src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRFCut_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_ElecRFCut_VdwLJSh_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_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     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     krf              = _mm_set1_ps(fr->ic->k_rf);
 118     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 119     crf              = _mm_set1_ps(fr->ic->c_rf);
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 127     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 128     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 129     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 130
 131     jq1              = _mm_set1_ps(charge[inr+1]);
 132     jq2              = _mm_set1_ps(charge[inr+2]);
 133     jq3              = _mm_set1_ps(charge[inr+3]);
 134     vdwjidx0A        = 2*vdwtype[inr+0];
 135     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 136     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 137     qq11             = _mm_mul_ps(iq1,jq1);
 138     qq12             = _mm_mul_ps(iq1,jq2);
 139     qq13             = _mm_mul_ps(iq1,jq3);
 140     qq21             = _mm_mul_ps(iq2,jq1);
 141     qq22             = _mm_mul_ps(iq2,jq2);
 142     qq23             = _mm_mul_ps(iq2,jq3);
 143     qq31             = _mm_mul_ps(iq3,jq1);
 144     qq32             = _mm_mul_ps(iq3,jq2);
 145     qq33             = _mm_mul_ps(iq3,jq3);
 146
 147     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 148     rcutoff_scalar   = fr->rcoulomb;
 149     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 150     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 151
 152     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 153     rvdw             = _mm_set1_ps(fr->rvdw);
 154
 155     /* Avoid stupid compiler warnings */
 156     jnrA = jnrB = jnrC = jnrD = 0;
 157     j_coord_offsetA = 0;
 158     j_coord_offsetB = 0;
 159     j_coord_offsetC = 0;
 160     j_coord_offsetD = 0;
 161
 162     outeriter        = 0;
 163     inneriter        = 0;
 164
 165     /* Start outer loop over neighborlists */
 166     for(iidx=0; iidx<nri; iidx++)
 167     {
 168         /* Load shift vector for this list */
 169         i_shift_offset   = DIM*shiftidx[iidx];
 170         shX              = shiftvec[i_shift_offset+XX];
 171         shY              = shiftvec[i_shift_offset+YY];
 172         shZ              = shiftvec[i_shift_offset+ZZ];
 173
 174         /* Load limits for loop over neighbors */
 175         j_index_start    = jindex[iidx];
 176         j_index_end      = jindex[iidx+1];
 177
 178         /* Get outer coordinate index */
 179         inr              = iinr[iidx];
 180         i_coord_offset   = DIM*inr;
 181
 182         /* Load i particle coords and add shift vector */
 183         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 184         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 185         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 186         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 187         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 188         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 189         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 190         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 191         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 192         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 193         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 194         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 195
 196         fix0             = _mm_setzero_ps();
 197         fiy0             = _mm_setzero_ps();
 198         fiz0             = _mm_setzero_ps();
 199         fix1             = _mm_setzero_ps();
 200         fiy1             = _mm_setzero_ps();
 201         fiz1             = _mm_setzero_ps();
 202         fix2             = _mm_setzero_ps();
 203         fiy2             = _mm_setzero_ps();
 204         fiz2             = _mm_setzero_ps();
 205         fix3             = _mm_setzero_ps();
 206         fiy3             = _mm_setzero_ps();
 207         fiz3             = _mm_setzero_ps();
 208
 209         /* Reset potential sums */
 210         velecsum         = _mm_setzero_ps();
 211         vvdwsum          = _mm_setzero_ps();
 212
 213         /* Start inner kernel loop */
 214         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 215         {
 216
 217             /* Get j neighbor index, and coordinate index */
 218             jnrA             = jjnr[jidx];
 219             jnrB             = jjnr[jidx+1];
 220             jnrC             = jjnr[jidx+2];
 221             jnrD             = jjnr[jidx+3];
 222
 223             j_coord_offsetA  = DIM*jnrA;
 224             j_coord_offsetB  = DIM*jnrB;
 225             j_coord_offsetC  = DIM*jnrC;
 226             j_coord_offsetD  = DIM*jnrD;
 227
 228             /* load j atom coordinates */
 229             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 230                                               x+j_coord_offsetC,x+j_coord_offsetD,
 231                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 232                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 233
 234             /* Calculate displacement vector */
 235             dx00             = _mm_sub_ps(ix0,jx0);
 236             dy00             = _mm_sub_ps(iy0,jy0);
 237             dz00             = _mm_sub_ps(iz0,jz0);
 238             dx11             = _mm_sub_ps(ix1,jx1);
 239             dy11             = _mm_sub_ps(iy1,jy1);
 240             dz11             = _mm_sub_ps(iz1,jz1);
 241             dx12             = _mm_sub_ps(ix1,jx2);
 242             dy12             = _mm_sub_ps(iy1,jy2);
 243             dz12             = _mm_sub_ps(iz1,jz2);
 244             dx13             = _mm_sub_ps(ix1,jx3);
 245             dy13             = _mm_sub_ps(iy1,jy3);
 246             dz13             = _mm_sub_ps(iz1,jz3);
 247             dx21             = _mm_sub_ps(ix2,jx1);
 248             dy21             = _mm_sub_ps(iy2,jy1);
 249             dz21             = _mm_sub_ps(iz2,jz1);
 250             dx22             = _mm_sub_ps(ix2,jx2);
 251             dy22             = _mm_sub_ps(iy2,jy2);
 252             dz22             = _mm_sub_ps(iz2,jz2);
 253             dx23             = _mm_sub_ps(ix2,jx3);
 254             dy23             = _mm_sub_ps(iy2,jy3);
 255             dz23             = _mm_sub_ps(iz2,jz3);
 256             dx31             = _mm_sub_ps(ix3,jx1);
 257             dy31             = _mm_sub_ps(iy3,jy1);
 258             dz31             = _mm_sub_ps(iz3,jz1);
 259             dx32             = _mm_sub_ps(ix3,jx2);
 260             dy32             = _mm_sub_ps(iy3,jy2);
 261             dz32             = _mm_sub_ps(iz3,jz2);
 262             dx33             = _mm_sub_ps(ix3,jx3);
 263             dy33             = _mm_sub_ps(iy3,jy3);
 264             dz33             = _mm_sub_ps(iz3,jz3);
 265
 266             /* Calculate squared distance and things based on it */
 267             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 268             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 269             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 270             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 271             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 272             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 273             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 274             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 275             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 276             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 277
 278             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 279             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 280             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 281             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 282             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 283             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 284             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 285             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 286             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 287
 288             rinvsq00         = gmx_mm_inv_ps(rsq00);
 289             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 290             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 291             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 292             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 293             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 294             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 295             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 296             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 297             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 298
 299             fjx0             = _mm_setzero_ps();
 300             fjy0             = _mm_setzero_ps();
 301             fjz0             = _mm_setzero_ps();
 302             fjx1             = _mm_setzero_ps();
 303             fjy1             = _mm_setzero_ps();
 304             fjz1             = _mm_setzero_ps();
 305             fjx2             = _mm_setzero_ps();
 306             fjy2             = _mm_setzero_ps();
 307             fjz2             = _mm_setzero_ps();
 308             fjx3             = _mm_setzero_ps();
 309             fjy3             = _mm_setzero_ps();
 310             fjz3             = _mm_setzero_ps();
 311
 312             /**************************
 313              * CALCULATE INTERACTIONS *
 314              **************************/
 315
 316             if (gmx_mm_any_lt(rsq00,rcutoff2))
 317             {
 318
 319             /* LENNARD-JONES DISPERSION/REPULSION */
 320
 321             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 322             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 323             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 324             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) ,
 325                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 326             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 327
 328             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 329
 330             /* Update potential sum for this i atom from the interaction with this j atom. */
 331             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 332             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 333
 334             fscal            = fvdw;
 335
 336             fscal            = _mm_and_ps(fscal,cutoff_mask);
 337
 338             /* Calculate temporary vectorial force */
 339             tx               = _mm_mul_ps(fscal,dx00);
 340             ty               = _mm_mul_ps(fscal,dy00);
 341             tz               = _mm_mul_ps(fscal,dz00);
 342
 343             /* Update vectorial force */
 344             fix0             = _mm_add_ps(fix0,tx);
 345             fiy0             = _mm_add_ps(fiy0,ty);
 346             fiz0             = _mm_add_ps(fiz0,tz);
 347
 348             fjx0             = _mm_add_ps(fjx0,tx);
 349             fjy0             = _mm_add_ps(fjy0,ty);
 350             fjz0             = _mm_add_ps(fjz0,tz);
 351
 352             }
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             if (gmx_mm_any_lt(rsq11,rcutoff2))
 359             {
 360
 361             /* REACTION-FIELD ELECTROSTATICS */
 362             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 363             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 364
 365             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velec            = _mm_and_ps(velec,cutoff_mask);
 369             velecsum         = _mm_add_ps(velecsum,velec);
 370
 371             fscal            = felec;
 372
 373             fscal            = _mm_and_ps(fscal,cutoff_mask);
 374
 375             /* Calculate temporary vectorial force */
 376             tx               = _mm_mul_ps(fscal,dx11);
 377             ty               = _mm_mul_ps(fscal,dy11);
 378             tz               = _mm_mul_ps(fscal,dz11);
 379
 380             /* Update vectorial force */
 381             fix1             = _mm_add_ps(fix1,tx);
 382             fiy1             = _mm_add_ps(fiy1,ty);
 383             fiz1             = _mm_add_ps(fiz1,tz);
 384
 385             fjx1             = _mm_add_ps(fjx1,tx);
 386             fjy1             = _mm_add_ps(fjy1,ty);
 387             fjz1             = _mm_add_ps(fjz1,tz);
 388
 389             }
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             if (gmx_mm_any_lt(rsq12,rcutoff2))
 396             {
 397
 398             /* REACTION-FIELD ELECTROSTATICS */
 399             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 400             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 401
 402             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velec            = _mm_and_ps(velec,cutoff_mask);
 406             velecsum         = _mm_add_ps(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             fscal            = _mm_and_ps(fscal,cutoff_mask);
 411
 412             /* Calculate temporary vectorial force */
 413             tx               = _mm_mul_ps(fscal,dx12);
 414             ty               = _mm_mul_ps(fscal,dy12);
 415             tz               = _mm_mul_ps(fscal,dz12);
 416
 417             /* Update vectorial force */
 418             fix1             = _mm_add_ps(fix1,tx);
 419             fiy1             = _mm_add_ps(fiy1,ty);
 420             fiz1             = _mm_add_ps(fiz1,tz);
 421
 422             fjx2             = _mm_add_ps(fjx2,tx);
 423             fjy2             = _mm_add_ps(fjy2,ty);
 424             fjz2             = _mm_add_ps(fjz2,tz);
 425
 426             }
 427
 428             /**************************
 429              * CALCULATE INTERACTIONS *
 430              **************************/
 431
 432             if (gmx_mm_any_lt(rsq13,rcutoff2))
 433             {
 434
 435             /* REACTION-FIELD ELECTROSTATICS */
 436             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 437             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 438
 439             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 440
 441             /* Update potential sum for this i atom from the interaction with this j atom. */
 442             velec            = _mm_and_ps(velec,cutoff_mask);
 443             velecsum         = _mm_add_ps(velecsum,velec);
 444
 445             fscal            = felec;
 446
 447             fscal            = _mm_and_ps(fscal,cutoff_mask);
 448
 449             /* Calculate temporary vectorial force */
 450             tx               = _mm_mul_ps(fscal,dx13);
 451             ty               = _mm_mul_ps(fscal,dy13);
 452             tz               = _mm_mul_ps(fscal,dz13);
 453
 454             /* Update vectorial force */
 455             fix1             = _mm_add_ps(fix1,tx);
 456             fiy1             = _mm_add_ps(fiy1,ty);
 457             fiz1             = _mm_add_ps(fiz1,tz);
 458
 459             fjx3             = _mm_add_ps(fjx3,tx);
 460             fjy3             = _mm_add_ps(fjy3,ty);
 461             fjz3             = _mm_add_ps(fjz3,tz);
 462
 463             }
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             if (gmx_mm_any_lt(rsq21,rcutoff2))
 470             {
 471
 472             /* REACTION-FIELD ELECTROSTATICS */
 473             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 474             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 475
 476             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 477
 478             /* Update potential sum for this i atom from the interaction with this j atom. */
 479             velec            = _mm_and_ps(velec,cutoff_mask);
 480             velecsum         = _mm_add_ps(velecsum,velec);
 481
 482             fscal            = felec;
 483
 484             fscal            = _mm_and_ps(fscal,cutoff_mask);
 485
 486             /* Calculate temporary vectorial force */
 487             tx               = _mm_mul_ps(fscal,dx21);
 488             ty               = _mm_mul_ps(fscal,dy21);
 489             tz               = _mm_mul_ps(fscal,dz21);
 490
 491             /* Update vectorial force */
 492             fix2             = _mm_add_ps(fix2,tx);
 493             fiy2             = _mm_add_ps(fiy2,ty);
 494             fiz2             = _mm_add_ps(fiz2,tz);
 495
 496             fjx1             = _mm_add_ps(fjx1,tx);
 497             fjy1             = _mm_add_ps(fjy1,ty);
 498             fjz1             = _mm_add_ps(fjz1,tz);
 499
 500             }
 501
 502             /**************************
 503              * CALCULATE INTERACTIONS *
 504              **************************/
 505
 506             if (gmx_mm_any_lt(rsq22,rcutoff2))
 507             {
 508
 509             /* REACTION-FIELD ELECTROSTATICS */
 510             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 511             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 512
 513             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 514
 515             /* Update potential sum for this i atom from the interaction with this j atom. */
 516             velec            = _mm_and_ps(velec,cutoff_mask);
 517             velecsum         = _mm_add_ps(velecsum,velec);
 518
 519             fscal            = felec;
 520
 521             fscal            = _mm_and_ps(fscal,cutoff_mask);
 522
 523             /* Calculate temporary vectorial force */
 524             tx               = _mm_mul_ps(fscal,dx22);
 525             ty               = _mm_mul_ps(fscal,dy22);
 526             tz               = _mm_mul_ps(fscal,dz22);
 527
 528             /* Update vectorial force */
 529             fix2             = _mm_add_ps(fix2,tx);
 530             fiy2             = _mm_add_ps(fiy2,ty);
 531             fiz2             = _mm_add_ps(fiz2,tz);
 532
 533             fjx2             = _mm_add_ps(fjx2,tx);
 534             fjy2             = _mm_add_ps(fjy2,ty);
 535             fjz2             = _mm_add_ps(fjz2,tz);
 536
 537             }
 538
 539             /**************************
 540              * CALCULATE INTERACTIONS *
 541              **************************/
 542
 543             if (gmx_mm_any_lt(rsq23,rcutoff2))
 544             {
 545
 546             /* REACTION-FIELD ELECTROSTATICS */
 547             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 548             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 549
 550             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 551
 552             /* Update potential sum for this i atom from the interaction with this j atom. */
 553             velec            = _mm_and_ps(velec,cutoff_mask);
 554             velecsum         = _mm_add_ps(velecsum,velec);
 555
 556             fscal            = felec;
 557
 558             fscal            = _mm_and_ps(fscal,cutoff_mask);
 559
 560             /* Calculate temporary vectorial force */
 561             tx               = _mm_mul_ps(fscal,dx23);
 562             ty               = _mm_mul_ps(fscal,dy23);
 563             tz               = _mm_mul_ps(fscal,dz23);
 564
 565             /* Update vectorial force */
 566             fix2             = _mm_add_ps(fix2,tx);
 567             fiy2             = _mm_add_ps(fiy2,ty);
 568             fiz2             = _mm_add_ps(fiz2,tz);
 569
 570             fjx3             = _mm_add_ps(fjx3,tx);
 571             fjy3             = _mm_add_ps(fjy3,ty);
 572             fjz3             = _mm_add_ps(fjz3,tz);
 573
 574             }
 575
 576             /**************************
 577              * CALCULATE INTERACTIONS *
 578              **************************/
 579
 580             if (gmx_mm_any_lt(rsq31,rcutoff2))
 581             {
 582
 583             /* REACTION-FIELD ELECTROSTATICS */
 584             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 585             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 586
 587             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 588
 589             /* Update potential sum for this i atom from the interaction with this j atom. */
 590             velec            = _mm_and_ps(velec,cutoff_mask);
 591             velecsum         = _mm_add_ps(velecsum,velec);
 592
 593             fscal            = felec;
 594
 595             fscal            = _mm_and_ps(fscal,cutoff_mask);
 596
 597             /* Calculate temporary vectorial force */
 598             tx               = _mm_mul_ps(fscal,dx31);
 599             ty               = _mm_mul_ps(fscal,dy31);
 600             tz               = _mm_mul_ps(fscal,dz31);
 601
 602             /* Update vectorial force */
 603             fix3             = _mm_add_ps(fix3,tx);
 604             fiy3             = _mm_add_ps(fiy3,ty);
 605             fiz3             = _mm_add_ps(fiz3,tz);
 606
 607             fjx1             = _mm_add_ps(fjx1,tx);
 608             fjy1             = _mm_add_ps(fjy1,ty);
 609             fjz1             = _mm_add_ps(fjz1,tz);
 610
 611             }
 612
 613             /**************************
 614              * CALCULATE INTERACTIONS *
 615              **************************/
 616
 617             if (gmx_mm_any_lt(rsq32,rcutoff2))
 618             {
 619
 620             /* REACTION-FIELD ELECTROSTATICS */
 621             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 622             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 623
 624             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 625
 626             /* Update potential sum for this i atom from the interaction with this j atom. */
 627             velec            = _mm_and_ps(velec,cutoff_mask);
 628             velecsum         = _mm_add_ps(velecsum,velec);
 629
 630             fscal            = felec;
 631
 632             fscal            = _mm_and_ps(fscal,cutoff_mask);
 633
 634             /* Calculate temporary vectorial force */
 635             tx               = _mm_mul_ps(fscal,dx32);
 636             ty               = _mm_mul_ps(fscal,dy32);
 637             tz               = _mm_mul_ps(fscal,dz32);
 638
 639             /* Update vectorial force */
 640             fix3             = _mm_add_ps(fix3,tx);
 641             fiy3             = _mm_add_ps(fiy3,ty);
 642             fiz3             = _mm_add_ps(fiz3,tz);
 643
 644             fjx2             = _mm_add_ps(fjx2,tx);
 645             fjy2             = _mm_add_ps(fjy2,ty);
 646             fjz2             = _mm_add_ps(fjz2,tz);
 647
 648             }
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             if (gmx_mm_any_lt(rsq33,rcutoff2))
 655             {
 656
 657             /* REACTION-FIELD ELECTROSTATICS */
 658             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 659             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 660
 661             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 662
 663             /* Update potential sum for this i atom from the interaction with this j atom. */
 664             velec            = _mm_and_ps(velec,cutoff_mask);
 665             velecsum         = _mm_add_ps(velecsum,velec);
 666
 667             fscal            = felec;
 668
 669             fscal            = _mm_and_ps(fscal,cutoff_mask);
 670
 671             /* Calculate temporary vectorial force */
 672             tx               = _mm_mul_ps(fscal,dx33);
 673             ty               = _mm_mul_ps(fscal,dy33);
 674             tz               = _mm_mul_ps(fscal,dz33);
 675
 676             /* Update vectorial force */
 677             fix3             = _mm_add_ps(fix3,tx);
 678             fiy3             = _mm_add_ps(fiy3,ty);
 679             fiz3             = _mm_add_ps(fiz3,tz);
 680
 681             fjx3             = _mm_add_ps(fjx3,tx);
 682             fjy3             = _mm_add_ps(fjy3,ty);
 683             fjz3             = _mm_add_ps(fjz3,tz);
 684
 685             }
 686
 687             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 688                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 689                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 690                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 691
 692             /* Inner loop uses 368 flops */
 693         }
 694
 695         if(jidx<j_index_end)
 696         {
 697
 698             /* Get j neighbor index, and coordinate index */
 699             jnrA             = jjnr[jidx];
 700             jnrB             = jjnr[jidx+1];
 701             jnrC             = jjnr[jidx+2];
 702             jnrD             = jjnr[jidx+3];
 703
 704             /* Sign of each element will be negative for non-real atoms.
 705              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 706              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 707              */
 708             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 709             jnrA       = (jnrA>=0) ? jnrA : 0;
 710             jnrB       = (jnrB>=0) ? jnrB : 0;
 711             jnrC       = (jnrC>=0) ? jnrC : 0;
 712             jnrD       = (jnrD>=0) ? jnrD : 0;
 713
 714             j_coord_offsetA  = DIM*jnrA;
 715             j_coord_offsetB  = DIM*jnrB;
 716             j_coord_offsetC  = DIM*jnrC;
 717             j_coord_offsetD  = DIM*jnrD;
 718
 719             /* load j atom coordinates */
 720             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 721                                               x+j_coord_offsetC,x+j_coord_offsetD,
 722                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 723                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 724
 725             /* Calculate displacement vector */
 726             dx00             = _mm_sub_ps(ix0,jx0);
 727             dy00             = _mm_sub_ps(iy0,jy0);
 728             dz00             = _mm_sub_ps(iz0,jz0);
 729             dx11             = _mm_sub_ps(ix1,jx1);
 730             dy11             = _mm_sub_ps(iy1,jy1);
 731             dz11             = _mm_sub_ps(iz1,jz1);
 732             dx12             = _mm_sub_ps(ix1,jx2);
 733             dy12             = _mm_sub_ps(iy1,jy2);
 734             dz12             = _mm_sub_ps(iz1,jz2);
 735             dx13             = _mm_sub_ps(ix1,jx3);
 736             dy13             = _mm_sub_ps(iy1,jy3);
 737             dz13             = _mm_sub_ps(iz1,jz3);
 738             dx21             = _mm_sub_ps(ix2,jx1);
 739             dy21             = _mm_sub_ps(iy2,jy1);
 740             dz21             = _mm_sub_ps(iz2,jz1);
 741             dx22             = _mm_sub_ps(ix2,jx2);
 742             dy22             = _mm_sub_ps(iy2,jy2);
 743             dz22             = _mm_sub_ps(iz2,jz2);
 744             dx23             = _mm_sub_ps(ix2,jx3);
 745             dy23             = _mm_sub_ps(iy2,jy3);
 746             dz23             = _mm_sub_ps(iz2,jz3);
 747             dx31             = _mm_sub_ps(ix3,jx1);
 748             dy31             = _mm_sub_ps(iy3,jy1);
 749             dz31             = _mm_sub_ps(iz3,jz1);
 750             dx32             = _mm_sub_ps(ix3,jx2);
 751             dy32             = _mm_sub_ps(iy3,jy2);
 752             dz32             = _mm_sub_ps(iz3,jz2);
 753             dx33             = _mm_sub_ps(ix3,jx3);
 754             dy33             = _mm_sub_ps(iy3,jy3);
 755             dz33             = _mm_sub_ps(iz3,jz3);
 756
 757             /* Calculate squared distance and things based on it */
 758             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 759             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 760             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 761             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 762             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 763             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 764             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 765             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 766             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 767             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 768
 769             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 770             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 771             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 772             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 773             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 774             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 775             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 776             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 777             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 778
 779             rinvsq00         = gmx_mm_inv_ps(rsq00);
 780             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 781             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 782             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 783             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 784             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 785             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 786             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 787             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 788             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 789
 790             fjx0             = _mm_setzero_ps();
 791             fjy0             = _mm_setzero_ps();
 792             fjz0             = _mm_setzero_ps();
 793             fjx1             = _mm_setzero_ps();
 794             fjy1             = _mm_setzero_ps();
 795             fjz1             = _mm_setzero_ps();
 796             fjx2             = _mm_setzero_ps();
 797             fjy2             = _mm_setzero_ps();
 798             fjz2             = _mm_setzero_ps();
 799             fjx3             = _mm_setzero_ps();
 800             fjy3             = _mm_setzero_ps();
 801             fjz3             = _mm_setzero_ps();
 802
 803             /**************************
 804              * CALCULATE INTERACTIONS *
 805              **************************/
 806
 807             if (gmx_mm_any_lt(rsq00,rcutoff2))
 808             {
 809
 810             /* LENNARD-JONES DISPERSION/REPULSION */
 811
 812             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 813             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 814             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 815             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) ,
 816                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 817             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 818
 819             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 820
 821             /* Update potential sum for this i atom from the interaction with this j atom. */
 822             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 823             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 824             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 825
 826             fscal            = fvdw;
 827
 828             fscal            = _mm_and_ps(fscal,cutoff_mask);
 829
 830             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 831
 832             /* Calculate temporary vectorial force */
 833             tx               = _mm_mul_ps(fscal,dx00);
 834             ty               = _mm_mul_ps(fscal,dy00);
 835             tz               = _mm_mul_ps(fscal,dz00);
 836
 837             /* Update vectorial force */
 838             fix0             = _mm_add_ps(fix0,tx);
 839             fiy0             = _mm_add_ps(fiy0,ty);
 840             fiz0             = _mm_add_ps(fiz0,tz);
 841
 842             fjx0             = _mm_add_ps(fjx0,tx);
 843             fjy0             = _mm_add_ps(fjy0,ty);
 844             fjz0             = _mm_add_ps(fjz0,tz);
 845
 846             }
 847
 848             /**************************
 849              * CALCULATE INTERACTIONS *
 850              **************************/
 851
 852             if (gmx_mm_any_lt(rsq11,rcutoff2))
 853             {
 854
 855             /* REACTION-FIELD ELECTROSTATICS */
 856             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 857             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 858
 859             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 860
 861             /* Update potential sum for this i atom from the interaction with this j atom. */
 862             velec            = _mm_and_ps(velec,cutoff_mask);
 863             velec            = _mm_andnot_ps(dummy_mask,velec);
 864             velecsum         = _mm_add_ps(velecsum,velec);
 865
 866             fscal            = felec;
 867
 868             fscal            = _mm_and_ps(fscal,cutoff_mask);
 869
 870             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 871
 872             /* Calculate temporary vectorial force */
 873             tx               = _mm_mul_ps(fscal,dx11);
 874             ty               = _mm_mul_ps(fscal,dy11);
 875             tz               = _mm_mul_ps(fscal,dz11);
 876
 877             /* Update vectorial force */
 878             fix1             = _mm_add_ps(fix1,tx);
 879             fiy1             = _mm_add_ps(fiy1,ty);
 880             fiz1             = _mm_add_ps(fiz1,tz);
 881
 882             fjx1             = _mm_add_ps(fjx1,tx);
 883             fjy1             = _mm_add_ps(fjy1,ty);
 884             fjz1             = _mm_add_ps(fjz1,tz);
 885
 886             }
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             if (gmx_mm_any_lt(rsq12,rcutoff2))
 893             {
 894
 895             /* REACTION-FIELD ELECTROSTATICS */
 896             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 897             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 898
 899             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 900
 901             /* Update potential sum for this i atom from the interaction with this j atom. */
 902             velec            = _mm_and_ps(velec,cutoff_mask);
 903             velec            = _mm_andnot_ps(dummy_mask,velec);
 904             velecsum         = _mm_add_ps(velecsum,velec);
 905
 906             fscal            = felec;
 907
 908             fscal            = _mm_and_ps(fscal,cutoff_mask);
 909
 910             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 911
 912             /* Calculate temporary vectorial force */
 913             tx               = _mm_mul_ps(fscal,dx12);
 914             ty               = _mm_mul_ps(fscal,dy12);
 915             tz               = _mm_mul_ps(fscal,dz12);
 916
 917             /* Update vectorial force */
 918             fix1             = _mm_add_ps(fix1,tx);
 919             fiy1             = _mm_add_ps(fiy1,ty);
 920             fiz1             = _mm_add_ps(fiz1,tz);
 921
 922             fjx2             = _mm_add_ps(fjx2,tx);
 923             fjy2             = _mm_add_ps(fjy2,ty);
 924             fjz2             = _mm_add_ps(fjz2,tz);
 925
 926             }
 927
 928             /**************************
 929              * CALCULATE INTERACTIONS *
 930              **************************/
 931
 932             if (gmx_mm_any_lt(rsq13,rcutoff2))
 933             {
 934
 935             /* REACTION-FIELD ELECTROSTATICS */
 936             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 937             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 938
 939             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 940
 941             /* Update potential sum for this i atom from the interaction with this j atom. */
 942             velec            = _mm_and_ps(velec,cutoff_mask);
 943             velec            = _mm_andnot_ps(dummy_mask,velec);
 944             velecsum         = _mm_add_ps(velecsum,velec);
 945
 946             fscal            = felec;
 947
 948             fscal            = _mm_and_ps(fscal,cutoff_mask);
 949
 950             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 951
 952             /* Calculate temporary vectorial force */
 953             tx               = _mm_mul_ps(fscal,dx13);
 954             ty               = _mm_mul_ps(fscal,dy13);
 955             tz               = _mm_mul_ps(fscal,dz13);
 956
 957             /* Update vectorial force */
 958             fix1             = _mm_add_ps(fix1,tx);
 959             fiy1             = _mm_add_ps(fiy1,ty);
 960             fiz1             = _mm_add_ps(fiz1,tz);
 961
 962             fjx3             = _mm_add_ps(fjx3,tx);
 963             fjy3             = _mm_add_ps(fjy3,ty);
 964             fjz3             = _mm_add_ps(fjz3,tz);
 965
 966             }
 967
 968             /**************************
 969              * CALCULATE INTERACTIONS *
 970              **************************/
 971
 972             if (gmx_mm_any_lt(rsq21,rcutoff2))
 973             {
 974
 975             /* REACTION-FIELD ELECTROSTATICS */
 976             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 977             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 978
 979             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 980
 981             /* Update potential sum for this i atom from the interaction with this j atom. */
 982             velec            = _mm_and_ps(velec,cutoff_mask);
 983             velec            = _mm_andnot_ps(dummy_mask,velec);
 984             velecsum         = _mm_add_ps(velecsum,velec);
 985
 986             fscal            = felec;
 987
 988             fscal            = _mm_and_ps(fscal,cutoff_mask);
 989
 990             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 991
 992             /* Calculate temporary vectorial force */
 993             tx               = _mm_mul_ps(fscal,dx21);
 994             ty               = _mm_mul_ps(fscal,dy21);
 995             tz               = _mm_mul_ps(fscal,dz21);
 996
 997             /* Update vectorial force */
 998             fix2             = _mm_add_ps(fix2,tx);
 999             fiy2             = _mm_add_ps(fiy2,ty);
1000             fiz2             = _mm_add_ps(fiz2,tz);
1001
1002             fjx1             = _mm_add_ps(fjx1,tx);
1003             fjy1             = _mm_add_ps(fjy1,ty);
1004             fjz1             = _mm_add_ps(fjz1,tz);
1005
1006             }
1007
1008             /**************************
1009              * CALCULATE INTERACTIONS *
1010              **************************/
1011
1012             if (gmx_mm_any_lt(rsq22,rcutoff2))
1013             {
1014
1015             /* REACTION-FIELD ELECTROSTATICS */
1016             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1017             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1018
1019             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1020
1021             /* Update potential sum for this i atom from the interaction with this j atom. */
1022             velec            = _mm_and_ps(velec,cutoff_mask);
1023             velec            = _mm_andnot_ps(dummy_mask,velec);
1024             velecsum         = _mm_add_ps(velecsum,velec);
1025
1026             fscal            = felec;
1027
1028             fscal            = _mm_and_ps(fscal,cutoff_mask);
1029
1030             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1031
1032             /* Calculate temporary vectorial force */
1033             tx               = _mm_mul_ps(fscal,dx22);
1034             ty               = _mm_mul_ps(fscal,dy22);
1035             tz               = _mm_mul_ps(fscal,dz22);
1036
1037             /* Update vectorial force */
1038             fix2             = _mm_add_ps(fix2,tx);
1039             fiy2             = _mm_add_ps(fiy2,ty);
1040             fiz2             = _mm_add_ps(fiz2,tz);
1041
1042             fjx2             = _mm_add_ps(fjx2,tx);
1043             fjy2             = _mm_add_ps(fjy2,ty);
1044             fjz2             = _mm_add_ps(fjz2,tz);
1045
1046             }
1047
1048             /**************************
1049              * CALCULATE INTERACTIONS *
1050              **************************/
1051
1052             if (gmx_mm_any_lt(rsq23,rcutoff2))
1053             {
1054
1055             /* REACTION-FIELD ELECTROSTATICS */
1056             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
1057             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1058
1059             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1060
1061             /* Update potential sum for this i atom from the interaction with this j atom. */
1062             velec            = _mm_and_ps(velec,cutoff_mask);
1063             velec            = _mm_andnot_ps(dummy_mask,velec);
1064             velecsum         = _mm_add_ps(velecsum,velec);
1065
1066             fscal            = felec;
1067
1068             fscal            = _mm_and_ps(fscal,cutoff_mask);
1069
1070             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1071
1072             /* Calculate temporary vectorial force */
1073             tx               = _mm_mul_ps(fscal,dx23);
1074             ty               = _mm_mul_ps(fscal,dy23);
1075             tz               = _mm_mul_ps(fscal,dz23);
1076
1077             /* Update vectorial force */
1078             fix2             = _mm_add_ps(fix2,tx);
1079             fiy2             = _mm_add_ps(fiy2,ty);
1080             fiz2             = _mm_add_ps(fiz2,tz);
1081
1082             fjx3             = _mm_add_ps(fjx3,tx);
1083             fjy3             = _mm_add_ps(fjy3,ty);
1084             fjz3             = _mm_add_ps(fjz3,tz);
1085
1086             }
1087
1088             /**************************
1089              * CALCULATE INTERACTIONS *
1090              **************************/
1091
1092             if (gmx_mm_any_lt(rsq31,rcutoff2))
1093             {
1094
1095             /* REACTION-FIELD ELECTROSTATICS */
1096             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
1097             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1098
1099             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1100
1101             /* Update potential sum for this i atom from the interaction with this j atom. */
1102             velec            = _mm_and_ps(velec,cutoff_mask);
1103             velec            = _mm_andnot_ps(dummy_mask,velec);
1104             velecsum         = _mm_add_ps(velecsum,velec);
1105
1106             fscal            = felec;
1107
1108             fscal            = _mm_and_ps(fscal,cutoff_mask);
1109
1110             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1111
1112             /* Calculate temporary vectorial force */
1113             tx               = _mm_mul_ps(fscal,dx31);
1114             ty               = _mm_mul_ps(fscal,dy31);
1115             tz               = _mm_mul_ps(fscal,dz31);
1116
1117             /* Update vectorial force */
1118             fix3             = _mm_add_ps(fix3,tx);
1119             fiy3             = _mm_add_ps(fiy3,ty);
1120             fiz3             = _mm_add_ps(fiz3,tz);
1121
1122             fjx1             = _mm_add_ps(fjx1,tx);
1123             fjy1             = _mm_add_ps(fjy1,ty);
1124             fjz1             = _mm_add_ps(fjz1,tz);
1125
1126             }
1127
1128             /**************************
1129              * CALCULATE INTERACTIONS *
1130              **************************/
1131
1132             if (gmx_mm_any_lt(rsq32,rcutoff2))
1133             {
1134
1135             /* REACTION-FIELD ELECTROSTATICS */
1136             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1137             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1138
1139             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1140
1141             /* Update potential sum for this i atom from the interaction with this j atom. */
1142             velec            = _mm_and_ps(velec,cutoff_mask);
1143             velec            = _mm_andnot_ps(dummy_mask,velec);
1144             velecsum         = _mm_add_ps(velecsum,velec);
1145
1146             fscal            = felec;
1147
1148             fscal            = _mm_and_ps(fscal,cutoff_mask);
1149
1150             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1151
1152             /* Calculate temporary vectorial force */
1153             tx               = _mm_mul_ps(fscal,dx32);
1154             ty               = _mm_mul_ps(fscal,dy32);
1155             tz               = _mm_mul_ps(fscal,dz32);
1156
1157             /* Update vectorial force */
1158             fix3             = _mm_add_ps(fix3,tx);
1159             fiy3             = _mm_add_ps(fiy3,ty);
1160             fiz3             = _mm_add_ps(fiz3,tz);
1161
1162             fjx2             = _mm_add_ps(fjx2,tx);
1163             fjy2             = _mm_add_ps(fjy2,ty);
1164             fjz2             = _mm_add_ps(fjz2,tz);
1165
1166             }
1167
1168             /**************************
1169              * CALCULATE INTERACTIONS *
1170              **************************/
1171
1172             if (gmx_mm_any_lt(rsq33,rcutoff2))
1173             {
1174
1175             /* REACTION-FIELD ELECTROSTATICS */
1176             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1177             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1178
1179             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1180
1181             /* Update potential sum for this i atom from the interaction with this j atom. */
1182             velec            = _mm_and_ps(velec,cutoff_mask);
1183             velec            = _mm_andnot_ps(dummy_mask,velec);
1184             velecsum         = _mm_add_ps(velecsum,velec);
1185
1186             fscal            = felec;
1187
1188             fscal            = _mm_and_ps(fscal,cutoff_mask);
1189
1190             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1191
1192             /* Calculate temporary vectorial force */
1193             tx               = _mm_mul_ps(fscal,dx33);
1194             ty               = _mm_mul_ps(fscal,dy33);
1195             tz               = _mm_mul_ps(fscal,dz33);
1196
1197             /* Update vectorial force */
1198             fix3             = _mm_add_ps(fix3,tx);
1199             fiy3             = _mm_add_ps(fiy3,ty);
1200             fiz3             = _mm_add_ps(fiz3,tz);
1201
1202             fjx3             = _mm_add_ps(fjx3,tx);
1203             fjy3             = _mm_add_ps(fjy3,ty);
1204             fjz3             = _mm_add_ps(fjz3,tz);
1205
1206             }
1207
1208             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1209                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1210                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1211                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1212
1213             /* Inner loop uses 368 flops */
1214         }
1215
1216         /* End of innermost loop */
1217
1218         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1219                                               f+i_coord_offset,fshift+i_shift_offset);
1220
1221         ggid                        = gid[iidx];
1222         /* Update potential energies */
1223         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1224         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1225
1226         /* Increment number of inner iterations */
1227         inneriter                  += j_index_end - j_index_start;
1228
1229         /* Outer loop uses 38 flops */
1230     }
1231
1232     /* Increment number of outer iterations */
1233     outeriter        += nri;
1234
1235     /* Update outer/inner flops */
1236
1237     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*38 + inneriter*368);
1238 }
1239 /*
1240  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_sse2_single
1241  * Electrostatics interaction: ReactionField
1242  * VdW interaction:            LennardJones
1243  * Geometry:                   Water4-Water4
1244  * Calculate force/pot:        Force
1245  */
1246 void
1247 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_sse2_single
1248                     (t_nblist * gmx_restrict                nlist,
1249                      rvec * gmx_restrict                    xx,
1250                      rvec * gmx_restrict                    ff,
1251                      t_forcerec * gmx_restrict              fr,
1252                      t_mdatoms * gmx_restrict               mdatoms,
1253                      nb_kernel_data_t * gmx_restrict        kernel_data,
1254                      t_nrnb * gmx_restrict                  nrnb)
1255 {
1256     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1257      * just 0 for non-waters.
1258      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1259      * jnr indices corresponding to data put in the four positions in the SIMD register.
1260      */
1261     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1262     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1263     int              jnrA,jnrB,jnrC,jnrD;
1264     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1265     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1266     real             shX,shY,shZ,rcutoff_scalar;
1267     real             *shiftvec,*fshift,*x,*f;
1268     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1269     int              vdwioffset0;
1270     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1271     int              vdwioffset1;
1272     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1273     int              vdwioffset2;
1274     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1275     int              vdwioffset3;
1276     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1277     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1278     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1279     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1280     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1281     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1282     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1283     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1284     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1285     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1286     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1287     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1288     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1289     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1290     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1291     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1292     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1293     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1294     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1295     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1296     real             *charge;
1297     int              nvdwtype;
1298     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1299     int              *vdwtype;
1300     real             *vdwparam;
1301     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1302     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1303     __m128           dummy_mask,cutoff_mask;
1304     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1305     __m128           one     = _mm_set1_ps(1.0);
1306     __m128           two     = _mm_set1_ps(2.0);
1307     x                = xx[0];
1308     f                = ff[0];
1309
1310     nri              = nlist->nri;
1311     iinr             = nlist->iinr;
1312     jindex           = nlist->jindex;
1313     jjnr             = nlist->jjnr;
1314     shiftidx         = nlist->shift;
1315     gid              = nlist->gid;
1316     shiftvec         = fr->shift_vec[0];
1317     fshift           = fr->fshift[0];
1318     facel            = _mm_set1_ps(fr->epsfac);
1319     charge           = mdatoms->chargeA;
1320     krf              = _mm_set1_ps(fr->ic->k_rf);
1321     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1322     crf              = _mm_set1_ps(fr->ic->c_rf);
1323     nvdwtype         = fr->ntype;
1324     vdwparam         = fr->nbfp;
1325     vdwtype          = mdatoms->typeA;
1326
1327     /* Setup water-specific parameters */
1328     inr              = nlist->iinr[0];
1329     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1330     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1331     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1332     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1333
1334     jq1              = _mm_set1_ps(charge[inr+1]);
1335     jq2              = _mm_set1_ps(charge[inr+2]);
1336     jq3              = _mm_set1_ps(charge[inr+3]);
1337     vdwjidx0A        = 2*vdwtype[inr+0];
1338     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1339     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1340     qq11             = _mm_mul_ps(iq1,jq1);
1341     qq12             = _mm_mul_ps(iq1,jq2);
1342     qq13             = _mm_mul_ps(iq1,jq3);
1343     qq21             = _mm_mul_ps(iq2,jq1);
1344     qq22             = _mm_mul_ps(iq2,jq2);
1345     qq23             = _mm_mul_ps(iq2,jq3);
1346     qq31             = _mm_mul_ps(iq3,jq1);
1347     qq32             = _mm_mul_ps(iq3,jq2);
1348     qq33             = _mm_mul_ps(iq3,jq3);
1349
1350     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1351     rcutoff_scalar   = fr->rcoulomb;
1352     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1353     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1354
1355     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1356     rvdw             = _mm_set1_ps(fr->rvdw);
1357
1358     /* Avoid stupid compiler warnings */
1359     jnrA = jnrB = jnrC = jnrD = 0;
1360     j_coord_offsetA = 0;
1361     j_coord_offsetB = 0;
1362     j_coord_offsetC = 0;
1363     j_coord_offsetD = 0;
1364
1365     outeriter        = 0;
1366     inneriter        = 0;
1367
1368     /* Start outer loop over neighborlists */
1369     for(iidx=0; iidx<nri; iidx++)
1370     {
1371         /* Load shift vector for this list */
1372         i_shift_offset   = DIM*shiftidx[iidx];
1373         shX              = shiftvec[i_shift_offset+XX];
1374         shY              = shiftvec[i_shift_offset+YY];
1375         shZ              = shiftvec[i_shift_offset+ZZ];
1376
1377         /* Load limits for loop over neighbors */
1378         j_index_start    = jindex[iidx];
1379         j_index_end      = jindex[iidx+1];
1380
1381         /* Get outer coordinate index */
1382         inr              = iinr[iidx];
1383         i_coord_offset   = DIM*inr;
1384
1385         /* Load i particle coords and add shift vector */
1386         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1387         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1388         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1389         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1390         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1391         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1392         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1393         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1394         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1395         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1396         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1397         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1398
1399         fix0             = _mm_setzero_ps();
1400         fiy0             = _mm_setzero_ps();
1401         fiz0             = _mm_setzero_ps();
1402         fix1             = _mm_setzero_ps();
1403         fiy1             = _mm_setzero_ps();
1404         fiz1             = _mm_setzero_ps();
1405         fix2             = _mm_setzero_ps();
1406         fiy2             = _mm_setzero_ps();
1407         fiz2             = _mm_setzero_ps();
1408         fix3             = _mm_setzero_ps();
1409         fiy3             = _mm_setzero_ps();
1410         fiz3             = _mm_setzero_ps();
1411
1412         /* Start inner kernel loop */
1413         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1414         {
1415
1416             /* Get j neighbor index, and coordinate index */
1417             jnrA             = jjnr[jidx];
1418             jnrB             = jjnr[jidx+1];
1419             jnrC             = jjnr[jidx+2];
1420             jnrD             = jjnr[jidx+3];
1421
1422             j_coord_offsetA  = DIM*jnrA;
1423             j_coord_offsetB  = DIM*jnrB;
1424             j_coord_offsetC  = DIM*jnrC;
1425             j_coord_offsetD  = DIM*jnrD;
1426
1427             /* load j atom coordinates */
1428             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1429                                               x+j_coord_offsetC,x+j_coord_offsetD,
1430                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1431                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1432
1433             /* Calculate displacement vector */
1434             dx00             = _mm_sub_ps(ix0,jx0);
1435             dy00             = _mm_sub_ps(iy0,jy0);
1436             dz00             = _mm_sub_ps(iz0,jz0);
1437             dx11             = _mm_sub_ps(ix1,jx1);
1438             dy11             = _mm_sub_ps(iy1,jy1);
1439             dz11             = _mm_sub_ps(iz1,jz1);
1440             dx12             = _mm_sub_ps(ix1,jx2);
1441             dy12             = _mm_sub_ps(iy1,jy2);
1442             dz12             = _mm_sub_ps(iz1,jz2);
1443             dx13             = _mm_sub_ps(ix1,jx3);
1444             dy13             = _mm_sub_ps(iy1,jy3);
1445             dz13             = _mm_sub_ps(iz1,jz3);
1446             dx21             = _mm_sub_ps(ix2,jx1);
1447             dy21             = _mm_sub_ps(iy2,jy1);
1448             dz21             = _mm_sub_ps(iz2,jz1);
1449             dx22             = _mm_sub_ps(ix2,jx2);
1450             dy22             = _mm_sub_ps(iy2,jy2);
1451             dz22             = _mm_sub_ps(iz2,jz2);
1452             dx23             = _mm_sub_ps(ix2,jx3);
1453             dy23             = _mm_sub_ps(iy2,jy3);
1454             dz23             = _mm_sub_ps(iz2,jz3);
1455             dx31             = _mm_sub_ps(ix3,jx1);
1456             dy31             = _mm_sub_ps(iy3,jy1);
1457             dz31             = _mm_sub_ps(iz3,jz1);
1458             dx32             = _mm_sub_ps(ix3,jx2);
1459             dy32             = _mm_sub_ps(iy3,jy2);
1460             dz32             = _mm_sub_ps(iz3,jz2);
1461             dx33             = _mm_sub_ps(ix3,jx3);
1462             dy33             = _mm_sub_ps(iy3,jy3);
1463             dz33             = _mm_sub_ps(iz3,jz3);
1464
1465             /* Calculate squared distance and things based on it */
1466             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1467             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1468             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1469             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1470             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1471             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1472             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1473             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1474             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1475             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1476
1477             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1478             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1479             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1480             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1481             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1482             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1483             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1484             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1485             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1486
1487             rinvsq00         = gmx_mm_inv_ps(rsq00);
1488             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1489             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1490             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1491             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1492             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1493             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1494             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1495             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1496             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1497
1498             fjx0             = _mm_setzero_ps();
1499             fjy0             = _mm_setzero_ps();
1500             fjz0             = _mm_setzero_ps();
1501             fjx1             = _mm_setzero_ps();
1502             fjy1             = _mm_setzero_ps();
1503             fjz1             = _mm_setzero_ps();
1504             fjx2             = _mm_setzero_ps();
1505             fjy2             = _mm_setzero_ps();
1506             fjz2             = _mm_setzero_ps();
1507             fjx3             = _mm_setzero_ps();
1508             fjy3             = _mm_setzero_ps();
1509             fjz3             = _mm_setzero_ps();
1510
1511             /**************************
1512              * CALCULATE INTERACTIONS *
1513              **************************/
1514
1515             if (gmx_mm_any_lt(rsq00,rcutoff2))
1516             {
1517
1518             /* LENNARD-JONES DISPERSION/REPULSION */
1519
1520             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1521             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1522
1523             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1524
1525             fscal            = fvdw;
1526
1527             fscal            = _mm_and_ps(fscal,cutoff_mask);
1528
1529             /* Calculate temporary vectorial force */
1530             tx               = _mm_mul_ps(fscal,dx00);
1531             ty               = _mm_mul_ps(fscal,dy00);
1532             tz               = _mm_mul_ps(fscal,dz00);
1533
1534             /* Update vectorial force */
1535             fix0             = _mm_add_ps(fix0,tx);
1536             fiy0             = _mm_add_ps(fiy0,ty);
1537             fiz0             = _mm_add_ps(fiz0,tz);
1538
1539             fjx0             = _mm_add_ps(fjx0,tx);
1540             fjy0             = _mm_add_ps(fjy0,ty);
1541             fjz0             = _mm_add_ps(fjz0,tz);
1542
1543             }
1544
1545             /**************************
1546              * CALCULATE INTERACTIONS *
1547              **************************/
1548
1549             if (gmx_mm_any_lt(rsq11,rcutoff2))
1550             {
1551
1552             /* REACTION-FIELD ELECTROSTATICS */
1553             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1554
1555             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1556
1557             fscal            = felec;
1558
1559             fscal            = _mm_and_ps(fscal,cutoff_mask);
1560
1561             /* Calculate temporary vectorial force */
1562             tx               = _mm_mul_ps(fscal,dx11);
1563             ty               = _mm_mul_ps(fscal,dy11);
1564             tz               = _mm_mul_ps(fscal,dz11);
1565
1566             /* Update vectorial force */
1567             fix1             = _mm_add_ps(fix1,tx);
1568             fiy1             = _mm_add_ps(fiy1,ty);
1569             fiz1             = _mm_add_ps(fiz1,tz);
1570
1571             fjx1             = _mm_add_ps(fjx1,tx);
1572             fjy1             = _mm_add_ps(fjy1,ty);
1573             fjz1             = _mm_add_ps(fjz1,tz);
1574
1575             }
1576
1577             /**************************
1578              * CALCULATE INTERACTIONS *
1579              **************************/
1580
1581             if (gmx_mm_any_lt(rsq12,rcutoff2))
1582             {
1583
1584             /* REACTION-FIELD ELECTROSTATICS */
1585             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1586
1587             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1588
1589             fscal            = felec;
1590
1591             fscal            = _mm_and_ps(fscal,cutoff_mask);
1592
1593             /* Calculate temporary vectorial force */
1594             tx               = _mm_mul_ps(fscal,dx12);
1595             ty               = _mm_mul_ps(fscal,dy12);
1596             tz               = _mm_mul_ps(fscal,dz12);
1597
1598             /* Update vectorial force */
1599             fix1             = _mm_add_ps(fix1,tx);
1600             fiy1             = _mm_add_ps(fiy1,ty);
1601             fiz1             = _mm_add_ps(fiz1,tz);
1602
1603             fjx2             = _mm_add_ps(fjx2,tx);
1604             fjy2             = _mm_add_ps(fjy2,ty);
1605             fjz2             = _mm_add_ps(fjz2,tz);
1606
1607             }
1608
1609             /**************************
1610              * CALCULATE INTERACTIONS *
1611              **************************/
1612
1613             if (gmx_mm_any_lt(rsq13,rcutoff2))
1614             {
1615
1616             /* REACTION-FIELD ELECTROSTATICS */
1617             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1618
1619             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1620
1621             fscal            = felec;
1622
1623             fscal            = _mm_and_ps(fscal,cutoff_mask);
1624
1625             /* Calculate temporary vectorial force */
1626             tx               = _mm_mul_ps(fscal,dx13);
1627             ty               = _mm_mul_ps(fscal,dy13);
1628             tz               = _mm_mul_ps(fscal,dz13);
1629
1630             /* Update vectorial force */
1631             fix1             = _mm_add_ps(fix1,tx);
1632             fiy1             = _mm_add_ps(fiy1,ty);
1633             fiz1             = _mm_add_ps(fiz1,tz);
1634
1635             fjx3             = _mm_add_ps(fjx3,tx);
1636             fjy3             = _mm_add_ps(fjy3,ty);
1637             fjz3             = _mm_add_ps(fjz3,tz);
1638
1639             }
1640
1641             /**************************
1642              * CALCULATE INTERACTIONS *
1643              **************************/
1644
1645             if (gmx_mm_any_lt(rsq21,rcutoff2))
1646             {
1647
1648             /* REACTION-FIELD ELECTROSTATICS */
1649             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1650
1651             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1652
1653             fscal            = felec;
1654
1655             fscal            = _mm_and_ps(fscal,cutoff_mask);
1656
1657             /* Calculate temporary vectorial force */
1658             tx               = _mm_mul_ps(fscal,dx21);
1659             ty               = _mm_mul_ps(fscal,dy21);
1660             tz               = _mm_mul_ps(fscal,dz21);
1661
1662             /* Update vectorial force */
1663             fix2             = _mm_add_ps(fix2,tx);
1664             fiy2             = _mm_add_ps(fiy2,ty);
1665             fiz2             = _mm_add_ps(fiz2,tz);
1666
1667             fjx1             = _mm_add_ps(fjx1,tx);
1668             fjy1             = _mm_add_ps(fjy1,ty);
1669             fjz1             = _mm_add_ps(fjz1,tz);
1670
1671             }
1672
1673             /**************************
1674              * CALCULATE INTERACTIONS *
1675              **************************/
1676
1677             if (gmx_mm_any_lt(rsq22,rcutoff2))
1678             {
1679
1680             /* REACTION-FIELD ELECTROSTATICS */
1681             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1682
1683             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1684
1685             fscal            = felec;
1686
1687             fscal            = _mm_and_ps(fscal,cutoff_mask);
1688
1689             /* Calculate temporary vectorial force */
1690             tx               = _mm_mul_ps(fscal,dx22);
1691             ty               = _mm_mul_ps(fscal,dy22);
1692             tz               = _mm_mul_ps(fscal,dz22);
1693
1694             /* Update vectorial force */
1695             fix2             = _mm_add_ps(fix2,tx);
1696             fiy2             = _mm_add_ps(fiy2,ty);
1697             fiz2             = _mm_add_ps(fiz2,tz);
1698
1699             fjx2             = _mm_add_ps(fjx2,tx);
1700             fjy2             = _mm_add_ps(fjy2,ty);
1701             fjz2             = _mm_add_ps(fjz2,tz);
1702
1703             }
1704
1705             /**************************
1706              * CALCULATE INTERACTIONS *
1707              **************************/
1708
1709             if (gmx_mm_any_lt(rsq23,rcutoff2))
1710             {
1711
1712             /* REACTION-FIELD ELECTROSTATICS */
1713             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1714
1715             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1716
1717             fscal            = felec;
1718
1719             fscal            = _mm_and_ps(fscal,cutoff_mask);
1720
1721             /* Calculate temporary vectorial force */
1722             tx               = _mm_mul_ps(fscal,dx23);
1723             ty               = _mm_mul_ps(fscal,dy23);
1724             tz               = _mm_mul_ps(fscal,dz23);
1725
1726             /* Update vectorial force */
1727             fix2             = _mm_add_ps(fix2,tx);
1728             fiy2             = _mm_add_ps(fiy2,ty);
1729             fiz2             = _mm_add_ps(fiz2,tz);
1730
1731             fjx3             = _mm_add_ps(fjx3,tx);
1732             fjy3             = _mm_add_ps(fjy3,ty);
1733             fjz3             = _mm_add_ps(fjz3,tz);
1734
1735             }
1736
1737             /**************************
1738              * CALCULATE INTERACTIONS *
1739              **************************/
1740
1741             if (gmx_mm_any_lt(rsq31,rcutoff2))
1742             {
1743
1744             /* REACTION-FIELD ELECTROSTATICS */
1745             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1746
1747             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1748
1749             fscal            = felec;
1750
1751             fscal            = _mm_and_ps(fscal,cutoff_mask);
1752
1753             /* Calculate temporary vectorial force */
1754             tx               = _mm_mul_ps(fscal,dx31);
1755             ty               = _mm_mul_ps(fscal,dy31);
1756             tz               = _mm_mul_ps(fscal,dz31);
1757
1758             /* Update vectorial force */
1759             fix3             = _mm_add_ps(fix3,tx);
1760             fiy3             = _mm_add_ps(fiy3,ty);
1761             fiz3             = _mm_add_ps(fiz3,tz);
1762
1763             fjx1             = _mm_add_ps(fjx1,tx);
1764             fjy1             = _mm_add_ps(fjy1,ty);
1765             fjz1             = _mm_add_ps(fjz1,tz);
1766
1767             }
1768
1769             /**************************
1770              * CALCULATE INTERACTIONS *
1771              **************************/
1772
1773             if (gmx_mm_any_lt(rsq32,rcutoff2))
1774             {
1775
1776             /* REACTION-FIELD ELECTROSTATICS */
1777             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1778
1779             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1780
1781             fscal            = felec;
1782
1783             fscal            = _mm_and_ps(fscal,cutoff_mask);
1784
1785             /* Calculate temporary vectorial force */
1786             tx               = _mm_mul_ps(fscal,dx32);
1787             ty               = _mm_mul_ps(fscal,dy32);
1788             tz               = _mm_mul_ps(fscal,dz32);
1789
1790             /* Update vectorial force */
1791             fix3             = _mm_add_ps(fix3,tx);
1792             fiy3             = _mm_add_ps(fiy3,ty);
1793             fiz3             = _mm_add_ps(fiz3,tz);
1794
1795             fjx2             = _mm_add_ps(fjx2,tx);
1796             fjy2             = _mm_add_ps(fjy2,ty);
1797             fjz2             = _mm_add_ps(fjz2,tz);
1798
1799             }
1800
1801             /**************************
1802              * CALCULATE INTERACTIONS *
1803              **************************/
1804
1805             if (gmx_mm_any_lt(rsq33,rcutoff2))
1806             {
1807
1808             /* REACTION-FIELD ELECTROSTATICS */
1809             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1810
1811             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1812
1813             fscal            = felec;
1814
1815             fscal            = _mm_and_ps(fscal,cutoff_mask);
1816
1817             /* Calculate temporary vectorial force */
1818             tx               = _mm_mul_ps(fscal,dx33);
1819             ty               = _mm_mul_ps(fscal,dy33);
1820             tz               = _mm_mul_ps(fscal,dz33);
1821
1822             /* Update vectorial force */
1823             fix3             = _mm_add_ps(fix3,tx);
1824             fiy3             = _mm_add_ps(fiy3,ty);
1825             fiz3             = _mm_add_ps(fiz3,tz);
1826
1827             fjx3             = _mm_add_ps(fjx3,tx);
1828             fjy3             = _mm_add_ps(fjy3,ty);
1829             fjz3             = _mm_add_ps(fjz3,tz);
1830
1831             }
1832
1833             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1834                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1835                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1836                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1837
1838             /* Inner loop uses 303 flops */
1839         }
1840
1841         if(jidx<j_index_end)
1842         {
1843
1844             /* Get j neighbor index, and coordinate index */
1845             jnrA             = jjnr[jidx];
1846             jnrB             = jjnr[jidx+1];
1847             jnrC             = jjnr[jidx+2];
1848             jnrD             = jjnr[jidx+3];
1849
1850             /* Sign of each element will be negative for non-real atoms.
1851              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1852              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1853              */
1854             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1855             jnrA       = (jnrA>=0) ? jnrA : 0;
1856             jnrB       = (jnrB>=0) ? jnrB : 0;
1857             jnrC       = (jnrC>=0) ? jnrC : 0;
1858             jnrD       = (jnrD>=0) ? jnrD : 0;
1859
1860             j_coord_offsetA  = DIM*jnrA;
1861             j_coord_offsetB  = DIM*jnrB;
1862             j_coord_offsetC  = DIM*jnrC;
1863             j_coord_offsetD  = DIM*jnrD;
1864
1865             /* load j atom coordinates */
1866             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1867                                               x+j_coord_offsetC,x+j_coord_offsetD,
1868                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1869                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1870
1871             /* Calculate displacement vector */
1872             dx00             = _mm_sub_ps(ix0,jx0);
1873             dy00             = _mm_sub_ps(iy0,jy0);
1874             dz00             = _mm_sub_ps(iz0,jz0);
1875             dx11             = _mm_sub_ps(ix1,jx1);
1876             dy11             = _mm_sub_ps(iy1,jy1);
1877             dz11             = _mm_sub_ps(iz1,jz1);
1878             dx12             = _mm_sub_ps(ix1,jx2);
1879             dy12             = _mm_sub_ps(iy1,jy2);
1880             dz12             = _mm_sub_ps(iz1,jz2);
1881             dx13             = _mm_sub_ps(ix1,jx3);
1882             dy13             = _mm_sub_ps(iy1,jy3);
1883             dz13             = _mm_sub_ps(iz1,jz3);
1884             dx21             = _mm_sub_ps(ix2,jx1);
1885             dy21             = _mm_sub_ps(iy2,jy1);
1886             dz21             = _mm_sub_ps(iz2,jz1);
1887             dx22             = _mm_sub_ps(ix2,jx2);
1888             dy22             = _mm_sub_ps(iy2,jy2);
1889             dz22             = _mm_sub_ps(iz2,jz2);
1890             dx23             = _mm_sub_ps(ix2,jx3);
1891             dy23             = _mm_sub_ps(iy2,jy3);
1892             dz23             = _mm_sub_ps(iz2,jz3);
1893             dx31             = _mm_sub_ps(ix3,jx1);
1894             dy31             = _mm_sub_ps(iy3,jy1);
1895             dz31             = _mm_sub_ps(iz3,jz1);
1896             dx32             = _mm_sub_ps(ix3,jx2);
1897             dy32             = _mm_sub_ps(iy3,jy2);
1898             dz32             = _mm_sub_ps(iz3,jz2);
1899             dx33             = _mm_sub_ps(ix3,jx3);
1900             dy33             = _mm_sub_ps(iy3,jy3);
1901             dz33             = _mm_sub_ps(iz3,jz3);
1902
1903             /* Calculate squared distance and things based on it */
1904             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1905             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1906             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1907             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1908             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1909             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1910             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1911             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1912             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1913             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1914
1915             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1916             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1917             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1918             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1919             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1920             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1921             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1922             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1923             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1924
1925             rinvsq00         = gmx_mm_inv_ps(rsq00);
1926             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1927             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1928             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1929             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1930             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1931             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1932             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1933             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1934             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1935
1936             fjx0             = _mm_setzero_ps();
1937             fjy0             = _mm_setzero_ps();
1938             fjz0             = _mm_setzero_ps();
1939             fjx1             = _mm_setzero_ps();
1940             fjy1             = _mm_setzero_ps();
1941             fjz1             = _mm_setzero_ps();
1942             fjx2             = _mm_setzero_ps();
1943             fjy2             = _mm_setzero_ps();
1944             fjz2             = _mm_setzero_ps();
1945             fjx3             = _mm_setzero_ps();
1946             fjy3             = _mm_setzero_ps();
1947             fjz3             = _mm_setzero_ps();
1948
1949             /**************************
1950              * CALCULATE INTERACTIONS *
1951              **************************/
1952
1953             if (gmx_mm_any_lt(rsq00,rcutoff2))
1954             {
1955
1956             /* LENNARD-JONES DISPERSION/REPULSION */
1957
1958             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1959             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1960
1961             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1962
1963             fscal            = fvdw;
1964
1965             fscal            = _mm_and_ps(fscal,cutoff_mask);
1966
1967             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1968
1969             /* Calculate temporary vectorial force */
1970             tx               = _mm_mul_ps(fscal,dx00);
1971             ty               = _mm_mul_ps(fscal,dy00);
1972             tz               = _mm_mul_ps(fscal,dz00);
1973
1974             /* Update vectorial force */
1975             fix0             = _mm_add_ps(fix0,tx);
1976             fiy0             = _mm_add_ps(fiy0,ty);
1977             fiz0             = _mm_add_ps(fiz0,tz);
1978
1979             fjx0             = _mm_add_ps(fjx0,tx);
1980             fjy0             = _mm_add_ps(fjy0,ty);
1981             fjz0             = _mm_add_ps(fjz0,tz);
1982
1983             }
1984
1985             /**************************
1986              * CALCULATE INTERACTIONS *
1987              **************************/
1988
1989             if (gmx_mm_any_lt(rsq11,rcutoff2))
1990             {
1991
1992             /* REACTION-FIELD ELECTROSTATICS */
1993             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1994
1995             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1996
1997             fscal            = felec;
1998
1999             fscal            = _mm_and_ps(fscal,cutoff_mask);
2000
2001             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2002
2003             /* Calculate temporary vectorial force */
2004             tx               = _mm_mul_ps(fscal,dx11);
2005             ty               = _mm_mul_ps(fscal,dy11);
2006             tz               = _mm_mul_ps(fscal,dz11);
2007
2008             /* Update vectorial force */
2009             fix1             = _mm_add_ps(fix1,tx);
2010             fiy1             = _mm_add_ps(fiy1,ty);
2011             fiz1             = _mm_add_ps(fiz1,tz);
2012
2013             fjx1             = _mm_add_ps(fjx1,tx);
2014             fjy1             = _mm_add_ps(fjy1,ty);
2015             fjz1             = _mm_add_ps(fjz1,tz);
2016
2017             }
2018
2019             /**************************
2020              * CALCULATE INTERACTIONS *
2021              **************************/
2022
2023             if (gmx_mm_any_lt(rsq12,rcutoff2))
2024             {
2025
2026             /* REACTION-FIELD ELECTROSTATICS */
2027             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
2028
2029             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2030
2031             fscal            = felec;
2032
2033             fscal            = _mm_and_ps(fscal,cutoff_mask);
2034
2035             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2036
2037             /* Calculate temporary vectorial force */
2038             tx               = _mm_mul_ps(fscal,dx12);
2039             ty               = _mm_mul_ps(fscal,dy12);
2040             tz               = _mm_mul_ps(fscal,dz12);
2041
2042             /* Update vectorial force */
2043             fix1             = _mm_add_ps(fix1,tx);
2044             fiy1             = _mm_add_ps(fiy1,ty);
2045             fiz1             = _mm_add_ps(fiz1,tz);
2046
2047             fjx2             = _mm_add_ps(fjx2,tx);
2048             fjy2             = _mm_add_ps(fjy2,ty);
2049             fjz2             = _mm_add_ps(fjz2,tz);
2050
2051             }
2052
2053             /**************************
2054              * CALCULATE INTERACTIONS *
2055              **************************/
2056
2057             if (gmx_mm_any_lt(rsq13,rcutoff2))
2058             {
2059
2060             /* REACTION-FIELD ELECTROSTATICS */
2061             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
2062
2063             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
2064
2065             fscal            = felec;
2066
2067             fscal            = _mm_and_ps(fscal,cutoff_mask);
2068
2069             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2070
2071             /* Calculate temporary vectorial force */
2072             tx               = _mm_mul_ps(fscal,dx13);
2073             ty               = _mm_mul_ps(fscal,dy13);
2074             tz               = _mm_mul_ps(fscal,dz13);
2075
2076             /* Update vectorial force */
2077             fix1             = _mm_add_ps(fix1,tx);
2078             fiy1             = _mm_add_ps(fiy1,ty);
2079             fiz1             = _mm_add_ps(fiz1,tz);
2080
2081             fjx3             = _mm_add_ps(fjx3,tx);
2082             fjy3             = _mm_add_ps(fjy3,ty);
2083             fjz3             = _mm_add_ps(fjz3,tz);
2084
2085             }
2086
2087             /**************************
2088              * CALCULATE INTERACTIONS *
2089              **************************/
2090
2091             if (gmx_mm_any_lt(rsq21,rcutoff2))
2092             {
2093
2094             /* REACTION-FIELD ELECTROSTATICS */
2095             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
2096
2097             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2098
2099             fscal            = felec;
2100
2101             fscal            = _mm_and_ps(fscal,cutoff_mask);
2102
2103             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2104
2105             /* Calculate temporary vectorial force */
2106             tx               = _mm_mul_ps(fscal,dx21);
2107             ty               = _mm_mul_ps(fscal,dy21);
2108             tz               = _mm_mul_ps(fscal,dz21);
2109
2110             /* Update vectorial force */
2111             fix2             = _mm_add_ps(fix2,tx);
2112             fiy2             = _mm_add_ps(fiy2,ty);
2113             fiz2             = _mm_add_ps(fiz2,tz);
2114
2115             fjx1             = _mm_add_ps(fjx1,tx);
2116             fjy1             = _mm_add_ps(fjy1,ty);
2117             fjz1             = _mm_add_ps(fjz1,tz);
2118
2119             }
2120
2121             /**************************
2122              * CALCULATE INTERACTIONS *
2123              **************************/
2124
2125             if (gmx_mm_any_lt(rsq22,rcutoff2))
2126             {
2127
2128             /* REACTION-FIELD ELECTROSTATICS */
2129             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2130
2131             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2132
2133             fscal            = felec;
2134
2135             fscal            = _mm_and_ps(fscal,cutoff_mask);
2136
2137             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2138
2139             /* Calculate temporary vectorial force */
2140             tx               = _mm_mul_ps(fscal,dx22);
2141             ty               = _mm_mul_ps(fscal,dy22);
2142             tz               = _mm_mul_ps(fscal,dz22);
2143
2144             /* Update vectorial force */
2145             fix2             = _mm_add_ps(fix2,tx);
2146             fiy2             = _mm_add_ps(fiy2,ty);
2147             fiz2             = _mm_add_ps(fiz2,tz);
2148
2149             fjx2             = _mm_add_ps(fjx2,tx);
2150             fjy2             = _mm_add_ps(fjy2,ty);
2151             fjz2             = _mm_add_ps(fjz2,tz);
2152
2153             }
2154
2155             /**************************
2156              * CALCULATE INTERACTIONS *
2157              **************************/
2158
2159             if (gmx_mm_any_lt(rsq23,rcutoff2))
2160             {
2161
2162             /* REACTION-FIELD ELECTROSTATICS */
2163             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
2164
2165             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2166
2167             fscal            = felec;
2168
2169             fscal            = _mm_and_ps(fscal,cutoff_mask);
2170
2171             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2172
2173             /* Calculate temporary vectorial force */
2174             tx               = _mm_mul_ps(fscal,dx23);
2175             ty               = _mm_mul_ps(fscal,dy23);
2176             tz               = _mm_mul_ps(fscal,dz23);
2177
2178             /* Update vectorial force */
2179             fix2             = _mm_add_ps(fix2,tx);
2180             fiy2             = _mm_add_ps(fiy2,ty);
2181             fiz2             = _mm_add_ps(fiz2,tz);
2182
2183             fjx3             = _mm_add_ps(fjx3,tx);
2184             fjy3             = _mm_add_ps(fjy3,ty);
2185             fjz3             = _mm_add_ps(fjz3,tz);
2186
2187             }
2188
2189             /**************************
2190              * CALCULATE INTERACTIONS *
2191              **************************/
2192
2193             if (gmx_mm_any_lt(rsq31,rcutoff2))
2194             {
2195
2196             /* REACTION-FIELD ELECTROSTATICS */
2197             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
2198
2199             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2200
2201             fscal            = felec;
2202
2203             fscal            = _mm_and_ps(fscal,cutoff_mask);
2204
2205             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2206
2207             /* Calculate temporary vectorial force */
2208             tx               = _mm_mul_ps(fscal,dx31);
2209             ty               = _mm_mul_ps(fscal,dy31);
2210             tz               = _mm_mul_ps(fscal,dz31);
2211
2212             /* Update vectorial force */
2213             fix3             = _mm_add_ps(fix3,tx);
2214             fiy3             = _mm_add_ps(fiy3,ty);
2215             fiz3             = _mm_add_ps(fiz3,tz);
2216
2217             fjx1             = _mm_add_ps(fjx1,tx);
2218             fjy1             = _mm_add_ps(fjy1,ty);
2219             fjz1             = _mm_add_ps(fjz1,tz);
2220
2221             }
2222
2223             /**************************
2224              * CALCULATE INTERACTIONS *
2225              **************************/
2226
2227             if (gmx_mm_any_lt(rsq32,rcutoff2))
2228             {
2229
2230             /* REACTION-FIELD ELECTROSTATICS */
2231             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
2232
2233             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2234
2235             fscal            = felec;
2236
2237             fscal            = _mm_and_ps(fscal,cutoff_mask);
2238
2239             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2240
2241             /* Calculate temporary vectorial force */
2242             tx               = _mm_mul_ps(fscal,dx32);
2243             ty               = _mm_mul_ps(fscal,dy32);
2244             tz               = _mm_mul_ps(fscal,dz32);
2245
2246             /* Update vectorial force */
2247             fix3             = _mm_add_ps(fix3,tx);
2248             fiy3             = _mm_add_ps(fiy3,ty);
2249             fiz3             = _mm_add_ps(fiz3,tz);
2250
2251             fjx2             = _mm_add_ps(fjx2,tx);
2252             fjy2             = _mm_add_ps(fjy2,ty);
2253             fjz2             = _mm_add_ps(fjz2,tz);
2254
2255             }
2256
2257             /**************************
2258              * CALCULATE INTERACTIONS *
2259              **************************/
2260
2261             if (gmx_mm_any_lt(rsq33,rcutoff2))
2262             {
2263
2264             /* REACTION-FIELD ELECTROSTATICS */
2265             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2266
2267             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2268
2269             fscal            = felec;
2270
2271             fscal            = _mm_and_ps(fscal,cutoff_mask);
2272
2273             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2274
2275             /* Calculate temporary vectorial force */
2276             tx               = _mm_mul_ps(fscal,dx33);
2277             ty               = _mm_mul_ps(fscal,dy33);
2278             tz               = _mm_mul_ps(fscal,dz33);
2279
2280             /* Update vectorial force */
2281             fix3             = _mm_add_ps(fix3,tx);
2282             fiy3             = _mm_add_ps(fiy3,ty);
2283             fiz3             = _mm_add_ps(fiz3,tz);
2284
2285             fjx3             = _mm_add_ps(fjx3,tx);
2286             fjy3             = _mm_add_ps(fjy3,ty);
2287             fjz3             = _mm_add_ps(fjz3,tz);
2288
2289             }
2290
2291             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
2292                                                    f+j_coord_offsetC,f+j_coord_offsetD,
2293                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2294                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2295
2296             /* Inner loop uses 303 flops */
2297         }
2298
2299         /* End of innermost loop */
2300
2301         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2302                                               f+i_coord_offset,fshift+i_shift_offset);
2303
2304         /* Increment number of inner iterations */
2305         inneriter                  += j_index_end - j_index_start;
2306
2307         /* Outer loop uses 36 flops */
2308     }
2309
2310     /* Increment number of outer iterations */
2311     outeriter        += nri;
2312
2313     /* Update outer/inner flops */
2314
2315     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*36 + inneriter*303);
2316 }