src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse2_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             shX,shY,shZ,rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwioffset3;
  73     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  75     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  77     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  79     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  81     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  99     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 100     __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     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     /* Setup water-specific parameters */
 122     inr              = nlist->iinr[0];
 123     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 124     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 125     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 126     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 127
 128     jq1              = _mm_set1_ps(charge[inr+1]);
 129     jq2              = _mm_set1_ps(charge[inr+2]);
 130     jq3              = _mm_set1_ps(charge[inr+3]);
 131     vdwjidx0A        = 2*vdwtype[inr+0];
 132     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 133     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 134     qq11             = _mm_mul_ps(iq1,jq1);
 135     qq12             = _mm_mul_ps(iq1,jq2);
 136     qq13             = _mm_mul_ps(iq1,jq3);
 137     qq21             = _mm_mul_ps(iq2,jq1);
 138     qq22             = _mm_mul_ps(iq2,jq2);
 139     qq23             = _mm_mul_ps(iq2,jq3);
 140     qq31             = _mm_mul_ps(iq3,jq1);
 141     qq32             = _mm_mul_ps(iq3,jq2);
 142     qq33             = _mm_mul_ps(iq3,jq3);
 143
 144     /* Avoid stupid compiler warnings */
 145     jnrA = jnrB = jnrC = jnrD = 0;
 146     j_coord_offsetA = 0;
 147     j_coord_offsetB = 0;
 148     j_coord_offsetC = 0;
 149     j_coord_offsetD = 0;
 150
 151     outeriter        = 0;
 152     inneriter        = 0;
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159         shX              = shiftvec[i_shift_offset+XX];
 160         shY              = shiftvec[i_shift_offset+YY];
 161         shZ              = shiftvec[i_shift_offset+ZZ];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 173         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 174         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 175         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 176         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 177         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 178         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 179         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 180         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 181         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 182         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 183         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 184
 185         fix0             = _mm_setzero_ps();
 186         fiy0             = _mm_setzero_ps();
 187         fiz0             = _mm_setzero_ps();
 188         fix1             = _mm_setzero_ps();
 189         fiy1             = _mm_setzero_ps();
 190         fiz1             = _mm_setzero_ps();
 191         fix2             = _mm_setzero_ps();
 192         fiy2             = _mm_setzero_ps();
 193         fiz2             = _mm_setzero_ps();
 194         fix3             = _mm_setzero_ps();
 195         fiy3             = _mm_setzero_ps();
 196         fiz3             = _mm_setzero_ps();
 197
 198         /* Reset potential sums */
 199         velecsum         = _mm_setzero_ps();
 200         vvdwsum          = _mm_setzero_ps();
 201
 202         /* Start inner kernel loop */
 203         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 204         {
 205
 206             /* Get j neighbor index, and coordinate index */
 207             jnrA             = jjnr[jidx];
 208             jnrB             = jjnr[jidx+1];
 209             jnrC             = jjnr[jidx+2];
 210             jnrD             = jjnr[jidx+3];
 211
 212             j_coord_offsetA  = DIM*jnrA;
 213             j_coord_offsetB  = DIM*jnrB;
 214             j_coord_offsetC  = DIM*jnrC;
 215             j_coord_offsetD  = DIM*jnrD;
 216
 217             /* load j atom coordinates */
 218             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 219                                               x+j_coord_offsetC,x+j_coord_offsetD,
 220                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 221                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 222
 223             /* Calculate displacement vector */
 224             dx00             = _mm_sub_ps(ix0,jx0);
 225             dy00             = _mm_sub_ps(iy0,jy0);
 226             dz00             = _mm_sub_ps(iz0,jz0);
 227             dx11             = _mm_sub_ps(ix1,jx1);
 228             dy11             = _mm_sub_ps(iy1,jy1);
 229             dz11             = _mm_sub_ps(iz1,jz1);
 230             dx12             = _mm_sub_ps(ix1,jx2);
 231             dy12             = _mm_sub_ps(iy1,jy2);
 232             dz12             = _mm_sub_ps(iz1,jz2);
 233             dx13             = _mm_sub_ps(ix1,jx3);
 234             dy13             = _mm_sub_ps(iy1,jy3);
 235             dz13             = _mm_sub_ps(iz1,jz3);
 236             dx21             = _mm_sub_ps(ix2,jx1);
 237             dy21             = _mm_sub_ps(iy2,jy1);
 238             dz21             = _mm_sub_ps(iz2,jz1);
 239             dx22             = _mm_sub_ps(ix2,jx2);
 240             dy22             = _mm_sub_ps(iy2,jy2);
 241             dz22             = _mm_sub_ps(iz2,jz2);
 242             dx23             = _mm_sub_ps(ix2,jx3);
 243             dy23             = _mm_sub_ps(iy2,jy3);
 244             dz23             = _mm_sub_ps(iz2,jz3);
 245             dx31             = _mm_sub_ps(ix3,jx1);
 246             dy31             = _mm_sub_ps(iy3,jy1);
 247             dz31             = _mm_sub_ps(iz3,jz1);
 248             dx32             = _mm_sub_ps(ix3,jx2);
 249             dy32             = _mm_sub_ps(iy3,jy2);
 250             dz32             = _mm_sub_ps(iz3,jz2);
 251             dx33             = _mm_sub_ps(ix3,jx3);
 252             dy33             = _mm_sub_ps(iy3,jy3);
 253             dz33             = _mm_sub_ps(iz3,jz3);
 254
 255             /* Calculate squared distance and things based on it */
 256             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 257             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 258             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 259             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 260             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 261             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 262             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 263             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 264             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 265             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 266
 267             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 268             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 269             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 270             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 271             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 272             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 273             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 274             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 275             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 276
 277             rinvsq00         = gmx_mm_inv_ps(rsq00);
 278             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 279             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 280             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 281             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 282             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 283             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 284             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 285             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 286             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 287
 288             fjx0             = _mm_setzero_ps();
 289             fjy0             = _mm_setzero_ps();
 290             fjz0             = _mm_setzero_ps();
 291             fjx1             = _mm_setzero_ps();
 292             fjy1             = _mm_setzero_ps();
 293             fjz1             = _mm_setzero_ps();
 294             fjx2             = _mm_setzero_ps();
 295             fjy2             = _mm_setzero_ps();
 296             fjz2             = _mm_setzero_ps();
 297             fjx3             = _mm_setzero_ps();
 298             fjy3             = _mm_setzero_ps();
 299             fjz3             = _mm_setzero_ps();
 300
 301             /**************************
 302              * CALCULATE INTERACTIONS *
 303              **************************/
 304
 305             /* LENNARD-JONES DISPERSION/REPULSION */
 306
 307             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 308             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 309             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 310             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 311             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 312
 313             /* Update potential sum for this i atom from the interaction with this j atom. */
 314             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 315
 316             fscal            = fvdw;
 317
 318             /* Calculate temporary vectorial force */
 319             tx               = _mm_mul_ps(fscal,dx00);
 320             ty               = _mm_mul_ps(fscal,dy00);
 321             tz               = _mm_mul_ps(fscal,dz00);
 322
 323             /* Update vectorial force */
 324             fix0             = _mm_add_ps(fix0,tx);
 325             fiy0             = _mm_add_ps(fiy0,ty);
 326             fiz0             = _mm_add_ps(fiz0,tz);
 327
 328             fjx0             = _mm_add_ps(fjx0,tx);
 329             fjy0             = _mm_add_ps(fjy0,ty);
 330             fjz0             = _mm_add_ps(fjz0,tz);
 331
 332             /**************************
 333              * CALCULATE INTERACTIONS *
 334              **************************/
 335
 336             /* COULOMB ELECTROSTATICS */
 337             velec            = _mm_mul_ps(qq11,rinv11);
 338             felec            = _mm_mul_ps(velec,rinvsq11);
 339
 340             /* Update potential sum for this i atom from the interaction with this j atom. */
 341             velecsum         = _mm_add_ps(velecsum,velec);
 342
 343             fscal            = felec;
 344
 345             /* Calculate temporary vectorial force */
 346             tx               = _mm_mul_ps(fscal,dx11);
 347             ty               = _mm_mul_ps(fscal,dy11);
 348             tz               = _mm_mul_ps(fscal,dz11);
 349
 350             /* Update vectorial force */
 351             fix1             = _mm_add_ps(fix1,tx);
 352             fiy1             = _mm_add_ps(fiy1,ty);
 353             fiz1             = _mm_add_ps(fiz1,tz);
 354
 355             fjx1             = _mm_add_ps(fjx1,tx);
 356             fjy1             = _mm_add_ps(fjy1,ty);
 357             fjz1             = _mm_add_ps(fjz1,tz);
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             /* COULOMB ELECTROSTATICS */
 364             velec            = _mm_mul_ps(qq12,rinv12);
 365             felec            = _mm_mul_ps(velec,rinvsq12);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velecsum         = _mm_add_ps(velecsum,velec);
 369
 370             fscal            = felec;
 371
 372             /* Calculate temporary vectorial force */
 373             tx               = _mm_mul_ps(fscal,dx12);
 374             ty               = _mm_mul_ps(fscal,dy12);
 375             tz               = _mm_mul_ps(fscal,dz12);
 376
 377             /* Update vectorial force */
 378             fix1             = _mm_add_ps(fix1,tx);
 379             fiy1             = _mm_add_ps(fiy1,ty);
 380             fiz1             = _mm_add_ps(fiz1,tz);
 381
 382             fjx2             = _mm_add_ps(fjx2,tx);
 383             fjy2             = _mm_add_ps(fjy2,ty);
 384             fjz2             = _mm_add_ps(fjz2,tz);
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             /* COULOMB ELECTROSTATICS */
 391             velec            = _mm_mul_ps(qq13,rinv13);
 392             felec            = _mm_mul_ps(velec,rinvsq13);
 393
 394             /* Update potential sum for this i atom from the interaction with this j atom. */
 395             velecsum         = _mm_add_ps(velecsum,velec);
 396
 397             fscal            = felec;
 398
 399             /* Calculate temporary vectorial force */
 400             tx               = _mm_mul_ps(fscal,dx13);
 401             ty               = _mm_mul_ps(fscal,dy13);
 402             tz               = _mm_mul_ps(fscal,dz13);
 403
 404             /* Update vectorial force */
 405             fix1             = _mm_add_ps(fix1,tx);
 406             fiy1             = _mm_add_ps(fiy1,ty);
 407             fiz1             = _mm_add_ps(fiz1,tz);
 408
 409             fjx3             = _mm_add_ps(fjx3,tx);
 410             fjy3             = _mm_add_ps(fjy3,ty);
 411             fjz3             = _mm_add_ps(fjz3,tz);
 412
 413             /**************************
 414              * CALCULATE INTERACTIONS *
 415              **************************/
 416
 417             /* COULOMB ELECTROSTATICS */
 418             velec            = _mm_mul_ps(qq21,rinv21);
 419             felec            = _mm_mul_ps(velec,rinvsq21);
 420
 421             /* Update potential sum for this i atom from the interaction with this j atom. */
 422             velecsum         = _mm_add_ps(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             /* Calculate temporary vectorial force */
 427             tx               = _mm_mul_ps(fscal,dx21);
 428             ty               = _mm_mul_ps(fscal,dy21);
 429             tz               = _mm_mul_ps(fscal,dz21);
 430
 431             /* Update vectorial force */
 432             fix2             = _mm_add_ps(fix2,tx);
 433             fiy2             = _mm_add_ps(fiy2,ty);
 434             fiz2             = _mm_add_ps(fiz2,tz);
 435
 436             fjx1             = _mm_add_ps(fjx1,tx);
 437             fjy1             = _mm_add_ps(fjy1,ty);
 438             fjz1             = _mm_add_ps(fjz1,tz);
 439
 440             /**************************
 441              * CALCULATE INTERACTIONS *
 442              **************************/
 443
 444             /* COULOMB ELECTROSTATICS */
 445             velec            = _mm_mul_ps(qq22,rinv22);
 446             felec            = _mm_mul_ps(velec,rinvsq22);
 447
 448             /* Update potential sum for this i atom from the interaction with this j atom. */
 449             velecsum         = _mm_add_ps(velecsum,velec);
 450
 451             fscal            = felec;
 452
 453             /* Calculate temporary vectorial force */
 454             tx               = _mm_mul_ps(fscal,dx22);
 455             ty               = _mm_mul_ps(fscal,dy22);
 456             tz               = _mm_mul_ps(fscal,dz22);
 457
 458             /* Update vectorial force */
 459             fix2             = _mm_add_ps(fix2,tx);
 460             fiy2             = _mm_add_ps(fiy2,ty);
 461             fiz2             = _mm_add_ps(fiz2,tz);
 462
 463             fjx2             = _mm_add_ps(fjx2,tx);
 464             fjy2             = _mm_add_ps(fjy2,ty);
 465             fjz2             = _mm_add_ps(fjz2,tz);
 466
 467             /**************************
 468              * CALCULATE INTERACTIONS *
 469              **************************/
 470
 471             /* COULOMB ELECTROSTATICS */
 472             velec            = _mm_mul_ps(qq23,rinv23);
 473             felec            = _mm_mul_ps(velec,rinvsq23);
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velecsum         = _mm_add_ps(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480             /* Calculate temporary vectorial force */
 481             tx               = _mm_mul_ps(fscal,dx23);
 482             ty               = _mm_mul_ps(fscal,dy23);
 483             tz               = _mm_mul_ps(fscal,dz23);
 484
 485             /* Update vectorial force */
 486             fix2             = _mm_add_ps(fix2,tx);
 487             fiy2             = _mm_add_ps(fiy2,ty);
 488             fiz2             = _mm_add_ps(fiz2,tz);
 489
 490             fjx3             = _mm_add_ps(fjx3,tx);
 491             fjy3             = _mm_add_ps(fjy3,ty);
 492             fjz3             = _mm_add_ps(fjz3,tz);
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             /* COULOMB ELECTROSTATICS */
 499             velec            = _mm_mul_ps(qq31,rinv31);
 500             felec            = _mm_mul_ps(velec,rinvsq31);
 501
 502             /* Update potential sum for this i atom from the interaction with this j atom. */
 503             velecsum         = _mm_add_ps(velecsum,velec);
 504
 505             fscal            = felec;
 506
 507             /* Calculate temporary vectorial force */
 508             tx               = _mm_mul_ps(fscal,dx31);
 509             ty               = _mm_mul_ps(fscal,dy31);
 510             tz               = _mm_mul_ps(fscal,dz31);
 511
 512             /* Update vectorial force */
 513             fix3             = _mm_add_ps(fix3,tx);
 514             fiy3             = _mm_add_ps(fiy3,ty);
 515             fiz3             = _mm_add_ps(fiz3,tz);
 516
 517             fjx1             = _mm_add_ps(fjx1,tx);
 518             fjy1             = _mm_add_ps(fjy1,ty);
 519             fjz1             = _mm_add_ps(fjz1,tz);
 520
 521             /**************************
 522              * CALCULATE INTERACTIONS *
 523              **************************/
 524
 525             /* COULOMB ELECTROSTATICS */
 526             velec            = _mm_mul_ps(qq32,rinv32);
 527             felec            = _mm_mul_ps(velec,rinvsq32);
 528
 529             /* Update potential sum for this i atom from the interaction with this j atom. */
 530             velecsum         = _mm_add_ps(velecsum,velec);
 531
 532             fscal            = felec;
 533
 534             /* Calculate temporary vectorial force */
 535             tx               = _mm_mul_ps(fscal,dx32);
 536             ty               = _mm_mul_ps(fscal,dy32);
 537             tz               = _mm_mul_ps(fscal,dz32);
 538
 539             /* Update vectorial force */
 540             fix3             = _mm_add_ps(fix3,tx);
 541             fiy3             = _mm_add_ps(fiy3,ty);
 542             fiz3             = _mm_add_ps(fiz3,tz);
 543
 544             fjx2             = _mm_add_ps(fjx2,tx);
 545             fjy2             = _mm_add_ps(fjy2,ty);
 546             fjz2             = _mm_add_ps(fjz2,tz);
 547
 548             /**************************
 549              * CALCULATE INTERACTIONS *
 550              **************************/
 551
 552             /* COULOMB ELECTROSTATICS */
 553             velec            = _mm_mul_ps(qq33,rinv33);
 554             felec            = _mm_mul_ps(velec,rinvsq33);
 555
 556             /* Update potential sum for this i atom from the interaction with this j atom. */
 557             velecsum         = _mm_add_ps(velecsum,velec);
 558
 559             fscal            = felec;
 560
 561             /* Calculate temporary vectorial force */
 562             tx               = _mm_mul_ps(fscal,dx33);
 563             ty               = _mm_mul_ps(fscal,dy33);
 564             tz               = _mm_mul_ps(fscal,dz33);
 565
 566             /* Update vectorial force */
 567             fix3             = _mm_add_ps(fix3,tx);
 568             fiy3             = _mm_add_ps(fiy3,ty);
 569             fiz3             = _mm_add_ps(fiz3,tz);
 570
 571             fjx3             = _mm_add_ps(fjx3,tx);
 572             fjy3             = _mm_add_ps(fjy3,ty);
 573             fjz3             = _mm_add_ps(fjz3,tz);
 574
 575             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 576                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 577                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 578                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 579
 580             /* Inner loop uses 287 flops */
 581         }
 582
 583         if(jidx<j_index_end)
 584         {
 585
 586             /* Get j neighbor index, and coordinate index */
 587             jnrA             = jjnr[jidx];
 588             jnrB             = jjnr[jidx+1];
 589             jnrC             = jjnr[jidx+2];
 590             jnrD             = jjnr[jidx+3];
 591
 592             /* Sign of each element will be negative for non-real atoms.
 593              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 594              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 595              */
 596             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 597             jnrA       = (jnrA>=0) ? jnrA : 0;
 598             jnrB       = (jnrB>=0) ? jnrB : 0;
 599             jnrC       = (jnrC>=0) ? jnrC : 0;
 600             jnrD       = (jnrD>=0) ? jnrD : 0;
 601
 602             j_coord_offsetA  = DIM*jnrA;
 603             j_coord_offsetB  = DIM*jnrB;
 604             j_coord_offsetC  = DIM*jnrC;
 605             j_coord_offsetD  = DIM*jnrD;
 606
 607             /* load j atom coordinates */
 608             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 609                                               x+j_coord_offsetC,x+j_coord_offsetD,
 610                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 611                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 612
 613             /* Calculate displacement vector */
 614             dx00             = _mm_sub_ps(ix0,jx0);
 615             dy00             = _mm_sub_ps(iy0,jy0);
 616             dz00             = _mm_sub_ps(iz0,jz0);
 617             dx11             = _mm_sub_ps(ix1,jx1);
 618             dy11             = _mm_sub_ps(iy1,jy1);
 619             dz11             = _mm_sub_ps(iz1,jz1);
 620             dx12             = _mm_sub_ps(ix1,jx2);
 621             dy12             = _mm_sub_ps(iy1,jy2);
 622             dz12             = _mm_sub_ps(iz1,jz2);
 623             dx13             = _mm_sub_ps(ix1,jx3);
 624             dy13             = _mm_sub_ps(iy1,jy3);
 625             dz13             = _mm_sub_ps(iz1,jz3);
 626             dx21             = _mm_sub_ps(ix2,jx1);
 627             dy21             = _mm_sub_ps(iy2,jy1);
 628             dz21             = _mm_sub_ps(iz2,jz1);
 629             dx22             = _mm_sub_ps(ix2,jx2);
 630             dy22             = _mm_sub_ps(iy2,jy2);
 631             dz22             = _mm_sub_ps(iz2,jz2);
 632             dx23             = _mm_sub_ps(ix2,jx3);
 633             dy23             = _mm_sub_ps(iy2,jy3);
 634             dz23             = _mm_sub_ps(iz2,jz3);
 635             dx31             = _mm_sub_ps(ix3,jx1);
 636             dy31             = _mm_sub_ps(iy3,jy1);
 637             dz31             = _mm_sub_ps(iz3,jz1);
 638             dx32             = _mm_sub_ps(ix3,jx2);
 639             dy32             = _mm_sub_ps(iy3,jy2);
 640             dz32             = _mm_sub_ps(iz3,jz2);
 641             dx33             = _mm_sub_ps(ix3,jx3);
 642             dy33             = _mm_sub_ps(iy3,jy3);
 643             dz33             = _mm_sub_ps(iz3,jz3);
 644
 645             /* Calculate squared distance and things based on it */
 646             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 647             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 648             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 649             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 650             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 651             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 652             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 653             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 654             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 655             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 656
 657             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 658             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 659             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 660             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 661             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 662             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 663             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 664             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 665             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 666
 667             rinvsq00         = gmx_mm_inv_ps(rsq00);
 668             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 669             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 670             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 671             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 672             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 673             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 674             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 675             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 676             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 677
 678             fjx0             = _mm_setzero_ps();
 679             fjy0             = _mm_setzero_ps();
 680             fjz0             = _mm_setzero_ps();
 681             fjx1             = _mm_setzero_ps();
 682             fjy1             = _mm_setzero_ps();
 683             fjz1             = _mm_setzero_ps();
 684             fjx2             = _mm_setzero_ps();
 685             fjy2             = _mm_setzero_ps();
 686             fjz2             = _mm_setzero_ps();
 687             fjx3             = _mm_setzero_ps();
 688             fjy3             = _mm_setzero_ps();
 689             fjz3             = _mm_setzero_ps();
 690
 691             /**************************
 692              * CALCULATE INTERACTIONS *
 693              **************************/
 694
 695             /* LENNARD-JONES DISPERSION/REPULSION */
 696
 697             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 698             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 699             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 700             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 701             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 702
 703             /* Update potential sum for this i atom from the interaction with this j atom. */
 704             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 705             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 706
 707             fscal            = fvdw;
 708
 709             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 710
 711             /* Calculate temporary vectorial force */
 712             tx               = _mm_mul_ps(fscal,dx00);
 713             ty               = _mm_mul_ps(fscal,dy00);
 714             tz               = _mm_mul_ps(fscal,dz00);
 715
 716             /* Update vectorial force */
 717             fix0             = _mm_add_ps(fix0,tx);
 718             fiy0             = _mm_add_ps(fiy0,ty);
 719             fiz0             = _mm_add_ps(fiz0,tz);
 720
 721             fjx0             = _mm_add_ps(fjx0,tx);
 722             fjy0             = _mm_add_ps(fjy0,ty);
 723             fjz0             = _mm_add_ps(fjz0,tz);
 724
 725             /**************************
 726              * CALCULATE INTERACTIONS *
 727              **************************/
 728
 729             /* COULOMB ELECTROSTATICS */
 730             velec            = _mm_mul_ps(qq11,rinv11);
 731             felec            = _mm_mul_ps(velec,rinvsq11);
 732
 733             /* Update potential sum for this i atom from the interaction with this j atom. */
 734             velec            = _mm_andnot_ps(dummy_mask,velec);
 735             velecsum         = _mm_add_ps(velecsum,velec);
 736
 737             fscal            = felec;
 738
 739             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 740
 741             /* Calculate temporary vectorial force */
 742             tx               = _mm_mul_ps(fscal,dx11);
 743             ty               = _mm_mul_ps(fscal,dy11);
 744             tz               = _mm_mul_ps(fscal,dz11);
 745
 746             /* Update vectorial force */
 747             fix1             = _mm_add_ps(fix1,tx);
 748             fiy1             = _mm_add_ps(fiy1,ty);
 749             fiz1             = _mm_add_ps(fiz1,tz);
 750
 751             fjx1             = _mm_add_ps(fjx1,tx);
 752             fjy1             = _mm_add_ps(fjy1,ty);
 753             fjz1             = _mm_add_ps(fjz1,tz);
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             /* COULOMB ELECTROSTATICS */
 760             velec            = _mm_mul_ps(qq12,rinv12);
 761             felec            = _mm_mul_ps(velec,rinvsq12);
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _mm_andnot_ps(dummy_mask,velec);
 765             velecsum         = _mm_add_ps(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 770
 771             /* Calculate temporary vectorial force */
 772             tx               = _mm_mul_ps(fscal,dx12);
 773             ty               = _mm_mul_ps(fscal,dy12);
 774             tz               = _mm_mul_ps(fscal,dz12);
 775
 776             /* Update vectorial force */
 777             fix1             = _mm_add_ps(fix1,tx);
 778             fiy1             = _mm_add_ps(fiy1,ty);
 779             fiz1             = _mm_add_ps(fiz1,tz);
 780
 781             fjx2             = _mm_add_ps(fjx2,tx);
 782             fjy2             = _mm_add_ps(fjy2,ty);
 783             fjz2             = _mm_add_ps(fjz2,tz);
 784
 785             /**************************
 786              * CALCULATE INTERACTIONS *
 787              **************************/
 788
 789             /* COULOMB ELECTROSTATICS */
 790             velec            = _mm_mul_ps(qq13,rinv13);
 791             felec            = _mm_mul_ps(velec,rinvsq13);
 792
 793             /* Update potential sum for this i atom from the interaction with this j atom. */
 794             velec            = _mm_andnot_ps(dummy_mask,velec);
 795             velecsum         = _mm_add_ps(velecsum,velec);
 796
 797             fscal            = felec;
 798
 799             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 800
 801             /* Calculate temporary vectorial force */
 802             tx               = _mm_mul_ps(fscal,dx13);
 803             ty               = _mm_mul_ps(fscal,dy13);
 804             tz               = _mm_mul_ps(fscal,dz13);
 805
 806             /* Update vectorial force */
 807             fix1             = _mm_add_ps(fix1,tx);
 808             fiy1             = _mm_add_ps(fiy1,ty);
 809             fiz1             = _mm_add_ps(fiz1,tz);
 810
 811             fjx3             = _mm_add_ps(fjx3,tx);
 812             fjy3             = _mm_add_ps(fjy3,ty);
 813             fjz3             = _mm_add_ps(fjz3,tz);
 814
 815             /**************************
 816              * CALCULATE INTERACTIONS *
 817              **************************/
 818
 819             /* COULOMB ELECTROSTATICS */
 820             velec            = _mm_mul_ps(qq21,rinv21);
 821             felec            = _mm_mul_ps(velec,rinvsq21);
 822
 823             /* Update potential sum for this i atom from the interaction with this j atom. */
 824             velec            = _mm_andnot_ps(dummy_mask,velec);
 825             velecsum         = _mm_add_ps(velecsum,velec);
 826
 827             fscal            = felec;
 828
 829             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 830
 831             /* Calculate temporary vectorial force */
 832             tx               = _mm_mul_ps(fscal,dx21);
 833             ty               = _mm_mul_ps(fscal,dy21);
 834             tz               = _mm_mul_ps(fscal,dz21);
 835
 836             /* Update vectorial force */
 837             fix2             = _mm_add_ps(fix2,tx);
 838             fiy2             = _mm_add_ps(fiy2,ty);
 839             fiz2             = _mm_add_ps(fiz2,tz);
 840
 841             fjx1             = _mm_add_ps(fjx1,tx);
 842             fjy1             = _mm_add_ps(fjy1,ty);
 843             fjz1             = _mm_add_ps(fjz1,tz);
 844
 845             /**************************
 846              * CALCULATE INTERACTIONS *
 847              **************************/
 848
 849             /* COULOMB ELECTROSTATICS */
 850             velec            = _mm_mul_ps(qq22,rinv22);
 851             felec            = _mm_mul_ps(velec,rinvsq22);
 852
 853             /* Update potential sum for this i atom from the interaction with this j atom. */
 854             velec            = _mm_andnot_ps(dummy_mask,velec);
 855             velecsum         = _mm_add_ps(velecsum,velec);
 856
 857             fscal            = felec;
 858
 859             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 860
 861             /* Calculate temporary vectorial force */
 862             tx               = _mm_mul_ps(fscal,dx22);
 863             ty               = _mm_mul_ps(fscal,dy22);
 864             tz               = _mm_mul_ps(fscal,dz22);
 865
 866             /* Update vectorial force */
 867             fix2             = _mm_add_ps(fix2,tx);
 868             fiy2             = _mm_add_ps(fiy2,ty);
 869             fiz2             = _mm_add_ps(fiz2,tz);
 870
 871             fjx2             = _mm_add_ps(fjx2,tx);
 872             fjy2             = _mm_add_ps(fjy2,ty);
 873             fjz2             = _mm_add_ps(fjz2,tz);
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             /* COULOMB ELECTROSTATICS */
 880             velec            = _mm_mul_ps(qq23,rinv23);
 881             felec            = _mm_mul_ps(velec,rinvsq23);
 882
 883             /* Update potential sum for this i atom from the interaction with this j atom. */
 884             velec            = _mm_andnot_ps(dummy_mask,velec);
 885             velecsum         = _mm_add_ps(velecsum,velec);
 886
 887             fscal            = felec;
 888
 889             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 890
 891             /* Calculate temporary vectorial force */
 892             tx               = _mm_mul_ps(fscal,dx23);
 893             ty               = _mm_mul_ps(fscal,dy23);
 894             tz               = _mm_mul_ps(fscal,dz23);
 895
 896             /* Update vectorial force */
 897             fix2             = _mm_add_ps(fix2,tx);
 898             fiy2             = _mm_add_ps(fiy2,ty);
 899             fiz2             = _mm_add_ps(fiz2,tz);
 900
 901             fjx3             = _mm_add_ps(fjx3,tx);
 902             fjy3             = _mm_add_ps(fjy3,ty);
 903             fjz3             = _mm_add_ps(fjz3,tz);
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             /* COULOMB ELECTROSTATICS */
 910             velec            = _mm_mul_ps(qq31,rinv31);
 911             felec            = _mm_mul_ps(velec,rinvsq31);
 912
 913             /* Update potential sum for this i atom from the interaction with this j atom. */
 914             velec            = _mm_andnot_ps(dummy_mask,velec);
 915             velecsum         = _mm_add_ps(velecsum,velec);
 916
 917             fscal            = felec;
 918
 919             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 920
 921             /* Calculate temporary vectorial force */
 922             tx               = _mm_mul_ps(fscal,dx31);
 923             ty               = _mm_mul_ps(fscal,dy31);
 924             tz               = _mm_mul_ps(fscal,dz31);
 925
 926             /* Update vectorial force */
 927             fix3             = _mm_add_ps(fix3,tx);
 928             fiy3             = _mm_add_ps(fiy3,ty);
 929             fiz3             = _mm_add_ps(fiz3,tz);
 930
 931             fjx1             = _mm_add_ps(fjx1,tx);
 932             fjy1             = _mm_add_ps(fjy1,ty);
 933             fjz1             = _mm_add_ps(fjz1,tz);
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             /* COULOMB ELECTROSTATICS */
 940             velec            = _mm_mul_ps(qq32,rinv32);
 941             felec            = _mm_mul_ps(velec,rinvsq32);
 942
 943             /* Update potential sum for this i atom from the interaction with this j atom. */
 944             velec            = _mm_andnot_ps(dummy_mask,velec);
 945             velecsum         = _mm_add_ps(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 950
 951             /* Calculate temporary vectorial force */
 952             tx               = _mm_mul_ps(fscal,dx32);
 953             ty               = _mm_mul_ps(fscal,dy32);
 954             tz               = _mm_mul_ps(fscal,dz32);
 955
 956             /* Update vectorial force */
 957             fix3             = _mm_add_ps(fix3,tx);
 958             fiy3             = _mm_add_ps(fiy3,ty);
 959             fiz3             = _mm_add_ps(fiz3,tz);
 960
 961             fjx2             = _mm_add_ps(fjx2,tx);
 962             fjy2             = _mm_add_ps(fjy2,ty);
 963             fjz2             = _mm_add_ps(fjz2,tz);
 964
 965             /**************************
 966              * CALCULATE INTERACTIONS *
 967              **************************/
 968
 969             /* COULOMB ELECTROSTATICS */
 970             velec            = _mm_mul_ps(qq33,rinv33);
 971             felec            = _mm_mul_ps(velec,rinvsq33);
 972
 973             /* Update potential sum for this i atom from the interaction with this j atom. */
 974             velec            = _mm_andnot_ps(dummy_mask,velec);
 975             velecsum         = _mm_add_ps(velecsum,velec);
 976
 977             fscal            = felec;
 978
 979             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 980
 981             /* Calculate temporary vectorial force */
 982             tx               = _mm_mul_ps(fscal,dx33);
 983             ty               = _mm_mul_ps(fscal,dy33);
 984             tz               = _mm_mul_ps(fscal,dz33);
 985
 986             /* Update vectorial force */
 987             fix3             = _mm_add_ps(fix3,tx);
 988             fiy3             = _mm_add_ps(fiy3,ty);
 989             fiz3             = _mm_add_ps(fiz3,tz);
 990
 991             fjx3             = _mm_add_ps(fjx3,tx);
 992             fjy3             = _mm_add_ps(fjy3,ty);
 993             fjz3             = _mm_add_ps(fjz3,tz);
 994
 995             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 996                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 997                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 998                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 999
1000             /* Inner loop uses 287 flops */
1001         }
1002
1003         /* End of innermost loop */
1004
1005         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1006                                               f+i_coord_offset,fshift+i_shift_offset);
1007
1008         ggid                        = gid[iidx];
1009         /* Update potential energies */
1010         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1011         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1012
1013         /* Increment number of inner iterations */
1014         inneriter                  += j_index_end - j_index_start;
1015
1016         /* Outer loop uses 38 flops */
1017     }
1018
1019     /* Increment number of outer iterations */
1020     outeriter        += nri;
1021
1022     /* Update outer/inner flops */
1023
1024     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*38 + inneriter*287);
1025 }
1026 /*
1027  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse2_single
1028  * Electrostatics interaction: Coulomb
1029  * VdW interaction:            LennardJones
1030  * Geometry:                   Water4-Water4
1031  * Calculate force/pot:        Force
1032  */
1033 void
1034 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse2_single
1035                     (t_nblist * gmx_restrict                nlist,
1036                      rvec * gmx_restrict                    xx,
1037                      rvec * gmx_restrict                    ff,
1038                      t_forcerec * gmx_restrict              fr,
1039                      t_mdatoms * gmx_restrict               mdatoms,
1040                      nb_kernel_data_t * gmx_restrict        kernel_data,
1041                      t_nrnb * gmx_restrict                  nrnb)
1042 {
1043     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1044      * just 0 for non-waters.
1045      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1046      * jnr indices corresponding to data put in the four positions in the SIMD register.
1047      */
1048     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1049     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1050     int              jnrA,jnrB,jnrC,jnrD;
1051     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1052     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1053     real             shX,shY,shZ,rcutoff_scalar;
1054     real             *shiftvec,*fshift,*x,*f;
1055     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1056     int              vdwioffset0;
1057     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1058     int              vdwioffset1;
1059     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1060     int              vdwioffset2;
1061     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1062     int              vdwioffset3;
1063     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1064     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1065     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1066     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1067     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1068     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1069     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1070     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1071     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1072     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1073     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1074     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1075     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1076     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1077     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1078     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1079     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1080     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1081     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1082     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1083     real             *charge;
1084     int              nvdwtype;
1085     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1086     int              *vdwtype;
1087     real             *vdwparam;
1088     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1089     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1090     __m128           dummy_mask,cutoff_mask;
1091     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1092     __m128           one     = _mm_set1_ps(1.0);
1093     __m128           two     = _mm_set1_ps(2.0);
1094     x                = xx[0];
1095     f                = ff[0];
1096
1097     nri              = nlist->nri;
1098     iinr             = nlist->iinr;
1099     jindex           = nlist->jindex;
1100     jjnr             = nlist->jjnr;
1101     shiftidx         = nlist->shift;
1102     gid              = nlist->gid;
1103     shiftvec         = fr->shift_vec[0];
1104     fshift           = fr->fshift[0];
1105     facel            = _mm_set1_ps(fr->epsfac);
1106     charge           = mdatoms->chargeA;
1107     nvdwtype         = fr->ntype;
1108     vdwparam         = fr->nbfp;
1109     vdwtype          = mdatoms->typeA;
1110
1111     /* Setup water-specific parameters */
1112     inr              = nlist->iinr[0];
1113     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1114     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1115     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1116     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1117
1118     jq1              = _mm_set1_ps(charge[inr+1]);
1119     jq2              = _mm_set1_ps(charge[inr+2]);
1120     jq3              = _mm_set1_ps(charge[inr+3]);
1121     vdwjidx0A        = 2*vdwtype[inr+0];
1122     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1123     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1124     qq11             = _mm_mul_ps(iq1,jq1);
1125     qq12             = _mm_mul_ps(iq1,jq2);
1126     qq13             = _mm_mul_ps(iq1,jq3);
1127     qq21             = _mm_mul_ps(iq2,jq1);
1128     qq22             = _mm_mul_ps(iq2,jq2);
1129     qq23             = _mm_mul_ps(iq2,jq3);
1130     qq31             = _mm_mul_ps(iq3,jq1);
1131     qq32             = _mm_mul_ps(iq3,jq2);
1132     qq33             = _mm_mul_ps(iq3,jq3);
1133
1134     /* Avoid stupid compiler warnings */
1135     jnrA = jnrB = jnrC = jnrD = 0;
1136     j_coord_offsetA = 0;
1137     j_coord_offsetB = 0;
1138     j_coord_offsetC = 0;
1139     j_coord_offsetD = 0;
1140
1141     outeriter        = 0;
1142     inneriter        = 0;
1143
1144     /* Start outer loop over neighborlists */
1145     for(iidx=0; iidx<nri; iidx++)
1146     {
1147         /* Load shift vector for this list */
1148         i_shift_offset   = DIM*shiftidx[iidx];
1149         shX              = shiftvec[i_shift_offset+XX];
1150         shY              = shiftvec[i_shift_offset+YY];
1151         shZ              = shiftvec[i_shift_offset+ZZ];
1152
1153         /* Load limits for loop over neighbors */
1154         j_index_start    = jindex[iidx];
1155         j_index_end      = jindex[iidx+1];
1156
1157         /* Get outer coordinate index */
1158         inr              = iinr[iidx];
1159         i_coord_offset   = DIM*inr;
1160
1161         /* Load i particle coords and add shift vector */
1162         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1163         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1164         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1165         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1166         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1167         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1168         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1169         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1170         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1171         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1172         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1173         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1174
1175         fix0             = _mm_setzero_ps();
1176         fiy0             = _mm_setzero_ps();
1177         fiz0             = _mm_setzero_ps();
1178         fix1             = _mm_setzero_ps();
1179         fiy1             = _mm_setzero_ps();
1180         fiz1             = _mm_setzero_ps();
1181         fix2             = _mm_setzero_ps();
1182         fiy2             = _mm_setzero_ps();
1183         fiz2             = _mm_setzero_ps();
1184         fix3             = _mm_setzero_ps();
1185         fiy3             = _mm_setzero_ps();
1186         fiz3             = _mm_setzero_ps();
1187
1188         /* Start inner kernel loop */
1189         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1190         {
1191
1192             /* Get j neighbor index, and coordinate index */
1193             jnrA             = jjnr[jidx];
1194             jnrB             = jjnr[jidx+1];
1195             jnrC             = jjnr[jidx+2];
1196             jnrD             = jjnr[jidx+3];
1197
1198             j_coord_offsetA  = DIM*jnrA;
1199             j_coord_offsetB  = DIM*jnrB;
1200             j_coord_offsetC  = DIM*jnrC;
1201             j_coord_offsetD  = DIM*jnrD;
1202
1203             /* load j atom coordinates */
1204             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1205                                               x+j_coord_offsetC,x+j_coord_offsetD,
1206                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1207                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1208
1209             /* Calculate displacement vector */
1210             dx00             = _mm_sub_ps(ix0,jx0);
1211             dy00             = _mm_sub_ps(iy0,jy0);
1212             dz00             = _mm_sub_ps(iz0,jz0);
1213             dx11             = _mm_sub_ps(ix1,jx1);
1214             dy11             = _mm_sub_ps(iy1,jy1);
1215             dz11             = _mm_sub_ps(iz1,jz1);
1216             dx12             = _mm_sub_ps(ix1,jx2);
1217             dy12             = _mm_sub_ps(iy1,jy2);
1218             dz12             = _mm_sub_ps(iz1,jz2);
1219             dx13             = _mm_sub_ps(ix1,jx3);
1220             dy13             = _mm_sub_ps(iy1,jy3);
1221             dz13             = _mm_sub_ps(iz1,jz3);
1222             dx21             = _mm_sub_ps(ix2,jx1);
1223             dy21             = _mm_sub_ps(iy2,jy1);
1224             dz21             = _mm_sub_ps(iz2,jz1);
1225             dx22             = _mm_sub_ps(ix2,jx2);
1226             dy22             = _mm_sub_ps(iy2,jy2);
1227             dz22             = _mm_sub_ps(iz2,jz2);
1228             dx23             = _mm_sub_ps(ix2,jx3);
1229             dy23             = _mm_sub_ps(iy2,jy3);
1230             dz23             = _mm_sub_ps(iz2,jz3);
1231             dx31             = _mm_sub_ps(ix3,jx1);
1232             dy31             = _mm_sub_ps(iy3,jy1);
1233             dz31             = _mm_sub_ps(iz3,jz1);
1234             dx32             = _mm_sub_ps(ix3,jx2);
1235             dy32             = _mm_sub_ps(iy3,jy2);
1236             dz32             = _mm_sub_ps(iz3,jz2);
1237             dx33             = _mm_sub_ps(ix3,jx3);
1238             dy33             = _mm_sub_ps(iy3,jy3);
1239             dz33             = _mm_sub_ps(iz3,jz3);
1240
1241             /* Calculate squared distance and things based on it */
1242             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1243             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1244             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1245             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1246             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1247             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1248             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1249             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1250             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1251             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1252
1253             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1254             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1255             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1256             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1257             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1258             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1259             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1260             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1261             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1262
1263             rinvsq00         = gmx_mm_inv_ps(rsq00);
1264             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1265             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1266             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1267             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1268             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1269             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1270             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1271             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1272             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1273
1274             fjx0             = _mm_setzero_ps();
1275             fjy0             = _mm_setzero_ps();
1276             fjz0             = _mm_setzero_ps();
1277             fjx1             = _mm_setzero_ps();
1278             fjy1             = _mm_setzero_ps();
1279             fjz1             = _mm_setzero_ps();
1280             fjx2             = _mm_setzero_ps();
1281             fjy2             = _mm_setzero_ps();
1282             fjz2             = _mm_setzero_ps();
1283             fjx3             = _mm_setzero_ps();
1284             fjy3             = _mm_setzero_ps();
1285             fjz3             = _mm_setzero_ps();
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             /* LENNARD-JONES DISPERSION/REPULSION */
1292
1293             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1294             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1295
1296             fscal            = fvdw;
1297
1298             /* Calculate temporary vectorial force */
1299             tx               = _mm_mul_ps(fscal,dx00);
1300             ty               = _mm_mul_ps(fscal,dy00);
1301             tz               = _mm_mul_ps(fscal,dz00);
1302
1303             /* Update vectorial force */
1304             fix0             = _mm_add_ps(fix0,tx);
1305             fiy0             = _mm_add_ps(fiy0,ty);
1306             fiz0             = _mm_add_ps(fiz0,tz);
1307
1308             fjx0             = _mm_add_ps(fjx0,tx);
1309             fjy0             = _mm_add_ps(fjy0,ty);
1310             fjz0             = _mm_add_ps(fjz0,tz);
1311
1312             /**************************
1313              * CALCULATE INTERACTIONS *
1314              **************************/
1315
1316             /* COULOMB ELECTROSTATICS */
1317             velec            = _mm_mul_ps(qq11,rinv11);
1318             felec            = _mm_mul_ps(velec,rinvsq11);
1319
1320             fscal            = felec;
1321
1322             /* Calculate temporary vectorial force */
1323             tx               = _mm_mul_ps(fscal,dx11);
1324             ty               = _mm_mul_ps(fscal,dy11);
1325             tz               = _mm_mul_ps(fscal,dz11);
1326
1327             /* Update vectorial force */
1328             fix1             = _mm_add_ps(fix1,tx);
1329             fiy1             = _mm_add_ps(fiy1,ty);
1330             fiz1             = _mm_add_ps(fiz1,tz);
1331
1332             fjx1             = _mm_add_ps(fjx1,tx);
1333             fjy1             = _mm_add_ps(fjy1,ty);
1334             fjz1             = _mm_add_ps(fjz1,tz);
1335
1336             /**************************
1337              * CALCULATE INTERACTIONS *
1338              **************************/
1339
1340             /* COULOMB ELECTROSTATICS */
1341             velec            = _mm_mul_ps(qq12,rinv12);
1342             felec            = _mm_mul_ps(velec,rinvsq12);
1343
1344             fscal            = felec;
1345
1346             /* Calculate temporary vectorial force */
1347             tx               = _mm_mul_ps(fscal,dx12);
1348             ty               = _mm_mul_ps(fscal,dy12);
1349             tz               = _mm_mul_ps(fscal,dz12);
1350
1351             /* Update vectorial force */
1352             fix1             = _mm_add_ps(fix1,tx);
1353             fiy1             = _mm_add_ps(fiy1,ty);
1354             fiz1             = _mm_add_ps(fiz1,tz);
1355
1356             fjx2             = _mm_add_ps(fjx2,tx);
1357             fjy2             = _mm_add_ps(fjy2,ty);
1358             fjz2             = _mm_add_ps(fjz2,tz);
1359
1360             /**************************
1361              * CALCULATE INTERACTIONS *
1362              **************************/
1363
1364             /* COULOMB ELECTROSTATICS */
1365             velec            = _mm_mul_ps(qq13,rinv13);
1366             felec            = _mm_mul_ps(velec,rinvsq13);
1367
1368             fscal            = felec;
1369
1370             /* Calculate temporary vectorial force */
1371             tx               = _mm_mul_ps(fscal,dx13);
1372             ty               = _mm_mul_ps(fscal,dy13);
1373             tz               = _mm_mul_ps(fscal,dz13);
1374
1375             /* Update vectorial force */
1376             fix1             = _mm_add_ps(fix1,tx);
1377             fiy1             = _mm_add_ps(fiy1,ty);
1378             fiz1             = _mm_add_ps(fiz1,tz);
1379
1380             fjx3             = _mm_add_ps(fjx3,tx);
1381             fjy3             = _mm_add_ps(fjy3,ty);
1382             fjz3             = _mm_add_ps(fjz3,tz);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             /* COULOMB ELECTROSTATICS */
1389             velec            = _mm_mul_ps(qq21,rinv21);
1390             felec            = _mm_mul_ps(velec,rinvsq21);
1391
1392             fscal            = felec;
1393
1394             /* Calculate temporary vectorial force */
1395             tx               = _mm_mul_ps(fscal,dx21);
1396             ty               = _mm_mul_ps(fscal,dy21);
1397             tz               = _mm_mul_ps(fscal,dz21);
1398
1399             /* Update vectorial force */
1400             fix2             = _mm_add_ps(fix2,tx);
1401             fiy2             = _mm_add_ps(fiy2,ty);
1402             fiz2             = _mm_add_ps(fiz2,tz);
1403
1404             fjx1             = _mm_add_ps(fjx1,tx);
1405             fjy1             = _mm_add_ps(fjy1,ty);
1406             fjz1             = _mm_add_ps(fjz1,tz);
1407
1408             /**************************
1409              * CALCULATE INTERACTIONS *
1410              **************************/
1411
1412             /* COULOMB ELECTROSTATICS */
1413             velec            = _mm_mul_ps(qq22,rinv22);
1414             felec            = _mm_mul_ps(velec,rinvsq22);
1415
1416             fscal            = felec;
1417
1418             /* Calculate temporary vectorial force */
1419             tx               = _mm_mul_ps(fscal,dx22);
1420             ty               = _mm_mul_ps(fscal,dy22);
1421             tz               = _mm_mul_ps(fscal,dz22);
1422
1423             /* Update vectorial force */
1424             fix2             = _mm_add_ps(fix2,tx);
1425             fiy2             = _mm_add_ps(fiy2,ty);
1426             fiz2             = _mm_add_ps(fiz2,tz);
1427
1428             fjx2             = _mm_add_ps(fjx2,tx);
1429             fjy2             = _mm_add_ps(fjy2,ty);
1430             fjz2             = _mm_add_ps(fjz2,tz);
1431
1432             /**************************
1433              * CALCULATE INTERACTIONS *
1434              **************************/
1435
1436             /* COULOMB ELECTROSTATICS */
1437             velec            = _mm_mul_ps(qq23,rinv23);
1438             felec            = _mm_mul_ps(velec,rinvsq23);
1439
1440             fscal            = felec;
1441
1442             /* Calculate temporary vectorial force */
1443             tx               = _mm_mul_ps(fscal,dx23);
1444             ty               = _mm_mul_ps(fscal,dy23);
1445             tz               = _mm_mul_ps(fscal,dz23);
1446
1447             /* Update vectorial force */
1448             fix2             = _mm_add_ps(fix2,tx);
1449             fiy2             = _mm_add_ps(fiy2,ty);
1450             fiz2             = _mm_add_ps(fiz2,tz);
1451
1452             fjx3             = _mm_add_ps(fjx3,tx);
1453             fjy3             = _mm_add_ps(fjy3,ty);
1454             fjz3             = _mm_add_ps(fjz3,tz);
1455
1456             /**************************
1457              * CALCULATE INTERACTIONS *
1458              **************************/
1459
1460             /* COULOMB ELECTROSTATICS */
1461             velec            = _mm_mul_ps(qq31,rinv31);
1462             felec            = _mm_mul_ps(velec,rinvsq31);
1463
1464             fscal            = felec;
1465
1466             /* Calculate temporary vectorial force */
1467             tx               = _mm_mul_ps(fscal,dx31);
1468             ty               = _mm_mul_ps(fscal,dy31);
1469             tz               = _mm_mul_ps(fscal,dz31);
1470
1471             /* Update vectorial force */
1472             fix3             = _mm_add_ps(fix3,tx);
1473             fiy3             = _mm_add_ps(fiy3,ty);
1474             fiz3             = _mm_add_ps(fiz3,tz);
1475
1476             fjx1             = _mm_add_ps(fjx1,tx);
1477             fjy1             = _mm_add_ps(fjy1,ty);
1478             fjz1             = _mm_add_ps(fjz1,tz);
1479
1480             /**************************
1481              * CALCULATE INTERACTIONS *
1482              **************************/
1483
1484             /* COULOMB ELECTROSTATICS */
1485             velec            = _mm_mul_ps(qq32,rinv32);
1486             felec            = _mm_mul_ps(velec,rinvsq32);
1487
1488             fscal            = felec;
1489
1490             /* Calculate temporary vectorial force */
1491             tx               = _mm_mul_ps(fscal,dx32);
1492             ty               = _mm_mul_ps(fscal,dy32);
1493             tz               = _mm_mul_ps(fscal,dz32);
1494
1495             /* Update vectorial force */
1496             fix3             = _mm_add_ps(fix3,tx);
1497             fiy3             = _mm_add_ps(fiy3,ty);
1498             fiz3             = _mm_add_ps(fiz3,tz);
1499
1500             fjx2             = _mm_add_ps(fjx2,tx);
1501             fjy2             = _mm_add_ps(fjy2,ty);
1502             fjz2             = _mm_add_ps(fjz2,tz);
1503
1504             /**************************
1505              * CALCULATE INTERACTIONS *
1506              **************************/
1507
1508             /* COULOMB ELECTROSTATICS */
1509             velec            = _mm_mul_ps(qq33,rinv33);
1510             felec            = _mm_mul_ps(velec,rinvsq33);
1511
1512             fscal            = felec;
1513
1514             /* Calculate temporary vectorial force */
1515             tx               = _mm_mul_ps(fscal,dx33);
1516             ty               = _mm_mul_ps(fscal,dy33);
1517             tz               = _mm_mul_ps(fscal,dz33);
1518
1519             /* Update vectorial force */
1520             fix3             = _mm_add_ps(fix3,tx);
1521             fiy3             = _mm_add_ps(fiy3,ty);
1522             fiz3             = _mm_add_ps(fiz3,tz);
1523
1524             fjx3             = _mm_add_ps(fjx3,tx);
1525             fjy3             = _mm_add_ps(fjy3,ty);
1526             fjz3             = _mm_add_ps(fjz3,tz);
1527
1528             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1529                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1530                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1531                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1532
1533             /* Inner loop uses 273 flops */
1534         }
1535
1536         if(jidx<j_index_end)
1537         {
1538
1539             /* Get j neighbor index, and coordinate index */
1540             jnrA             = jjnr[jidx];
1541             jnrB             = jjnr[jidx+1];
1542             jnrC             = jjnr[jidx+2];
1543             jnrD             = jjnr[jidx+3];
1544
1545             /* Sign of each element will be negative for non-real atoms.
1546              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1547              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1548              */
1549             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1550             jnrA       = (jnrA>=0) ? jnrA : 0;
1551             jnrB       = (jnrB>=0) ? jnrB : 0;
1552             jnrC       = (jnrC>=0) ? jnrC : 0;
1553             jnrD       = (jnrD>=0) ? jnrD : 0;
1554
1555             j_coord_offsetA  = DIM*jnrA;
1556             j_coord_offsetB  = DIM*jnrB;
1557             j_coord_offsetC  = DIM*jnrC;
1558             j_coord_offsetD  = DIM*jnrD;
1559
1560             /* load j atom coordinates */
1561             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1562                                               x+j_coord_offsetC,x+j_coord_offsetD,
1563                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1564                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1565
1566             /* Calculate displacement vector */
1567             dx00             = _mm_sub_ps(ix0,jx0);
1568             dy00             = _mm_sub_ps(iy0,jy0);
1569             dz00             = _mm_sub_ps(iz0,jz0);
1570             dx11             = _mm_sub_ps(ix1,jx1);
1571             dy11             = _mm_sub_ps(iy1,jy1);
1572             dz11             = _mm_sub_ps(iz1,jz1);
1573             dx12             = _mm_sub_ps(ix1,jx2);
1574             dy12             = _mm_sub_ps(iy1,jy2);
1575             dz12             = _mm_sub_ps(iz1,jz2);
1576             dx13             = _mm_sub_ps(ix1,jx3);
1577             dy13             = _mm_sub_ps(iy1,jy3);
1578             dz13             = _mm_sub_ps(iz1,jz3);
1579             dx21             = _mm_sub_ps(ix2,jx1);
1580             dy21             = _mm_sub_ps(iy2,jy1);
1581             dz21             = _mm_sub_ps(iz2,jz1);
1582             dx22             = _mm_sub_ps(ix2,jx2);
1583             dy22             = _mm_sub_ps(iy2,jy2);
1584             dz22             = _mm_sub_ps(iz2,jz2);
1585             dx23             = _mm_sub_ps(ix2,jx3);
1586             dy23             = _mm_sub_ps(iy2,jy3);
1587             dz23             = _mm_sub_ps(iz2,jz3);
1588             dx31             = _mm_sub_ps(ix3,jx1);
1589             dy31             = _mm_sub_ps(iy3,jy1);
1590             dz31             = _mm_sub_ps(iz3,jz1);
1591             dx32             = _mm_sub_ps(ix3,jx2);
1592             dy32             = _mm_sub_ps(iy3,jy2);
1593             dz32             = _mm_sub_ps(iz3,jz2);
1594             dx33             = _mm_sub_ps(ix3,jx3);
1595             dy33             = _mm_sub_ps(iy3,jy3);
1596             dz33             = _mm_sub_ps(iz3,jz3);
1597
1598             /* Calculate squared distance and things based on it */
1599             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1600             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1601             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1602             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1603             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1604             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1605             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1606             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1607             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1608             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1609
1610             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1611             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1612             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1613             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1614             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1615             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1616             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1617             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1618             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1619
1620             rinvsq00         = gmx_mm_inv_ps(rsq00);
1621             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1622             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1623             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1624             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1625             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1626             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1627             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1628             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1629             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1630
1631             fjx0             = _mm_setzero_ps();
1632             fjy0             = _mm_setzero_ps();
1633             fjz0             = _mm_setzero_ps();
1634             fjx1             = _mm_setzero_ps();
1635             fjy1             = _mm_setzero_ps();
1636             fjz1             = _mm_setzero_ps();
1637             fjx2             = _mm_setzero_ps();
1638             fjy2             = _mm_setzero_ps();
1639             fjz2             = _mm_setzero_ps();
1640             fjx3             = _mm_setzero_ps();
1641             fjy3             = _mm_setzero_ps();
1642             fjz3             = _mm_setzero_ps();
1643
1644             /**************************
1645              * CALCULATE INTERACTIONS *
1646              **************************/
1647
1648             /* LENNARD-JONES DISPERSION/REPULSION */
1649
1650             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1651             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1652
1653             fscal            = fvdw;
1654
1655             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1656
1657             /* Calculate temporary vectorial force */
1658             tx               = _mm_mul_ps(fscal,dx00);
1659             ty               = _mm_mul_ps(fscal,dy00);
1660             tz               = _mm_mul_ps(fscal,dz00);
1661
1662             /* Update vectorial force */
1663             fix0             = _mm_add_ps(fix0,tx);
1664             fiy0             = _mm_add_ps(fiy0,ty);
1665             fiz0             = _mm_add_ps(fiz0,tz);
1666
1667             fjx0             = _mm_add_ps(fjx0,tx);
1668             fjy0             = _mm_add_ps(fjy0,ty);
1669             fjz0             = _mm_add_ps(fjz0,tz);
1670
1671             /**************************
1672              * CALCULATE INTERACTIONS *
1673              **************************/
1674
1675             /* COULOMB ELECTROSTATICS */
1676             velec            = _mm_mul_ps(qq11,rinv11);
1677             felec            = _mm_mul_ps(velec,rinvsq11);
1678
1679             fscal            = felec;
1680
1681             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1682
1683             /* Calculate temporary vectorial force */
1684             tx               = _mm_mul_ps(fscal,dx11);
1685             ty               = _mm_mul_ps(fscal,dy11);
1686             tz               = _mm_mul_ps(fscal,dz11);
1687
1688             /* Update vectorial force */
1689             fix1             = _mm_add_ps(fix1,tx);
1690             fiy1             = _mm_add_ps(fiy1,ty);
1691             fiz1             = _mm_add_ps(fiz1,tz);
1692
1693             fjx1             = _mm_add_ps(fjx1,tx);
1694             fjy1             = _mm_add_ps(fjy1,ty);
1695             fjz1             = _mm_add_ps(fjz1,tz);
1696
1697             /**************************
1698              * CALCULATE INTERACTIONS *
1699              **************************/
1700
1701             /* COULOMB ELECTROSTATICS */
1702             velec            = _mm_mul_ps(qq12,rinv12);
1703             felec            = _mm_mul_ps(velec,rinvsq12);
1704
1705             fscal            = felec;
1706
1707             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1708
1709             /* Calculate temporary vectorial force */
1710             tx               = _mm_mul_ps(fscal,dx12);
1711             ty               = _mm_mul_ps(fscal,dy12);
1712             tz               = _mm_mul_ps(fscal,dz12);
1713
1714             /* Update vectorial force */
1715             fix1             = _mm_add_ps(fix1,tx);
1716             fiy1             = _mm_add_ps(fiy1,ty);
1717             fiz1             = _mm_add_ps(fiz1,tz);
1718
1719             fjx2             = _mm_add_ps(fjx2,tx);
1720             fjy2             = _mm_add_ps(fjy2,ty);
1721             fjz2             = _mm_add_ps(fjz2,tz);
1722
1723             /**************************
1724              * CALCULATE INTERACTIONS *
1725              **************************/
1726
1727             /* COULOMB ELECTROSTATICS */
1728             velec            = _mm_mul_ps(qq13,rinv13);
1729             felec            = _mm_mul_ps(velec,rinvsq13);
1730
1731             fscal            = felec;
1732
1733             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1734
1735             /* Calculate temporary vectorial force */
1736             tx               = _mm_mul_ps(fscal,dx13);
1737             ty               = _mm_mul_ps(fscal,dy13);
1738             tz               = _mm_mul_ps(fscal,dz13);
1739
1740             /* Update vectorial force */
1741             fix1             = _mm_add_ps(fix1,tx);
1742             fiy1             = _mm_add_ps(fiy1,ty);
1743             fiz1             = _mm_add_ps(fiz1,tz);
1744
1745             fjx3             = _mm_add_ps(fjx3,tx);
1746             fjy3             = _mm_add_ps(fjy3,ty);
1747             fjz3             = _mm_add_ps(fjz3,tz);
1748
1749             /**************************
1750              * CALCULATE INTERACTIONS *
1751              **************************/
1752
1753             /* COULOMB ELECTROSTATICS */
1754             velec            = _mm_mul_ps(qq21,rinv21);
1755             felec            = _mm_mul_ps(velec,rinvsq21);
1756
1757             fscal            = felec;
1758
1759             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1760
1761             /* Calculate temporary vectorial force */
1762             tx               = _mm_mul_ps(fscal,dx21);
1763             ty               = _mm_mul_ps(fscal,dy21);
1764             tz               = _mm_mul_ps(fscal,dz21);
1765
1766             /* Update vectorial force */
1767             fix2             = _mm_add_ps(fix2,tx);
1768             fiy2             = _mm_add_ps(fiy2,ty);
1769             fiz2             = _mm_add_ps(fiz2,tz);
1770
1771             fjx1             = _mm_add_ps(fjx1,tx);
1772             fjy1             = _mm_add_ps(fjy1,ty);
1773             fjz1             = _mm_add_ps(fjz1,tz);
1774
1775             /**************************
1776              * CALCULATE INTERACTIONS *
1777              **************************/
1778
1779             /* COULOMB ELECTROSTATICS */
1780             velec            = _mm_mul_ps(qq22,rinv22);
1781             felec            = _mm_mul_ps(velec,rinvsq22);
1782
1783             fscal            = felec;
1784
1785             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1786
1787             /* Calculate temporary vectorial force */
1788             tx               = _mm_mul_ps(fscal,dx22);
1789             ty               = _mm_mul_ps(fscal,dy22);
1790             tz               = _mm_mul_ps(fscal,dz22);
1791
1792             /* Update vectorial force */
1793             fix2             = _mm_add_ps(fix2,tx);
1794             fiy2             = _mm_add_ps(fiy2,ty);
1795             fiz2             = _mm_add_ps(fiz2,tz);
1796
1797             fjx2             = _mm_add_ps(fjx2,tx);
1798             fjy2             = _mm_add_ps(fjy2,ty);
1799             fjz2             = _mm_add_ps(fjz2,tz);
1800
1801             /**************************
1802              * CALCULATE INTERACTIONS *
1803              **************************/
1804
1805             /* COULOMB ELECTROSTATICS */
1806             velec            = _mm_mul_ps(qq23,rinv23);
1807             felec            = _mm_mul_ps(velec,rinvsq23);
1808
1809             fscal            = felec;
1810
1811             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1812
1813             /* Calculate temporary vectorial force */
1814             tx               = _mm_mul_ps(fscal,dx23);
1815             ty               = _mm_mul_ps(fscal,dy23);
1816             tz               = _mm_mul_ps(fscal,dz23);
1817
1818             /* Update vectorial force */
1819             fix2             = _mm_add_ps(fix2,tx);
1820             fiy2             = _mm_add_ps(fiy2,ty);
1821             fiz2             = _mm_add_ps(fiz2,tz);
1822
1823             fjx3             = _mm_add_ps(fjx3,tx);
1824             fjy3             = _mm_add_ps(fjy3,ty);
1825             fjz3             = _mm_add_ps(fjz3,tz);
1826
1827             /**************************
1828              * CALCULATE INTERACTIONS *
1829              **************************/
1830
1831             /* COULOMB ELECTROSTATICS */
1832             velec            = _mm_mul_ps(qq31,rinv31);
1833             felec            = _mm_mul_ps(velec,rinvsq31);
1834
1835             fscal            = felec;
1836
1837             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1838
1839             /* Calculate temporary vectorial force */
1840             tx               = _mm_mul_ps(fscal,dx31);
1841             ty               = _mm_mul_ps(fscal,dy31);
1842             tz               = _mm_mul_ps(fscal,dz31);
1843
1844             /* Update vectorial force */
1845             fix3             = _mm_add_ps(fix3,tx);
1846             fiy3             = _mm_add_ps(fiy3,ty);
1847             fiz3             = _mm_add_ps(fiz3,tz);
1848
1849             fjx1             = _mm_add_ps(fjx1,tx);
1850             fjy1             = _mm_add_ps(fjy1,ty);
1851             fjz1             = _mm_add_ps(fjz1,tz);
1852
1853             /**************************
1854              * CALCULATE INTERACTIONS *
1855              **************************/
1856
1857             /* COULOMB ELECTROSTATICS */
1858             velec            = _mm_mul_ps(qq32,rinv32);
1859             felec            = _mm_mul_ps(velec,rinvsq32);
1860
1861             fscal            = felec;
1862
1863             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1864
1865             /* Calculate temporary vectorial force */
1866             tx               = _mm_mul_ps(fscal,dx32);
1867             ty               = _mm_mul_ps(fscal,dy32);
1868             tz               = _mm_mul_ps(fscal,dz32);
1869
1870             /* Update vectorial force */
1871             fix3             = _mm_add_ps(fix3,tx);
1872             fiy3             = _mm_add_ps(fiy3,ty);
1873             fiz3             = _mm_add_ps(fiz3,tz);
1874
1875             fjx2             = _mm_add_ps(fjx2,tx);
1876             fjy2             = _mm_add_ps(fjy2,ty);
1877             fjz2             = _mm_add_ps(fjz2,tz);
1878
1879             /**************************
1880              * CALCULATE INTERACTIONS *
1881              **************************/
1882
1883             /* COULOMB ELECTROSTATICS */
1884             velec            = _mm_mul_ps(qq33,rinv33);
1885             felec            = _mm_mul_ps(velec,rinvsq33);
1886
1887             fscal            = felec;
1888
1889             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1890
1891             /* Calculate temporary vectorial force */
1892             tx               = _mm_mul_ps(fscal,dx33);
1893             ty               = _mm_mul_ps(fscal,dy33);
1894             tz               = _mm_mul_ps(fscal,dz33);
1895
1896             /* Update vectorial force */
1897             fix3             = _mm_add_ps(fix3,tx);
1898             fiy3             = _mm_add_ps(fiy3,ty);
1899             fiz3             = _mm_add_ps(fiz3,tz);
1900
1901             fjx3             = _mm_add_ps(fjx3,tx);
1902             fjy3             = _mm_add_ps(fjy3,ty);
1903             fjz3             = _mm_add_ps(fjz3,tz);
1904
1905             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1906                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1907                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1908                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1909
1910             /* Inner loop uses 273 flops */
1911         }
1912
1913         /* End of innermost loop */
1914
1915         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1916                                               f+i_coord_offset,fshift+i_shift_offset);
1917
1918         /* Increment number of inner iterations */
1919         inneriter                  += j_index_end - j_index_start;
1920
1921         /* Outer loop uses 36 flops */
1922     }
1923
1924     /* Increment number of outer iterations */
1925     outeriter        += nri;
1926
1927     /* Update outer/inner flops */
1928
1929     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*36 + inneriter*273);
1930 }