src/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              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwioffset3;
  76     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  77     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  78     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  79     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  80     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  81     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  82     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  83     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  84     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  85     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  86     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  87     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  88     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  89     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  90     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  91     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  92     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  93     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  94     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  95     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 102     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 103     __m128           dummy_mask,cutoff_mask;
 104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 105     __m128           one     = _mm_set1_ps(1.0);
 106     __m128           two     = _mm_set1_ps(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_ps(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     /* 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     /* Avoid stupid compiler warnings */
 148     jnrA = jnrB = jnrC = jnrD = 0;
 149     j_coord_offsetA = 0;
 150     j_coord_offsetB = 0;
 151     j_coord_offsetC = 0;
 152     j_coord_offsetD = 0;
 153
 154     outeriter        = 0;
 155     inneriter        = 0;
 156
 157     for(iidx=0;iidx<4*DIM;iidx++)
 158     {
 159         scratch[iidx] = 0.0;
 160     }
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 178                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 179
 180         fix0             = _mm_setzero_ps();
 181         fiy0             = _mm_setzero_ps();
 182         fiz0             = _mm_setzero_ps();
 183         fix1             = _mm_setzero_ps();
 184         fiy1             = _mm_setzero_ps();
 185         fiz1             = _mm_setzero_ps();
 186         fix2             = _mm_setzero_ps();
 187         fiy2             = _mm_setzero_ps();
 188         fiz2             = _mm_setzero_ps();
 189         fix3             = _mm_setzero_ps();
 190         fiy3             = _mm_setzero_ps();
 191         fiz3             = _mm_setzero_ps();
 192
 193         /* Reset potential sums */
 194         velecsum         = _mm_setzero_ps();
 195         vvdwsum          = _mm_setzero_ps();
 196
 197         /* Start inner kernel loop */
 198         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 199         {
 200
 201             /* Get j neighbor index, and coordinate index */
 202             jnrA             = jjnr[jidx];
 203             jnrB             = jjnr[jidx+1];
 204             jnrC             = jjnr[jidx+2];
 205             jnrD             = jjnr[jidx+3];
 206             j_coord_offsetA  = DIM*jnrA;
 207             j_coord_offsetB  = DIM*jnrB;
 208             j_coord_offsetC  = DIM*jnrC;
 209             j_coord_offsetD  = DIM*jnrD;
 210
 211             /* load j atom coordinates */
 212             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 213                                               x+j_coord_offsetC,x+j_coord_offsetD,
 214                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 215                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 216
 217             /* Calculate displacement vector */
 218             dx00             = _mm_sub_ps(ix0,jx0);
 219             dy00             = _mm_sub_ps(iy0,jy0);
 220             dz00             = _mm_sub_ps(iz0,jz0);
 221             dx11             = _mm_sub_ps(ix1,jx1);
 222             dy11             = _mm_sub_ps(iy1,jy1);
 223             dz11             = _mm_sub_ps(iz1,jz1);
 224             dx12             = _mm_sub_ps(ix1,jx2);
 225             dy12             = _mm_sub_ps(iy1,jy2);
 226             dz12             = _mm_sub_ps(iz1,jz2);
 227             dx13             = _mm_sub_ps(ix1,jx3);
 228             dy13             = _mm_sub_ps(iy1,jy3);
 229             dz13             = _mm_sub_ps(iz1,jz3);
 230             dx21             = _mm_sub_ps(ix2,jx1);
 231             dy21             = _mm_sub_ps(iy2,jy1);
 232             dz21             = _mm_sub_ps(iz2,jz1);
 233             dx22             = _mm_sub_ps(ix2,jx2);
 234             dy22             = _mm_sub_ps(iy2,jy2);
 235             dz22             = _mm_sub_ps(iz2,jz2);
 236             dx23             = _mm_sub_ps(ix2,jx3);
 237             dy23             = _mm_sub_ps(iy2,jy3);
 238             dz23             = _mm_sub_ps(iz2,jz3);
 239             dx31             = _mm_sub_ps(ix3,jx1);
 240             dy31             = _mm_sub_ps(iy3,jy1);
 241             dz31             = _mm_sub_ps(iz3,jz1);
 242             dx32             = _mm_sub_ps(ix3,jx2);
 243             dy32             = _mm_sub_ps(iy3,jy2);
 244             dz32             = _mm_sub_ps(iz3,jz2);
 245             dx33             = _mm_sub_ps(ix3,jx3);
 246             dy33             = _mm_sub_ps(iy3,jy3);
 247             dz33             = _mm_sub_ps(iz3,jz3);
 248
 249             /* Calculate squared distance and things based on it */
 250             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 251             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 252             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 253             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 254             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 255             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 256             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 257             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 258             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 259             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 260
 261             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 262             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 263             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 264             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 265             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 266             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 267             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 268             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 269             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 270
 271             rinvsq00         = gmx_mm_inv_ps(rsq00);
 272             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 273             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 274             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 275             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 276             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 277             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 278             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 279             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 280             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 281
 282             fjx0             = _mm_setzero_ps();
 283             fjy0             = _mm_setzero_ps();
 284             fjz0             = _mm_setzero_ps();
 285             fjx1             = _mm_setzero_ps();
 286             fjy1             = _mm_setzero_ps();
 287             fjz1             = _mm_setzero_ps();
 288             fjx2             = _mm_setzero_ps();
 289             fjy2             = _mm_setzero_ps();
 290             fjz2             = _mm_setzero_ps();
 291             fjx3             = _mm_setzero_ps();
 292             fjy3             = _mm_setzero_ps();
 293             fjz3             = _mm_setzero_ps();
 294
 295             /**************************
 296              * CALCULATE INTERACTIONS *
 297              **************************/
 298
 299             /* LENNARD-JONES DISPERSION/REPULSION */
 300
 301             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 302             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 303             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 304             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 305             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 306
 307             /* Update potential sum for this i atom from the interaction with this j atom. */
 308             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 309
 310             fscal            = fvdw;
 311
 312             /* Calculate temporary vectorial force */
 313             tx               = _mm_mul_ps(fscal,dx00);
 314             ty               = _mm_mul_ps(fscal,dy00);
 315             tz               = _mm_mul_ps(fscal,dz00);
 316
 317             /* Update vectorial force */
 318             fix0             = _mm_add_ps(fix0,tx);
 319             fiy0             = _mm_add_ps(fiy0,ty);
 320             fiz0             = _mm_add_ps(fiz0,tz);
 321
 322             fjx0             = _mm_add_ps(fjx0,tx);
 323             fjy0             = _mm_add_ps(fjy0,ty);
 324             fjz0             = _mm_add_ps(fjz0,tz);
 325
 326             /**************************
 327              * CALCULATE INTERACTIONS *
 328              **************************/
 329
 330             /* COULOMB ELECTROSTATICS */
 331             velec            = _mm_mul_ps(qq11,rinv11);
 332             felec            = _mm_mul_ps(velec,rinvsq11);
 333
 334             /* Update potential sum for this i atom from the interaction with this j atom. */
 335             velecsum         = _mm_add_ps(velecsum,velec);
 336
 337             fscal            = felec;
 338
 339             /* Calculate temporary vectorial force */
 340             tx               = _mm_mul_ps(fscal,dx11);
 341             ty               = _mm_mul_ps(fscal,dy11);
 342             tz               = _mm_mul_ps(fscal,dz11);
 343
 344             /* Update vectorial force */
 345             fix1             = _mm_add_ps(fix1,tx);
 346             fiy1             = _mm_add_ps(fiy1,ty);
 347             fiz1             = _mm_add_ps(fiz1,tz);
 348
 349             fjx1             = _mm_add_ps(fjx1,tx);
 350             fjy1             = _mm_add_ps(fjy1,ty);
 351             fjz1             = _mm_add_ps(fjz1,tz);
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             /* COULOMB ELECTROSTATICS */
 358             velec            = _mm_mul_ps(qq12,rinv12);
 359             felec            = _mm_mul_ps(velec,rinvsq12);
 360
 361             /* Update potential sum for this i atom from the interaction with this j atom. */
 362             velecsum         = _mm_add_ps(velecsum,velec);
 363
 364             fscal            = felec;
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = _mm_mul_ps(fscal,dx12);
 368             ty               = _mm_mul_ps(fscal,dy12);
 369             tz               = _mm_mul_ps(fscal,dz12);
 370
 371             /* Update vectorial force */
 372             fix1             = _mm_add_ps(fix1,tx);
 373             fiy1             = _mm_add_ps(fiy1,ty);
 374             fiz1             = _mm_add_ps(fiz1,tz);
 375
 376             fjx2             = _mm_add_ps(fjx2,tx);
 377             fjy2             = _mm_add_ps(fjy2,ty);
 378             fjz2             = _mm_add_ps(fjz2,tz);
 379
 380             /**************************
 381              * CALCULATE INTERACTIONS *
 382              **************************/
 383
 384             /* COULOMB ELECTROSTATICS */
 385             velec            = _mm_mul_ps(qq13,rinv13);
 386             felec            = _mm_mul_ps(velec,rinvsq13);
 387
 388             /* Update potential sum for this i atom from the interaction with this j atom. */
 389             velecsum         = _mm_add_ps(velecsum,velec);
 390
 391             fscal            = felec;
 392
 393             /* Calculate temporary vectorial force */
 394             tx               = _mm_mul_ps(fscal,dx13);
 395             ty               = _mm_mul_ps(fscal,dy13);
 396             tz               = _mm_mul_ps(fscal,dz13);
 397
 398             /* Update vectorial force */
 399             fix1             = _mm_add_ps(fix1,tx);
 400             fiy1             = _mm_add_ps(fiy1,ty);
 401             fiz1             = _mm_add_ps(fiz1,tz);
 402
 403             fjx3             = _mm_add_ps(fjx3,tx);
 404             fjy3             = _mm_add_ps(fjy3,ty);
 405             fjz3             = _mm_add_ps(fjz3,tz);
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             /* COULOMB ELECTROSTATICS */
 412             velec            = _mm_mul_ps(qq21,rinv21);
 413             felec            = _mm_mul_ps(velec,rinvsq21);
 414
 415             /* Update potential sum for this i atom from the interaction with this j atom. */
 416             velecsum         = _mm_add_ps(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             /* Calculate temporary vectorial force */
 421             tx               = _mm_mul_ps(fscal,dx21);
 422             ty               = _mm_mul_ps(fscal,dy21);
 423             tz               = _mm_mul_ps(fscal,dz21);
 424
 425             /* Update vectorial force */
 426             fix2             = _mm_add_ps(fix2,tx);
 427             fiy2             = _mm_add_ps(fiy2,ty);
 428             fiz2             = _mm_add_ps(fiz2,tz);
 429
 430             fjx1             = _mm_add_ps(fjx1,tx);
 431             fjy1             = _mm_add_ps(fjy1,ty);
 432             fjz1             = _mm_add_ps(fjz1,tz);
 433
 434             /**************************
 435              * CALCULATE INTERACTIONS *
 436              **************************/
 437
 438             /* COULOMB ELECTROSTATICS */
 439             velec            = _mm_mul_ps(qq22,rinv22);
 440             felec            = _mm_mul_ps(velec,rinvsq22);
 441
 442             /* Update potential sum for this i atom from the interaction with this j atom. */
 443             velecsum         = _mm_add_ps(velecsum,velec);
 444
 445             fscal            = felec;
 446
 447             /* Calculate temporary vectorial force */
 448             tx               = _mm_mul_ps(fscal,dx22);
 449             ty               = _mm_mul_ps(fscal,dy22);
 450             tz               = _mm_mul_ps(fscal,dz22);
 451
 452             /* Update vectorial force */
 453             fix2             = _mm_add_ps(fix2,tx);
 454             fiy2             = _mm_add_ps(fiy2,ty);
 455             fiz2             = _mm_add_ps(fiz2,tz);
 456
 457             fjx2             = _mm_add_ps(fjx2,tx);
 458             fjy2             = _mm_add_ps(fjy2,ty);
 459             fjz2             = _mm_add_ps(fjz2,tz);
 460
 461             /**************************
 462              * CALCULATE INTERACTIONS *
 463              **************************/
 464
 465             /* COULOMB ELECTROSTATICS */
 466             velec            = _mm_mul_ps(qq23,rinv23);
 467             felec            = _mm_mul_ps(velec,rinvsq23);
 468
 469             /* Update potential sum for this i atom from the interaction with this j atom. */
 470             velecsum         = _mm_add_ps(velecsum,velec);
 471
 472             fscal            = felec;
 473
 474             /* Calculate temporary vectorial force */
 475             tx               = _mm_mul_ps(fscal,dx23);
 476             ty               = _mm_mul_ps(fscal,dy23);
 477             tz               = _mm_mul_ps(fscal,dz23);
 478
 479             /* Update vectorial force */
 480             fix2             = _mm_add_ps(fix2,tx);
 481             fiy2             = _mm_add_ps(fiy2,ty);
 482             fiz2             = _mm_add_ps(fiz2,tz);
 483
 484             fjx3             = _mm_add_ps(fjx3,tx);
 485             fjy3             = _mm_add_ps(fjy3,ty);
 486             fjz3             = _mm_add_ps(fjz3,tz);
 487
 488             /**************************
 489              * CALCULATE INTERACTIONS *
 490              **************************/
 491
 492             /* COULOMB ELECTROSTATICS */
 493             velec            = _mm_mul_ps(qq31,rinv31);
 494             felec            = _mm_mul_ps(velec,rinvsq31);
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velecsum         = _mm_add_ps(velecsum,velec);
 498
 499             fscal            = felec;
 500
 501             /* Calculate temporary vectorial force */
 502             tx               = _mm_mul_ps(fscal,dx31);
 503             ty               = _mm_mul_ps(fscal,dy31);
 504             tz               = _mm_mul_ps(fscal,dz31);
 505
 506             /* Update vectorial force */
 507             fix3             = _mm_add_ps(fix3,tx);
 508             fiy3             = _mm_add_ps(fiy3,ty);
 509             fiz3             = _mm_add_ps(fiz3,tz);
 510
 511             fjx1             = _mm_add_ps(fjx1,tx);
 512             fjy1             = _mm_add_ps(fjy1,ty);
 513             fjz1             = _mm_add_ps(fjz1,tz);
 514
 515             /**************************
 516              * CALCULATE INTERACTIONS *
 517              **************************/
 518
 519             /* COULOMB ELECTROSTATICS */
 520             velec            = _mm_mul_ps(qq32,rinv32);
 521             felec            = _mm_mul_ps(velec,rinvsq32);
 522
 523             /* Update potential sum for this i atom from the interaction with this j atom. */
 524             velecsum         = _mm_add_ps(velecsum,velec);
 525
 526             fscal            = felec;
 527
 528             /* Calculate temporary vectorial force */
 529             tx               = _mm_mul_ps(fscal,dx32);
 530             ty               = _mm_mul_ps(fscal,dy32);
 531             tz               = _mm_mul_ps(fscal,dz32);
 532
 533             /* Update vectorial force */
 534             fix3             = _mm_add_ps(fix3,tx);
 535             fiy3             = _mm_add_ps(fiy3,ty);
 536             fiz3             = _mm_add_ps(fiz3,tz);
 537
 538             fjx2             = _mm_add_ps(fjx2,tx);
 539             fjy2             = _mm_add_ps(fjy2,ty);
 540             fjz2             = _mm_add_ps(fjz2,tz);
 541
 542             /**************************
 543              * CALCULATE INTERACTIONS *
 544              **************************/
 545
 546             /* COULOMB ELECTROSTATICS */
 547             velec            = _mm_mul_ps(qq33,rinv33);
 548             felec            = _mm_mul_ps(velec,rinvsq33);
 549
 550             /* Update potential sum for this i atom from the interaction with this j atom. */
 551             velecsum         = _mm_add_ps(velecsum,velec);
 552
 553             fscal            = felec;
 554
 555             /* Calculate temporary vectorial force */
 556             tx               = _mm_mul_ps(fscal,dx33);
 557             ty               = _mm_mul_ps(fscal,dy33);
 558             tz               = _mm_mul_ps(fscal,dz33);
 559
 560             /* Update vectorial force */
 561             fix3             = _mm_add_ps(fix3,tx);
 562             fiy3             = _mm_add_ps(fiy3,ty);
 563             fiz3             = _mm_add_ps(fiz3,tz);
 564
 565             fjx3             = _mm_add_ps(fjx3,tx);
 566             fjy3             = _mm_add_ps(fjy3,ty);
 567             fjz3             = _mm_add_ps(fjz3,tz);
 568
 569             fjptrA             = f+j_coord_offsetA;
 570             fjptrB             = f+j_coord_offsetB;
 571             fjptrC             = f+j_coord_offsetC;
 572             fjptrD             = f+j_coord_offsetD;
 573
 574             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 575                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 576                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 577
 578             /* Inner loop uses 287 flops */
 579         }
 580
 581         if(jidx<j_index_end)
 582         {
 583
 584             /* Get j neighbor index, and coordinate index */
 585             jnrlistA         = jjnr[jidx];
 586             jnrlistB         = jjnr[jidx+1];
 587             jnrlistC         = jjnr[jidx+2];
 588             jnrlistD         = jjnr[jidx+3];
 589             /* Sign of each element will be negative for non-real atoms.
 590              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 591              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 592              */
 593             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 594             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 595             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 596             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 597             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 598             j_coord_offsetA  = DIM*jnrA;
 599             j_coord_offsetB  = DIM*jnrB;
 600             j_coord_offsetC  = DIM*jnrC;
 601             j_coord_offsetD  = DIM*jnrD;
 602
 603             /* load j atom coordinates */
 604             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 605                                               x+j_coord_offsetC,x+j_coord_offsetD,
 606                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 607                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 608
 609             /* Calculate displacement vector */
 610             dx00             = _mm_sub_ps(ix0,jx0);
 611             dy00             = _mm_sub_ps(iy0,jy0);
 612             dz00             = _mm_sub_ps(iz0,jz0);
 613             dx11             = _mm_sub_ps(ix1,jx1);
 614             dy11             = _mm_sub_ps(iy1,jy1);
 615             dz11             = _mm_sub_ps(iz1,jz1);
 616             dx12             = _mm_sub_ps(ix1,jx2);
 617             dy12             = _mm_sub_ps(iy1,jy2);
 618             dz12             = _mm_sub_ps(iz1,jz2);
 619             dx13             = _mm_sub_ps(ix1,jx3);
 620             dy13             = _mm_sub_ps(iy1,jy3);
 621             dz13             = _mm_sub_ps(iz1,jz3);
 622             dx21             = _mm_sub_ps(ix2,jx1);
 623             dy21             = _mm_sub_ps(iy2,jy1);
 624             dz21             = _mm_sub_ps(iz2,jz1);
 625             dx22             = _mm_sub_ps(ix2,jx2);
 626             dy22             = _mm_sub_ps(iy2,jy2);
 627             dz22             = _mm_sub_ps(iz2,jz2);
 628             dx23             = _mm_sub_ps(ix2,jx3);
 629             dy23             = _mm_sub_ps(iy2,jy3);
 630             dz23             = _mm_sub_ps(iz2,jz3);
 631             dx31             = _mm_sub_ps(ix3,jx1);
 632             dy31             = _mm_sub_ps(iy3,jy1);
 633             dz31             = _mm_sub_ps(iz3,jz1);
 634             dx32             = _mm_sub_ps(ix3,jx2);
 635             dy32             = _mm_sub_ps(iy3,jy2);
 636             dz32             = _mm_sub_ps(iz3,jz2);
 637             dx33             = _mm_sub_ps(ix3,jx3);
 638             dy33             = _mm_sub_ps(iy3,jy3);
 639             dz33             = _mm_sub_ps(iz3,jz3);
 640
 641             /* Calculate squared distance and things based on it */
 642             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 643             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 644             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 645             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 646             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 647             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 648             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 649             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 650             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 651             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 652
 653             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 654             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 655             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 656             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 657             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 658             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 659             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 660             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 661             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 662
 663             rinvsq00         = gmx_mm_inv_ps(rsq00);
 664             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 665             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 666             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 667             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 668             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 669             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 670             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 671             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 672             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 673
 674             fjx0             = _mm_setzero_ps();
 675             fjy0             = _mm_setzero_ps();
 676             fjz0             = _mm_setzero_ps();
 677             fjx1             = _mm_setzero_ps();
 678             fjy1             = _mm_setzero_ps();
 679             fjz1             = _mm_setzero_ps();
 680             fjx2             = _mm_setzero_ps();
 681             fjy2             = _mm_setzero_ps();
 682             fjz2             = _mm_setzero_ps();
 683             fjx3             = _mm_setzero_ps();
 684             fjy3             = _mm_setzero_ps();
 685             fjz3             = _mm_setzero_ps();
 686
 687             /**************************
 688              * CALCULATE INTERACTIONS *
 689              **************************/
 690
 691             /* LENNARD-JONES DISPERSION/REPULSION */
 692
 693             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 694             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 695             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 696             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 697             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 698
 699             /* Update potential sum for this i atom from the interaction with this j atom. */
 700             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 701             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 702
 703             fscal            = fvdw;
 704
 705             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 706
 707             /* Calculate temporary vectorial force */
 708             tx               = _mm_mul_ps(fscal,dx00);
 709             ty               = _mm_mul_ps(fscal,dy00);
 710             tz               = _mm_mul_ps(fscal,dz00);
 711
 712             /* Update vectorial force */
 713             fix0             = _mm_add_ps(fix0,tx);
 714             fiy0             = _mm_add_ps(fiy0,ty);
 715             fiz0             = _mm_add_ps(fiz0,tz);
 716
 717             fjx0             = _mm_add_ps(fjx0,tx);
 718             fjy0             = _mm_add_ps(fjy0,ty);
 719             fjz0             = _mm_add_ps(fjz0,tz);
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             /* COULOMB ELECTROSTATICS */
 726             velec            = _mm_mul_ps(qq11,rinv11);
 727             felec            = _mm_mul_ps(velec,rinvsq11);
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_andnot_ps(dummy_mask,velec);
 731             velecsum         = _mm_add_ps(velecsum,velec);
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 736
 737             /* Calculate temporary vectorial force */
 738             tx               = _mm_mul_ps(fscal,dx11);
 739             ty               = _mm_mul_ps(fscal,dy11);
 740             tz               = _mm_mul_ps(fscal,dz11);
 741
 742             /* Update vectorial force */
 743             fix1             = _mm_add_ps(fix1,tx);
 744             fiy1             = _mm_add_ps(fiy1,ty);
 745             fiz1             = _mm_add_ps(fiz1,tz);
 746
 747             fjx1             = _mm_add_ps(fjx1,tx);
 748             fjy1             = _mm_add_ps(fjy1,ty);
 749             fjz1             = _mm_add_ps(fjz1,tz);
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             /* COULOMB ELECTROSTATICS */
 756             velec            = _mm_mul_ps(qq12,rinv12);
 757             felec            = _mm_mul_ps(velec,rinvsq12);
 758
 759             /* Update potential sum for this i atom from the interaction with this j atom. */
 760             velec            = _mm_andnot_ps(dummy_mask,velec);
 761             velecsum         = _mm_add_ps(velecsum,velec);
 762
 763             fscal            = felec;
 764
 765             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 766
 767             /* Calculate temporary vectorial force */
 768             tx               = _mm_mul_ps(fscal,dx12);
 769             ty               = _mm_mul_ps(fscal,dy12);
 770             tz               = _mm_mul_ps(fscal,dz12);
 771
 772             /* Update vectorial force */
 773             fix1             = _mm_add_ps(fix1,tx);
 774             fiy1             = _mm_add_ps(fiy1,ty);
 775             fiz1             = _mm_add_ps(fiz1,tz);
 776
 777             fjx2             = _mm_add_ps(fjx2,tx);
 778             fjy2             = _mm_add_ps(fjy2,ty);
 779             fjz2             = _mm_add_ps(fjz2,tz);
 780
 781             /**************************
 782              * CALCULATE INTERACTIONS *
 783              **************************/
 784
 785             /* COULOMB ELECTROSTATICS */
 786             velec            = _mm_mul_ps(qq13,rinv13);
 787             felec            = _mm_mul_ps(velec,rinvsq13);
 788
 789             /* Update potential sum for this i atom from the interaction with this j atom. */
 790             velec            = _mm_andnot_ps(dummy_mask,velec);
 791             velecsum         = _mm_add_ps(velecsum,velec);
 792
 793             fscal            = felec;
 794
 795             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 796
 797             /* Calculate temporary vectorial force */
 798             tx               = _mm_mul_ps(fscal,dx13);
 799             ty               = _mm_mul_ps(fscal,dy13);
 800             tz               = _mm_mul_ps(fscal,dz13);
 801
 802             /* Update vectorial force */
 803             fix1             = _mm_add_ps(fix1,tx);
 804             fiy1             = _mm_add_ps(fiy1,ty);
 805             fiz1             = _mm_add_ps(fiz1,tz);
 806
 807             fjx3             = _mm_add_ps(fjx3,tx);
 808             fjy3             = _mm_add_ps(fjy3,ty);
 809             fjz3             = _mm_add_ps(fjz3,tz);
 810
 811             /**************************
 812              * CALCULATE INTERACTIONS *
 813              **************************/
 814
 815             /* COULOMB ELECTROSTATICS */
 816             velec            = _mm_mul_ps(qq21,rinv21);
 817             felec            = _mm_mul_ps(velec,rinvsq21);
 818
 819             /* Update potential sum for this i atom from the interaction with this j atom. */
 820             velec            = _mm_andnot_ps(dummy_mask,velec);
 821             velecsum         = _mm_add_ps(velecsum,velec);
 822
 823             fscal            = felec;
 824
 825             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 826
 827             /* Calculate temporary vectorial force */
 828             tx               = _mm_mul_ps(fscal,dx21);
 829             ty               = _mm_mul_ps(fscal,dy21);
 830             tz               = _mm_mul_ps(fscal,dz21);
 831
 832             /* Update vectorial force */
 833             fix2             = _mm_add_ps(fix2,tx);
 834             fiy2             = _mm_add_ps(fiy2,ty);
 835             fiz2             = _mm_add_ps(fiz2,tz);
 836
 837             fjx1             = _mm_add_ps(fjx1,tx);
 838             fjy1             = _mm_add_ps(fjy1,ty);
 839             fjz1             = _mm_add_ps(fjz1,tz);
 840
 841             /**************************
 842              * CALCULATE INTERACTIONS *
 843              **************************/
 844
 845             /* COULOMB ELECTROSTATICS */
 846             velec            = _mm_mul_ps(qq22,rinv22);
 847             felec            = _mm_mul_ps(velec,rinvsq22);
 848
 849             /* Update potential sum for this i atom from the interaction with this j atom. */
 850             velec            = _mm_andnot_ps(dummy_mask,velec);
 851             velecsum         = _mm_add_ps(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 856
 857             /* Calculate temporary vectorial force */
 858             tx               = _mm_mul_ps(fscal,dx22);
 859             ty               = _mm_mul_ps(fscal,dy22);
 860             tz               = _mm_mul_ps(fscal,dz22);
 861
 862             /* Update vectorial force */
 863             fix2             = _mm_add_ps(fix2,tx);
 864             fiy2             = _mm_add_ps(fiy2,ty);
 865             fiz2             = _mm_add_ps(fiz2,tz);
 866
 867             fjx2             = _mm_add_ps(fjx2,tx);
 868             fjy2             = _mm_add_ps(fjy2,ty);
 869             fjz2             = _mm_add_ps(fjz2,tz);
 870
 871             /**************************
 872              * CALCULATE INTERACTIONS *
 873              **************************/
 874
 875             /* COULOMB ELECTROSTATICS */
 876             velec            = _mm_mul_ps(qq23,rinv23);
 877             felec            = _mm_mul_ps(velec,rinvsq23);
 878
 879             /* Update potential sum for this i atom from the interaction with this j atom. */
 880             velec            = _mm_andnot_ps(dummy_mask,velec);
 881             velecsum         = _mm_add_ps(velecsum,velec);
 882
 883             fscal            = felec;
 884
 885             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 886
 887             /* Calculate temporary vectorial force */
 888             tx               = _mm_mul_ps(fscal,dx23);
 889             ty               = _mm_mul_ps(fscal,dy23);
 890             tz               = _mm_mul_ps(fscal,dz23);
 891
 892             /* Update vectorial force */
 893             fix2             = _mm_add_ps(fix2,tx);
 894             fiy2             = _mm_add_ps(fiy2,ty);
 895             fiz2             = _mm_add_ps(fiz2,tz);
 896
 897             fjx3             = _mm_add_ps(fjx3,tx);
 898             fjy3             = _mm_add_ps(fjy3,ty);
 899             fjz3             = _mm_add_ps(fjz3,tz);
 900
 901             /**************************
 902              * CALCULATE INTERACTIONS *
 903              **************************/
 904
 905             /* COULOMB ELECTROSTATICS */
 906             velec            = _mm_mul_ps(qq31,rinv31);
 907             felec            = _mm_mul_ps(velec,rinvsq31);
 908
 909             /* Update potential sum for this i atom from the interaction with this j atom. */
 910             velec            = _mm_andnot_ps(dummy_mask,velec);
 911             velecsum         = _mm_add_ps(velecsum,velec);
 912
 913             fscal            = felec;
 914
 915             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 916
 917             /* Calculate temporary vectorial force */
 918             tx               = _mm_mul_ps(fscal,dx31);
 919             ty               = _mm_mul_ps(fscal,dy31);
 920             tz               = _mm_mul_ps(fscal,dz31);
 921
 922             /* Update vectorial force */
 923             fix3             = _mm_add_ps(fix3,tx);
 924             fiy3             = _mm_add_ps(fiy3,ty);
 925             fiz3             = _mm_add_ps(fiz3,tz);
 926
 927             fjx1             = _mm_add_ps(fjx1,tx);
 928             fjy1             = _mm_add_ps(fjy1,ty);
 929             fjz1             = _mm_add_ps(fjz1,tz);
 930
 931             /**************************
 932              * CALCULATE INTERACTIONS *
 933              **************************/
 934
 935             /* COULOMB ELECTROSTATICS */
 936             velec            = _mm_mul_ps(qq32,rinv32);
 937             felec            = _mm_mul_ps(velec,rinvsq32);
 938
 939             /* Update potential sum for this i atom from the interaction with this j atom. */
 940             velec            = _mm_andnot_ps(dummy_mask,velec);
 941             velecsum         = _mm_add_ps(velecsum,velec);
 942
 943             fscal            = felec;
 944
 945             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 946
 947             /* Calculate temporary vectorial force */
 948             tx               = _mm_mul_ps(fscal,dx32);
 949             ty               = _mm_mul_ps(fscal,dy32);
 950             tz               = _mm_mul_ps(fscal,dz32);
 951
 952             /* Update vectorial force */
 953             fix3             = _mm_add_ps(fix3,tx);
 954             fiy3             = _mm_add_ps(fiy3,ty);
 955             fiz3             = _mm_add_ps(fiz3,tz);
 956
 957             fjx2             = _mm_add_ps(fjx2,tx);
 958             fjy2             = _mm_add_ps(fjy2,ty);
 959             fjz2             = _mm_add_ps(fjz2,tz);
 960
 961             /**************************
 962              * CALCULATE INTERACTIONS *
 963              **************************/
 964
 965             /* COULOMB ELECTROSTATICS */
 966             velec            = _mm_mul_ps(qq33,rinv33);
 967             felec            = _mm_mul_ps(velec,rinvsq33);
 968
 969             /* Update potential sum for this i atom from the interaction with this j atom. */
 970             velec            = _mm_andnot_ps(dummy_mask,velec);
 971             velecsum         = _mm_add_ps(velecsum,velec);
 972
 973             fscal            = felec;
 974
 975             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 976
 977             /* Calculate temporary vectorial force */
 978             tx               = _mm_mul_ps(fscal,dx33);
 979             ty               = _mm_mul_ps(fscal,dy33);
 980             tz               = _mm_mul_ps(fscal,dz33);
 981
 982             /* Update vectorial force */
 983             fix3             = _mm_add_ps(fix3,tx);
 984             fiy3             = _mm_add_ps(fiy3,ty);
 985             fiz3             = _mm_add_ps(fiz3,tz);
 986
 987             fjx3             = _mm_add_ps(fjx3,tx);
 988             fjy3             = _mm_add_ps(fjy3,ty);
 989             fjz3             = _mm_add_ps(fjz3,tz);
 990
 991             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 992             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 993             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 994             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 995
 996             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 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 26 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*26 + 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              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1052     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1053     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1054     real             rcutoff_scalar;
1055     real             *shiftvec,*fshift,*x,*f;
1056     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1057     real             scratch[4*DIM];
1058     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1059     int              vdwioffset0;
1060     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1061     int              vdwioffset1;
1062     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1063     int              vdwioffset2;
1064     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1065     int              vdwioffset3;
1066     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1067     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1068     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1069     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1070     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1071     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1072     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1073     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1074     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1075     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1076     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1077     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1078     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1079     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1080     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1081     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1082     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1083     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1084     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1085     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1086     real             *charge;
1087     int              nvdwtype;
1088     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1089     int              *vdwtype;
1090     real             *vdwparam;
1091     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1092     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1093     __m128           dummy_mask,cutoff_mask;
1094     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1095     __m128           one     = _mm_set1_ps(1.0);
1096     __m128           two     = _mm_set1_ps(2.0);
1097     x                = xx[0];
1098     f                = ff[0];
1099
1100     nri              = nlist->nri;
1101     iinr             = nlist->iinr;
1102     jindex           = nlist->jindex;
1103     jjnr             = nlist->jjnr;
1104     shiftidx         = nlist->shift;
1105     gid              = nlist->gid;
1106     shiftvec         = fr->shift_vec[0];
1107     fshift           = fr->fshift[0];
1108     facel            = _mm_set1_ps(fr->epsfac);
1109     charge           = mdatoms->chargeA;
1110     nvdwtype         = fr->ntype;
1111     vdwparam         = fr->nbfp;
1112     vdwtype          = mdatoms->typeA;
1113
1114     /* Setup water-specific parameters */
1115     inr              = nlist->iinr[0];
1116     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1117     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1118     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1119     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1120
1121     jq1              = _mm_set1_ps(charge[inr+1]);
1122     jq2              = _mm_set1_ps(charge[inr+2]);
1123     jq3              = _mm_set1_ps(charge[inr+3]);
1124     vdwjidx0A        = 2*vdwtype[inr+0];
1125     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1126     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1127     qq11             = _mm_mul_ps(iq1,jq1);
1128     qq12             = _mm_mul_ps(iq1,jq2);
1129     qq13             = _mm_mul_ps(iq1,jq3);
1130     qq21             = _mm_mul_ps(iq2,jq1);
1131     qq22             = _mm_mul_ps(iq2,jq2);
1132     qq23             = _mm_mul_ps(iq2,jq3);
1133     qq31             = _mm_mul_ps(iq3,jq1);
1134     qq32             = _mm_mul_ps(iq3,jq2);
1135     qq33             = _mm_mul_ps(iq3,jq3);
1136
1137     /* Avoid stupid compiler warnings */
1138     jnrA = jnrB = jnrC = jnrD = 0;
1139     j_coord_offsetA = 0;
1140     j_coord_offsetB = 0;
1141     j_coord_offsetC = 0;
1142     j_coord_offsetD = 0;
1143
1144     outeriter        = 0;
1145     inneriter        = 0;
1146
1147     for(iidx=0;iidx<4*DIM;iidx++)
1148     {
1149         scratch[iidx] = 0.0;
1150     }
1151
1152     /* Start outer loop over neighborlists */
1153     for(iidx=0; iidx<nri; iidx++)
1154     {
1155         /* Load shift vector for this list */
1156         i_shift_offset   = DIM*shiftidx[iidx];
1157
1158         /* Load limits for loop over neighbors */
1159         j_index_start    = jindex[iidx];
1160         j_index_end      = jindex[iidx+1];
1161
1162         /* Get outer coordinate index */
1163         inr              = iinr[iidx];
1164         i_coord_offset   = DIM*inr;
1165
1166         /* Load i particle coords and add shift vector */
1167         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1168                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1169
1170         fix0             = _mm_setzero_ps();
1171         fiy0             = _mm_setzero_ps();
1172         fiz0             = _mm_setzero_ps();
1173         fix1             = _mm_setzero_ps();
1174         fiy1             = _mm_setzero_ps();
1175         fiz1             = _mm_setzero_ps();
1176         fix2             = _mm_setzero_ps();
1177         fiy2             = _mm_setzero_ps();
1178         fiz2             = _mm_setzero_ps();
1179         fix3             = _mm_setzero_ps();
1180         fiy3             = _mm_setzero_ps();
1181         fiz3             = _mm_setzero_ps();
1182
1183         /* Start inner kernel loop */
1184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1185         {
1186
1187             /* Get j neighbor index, and coordinate index */
1188             jnrA             = jjnr[jidx];
1189             jnrB             = jjnr[jidx+1];
1190             jnrC             = jjnr[jidx+2];
1191             jnrD             = jjnr[jidx+3];
1192             j_coord_offsetA  = DIM*jnrA;
1193             j_coord_offsetB  = DIM*jnrB;
1194             j_coord_offsetC  = DIM*jnrC;
1195             j_coord_offsetD  = DIM*jnrD;
1196
1197             /* load j atom coordinates */
1198             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1199                                               x+j_coord_offsetC,x+j_coord_offsetD,
1200                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1201                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1202
1203             /* Calculate displacement vector */
1204             dx00             = _mm_sub_ps(ix0,jx0);
1205             dy00             = _mm_sub_ps(iy0,jy0);
1206             dz00             = _mm_sub_ps(iz0,jz0);
1207             dx11             = _mm_sub_ps(ix1,jx1);
1208             dy11             = _mm_sub_ps(iy1,jy1);
1209             dz11             = _mm_sub_ps(iz1,jz1);
1210             dx12             = _mm_sub_ps(ix1,jx2);
1211             dy12             = _mm_sub_ps(iy1,jy2);
1212             dz12             = _mm_sub_ps(iz1,jz2);
1213             dx13             = _mm_sub_ps(ix1,jx3);
1214             dy13             = _mm_sub_ps(iy1,jy3);
1215             dz13             = _mm_sub_ps(iz1,jz3);
1216             dx21             = _mm_sub_ps(ix2,jx1);
1217             dy21             = _mm_sub_ps(iy2,jy1);
1218             dz21             = _mm_sub_ps(iz2,jz1);
1219             dx22             = _mm_sub_ps(ix2,jx2);
1220             dy22             = _mm_sub_ps(iy2,jy2);
1221             dz22             = _mm_sub_ps(iz2,jz2);
1222             dx23             = _mm_sub_ps(ix2,jx3);
1223             dy23             = _mm_sub_ps(iy2,jy3);
1224             dz23             = _mm_sub_ps(iz2,jz3);
1225             dx31             = _mm_sub_ps(ix3,jx1);
1226             dy31             = _mm_sub_ps(iy3,jy1);
1227             dz31             = _mm_sub_ps(iz3,jz1);
1228             dx32             = _mm_sub_ps(ix3,jx2);
1229             dy32             = _mm_sub_ps(iy3,jy2);
1230             dz32             = _mm_sub_ps(iz3,jz2);
1231             dx33             = _mm_sub_ps(ix3,jx3);
1232             dy33             = _mm_sub_ps(iy3,jy3);
1233             dz33             = _mm_sub_ps(iz3,jz3);
1234
1235             /* Calculate squared distance and things based on it */
1236             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1237             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1238             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1239             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1240             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1241             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1242             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1243             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1244             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1245             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1246
1247             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1248             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1249             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1250             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1251             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1252             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1253             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1254             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1255             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1256
1257             rinvsq00         = gmx_mm_inv_ps(rsq00);
1258             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1259             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1260             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1261             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1262             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1263             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1264             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1265             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1266             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1267
1268             fjx0             = _mm_setzero_ps();
1269             fjy0             = _mm_setzero_ps();
1270             fjz0             = _mm_setzero_ps();
1271             fjx1             = _mm_setzero_ps();
1272             fjy1             = _mm_setzero_ps();
1273             fjz1             = _mm_setzero_ps();
1274             fjx2             = _mm_setzero_ps();
1275             fjy2             = _mm_setzero_ps();
1276             fjz2             = _mm_setzero_ps();
1277             fjx3             = _mm_setzero_ps();
1278             fjy3             = _mm_setzero_ps();
1279             fjz3             = _mm_setzero_ps();
1280
1281             /**************************
1282              * CALCULATE INTERACTIONS *
1283              **************************/
1284
1285             /* LENNARD-JONES DISPERSION/REPULSION */
1286
1287             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1288             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1289
1290             fscal            = fvdw;
1291
1292             /* Calculate temporary vectorial force */
1293             tx               = _mm_mul_ps(fscal,dx00);
1294             ty               = _mm_mul_ps(fscal,dy00);
1295             tz               = _mm_mul_ps(fscal,dz00);
1296
1297             /* Update vectorial force */
1298             fix0             = _mm_add_ps(fix0,tx);
1299             fiy0             = _mm_add_ps(fiy0,ty);
1300             fiz0             = _mm_add_ps(fiz0,tz);
1301
1302             fjx0             = _mm_add_ps(fjx0,tx);
1303             fjy0             = _mm_add_ps(fjy0,ty);
1304             fjz0             = _mm_add_ps(fjz0,tz);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* COULOMB ELECTROSTATICS */
1311             velec            = _mm_mul_ps(qq11,rinv11);
1312             felec            = _mm_mul_ps(velec,rinvsq11);
1313
1314             fscal            = felec;
1315
1316             /* Calculate temporary vectorial force */
1317             tx               = _mm_mul_ps(fscal,dx11);
1318             ty               = _mm_mul_ps(fscal,dy11);
1319             tz               = _mm_mul_ps(fscal,dz11);
1320
1321             /* Update vectorial force */
1322             fix1             = _mm_add_ps(fix1,tx);
1323             fiy1             = _mm_add_ps(fiy1,ty);
1324             fiz1             = _mm_add_ps(fiz1,tz);
1325
1326             fjx1             = _mm_add_ps(fjx1,tx);
1327             fjy1             = _mm_add_ps(fjy1,ty);
1328             fjz1             = _mm_add_ps(fjz1,tz);
1329
1330             /**************************
1331              * CALCULATE INTERACTIONS *
1332              **************************/
1333
1334             /* COULOMB ELECTROSTATICS */
1335             velec            = _mm_mul_ps(qq12,rinv12);
1336             felec            = _mm_mul_ps(velec,rinvsq12);
1337
1338             fscal            = felec;
1339
1340             /* Calculate temporary vectorial force */
1341             tx               = _mm_mul_ps(fscal,dx12);
1342             ty               = _mm_mul_ps(fscal,dy12);
1343             tz               = _mm_mul_ps(fscal,dz12);
1344
1345             /* Update vectorial force */
1346             fix1             = _mm_add_ps(fix1,tx);
1347             fiy1             = _mm_add_ps(fiy1,ty);
1348             fiz1             = _mm_add_ps(fiz1,tz);
1349
1350             fjx2             = _mm_add_ps(fjx2,tx);
1351             fjy2             = _mm_add_ps(fjy2,ty);
1352             fjz2             = _mm_add_ps(fjz2,tz);
1353
1354             /**************************
1355              * CALCULATE INTERACTIONS *
1356              **************************/
1357
1358             /* COULOMB ELECTROSTATICS */
1359             velec            = _mm_mul_ps(qq13,rinv13);
1360             felec            = _mm_mul_ps(velec,rinvsq13);
1361
1362             fscal            = felec;
1363
1364             /* Calculate temporary vectorial force */
1365             tx               = _mm_mul_ps(fscal,dx13);
1366             ty               = _mm_mul_ps(fscal,dy13);
1367             tz               = _mm_mul_ps(fscal,dz13);
1368
1369             /* Update vectorial force */
1370             fix1             = _mm_add_ps(fix1,tx);
1371             fiy1             = _mm_add_ps(fiy1,ty);
1372             fiz1             = _mm_add_ps(fiz1,tz);
1373
1374             fjx3             = _mm_add_ps(fjx3,tx);
1375             fjy3             = _mm_add_ps(fjy3,ty);
1376             fjz3             = _mm_add_ps(fjz3,tz);
1377
1378             /**************************
1379              * CALCULATE INTERACTIONS *
1380              **************************/
1381
1382             /* COULOMB ELECTROSTATICS */
1383             velec            = _mm_mul_ps(qq21,rinv21);
1384             felec            = _mm_mul_ps(velec,rinvsq21);
1385
1386             fscal            = felec;
1387
1388             /* Calculate temporary vectorial force */
1389             tx               = _mm_mul_ps(fscal,dx21);
1390             ty               = _mm_mul_ps(fscal,dy21);
1391             tz               = _mm_mul_ps(fscal,dz21);
1392
1393             /* Update vectorial force */
1394             fix2             = _mm_add_ps(fix2,tx);
1395             fiy2             = _mm_add_ps(fiy2,ty);
1396             fiz2             = _mm_add_ps(fiz2,tz);
1397
1398             fjx1             = _mm_add_ps(fjx1,tx);
1399             fjy1             = _mm_add_ps(fjy1,ty);
1400             fjz1             = _mm_add_ps(fjz1,tz);
1401
1402             /**************************
1403              * CALCULATE INTERACTIONS *
1404              **************************/
1405
1406             /* COULOMB ELECTROSTATICS */
1407             velec            = _mm_mul_ps(qq22,rinv22);
1408             felec            = _mm_mul_ps(velec,rinvsq22);
1409
1410             fscal            = felec;
1411
1412             /* Calculate temporary vectorial force */
1413             tx               = _mm_mul_ps(fscal,dx22);
1414             ty               = _mm_mul_ps(fscal,dy22);
1415             tz               = _mm_mul_ps(fscal,dz22);
1416
1417             /* Update vectorial force */
1418             fix2             = _mm_add_ps(fix2,tx);
1419             fiy2             = _mm_add_ps(fiy2,ty);
1420             fiz2             = _mm_add_ps(fiz2,tz);
1421
1422             fjx2             = _mm_add_ps(fjx2,tx);
1423             fjy2             = _mm_add_ps(fjy2,ty);
1424             fjz2             = _mm_add_ps(fjz2,tz);
1425
1426             /**************************
1427              * CALCULATE INTERACTIONS *
1428              **************************/
1429
1430             /* COULOMB ELECTROSTATICS */
1431             velec            = _mm_mul_ps(qq23,rinv23);
1432             felec            = _mm_mul_ps(velec,rinvsq23);
1433
1434             fscal            = felec;
1435
1436             /* Calculate temporary vectorial force */
1437             tx               = _mm_mul_ps(fscal,dx23);
1438             ty               = _mm_mul_ps(fscal,dy23);
1439             tz               = _mm_mul_ps(fscal,dz23);
1440
1441             /* Update vectorial force */
1442             fix2             = _mm_add_ps(fix2,tx);
1443             fiy2             = _mm_add_ps(fiy2,ty);
1444             fiz2             = _mm_add_ps(fiz2,tz);
1445
1446             fjx3             = _mm_add_ps(fjx3,tx);
1447             fjy3             = _mm_add_ps(fjy3,ty);
1448             fjz3             = _mm_add_ps(fjz3,tz);
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             /* COULOMB ELECTROSTATICS */
1455             velec            = _mm_mul_ps(qq31,rinv31);
1456             felec            = _mm_mul_ps(velec,rinvsq31);
1457
1458             fscal            = felec;
1459
1460             /* Calculate temporary vectorial force */
1461             tx               = _mm_mul_ps(fscal,dx31);
1462             ty               = _mm_mul_ps(fscal,dy31);
1463             tz               = _mm_mul_ps(fscal,dz31);
1464
1465             /* Update vectorial force */
1466             fix3             = _mm_add_ps(fix3,tx);
1467             fiy3             = _mm_add_ps(fiy3,ty);
1468             fiz3             = _mm_add_ps(fiz3,tz);
1469
1470             fjx1             = _mm_add_ps(fjx1,tx);
1471             fjy1             = _mm_add_ps(fjy1,ty);
1472             fjz1             = _mm_add_ps(fjz1,tz);
1473
1474             /**************************
1475              * CALCULATE INTERACTIONS *
1476              **************************/
1477
1478             /* COULOMB ELECTROSTATICS */
1479             velec            = _mm_mul_ps(qq32,rinv32);
1480             felec            = _mm_mul_ps(velec,rinvsq32);
1481
1482             fscal            = felec;
1483
1484             /* Calculate temporary vectorial force */
1485             tx               = _mm_mul_ps(fscal,dx32);
1486             ty               = _mm_mul_ps(fscal,dy32);
1487             tz               = _mm_mul_ps(fscal,dz32);
1488
1489             /* Update vectorial force */
1490             fix3             = _mm_add_ps(fix3,tx);
1491             fiy3             = _mm_add_ps(fiy3,ty);
1492             fiz3             = _mm_add_ps(fiz3,tz);
1493
1494             fjx2             = _mm_add_ps(fjx2,tx);
1495             fjy2             = _mm_add_ps(fjy2,ty);
1496             fjz2             = _mm_add_ps(fjz2,tz);
1497
1498             /**************************
1499              * CALCULATE INTERACTIONS *
1500              **************************/
1501
1502             /* COULOMB ELECTROSTATICS */
1503             velec            = _mm_mul_ps(qq33,rinv33);
1504             felec            = _mm_mul_ps(velec,rinvsq33);
1505
1506             fscal            = felec;
1507
1508             /* Calculate temporary vectorial force */
1509             tx               = _mm_mul_ps(fscal,dx33);
1510             ty               = _mm_mul_ps(fscal,dy33);
1511             tz               = _mm_mul_ps(fscal,dz33);
1512
1513             /* Update vectorial force */
1514             fix3             = _mm_add_ps(fix3,tx);
1515             fiy3             = _mm_add_ps(fiy3,ty);
1516             fiz3             = _mm_add_ps(fiz3,tz);
1517
1518             fjx3             = _mm_add_ps(fjx3,tx);
1519             fjy3             = _mm_add_ps(fjy3,ty);
1520             fjz3             = _mm_add_ps(fjz3,tz);
1521
1522             fjptrA             = f+j_coord_offsetA;
1523             fjptrB             = f+j_coord_offsetB;
1524             fjptrC             = f+j_coord_offsetC;
1525             fjptrD             = f+j_coord_offsetD;
1526
1527             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1528                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1529                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1530
1531             /* Inner loop uses 273 flops */
1532         }
1533
1534         if(jidx<j_index_end)
1535         {
1536
1537             /* Get j neighbor index, and coordinate index */
1538             jnrlistA         = jjnr[jidx];
1539             jnrlistB         = jjnr[jidx+1];
1540             jnrlistC         = jjnr[jidx+2];
1541             jnrlistD         = jjnr[jidx+3];
1542             /* Sign of each element will be negative for non-real atoms.
1543              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1544              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1545              */
1546             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1547             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1548             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1549             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1550             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1551             j_coord_offsetA  = DIM*jnrA;
1552             j_coord_offsetB  = DIM*jnrB;
1553             j_coord_offsetC  = DIM*jnrC;
1554             j_coord_offsetD  = DIM*jnrD;
1555
1556             /* load j atom coordinates */
1557             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1558                                               x+j_coord_offsetC,x+j_coord_offsetD,
1559                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1560                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1561
1562             /* Calculate displacement vector */
1563             dx00             = _mm_sub_ps(ix0,jx0);
1564             dy00             = _mm_sub_ps(iy0,jy0);
1565             dz00             = _mm_sub_ps(iz0,jz0);
1566             dx11             = _mm_sub_ps(ix1,jx1);
1567             dy11             = _mm_sub_ps(iy1,jy1);
1568             dz11             = _mm_sub_ps(iz1,jz1);
1569             dx12             = _mm_sub_ps(ix1,jx2);
1570             dy12             = _mm_sub_ps(iy1,jy2);
1571             dz12             = _mm_sub_ps(iz1,jz2);
1572             dx13             = _mm_sub_ps(ix1,jx3);
1573             dy13             = _mm_sub_ps(iy1,jy3);
1574             dz13             = _mm_sub_ps(iz1,jz3);
1575             dx21             = _mm_sub_ps(ix2,jx1);
1576             dy21             = _mm_sub_ps(iy2,jy1);
1577             dz21             = _mm_sub_ps(iz2,jz1);
1578             dx22             = _mm_sub_ps(ix2,jx2);
1579             dy22             = _mm_sub_ps(iy2,jy2);
1580             dz22             = _mm_sub_ps(iz2,jz2);
1581             dx23             = _mm_sub_ps(ix2,jx3);
1582             dy23             = _mm_sub_ps(iy2,jy3);
1583             dz23             = _mm_sub_ps(iz2,jz3);
1584             dx31             = _mm_sub_ps(ix3,jx1);
1585             dy31             = _mm_sub_ps(iy3,jy1);
1586             dz31             = _mm_sub_ps(iz3,jz1);
1587             dx32             = _mm_sub_ps(ix3,jx2);
1588             dy32             = _mm_sub_ps(iy3,jy2);
1589             dz32             = _mm_sub_ps(iz3,jz2);
1590             dx33             = _mm_sub_ps(ix3,jx3);
1591             dy33             = _mm_sub_ps(iy3,jy3);
1592             dz33             = _mm_sub_ps(iz3,jz3);
1593
1594             /* Calculate squared distance and things based on it */
1595             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1596             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1597             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1598             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1599             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1600             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1601             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1602             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1603             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1604             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1605
1606             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1607             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1608             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1609             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1610             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1611             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1612             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1613             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1614             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1615
1616             rinvsq00         = gmx_mm_inv_ps(rsq00);
1617             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1618             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1619             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1620             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1621             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1622             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1623             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1624             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1625             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1626
1627             fjx0             = _mm_setzero_ps();
1628             fjy0             = _mm_setzero_ps();
1629             fjz0             = _mm_setzero_ps();
1630             fjx1             = _mm_setzero_ps();
1631             fjy1             = _mm_setzero_ps();
1632             fjz1             = _mm_setzero_ps();
1633             fjx2             = _mm_setzero_ps();
1634             fjy2             = _mm_setzero_ps();
1635             fjz2             = _mm_setzero_ps();
1636             fjx3             = _mm_setzero_ps();
1637             fjy3             = _mm_setzero_ps();
1638             fjz3             = _mm_setzero_ps();
1639
1640             /**************************
1641              * CALCULATE INTERACTIONS *
1642              **************************/
1643
1644             /* LENNARD-JONES DISPERSION/REPULSION */
1645
1646             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1647             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1648
1649             fscal            = fvdw;
1650
1651             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1652
1653             /* Calculate temporary vectorial force */
1654             tx               = _mm_mul_ps(fscal,dx00);
1655             ty               = _mm_mul_ps(fscal,dy00);
1656             tz               = _mm_mul_ps(fscal,dz00);
1657
1658             /* Update vectorial force */
1659             fix0             = _mm_add_ps(fix0,tx);
1660             fiy0             = _mm_add_ps(fiy0,ty);
1661             fiz0             = _mm_add_ps(fiz0,tz);
1662
1663             fjx0             = _mm_add_ps(fjx0,tx);
1664             fjy0             = _mm_add_ps(fjy0,ty);
1665             fjz0             = _mm_add_ps(fjz0,tz);
1666
1667             /**************************
1668              * CALCULATE INTERACTIONS *
1669              **************************/
1670
1671             /* COULOMB ELECTROSTATICS */
1672             velec            = _mm_mul_ps(qq11,rinv11);
1673             felec            = _mm_mul_ps(velec,rinvsq11);
1674
1675             fscal            = felec;
1676
1677             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1678
1679             /* Calculate temporary vectorial force */
1680             tx               = _mm_mul_ps(fscal,dx11);
1681             ty               = _mm_mul_ps(fscal,dy11);
1682             tz               = _mm_mul_ps(fscal,dz11);
1683
1684             /* Update vectorial force */
1685             fix1             = _mm_add_ps(fix1,tx);
1686             fiy1             = _mm_add_ps(fiy1,ty);
1687             fiz1             = _mm_add_ps(fiz1,tz);
1688
1689             fjx1             = _mm_add_ps(fjx1,tx);
1690             fjy1             = _mm_add_ps(fjy1,ty);
1691             fjz1             = _mm_add_ps(fjz1,tz);
1692
1693             /**************************
1694              * CALCULATE INTERACTIONS *
1695              **************************/
1696
1697             /* COULOMB ELECTROSTATICS */
1698             velec            = _mm_mul_ps(qq12,rinv12);
1699             felec            = _mm_mul_ps(velec,rinvsq12);
1700
1701             fscal            = felec;
1702
1703             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1704
1705             /* Calculate temporary vectorial force */
1706             tx               = _mm_mul_ps(fscal,dx12);
1707             ty               = _mm_mul_ps(fscal,dy12);
1708             tz               = _mm_mul_ps(fscal,dz12);
1709
1710             /* Update vectorial force */
1711             fix1             = _mm_add_ps(fix1,tx);
1712             fiy1             = _mm_add_ps(fiy1,ty);
1713             fiz1             = _mm_add_ps(fiz1,tz);
1714
1715             fjx2             = _mm_add_ps(fjx2,tx);
1716             fjy2             = _mm_add_ps(fjy2,ty);
1717             fjz2             = _mm_add_ps(fjz2,tz);
1718
1719             /**************************
1720              * CALCULATE INTERACTIONS *
1721              **************************/
1722
1723             /* COULOMB ELECTROSTATICS */
1724             velec            = _mm_mul_ps(qq13,rinv13);
1725             felec            = _mm_mul_ps(velec,rinvsq13);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1730
1731             /* Calculate temporary vectorial force */
1732             tx               = _mm_mul_ps(fscal,dx13);
1733             ty               = _mm_mul_ps(fscal,dy13);
1734             tz               = _mm_mul_ps(fscal,dz13);
1735
1736             /* Update vectorial force */
1737             fix1             = _mm_add_ps(fix1,tx);
1738             fiy1             = _mm_add_ps(fiy1,ty);
1739             fiz1             = _mm_add_ps(fiz1,tz);
1740
1741             fjx3             = _mm_add_ps(fjx3,tx);
1742             fjy3             = _mm_add_ps(fjy3,ty);
1743             fjz3             = _mm_add_ps(fjz3,tz);
1744
1745             /**************************
1746              * CALCULATE INTERACTIONS *
1747              **************************/
1748
1749             /* COULOMB ELECTROSTATICS */
1750             velec            = _mm_mul_ps(qq21,rinv21);
1751             felec            = _mm_mul_ps(velec,rinvsq21);
1752
1753             fscal            = felec;
1754
1755             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1756
1757             /* Calculate temporary vectorial force */
1758             tx               = _mm_mul_ps(fscal,dx21);
1759             ty               = _mm_mul_ps(fscal,dy21);
1760             tz               = _mm_mul_ps(fscal,dz21);
1761
1762             /* Update vectorial force */
1763             fix2             = _mm_add_ps(fix2,tx);
1764             fiy2             = _mm_add_ps(fiy2,ty);
1765             fiz2             = _mm_add_ps(fiz2,tz);
1766
1767             fjx1             = _mm_add_ps(fjx1,tx);
1768             fjy1             = _mm_add_ps(fjy1,ty);
1769             fjz1             = _mm_add_ps(fjz1,tz);
1770
1771             /**************************
1772              * CALCULATE INTERACTIONS *
1773              **************************/
1774
1775             /* COULOMB ELECTROSTATICS */
1776             velec            = _mm_mul_ps(qq22,rinv22);
1777             felec            = _mm_mul_ps(velec,rinvsq22);
1778
1779             fscal            = felec;
1780
1781             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1782
1783             /* Calculate temporary vectorial force */
1784             tx               = _mm_mul_ps(fscal,dx22);
1785             ty               = _mm_mul_ps(fscal,dy22);
1786             tz               = _mm_mul_ps(fscal,dz22);
1787
1788             /* Update vectorial force */
1789             fix2             = _mm_add_ps(fix2,tx);
1790             fiy2             = _mm_add_ps(fiy2,ty);
1791             fiz2             = _mm_add_ps(fiz2,tz);
1792
1793             fjx2             = _mm_add_ps(fjx2,tx);
1794             fjy2             = _mm_add_ps(fjy2,ty);
1795             fjz2             = _mm_add_ps(fjz2,tz);
1796
1797             /**************************
1798              * CALCULATE INTERACTIONS *
1799              **************************/
1800
1801             /* COULOMB ELECTROSTATICS */
1802             velec            = _mm_mul_ps(qq23,rinv23);
1803             felec            = _mm_mul_ps(velec,rinvsq23);
1804
1805             fscal            = felec;
1806
1807             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1808
1809             /* Calculate temporary vectorial force */
1810             tx               = _mm_mul_ps(fscal,dx23);
1811             ty               = _mm_mul_ps(fscal,dy23);
1812             tz               = _mm_mul_ps(fscal,dz23);
1813
1814             /* Update vectorial force */
1815             fix2             = _mm_add_ps(fix2,tx);
1816             fiy2             = _mm_add_ps(fiy2,ty);
1817             fiz2             = _mm_add_ps(fiz2,tz);
1818
1819             fjx3             = _mm_add_ps(fjx3,tx);
1820             fjy3             = _mm_add_ps(fjy3,ty);
1821             fjz3             = _mm_add_ps(fjz3,tz);
1822
1823             /**************************
1824              * CALCULATE INTERACTIONS *
1825              **************************/
1826
1827             /* COULOMB ELECTROSTATICS */
1828             velec            = _mm_mul_ps(qq31,rinv31);
1829             felec            = _mm_mul_ps(velec,rinvsq31);
1830
1831             fscal            = felec;
1832
1833             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1834
1835             /* Calculate temporary vectorial force */
1836             tx               = _mm_mul_ps(fscal,dx31);
1837             ty               = _mm_mul_ps(fscal,dy31);
1838             tz               = _mm_mul_ps(fscal,dz31);
1839
1840             /* Update vectorial force */
1841             fix3             = _mm_add_ps(fix3,tx);
1842             fiy3             = _mm_add_ps(fiy3,ty);
1843             fiz3             = _mm_add_ps(fiz3,tz);
1844
1845             fjx1             = _mm_add_ps(fjx1,tx);
1846             fjy1             = _mm_add_ps(fjy1,ty);
1847             fjz1             = _mm_add_ps(fjz1,tz);
1848
1849             /**************************
1850              * CALCULATE INTERACTIONS *
1851              **************************/
1852
1853             /* COULOMB ELECTROSTATICS */
1854             velec            = _mm_mul_ps(qq32,rinv32);
1855             felec            = _mm_mul_ps(velec,rinvsq32);
1856
1857             fscal            = felec;
1858
1859             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1860
1861             /* Calculate temporary vectorial force */
1862             tx               = _mm_mul_ps(fscal,dx32);
1863             ty               = _mm_mul_ps(fscal,dy32);
1864             tz               = _mm_mul_ps(fscal,dz32);
1865
1866             /* Update vectorial force */
1867             fix3             = _mm_add_ps(fix3,tx);
1868             fiy3             = _mm_add_ps(fiy3,ty);
1869             fiz3             = _mm_add_ps(fiz3,tz);
1870
1871             fjx2             = _mm_add_ps(fjx2,tx);
1872             fjy2             = _mm_add_ps(fjy2,ty);
1873             fjz2             = _mm_add_ps(fjz2,tz);
1874
1875             /**************************
1876              * CALCULATE INTERACTIONS *
1877              **************************/
1878
1879             /* COULOMB ELECTROSTATICS */
1880             velec            = _mm_mul_ps(qq33,rinv33);
1881             felec            = _mm_mul_ps(velec,rinvsq33);
1882
1883             fscal            = felec;
1884
1885             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1886
1887             /* Calculate temporary vectorial force */
1888             tx               = _mm_mul_ps(fscal,dx33);
1889             ty               = _mm_mul_ps(fscal,dy33);
1890             tz               = _mm_mul_ps(fscal,dz33);
1891
1892             /* Update vectorial force */
1893             fix3             = _mm_add_ps(fix3,tx);
1894             fiy3             = _mm_add_ps(fiy3,ty);
1895             fiz3             = _mm_add_ps(fiz3,tz);
1896
1897             fjx3             = _mm_add_ps(fjx3,tx);
1898             fjy3             = _mm_add_ps(fjy3,ty);
1899             fjz3             = _mm_add_ps(fjz3,tz);
1900
1901             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1902             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1903             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1904             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1905
1906             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
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 24 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*24 + inneriter*273);
1930 }