src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecEw_VdwLJ_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_VdwLJ_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_VdwLJ_GeomW4W4_VF_sse2_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             shX,shY,shZ,rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwioffset3;
  73     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  75     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  77     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  79     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  81     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  99     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 100     __m128i          ewitab;
 101     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 102     real             *ewtab;
 103     __m128           dummy_mask,cutoff_mask;
 104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 105     __m128           one     = _mm_set1_ps(1.0);
 106     __m128           two     = _mm_set1_ps(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_ps(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 125     ewtab            = fr->ic->tabq_coul_FDV0;
 126     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 127     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 132     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 133     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 134     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 135
 136     jq1              = _mm_set1_ps(charge[inr+1]);
 137     jq2              = _mm_set1_ps(charge[inr+2]);
 138     jq3              = _mm_set1_ps(charge[inr+3]);
 139     vdwjidx0A        = 2*vdwtype[inr+0];
 140     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 141     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 142     qq11             = _mm_mul_ps(iq1,jq1);
 143     qq12             = _mm_mul_ps(iq1,jq2);
 144     qq13             = _mm_mul_ps(iq1,jq3);
 145     qq21             = _mm_mul_ps(iq2,jq1);
 146     qq22             = _mm_mul_ps(iq2,jq2);
 147     qq23             = _mm_mul_ps(iq2,jq3);
 148     qq31             = _mm_mul_ps(iq3,jq1);
 149     qq32             = _mm_mul_ps(iq3,jq2);
 150     qq33             = _mm_mul_ps(iq3,jq3);
 151
 152     /* Avoid stupid compiler warnings */
 153     jnrA = jnrB = jnrC = jnrD = 0;
 154     j_coord_offsetA = 0;
 155     j_coord_offsetB = 0;
 156     j_coord_offsetC = 0;
 157     j_coord_offsetD = 0;
 158
 159     outeriter        = 0;
 160     inneriter        = 0;
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167         shX              = shiftvec[i_shift_offset+XX];
 168         shY              = shiftvec[i_shift_offset+YY];
 169         shZ              = shiftvec[i_shift_offset+ZZ];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 181         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 182         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 183         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 184         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 185         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 186         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 187         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 188         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 189         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 190         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 191         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 192
 193         fix0             = _mm_setzero_ps();
 194         fiy0             = _mm_setzero_ps();
 195         fiz0             = _mm_setzero_ps();
 196         fix1             = _mm_setzero_ps();
 197         fiy1             = _mm_setzero_ps();
 198         fiz1             = _mm_setzero_ps();
 199         fix2             = _mm_setzero_ps();
 200         fiy2             = _mm_setzero_ps();
 201         fiz2             = _mm_setzero_ps();
 202         fix3             = _mm_setzero_ps();
 203         fiy3             = _mm_setzero_ps();
 204         fiz3             = _mm_setzero_ps();
 205
 206         /* Reset potential sums */
 207         velecsum         = _mm_setzero_ps();
 208         vvdwsum          = _mm_setzero_ps();
 209
 210         /* Start inner kernel loop */
 211         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 212         {
 213
 214             /* Get j neighbor index, and coordinate index */
 215             jnrA             = jjnr[jidx];
 216             jnrB             = jjnr[jidx+1];
 217             jnrC             = jjnr[jidx+2];
 218             jnrD             = jjnr[jidx+3];
 219
 220             j_coord_offsetA  = DIM*jnrA;
 221             j_coord_offsetB  = DIM*jnrB;
 222             j_coord_offsetC  = DIM*jnrC;
 223             j_coord_offsetD  = DIM*jnrD;
 224
 225             /* load j atom coordinates */
 226             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 227                                               x+j_coord_offsetC,x+j_coord_offsetD,
 228                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 229                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 230
 231             /* Calculate displacement vector */
 232             dx00             = _mm_sub_ps(ix0,jx0);
 233             dy00             = _mm_sub_ps(iy0,jy0);
 234             dz00             = _mm_sub_ps(iz0,jz0);
 235             dx11             = _mm_sub_ps(ix1,jx1);
 236             dy11             = _mm_sub_ps(iy1,jy1);
 237             dz11             = _mm_sub_ps(iz1,jz1);
 238             dx12             = _mm_sub_ps(ix1,jx2);
 239             dy12             = _mm_sub_ps(iy1,jy2);
 240             dz12             = _mm_sub_ps(iz1,jz2);
 241             dx13             = _mm_sub_ps(ix1,jx3);
 242             dy13             = _mm_sub_ps(iy1,jy3);
 243             dz13             = _mm_sub_ps(iz1,jz3);
 244             dx21             = _mm_sub_ps(ix2,jx1);
 245             dy21             = _mm_sub_ps(iy2,jy1);
 246             dz21             = _mm_sub_ps(iz2,jz1);
 247             dx22             = _mm_sub_ps(ix2,jx2);
 248             dy22             = _mm_sub_ps(iy2,jy2);
 249             dz22             = _mm_sub_ps(iz2,jz2);
 250             dx23             = _mm_sub_ps(ix2,jx3);
 251             dy23             = _mm_sub_ps(iy2,jy3);
 252             dz23             = _mm_sub_ps(iz2,jz3);
 253             dx31             = _mm_sub_ps(ix3,jx1);
 254             dy31             = _mm_sub_ps(iy3,jy1);
 255             dz31             = _mm_sub_ps(iz3,jz1);
 256             dx32             = _mm_sub_ps(ix3,jx2);
 257             dy32             = _mm_sub_ps(iy3,jy2);
 258             dz32             = _mm_sub_ps(iz3,jz2);
 259             dx33             = _mm_sub_ps(ix3,jx3);
 260             dy33             = _mm_sub_ps(iy3,jy3);
 261             dz33             = _mm_sub_ps(iz3,jz3);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 265             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 266             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 267             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 268             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 269             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 270             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 271             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 272             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 273             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 274
 275             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 276             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 277             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 278             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 279             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 280             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 281             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 282             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 283             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 284
 285             rinvsq00         = gmx_mm_inv_ps(rsq00);
 286             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 287             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 288             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 289             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 290             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 291             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 292             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 293             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 294             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 295
 296             fjx0             = _mm_setzero_ps();
 297             fjy0             = _mm_setzero_ps();
 298             fjz0             = _mm_setzero_ps();
 299             fjx1             = _mm_setzero_ps();
 300             fjy1             = _mm_setzero_ps();
 301             fjz1             = _mm_setzero_ps();
 302             fjx2             = _mm_setzero_ps();
 303             fjy2             = _mm_setzero_ps();
 304             fjz2             = _mm_setzero_ps();
 305             fjx3             = _mm_setzero_ps();
 306             fjy3             = _mm_setzero_ps();
 307             fjz3             = _mm_setzero_ps();
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             /* LENNARD-JONES DISPERSION/REPULSION */
 314
 315             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 316             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 317             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 318             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 319             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 320
 321             /* Update potential sum for this i atom from the interaction with this j atom. */
 322             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 323
 324             fscal            = fvdw;
 325
 326             /* Calculate temporary vectorial force */
 327             tx               = _mm_mul_ps(fscal,dx00);
 328             ty               = _mm_mul_ps(fscal,dy00);
 329             tz               = _mm_mul_ps(fscal,dz00);
 330
 331             /* Update vectorial force */
 332             fix0             = _mm_add_ps(fix0,tx);
 333             fiy0             = _mm_add_ps(fiy0,ty);
 334             fiz0             = _mm_add_ps(fiz0,tz);
 335
 336             fjx0             = _mm_add_ps(fjx0,tx);
 337             fjy0             = _mm_add_ps(fjy0,ty);
 338             fjz0             = _mm_add_ps(fjz0,tz);
 339
 340             /**************************
 341              * CALCULATE INTERACTIONS *
 342              **************************/
 343
 344             r11              = _mm_mul_ps(rsq11,rinv11);
 345
 346             /* EWALD ELECTROSTATICS */
 347
 348             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 349             ewrt             = _mm_mul_ps(r11,ewtabscale);
 350             ewitab           = _mm_cvttps_epi32(ewrt);
 351             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 352             ewitab           = _mm_slli_epi32(ewitab,2);
 353             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 354             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 355             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 356             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 357             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 358             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 359             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 360             velec            = _mm_mul_ps(qq11,_mm_sub_ps(rinv11,velec));
 361             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368             /* Calculate temporary vectorial force */
 369             tx               = _mm_mul_ps(fscal,dx11);
 370             ty               = _mm_mul_ps(fscal,dy11);
 371             tz               = _mm_mul_ps(fscal,dz11);
 372
 373             /* Update vectorial force */
 374             fix1             = _mm_add_ps(fix1,tx);
 375             fiy1             = _mm_add_ps(fiy1,ty);
 376             fiz1             = _mm_add_ps(fiz1,tz);
 377
 378             fjx1             = _mm_add_ps(fjx1,tx);
 379             fjy1             = _mm_add_ps(fjy1,ty);
 380             fjz1             = _mm_add_ps(fjz1,tz);
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             r12              = _mm_mul_ps(rsq12,rinv12);
 387
 388             /* EWALD ELECTROSTATICS */
 389
 390             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 391             ewrt             = _mm_mul_ps(r12,ewtabscale);
 392             ewitab           = _mm_cvttps_epi32(ewrt);
 393             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 394             ewitab           = _mm_slli_epi32(ewitab,2);
 395             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 396             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 397             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 398             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 399             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 400             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 401             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 402             velec            = _mm_mul_ps(qq12,_mm_sub_ps(rinv12,velec));
 403             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
 404
 405             /* Update potential sum for this i atom from the interaction with this j atom. */
 406             velecsum         = _mm_add_ps(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             /* Calculate temporary vectorial force */
 411             tx               = _mm_mul_ps(fscal,dx12);
 412             ty               = _mm_mul_ps(fscal,dy12);
 413             tz               = _mm_mul_ps(fscal,dz12);
 414
 415             /* Update vectorial force */
 416             fix1             = _mm_add_ps(fix1,tx);
 417             fiy1             = _mm_add_ps(fiy1,ty);
 418             fiz1             = _mm_add_ps(fiz1,tz);
 419
 420             fjx2             = _mm_add_ps(fjx2,tx);
 421             fjy2             = _mm_add_ps(fjy2,ty);
 422             fjz2             = _mm_add_ps(fjz2,tz);
 423
 424             /**************************
 425              * CALCULATE INTERACTIONS *
 426              **************************/
 427
 428             r13              = _mm_mul_ps(rsq13,rinv13);
 429
 430             /* EWALD ELECTROSTATICS */
 431
 432             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 433             ewrt             = _mm_mul_ps(r13,ewtabscale);
 434             ewitab           = _mm_cvttps_epi32(ewrt);
 435             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 436             ewitab           = _mm_slli_epi32(ewitab,2);
 437             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 438             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 439             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 440             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 441             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 442             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 443             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 444             velec            = _mm_mul_ps(qq13,_mm_sub_ps(rinv13,velec));
 445             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
 446
 447             /* Update potential sum for this i atom from the interaction with this j atom. */
 448             velecsum         = _mm_add_ps(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             /* Calculate temporary vectorial force */
 453             tx               = _mm_mul_ps(fscal,dx13);
 454             ty               = _mm_mul_ps(fscal,dy13);
 455             tz               = _mm_mul_ps(fscal,dz13);
 456
 457             /* Update vectorial force */
 458             fix1             = _mm_add_ps(fix1,tx);
 459             fiy1             = _mm_add_ps(fiy1,ty);
 460             fiz1             = _mm_add_ps(fiz1,tz);
 461
 462             fjx3             = _mm_add_ps(fjx3,tx);
 463             fjy3             = _mm_add_ps(fjy3,ty);
 464             fjz3             = _mm_add_ps(fjz3,tz);
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             r21              = _mm_mul_ps(rsq21,rinv21);
 471
 472             /* EWALD ELECTROSTATICS */
 473
 474             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 475             ewrt             = _mm_mul_ps(r21,ewtabscale);
 476             ewitab           = _mm_cvttps_epi32(ewrt);
 477             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 478             ewitab           = _mm_slli_epi32(ewitab,2);
 479             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 480             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 481             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 482             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 483             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 484             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 485             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 486             velec            = _mm_mul_ps(qq21,_mm_sub_ps(rinv21,velec));
 487             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
 488
 489             /* Update potential sum for this i atom from the interaction with this j atom. */
 490             velecsum         = _mm_add_ps(velecsum,velec);
 491
 492             fscal            = felec;
 493
 494             /* Calculate temporary vectorial force */
 495             tx               = _mm_mul_ps(fscal,dx21);
 496             ty               = _mm_mul_ps(fscal,dy21);
 497             tz               = _mm_mul_ps(fscal,dz21);
 498
 499             /* Update vectorial force */
 500             fix2             = _mm_add_ps(fix2,tx);
 501             fiy2             = _mm_add_ps(fiy2,ty);
 502             fiz2             = _mm_add_ps(fiz2,tz);
 503
 504             fjx1             = _mm_add_ps(fjx1,tx);
 505             fjy1             = _mm_add_ps(fjy1,ty);
 506             fjz1             = _mm_add_ps(fjz1,tz);
 507
 508             /**************************
 509              * CALCULATE INTERACTIONS *
 510              **************************/
 511
 512             r22              = _mm_mul_ps(rsq22,rinv22);
 513
 514             /* EWALD ELECTROSTATICS */
 515
 516             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 517             ewrt             = _mm_mul_ps(r22,ewtabscale);
 518             ewitab           = _mm_cvttps_epi32(ewrt);
 519             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 520             ewitab           = _mm_slli_epi32(ewitab,2);
 521             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 522             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 523             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 524             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 525             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 526             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 527             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 528             velec            = _mm_mul_ps(qq22,_mm_sub_ps(rinv22,velec));
 529             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
 530
 531             /* Update potential sum for this i atom from the interaction with this j atom. */
 532             velecsum         = _mm_add_ps(velecsum,velec);
 533
 534             fscal            = felec;
 535
 536             /* Calculate temporary vectorial force */
 537             tx               = _mm_mul_ps(fscal,dx22);
 538             ty               = _mm_mul_ps(fscal,dy22);
 539             tz               = _mm_mul_ps(fscal,dz22);
 540
 541             /* Update vectorial force */
 542             fix2             = _mm_add_ps(fix2,tx);
 543             fiy2             = _mm_add_ps(fiy2,ty);
 544             fiz2             = _mm_add_ps(fiz2,tz);
 545
 546             fjx2             = _mm_add_ps(fjx2,tx);
 547             fjy2             = _mm_add_ps(fjy2,ty);
 548             fjz2             = _mm_add_ps(fjz2,tz);
 549
 550             /**************************
 551              * CALCULATE INTERACTIONS *
 552              **************************/
 553
 554             r23              = _mm_mul_ps(rsq23,rinv23);
 555
 556             /* EWALD ELECTROSTATICS */
 557
 558             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 559             ewrt             = _mm_mul_ps(r23,ewtabscale);
 560             ewitab           = _mm_cvttps_epi32(ewrt);
 561             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 562             ewitab           = _mm_slli_epi32(ewitab,2);
 563             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 564             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 565             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 566             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 567             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 568             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 569             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 570             velec            = _mm_mul_ps(qq23,_mm_sub_ps(rinv23,velec));
 571             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
 572
 573             /* Update potential sum for this i atom from the interaction with this j atom. */
 574             velecsum         = _mm_add_ps(velecsum,velec);
 575
 576             fscal            = felec;
 577
 578             /* Calculate temporary vectorial force */
 579             tx               = _mm_mul_ps(fscal,dx23);
 580             ty               = _mm_mul_ps(fscal,dy23);
 581             tz               = _mm_mul_ps(fscal,dz23);
 582
 583             /* Update vectorial force */
 584             fix2             = _mm_add_ps(fix2,tx);
 585             fiy2             = _mm_add_ps(fiy2,ty);
 586             fiz2             = _mm_add_ps(fiz2,tz);
 587
 588             fjx3             = _mm_add_ps(fjx3,tx);
 589             fjy3             = _mm_add_ps(fjy3,ty);
 590             fjz3             = _mm_add_ps(fjz3,tz);
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             r31              = _mm_mul_ps(rsq31,rinv31);
 597
 598             /* EWALD ELECTROSTATICS */
 599
 600             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 601             ewrt             = _mm_mul_ps(r31,ewtabscale);
 602             ewitab           = _mm_cvttps_epi32(ewrt);
 603             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 604             ewitab           = _mm_slli_epi32(ewitab,2);
 605             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 606             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 607             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 608             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 609             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 610             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 611             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 612             velec            = _mm_mul_ps(qq31,_mm_sub_ps(rinv31,velec));
 613             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
 614
 615             /* Update potential sum for this i atom from the interaction with this j atom. */
 616             velecsum         = _mm_add_ps(velecsum,velec);
 617
 618             fscal            = felec;
 619
 620             /* Calculate temporary vectorial force */
 621             tx               = _mm_mul_ps(fscal,dx31);
 622             ty               = _mm_mul_ps(fscal,dy31);
 623             tz               = _mm_mul_ps(fscal,dz31);
 624
 625             /* Update vectorial force */
 626             fix3             = _mm_add_ps(fix3,tx);
 627             fiy3             = _mm_add_ps(fiy3,ty);
 628             fiz3             = _mm_add_ps(fiz3,tz);
 629
 630             fjx1             = _mm_add_ps(fjx1,tx);
 631             fjy1             = _mm_add_ps(fjy1,ty);
 632             fjz1             = _mm_add_ps(fjz1,tz);
 633
 634             /**************************
 635              * CALCULATE INTERACTIONS *
 636              **************************/
 637
 638             r32              = _mm_mul_ps(rsq32,rinv32);
 639
 640             /* EWALD ELECTROSTATICS */
 641
 642             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 643             ewrt             = _mm_mul_ps(r32,ewtabscale);
 644             ewitab           = _mm_cvttps_epi32(ewrt);
 645             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 646             ewitab           = _mm_slli_epi32(ewitab,2);
 647             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 648             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 649             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 650             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 651             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 652             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 653             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 654             velec            = _mm_mul_ps(qq32,_mm_sub_ps(rinv32,velec));
 655             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
 656
 657             /* Update potential sum for this i atom from the interaction with this j atom. */
 658             velecsum         = _mm_add_ps(velecsum,velec);
 659
 660             fscal            = felec;
 661
 662             /* Calculate temporary vectorial force */
 663             tx               = _mm_mul_ps(fscal,dx32);
 664             ty               = _mm_mul_ps(fscal,dy32);
 665             tz               = _mm_mul_ps(fscal,dz32);
 666
 667             /* Update vectorial force */
 668             fix3             = _mm_add_ps(fix3,tx);
 669             fiy3             = _mm_add_ps(fiy3,ty);
 670             fiz3             = _mm_add_ps(fiz3,tz);
 671
 672             fjx2             = _mm_add_ps(fjx2,tx);
 673             fjy2             = _mm_add_ps(fjy2,ty);
 674             fjz2             = _mm_add_ps(fjz2,tz);
 675
 676             /**************************
 677              * CALCULATE INTERACTIONS *
 678              **************************/
 679
 680             r33              = _mm_mul_ps(rsq33,rinv33);
 681
 682             /* EWALD ELECTROSTATICS */
 683
 684             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 685             ewrt             = _mm_mul_ps(r33,ewtabscale);
 686             ewitab           = _mm_cvttps_epi32(ewrt);
 687             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 688             ewitab           = _mm_slli_epi32(ewitab,2);
 689             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 690             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 691             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 692             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 693             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 694             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 695             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 696             velec            = _mm_mul_ps(qq33,_mm_sub_ps(rinv33,velec));
 697             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
 698
 699             /* Update potential sum for this i atom from the interaction with this j atom. */
 700             velecsum         = _mm_add_ps(velecsum,velec);
 701
 702             fscal            = felec;
 703
 704             /* Calculate temporary vectorial force */
 705             tx               = _mm_mul_ps(fscal,dx33);
 706             ty               = _mm_mul_ps(fscal,dy33);
 707             tz               = _mm_mul_ps(fscal,dz33);
 708
 709             /* Update vectorial force */
 710             fix3             = _mm_add_ps(fix3,tx);
 711             fiy3             = _mm_add_ps(fiy3,ty);
 712             fiz3             = _mm_add_ps(fiz3,tz);
 713
 714             fjx3             = _mm_add_ps(fjx3,tx);
 715             fjy3             = _mm_add_ps(fjy3,ty);
 716             fjz3             = _mm_add_ps(fjz3,tz);
 717
 718             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 719                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 720                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 721                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 722
 723             /* Inner loop uses 404 flops */
 724         }
 725
 726         if(jidx<j_index_end)
 727         {
 728
 729             /* Get j neighbor index, and coordinate index */
 730             jnrA             = jjnr[jidx];
 731             jnrB             = jjnr[jidx+1];
 732             jnrC             = jjnr[jidx+2];
 733             jnrD             = jjnr[jidx+3];
 734
 735             /* Sign of each element will be negative for non-real atoms.
 736              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 737              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 738              */
 739             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 740             jnrA       = (jnrA>=0) ? jnrA : 0;
 741             jnrB       = (jnrB>=0) ? jnrB : 0;
 742             jnrC       = (jnrC>=0) ? jnrC : 0;
 743             jnrD       = (jnrD>=0) ? jnrD : 0;
 744
 745             j_coord_offsetA  = DIM*jnrA;
 746             j_coord_offsetB  = DIM*jnrB;
 747             j_coord_offsetC  = DIM*jnrC;
 748             j_coord_offsetD  = DIM*jnrD;
 749
 750             /* load j atom coordinates */
 751             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 752                                               x+j_coord_offsetC,x+j_coord_offsetD,
 753                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 754                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 755
 756             /* Calculate displacement vector */
 757             dx00             = _mm_sub_ps(ix0,jx0);
 758             dy00             = _mm_sub_ps(iy0,jy0);
 759             dz00             = _mm_sub_ps(iz0,jz0);
 760             dx11             = _mm_sub_ps(ix1,jx1);
 761             dy11             = _mm_sub_ps(iy1,jy1);
 762             dz11             = _mm_sub_ps(iz1,jz1);
 763             dx12             = _mm_sub_ps(ix1,jx2);
 764             dy12             = _mm_sub_ps(iy1,jy2);
 765             dz12             = _mm_sub_ps(iz1,jz2);
 766             dx13             = _mm_sub_ps(ix1,jx3);
 767             dy13             = _mm_sub_ps(iy1,jy3);
 768             dz13             = _mm_sub_ps(iz1,jz3);
 769             dx21             = _mm_sub_ps(ix2,jx1);
 770             dy21             = _mm_sub_ps(iy2,jy1);
 771             dz21             = _mm_sub_ps(iz2,jz1);
 772             dx22             = _mm_sub_ps(ix2,jx2);
 773             dy22             = _mm_sub_ps(iy2,jy2);
 774             dz22             = _mm_sub_ps(iz2,jz2);
 775             dx23             = _mm_sub_ps(ix2,jx3);
 776             dy23             = _mm_sub_ps(iy2,jy3);
 777             dz23             = _mm_sub_ps(iz2,jz3);
 778             dx31             = _mm_sub_ps(ix3,jx1);
 779             dy31             = _mm_sub_ps(iy3,jy1);
 780             dz31             = _mm_sub_ps(iz3,jz1);
 781             dx32             = _mm_sub_ps(ix3,jx2);
 782             dy32             = _mm_sub_ps(iy3,jy2);
 783             dz32             = _mm_sub_ps(iz3,jz2);
 784             dx33             = _mm_sub_ps(ix3,jx3);
 785             dy33             = _mm_sub_ps(iy3,jy3);
 786             dz33             = _mm_sub_ps(iz3,jz3);
 787
 788             /* Calculate squared distance and things based on it */
 789             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 790             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 791             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 792             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 793             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 794             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 795             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 796             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 797             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 798             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 799
 800             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 801             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 802             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 803             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 804             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 805             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 806             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 807             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 808             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 809
 810             rinvsq00         = gmx_mm_inv_ps(rsq00);
 811             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 812             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 813             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 814             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 815             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 816             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 817             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 818             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 819             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 820
 821             fjx0             = _mm_setzero_ps();
 822             fjy0             = _mm_setzero_ps();
 823             fjz0             = _mm_setzero_ps();
 824             fjx1             = _mm_setzero_ps();
 825             fjy1             = _mm_setzero_ps();
 826             fjz1             = _mm_setzero_ps();
 827             fjx2             = _mm_setzero_ps();
 828             fjy2             = _mm_setzero_ps();
 829             fjz2             = _mm_setzero_ps();
 830             fjx3             = _mm_setzero_ps();
 831             fjy3             = _mm_setzero_ps();
 832             fjz3             = _mm_setzero_ps();
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             /* LENNARD-JONES DISPERSION/REPULSION */
 839
 840             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 841             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 842             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 843             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 844             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 845
 846             /* Update potential sum for this i atom from the interaction with this j atom. */
 847             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 848             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 849
 850             fscal            = fvdw;
 851
 852             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 853
 854             /* Calculate temporary vectorial force */
 855             tx               = _mm_mul_ps(fscal,dx00);
 856             ty               = _mm_mul_ps(fscal,dy00);
 857             tz               = _mm_mul_ps(fscal,dz00);
 858
 859             /* Update vectorial force */
 860             fix0             = _mm_add_ps(fix0,tx);
 861             fiy0             = _mm_add_ps(fiy0,ty);
 862             fiz0             = _mm_add_ps(fiz0,tz);
 863
 864             fjx0             = _mm_add_ps(fjx0,tx);
 865             fjy0             = _mm_add_ps(fjy0,ty);
 866             fjz0             = _mm_add_ps(fjz0,tz);
 867
 868             /**************************
 869              * CALCULATE INTERACTIONS *
 870              **************************/
 871
 872             r11              = _mm_mul_ps(rsq11,rinv11);
 873             r11              = _mm_andnot_ps(dummy_mask,r11);
 874
 875             /* EWALD ELECTROSTATICS */
 876
 877             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 878             ewrt             = _mm_mul_ps(r11,ewtabscale);
 879             ewitab           = _mm_cvttps_epi32(ewrt);
 880             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 881             ewitab           = _mm_slli_epi32(ewitab,2);
 882             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 883             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 884             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 885             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 886             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 887             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 888             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 889             velec            = _mm_mul_ps(qq11,_mm_sub_ps(rinv11,velec));
 890             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
 891
 892             /* Update potential sum for this i atom from the interaction with this j atom. */
 893             velec            = _mm_andnot_ps(dummy_mask,velec);
 894             velecsum         = _mm_add_ps(velecsum,velec);
 895
 896             fscal            = felec;
 897
 898             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 899
 900             /* Calculate temporary vectorial force */
 901             tx               = _mm_mul_ps(fscal,dx11);
 902             ty               = _mm_mul_ps(fscal,dy11);
 903             tz               = _mm_mul_ps(fscal,dz11);
 904
 905             /* Update vectorial force */
 906             fix1             = _mm_add_ps(fix1,tx);
 907             fiy1             = _mm_add_ps(fiy1,ty);
 908             fiz1             = _mm_add_ps(fiz1,tz);
 909
 910             fjx1             = _mm_add_ps(fjx1,tx);
 911             fjy1             = _mm_add_ps(fjy1,ty);
 912             fjz1             = _mm_add_ps(fjz1,tz);
 913
 914             /**************************
 915              * CALCULATE INTERACTIONS *
 916              **************************/
 917
 918             r12              = _mm_mul_ps(rsq12,rinv12);
 919             r12              = _mm_andnot_ps(dummy_mask,r12);
 920
 921             /* EWALD ELECTROSTATICS */
 922
 923             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 924             ewrt             = _mm_mul_ps(r12,ewtabscale);
 925             ewitab           = _mm_cvttps_epi32(ewrt);
 926             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 927             ewitab           = _mm_slli_epi32(ewitab,2);
 928             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 929             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 930             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 931             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 932             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 933             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 934             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 935             velec            = _mm_mul_ps(qq12,_mm_sub_ps(rinv12,velec));
 936             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
 937
 938             /* Update potential sum for this i atom from the interaction with this j atom. */
 939             velec            = _mm_andnot_ps(dummy_mask,velec);
 940             velecsum         = _mm_add_ps(velecsum,velec);
 941
 942             fscal            = felec;
 943
 944             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 945
 946             /* Calculate temporary vectorial force */
 947             tx               = _mm_mul_ps(fscal,dx12);
 948             ty               = _mm_mul_ps(fscal,dy12);
 949             tz               = _mm_mul_ps(fscal,dz12);
 950
 951             /* Update vectorial force */
 952             fix1             = _mm_add_ps(fix1,tx);
 953             fiy1             = _mm_add_ps(fiy1,ty);
 954             fiz1             = _mm_add_ps(fiz1,tz);
 955
 956             fjx2             = _mm_add_ps(fjx2,tx);
 957             fjy2             = _mm_add_ps(fjy2,ty);
 958             fjz2             = _mm_add_ps(fjz2,tz);
 959
 960             /**************************
 961              * CALCULATE INTERACTIONS *
 962              **************************/
 963
 964             r13              = _mm_mul_ps(rsq13,rinv13);
 965             r13              = _mm_andnot_ps(dummy_mask,r13);
 966
 967             /* EWALD ELECTROSTATICS */
 968
 969             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 970             ewrt             = _mm_mul_ps(r13,ewtabscale);
 971             ewitab           = _mm_cvttps_epi32(ewrt);
 972             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 973             ewitab           = _mm_slli_epi32(ewitab,2);
 974             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 975             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 976             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 977             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 978             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 979             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 980             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 981             velec            = _mm_mul_ps(qq13,_mm_sub_ps(rinv13,velec));
 982             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
 983
 984             /* Update potential sum for this i atom from the interaction with this j atom. */
 985             velec            = _mm_andnot_ps(dummy_mask,velec);
 986             velecsum         = _mm_add_ps(velecsum,velec);
 987
 988             fscal            = felec;
 989
 990             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 991
 992             /* Calculate temporary vectorial force */
 993             tx               = _mm_mul_ps(fscal,dx13);
 994             ty               = _mm_mul_ps(fscal,dy13);
 995             tz               = _mm_mul_ps(fscal,dz13);
 996
 997             /* Update vectorial force */
 998             fix1             = _mm_add_ps(fix1,tx);
 999             fiy1             = _mm_add_ps(fiy1,ty);
1000             fiz1             = _mm_add_ps(fiz1,tz);
1001
1002             fjx3             = _mm_add_ps(fjx3,tx);
1003             fjy3             = _mm_add_ps(fjy3,ty);
1004             fjz3             = _mm_add_ps(fjz3,tz);
1005
1006             /**************************
1007              * CALCULATE INTERACTIONS *
1008              **************************/
1009
1010             r21              = _mm_mul_ps(rsq21,rinv21);
1011             r21              = _mm_andnot_ps(dummy_mask,r21);
1012
1013             /* EWALD ELECTROSTATICS */
1014
1015             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1016             ewrt             = _mm_mul_ps(r21,ewtabscale);
1017             ewitab           = _mm_cvttps_epi32(ewrt);
1018             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1019             ewitab           = _mm_slli_epi32(ewitab,2);
1020             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1021             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1022             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1023             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1024             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1025             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1026             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1027             velec            = _mm_mul_ps(qq21,_mm_sub_ps(rinv21,velec));
1028             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1029
1030             /* Update potential sum for this i atom from the interaction with this j atom. */
1031             velec            = _mm_andnot_ps(dummy_mask,velec);
1032             velecsum         = _mm_add_ps(velecsum,velec);
1033
1034             fscal            = felec;
1035
1036             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1037
1038             /* Calculate temporary vectorial force */
1039             tx               = _mm_mul_ps(fscal,dx21);
1040             ty               = _mm_mul_ps(fscal,dy21);
1041             tz               = _mm_mul_ps(fscal,dz21);
1042
1043             /* Update vectorial force */
1044             fix2             = _mm_add_ps(fix2,tx);
1045             fiy2             = _mm_add_ps(fiy2,ty);
1046             fiz2             = _mm_add_ps(fiz2,tz);
1047
1048             fjx1             = _mm_add_ps(fjx1,tx);
1049             fjy1             = _mm_add_ps(fjy1,ty);
1050             fjz1             = _mm_add_ps(fjz1,tz);
1051
1052             /**************************
1053              * CALCULATE INTERACTIONS *
1054              **************************/
1055
1056             r22              = _mm_mul_ps(rsq22,rinv22);
1057             r22              = _mm_andnot_ps(dummy_mask,r22);
1058
1059             /* EWALD ELECTROSTATICS */
1060
1061             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1062             ewrt             = _mm_mul_ps(r22,ewtabscale);
1063             ewitab           = _mm_cvttps_epi32(ewrt);
1064             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1065             ewitab           = _mm_slli_epi32(ewitab,2);
1066             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1067             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1068             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1069             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1070             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1071             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1072             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1073             velec            = _mm_mul_ps(qq22,_mm_sub_ps(rinv22,velec));
1074             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
1075
1076             /* Update potential sum for this i atom from the interaction with this j atom. */
1077             velec            = _mm_andnot_ps(dummy_mask,velec);
1078             velecsum         = _mm_add_ps(velecsum,velec);
1079
1080             fscal            = felec;
1081
1082             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1083
1084             /* Calculate temporary vectorial force */
1085             tx               = _mm_mul_ps(fscal,dx22);
1086             ty               = _mm_mul_ps(fscal,dy22);
1087             tz               = _mm_mul_ps(fscal,dz22);
1088
1089             /* Update vectorial force */
1090             fix2             = _mm_add_ps(fix2,tx);
1091             fiy2             = _mm_add_ps(fiy2,ty);
1092             fiz2             = _mm_add_ps(fiz2,tz);
1093
1094             fjx2             = _mm_add_ps(fjx2,tx);
1095             fjy2             = _mm_add_ps(fjy2,ty);
1096             fjz2             = _mm_add_ps(fjz2,tz);
1097
1098             /**************************
1099              * CALCULATE INTERACTIONS *
1100              **************************/
1101
1102             r23              = _mm_mul_ps(rsq23,rinv23);
1103             r23              = _mm_andnot_ps(dummy_mask,r23);
1104
1105             /* EWALD ELECTROSTATICS */
1106
1107             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1108             ewrt             = _mm_mul_ps(r23,ewtabscale);
1109             ewitab           = _mm_cvttps_epi32(ewrt);
1110             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1111             ewitab           = _mm_slli_epi32(ewitab,2);
1112             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1113             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1114             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1115             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1116             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1117             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1118             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1119             velec            = _mm_mul_ps(qq23,_mm_sub_ps(rinv23,velec));
1120             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
1121
1122             /* Update potential sum for this i atom from the interaction with this j atom. */
1123             velec            = _mm_andnot_ps(dummy_mask,velec);
1124             velecsum         = _mm_add_ps(velecsum,velec);
1125
1126             fscal            = felec;
1127
1128             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1129
1130             /* Calculate temporary vectorial force */
1131             tx               = _mm_mul_ps(fscal,dx23);
1132             ty               = _mm_mul_ps(fscal,dy23);
1133             tz               = _mm_mul_ps(fscal,dz23);
1134
1135             /* Update vectorial force */
1136             fix2             = _mm_add_ps(fix2,tx);
1137             fiy2             = _mm_add_ps(fiy2,ty);
1138             fiz2             = _mm_add_ps(fiz2,tz);
1139
1140             fjx3             = _mm_add_ps(fjx3,tx);
1141             fjy3             = _mm_add_ps(fjy3,ty);
1142             fjz3             = _mm_add_ps(fjz3,tz);
1143
1144             /**************************
1145              * CALCULATE INTERACTIONS *
1146              **************************/
1147
1148             r31              = _mm_mul_ps(rsq31,rinv31);
1149             r31              = _mm_andnot_ps(dummy_mask,r31);
1150
1151             /* EWALD ELECTROSTATICS */
1152
1153             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1154             ewrt             = _mm_mul_ps(r31,ewtabscale);
1155             ewitab           = _mm_cvttps_epi32(ewrt);
1156             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1157             ewitab           = _mm_slli_epi32(ewitab,2);
1158             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1159             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1160             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1161             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1162             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1163             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1164             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1165             velec            = _mm_mul_ps(qq31,_mm_sub_ps(rinv31,velec));
1166             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
1167
1168             /* Update potential sum for this i atom from the interaction with this j atom. */
1169             velec            = _mm_andnot_ps(dummy_mask,velec);
1170             velecsum         = _mm_add_ps(velecsum,velec);
1171
1172             fscal            = felec;
1173
1174             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1175
1176             /* Calculate temporary vectorial force */
1177             tx               = _mm_mul_ps(fscal,dx31);
1178             ty               = _mm_mul_ps(fscal,dy31);
1179             tz               = _mm_mul_ps(fscal,dz31);
1180
1181             /* Update vectorial force */
1182             fix3             = _mm_add_ps(fix3,tx);
1183             fiy3             = _mm_add_ps(fiy3,ty);
1184             fiz3             = _mm_add_ps(fiz3,tz);
1185
1186             fjx1             = _mm_add_ps(fjx1,tx);
1187             fjy1             = _mm_add_ps(fjy1,ty);
1188             fjz1             = _mm_add_ps(fjz1,tz);
1189
1190             /**************************
1191              * CALCULATE INTERACTIONS *
1192              **************************/
1193
1194             r32              = _mm_mul_ps(rsq32,rinv32);
1195             r32              = _mm_andnot_ps(dummy_mask,r32);
1196
1197             /* EWALD ELECTROSTATICS */
1198
1199             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1200             ewrt             = _mm_mul_ps(r32,ewtabscale);
1201             ewitab           = _mm_cvttps_epi32(ewrt);
1202             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1203             ewitab           = _mm_slli_epi32(ewitab,2);
1204             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1205             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1206             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1207             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1208             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1209             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1210             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1211             velec            = _mm_mul_ps(qq32,_mm_sub_ps(rinv32,velec));
1212             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
1213
1214             /* Update potential sum for this i atom from the interaction with this j atom. */
1215             velec            = _mm_andnot_ps(dummy_mask,velec);
1216             velecsum         = _mm_add_ps(velecsum,velec);
1217
1218             fscal            = felec;
1219
1220             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1221
1222             /* Calculate temporary vectorial force */
1223             tx               = _mm_mul_ps(fscal,dx32);
1224             ty               = _mm_mul_ps(fscal,dy32);
1225             tz               = _mm_mul_ps(fscal,dz32);
1226
1227             /* Update vectorial force */
1228             fix3             = _mm_add_ps(fix3,tx);
1229             fiy3             = _mm_add_ps(fiy3,ty);
1230             fiz3             = _mm_add_ps(fiz3,tz);
1231
1232             fjx2             = _mm_add_ps(fjx2,tx);
1233             fjy2             = _mm_add_ps(fjy2,ty);
1234             fjz2             = _mm_add_ps(fjz2,tz);
1235
1236             /**************************
1237              * CALCULATE INTERACTIONS *
1238              **************************/
1239
1240             r33              = _mm_mul_ps(rsq33,rinv33);
1241             r33              = _mm_andnot_ps(dummy_mask,r33);
1242
1243             /* EWALD ELECTROSTATICS */
1244
1245             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1246             ewrt             = _mm_mul_ps(r33,ewtabscale);
1247             ewitab           = _mm_cvttps_epi32(ewrt);
1248             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1249             ewitab           = _mm_slli_epi32(ewitab,2);
1250             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1251             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1252             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1253             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1254             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1255             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1256             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1257             velec            = _mm_mul_ps(qq33,_mm_sub_ps(rinv33,velec));
1258             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
1259
1260             /* Update potential sum for this i atom from the interaction with this j atom. */
1261             velec            = _mm_andnot_ps(dummy_mask,velec);
1262             velecsum         = _mm_add_ps(velecsum,velec);
1263
1264             fscal            = felec;
1265
1266             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1267
1268             /* Calculate temporary vectorial force */
1269             tx               = _mm_mul_ps(fscal,dx33);
1270             ty               = _mm_mul_ps(fscal,dy33);
1271             tz               = _mm_mul_ps(fscal,dz33);
1272
1273             /* Update vectorial force */
1274             fix3             = _mm_add_ps(fix3,tx);
1275             fiy3             = _mm_add_ps(fiy3,ty);
1276             fiz3             = _mm_add_ps(fiz3,tz);
1277
1278             fjx3             = _mm_add_ps(fjx3,tx);
1279             fjy3             = _mm_add_ps(fjy3,ty);
1280             fjz3             = _mm_add_ps(fjz3,tz);
1281
1282             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1283                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1284                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1285                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1286
1287             /* Inner loop uses 413 flops */
1288         }
1289
1290         /* End of innermost loop */
1291
1292         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1293                                               f+i_coord_offset,fshift+i_shift_offset);
1294
1295         ggid                        = gid[iidx];
1296         /* Update potential energies */
1297         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1298         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1299
1300         /* Increment number of inner iterations */
1301         inneriter                  += j_index_end - j_index_start;
1302
1303         /* Outer loop uses 38 flops */
1304     }
1305
1306     /* Increment number of outer iterations */
1307     outeriter        += nri;
1308
1309     /* Update outer/inner flops */
1310
1311     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*38 + inneriter*413);
1312 }
1313 /*
1314  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW4W4_F_sse2_single
1315  * Electrostatics interaction: Ewald
1316  * VdW interaction:            LennardJones
1317  * Geometry:                   Water4-Water4
1318  * Calculate force/pot:        Force
1319  */
1320 void
1321 nb_kernel_ElecEw_VdwLJ_GeomW4W4_F_sse2_single
1322                     (t_nblist * gmx_restrict                nlist,
1323                      rvec * gmx_restrict                    xx,
1324                      rvec * gmx_restrict                    ff,
1325                      t_forcerec * gmx_restrict              fr,
1326                      t_mdatoms * gmx_restrict               mdatoms,
1327                      nb_kernel_data_t * gmx_restrict        kernel_data,
1328                      t_nrnb * gmx_restrict                  nrnb)
1329 {
1330     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1331      * just 0 for non-waters.
1332      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1333      * jnr indices corresponding to data put in the four positions in the SIMD register.
1334      */
1335     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1336     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1337     int              jnrA,jnrB,jnrC,jnrD;
1338     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1339     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1340     real             shX,shY,shZ,rcutoff_scalar;
1341     real             *shiftvec,*fshift,*x,*f;
1342     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1343     int              vdwioffset0;
1344     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1345     int              vdwioffset1;
1346     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1347     int              vdwioffset2;
1348     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1349     int              vdwioffset3;
1350     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1351     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1352     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1353     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1354     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1355     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1356     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1357     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1358     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1359     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1360     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1361     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1362     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1363     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1364     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1365     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1366     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1367     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1368     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1369     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1370     real             *charge;
1371     int              nvdwtype;
1372     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1373     int              *vdwtype;
1374     real             *vdwparam;
1375     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1376     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1377     __m128i          ewitab;
1378     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1379     real             *ewtab;
1380     __m128           dummy_mask,cutoff_mask;
1381     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1382     __m128           one     = _mm_set1_ps(1.0);
1383     __m128           two     = _mm_set1_ps(2.0);
1384     x                = xx[0];
1385     f                = ff[0];
1386
1387     nri              = nlist->nri;
1388     iinr             = nlist->iinr;
1389     jindex           = nlist->jindex;
1390     jjnr             = nlist->jjnr;
1391     shiftidx         = nlist->shift;
1392     gid              = nlist->gid;
1393     shiftvec         = fr->shift_vec[0];
1394     fshift           = fr->fshift[0];
1395     facel            = _mm_set1_ps(fr->epsfac);
1396     charge           = mdatoms->chargeA;
1397     nvdwtype         = fr->ntype;
1398     vdwparam         = fr->nbfp;
1399     vdwtype          = mdatoms->typeA;
1400
1401     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1402     ewtab            = fr->ic->tabq_coul_F;
1403     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1404     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1405
1406     /* Setup water-specific parameters */
1407     inr              = nlist->iinr[0];
1408     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1409     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1410     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1411     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1412
1413     jq1              = _mm_set1_ps(charge[inr+1]);
1414     jq2              = _mm_set1_ps(charge[inr+2]);
1415     jq3              = _mm_set1_ps(charge[inr+3]);
1416     vdwjidx0A        = 2*vdwtype[inr+0];
1417     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1418     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1419     qq11             = _mm_mul_ps(iq1,jq1);
1420     qq12             = _mm_mul_ps(iq1,jq2);
1421     qq13             = _mm_mul_ps(iq1,jq3);
1422     qq21             = _mm_mul_ps(iq2,jq1);
1423     qq22             = _mm_mul_ps(iq2,jq2);
1424     qq23             = _mm_mul_ps(iq2,jq3);
1425     qq31             = _mm_mul_ps(iq3,jq1);
1426     qq32             = _mm_mul_ps(iq3,jq2);
1427     qq33             = _mm_mul_ps(iq3,jq3);
1428
1429     /* Avoid stupid compiler warnings */
1430     jnrA = jnrB = jnrC = jnrD = 0;
1431     j_coord_offsetA = 0;
1432     j_coord_offsetB = 0;
1433     j_coord_offsetC = 0;
1434     j_coord_offsetD = 0;
1435
1436     outeriter        = 0;
1437     inneriter        = 0;
1438
1439     /* Start outer loop over neighborlists */
1440     for(iidx=0; iidx<nri; iidx++)
1441     {
1442         /* Load shift vector for this list */
1443         i_shift_offset   = DIM*shiftidx[iidx];
1444         shX              = shiftvec[i_shift_offset+XX];
1445         shY              = shiftvec[i_shift_offset+YY];
1446         shZ              = shiftvec[i_shift_offset+ZZ];
1447
1448         /* Load limits for loop over neighbors */
1449         j_index_start    = jindex[iidx];
1450         j_index_end      = jindex[iidx+1];
1451
1452         /* Get outer coordinate index */
1453         inr              = iinr[iidx];
1454         i_coord_offset   = DIM*inr;
1455
1456         /* Load i particle coords and add shift vector */
1457         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1458         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1459         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1460         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1461         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1462         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1463         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1464         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1465         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1466         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1467         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1468         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1469
1470         fix0             = _mm_setzero_ps();
1471         fiy0             = _mm_setzero_ps();
1472         fiz0             = _mm_setzero_ps();
1473         fix1             = _mm_setzero_ps();
1474         fiy1             = _mm_setzero_ps();
1475         fiz1             = _mm_setzero_ps();
1476         fix2             = _mm_setzero_ps();
1477         fiy2             = _mm_setzero_ps();
1478         fiz2             = _mm_setzero_ps();
1479         fix3             = _mm_setzero_ps();
1480         fiy3             = _mm_setzero_ps();
1481         fiz3             = _mm_setzero_ps();
1482
1483         /* Start inner kernel loop */
1484         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1485         {
1486
1487             /* Get j neighbor index, and coordinate index */
1488             jnrA             = jjnr[jidx];
1489             jnrB             = jjnr[jidx+1];
1490             jnrC             = jjnr[jidx+2];
1491             jnrD             = jjnr[jidx+3];
1492
1493             j_coord_offsetA  = DIM*jnrA;
1494             j_coord_offsetB  = DIM*jnrB;
1495             j_coord_offsetC  = DIM*jnrC;
1496             j_coord_offsetD  = DIM*jnrD;
1497
1498             /* load j atom coordinates */
1499             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1500                                               x+j_coord_offsetC,x+j_coord_offsetD,
1501                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1502                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1503
1504             /* Calculate displacement vector */
1505             dx00             = _mm_sub_ps(ix0,jx0);
1506             dy00             = _mm_sub_ps(iy0,jy0);
1507             dz00             = _mm_sub_ps(iz0,jz0);
1508             dx11             = _mm_sub_ps(ix1,jx1);
1509             dy11             = _mm_sub_ps(iy1,jy1);
1510             dz11             = _mm_sub_ps(iz1,jz1);
1511             dx12             = _mm_sub_ps(ix1,jx2);
1512             dy12             = _mm_sub_ps(iy1,jy2);
1513             dz12             = _mm_sub_ps(iz1,jz2);
1514             dx13             = _mm_sub_ps(ix1,jx3);
1515             dy13             = _mm_sub_ps(iy1,jy3);
1516             dz13             = _mm_sub_ps(iz1,jz3);
1517             dx21             = _mm_sub_ps(ix2,jx1);
1518             dy21             = _mm_sub_ps(iy2,jy1);
1519             dz21             = _mm_sub_ps(iz2,jz1);
1520             dx22             = _mm_sub_ps(ix2,jx2);
1521             dy22             = _mm_sub_ps(iy2,jy2);
1522             dz22             = _mm_sub_ps(iz2,jz2);
1523             dx23             = _mm_sub_ps(ix2,jx3);
1524             dy23             = _mm_sub_ps(iy2,jy3);
1525             dz23             = _mm_sub_ps(iz2,jz3);
1526             dx31             = _mm_sub_ps(ix3,jx1);
1527             dy31             = _mm_sub_ps(iy3,jy1);
1528             dz31             = _mm_sub_ps(iz3,jz1);
1529             dx32             = _mm_sub_ps(ix3,jx2);
1530             dy32             = _mm_sub_ps(iy3,jy2);
1531             dz32             = _mm_sub_ps(iz3,jz2);
1532             dx33             = _mm_sub_ps(ix3,jx3);
1533             dy33             = _mm_sub_ps(iy3,jy3);
1534             dz33             = _mm_sub_ps(iz3,jz3);
1535
1536             /* Calculate squared distance and things based on it */
1537             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1538             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1539             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1540             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1541             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1542             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1543             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1544             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1545             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1546             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1547
1548             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1549             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1550             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1551             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1552             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1553             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1554             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1555             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1556             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1557
1558             rinvsq00         = gmx_mm_inv_ps(rsq00);
1559             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1560             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1561             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1562             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1563             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1564             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1565             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1566             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1567             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1568
1569             fjx0             = _mm_setzero_ps();
1570             fjy0             = _mm_setzero_ps();
1571             fjz0             = _mm_setzero_ps();
1572             fjx1             = _mm_setzero_ps();
1573             fjy1             = _mm_setzero_ps();
1574             fjz1             = _mm_setzero_ps();
1575             fjx2             = _mm_setzero_ps();
1576             fjy2             = _mm_setzero_ps();
1577             fjz2             = _mm_setzero_ps();
1578             fjx3             = _mm_setzero_ps();
1579             fjy3             = _mm_setzero_ps();
1580             fjz3             = _mm_setzero_ps();
1581
1582             /**************************
1583              * CALCULATE INTERACTIONS *
1584              **************************/
1585
1586             /* LENNARD-JONES DISPERSION/REPULSION */
1587
1588             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1589             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1590
1591             fscal            = fvdw;
1592
1593             /* Calculate temporary vectorial force */
1594             tx               = _mm_mul_ps(fscal,dx00);
1595             ty               = _mm_mul_ps(fscal,dy00);
1596             tz               = _mm_mul_ps(fscal,dz00);
1597
1598             /* Update vectorial force */
1599             fix0             = _mm_add_ps(fix0,tx);
1600             fiy0             = _mm_add_ps(fiy0,ty);
1601             fiz0             = _mm_add_ps(fiz0,tz);
1602
1603             fjx0             = _mm_add_ps(fjx0,tx);
1604             fjy0             = _mm_add_ps(fjy0,ty);
1605             fjz0             = _mm_add_ps(fjz0,tz);
1606
1607             /**************************
1608              * CALCULATE INTERACTIONS *
1609              **************************/
1610
1611             r11              = _mm_mul_ps(rsq11,rinv11);
1612
1613             /* EWALD ELECTROSTATICS */
1614
1615             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1616             ewrt             = _mm_mul_ps(r11,ewtabscale);
1617             ewitab           = _mm_cvttps_epi32(ewrt);
1618             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1619             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1620                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1621                                          &ewtabF,&ewtabFn);
1622             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1623             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
1624
1625             fscal            = felec;
1626
1627             /* Calculate temporary vectorial force */
1628             tx               = _mm_mul_ps(fscal,dx11);
1629             ty               = _mm_mul_ps(fscal,dy11);
1630             tz               = _mm_mul_ps(fscal,dz11);
1631
1632             /* Update vectorial force */
1633             fix1             = _mm_add_ps(fix1,tx);
1634             fiy1             = _mm_add_ps(fiy1,ty);
1635             fiz1             = _mm_add_ps(fiz1,tz);
1636
1637             fjx1             = _mm_add_ps(fjx1,tx);
1638             fjy1             = _mm_add_ps(fjy1,ty);
1639             fjz1             = _mm_add_ps(fjz1,tz);
1640
1641             /**************************
1642              * CALCULATE INTERACTIONS *
1643              **************************/
1644
1645             r12              = _mm_mul_ps(rsq12,rinv12);
1646
1647             /* EWALD ELECTROSTATICS */
1648
1649             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1650             ewrt             = _mm_mul_ps(r12,ewtabscale);
1651             ewitab           = _mm_cvttps_epi32(ewrt);
1652             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1653             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1654                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1655                                          &ewtabF,&ewtabFn);
1656             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1657             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
1658
1659             fscal            = felec;
1660
1661             /* Calculate temporary vectorial force */
1662             tx               = _mm_mul_ps(fscal,dx12);
1663             ty               = _mm_mul_ps(fscal,dy12);
1664             tz               = _mm_mul_ps(fscal,dz12);
1665
1666             /* Update vectorial force */
1667             fix1             = _mm_add_ps(fix1,tx);
1668             fiy1             = _mm_add_ps(fiy1,ty);
1669             fiz1             = _mm_add_ps(fiz1,tz);
1670
1671             fjx2             = _mm_add_ps(fjx2,tx);
1672             fjy2             = _mm_add_ps(fjy2,ty);
1673             fjz2             = _mm_add_ps(fjz2,tz);
1674
1675             /**************************
1676              * CALCULATE INTERACTIONS *
1677              **************************/
1678
1679             r13              = _mm_mul_ps(rsq13,rinv13);
1680
1681             /* EWALD ELECTROSTATICS */
1682
1683             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1684             ewrt             = _mm_mul_ps(r13,ewtabscale);
1685             ewitab           = _mm_cvttps_epi32(ewrt);
1686             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1687             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1688                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1689                                          &ewtabF,&ewtabFn);
1690             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1691             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
1692
1693             fscal            = felec;
1694
1695             /* Calculate temporary vectorial force */
1696             tx               = _mm_mul_ps(fscal,dx13);
1697             ty               = _mm_mul_ps(fscal,dy13);
1698             tz               = _mm_mul_ps(fscal,dz13);
1699
1700             /* Update vectorial force */
1701             fix1             = _mm_add_ps(fix1,tx);
1702             fiy1             = _mm_add_ps(fiy1,ty);
1703             fiz1             = _mm_add_ps(fiz1,tz);
1704
1705             fjx3             = _mm_add_ps(fjx3,tx);
1706             fjy3             = _mm_add_ps(fjy3,ty);
1707             fjz3             = _mm_add_ps(fjz3,tz);
1708
1709             /**************************
1710              * CALCULATE INTERACTIONS *
1711              **************************/
1712
1713             r21              = _mm_mul_ps(rsq21,rinv21);
1714
1715             /* EWALD ELECTROSTATICS */
1716
1717             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1718             ewrt             = _mm_mul_ps(r21,ewtabscale);
1719             ewitab           = _mm_cvttps_epi32(ewrt);
1720             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1721             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1722                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1723                                          &ewtabF,&ewtabFn);
1724             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1725             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1726
1727             fscal            = felec;
1728
1729             /* Calculate temporary vectorial force */
1730             tx               = _mm_mul_ps(fscal,dx21);
1731             ty               = _mm_mul_ps(fscal,dy21);
1732             tz               = _mm_mul_ps(fscal,dz21);
1733
1734             /* Update vectorial force */
1735             fix2             = _mm_add_ps(fix2,tx);
1736             fiy2             = _mm_add_ps(fiy2,ty);
1737             fiz2             = _mm_add_ps(fiz2,tz);
1738
1739             fjx1             = _mm_add_ps(fjx1,tx);
1740             fjy1             = _mm_add_ps(fjy1,ty);
1741             fjz1             = _mm_add_ps(fjz1,tz);
1742
1743             /**************************
1744              * CALCULATE INTERACTIONS *
1745              **************************/
1746
1747             r22              = _mm_mul_ps(rsq22,rinv22);
1748
1749             /* EWALD ELECTROSTATICS */
1750
1751             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1752             ewrt             = _mm_mul_ps(r22,ewtabscale);
1753             ewitab           = _mm_cvttps_epi32(ewrt);
1754             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1755             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1756                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1757                                          &ewtabF,&ewtabFn);
1758             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1759             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
1760
1761             fscal            = felec;
1762
1763             /* Calculate temporary vectorial force */
1764             tx               = _mm_mul_ps(fscal,dx22);
1765             ty               = _mm_mul_ps(fscal,dy22);
1766             tz               = _mm_mul_ps(fscal,dz22);
1767
1768             /* Update vectorial force */
1769             fix2             = _mm_add_ps(fix2,tx);
1770             fiy2             = _mm_add_ps(fiy2,ty);
1771             fiz2             = _mm_add_ps(fiz2,tz);
1772
1773             fjx2             = _mm_add_ps(fjx2,tx);
1774             fjy2             = _mm_add_ps(fjy2,ty);
1775             fjz2             = _mm_add_ps(fjz2,tz);
1776
1777             /**************************
1778              * CALCULATE INTERACTIONS *
1779              **************************/
1780
1781             r23              = _mm_mul_ps(rsq23,rinv23);
1782
1783             /* EWALD ELECTROSTATICS */
1784
1785             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1786             ewrt             = _mm_mul_ps(r23,ewtabscale);
1787             ewitab           = _mm_cvttps_epi32(ewrt);
1788             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1789             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1790                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1791                                          &ewtabF,&ewtabFn);
1792             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1793             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
1794
1795             fscal            = felec;
1796
1797             /* Calculate temporary vectorial force */
1798             tx               = _mm_mul_ps(fscal,dx23);
1799             ty               = _mm_mul_ps(fscal,dy23);
1800             tz               = _mm_mul_ps(fscal,dz23);
1801
1802             /* Update vectorial force */
1803             fix2             = _mm_add_ps(fix2,tx);
1804             fiy2             = _mm_add_ps(fiy2,ty);
1805             fiz2             = _mm_add_ps(fiz2,tz);
1806
1807             fjx3             = _mm_add_ps(fjx3,tx);
1808             fjy3             = _mm_add_ps(fjy3,ty);
1809             fjz3             = _mm_add_ps(fjz3,tz);
1810
1811             /**************************
1812              * CALCULATE INTERACTIONS *
1813              **************************/
1814
1815             r31              = _mm_mul_ps(rsq31,rinv31);
1816
1817             /* EWALD ELECTROSTATICS */
1818
1819             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1820             ewrt             = _mm_mul_ps(r31,ewtabscale);
1821             ewitab           = _mm_cvttps_epi32(ewrt);
1822             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1823             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1824                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1825                                          &ewtabF,&ewtabFn);
1826             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1827             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
1828
1829             fscal            = felec;
1830
1831             /* Calculate temporary vectorial force */
1832             tx               = _mm_mul_ps(fscal,dx31);
1833             ty               = _mm_mul_ps(fscal,dy31);
1834             tz               = _mm_mul_ps(fscal,dz31);
1835
1836             /* Update vectorial force */
1837             fix3             = _mm_add_ps(fix3,tx);
1838             fiy3             = _mm_add_ps(fiy3,ty);
1839             fiz3             = _mm_add_ps(fiz3,tz);
1840
1841             fjx1             = _mm_add_ps(fjx1,tx);
1842             fjy1             = _mm_add_ps(fjy1,ty);
1843             fjz1             = _mm_add_ps(fjz1,tz);
1844
1845             /**************************
1846              * CALCULATE INTERACTIONS *
1847              **************************/
1848
1849             r32              = _mm_mul_ps(rsq32,rinv32);
1850
1851             /* EWALD ELECTROSTATICS */
1852
1853             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1854             ewrt             = _mm_mul_ps(r32,ewtabscale);
1855             ewitab           = _mm_cvttps_epi32(ewrt);
1856             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1857             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1858                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1859                                          &ewtabF,&ewtabFn);
1860             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1861             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
1862
1863             fscal            = felec;
1864
1865             /* Calculate temporary vectorial force */
1866             tx               = _mm_mul_ps(fscal,dx32);
1867             ty               = _mm_mul_ps(fscal,dy32);
1868             tz               = _mm_mul_ps(fscal,dz32);
1869
1870             /* Update vectorial force */
1871             fix3             = _mm_add_ps(fix3,tx);
1872             fiy3             = _mm_add_ps(fiy3,ty);
1873             fiz3             = _mm_add_ps(fiz3,tz);
1874
1875             fjx2             = _mm_add_ps(fjx2,tx);
1876             fjy2             = _mm_add_ps(fjy2,ty);
1877             fjz2             = _mm_add_ps(fjz2,tz);
1878
1879             /**************************
1880              * CALCULATE INTERACTIONS *
1881              **************************/
1882
1883             r33              = _mm_mul_ps(rsq33,rinv33);
1884
1885             /* EWALD ELECTROSTATICS */
1886
1887             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1888             ewrt             = _mm_mul_ps(r33,ewtabscale);
1889             ewitab           = _mm_cvttps_epi32(ewrt);
1890             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1891             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1892                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1893                                          &ewtabF,&ewtabFn);
1894             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1895             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
1896
1897             fscal            = felec;
1898
1899             /* Calculate temporary vectorial force */
1900             tx               = _mm_mul_ps(fscal,dx33);
1901             ty               = _mm_mul_ps(fscal,dy33);
1902             tz               = _mm_mul_ps(fscal,dz33);
1903
1904             /* Update vectorial force */
1905             fix3             = _mm_add_ps(fix3,tx);
1906             fiy3             = _mm_add_ps(fiy3,ty);
1907             fiz3             = _mm_add_ps(fiz3,tz);
1908
1909             fjx3             = _mm_add_ps(fjx3,tx);
1910             fjy3             = _mm_add_ps(fjy3,ty);
1911             fjz3             = _mm_add_ps(fjz3,tz);
1912
1913             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1914                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1915                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1916                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1917
1918             /* Inner loop uses 354 flops */
1919         }
1920
1921         if(jidx<j_index_end)
1922         {
1923
1924             /* Get j neighbor index, and coordinate index */
1925             jnrA             = jjnr[jidx];
1926             jnrB             = jjnr[jidx+1];
1927             jnrC             = jjnr[jidx+2];
1928             jnrD             = jjnr[jidx+3];
1929
1930             /* Sign of each element will be negative for non-real atoms.
1931              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1932              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1933              */
1934             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1935             jnrA       = (jnrA>=0) ? jnrA : 0;
1936             jnrB       = (jnrB>=0) ? jnrB : 0;
1937             jnrC       = (jnrC>=0) ? jnrC : 0;
1938             jnrD       = (jnrD>=0) ? jnrD : 0;
1939
1940             j_coord_offsetA  = DIM*jnrA;
1941             j_coord_offsetB  = DIM*jnrB;
1942             j_coord_offsetC  = DIM*jnrC;
1943             j_coord_offsetD  = DIM*jnrD;
1944
1945             /* load j atom coordinates */
1946             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1947                                               x+j_coord_offsetC,x+j_coord_offsetD,
1948                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1949                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1950
1951             /* Calculate displacement vector */
1952             dx00             = _mm_sub_ps(ix0,jx0);
1953             dy00             = _mm_sub_ps(iy0,jy0);
1954             dz00             = _mm_sub_ps(iz0,jz0);
1955             dx11             = _mm_sub_ps(ix1,jx1);
1956             dy11             = _mm_sub_ps(iy1,jy1);
1957             dz11             = _mm_sub_ps(iz1,jz1);
1958             dx12             = _mm_sub_ps(ix1,jx2);
1959             dy12             = _mm_sub_ps(iy1,jy2);
1960             dz12             = _mm_sub_ps(iz1,jz2);
1961             dx13             = _mm_sub_ps(ix1,jx3);
1962             dy13             = _mm_sub_ps(iy1,jy3);
1963             dz13             = _mm_sub_ps(iz1,jz3);
1964             dx21             = _mm_sub_ps(ix2,jx1);
1965             dy21             = _mm_sub_ps(iy2,jy1);
1966             dz21             = _mm_sub_ps(iz2,jz1);
1967             dx22             = _mm_sub_ps(ix2,jx2);
1968             dy22             = _mm_sub_ps(iy2,jy2);
1969             dz22             = _mm_sub_ps(iz2,jz2);
1970             dx23             = _mm_sub_ps(ix2,jx3);
1971             dy23             = _mm_sub_ps(iy2,jy3);
1972             dz23             = _mm_sub_ps(iz2,jz3);
1973             dx31             = _mm_sub_ps(ix3,jx1);
1974             dy31             = _mm_sub_ps(iy3,jy1);
1975             dz31             = _mm_sub_ps(iz3,jz1);
1976             dx32             = _mm_sub_ps(ix3,jx2);
1977             dy32             = _mm_sub_ps(iy3,jy2);
1978             dz32             = _mm_sub_ps(iz3,jz2);
1979             dx33             = _mm_sub_ps(ix3,jx3);
1980             dy33             = _mm_sub_ps(iy3,jy3);
1981             dz33             = _mm_sub_ps(iz3,jz3);
1982
1983             /* Calculate squared distance and things based on it */
1984             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1985             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1986             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1987             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1988             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1989             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1990             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1991             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1992             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1993             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1994
1995             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1996             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1997             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1998             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1999             rinv22           = gmx_mm_invsqrt_ps(rsq22);
2000             rinv23           = gmx_mm_invsqrt_ps(rsq23);
2001             rinv31           = gmx_mm_invsqrt_ps(rsq31);
2002             rinv32           = gmx_mm_invsqrt_ps(rsq32);
2003             rinv33           = gmx_mm_invsqrt_ps(rsq33);
2004
2005             rinvsq00         = gmx_mm_inv_ps(rsq00);
2006             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
2007             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
2008             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
2009             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
2010             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
2011             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
2012             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
2013             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
2014             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
2015
2016             fjx0             = _mm_setzero_ps();
2017             fjy0             = _mm_setzero_ps();
2018             fjz0             = _mm_setzero_ps();
2019             fjx1             = _mm_setzero_ps();
2020             fjy1             = _mm_setzero_ps();
2021             fjz1             = _mm_setzero_ps();
2022             fjx2             = _mm_setzero_ps();
2023             fjy2             = _mm_setzero_ps();
2024             fjz2             = _mm_setzero_ps();
2025             fjx3             = _mm_setzero_ps();
2026             fjy3             = _mm_setzero_ps();
2027             fjz3             = _mm_setzero_ps();
2028
2029             /**************************
2030              * CALCULATE INTERACTIONS *
2031              **************************/
2032
2033             /* LENNARD-JONES DISPERSION/REPULSION */
2034
2035             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
2036             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
2037
2038             fscal            = fvdw;
2039
2040             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2041
2042             /* Calculate temporary vectorial force */
2043             tx               = _mm_mul_ps(fscal,dx00);
2044             ty               = _mm_mul_ps(fscal,dy00);
2045             tz               = _mm_mul_ps(fscal,dz00);
2046
2047             /* Update vectorial force */
2048             fix0             = _mm_add_ps(fix0,tx);
2049             fiy0             = _mm_add_ps(fiy0,ty);
2050             fiz0             = _mm_add_ps(fiz0,tz);
2051
2052             fjx0             = _mm_add_ps(fjx0,tx);
2053             fjy0             = _mm_add_ps(fjy0,ty);
2054             fjz0             = _mm_add_ps(fjz0,tz);
2055
2056             /**************************
2057              * CALCULATE INTERACTIONS *
2058              **************************/
2059
2060             r11              = _mm_mul_ps(rsq11,rinv11);
2061             r11              = _mm_andnot_ps(dummy_mask,r11);
2062
2063             /* EWALD ELECTROSTATICS */
2064
2065             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2066             ewrt             = _mm_mul_ps(r11,ewtabscale);
2067             ewitab           = _mm_cvttps_epi32(ewrt);
2068             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2069             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2070                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2071                                          &ewtabF,&ewtabFn);
2072             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2073             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
2074
2075             fscal            = felec;
2076
2077             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2078
2079             /* Calculate temporary vectorial force */
2080             tx               = _mm_mul_ps(fscal,dx11);
2081             ty               = _mm_mul_ps(fscal,dy11);
2082             tz               = _mm_mul_ps(fscal,dz11);
2083
2084             /* Update vectorial force */
2085             fix1             = _mm_add_ps(fix1,tx);
2086             fiy1             = _mm_add_ps(fiy1,ty);
2087             fiz1             = _mm_add_ps(fiz1,tz);
2088
2089             fjx1             = _mm_add_ps(fjx1,tx);
2090             fjy1             = _mm_add_ps(fjy1,ty);
2091             fjz1             = _mm_add_ps(fjz1,tz);
2092
2093             /**************************
2094              * CALCULATE INTERACTIONS *
2095              **************************/
2096
2097             r12              = _mm_mul_ps(rsq12,rinv12);
2098             r12              = _mm_andnot_ps(dummy_mask,r12);
2099
2100             /* EWALD ELECTROSTATICS */
2101
2102             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2103             ewrt             = _mm_mul_ps(r12,ewtabscale);
2104             ewitab           = _mm_cvttps_epi32(ewrt);
2105             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2106             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2107                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2108                                          &ewtabF,&ewtabFn);
2109             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2110             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
2111
2112             fscal            = felec;
2113
2114             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2115
2116             /* Calculate temporary vectorial force */
2117             tx               = _mm_mul_ps(fscal,dx12);
2118             ty               = _mm_mul_ps(fscal,dy12);
2119             tz               = _mm_mul_ps(fscal,dz12);
2120
2121             /* Update vectorial force */
2122             fix1             = _mm_add_ps(fix1,tx);
2123             fiy1             = _mm_add_ps(fiy1,ty);
2124             fiz1             = _mm_add_ps(fiz1,tz);
2125
2126             fjx2             = _mm_add_ps(fjx2,tx);
2127             fjy2             = _mm_add_ps(fjy2,ty);
2128             fjz2             = _mm_add_ps(fjz2,tz);
2129
2130             /**************************
2131              * CALCULATE INTERACTIONS *
2132              **************************/
2133
2134             r13              = _mm_mul_ps(rsq13,rinv13);
2135             r13              = _mm_andnot_ps(dummy_mask,r13);
2136
2137             /* EWALD ELECTROSTATICS */
2138
2139             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2140             ewrt             = _mm_mul_ps(r13,ewtabscale);
2141             ewitab           = _mm_cvttps_epi32(ewrt);
2142             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2143             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2144                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2145                                          &ewtabF,&ewtabFn);
2146             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2147             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
2148
2149             fscal            = felec;
2150
2151             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2152
2153             /* Calculate temporary vectorial force */
2154             tx               = _mm_mul_ps(fscal,dx13);
2155             ty               = _mm_mul_ps(fscal,dy13);
2156             tz               = _mm_mul_ps(fscal,dz13);
2157
2158             /* Update vectorial force */
2159             fix1             = _mm_add_ps(fix1,tx);
2160             fiy1             = _mm_add_ps(fiy1,ty);
2161             fiz1             = _mm_add_ps(fiz1,tz);
2162
2163             fjx3             = _mm_add_ps(fjx3,tx);
2164             fjy3             = _mm_add_ps(fjy3,ty);
2165             fjz3             = _mm_add_ps(fjz3,tz);
2166
2167             /**************************
2168              * CALCULATE INTERACTIONS *
2169              **************************/
2170
2171             r21              = _mm_mul_ps(rsq21,rinv21);
2172             r21              = _mm_andnot_ps(dummy_mask,r21);
2173
2174             /* EWALD ELECTROSTATICS */
2175
2176             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2177             ewrt             = _mm_mul_ps(r21,ewtabscale);
2178             ewitab           = _mm_cvttps_epi32(ewrt);
2179             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2180             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2181                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2182                                          &ewtabF,&ewtabFn);
2183             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2184             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
2185
2186             fscal            = felec;
2187
2188             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2189
2190             /* Calculate temporary vectorial force */
2191             tx               = _mm_mul_ps(fscal,dx21);
2192             ty               = _mm_mul_ps(fscal,dy21);
2193             tz               = _mm_mul_ps(fscal,dz21);
2194
2195             /* Update vectorial force */
2196             fix2             = _mm_add_ps(fix2,tx);
2197             fiy2             = _mm_add_ps(fiy2,ty);
2198             fiz2             = _mm_add_ps(fiz2,tz);
2199
2200             fjx1             = _mm_add_ps(fjx1,tx);
2201             fjy1             = _mm_add_ps(fjy1,ty);
2202             fjz1             = _mm_add_ps(fjz1,tz);
2203
2204             /**************************
2205              * CALCULATE INTERACTIONS *
2206              **************************/
2207
2208             r22              = _mm_mul_ps(rsq22,rinv22);
2209             r22              = _mm_andnot_ps(dummy_mask,r22);
2210
2211             /* EWALD ELECTROSTATICS */
2212
2213             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2214             ewrt             = _mm_mul_ps(r22,ewtabscale);
2215             ewitab           = _mm_cvttps_epi32(ewrt);
2216             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2217             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2218                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2219                                          &ewtabF,&ewtabFn);
2220             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2221             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
2222
2223             fscal            = felec;
2224
2225             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2226
2227             /* Calculate temporary vectorial force */
2228             tx               = _mm_mul_ps(fscal,dx22);
2229             ty               = _mm_mul_ps(fscal,dy22);
2230             tz               = _mm_mul_ps(fscal,dz22);
2231
2232             /* Update vectorial force */
2233             fix2             = _mm_add_ps(fix2,tx);
2234             fiy2             = _mm_add_ps(fiy2,ty);
2235             fiz2             = _mm_add_ps(fiz2,tz);
2236
2237             fjx2             = _mm_add_ps(fjx2,tx);
2238             fjy2             = _mm_add_ps(fjy2,ty);
2239             fjz2             = _mm_add_ps(fjz2,tz);
2240
2241             /**************************
2242              * CALCULATE INTERACTIONS *
2243              **************************/
2244
2245             r23              = _mm_mul_ps(rsq23,rinv23);
2246             r23              = _mm_andnot_ps(dummy_mask,r23);
2247
2248             /* EWALD ELECTROSTATICS */
2249
2250             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2251             ewrt             = _mm_mul_ps(r23,ewtabscale);
2252             ewitab           = _mm_cvttps_epi32(ewrt);
2253             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2254             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2255                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2256                                          &ewtabF,&ewtabFn);
2257             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2258             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
2259
2260             fscal            = felec;
2261
2262             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2263
2264             /* Calculate temporary vectorial force */
2265             tx               = _mm_mul_ps(fscal,dx23);
2266             ty               = _mm_mul_ps(fscal,dy23);
2267             tz               = _mm_mul_ps(fscal,dz23);
2268
2269             /* Update vectorial force */
2270             fix2             = _mm_add_ps(fix2,tx);
2271             fiy2             = _mm_add_ps(fiy2,ty);
2272             fiz2             = _mm_add_ps(fiz2,tz);
2273
2274             fjx3             = _mm_add_ps(fjx3,tx);
2275             fjy3             = _mm_add_ps(fjy3,ty);
2276             fjz3             = _mm_add_ps(fjz3,tz);
2277
2278             /**************************
2279              * CALCULATE INTERACTIONS *
2280              **************************/
2281
2282             r31              = _mm_mul_ps(rsq31,rinv31);
2283             r31              = _mm_andnot_ps(dummy_mask,r31);
2284
2285             /* EWALD ELECTROSTATICS */
2286
2287             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2288             ewrt             = _mm_mul_ps(r31,ewtabscale);
2289             ewitab           = _mm_cvttps_epi32(ewrt);
2290             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2291             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2292                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2293                                          &ewtabF,&ewtabFn);
2294             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2295             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
2296
2297             fscal            = felec;
2298
2299             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2300
2301             /* Calculate temporary vectorial force */
2302             tx               = _mm_mul_ps(fscal,dx31);
2303             ty               = _mm_mul_ps(fscal,dy31);
2304             tz               = _mm_mul_ps(fscal,dz31);
2305
2306             /* Update vectorial force */
2307             fix3             = _mm_add_ps(fix3,tx);
2308             fiy3             = _mm_add_ps(fiy3,ty);
2309             fiz3             = _mm_add_ps(fiz3,tz);
2310
2311             fjx1             = _mm_add_ps(fjx1,tx);
2312             fjy1             = _mm_add_ps(fjy1,ty);
2313             fjz1             = _mm_add_ps(fjz1,tz);
2314
2315             /**************************
2316              * CALCULATE INTERACTIONS *
2317              **************************/
2318
2319             r32              = _mm_mul_ps(rsq32,rinv32);
2320             r32              = _mm_andnot_ps(dummy_mask,r32);
2321
2322             /* EWALD ELECTROSTATICS */
2323
2324             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2325             ewrt             = _mm_mul_ps(r32,ewtabscale);
2326             ewitab           = _mm_cvttps_epi32(ewrt);
2327             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2328             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2329                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2330                                          &ewtabF,&ewtabFn);
2331             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2332             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
2333
2334             fscal            = felec;
2335
2336             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2337
2338             /* Calculate temporary vectorial force */
2339             tx               = _mm_mul_ps(fscal,dx32);
2340             ty               = _mm_mul_ps(fscal,dy32);
2341             tz               = _mm_mul_ps(fscal,dz32);
2342
2343             /* Update vectorial force */
2344             fix3             = _mm_add_ps(fix3,tx);
2345             fiy3             = _mm_add_ps(fiy3,ty);
2346             fiz3             = _mm_add_ps(fiz3,tz);
2347
2348             fjx2             = _mm_add_ps(fjx2,tx);
2349             fjy2             = _mm_add_ps(fjy2,ty);
2350             fjz2             = _mm_add_ps(fjz2,tz);
2351
2352             /**************************
2353              * CALCULATE INTERACTIONS *
2354              **************************/
2355
2356             r33              = _mm_mul_ps(rsq33,rinv33);
2357             r33              = _mm_andnot_ps(dummy_mask,r33);
2358
2359             /* EWALD ELECTROSTATICS */
2360
2361             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2362             ewrt             = _mm_mul_ps(r33,ewtabscale);
2363             ewitab           = _mm_cvttps_epi32(ewrt);
2364             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2365             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2366                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2367                                          &ewtabF,&ewtabFn);
2368             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2369             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
2370
2371             fscal            = felec;
2372
2373             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2374
2375             /* Calculate temporary vectorial force */
2376             tx               = _mm_mul_ps(fscal,dx33);
2377             ty               = _mm_mul_ps(fscal,dy33);
2378             tz               = _mm_mul_ps(fscal,dz33);
2379
2380             /* Update vectorial force */
2381             fix3             = _mm_add_ps(fix3,tx);
2382             fiy3             = _mm_add_ps(fiy3,ty);
2383             fiz3             = _mm_add_ps(fiz3,tz);
2384
2385             fjx3             = _mm_add_ps(fjx3,tx);
2386             fjy3             = _mm_add_ps(fjy3,ty);
2387             fjz3             = _mm_add_ps(fjz3,tz);
2388
2389             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
2390                                                    f+j_coord_offsetC,f+j_coord_offsetD,
2391                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2392                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2393
2394             /* Inner loop uses 363 flops */
2395         }
2396
2397         /* End of innermost loop */
2398
2399         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2400                                               f+i_coord_offset,fshift+i_shift_offset);
2401
2402         /* Increment number of inner iterations */
2403         inneriter                  += j_index_end - j_index_start;
2404
2405         /* Outer loop uses 36 flops */
2406     }
2407
2408     /* Increment number of outer iterations */
2409     outeriter        += nri;
2410
2411     /* Update outer/inner flops */
2412
2413     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*36 + inneriter*363);
2414 }