src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecEwSh_VdwLJSh_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_ElecEwSh_VdwLJSh_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_ElecEwSh_VdwLJSh_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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 153     rcutoff_scalar   = fr->rcoulomb;
 154     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 155     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 156
 157     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 158     rvdw             = _mm_set1_ps(fr->rvdw);
 159
 160     /* Avoid stupid compiler warnings */
 161     jnrA = jnrB = jnrC = jnrD = 0;
 162     j_coord_offsetA = 0;
 163     j_coord_offsetB = 0;
 164     j_coord_offsetC = 0;
 165     j_coord_offsetD = 0;
 166
 167     outeriter        = 0;
 168     inneriter        = 0;
 169
 170     /* Start outer loop over neighborlists */
 171     for(iidx=0; iidx<nri; iidx++)
 172     {
 173         /* Load shift vector for this list */
 174         i_shift_offset   = DIM*shiftidx[iidx];
 175         shX              = shiftvec[i_shift_offset+XX];
 176         shY              = shiftvec[i_shift_offset+YY];
 177         shZ              = shiftvec[i_shift_offset+ZZ];
 178
 179         /* Load limits for loop over neighbors */
 180         j_index_start    = jindex[iidx];
 181         j_index_end      = jindex[iidx+1];
 182
 183         /* Get outer coordinate index */
 184         inr              = iinr[iidx];
 185         i_coord_offset   = DIM*inr;
 186
 187         /* Load i particle coords and add shift vector */
 188         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 189         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 190         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 191         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 192         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 193         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 194         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 195         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 196         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 197         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 198         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 199         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 200
 201         fix0             = _mm_setzero_ps();
 202         fiy0             = _mm_setzero_ps();
 203         fiz0             = _mm_setzero_ps();
 204         fix1             = _mm_setzero_ps();
 205         fiy1             = _mm_setzero_ps();
 206         fiz1             = _mm_setzero_ps();
 207         fix2             = _mm_setzero_ps();
 208         fiy2             = _mm_setzero_ps();
 209         fiz2             = _mm_setzero_ps();
 210         fix3             = _mm_setzero_ps();
 211         fiy3             = _mm_setzero_ps();
 212         fiz3             = _mm_setzero_ps();
 213
 214         /* Reset potential sums */
 215         velecsum         = _mm_setzero_ps();
 216         vvdwsum          = _mm_setzero_ps();
 217
 218         /* Start inner kernel loop */
 219         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 220         {
 221
 222             /* Get j neighbor index, and coordinate index */
 223             jnrA             = jjnr[jidx];
 224             jnrB             = jjnr[jidx+1];
 225             jnrC             = jjnr[jidx+2];
 226             jnrD             = jjnr[jidx+3];
 227
 228             j_coord_offsetA  = DIM*jnrA;
 229             j_coord_offsetB  = DIM*jnrB;
 230             j_coord_offsetC  = DIM*jnrC;
 231             j_coord_offsetD  = DIM*jnrD;
 232
 233             /* load j atom coordinates */
 234             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 235                                               x+j_coord_offsetC,x+j_coord_offsetD,
 236                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 237                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 238
 239             /* Calculate displacement vector */
 240             dx00             = _mm_sub_ps(ix0,jx0);
 241             dy00             = _mm_sub_ps(iy0,jy0);
 242             dz00             = _mm_sub_ps(iz0,jz0);
 243             dx11             = _mm_sub_ps(ix1,jx1);
 244             dy11             = _mm_sub_ps(iy1,jy1);
 245             dz11             = _mm_sub_ps(iz1,jz1);
 246             dx12             = _mm_sub_ps(ix1,jx2);
 247             dy12             = _mm_sub_ps(iy1,jy2);
 248             dz12             = _mm_sub_ps(iz1,jz2);
 249             dx13             = _mm_sub_ps(ix1,jx3);
 250             dy13             = _mm_sub_ps(iy1,jy3);
 251             dz13             = _mm_sub_ps(iz1,jz3);
 252             dx21             = _mm_sub_ps(ix2,jx1);
 253             dy21             = _mm_sub_ps(iy2,jy1);
 254             dz21             = _mm_sub_ps(iz2,jz1);
 255             dx22             = _mm_sub_ps(ix2,jx2);
 256             dy22             = _mm_sub_ps(iy2,jy2);
 257             dz22             = _mm_sub_ps(iz2,jz2);
 258             dx23             = _mm_sub_ps(ix2,jx3);
 259             dy23             = _mm_sub_ps(iy2,jy3);
 260             dz23             = _mm_sub_ps(iz2,jz3);
 261             dx31             = _mm_sub_ps(ix3,jx1);
 262             dy31             = _mm_sub_ps(iy3,jy1);
 263             dz31             = _mm_sub_ps(iz3,jz1);
 264             dx32             = _mm_sub_ps(ix3,jx2);
 265             dy32             = _mm_sub_ps(iy3,jy2);
 266             dz32             = _mm_sub_ps(iz3,jz2);
 267             dx33             = _mm_sub_ps(ix3,jx3);
 268             dy33             = _mm_sub_ps(iy3,jy3);
 269             dz33             = _mm_sub_ps(iz3,jz3);
 270
 271             /* Calculate squared distance and things based on it */
 272             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 273             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 274             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 275             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 276             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 277             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 278             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 279             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 280             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 281             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 282
 283             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 284             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 285             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 286             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 287             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 288             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 289             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 290             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 291             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 292
 293             rinvsq00         = gmx_mm_inv_ps(rsq00);
 294             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 295             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 296             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 297             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 298             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 299             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 300             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 301             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 302             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 303
 304             fjx0             = _mm_setzero_ps();
 305             fjy0             = _mm_setzero_ps();
 306             fjz0             = _mm_setzero_ps();
 307             fjx1             = _mm_setzero_ps();
 308             fjy1             = _mm_setzero_ps();
 309             fjz1             = _mm_setzero_ps();
 310             fjx2             = _mm_setzero_ps();
 311             fjy2             = _mm_setzero_ps();
 312             fjz2             = _mm_setzero_ps();
 313             fjx3             = _mm_setzero_ps();
 314             fjy3             = _mm_setzero_ps();
 315             fjz3             = _mm_setzero_ps();
 316
 317             /**************************
 318              * CALCULATE INTERACTIONS *
 319              **************************/
 320
 321             if (gmx_mm_any_lt(rsq00,rcutoff2))
 322             {
 323
 324             /* LENNARD-JONES DISPERSION/REPULSION */
 325
 326             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 327             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 328             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 329             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 330                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 331             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 332
 333             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 334
 335             /* Update potential sum for this i atom from the interaction with this j atom. */
 336             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 337             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 338
 339             fscal            = fvdw;
 340
 341             fscal            = _mm_and_ps(fscal,cutoff_mask);
 342
 343             /* Calculate temporary vectorial force */
 344             tx               = _mm_mul_ps(fscal,dx00);
 345             ty               = _mm_mul_ps(fscal,dy00);
 346             tz               = _mm_mul_ps(fscal,dz00);
 347
 348             /* Update vectorial force */
 349             fix0             = _mm_add_ps(fix0,tx);
 350             fiy0             = _mm_add_ps(fiy0,ty);
 351             fiz0             = _mm_add_ps(fiz0,tz);
 352
 353             fjx0             = _mm_add_ps(fjx0,tx);
 354             fjy0             = _mm_add_ps(fjy0,ty);
 355             fjz0             = _mm_add_ps(fjz0,tz);
 356
 357             }
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             if (gmx_mm_any_lt(rsq11,rcutoff2))
 364             {
 365
 366             r11              = _mm_mul_ps(rsq11,rinv11);
 367
 368             /* EWALD ELECTROSTATICS */
 369
 370             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 371             ewrt             = _mm_mul_ps(r11,ewtabscale);
 372             ewitab           = _mm_cvttps_epi32(ewrt);
 373             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 374             ewitab           = _mm_slli_epi32(ewitab,2);
 375             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 376             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 377             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 378             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 379             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 380             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 381             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 382             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_sub_ps(rinv11,sh_ewald),velec));
 383             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
 384
 385             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velec            = _mm_and_ps(velec,cutoff_mask);
 389             velecsum         = _mm_add_ps(velecsum,velec);
 390
 391             fscal            = felec;
 392
 393             fscal            = _mm_and_ps(fscal,cutoff_mask);
 394
 395             /* Calculate temporary vectorial force */
 396             tx               = _mm_mul_ps(fscal,dx11);
 397             ty               = _mm_mul_ps(fscal,dy11);
 398             tz               = _mm_mul_ps(fscal,dz11);
 399
 400             /* Update vectorial force */
 401             fix1             = _mm_add_ps(fix1,tx);
 402             fiy1             = _mm_add_ps(fiy1,ty);
 403             fiz1             = _mm_add_ps(fiz1,tz);
 404
 405             fjx1             = _mm_add_ps(fjx1,tx);
 406             fjy1             = _mm_add_ps(fjy1,ty);
 407             fjz1             = _mm_add_ps(fjz1,tz);
 408
 409             }
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             if (gmx_mm_any_lt(rsq12,rcutoff2))
 416             {
 417
 418             r12              = _mm_mul_ps(rsq12,rinv12);
 419
 420             /* EWALD ELECTROSTATICS */
 421
 422             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 423             ewrt             = _mm_mul_ps(r12,ewtabscale);
 424             ewitab           = _mm_cvttps_epi32(ewrt);
 425             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 426             ewitab           = _mm_slli_epi32(ewitab,2);
 427             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 428             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 429             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 430             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 431             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 432             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 433             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 434             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_sub_ps(rinv12,sh_ewald),velec));
 435             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
 436
 437             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 438
 439             /* Update potential sum for this i atom from the interaction with this j atom. */
 440             velec            = _mm_and_ps(velec,cutoff_mask);
 441             velecsum         = _mm_add_ps(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445             fscal            = _mm_and_ps(fscal,cutoff_mask);
 446
 447             /* Calculate temporary vectorial force */
 448             tx               = _mm_mul_ps(fscal,dx12);
 449             ty               = _mm_mul_ps(fscal,dy12);
 450             tz               = _mm_mul_ps(fscal,dz12);
 451
 452             /* Update vectorial force */
 453             fix1             = _mm_add_ps(fix1,tx);
 454             fiy1             = _mm_add_ps(fiy1,ty);
 455             fiz1             = _mm_add_ps(fiz1,tz);
 456
 457             fjx2             = _mm_add_ps(fjx2,tx);
 458             fjy2             = _mm_add_ps(fjy2,ty);
 459             fjz2             = _mm_add_ps(fjz2,tz);
 460
 461             }
 462
 463             /**************************
 464              * CALCULATE INTERACTIONS *
 465              **************************/
 466
 467             if (gmx_mm_any_lt(rsq13,rcutoff2))
 468             {
 469
 470             r13              = _mm_mul_ps(rsq13,rinv13);
 471
 472             /* EWALD ELECTROSTATICS */
 473
 474             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 475             ewrt             = _mm_mul_ps(r13,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(qq13,_mm_sub_ps(_mm_sub_ps(rinv13,sh_ewald),velec));
 487             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
 488
 489             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 490
 491             /* Update potential sum for this i atom from the interaction with this j atom. */
 492             velec            = _mm_and_ps(velec,cutoff_mask);
 493             velecsum         = _mm_add_ps(velecsum,velec);
 494
 495             fscal            = felec;
 496
 497             fscal            = _mm_and_ps(fscal,cutoff_mask);
 498
 499             /* Calculate temporary vectorial force */
 500             tx               = _mm_mul_ps(fscal,dx13);
 501             ty               = _mm_mul_ps(fscal,dy13);
 502             tz               = _mm_mul_ps(fscal,dz13);
 503
 504             /* Update vectorial force */
 505             fix1             = _mm_add_ps(fix1,tx);
 506             fiy1             = _mm_add_ps(fiy1,ty);
 507             fiz1             = _mm_add_ps(fiz1,tz);
 508
 509             fjx3             = _mm_add_ps(fjx3,tx);
 510             fjy3             = _mm_add_ps(fjy3,ty);
 511             fjz3             = _mm_add_ps(fjz3,tz);
 512
 513             }
 514
 515             /**************************
 516              * CALCULATE INTERACTIONS *
 517              **************************/
 518
 519             if (gmx_mm_any_lt(rsq21,rcutoff2))
 520             {
 521
 522             r21              = _mm_mul_ps(rsq21,rinv21);
 523
 524             /* EWALD ELECTROSTATICS */
 525
 526             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 527             ewrt             = _mm_mul_ps(r21,ewtabscale);
 528             ewitab           = _mm_cvttps_epi32(ewrt);
 529             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 530             ewitab           = _mm_slli_epi32(ewitab,2);
 531             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 532             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 533             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 534             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 535             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 536             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 537             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 538             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_sub_ps(rinv21,sh_ewald),velec));
 539             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
 540
 541             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 542
 543             /* Update potential sum for this i atom from the interaction with this j atom. */
 544             velec            = _mm_and_ps(velec,cutoff_mask);
 545             velecsum         = _mm_add_ps(velecsum,velec);
 546
 547             fscal            = felec;
 548
 549             fscal            = _mm_and_ps(fscal,cutoff_mask);
 550
 551             /* Calculate temporary vectorial force */
 552             tx               = _mm_mul_ps(fscal,dx21);
 553             ty               = _mm_mul_ps(fscal,dy21);
 554             tz               = _mm_mul_ps(fscal,dz21);
 555
 556             /* Update vectorial force */
 557             fix2             = _mm_add_ps(fix2,tx);
 558             fiy2             = _mm_add_ps(fiy2,ty);
 559             fiz2             = _mm_add_ps(fiz2,tz);
 560
 561             fjx1             = _mm_add_ps(fjx1,tx);
 562             fjy1             = _mm_add_ps(fjy1,ty);
 563             fjz1             = _mm_add_ps(fjz1,tz);
 564
 565             }
 566
 567             /**************************
 568              * CALCULATE INTERACTIONS *
 569              **************************/
 570
 571             if (gmx_mm_any_lt(rsq22,rcutoff2))
 572             {
 573
 574             r22              = _mm_mul_ps(rsq22,rinv22);
 575
 576             /* EWALD ELECTROSTATICS */
 577
 578             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 579             ewrt             = _mm_mul_ps(r22,ewtabscale);
 580             ewitab           = _mm_cvttps_epi32(ewrt);
 581             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 582             ewitab           = _mm_slli_epi32(ewitab,2);
 583             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 584             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 585             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 586             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 587             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 588             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 589             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 590             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_sub_ps(rinv22,sh_ewald),velec));
 591             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
 592
 593             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 594
 595             /* Update potential sum for this i atom from the interaction with this j atom. */
 596             velec            = _mm_and_ps(velec,cutoff_mask);
 597             velecsum         = _mm_add_ps(velecsum,velec);
 598
 599             fscal            = felec;
 600
 601             fscal            = _mm_and_ps(fscal,cutoff_mask);
 602
 603             /* Calculate temporary vectorial force */
 604             tx               = _mm_mul_ps(fscal,dx22);
 605             ty               = _mm_mul_ps(fscal,dy22);
 606             tz               = _mm_mul_ps(fscal,dz22);
 607
 608             /* Update vectorial force */
 609             fix2             = _mm_add_ps(fix2,tx);
 610             fiy2             = _mm_add_ps(fiy2,ty);
 611             fiz2             = _mm_add_ps(fiz2,tz);
 612
 613             fjx2             = _mm_add_ps(fjx2,tx);
 614             fjy2             = _mm_add_ps(fjy2,ty);
 615             fjz2             = _mm_add_ps(fjz2,tz);
 616
 617             }
 618
 619             /**************************
 620              * CALCULATE INTERACTIONS *
 621              **************************/
 622
 623             if (gmx_mm_any_lt(rsq23,rcutoff2))
 624             {
 625
 626             r23              = _mm_mul_ps(rsq23,rinv23);
 627
 628             /* EWALD ELECTROSTATICS */
 629
 630             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 631             ewrt             = _mm_mul_ps(r23,ewtabscale);
 632             ewitab           = _mm_cvttps_epi32(ewrt);
 633             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 634             ewitab           = _mm_slli_epi32(ewitab,2);
 635             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 636             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 637             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 638             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 639             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 640             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 641             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 642             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_sub_ps(rinv23,sh_ewald),velec));
 643             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
 644
 645             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 646
 647             /* Update potential sum for this i atom from the interaction with this j atom. */
 648             velec            = _mm_and_ps(velec,cutoff_mask);
 649             velecsum         = _mm_add_ps(velecsum,velec);
 650
 651             fscal            = felec;
 652
 653             fscal            = _mm_and_ps(fscal,cutoff_mask);
 654
 655             /* Calculate temporary vectorial force */
 656             tx               = _mm_mul_ps(fscal,dx23);
 657             ty               = _mm_mul_ps(fscal,dy23);
 658             tz               = _mm_mul_ps(fscal,dz23);
 659
 660             /* Update vectorial force */
 661             fix2             = _mm_add_ps(fix2,tx);
 662             fiy2             = _mm_add_ps(fiy2,ty);
 663             fiz2             = _mm_add_ps(fiz2,tz);
 664
 665             fjx3             = _mm_add_ps(fjx3,tx);
 666             fjy3             = _mm_add_ps(fjy3,ty);
 667             fjz3             = _mm_add_ps(fjz3,tz);
 668
 669             }
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             if (gmx_mm_any_lt(rsq31,rcutoff2))
 676             {
 677
 678             r31              = _mm_mul_ps(rsq31,rinv31);
 679
 680             /* EWALD ELECTROSTATICS */
 681
 682             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 683             ewrt             = _mm_mul_ps(r31,ewtabscale);
 684             ewitab           = _mm_cvttps_epi32(ewrt);
 685             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 686             ewitab           = _mm_slli_epi32(ewitab,2);
 687             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 688             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 689             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 690             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 691             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 692             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 693             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 694             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_sub_ps(rinv31,sh_ewald),velec));
 695             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
 696
 697             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 698
 699             /* Update potential sum for this i atom from the interaction with this j atom. */
 700             velec            = _mm_and_ps(velec,cutoff_mask);
 701             velecsum         = _mm_add_ps(velecsum,velec);
 702
 703             fscal            = felec;
 704
 705             fscal            = _mm_and_ps(fscal,cutoff_mask);
 706
 707             /* Calculate temporary vectorial force */
 708             tx               = _mm_mul_ps(fscal,dx31);
 709             ty               = _mm_mul_ps(fscal,dy31);
 710             tz               = _mm_mul_ps(fscal,dz31);
 711
 712             /* Update vectorial force */
 713             fix3             = _mm_add_ps(fix3,tx);
 714             fiy3             = _mm_add_ps(fiy3,ty);
 715             fiz3             = _mm_add_ps(fiz3,tz);
 716
 717             fjx1             = _mm_add_ps(fjx1,tx);
 718             fjy1             = _mm_add_ps(fjy1,ty);
 719             fjz1             = _mm_add_ps(fjz1,tz);
 720
 721             }
 722
 723             /**************************
 724              * CALCULATE INTERACTIONS *
 725              **************************/
 726
 727             if (gmx_mm_any_lt(rsq32,rcutoff2))
 728             {
 729
 730             r32              = _mm_mul_ps(rsq32,rinv32);
 731
 732             /* EWALD ELECTROSTATICS */
 733
 734             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 735             ewrt             = _mm_mul_ps(r32,ewtabscale);
 736             ewitab           = _mm_cvttps_epi32(ewrt);
 737             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 738             ewitab           = _mm_slli_epi32(ewitab,2);
 739             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 740             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 741             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 742             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 743             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 744             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 745             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 746             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_sub_ps(rinv32,sh_ewald),velec));
 747             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
 748
 749             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 750
 751             /* Update potential sum for this i atom from the interaction with this j atom. */
 752             velec            = _mm_and_ps(velec,cutoff_mask);
 753             velecsum         = _mm_add_ps(velecsum,velec);
 754
 755             fscal            = felec;
 756
 757             fscal            = _mm_and_ps(fscal,cutoff_mask);
 758
 759             /* Calculate temporary vectorial force */
 760             tx               = _mm_mul_ps(fscal,dx32);
 761             ty               = _mm_mul_ps(fscal,dy32);
 762             tz               = _mm_mul_ps(fscal,dz32);
 763
 764             /* Update vectorial force */
 765             fix3             = _mm_add_ps(fix3,tx);
 766             fiy3             = _mm_add_ps(fiy3,ty);
 767             fiz3             = _mm_add_ps(fiz3,tz);
 768
 769             fjx2             = _mm_add_ps(fjx2,tx);
 770             fjy2             = _mm_add_ps(fjy2,ty);
 771             fjz2             = _mm_add_ps(fjz2,tz);
 772
 773             }
 774
 775             /**************************
 776              * CALCULATE INTERACTIONS *
 777              **************************/
 778
 779             if (gmx_mm_any_lt(rsq33,rcutoff2))
 780             {
 781
 782             r33              = _mm_mul_ps(rsq33,rinv33);
 783
 784             /* EWALD ELECTROSTATICS */
 785
 786             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 787             ewrt             = _mm_mul_ps(r33,ewtabscale);
 788             ewitab           = _mm_cvttps_epi32(ewrt);
 789             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 790             ewitab           = _mm_slli_epi32(ewitab,2);
 791             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 792             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 793             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 794             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 795             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 796             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 797             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 798             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_sub_ps(rinv33,sh_ewald),velec));
 799             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
 800
 801             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 802
 803             /* Update potential sum for this i atom from the interaction with this j atom. */
 804             velec            = _mm_and_ps(velec,cutoff_mask);
 805             velecsum         = _mm_add_ps(velecsum,velec);
 806
 807             fscal            = felec;
 808
 809             fscal            = _mm_and_ps(fscal,cutoff_mask);
 810
 811             /* Calculate temporary vectorial force */
 812             tx               = _mm_mul_ps(fscal,dx33);
 813             ty               = _mm_mul_ps(fscal,dy33);
 814             tz               = _mm_mul_ps(fscal,dz33);
 815
 816             /* Update vectorial force */
 817             fix3             = _mm_add_ps(fix3,tx);
 818             fiy3             = _mm_add_ps(fiy3,ty);
 819             fiz3             = _mm_add_ps(fiz3,tz);
 820
 821             fjx3             = _mm_add_ps(fjx3,tx);
 822             fjy3             = _mm_add_ps(fjy3,ty);
 823             fjz3             = _mm_add_ps(fjz3,tz);
 824
 825             }
 826
 827             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 828                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 829                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 830                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 831
 832             /* Inner loop uses 458 flops */
 833         }
 834
 835         if(jidx<j_index_end)
 836         {
 837
 838             /* Get j neighbor index, and coordinate index */
 839             jnrA             = jjnr[jidx];
 840             jnrB             = jjnr[jidx+1];
 841             jnrC             = jjnr[jidx+2];
 842             jnrD             = jjnr[jidx+3];
 843
 844             /* Sign of each element will be negative for non-real atoms.
 845              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 846              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 847              */
 848             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 849             jnrA       = (jnrA>=0) ? jnrA : 0;
 850             jnrB       = (jnrB>=0) ? jnrB : 0;
 851             jnrC       = (jnrC>=0) ? jnrC : 0;
 852             jnrD       = (jnrD>=0) ? jnrD : 0;
 853
 854             j_coord_offsetA  = DIM*jnrA;
 855             j_coord_offsetB  = DIM*jnrB;
 856             j_coord_offsetC  = DIM*jnrC;
 857             j_coord_offsetD  = DIM*jnrD;
 858
 859             /* load j atom coordinates */
 860             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 861                                               x+j_coord_offsetC,x+j_coord_offsetD,
 862                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 863                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 864
 865             /* Calculate displacement vector */
 866             dx00             = _mm_sub_ps(ix0,jx0);
 867             dy00             = _mm_sub_ps(iy0,jy0);
 868             dz00             = _mm_sub_ps(iz0,jz0);
 869             dx11             = _mm_sub_ps(ix1,jx1);
 870             dy11             = _mm_sub_ps(iy1,jy1);
 871             dz11             = _mm_sub_ps(iz1,jz1);
 872             dx12             = _mm_sub_ps(ix1,jx2);
 873             dy12             = _mm_sub_ps(iy1,jy2);
 874             dz12             = _mm_sub_ps(iz1,jz2);
 875             dx13             = _mm_sub_ps(ix1,jx3);
 876             dy13             = _mm_sub_ps(iy1,jy3);
 877             dz13             = _mm_sub_ps(iz1,jz3);
 878             dx21             = _mm_sub_ps(ix2,jx1);
 879             dy21             = _mm_sub_ps(iy2,jy1);
 880             dz21             = _mm_sub_ps(iz2,jz1);
 881             dx22             = _mm_sub_ps(ix2,jx2);
 882             dy22             = _mm_sub_ps(iy2,jy2);
 883             dz22             = _mm_sub_ps(iz2,jz2);
 884             dx23             = _mm_sub_ps(ix2,jx3);
 885             dy23             = _mm_sub_ps(iy2,jy3);
 886             dz23             = _mm_sub_ps(iz2,jz3);
 887             dx31             = _mm_sub_ps(ix3,jx1);
 888             dy31             = _mm_sub_ps(iy3,jy1);
 889             dz31             = _mm_sub_ps(iz3,jz1);
 890             dx32             = _mm_sub_ps(ix3,jx2);
 891             dy32             = _mm_sub_ps(iy3,jy2);
 892             dz32             = _mm_sub_ps(iz3,jz2);
 893             dx33             = _mm_sub_ps(ix3,jx3);
 894             dy33             = _mm_sub_ps(iy3,jy3);
 895             dz33             = _mm_sub_ps(iz3,jz3);
 896
 897             /* Calculate squared distance and things based on it */
 898             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 899             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 900             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 901             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 902             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 903             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 904             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 905             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 906             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 907             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 908
 909             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 910             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 911             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 912             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 913             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 914             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 915             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 916             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 917             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 918
 919             rinvsq00         = gmx_mm_inv_ps(rsq00);
 920             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 921             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 922             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 923             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 924             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 925             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 926             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 927             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 928             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 929
 930             fjx0             = _mm_setzero_ps();
 931             fjy0             = _mm_setzero_ps();
 932             fjz0             = _mm_setzero_ps();
 933             fjx1             = _mm_setzero_ps();
 934             fjy1             = _mm_setzero_ps();
 935             fjz1             = _mm_setzero_ps();
 936             fjx2             = _mm_setzero_ps();
 937             fjy2             = _mm_setzero_ps();
 938             fjz2             = _mm_setzero_ps();
 939             fjx3             = _mm_setzero_ps();
 940             fjy3             = _mm_setzero_ps();
 941             fjz3             = _mm_setzero_ps();
 942
 943             /**************************
 944              * CALCULATE INTERACTIONS *
 945              **************************/
 946
 947             if (gmx_mm_any_lt(rsq00,rcutoff2))
 948             {
 949
 950             /* LENNARD-JONES DISPERSION/REPULSION */
 951
 952             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 953             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 954             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 955             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 956                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 957             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 958
 959             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 960
 961             /* Update potential sum for this i atom from the interaction with this j atom. */
 962             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 963             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 964             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 965
 966             fscal            = fvdw;
 967
 968             fscal            = _mm_and_ps(fscal,cutoff_mask);
 969
 970             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 971
 972             /* Calculate temporary vectorial force */
 973             tx               = _mm_mul_ps(fscal,dx00);
 974             ty               = _mm_mul_ps(fscal,dy00);
 975             tz               = _mm_mul_ps(fscal,dz00);
 976
 977             /* Update vectorial force */
 978             fix0             = _mm_add_ps(fix0,tx);
 979             fiy0             = _mm_add_ps(fiy0,ty);
 980             fiz0             = _mm_add_ps(fiz0,tz);
 981
 982             fjx0             = _mm_add_ps(fjx0,tx);
 983             fjy0             = _mm_add_ps(fjy0,ty);
 984             fjz0             = _mm_add_ps(fjz0,tz);
 985
 986             }
 987
 988             /**************************
 989              * CALCULATE INTERACTIONS *
 990              **************************/
 991
 992             if (gmx_mm_any_lt(rsq11,rcutoff2))
 993             {
 994
 995             r11              = _mm_mul_ps(rsq11,rinv11);
 996             r11              = _mm_andnot_ps(dummy_mask,r11);
 997
 998             /* EWALD ELECTROSTATICS */
 999
1000             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1001             ewrt             = _mm_mul_ps(r11,ewtabscale);
1002             ewitab           = _mm_cvttps_epi32(ewrt);
1003             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1004             ewitab           = _mm_slli_epi32(ewitab,2);
1005             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1006             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1007             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1008             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1009             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1010             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1011             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1012             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_sub_ps(rinv11,sh_ewald),velec));
1013             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
1014
1015             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1016
1017             /* Update potential sum for this i atom from the interaction with this j atom. */
1018             velec            = _mm_and_ps(velec,cutoff_mask);
1019             velec            = _mm_andnot_ps(dummy_mask,velec);
1020             velecsum         = _mm_add_ps(velecsum,velec);
1021
1022             fscal            = felec;
1023
1024             fscal            = _mm_and_ps(fscal,cutoff_mask);
1025
1026             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1027
1028             /* Calculate temporary vectorial force */
1029             tx               = _mm_mul_ps(fscal,dx11);
1030             ty               = _mm_mul_ps(fscal,dy11);
1031             tz               = _mm_mul_ps(fscal,dz11);
1032
1033             /* Update vectorial force */
1034             fix1             = _mm_add_ps(fix1,tx);
1035             fiy1             = _mm_add_ps(fiy1,ty);
1036             fiz1             = _mm_add_ps(fiz1,tz);
1037
1038             fjx1             = _mm_add_ps(fjx1,tx);
1039             fjy1             = _mm_add_ps(fjy1,ty);
1040             fjz1             = _mm_add_ps(fjz1,tz);
1041
1042             }
1043
1044             /**************************
1045              * CALCULATE INTERACTIONS *
1046              **************************/
1047
1048             if (gmx_mm_any_lt(rsq12,rcutoff2))
1049             {
1050
1051             r12              = _mm_mul_ps(rsq12,rinv12);
1052             r12              = _mm_andnot_ps(dummy_mask,r12);
1053
1054             /* EWALD ELECTROSTATICS */
1055
1056             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1057             ewrt             = _mm_mul_ps(r12,ewtabscale);
1058             ewitab           = _mm_cvttps_epi32(ewrt);
1059             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1060             ewitab           = _mm_slli_epi32(ewitab,2);
1061             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1062             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1063             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1064             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1065             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1066             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1067             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1068             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_sub_ps(rinv12,sh_ewald),velec));
1069             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
1070
1071             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1072
1073             /* Update potential sum for this i atom from the interaction with this j atom. */
1074             velec            = _mm_and_ps(velec,cutoff_mask);
1075             velec            = _mm_andnot_ps(dummy_mask,velec);
1076             velecsum         = _mm_add_ps(velecsum,velec);
1077
1078             fscal            = felec;
1079
1080             fscal            = _mm_and_ps(fscal,cutoff_mask);
1081
1082             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1083
1084             /* Calculate temporary vectorial force */
1085             tx               = _mm_mul_ps(fscal,dx12);
1086             ty               = _mm_mul_ps(fscal,dy12);
1087             tz               = _mm_mul_ps(fscal,dz12);
1088
1089             /* Update vectorial force */
1090             fix1             = _mm_add_ps(fix1,tx);
1091             fiy1             = _mm_add_ps(fiy1,ty);
1092             fiz1             = _mm_add_ps(fiz1,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
1100             /**************************
1101              * CALCULATE INTERACTIONS *
1102              **************************/
1103
1104             if (gmx_mm_any_lt(rsq13,rcutoff2))
1105             {
1106
1107             r13              = _mm_mul_ps(rsq13,rinv13);
1108             r13              = _mm_andnot_ps(dummy_mask,r13);
1109
1110             /* EWALD ELECTROSTATICS */
1111
1112             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1113             ewrt             = _mm_mul_ps(r13,ewtabscale);
1114             ewitab           = _mm_cvttps_epi32(ewrt);
1115             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1116             ewitab           = _mm_slli_epi32(ewitab,2);
1117             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1118             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1119             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1120             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1121             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1122             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1123             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1124             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_sub_ps(rinv13,sh_ewald),velec));
1125             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
1126
1127             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1128
1129             /* Update potential sum for this i atom from the interaction with this j atom. */
1130             velec            = _mm_and_ps(velec,cutoff_mask);
1131             velec            = _mm_andnot_ps(dummy_mask,velec);
1132             velecsum         = _mm_add_ps(velecsum,velec);
1133
1134             fscal            = felec;
1135
1136             fscal            = _mm_and_ps(fscal,cutoff_mask);
1137
1138             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1139
1140             /* Calculate temporary vectorial force */
1141             tx               = _mm_mul_ps(fscal,dx13);
1142             ty               = _mm_mul_ps(fscal,dy13);
1143             tz               = _mm_mul_ps(fscal,dz13);
1144
1145             /* Update vectorial force */
1146             fix1             = _mm_add_ps(fix1,tx);
1147             fiy1             = _mm_add_ps(fiy1,ty);
1148             fiz1             = _mm_add_ps(fiz1,tz);
1149
1150             fjx3             = _mm_add_ps(fjx3,tx);
1151             fjy3             = _mm_add_ps(fjy3,ty);
1152             fjz3             = _mm_add_ps(fjz3,tz);
1153
1154             }
1155
1156             /**************************
1157              * CALCULATE INTERACTIONS *
1158              **************************/
1159
1160             if (gmx_mm_any_lt(rsq21,rcutoff2))
1161             {
1162
1163             r21              = _mm_mul_ps(rsq21,rinv21);
1164             r21              = _mm_andnot_ps(dummy_mask,r21);
1165
1166             /* EWALD ELECTROSTATICS */
1167
1168             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1169             ewrt             = _mm_mul_ps(r21,ewtabscale);
1170             ewitab           = _mm_cvttps_epi32(ewrt);
1171             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1172             ewitab           = _mm_slli_epi32(ewitab,2);
1173             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1174             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1175             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1176             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1177             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1178             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1179             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1180             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_sub_ps(rinv21,sh_ewald),velec));
1181             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1182
1183             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1184
1185             /* Update potential sum for this i atom from the interaction with this j atom. */
1186             velec            = _mm_and_ps(velec,cutoff_mask);
1187             velec            = _mm_andnot_ps(dummy_mask,velec);
1188             velecsum         = _mm_add_ps(velecsum,velec);
1189
1190             fscal            = felec;
1191
1192             fscal            = _mm_and_ps(fscal,cutoff_mask);
1193
1194             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1195
1196             /* Calculate temporary vectorial force */
1197             tx               = _mm_mul_ps(fscal,dx21);
1198             ty               = _mm_mul_ps(fscal,dy21);
1199             tz               = _mm_mul_ps(fscal,dz21);
1200
1201             /* Update vectorial force */
1202             fix2             = _mm_add_ps(fix2,tx);
1203             fiy2             = _mm_add_ps(fiy2,ty);
1204             fiz2             = _mm_add_ps(fiz2,tz);
1205
1206             fjx1             = _mm_add_ps(fjx1,tx);
1207             fjy1             = _mm_add_ps(fjy1,ty);
1208             fjz1             = _mm_add_ps(fjz1,tz);
1209
1210             }
1211
1212             /**************************
1213              * CALCULATE INTERACTIONS *
1214              **************************/
1215
1216             if (gmx_mm_any_lt(rsq22,rcutoff2))
1217             {
1218
1219             r22              = _mm_mul_ps(rsq22,rinv22);
1220             r22              = _mm_andnot_ps(dummy_mask,r22);
1221
1222             /* EWALD ELECTROSTATICS */
1223
1224             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1225             ewrt             = _mm_mul_ps(r22,ewtabscale);
1226             ewitab           = _mm_cvttps_epi32(ewrt);
1227             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1228             ewitab           = _mm_slli_epi32(ewitab,2);
1229             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1230             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1231             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1232             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1233             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1234             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1235             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1236             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_sub_ps(rinv22,sh_ewald),velec));
1237             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
1238
1239             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1240
1241             /* Update potential sum for this i atom from the interaction with this j atom. */
1242             velec            = _mm_and_ps(velec,cutoff_mask);
1243             velec            = _mm_andnot_ps(dummy_mask,velec);
1244             velecsum         = _mm_add_ps(velecsum,velec);
1245
1246             fscal            = felec;
1247
1248             fscal            = _mm_and_ps(fscal,cutoff_mask);
1249
1250             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1251
1252             /* Calculate temporary vectorial force */
1253             tx               = _mm_mul_ps(fscal,dx22);
1254             ty               = _mm_mul_ps(fscal,dy22);
1255             tz               = _mm_mul_ps(fscal,dz22);
1256
1257             /* Update vectorial force */
1258             fix2             = _mm_add_ps(fix2,tx);
1259             fiy2             = _mm_add_ps(fiy2,ty);
1260             fiz2             = _mm_add_ps(fiz2,tz);
1261
1262             fjx2             = _mm_add_ps(fjx2,tx);
1263             fjy2             = _mm_add_ps(fjy2,ty);
1264             fjz2             = _mm_add_ps(fjz2,tz);
1265
1266             }
1267
1268             /**************************
1269              * CALCULATE INTERACTIONS *
1270              **************************/
1271
1272             if (gmx_mm_any_lt(rsq23,rcutoff2))
1273             {
1274
1275             r23              = _mm_mul_ps(rsq23,rinv23);
1276             r23              = _mm_andnot_ps(dummy_mask,r23);
1277
1278             /* EWALD ELECTROSTATICS */
1279
1280             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1281             ewrt             = _mm_mul_ps(r23,ewtabscale);
1282             ewitab           = _mm_cvttps_epi32(ewrt);
1283             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1284             ewitab           = _mm_slli_epi32(ewitab,2);
1285             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1286             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1287             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1288             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1289             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1290             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1291             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1292             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_sub_ps(rinv23,sh_ewald),velec));
1293             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
1294
1295             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1296
1297             /* Update potential sum for this i atom from the interaction with this j atom. */
1298             velec            = _mm_and_ps(velec,cutoff_mask);
1299             velec            = _mm_andnot_ps(dummy_mask,velec);
1300             velecsum         = _mm_add_ps(velecsum,velec);
1301
1302             fscal            = felec;
1303
1304             fscal            = _mm_and_ps(fscal,cutoff_mask);
1305
1306             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1307
1308             /* Calculate temporary vectorial force */
1309             tx               = _mm_mul_ps(fscal,dx23);
1310             ty               = _mm_mul_ps(fscal,dy23);
1311             tz               = _mm_mul_ps(fscal,dz23);
1312
1313             /* Update vectorial force */
1314             fix2             = _mm_add_ps(fix2,tx);
1315             fiy2             = _mm_add_ps(fiy2,ty);
1316             fiz2             = _mm_add_ps(fiz2,tz);
1317
1318             fjx3             = _mm_add_ps(fjx3,tx);
1319             fjy3             = _mm_add_ps(fjy3,ty);
1320             fjz3             = _mm_add_ps(fjz3,tz);
1321
1322             }
1323
1324             /**************************
1325              * CALCULATE INTERACTIONS *
1326              **************************/
1327
1328             if (gmx_mm_any_lt(rsq31,rcutoff2))
1329             {
1330
1331             r31              = _mm_mul_ps(rsq31,rinv31);
1332             r31              = _mm_andnot_ps(dummy_mask,r31);
1333
1334             /* EWALD ELECTROSTATICS */
1335
1336             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1337             ewrt             = _mm_mul_ps(r31,ewtabscale);
1338             ewitab           = _mm_cvttps_epi32(ewrt);
1339             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1340             ewitab           = _mm_slli_epi32(ewitab,2);
1341             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1342             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1343             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1344             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1345             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1346             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1347             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1348             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_sub_ps(rinv31,sh_ewald),velec));
1349             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
1350
1351             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1352
1353             /* Update potential sum for this i atom from the interaction with this j atom. */
1354             velec            = _mm_and_ps(velec,cutoff_mask);
1355             velec            = _mm_andnot_ps(dummy_mask,velec);
1356             velecsum         = _mm_add_ps(velecsum,velec);
1357
1358             fscal            = felec;
1359
1360             fscal            = _mm_and_ps(fscal,cutoff_mask);
1361
1362             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1363
1364             /* Calculate temporary vectorial force */
1365             tx               = _mm_mul_ps(fscal,dx31);
1366             ty               = _mm_mul_ps(fscal,dy31);
1367             tz               = _mm_mul_ps(fscal,dz31);
1368
1369             /* Update vectorial force */
1370             fix3             = _mm_add_ps(fix3,tx);
1371             fiy3             = _mm_add_ps(fiy3,ty);
1372             fiz3             = _mm_add_ps(fiz3,tz);
1373
1374             fjx1             = _mm_add_ps(fjx1,tx);
1375             fjy1             = _mm_add_ps(fjy1,ty);
1376             fjz1             = _mm_add_ps(fjz1,tz);
1377
1378             }
1379
1380             /**************************
1381              * CALCULATE INTERACTIONS *
1382              **************************/
1383
1384             if (gmx_mm_any_lt(rsq32,rcutoff2))
1385             {
1386
1387             r32              = _mm_mul_ps(rsq32,rinv32);
1388             r32              = _mm_andnot_ps(dummy_mask,r32);
1389
1390             /* EWALD ELECTROSTATICS */
1391
1392             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1393             ewrt             = _mm_mul_ps(r32,ewtabscale);
1394             ewitab           = _mm_cvttps_epi32(ewrt);
1395             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1396             ewitab           = _mm_slli_epi32(ewitab,2);
1397             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1398             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1399             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1400             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1401             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1402             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1403             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1404             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_sub_ps(rinv32,sh_ewald),velec));
1405             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
1406
1407             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1408
1409             /* Update potential sum for this i atom from the interaction with this j atom. */
1410             velec            = _mm_and_ps(velec,cutoff_mask);
1411             velec            = _mm_andnot_ps(dummy_mask,velec);
1412             velecsum         = _mm_add_ps(velecsum,velec);
1413
1414             fscal            = felec;
1415
1416             fscal            = _mm_and_ps(fscal,cutoff_mask);
1417
1418             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1419
1420             /* Calculate temporary vectorial force */
1421             tx               = _mm_mul_ps(fscal,dx32);
1422             ty               = _mm_mul_ps(fscal,dy32);
1423             tz               = _mm_mul_ps(fscal,dz32);
1424
1425             /* Update vectorial force */
1426             fix3             = _mm_add_ps(fix3,tx);
1427             fiy3             = _mm_add_ps(fiy3,ty);
1428             fiz3             = _mm_add_ps(fiz3,tz);
1429
1430             fjx2             = _mm_add_ps(fjx2,tx);
1431             fjy2             = _mm_add_ps(fjy2,ty);
1432             fjz2             = _mm_add_ps(fjz2,tz);
1433
1434             }
1435
1436             /**************************
1437              * CALCULATE INTERACTIONS *
1438              **************************/
1439
1440             if (gmx_mm_any_lt(rsq33,rcutoff2))
1441             {
1442
1443             r33              = _mm_mul_ps(rsq33,rinv33);
1444             r33              = _mm_andnot_ps(dummy_mask,r33);
1445
1446             /* EWALD ELECTROSTATICS */
1447
1448             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1449             ewrt             = _mm_mul_ps(r33,ewtabscale);
1450             ewitab           = _mm_cvttps_epi32(ewrt);
1451             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1452             ewitab           = _mm_slli_epi32(ewitab,2);
1453             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1454             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1455             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1456             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1457             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1458             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1459             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1460             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_sub_ps(rinv33,sh_ewald),velec));
1461             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
1462
1463             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1464
1465             /* Update potential sum for this i atom from the interaction with this j atom. */
1466             velec            = _mm_and_ps(velec,cutoff_mask);
1467             velec            = _mm_andnot_ps(dummy_mask,velec);
1468             velecsum         = _mm_add_ps(velecsum,velec);
1469
1470             fscal            = felec;
1471
1472             fscal            = _mm_and_ps(fscal,cutoff_mask);
1473
1474             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1475
1476             /* Calculate temporary vectorial force */
1477             tx               = _mm_mul_ps(fscal,dx33);
1478             ty               = _mm_mul_ps(fscal,dy33);
1479             tz               = _mm_mul_ps(fscal,dz33);
1480
1481             /* Update vectorial force */
1482             fix3             = _mm_add_ps(fix3,tx);
1483             fiy3             = _mm_add_ps(fiy3,ty);
1484             fiz3             = _mm_add_ps(fiz3,tz);
1485
1486             fjx3             = _mm_add_ps(fjx3,tx);
1487             fjy3             = _mm_add_ps(fjy3,ty);
1488             fjz3             = _mm_add_ps(fjz3,tz);
1489
1490             }
1491
1492             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1493                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1494                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1495                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1496
1497             /* Inner loop uses 467 flops */
1498         }
1499
1500         /* End of innermost loop */
1501
1502         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1503                                               f+i_coord_offset,fshift+i_shift_offset);
1504
1505         ggid                        = gid[iidx];
1506         /* Update potential energies */
1507         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1508         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1509
1510         /* Increment number of inner iterations */
1511         inneriter                  += j_index_end - j_index_start;
1512
1513         /* Outer loop uses 38 flops */
1514     }
1515
1516     /* Increment number of outer iterations */
1517     outeriter        += nri;
1518
1519     /* Update outer/inner flops */
1520
1521     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*38 + inneriter*467);
1522 }
1523 /*
1524  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomW4W4_F_sse2_single
1525  * Electrostatics interaction: Ewald
1526  * VdW interaction:            LennardJones
1527  * Geometry:                   Water4-Water4
1528  * Calculate force/pot:        Force
1529  */
1530 void
1531 nb_kernel_ElecEwSh_VdwLJSh_GeomW4W4_F_sse2_single
1532                     (t_nblist * gmx_restrict                nlist,
1533                      rvec * gmx_restrict                    xx,
1534                      rvec * gmx_restrict                    ff,
1535                      t_forcerec * gmx_restrict              fr,
1536                      t_mdatoms * gmx_restrict               mdatoms,
1537                      nb_kernel_data_t * gmx_restrict        kernel_data,
1538                      t_nrnb * gmx_restrict                  nrnb)
1539 {
1540     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1541      * just 0 for non-waters.
1542      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1543      * jnr indices corresponding to data put in the four positions in the SIMD register.
1544      */
1545     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1546     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1547     int              jnrA,jnrB,jnrC,jnrD;
1548     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1549     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1550     real             shX,shY,shZ,rcutoff_scalar;
1551     real             *shiftvec,*fshift,*x,*f;
1552     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1553     int              vdwioffset0;
1554     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1555     int              vdwioffset1;
1556     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1557     int              vdwioffset2;
1558     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1559     int              vdwioffset3;
1560     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1561     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1562     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1563     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1564     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1565     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1566     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1567     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1568     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1569     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1570     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1571     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1572     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1573     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1574     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1575     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1576     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1577     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1578     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1579     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1580     real             *charge;
1581     int              nvdwtype;
1582     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1583     int              *vdwtype;
1584     real             *vdwparam;
1585     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1586     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1587     __m128i          ewitab;
1588     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1589     real             *ewtab;
1590     __m128           dummy_mask,cutoff_mask;
1591     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1592     __m128           one     = _mm_set1_ps(1.0);
1593     __m128           two     = _mm_set1_ps(2.0);
1594     x                = xx[0];
1595     f                = ff[0];
1596
1597     nri              = nlist->nri;
1598     iinr             = nlist->iinr;
1599     jindex           = nlist->jindex;
1600     jjnr             = nlist->jjnr;
1601     shiftidx         = nlist->shift;
1602     gid              = nlist->gid;
1603     shiftvec         = fr->shift_vec[0];
1604     fshift           = fr->fshift[0];
1605     facel            = _mm_set1_ps(fr->epsfac);
1606     charge           = mdatoms->chargeA;
1607     nvdwtype         = fr->ntype;
1608     vdwparam         = fr->nbfp;
1609     vdwtype          = mdatoms->typeA;
1610
1611     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1612     ewtab            = fr->ic->tabq_coul_F;
1613     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1614     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1615
1616     /* Setup water-specific parameters */
1617     inr              = nlist->iinr[0];
1618     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1619     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1620     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1621     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1622
1623     jq1              = _mm_set1_ps(charge[inr+1]);
1624     jq2              = _mm_set1_ps(charge[inr+2]);
1625     jq3              = _mm_set1_ps(charge[inr+3]);
1626     vdwjidx0A        = 2*vdwtype[inr+0];
1627     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1628     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1629     qq11             = _mm_mul_ps(iq1,jq1);
1630     qq12             = _mm_mul_ps(iq1,jq2);
1631     qq13             = _mm_mul_ps(iq1,jq3);
1632     qq21             = _mm_mul_ps(iq2,jq1);
1633     qq22             = _mm_mul_ps(iq2,jq2);
1634     qq23             = _mm_mul_ps(iq2,jq3);
1635     qq31             = _mm_mul_ps(iq3,jq1);
1636     qq32             = _mm_mul_ps(iq3,jq2);
1637     qq33             = _mm_mul_ps(iq3,jq3);
1638
1639     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1640     rcutoff_scalar   = fr->rcoulomb;
1641     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1642     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1643
1644     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1645     rvdw             = _mm_set1_ps(fr->rvdw);
1646
1647     /* Avoid stupid compiler warnings */
1648     jnrA = jnrB = jnrC = jnrD = 0;
1649     j_coord_offsetA = 0;
1650     j_coord_offsetB = 0;
1651     j_coord_offsetC = 0;
1652     j_coord_offsetD = 0;
1653
1654     outeriter        = 0;
1655     inneriter        = 0;
1656
1657     /* Start outer loop over neighborlists */
1658     for(iidx=0; iidx<nri; iidx++)
1659     {
1660         /* Load shift vector for this list */
1661         i_shift_offset   = DIM*shiftidx[iidx];
1662         shX              = shiftvec[i_shift_offset+XX];
1663         shY              = shiftvec[i_shift_offset+YY];
1664         shZ              = shiftvec[i_shift_offset+ZZ];
1665
1666         /* Load limits for loop over neighbors */
1667         j_index_start    = jindex[iidx];
1668         j_index_end      = jindex[iidx+1];
1669
1670         /* Get outer coordinate index */
1671         inr              = iinr[iidx];
1672         i_coord_offset   = DIM*inr;
1673
1674         /* Load i particle coords and add shift vector */
1675         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1676         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1677         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1678         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1679         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1680         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1681         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1682         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1683         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1684         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1685         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1686         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1687
1688         fix0             = _mm_setzero_ps();
1689         fiy0             = _mm_setzero_ps();
1690         fiz0             = _mm_setzero_ps();
1691         fix1             = _mm_setzero_ps();
1692         fiy1             = _mm_setzero_ps();
1693         fiz1             = _mm_setzero_ps();
1694         fix2             = _mm_setzero_ps();
1695         fiy2             = _mm_setzero_ps();
1696         fiz2             = _mm_setzero_ps();
1697         fix3             = _mm_setzero_ps();
1698         fiy3             = _mm_setzero_ps();
1699         fiz3             = _mm_setzero_ps();
1700
1701         /* Start inner kernel loop */
1702         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1703         {
1704
1705             /* Get j neighbor index, and coordinate index */
1706             jnrA             = jjnr[jidx];
1707             jnrB             = jjnr[jidx+1];
1708             jnrC             = jjnr[jidx+2];
1709             jnrD             = jjnr[jidx+3];
1710
1711             j_coord_offsetA  = DIM*jnrA;
1712             j_coord_offsetB  = DIM*jnrB;
1713             j_coord_offsetC  = DIM*jnrC;
1714             j_coord_offsetD  = DIM*jnrD;
1715
1716             /* load j atom coordinates */
1717             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1718                                               x+j_coord_offsetC,x+j_coord_offsetD,
1719                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1720                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1721
1722             /* Calculate displacement vector */
1723             dx00             = _mm_sub_ps(ix0,jx0);
1724             dy00             = _mm_sub_ps(iy0,jy0);
1725             dz00             = _mm_sub_ps(iz0,jz0);
1726             dx11             = _mm_sub_ps(ix1,jx1);
1727             dy11             = _mm_sub_ps(iy1,jy1);
1728             dz11             = _mm_sub_ps(iz1,jz1);
1729             dx12             = _mm_sub_ps(ix1,jx2);
1730             dy12             = _mm_sub_ps(iy1,jy2);
1731             dz12             = _mm_sub_ps(iz1,jz2);
1732             dx13             = _mm_sub_ps(ix1,jx3);
1733             dy13             = _mm_sub_ps(iy1,jy3);
1734             dz13             = _mm_sub_ps(iz1,jz3);
1735             dx21             = _mm_sub_ps(ix2,jx1);
1736             dy21             = _mm_sub_ps(iy2,jy1);
1737             dz21             = _mm_sub_ps(iz2,jz1);
1738             dx22             = _mm_sub_ps(ix2,jx2);
1739             dy22             = _mm_sub_ps(iy2,jy2);
1740             dz22             = _mm_sub_ps(iz2,jz2);
1741             dx23             = _mm_sub_ps(ix2,jx3);
1742             dy23             = _mm_sub_ps(iy2,jy3);
1743             dz23             = _mm_sub_ps(iz2,jz3);
1744             dx31             = _mm_sub_ps(ix3,jx1);
1745             dy31             = _mm_sub_ps(iy3,jy1);
1746             dz31             = _mm_sub_ps(iz3,jz1);
1747             dx32             = _mm_sub_ps(ix3,jx2);
1748             dy32             = _mm_sub_ps(iy3,jy2);
1749             dz32             = _mm_sub_ps(iz3,jz2);
1750             dx33             = _mm_sub_ps(ix3,jx3);
1751             dy33             = _mm_sub_ps(iy3,jy3);
1752             dz33             = _mm_sub_ps(iz3,jz3);
1753
1754             /* Calculate squared distance and things based on it */
1755             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1756             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1757             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1758             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1759             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1760             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1761             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1762             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1763             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1764             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1765
1766             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1767             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1768             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1769             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1770             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1771             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1772             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1773             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1774             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1775
1776             rinvsq00         = gmx_mm_inv_ps(rsq00);
1777             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1778             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1779             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1780             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1781             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1782             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1783             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1784             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1785             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1786
1787             fjx0             = _mm_setzero_ps();
1788             fjy0             = _mm_setzero_ps();
1789             fjz0             = _mm_setzero_ps();
1790             fjx1             = _mm_setzero_ps();
1791             fjy1             = _mm_setzero_ps();
1792             fjz1             = _mm_setzero_ps();
1793             fjx2             = _mm_setzero_ps();
1794             fjy2             = _mm_setzero_ps();
1795             fjz2             = _mm_setzero_ps();
1796             fjx3             = _mm_setzero_ps();
1797             fjy3             = _mm_setzero_ps();
1798             fjz3             = _mm_setzero_ps();
1799
1800             /**************************
1801              * CALCULATE INTERACTIONS *
1802              **************************/
1803
1804             if (gmx_mm_any_lt(rsq00,rcutoff2))
1805             {
1806
1807             /* LENNARD-JONES DISPERSION/REPULSION */
1808
1809             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1810             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1811
1812             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1813
1814             fscal            = fvdw;
1815
1816             fscal            = _mm_and_ps(fscal,cutoff_mask);
1817
1818             /* Calculate temporary vectorial force */
1819             tx               = _mm_mul_ps(fscal,dx00);
1820             ty               = _mm_mul_ps(fscal,dy00);
1821             tz               = _mm_mul_ps(fscal,dz00);
1822
1823             /* Update vectorial force */
1824             fix0             = _mm_add_ps(fix0,tx);
1825             fiy0             = _mm_add_ps(fiy0,ty);
1826             fiz0             = _mm_add_ps(fiz0,tz);
1827
1828             fjx0             = _mm_add_ps(fjx0,tx);
1829             fjy0             = _mm_add_ps(fjy0,ty);
1830             fjz0             = _mm_add_ps(fjz0,tz);
1831
1832             }
1833
1834             /**************************
1835              * CALCULATE INTERACTIONS *
1836              **************************/
1837
1838             if (gmx_mm_any_lt(rsq11,rcutoff2))
1839             {
1840
1841             r11              = _mm_mul_ps(rsq11,rinv11);
1842
1843             /* EWALD ELECTROSTATICS */
1844
1845             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1846             ewrt             = _mm_mul_ps(r11,ewtabscale);
1847             ewitab           = _mm_cvttps_epi32(ewrt);
1848             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1849             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1850                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1851                                          &ewtabF,&ewtabFn);
1852             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1853             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
1854
1855             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1856
1857             fscal            = felec;
1858
1859             fscal            = _mm_and_ps(fscal,cutoff_mask);
1860
1861             /* Calculate temporary vectorial force */
1862             tx               = _mm_mul_ps(fscal,dx11);
1863             ty               = _mm_mul_ps(fscal,dy11);
1864             tz               = _mm_mul_ps(fscal,dz11);
1865
1866             /* Update vectorial force */
1867             fix1             = _mm_add_ps(fix1,tx);
1868             fiy1             = _mm_add_ps(fiy1,ty);
1869             fiz1             = _mm_add_ps(fiz1,tz);
1870
1871             fjx1             = _mm_add_ps(fjx1,tx);
1872             fjy1             = _mm_add_ps(fjy1,ty);
1873             fjz1             = _mm_add_ps(fjz1,tz);
1874
1875             }
1876
1877             /**************************
1878              * CALCULATE INTERACTIONS *
1879              **************************/
1880
1881             if (gmx_mm_any_lt(rsq12,rcutoff2))
1882             {
1883
1884             r12              = _mm_mul_ps(rsq12,rinv12);
1885
1886             /* EWALD ELECTROSTATICS */
1887
1888             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1889             ewrt             = _mm_mul_ps(r12,ewtabscale);
1890             ewitab           = _mm_cvttps_epi32(ewrt);
1891             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1892             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1893                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1894                                          &ewtabF,&ewtabFn);
1895             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1896             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
1897
1898             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1899
1900             fscal            = felec;
1901
1902             fscal            = _mm_and_ps(fscal,cutoff_mask);
1903
1904             /* Calculate temporary vectorial force */
1905             tx               = _mm_mul_ps(fscal,dx12);
1906             ty               = _mm_mul_ps(fscal,dy12);
1907             tz               = _mm_mul_ps(fscal,dz12);
1908
1909             /* Update vectorial force */
1910             fix1             = _mm_add_ps(fix1,tx);
1911             fiy1             = _mm_add_ps(fiy1,ty);
1912             fiz1             = _mm_add_ps(fiz1,tz);
1913
1914             fjx2             = _mm_add_ps(fjx2,tx);
1915             fjy2             = _mm_add_ps(fjy2,ty);
1916             fjz2             = _mm_add_ps(fjz2,tz);
1917
1918             }
1919
1920             /**************************
1921              * CALCULATE INTERACTIONS *
1922              **************************/
1923
1924             if (gmx_mm_any_lt(rsq13,rcutoff2))
1925             {
1926
1927             r13              = _mm_mul_ps(rsq13,rinv13);
1928
1929             /* EWALD ELECTROSTATICS */
1930
1931             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1932             ewrt             = _mm_mul_ps(r13,ewtabscale);
1933             ewitab           = _mm_cvttps_epi32(ewrt);
1934             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1935             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1936                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1937                                          &ewtabF,&ewtabFn);
1938             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1939             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
1940
1941             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1942
1943             fscal            = felec;
1944
1945             fscal            = _mm_and_ps(fscal,cutoff_mask);
1946
1947             /* Calculate temporary vectorial force */
1948             tx               = _mm_mul_ps(fscal,dx13);
1949             ty               = _mm_mul_ps(fscal,dy13);
1950             tz               = _mm_mul_ps(fscal,dz13);
1951
1952             /* Update vectorial force */
1953             fix1             = _mm_add_ps(fix1,tx);
1954             fiy1             = _mm_add_ps(fiy1,ty);
1955             fiz1             = _mm_add_ps(fiz1,tz);
1956
1957             fjx3             = _mm_add_ps(fjx3,tx);
1958             fjy3             = _mm_add_ps(fjy3,ty);
1959             fjz3             = _mm_add_ps(fjz3,tz);
1960
1961             }
1962
1963             /**************************
1964              * CALCULATE INTERACTIONS *
1965              **************************/
1966
1967             if (gmx_mm_any_lt(rsq21,rcutoff2))
1968             {
1969
1970             r21              = _mm_mul_ps(rsq21,rinv21);
1971
1972             /* EWALD ELECTROSTATICS */
1973
1974             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1975             ewrt             = _mm_mul_ps(r21,ewtabscale);
1976             ewitab           = _mm_cvttps_epi32(ewrt);
1977             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1978             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1979                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1980                                          &ewtabF,&ewtabFn);
1981             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1982             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1983
1984             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1985
1986             fscal            = felec;
1987
1988             fscal            = _mm_and_ps(fscal,cutoff_mask);
1989
1990             /* Calculate temporary vectorial force */
1991             tx               = _mm_mul_ps(fscal,dx21);
1992             ty               = _mm_mul_ps(fscal,dy21);
1993             tz               = _mm_mul_ps(fscal,dz21);
1994
1995             /* Update vectorial force */
1996             fix2             = _mm_add_ps(fix2,tx);
1997             fiy2             = _mm_add_ps(fiy2,ty);
1998             fiz2             = _mm_add_ps(fiz2,tz);
1999
2000             fjx1             = _mm_add_ps(fjx1,tx);
2001             fjy1             = _mm_add_ps(fjy1,ty);
2002             fjz1             = _mm_add_ps(fjz1,tz);
2003
2004             }
2005
2006             /**************************
2007              * CALCULATE INTERACTIONS *
2008              **************************/
2009
2010             if (gmx_mm_any_lt(rsq22,rcutoff2))
2011             {
2012
2013             r22              = _mm_mul_ps(rsq22,rinv22);
2014
2015             /* EWALD ELECTROSTATICS */
2016
2017             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2018             ewrt             = _mm_mul_ps(r22,ewtabscale);
2019             ewitab           = _mm_cvttps_epi32(ewrt);
2020             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2021             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2022                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2023                                          &ewtabF,&ewtabFn);
2024             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2025             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
2026
2027             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2028
2029             fscal            = felec;
2030
2031             fscal            = _mm_and_ps(fscal,cutoff_mask);
2032
2033             /* Calculate temporary vectorial force */
2034             tx               = _mm_mul_ps(fscal,dx22);
2035             ty               = _mm_mul_ps(fscal,dy22);
2036             tz               = _mm_mul_ps(fscal,dz22);
2037
2038             /* Update vectorial force */
2039             fix2             = _mm_add_ps(fix2,tx);
2040             fiy2             = _mm_add_ps(fiy2,ty);
2041             fiz2             = _mm_add_ps(fiz2,tz);
2042
2043             fjx2             = _mm_add_ps(fjx2,tx);
2044             fjy2             = _mm_add_ps(fjy2,ty);
2045             fjz2             = _mm_add_ps(fjz2,tz);
2046
2047             }
2048
2049             /**************************
2050              * CALCULATE INTERACTIONS *
2051              **************************/
2052
2053             if (gmx_mm_any_lt(rsq23,rcutoff2))
2054             {
2055
2056             r23              = _mm_mul_ps(rsq23,rinv23);
2057
2058             /* EWALD ELECTROSTATICS */
2059
2060             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2061             ewrt             = _mm_mul_ps(r23,ewtabscale);
2062             ewitab           = _mm_cvttps_epi32(ewrt);
2063             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2064             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2065                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2066                                          &ewtabF,&ewtabFn);
2067             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2068             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
2069
2070             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2071
2072             fscal            = felec;
2073
2074             fscal            = _mm_and_ps(fscal,cutoff_mask);
2075
2076             /* Calculate temporary vectorial force */
2077             tx               = _mm_mul_ps(fscal,dx23);
2078             ty               = _mm_mul_ps(fscal,dy23);
2079             tz               = _mm_mul_ps(fscal,dz23);
2080
2081             /* Update vectorial force */
2082             fix2             = _mm_add_ps(fix2,tx);
2083             fiy2             = _mm_add_ps(fiy2,ty);
2084             fiz2             = _mm_add_ps(fiz2,tz);
2085
2086             fjx3             = _mm_add_ps(fjx3,tx);
2087             fjy3             = _mm_add_ps(fjy3,ty);
2088             fjz3             = _mm_add_ps(fjz3,tz);
2089
2090             }
2091
2092             /**************************
2093              * CALCULATE INTERACTIONS *
2094              **************************/
2095
2096             if (gmx_mm_any_lt(rsq31,rcutoff2))
2097             {
2098
2099             r31              = _mm_mul_ps(rsq31,rinv31);
2100
2101             /* EWALD ELECTROSTATICS */
2102
2103             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2104             ewrt             = _mm_mul_ps(r31,ewtabscale);
2105             ewitab           = _mm_cvttps_epi32(ewrt);
2106             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2107             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2108                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2109                                          &ewtabF,&ewtabFn);
2110             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2111             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
2112
2113             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2114
2115             fscal            = felec;
2116
2117             fscal            = _mm_and_ps(fscal,cutoff_mask);
2118
2119             /* Calculate temporary vectorial force */
2120             tx               = _mm_mul_ps(fscal,dx31);
2121             ty               = _mm_mul_ps(fscal,dy31);
2122             tz               = _mm_mul_ps(fscal,dz31);
2123
2124             /* Update vectorial force */
2125             fix3             = _mm_add_ps(fix3,tx);
2126             fiy3             = _mm_add_ps(fiy3,ty);
2127             fiz3             = _mm_add_ps(fiz3,tz);
2128
2129             fjx1             = _mm_add_ps(fjx1,tx);
2130             fjy1             = _mm_add_ps(fjy1,ty);
2131             fjz1             = _mm_add_ps(fjz1,tz);
2132
2133             }
2134
2135             /**************************
2136              * CALCULATE INTERACTIONS *
2137              **************************/
2138
2139             if (gmx_mm_any_lt(rsq32,rcutoff2))
2140             {
2141
2142             r32              = _mm_mul_ps(rsq32,rinv32);
2143
2144             /* EWALD ELECTROSTATICS */
2145
2146             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2147             ewrt             = _mm_mul_ps(r32,ewtabscale);
2148             ewitab           = _mm_cvttps_epi32(ewrt);
2149             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2150             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2151                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2152                                          &ewtabF,&ewtabFn);
2153             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2154             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
2155
2156             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2157
2158             fscal            = felec;
2159
2160             fscal            = _mm_and_ps(fscal,cutoff_mask);
2161
2162             /* Calculate temporary vectorial force */
2163             tx               = _mm_mul_ps(fscal,dx32);
2164             ty               = _mm_mul_ps(fscal,dy32);
2165             tz               = _mm_mul_ps(fscal,dz32);
2166
2167             /* Update vectorial force */
2168             fix3             = _mm_add_ps(fix3,tx);
2169             fiy3             = _mm_add_ps(fiy3,ty);
2170             fiz3             = _mm_add_ps(fiz3,tz);
2171
2172             fjx2             = _mm_add_ps(fjx2,tx);
2173             fjy2             = _mm_add_ps(fjy2,ty);
2174             fjz2             = _mm_add_ps(fjz2,tz);
2175
2176             }
2177
2178             /**************************
2179              * CALCULATE INTERACTIONS *
2180              **************************/
2181
2182             if (gmx_mm_any_lt(rsq33,rcutoff2))
2183             {
2184
2185             r33              = _mm_mul_ps(rsq33,rinv33);
2186
2187             /* EWALD ELECTROSTATICS */
2188
2189             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2190             ewrt             = _mm_mul_ps(r33,ewtabscale);
2191             ewitab           = _mm_cvttps_epi32(ewrt);
2192             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2193             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2194                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2195                                          &ewtabF,&ewtabFn);
2196             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2197             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
2198
2199             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2200
2201             fscal            = felec;
2202
2203             fscal            = _mm_and_ps(fscal,cutoff_mask);
2204
2205             /* Calculate temporary vectorial force */
2206             tx               = _mm_mul_ps(fscal,dx33);
2207             ty               = _mm_mul_ps(fscal,dy33);
2208             tz               = _mm_mul_ps(fscal,dz33);
2209
2210             /* Update vectorial force */
2211             fix3             = _mm_add_ps(fix3,tx);
2212             fiy3             = _mm_add_ps(fiy3,ty);
2213             fiz3             = _mm_add_ps(fiz3,tz);
2214
2215             fjx3             = _mm_add_ps(fjx3,tx);
2216             fjy3             = _mm_add_ps(fjy3,ty);
2217             fjz3             = _mm_add_ps(fjz3,tz);
2218
2219             }
2220
2221             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
2222                                                    f+j_coord_offsetC,f+j_coord_offsetD,
2223                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2224                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2225
2226             /* Inner loop uses 384 flops */
2227         }
2228
2229         if(jidx<j_index_end)
2230         {
2231
2232             /* Get j neighbor index, and coordinate index */
2233             jnrA             = jjnr[jidx];
2234             jnrB             = jjnr[jidx+1];
2235             jnrC             = jjnr[jidx+2];
2236             jnrD             = jjnr[jidx+3];
2237
2238             /* Sign of each element will be negative for non-real atoms.
2239              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
2240              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
2241              */
2242             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
2243             jnrA       = (jnrA>=0) ? jnrA : 0;
2244             jnrB       = (jnrB>=0) ? jnrB : 0;
2245             jnrC       = (jnrC>=0) ? jnrC : 0;
2246             jnrD       = (jnrD>=0) ? jnrD : 0;
2247
2248             j_coord_offsetA  = DIM*jnrA;
2249             j_coord_offsetB  = DIM*jnrB;
2250             j_coord_offsetC  = DIM*jnrC;
2251             j_coord_offsetD  = DIM*jnrD;
2252
2253             /* load j atom coordinates */
2254             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
2255                                               x+j_coord_offsetC,x+j_coord_offsetD,
2256                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
2257                                               &jy2,&jz2,&jx3,&jy3,&jz3);
2258
2259             /* Calculate displacement vector */
2260             dx00             = _mm_sub_ps(ix0,jx0);
2261             dy00             = _mm_sub_ps(iy0,jy0);
2262             dz00             = _mm_sub_ps(iz0,jz0);
2263             dx11             = _mm_sub_ps(ix1,jx1);
2264             dy11             = _mm_sub_ps(iy1,jy1);
2265             dz11             = _mm_sub_ps(iz1,jz1);
2266             dx12             = _mm_sub_ps(ix1,jx2);
2267             dy12             = _mm_sub_ps(iy1,jy2);
2268             dz12             = _mm_sub_ps(iz1,jz2);
2269             dx13             = _mm_sub_ps(ix1,jx3);
2270             dy13             = _mm_sub_ps(iy1,jy3);
2271             dz13             = _mm_sub_ps(iz1,jz3);
2272             dx21             = _mm_sub_ps(ix2,jx1);
2273             dy21             = _mm_sub_ps(iy2,jy1);
2274             dz21             = _mm_sub_ps(iz2,jz1);
2275             dx22             = _mm_sub_ps(ix2,jx2);
2276             dy22             = _mm_sub_ps(iy2,jy2);
2277             dz22             = _mm_sub_ps(iz2,jz2);
2278             dx23             = _mm_sub_ps(ix2,jx3);
2279             dy23             = _mm_sub_ps(iy2,jy3);
2280             dz23             = _mm_sub_ps(iz2,jz3);
2281             dx31             = _mm_sub_ps(ix3,jx1);
2282             dy31             = _mm_sub_ps(iy3,jy1);
2283             dz31             = _mm_sub_ps(iz3,jz1);
2284             dx32             = _mm_sub_ps(ix3,jx2);
2285             dy32             = _mm_sub_ps(iy3,jy2);
2286             dz32             = _mm_sub_ps(iz3,jz2);
2287             dx33             = _mm_sub_ps(ix3,jx3);
2288             dy33             = _mm_sub_ps(iy3,jy3);
2289             dz33             = _mm_sub_ps(iz3,jz3);
2290
2291             /* Calculate squared distance and things based on it */
2292             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
2293             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
2294             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
2295             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
2296             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
2297             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
2298             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
2299             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
2300             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
2301             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
2302
2303             rinv11           = gmx_mm_invsqrt_ps(rsq11);
2304             rinv12           = gmx_mm_invsqrt_ps(rsq12);
2305             rinv13           = gmx_mm_invsqrt_ps(rsq13);
2306             rinv21           = gmx_mm_invsqrt_ps(rsq21);
2307             rinv22           = gmx_mm_invsqrt_ps(rsq22);
2308             rinv23           = gmx_mm_invsqrt_ps(rsq23);
2309             rinv31           = gmx_mm_invsqrt_ps(rsq31);
2310             rinv32           = gmx_mm_invsqrt_ps(rsq32);
2311             rinv33           = gmx_mm_invsqrt_ps(rsq33);
2312
2313             rinvsq00         = gmx_mm_inv_ps(rsq00);
2314             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
2315             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
2316             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
2317             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
2318             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
2319             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
2320             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
2321             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
2322             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
2323
2324             fjx0             = _mm_setzero_ps();
2325             fjy0             = _mm_setzero_ps();
2326             fjz0             = _mm_setzero_ps();
2327             fjx1             = _mm_setzero_ps();
2328             fjy1             = _mm_setzero_ps();
2329             fjz1             = _mm_setzero_ps();
2330             fjx2             = _mm_setzero_ps();
2331             fjy2             = _mm_setzero_ps();
2332             fjz2             = _mm_setzero_ps();
2333             fjx3             = _mm_setzero_ps();
2334             fjy3             = _mm_setzero_ps();
2335             fjz3             = _mm_setzero_ps();
2336
2337             /**************************
2338              * CALCULATE INTERACTIONS *
2339              **************************/
2340
2341             if (gmx_mm_any_lt(rsq00,rcutoff2))
2342             {
2343
2344             /* LENNARD-JONES DISPERSION/REPULSION */
2345
2346             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
2347             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
2348
2349             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
2350
2351             fscal            = fvdw;
2352
2353             fscal            = _mm_and_ps(fscal,cutoff_mask);
2354
2355             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2356
2357             /* Calculate temporary vectorial force */
2358             tx               = _mm_mul_ps(fscal,dx00);
2359             ty               = _mm_mul_ps(fscal,dy00);
2360             tz               = _mm_mul_ps(fscal,dz00);
2361
2362             /* Update vectorial force */
2363             fix0             = _mm_add_ps(fix0,tx);
2364             fiy0             = _mm_add_ps(fiy0,ty);
2365             fiz0             = _mm_add_ps(fiz0,tz);
2366
2367             fjx0             = _mm_add_ps(fjx0,tx);
2368             fjy0             = _mm_add_ps(fjy0,ty);
2369             fjz0             = _mm_add_ps(fjz0,tz);
2370
2371             }
2372
2373             /**************************
2374              * CALCULATE INTERACTIONS *
2375              **************************/
2376
2377             if (gmx_mm_any_lt(rsq11,rcutoff2))
2378             {
2379
2380             r11              = _mm_mul_ps(rsq11,rinv11);
2381             r11              = _mm_andnot_ps(dummy_mask,r11);
2382
2383             /* EWALD ELECTROSTATICS */
2384
2385             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2386             ewrt             = _mm_mul_ps(r11,ewtabscale);
2387             ewitab           = _mm_cvttps_epi32(ewrt);
2388             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2389             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2390                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2391                                          &ewtabF,&ewtabFn);
2392             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2393             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
2394
2395             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2396
2397             fscal            = felec;
2398
2399             fscal            = _mm_and_ps(fscal,cutoff_mask);
2400
2401             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2402
2403             /* Calculate temporary vectorial force */
2404             tx               = _mm_mul_ps(fscal,dx11);
2405             ty               = _mm_mul_ps(fscal,dy11);
2406             tz               = _mm_mul_ps(fscal,dz11);
2407
2408             /* Update vectorial force */
2409             fix1             = _mm_add_ps(fix1,tx);
2410             fiy1             = _mm_add_ps(fiy1,ty);
2411             fiz1             = _mm_add_ps(fiz1,tz);
2412
2413             fjx1             = _mm_add_ps(fjx1,tx);
2414             fjy1             = _mm_add_ps(fjy1,ty);
2415             fjz1             = _mm_add_ps(fjz1,tz);
2416
2417             }
2418
2419             /**************************
2420              * CALCULATE INTERACTIONS *
2421              **************************/
2422
2423             if (gmx_mm_any_lt(rsq12,rcutoff2))
2424             {
2425
2426             r12              = _mm_mul_ps(rsq12,rinv12);
2427             r12              = _mm_andnot_ps(dummy_mask,r12);
2428
2429             /* EWALD ELECTROSTATICS */
2430
2431             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2432             ewrt             = _mm_mul_ps(r12,ewtabscale);
2433             ewitab           = _mm_cvttps_epi32(ewrt);
2434             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2435             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2436                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2437                                          &ewtabF,&ewtabFn);
2438             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2439             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
2440
2441             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2442
2443             fscal            = felec;
2444
2445             fscal            = _mm_and_ps(fscal,cutoff_mask);
2446
2447             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2448
2449             /* Calculate temporary vectorial force */
2450             tx               = _mm_mul_ps(fscal,dx12);
2451             ty               = _mm_mul_ps(fscal,dy12);
2452             tz               = _mm_mul_ps(fscal,dz12);
2453
2454             /* Update vectorial force */
2455             fix1             = _mm_add_ps(fix1,tx);
2456             fiy1             = _mm_add_ps(fiy1,ty);
2457             fiz1             = _mm_add_ps(fiz1,tz);
2458
2459             fjx2             = _mm_add_ps(fjx2,tx);
2460             fjy2             = _mm_add_ps(fjy2,ty);
2461             fjz2             = _mm_add_ps(fjz2,tz);
2462
2463             }
2464
2465             /**************************
2466              * CALCULATE INTERACTIONS *
2467              **************************/
2468
2469             if (gmx_mm_any_lt(rsq13,rcutoff2))
2470             {
2471
2472             r13              = _mm_mul_ps(rsq13,rinv13);
2473             r13              = _mm_andnot_ps(dummy_mask,r13);
2474
2475             /* EWALD ELECTROSTATICS */
2476
2477             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2478             ewrt             = _mm_mul_ps(r13,ewtabscale);
2479             ewitab           = _mm_cvttps_epi32(ewrt);
2480             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2481             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2482                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2483                                          &ewtabF,&ewtabFn);
2484             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2485             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
2486
2487             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
2488
2489             fscal            = felec;
2490
2491             fscal            = _mm_and_ps(fscal,cutoff_mask);
2492
2493             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2494
2495             /* Calculate temporary vectorial force */
2496             tx               = _mm_mul_ps(fscal,dx13);
2497             ty               = _mm_mul_ps(fscal,dy13);
2498             tz               = _mm_mul_ps(fscal,dz13);
2499
2500             /* Update vectorial force */
2501             fix1             = _mm_add_ps(fix1,tx);
2502             fiy1             = _mm_add_ps(fiy1,ty);
2503             fiz1             = _mm_add_ps(fiz1,tz);
2504
2505             fjx3             = _mm_add_ps(fjx3,tx);
2506             fjy3             = _mm_add_ps(fjy3,ty);
2507             fjz3             = _mm_add_ps(fjz3,tz);
2508
2509             }
2510
2511             /**************************
2512              * CALCULATE INTERACTIONS *
2513              **************************/
2514
2515             if (gmx_mm_any_lt(rsq21,rcutoff2))
2516             {
2517
2518             r21              = _mm_mul_ps(rsq21,rinv21);
2519             r21              = _mm_andnot_ps(dummy_mask,r21);
2520
2521             /* EWALD ELECTROSTATICS */
2522
2523             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2524             ewrt             = _mm_mul_ps(r21,ewtabscale);
2525             ewitab           = _mm_cvttps_epi32(ewrt);
2526             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2527             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2528                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2529                                          &ewtabF,&ewtabFn);
2530             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2531             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
2532
2533             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2534
2535             fscal            = felec;
2536
2537             fscal            = _mm_and_ps(fscal,cutoff_mask);
2538
2539             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2540
2541             /* Calculate temporary vectorial force */
2542             tx               = _mm_mul_ps(fscal,dx21);
2543             ty               = _mm_mul_ps(fscal,dy21);
2544             tz               = _mm_mul_ps(fscal,dz21);
2545
2546             /* Update vectorial force */
2547             fix2             = _mm_add_ps(fix2,tx);
2548             fiy2             = _mm_add_ps(fiy2,ty);
2549             fiz2             = _mm_add_ps(fiz2,tz);
2550
2551             fjx1             = _mm_add_ps(fjx1,tx);
2552             fjy1             = _mm_add_ps(fjy1,ty);
2553             fjz1             = _mm_add_ps(fjz1,tz);
2554
2555             }
2556
2557             /**************************
2558              * CALCULATE INTERACTIONS *
2559              **************************/
2560
2561             if (gmx_mm_any_lt(rsq22,rcutoff2))
2562             {
2563
2564             r22              = _mm_mul_ps(rsq22,rinv22);
2565             r22              = _mm_andnot_ps(dummy_mask,r22);
2566
2567             /* EWALD ELECTROSTATICS */
2568
2569             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2570             ewrt             = _mm_mul_ps(r22,ewtabscale);
2571             ewitab           = _mm_cvttps_epi32(ewrt);
2572             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2573             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2574                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2575                                          &ewtabF,&ewtabFn);
2576             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2577             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
2578
2579             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2580
2581             fscal            = felec;
2582
2583             fscal            = _mm_and_ps(fscal,cutoff_mask);
2584
2585             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2586
2587             /* Calculate temporary vectorial force */
2588             tx               = _mm_mul_ps(fscal,dx22);
2589             ty               = _mm_mul_ps(fscal,dy22);
2590             tz               = _mm_mul_ps(fscal,dz22);
2591
2592             /* Update vectorial force */
2593             fix2             = _mm_add_ps(fix2,tx);
2594             fiy2             = _mm_add_ps(fiy2,ty);
2595             fiz2             = _mm_add_ps(fiz2,tz);
2596
2597             fjx2             = _mm_add_ps(fjx2,tx);
2598             fjy2             = _mm_add_ps(fjy2,ty);
2599             fjz2             = _mm_add_ps(fjz2,tz);
2600
2601             }
2602
2603             /**************************
2604              * CALCULATE INTERACTIONS *
2605              **************************/
2606
2607             if (gmx_mm_any_lt(rsq23,rcutoff2))
2608             {
2609
2610             r23              = _mm_mul_ps(rsq23,rinv23);
2611             r23              = _mm_andnot_ps(dummy_mask,r23);
2612
2613             /* EWALD ELECTROSTATICS */
2614
2615             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2616             ewrt             = _mm_mul_ps(r23,ewtabscale);
2617             ewitab           = _mm_cvttps_epi32(ewrt);
2618             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2619             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2620                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2621                                          &ewtabF,&ewtabFn);
2622             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2623             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
2624
2625             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2626
2627             fscal            = felec;
2628
2629             fscal            = _mm_and_ps(fscal,cutoff_mask);
2630
2631             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2632
2633             /* Calculate temporary vectorial force */
2634             tx               = _mm_mul_ps(fscal,dx23);
2635             ty               = _mm_mul_ps(fscal,dy23);
2636             tz               = _mm_mul_ps(fscal,dz23);
2637
2638             /* Update vectorial force */
2639             fix2             = _mm_add_ps(fix2,tx);
2640             fiy2             = _mm_add_ps(fiy2,ty);
2641             fiz2             = _mm_add_ps(fiz2,tz);
2642
2643             fjx3             = _mm_add_ps(fjx3,tx);
2644             fjy3             = _mm_add_ps(fjy3,ty);
2645             fjz3             = _mm_add_ps(fjz3,tz);
2646
2647             }
2648
2649             /**************************
2650              * CALCULATE INTERACTIONS *
2651              **************************/
2652
2653             if (gmx_mm_any_lt(rsq31,rcutoff2))
2654             {
2655
2656             r31              = _mm_mul_ps(rsq31,rinv31);
2657             r31              = _mm_andnot_ps(dummy_mask,r31);
2658
2659             /* EWALD ELECTROSTATICS */
2660
2661             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2662             ewrt             = _mm_mul_ps(r31,ewtabscale);
2663             ewitab           = _mm_cvttps_epi32(ewrt);
2664             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2665             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2666                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2667                                          &ewtabF,&ewtabFn);
2668             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2669             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
2670
2671             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2672
2673             fscal            = felec;
2674
2675             fscal            = _mm_and_ps(fscal,cutoff_mask);
2676
2677             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2678
2679             /* Calculate temporary vectorial force */
2680             tx               = _mm_mul_ps(fscal,dx31);
2681             ty               = _mm_mul_ps(fscal,dy31);
2682             tz               = _mm_mul_ps(fscal,dz31);
2683
2684             /* Update vectorial force */
2685             fix3             = _mm_add_ps(fix3,tx);
2686             fiy3             = _mm_add_ps(fiy3,ty);
2687             fiz3             = _mm_add_ps(fiz3,tz);
2688
2689             fjx1             = _mm_add_ps(fjx1,tx);
2690             fjy1             = _mm_add_ps(fjy1,ty);
2691             fjz1             = _mm_add_ps(fjz1,tz);
2692
2693             }
2694
2695             /**************************
2696              * CALCULATE INTERACTIONS *
2697              **************************/
2698
2699             if (gmx_mm_any_lt(rsq32,rcutoff2))
2700             {
2701
2702             r32              = _mm_mul_ps(rsq32,rinv32);
2703             r32              = _mm_andnot_ps(dummy_mask,r32);
2704
2705             /* EWALD ELECTROSTATICS */
2706
2707             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2708             ewrt             = _mm_mul_ps(r32,ewtabscale);
2709             ewitab           = _mm_cvttps_epi32(ewrt);
2710             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2711             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2712                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2713                                          &ewtabF,&ewtabFn);
2714             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2715             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
2716
2717             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2718
2719             fscal            = felec;
2720
2721             fscal            = _mm_and_ps(fscal,cutoff_mask);
2722
2723             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2724
2725             /* Calculate temporary vectorial force */
2726             tx               = _mm_mul_ps(fscal,dx32);
2727             ty               = _mm_mul_ps(fscal,dy32);
2728             tz               = _mm_mul_ps(fscal,dz32);
2729
2730             /* Update vectorial force */
2731             fix3             = _mm_add_ps(fix3,tx);
2732             fiy3             = _mm_add_ps(fiy3,ty);
2733             fiz3             = _mm_add_ps(fiz3,tz);
2734
2735             fjx2             = _mm_add_ps(fjx2,tx);
2736             fjy2             = _mm_add_ps(fjy2,ty);
2737             fjz2             = _mm_add_ps(fjz2,tz);
2738
2739             }
2740
2741             /**************************
2742              * CALCULATE INTERACTIONS *
2743              **************************/
2744
2745             if (gmx_mm_any_lt(rsq33,rcutoff2))
2746             {
2747
2748             r33              = _mm_mul_ps(rsq33,rinv33);
2749             r33              = _mm_andnot_ps(dummy_mask,r33);
2750
2751             /* EWALD ELECTROSTATICS */
2752
2753             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2754             ewrt             = _mm_mul_ps(r33,ewtabscale);
2755             ewitab           = _mm_cvttps_epi32(ewrt);
2756             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2757             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2758                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2759                                          &ewtabF,&ewtabFn);
2760             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2761             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
2762
2763             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2764
2765             fscal            = felec;
2766
2767             fscal            = _mm_and_ps(fscal,cutoff_mask);
2768
2769             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2770
2771             /* Calculate temporary vectorial force */
2772             tx               = _mm_mul_ps(fscal,dx33);
2773             ty               = _mm_mul_ps(fscal,dy33);
2774             tz               = _mm_mul_ps(fscal,dz33);
2775
2776             /* Update vectorial force */
2777             fix3             = _mm_add_ps(fix3,tx);
2778             fiy3             = _mm_add_ps(fiy3,ty);
2779             fiz3             = _mm_add_ps(fiz3,tz);
2780
2781             fjx3             = _mm_add_ps(fjx3,tx);
2782             fjy3             = _mm_add_ps(fjy3,ty);
2783             fjz3             = _mm_add_ps(fjz3,tz);
2784
2785             }
2786
2787             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
2788                                                    f+j_coord_offsetC,f+j_coord_offsetD,
2789                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2790                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2791
2792             /* Inner loop uses 393 flops */
2793         }
2794
2795         /* End of innermost loop */
2796
2797         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2798                                               f+i_coord_offset,fshift+i_shift_offset);
2799
2800         /* Increment number of inner iterations */
2801         inneriter                  += j_index_end - j_index_start;
2802
2803         /* Outer loop uses 36 flops */
2804     }
2805
2806     /* Increment number of outer iterations */
2807     outeriter        += nri;
2808
2809     /* Update outer/inner flops */
2810
2811     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*36 + inneriter*393);
2812 }