     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1263     real             *ewtab;
1264     __m128           dummy_mask,cutoff_mask;
1265     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1266     __m128           one     = _mm_set1_ps(1.0);
1267     __m128           two     = _mm_set1_ps(2.0);
1268     x                = xx[0];
1269     f                = ff[0];
1270
1271     nri              = nlist->nri;
1272     iinr             = nlist->iinr;
1273     jindex           = nlist->jindex;
1274     jjnr             = nlist->jjnr;
1275     shiftidx         = nlist->shift;
1276     gid              = nlist->gid;
1277     shiftvec         = fr->shift_vec[0];
1278     fshift           = fr->fshift[0];
1279     facel            = _mm_set1_ps(fr->epsfac);
1280     charge           = mdatoms->chargeA;
1281
1282     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1283     ewtab            = fr->ic->tabq_coul_F;
1284     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1285     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1286
1287     /* Setup water-specific parameters */
1288     inr              = nlist->iinr[0];
1289     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1290     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1291     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1292
1293     jq1              = _mm_set1_ps(charge[inr+1]);
1294     jq2              = _mm_set1_ps(charge[inr+2]);
1295     jq3              = _mm_set1_ps(charge[inr+3]);
1296     qq11             = _mm_mul_ps(iq1,jq1);
1297     qq12             = _mm_mul_ps(iq1,jq2);
1298     qq13             = _mm_mul_ps(iq1,jq3);
1299     qq21             = _mm_mul_ps(iq2,jq1);
1300     qq22             = _mm_mul_ps(iq2,jq2);
1301     qq23             = _mm_mul_ps(iq2,jq3);
1302     qq31             = _mm_mul_ps(iq3,jq1);
1303     qq32             = _mm_mul_ps(iq3,jq2);
1304     qq33             = _mm_mul_ps(iq3,jq3);
1305
1306     /* Avoid stupid compiler warnings */
1307     jnrA = jnrB = jnrC = jnrD = 0;
1308     j_coord_offsetA = 0;
1309     j_coord_offsetB = 0;
1310     j_coord_offsetC = 0;
1311     j_coord_offsetD = 0;
1312
1313     outeriter        = 0;
1314     inneriter        = 0;
1315
1316     for(iidx=0;iidx<4*DIM;iidx++)
1317     {
1318         scratch[iidx] = 0.0;
1319     }
1320
1321     /* Start outer loop over neighborlists */
1322     for(iidx=0; iidx<nri; iidx++)
1323     {
1324         /* Load shift vector for this list */
1325         i_shift_offset   = DIM*shiftidx[iidx];
1326
1327         /* Load limits for loop over neighbors */
1328         j_index_start    = jindex[iidx];
1329         j_index_end      = jindex[iidx+1];
1330
1331         /* Get outer coordinate index */
1332         inr              = iinr[iidx];
1333         i_coord_offset   = DIM*inr;
1334
1335         /* Load i particle coords and add shift vector */
1336         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1337                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1338
1339         fix1             = _mm_setzero_ps();
1340         fiy1             = _mm_setzero_ps();
1341         fiz1             = _mm_setzero_ps();
1342         fix2             = _mm_setzero_ps();
1343         fiy2             = _mm_setzero_ps();
1344         fiz2             = _mm_setzero_ps();
1345         fix3             = _mm_setzero_ps();
1346         fiy3             = _mm_setzero_ps();
1347         fiz3             = _mm_setzero_ps();
1348
1349         /* Start inner kernel loop */
1350         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1351         {
1352
1353             /* Get j neighbor index, and coordinate index */
1354             jnrA             = jjnr[jidx];
1355             jnrB             = jjnr[jidx+1];
1356             jnrC             = jjnr[jidx+2];
1357             jnrD             = jjnr[jidx+3];
1358             j_coord_offsetA  = DIM*jnrA;
1359             j_coord_offsetB  = DIM*jnrB;
1360             j_coord_offsetC  = DIM*jnrC;
1361             j_coord_offsetD  = DIM*jnrD;
1362
1363             /* load j atom coordinates */
1364             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1365                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1366                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1367
1368             /* Calculate displacement vector */
1369             dx11             = _mm_sub_ps(ix1,jx1);
1370             dy11             = _mm_sub_ps(iy1,jy1);
1371             dz11             = _mm_sub_ps(iz1,jz1);
1372             dx12             = _mm_sub_ps(ix1,jx2);
1373             dy12             = _mm_sub_ps(iy1,jy2);
1374             dz12             = _mm_sub_ps(iz1,jz2);
1375             dx13             = _mm_sub_ps(ix1,jx3);
1376             dy13             = _mm_sub_ps(iy1,jy3);
1377             dz13             = _mm_sub_ps(iz1,jz3);
1378             dx21             = _mm_sub_ps(ix2,jx1);
1379             dy21             = _mm_sub_ps(iy2,jy1);
1380             dz21             = _mm_sub_ps(iz2,jz1);
1381             dx22             = _mm_sub_ps(ix2,jx2);
1382             dy22             = _mm_sub_ps(iy2,jy2);
1383             dz22             = _mm_sub_ps(iz2,jz2);
1384             dx23             = _mm_sub_ps(ix2,jx3);
1385             dy23             = _mm_sub_ps(iy2,jy3);
1386             dz23             = _mm_sub_ps(iz2,jz3);
1387             dx31             = _mm_sub_ps(ix3,jx1);
1388             dy31             = _mm_sub_ps(iy3,jy1);
1389             dz31             = _mm_sub_ps(iz3,jz1);
1390             dx32             = _mm_sub_ps(ix3,jx2);
1391             dy32             = _mm_sub_ps(iy3,jy2);
1392             dz32             = _mm_sub_ps(iz3,jz2);
1393             dx33             = _mm_sub_ps(ix3,jx3);
1394             dy33             = _mm_sub_ps(iy3,jy3);
1395             dz33             = _mm_sub_ps(iz3,jz3);
1396
1397             /* Calculate squared distance and things based on it */
1398             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1399             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1400             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1401             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1402             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1403             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1404             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1405             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1406             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1407
1408             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1409             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1410             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1411             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1412             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1413             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1414             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1415             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1416             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1417
1418             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1419             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1420             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1421             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1422             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1423             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1424             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1425             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1426             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1427
1428             fjx1             = _mm_setzero_ps();
1429             fjy1             = _mm_setzero_ps();
1430             fjz1             = _mm_setzero_ps();
1431             fjx2             = _mm_setzero_ps();
1432             fjy2             = _mm_setzero_ps();
1433             fjz2             = _mm_setzero_ps();
1434             fjx3             = _mm_setzero_ps();
1435             fjy3             = _mm_setzero_ps();
1436             fjz3             = _mm_setzero_ps();
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             r11              = _mm_mul_ps(rsq11,rinv11);
1443
1444             /* EWALD ELECTROSTATICS */
1445
1446             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1447             ewrt             = _mm_mul_ps(r11,ewtabscale);
1448             ewitab           = _mm_cvttps_epi32(ewrt);
1449             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1450             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1451                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1452                                          &ewtabF,&ewtabFn);
1453             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1454             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
1455
1456             fscal            = felec;
1457
1458             /* Calculate temporary vectorial force */
1459             tx               = _mm_mul_ps(fscal,dx11);
1460             ty               = _mm_mul_ps(fscal,dy11);
1461             tz               = _mm_mul_ps(fscal,dz11);
1462
1463             /* Update vectorial force */
1464             fix1             = _mm_add_ps(fix1,tx);
1465             fiy1             = _mm_add_ps(fiy1,ty);
1466             fiz1             = _mm_add_ps(fiz1,tz);
1467
1468             fjx1             = _mm_add_ps(fjx1,tx);
1469             fjy1             = _mm_add_ps(fjy1,ty);
1470             fjz1             = _mm_add_ps(fjz1,tz);
1471
1472             /**************************
1473              * CALCULATE INTERACTIONS *
1474              **************************/
1475
1476             r12              = _mm_mul_ps(rsq12,rinv12);
1477
1478             /* EWALD ELECTROSTATICS */
1479
1480             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1481             ewrt             = _mm_mul_ps(r12,ewtabscale);
1482             ewitab           = _mm_cvttps_epi32(ewrt);
1483             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1484             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1485                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1486                                          &ewtabF,&ewtabFn);
1487             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1488             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
1489
1490             fscal            = felec;
1491
1492             /* Calculate temporary vectorial force */
1493             tx               = _mm_mul_ps(fscal,dx12);
1494             ty               = _mm_mul_ps(fscal,dy12);
1495             tz               = _mm_mul_ps(fscal,dz12);
1496
1497             /* Update vectorial force */
1498             fix1             = _mm_add_ps(fix1,tx);
1499             fiy1             = _mm_add_ps(fiy1,ty);
1500             fiz1             = _mm_add_ps(fiz1,tz);
1501
1502             fjx2             = _mm_add_ps(fjx2,tx);
1503             fjy2             = _mm_add_ps(fjy2,ty);
1504             fjz2             = _mm_add_ps(fjz2,tz);
1505
1506             /**************************
1507              * CALCULATE INTERACTIONS *
1508              **************************/
1509
1510             r13              = _mm_mul_ps(rsq13,rinv13);
1511
1512             /* EWALD ELECTROSTATICS */
1513
1514             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1515             ewrt             = _mm_mul_ps(r13,ewtabscale);
1516             ewitab           = _mm_cvttps_epi32(ewrt);
1517             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1518             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1519                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1520                                          &ewtabF,&ewtabFn);
1521             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1522             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
1523
1524             fscal            = felec;
1525
1526             /* Calculate temporary vectorial force */
1527             tx               = _mm_mul_ps(fscal,dx13);
1528             ty               = _mm_mul_ps(fscal,dy13);
1529             tz               = _mm_mul_ps(fscal,dz13);
1530
1531             /* Update vectorial force */
1532             fix1             = _mm_add_ps(fix1,tx);
1533             fiy1             = _mm_add_ps(fiy1,ty);
1534             fiz1             = _mm_add_ps(fiz1,tz);
1535
1536             fjx3             = _mm_add_ps(fjx3,tx);
1537             fjy3             = _mm_add_ps(fjy3,ty);
1538             fjz3             = _mm_add_ps(fjz3,tz);
1539
1540             /**************************
1541              * CALCULATE INTERACTIONS *
1542              **************************/
1543
1544             r21              = _mm_mul_ps(rsq21,rinv21);
1545
1546             /* EWALD ELECTROSTATICS */
1547
1548             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1549             ewrt             = _mm_mul_ps(r21,ewtabscale);
1550             ewitab           = _mm_cvttps_epi32(ewrt);
1551             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1552             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1553                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1554                                          &ewtabF,&ewtabFn);
1555             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1556             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1557
1558             fscal            = felec;
1559
1560             /* Calculate temporary vectorial force */
1561             tx               = _mm_mul_ps(fscal,dx21);
1562             ty               = _mm_mul_ps(fscal,dy21);
1563             tz               = _mm_mul_ps(fscal,dz21);
1564
1565             /* Update vectorial force */
1566             fix2             = _mm_add_ps(fix2,tx);
1567             fiy2             = _mm_add_ps(fiy2,ty);
1568             fiz2             = _mm_add_ps(fiz2,tz);
1569
1570             fjx1             = _mm_add_ps(fjx1,tx);
1571             fjy1             = _mm_add_ps(fjy1,ty);
1572             fjz1             = _mm_add_ps(fjz1,tz);
1573
1574             /**************************
1575              * CALCULATE INTERACTIONS *
1576              **************************/
1577
1578             r22              = _mm_mul_ps(rsq22,rinv22);
1579
1580             /* EWALD ELECTROSTATICS */
1581
1582             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1583             ewrt             = _mm_mul_ps(r22,ewtabscale);
1584             ewitab           = _mm_cvttps_epi32(ewrt);
1585             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1586             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1587                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1588                                          &ewtabF,&ewtabFn);
1589             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1590             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
1591
1592             fscal            = felec;
1593
1594             /* Calculate temporary vectorial force */
1595             tx               = _mm_mul_ps(fscal,dx22);
1596             ty               = _mm_mul_ps(fscal,dy22);
1597             tz               = _mm_mul_ps(fscal,dz22);
1598
1599             /* Update vectorial force */
1600             fix2             = _mm_add_ps(fix2,tx);
1601             fiy2             = _mm_add_ps(fiy2,ty);
1602             fiz2             = _mm_add_ps(fiz2,tz);
1603
1604             fjx2             = _mm_add_ps(fjx2,tx);
1605             fjy2             = _mm_add_ps(fjy2,ty);
1606             fjz2             = _mm_add_ps(fjz2,tz);
1607
1608             /**************************
1609              * CALCULATE INTERACTIONS *
1610              **************************/
1611
1612             r23              = _mm_mul_ps(rsq23,rinv23);
1613
1614             /* EWALD ELECTROSTATICS */
1615
1616             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1617             ewrt             = _mm_mul_ps(r23,ewtabscale);
1618             ewitab           = _mm_cvttps_epi32(ewrt);
1619             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1620             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1621                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1622                                          &ewtabF,&ewtabFn);
1623             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1624             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
1625
1626             fscal            = felec;
1627
1628             /* Calculate temporary vectorial force */
1629             tx               = _mm_mul_ps(fscal,dx23);
1630             ty               = _mm_mul_ps(fscal,dy23);
1631             tz               = _mm_mul_ps(fscal,dz23);
1632
1633             /* Update vectorial force */
1634             fix2             = _mm_add_ps(fix2,tx);
1635             fiy2             = _mm_add_ps(fiy2,ty);
1636             fiz2             = _mm_add_ps(fiz2,tz);
1637
1638             fjx3             = _mm_add_ps(fjx3,tx);
1639             fjy3             = _mm_add_ps(fjy3,ty);
1640             fjz3             = _mm_add_ps(fjz3,tz);
1641
1642             /**************************
1643              * CALCULATE INTERACTIONS *
1644              **************************/
1645
1646             r31              = _mm_mul_ps(rsq31,rinv31);
1647
1648             /* EWALD ELECTROSTATICS */
1649
1650             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1651             ewrt             = _mm_mul_ps(r31,ewtabscale);
1652             ewitab           = _mm_cvttps_epi32(ewrt);
1653             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1654             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1655                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1656                                          &ewtabF,&ewtabFn);
1657             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1658             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
1659
1660             fscal            = felec;
1661
1662             /* Calculate temporary vectorial force */
1663             tx               = _mm_mul_ps(fscal,dx31);
1664             ty               = _mm_mul_ps(fscal,dy31);
1665             tz               = _mm_mul_ps(fscal,dz31);
1666
1667             /* Update vectorial force */
1668             fix3             = _mm_add_ps(fix3,tx);
1669             fiy3             = _mm_add_ps(fiy3,ty);
1670             fiz3             = _mm_add_ps(fiz3,tz);
1671
1672             fjx1             = _mm_add_ps(fjx1,tx);
1673             fjy1             = _mm_add_ps(fjy1,ty);
1674             fjz1             = _mm_add_ps(fjz1,tz);
1675
1676             /**************************
1677              * CALCULATE INTERACTIONS *
1678              **************************/
1679
1680             r32              = _mm_mul_ps(rsq32,rinv32);
1681
1682             /* EWALD ELECTROSTATICS */
1683
1684             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1685             ewrt             = _mm_mul_ps(r32,ewtabscale);
1686             ewitab           = _mm_cvttps_epi32(ewrt);
1687             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1688             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1689                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1690                                          &ewtabF,&ewtabFn);
1691             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1692             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
1693
1694             fscal            = felec;
1695
1696             /* Calculate temporary vectorial force */
1697             tx               = _mm_mul_ps(fscal,dx32);
1698             ty               = _mm_mul_ps(fscal,dy32);
1699             tz               = _mm_mul_ps(fscal,dz32);
1700
1701             /* Update vectorial force */
1702             fix3             = _mm_add_ps(fix3,tx);
1703             fiy3             = _mm_add_ps(fiy3,ty);
1704             fiz3             = _mm_add_ps(fiz3,tz);
1705
1706             fjx2             = _mm_add_ps(fjx2,tx);
1707             fjy2             = _mm_add_ps(fjy2,ty);
1708             fjz2             = _mm_add_ps(fjz2,tz);
1709
1710             /**************************
1711              * CALCULATE INTERACTIONS *
1712              **************************/
1713
1714             r33              = _mm_mul_ps(rsq33,rinv33);
1715
1716             /* EWALD ELECTROSTATICS */
1717
1718             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1719             ewrt             = _mm_mul_ps(r33,ewtabscale);
1720             ewitab           = _mm_cvttps_epi32(ewrt);
1721             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1722             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1723                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1724                                          &ewtabF,&ewtabFn);
1725             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1726             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
1727
1728             fscal            = felec;
1729
1730             /* Calculate temporary vectorial force */
1731             tx               = _mm_mul_ps(fscal,dx33);
1732             ty               = _mm_mul_ps(fscal,dy33);
1733             tz               = _mm_mul_ps(fscal,dz33);
1734
1735             /* Update vectorial force */
1736             fix3             = _mm_add_ps(fix3,tx);
1737             fiy3             = _mm_add_ps(fiy3,ty);
1738             fiz3             = _mm_add_ps(fiz3,tz);
1739
1740             fjx3             = _mm_add_ps(fjx3,tx);
1741             fjy3             = _mm_add_ps(fjy3,ty);
1742             fjz3             = _mm_add_ps(fjz3,tz);
1743
1744             fjptrA             = f+j_coord_offsetA;
1745             fjptrB             = f+j_coord_offsetB;
1746             fjptrC             = f+j_coord_offsetC;
1747             fjptrD             = f+j_coord_offsetD;
1748
1749             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1750                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1751
1752             /* Inner loop uses 324 flops */
1753         }
1754
1755         if(jidx<j_index_end)
1756         {
1757
1758             /* Get j neighbor index, and coordinate index */
1759             jnrlistA         = jjnr[jidx];
1760             jnrlistB         = jjnr[jidx+1];
1761             jnrlistC         = jjnr[jidx+2];
1762             jnrlistD         = jjnr[jidx+3];
1763             /* Sign of each element will be negative for non-real atoms.
1764              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1765              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1766              */
1767             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1768             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1769             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1770             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1771             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1772             j_coord_offsetA  = DIM*jnrA;
1773             j_coord_offsetB  = DIM*jnrB;
1774             j_coord_offsetC  = DIM*jnrC;
1775             j_coord_offsetD  = DIM*jnrD;
1776
1777             /* load j atom coordinates */
1778             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1779                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1780                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1781
1782             /* Calculate displacement vector */
1783             dx11             = _mm_sub_ps(ix1,jx1);
1784             dy11             = _mm_sub_ps(iy1,jy1);
1785             dz11             = _mm_sub_ps(iz1,jz1);
1786             dx12             = _mm_sub_ps(ix1,jx2);
1787             dy12             = _mm_sub_ps(iy1,jy2);
1788             dz12             = _mm_sub_ps(iz1,jz2);
1789             dx13             = _mm_sub_ps(ix1,jx3);
1790             dy13             = _mm_sub_ps(iy1,jy3);
1791             dz13             = _mm_sub_ps(iz1,jz3);
1792             dx21             = _mm_sub_ps(ix2,jx1);
1793             dy21             = _mm_sub_ps(iy2,jy1);
1794             dz21             = _mm_sub_ps(iz2,jz1);
1795             dx22             = _mm_sub_ps(ix2,jx2);
1796             dy22             = _mm_sub_ps(iy2,jy2);
1797             dz22             = _mm_sub_ps(iz2,jz2);
1798             dx23             = _mm_sub_ps(ix2,jx3);
1799             dy23             = _mm_sub_ps(iy2,jy3);
1800             dz23             = _mm_sub_ps(iz2,jz3);
1801             dx31             = _mm_sub_ps(ix3,jx1);
1802             dy31             = _mm_sub_ps(iy3,jy1);
1803             dz31             = _mm_sub_ps(iz3,jz1);
1804             dx32             = _mm_sub_ps(ix3,jx2);
1805             dy32             = _mm_sub_ps(iy3,jy2);
1806             dz32             = _mm_sub_ps(iz3,jz2);
1807             dx33             = _mm_sub_ps(ix3,jx3);
1808             dy33             = _mm_sub_ps(iy3,jy3);
1809             dz33             = _mm_sub_ps(iz3,jz3);
1810
1811             /* Calculate squared distance and things based on it */
1812             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1813             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1814             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1815             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1816             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1817             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1818             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1819             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1820             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1821
1822             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1823             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1824             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1825             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1826             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1827             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1828             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1829             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1830             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1831
1832             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1833             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1834             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1835             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1836             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1837             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1838             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1839             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1840             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1841
1842             fjx1             = _mm_setzero_ps();
1843             fjy1             = _mm_setzero_ps();
1844             fjz1             = _mm_setzero_ps();
1845             fjx2             = _mm_setzero_ps();
1846             fjy2             = _mm_setzero_ps();
1847             fjz2             = _mm_setzero_ps();
1848             fjx3             = _mm_setzero_ps();
1849             fjy3             = _mm_setzero_ps();
1850             fjz3             = _mm_setzero_ps();
1851
1852             /**************************
1853              * CALCULATE INTERACTIONS *
1854              **************************/
1855
1856             r11              = _mm_mul_ps(rsq11,rinv11);
1857             r11              = _mm_andnot_ps(dummy_mask,r11);
1858
1859             /* EWALD ELECTROSTATICS */
1860
1861             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1862             ewrt             = _mm_mul_ps(r11,ewtabscale);
1863             ewitab           = _mm_cvttps_epi32(ewrt);
1864             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1865             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1866                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1867                                          &ewtabF,&ewtabFn);
1868             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1869             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
1870
1871             fscal            = felec;
1872
1873             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1874
1875             /* Calculate temporary vectorial force */
1876             tx               = _mm_mul_ps(fscal,dx11);
1877             ty               = _mm_mul_ps(fscal,dy11);
1878             tz               = _mm_mul_ps(fscal,dz11);
1879
1880             /* Update vectorial force */
1881             fix1             = _mm_add_ps(fix1,tx);
1882             fiy1             = _mm_add_ps(fiy1,ty);
1883             fiz1             = _mm_add_ps(fiz1,tz);
1884
1885             fjx1             = _mm_add_ps(fjx1,tx);
1886             fjy1             = _mm_add_ps(fjy1,ty);
1887             fjz1             = _mm_add_ps(fjz1,tz);
1888
1889             /**************************
1890              * CALCULATE INTERACTIONS *
1891              **************************/
1892
1893             r12              = _mm_mul_ps(rsq12,rinv12);
1894             r12              = _mm_andnot_ps(dummy_mask,r12);
1895
1896             /* EWALD ELECTROSTATICS */
1897
1898             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1899             ewrt             = _mm_mul_ps(r12,ewtabscale);
1900             ewitab           = _mm_cvttps_epi32(ewrt);
1901             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1902             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1903                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1904                                          &ewtabF,&ewtabFn);
1905             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1906             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
1907
1908             fscal            = felec;
1909
1910             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1911
1912             /* Calculate temporary vectorial force */
1913             tx               = _mm_mul_ps(fscal,dx12);
1914             ty               = _mm_mul_ps(fscal,dy12);
1915             tz               = _mm_mul_ps(fscal,dz12);
1916
1917             /* Update vectorial force */
1918             fix1             = _mm_add_ps(fix1,tx);
1919             fiy1             = _mm_add_ps(fiy1,ty);
1920             fiz1             = _mm_add_ps(fiz1,tz);
1921
1922             fjx2             = _mm_add_ps(fjx2,tx);
1923             fjy2             = _mm_add_ps(fjy2,ty);
1924             fjz2             = _mm_add_ps(fjz2,tz);
1925
1926             /**************************
1927              * CALCULATE INTERACTIONS *
1928              **************************/
1929
1930             r13              = _mm_mul_ps(rsq13,rinv13);
1931             r13              = _mm_andnot_ps(dummy_mask,r13);
1932
1933             /* EWALD ELECTROSTATICS */
1934
1935             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1936             ewrt             = _mm_mul_ps(r13,ewtabscale);
1937             ewitab           = _mm_cvttps_epi32(ewrt);
1938             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1939             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1940                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1941                                          &ewtabF,&ewtabFn);
1942             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1943             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
1944
1945             fscal            = felec;
1946
1947             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1948
1949             /* Calculate temporary vectorial force */
1950             tx               = _mm_mul_ps(fscal,dx13);
1951             ty               = _mm_mul_ps(fscal,dy13);
1952             tz               = _mm_mul_ps(fscal,dz13);
1953
1954             /* Update vectorial force */
1955             fix1             = _mm_add_ps(fix1,tx);
1956             fiy1             = _mm_add_ps(fiy1,ty);
1957             fiz1             = _mm_add_ps(fiz1,tz);
1958
1959             fjx3             = _mm_add_ps(fjx3,tx);
1960             fjy3             = _mm_add_ps(fjy3,ty);
1961             fjz3             = _mm_add_ps(fjz3,tz);
1962
1963             /**************************
1964              * CALCULATE INTERACTIONS *
1965              **************************/
1966
1967             r21              = _mm_mul_ps(rsq21,rinv21);
1968             r21              = _mm_andnot_ps(dummy_mask,r21);
1969
1970             /* EWALD ELECTROSTATICS */
1971
1972             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1973             ewrt             = _mm_mul_ps(r21,ewtabscale);
1974             ewitab           = _mm_cvttps_epi32(ewrt);
1975             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1976             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1977                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1978                                          &ewtabF,&ewtabFn);
1979             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1980             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1981
1982             fscal            = felec;
1983
1984             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1985
1986             /* Calculate temporary vectorial force */
1987             tx               = _mm_mul_ps(fscal,dx21);
1988             ty               = _mm_mul_ps(fscal,dy21);
1989             tz               = _mm_mul_ps(fscal,dz21);
1990
1991             /* Update vectorial force */
1992             fix2             = _mm_add_ps(fix2,tx);
1993             fiy2             = _mm_add_ps(fiy2,ty);
1994             fiz2             = _mm_add_ps(fiz2,tz);
1995
1996             fjx1             = _mm_add_ps(fjx1,tx);
1997             fjy1             = _mm_add_ps(fjy1,ty);
1998             fjz1             = _mm_add_ps(fjz1,tz);
1999
2000             /**************************
2001              * CALCULATE INTERACTIONS *
2002              **************************/
2003
2004             r22              = _mm_mul_ps(rsq22,rinv22);
2005             r22              = _mm_andnot_ps(dummy_mask,r22);
2006
2007             /* EWALD ELECTROSTATICS */
2008
2009             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2010             ewrt             = _mm_mul_ps(r22,ewtabscale);
2011             ewitab           = _mm_cvttps_epi32(ewrt);
2012             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2013             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2014                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2015                                          &ewtabF,&ewtabFn);
2016             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2017             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
2018
2019             fscal            = felec;
2020
2021             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2022
2023             /* Calculate temporary vectorial force */
2024             tx               = _mm_mul_ps(fscal,dx22);
2025             ty               = _mm_mul_ps(fscal,dy22);
2026             tz               = _mm_mul_ps(fscal,dz22);
2027
2028             /* Update vectorial force */
2029             fix2             = _mm_add_ps(fix2,tx);
2030             fiy2             = _mm_add_ps(fiy2,ty);
2031             fiz2             = _mm_add_ps(fiz2,tz);
2032
2033             fjx2             = _mm_add_ps(fjx2,tx);
2034             fjy2             = _mm_add_ps(fjy2,ty);
2035             fjz2             = _mm_add_ps(fjz2,tz);
2036
2037             /**************************
2038              * CALCULATE INTERACTIONS *
2039              **************************/
2040
2041             r23              = _mm_mul_ps(rsq23,rinv23);
2042             r23              = _mm_andnot_ps(dummy_mask,r23);
2043
2044             /* EWALD ELECTROSTATICS */
2045
2046             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2047             ewrt             = _mm_mul_ps(r23,ewtabscale);
2048             ewitab           = _mm_cvttps_epi32(ewrt);
2049             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2050             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2051                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2052                                          &ewtabF,&ewtabFn);
2053             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2054             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
2055
2056             fscal            = felec;
2057
2058             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2059
2060             /* Calculate temporary vectorial force */
2061             tx               = _mm_mul_ps(fscal,dx23);
2062             ty               = _mm_mul_ps(fscal,dy23);
2063             tz               = _mm_mul_ps(fscal,dz23);
2064
2065             /* Update vectorial force */
2066             fix2             = _mm_add_ps(fix2,tx);
2067             fiy2             = _mm_add_ps(fiy2,ty);
2068             fiz2             = _mm_add_ps(fiz2,tz);
2069
2070             fjx3             = _mm_add_ps(fjx3,tx);
2071             fjy3             = _mm_add_ps(fjy3,ty);
2072             fjz3             = _mm_add_ps(fjz3,tz);
2073
2074             /**************************
2075              * CALCULATE INTERACTIONS *
2076              **************************/
2077
2078             r31              = _mm_mul_ps(rsq31,rinv31);
2079             r31              = _mm_andnot_ps(dummy_mask,r31);
2080
2081             /* EWALD ELECTROSTATICS */
2082
2083             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2084             ewrt             = _mm_mul_ps(r31,ewtabscale);
2085             ewitab           = _mm_cvttps_epi32(ewrt);
2086             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2087             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2088                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2089                                          &ewtabF,&ewtabFn);
2090             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2091             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
2092
2093             fscal            = felec;
2094
2095             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2096
2097             /* Calculate temporary vectorial force */
2098             tx               = _mm_mul_ps(fscal,dx31);
2099             ty               = _mm_mul_ps(fscal,dy31);
2100             tz               = _mm_mul_ps(fscal,dz31);
2101
2102             /* Update vectorial force */
2103             fix3             = _mm_add_ps(fix3,tx);
2104             fiy3             = _mm_add_ps(fiy3,ty);
2105             fiz3             = _mm_add_ps(fiz3,tz);
2106
2107             fjx1             = _mm_add_ps(fjx1,tx);
2108             fjy1             = _mm_add_ps(fjy1,ty);
2109             fjz1             = _mm_add_ps(fjz1,tz);
2110
2111             /**************************
2112              * CALCULATE INTERACTIONS *
2113              **************************/
2114
2115             r32              = _mm_mul_ps(rsq32,rinv32);
2116             r32              = _mm_andnot_ps(dummy_mask,r32);
2117
2118             /* EWALD ELECTROSTATICS */
2119
2120             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2121             ewrt             = _mm_mul_ps(r32,ewtabscale);
2122             ewitab           = _mm_cvttps_epi32(ewrt);
2123             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2124             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2125                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2126                                          &ewtabF,&ewtabFn);
2127             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2128             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
2129
2130             fscal            = felec;
2131
2132             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2133
2134             /* Calculate temporary vectorial force */
2135             tx               = _mm_mul_ps(fscal,dx32);
2136             ty               = _mm_mul_ps(fscal,dy32);
2137             tz               = _mm_mul_ps(fscal,dz32);
2138
2139             /* Update vectorial force */
2140             fix3             = _mm_add_ps(fix3,tx);
2141             fiy3             = _mm_add_ps(fiy3,ty);
2142             fiz3             = _mm_add_ps(fiz3,tz);
2143
2144             fjx2             = _mm_add_ps(fjx2,tx);
2145             fjy2             = _mm_add_ps(fjy2,ty);
2146             fjz2             = _mm_add_ps(fjz2,tz);
2147
2148             /**************************
2149              * CALCULATE INTERACTIONS *
2150              **************************/
2151
2152             r33              = _mm_mul_ps(rsq33,rinv33);
2153             r33              = _mm_andnot_ps(dummy_mask,r33);
2154
2155             /* EWALD ELECTROSTATICS */
2156
2157             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2158             ewrt             = _mm_mul_ps(r33,ewtabscale);
2159             ewitab           = _mm_cvttps_epi32(ewrt);
2160             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2161             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2162                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2163                                          &ewtabF,&ewtabFn);
2164             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2165             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
2166
2167             fscal            = felec;
2168
2169             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2170
2171             /* Calculate temporary vectorial force */
2172             tx               = _mm_mul_ps(fscal,dx33);
2173             ty               = _mm_mul_ps(fscal,dy33);
2174             tz               = _mm_mul_ps(fscal,dz33);
2175
2176             /* Update vectorial force */
2177             fix3             = _mm_add_ps(fix3,tx);
2178             fiy3             = _mm_add_ps(fiy3,ty);
2179             fiz3             = _mm_add_ps(fiz3,tz);
2180
2181             fjx3             = _mm_add_ps(fjx3,tx);
2182             fjy3             = _mm_add_ps(fjy3,ty);
2183             fjz3             = _mm_add_ps(fjz3,tz);
2184
2185             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2186             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2187             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2188             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2189
2190             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
2191                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2192
2193             /* Inner loop uses 333 flops */
2194         }
2195
2196         /* End of innermost loop */
2197
2198         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2199                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2200
2201         /* Increment number of inner iterations */
2202         inneriter                  += j_index_end - j_index_start;
2203
2204         /* Outer loop uses 18 flops */
2205     }
2206
2207     /* Increment number of outer iterations */
2208     outeriter        += nri;
2209
2210     /* Update outer/inner flops */
2211
2212     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*333);
2213 }