src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRF_VdwLJ_GeomW3W3_sse4_1_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse4_1_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse4_1_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_sse4_1_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_sse4_1_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  91     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  93     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     int              nvdwtype;
 106     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 107     int              *vdwtype;
 108     real             *vdwparam;
 109     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 110     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 111     __m128           dummy_mask,cutoff_mask;
 112     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 113     __m128           one     = _mm_set1_ps(1.0);
 114     __m128           two     = _mm_set1_ps(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm_set1_ps(fr->ic->epsfac);
 127     charge           = mdatoms->chargeA;
 128     krf              = _mm_set1_ps(fr->ic->k_rf);
 129     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 130     crf              = _mm_set1_ps(fr->ic->c_rf);
 131     nvdwtype         = fr->ntype;
 132     vdwparam         = fr->nbfp;
 133     vdwtype          = mdatoms->typeA;
 134
 135     /* Setup water-specific parameters */
 136     inr              = nlist->iinr[0];
 137     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 138     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 139     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 140     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 141
 142     jq0              = _mm_set1_ps(charge[inr+0]);
 143     jq1              = _mm_set1_ps(charge[inr+1]);
 144     jq2              = _mm_set1_ps(charge[inr+2]);
 145     vdwjidx0A        = 2*vdwtype[inr+0];
 146     qq00             = _mm_mul_ps(iq0,jq0);
 147     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 148     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 149     qq01             = _mm_mul_ps(iq0,jq1);
 150     qq02             = _mm_mul_ps(iq0,jq2);
 151     qq10             = _mm_mul_ps(iq1,jq0);
 152     qq11             = _mm_mul_ps(iq1,jq1);
 153     qq12             = _mm_mul_ps(iq1,jq2);
 154     qq20             = _mm_mul_ps(iq2,jq0);
 155     qq21             = _mm_mul_ps(iq2,jq1);
 156     qq22             = _mm_mul_ps(iq2,jq2);
 157
 158     /* Avoid stupid compiler warnings */
 159     jnrA = jnrB = jnrC = jnrD = 0;
 160     j_coord_offsetA = 0;
 161     j_coord_offsetB = 0;
 162     j_coord_offsetC = 0;
 163     j_coord_offsetD = 0;
 164
 165     outeriter        = 0;
 166     inneriter        = 0;
 167
 168     for(iidx=0;iidx<4*DIM;iidx++)
 169     {
 170         scratch[iidx] = 0.0;
 171     }
 172
 173     /* Start outer loop over neighborlists */
 174     for(iidx=0; iidx<nri; iidx++)
 175     {
 176         /* Load shift vector for this list */
 177         i_shift_offset   = DIM*shiftidx[iidx];
 178
 179         /* Load limits for loop over neighbors */
 180         j_index_start    = jindex[iidx];
 181         j_index_end      = jindex[iidx+1];
 182
 183         /* Get outer coordinate index */
 184         inr              = iinr[iidx];
 185         i_coord_offset   = DIM*inr;
 186
 187         /* Load i particle coords and add shift vector */
 188         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 189                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 190
 191         fix0             = _mm_setzero_ps();
 192         fiy0             = _mm_setzero_ps();
 193         fiz0             = _mm_setzero_ps();
 194         fix1             = _mm_setzero_ps();
 195         fiy1             = _mm_setzero_ps();
 196         fiz1             = _mm_setzero_ps();
 197         fix2             = _mm_setzero_ps();
 198         fiy2             = _mm_setzero_ps();
 199         fiz2             = _mm_setzero_ps();
 200
 201         /* Reset potential sums */
 202         velecsum         = _mm_setzero_ps();
 203         vvdwsum          = _mm_setzero_ps();
 204
 205         /* Start inner kernel loop */
 206         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 207         {
 208
 209             /* Get j neighbor index, and coordinate index */
 210             jnrA             = jjnr[jidx];
 211             jnrB             = jjnr[jidx+1];
 212             jnrC             = jjnr[jidx+2];
 213             jnrD             = jjnr[jidx+3];
 214             j_coord_offsetA  = DIM*jnrA;
 215             j_coord_offsetB  = DIM*jnrB;
 216             j_coord_offsetC  = DIM*jnrC;
 217             j_coord_offsetD  = DIM*jnrD;
 218
 219             /* load j atom coordinates */
 220             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 221                                               x+j_coord_offsetC,x+j_coord_offsetD,
 222                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 223
 224             /* Calculate displacement vector */
 225             dx00             = _mm_sub_ps(ix0,jx0);
 226             dy00             = _mm_sub_ps(iy0,jy0);
 227             dz00             = _mm_sub_ps(iz0,jz0);
 228             dx01             = _mm_sub_ps(ix0,jx1);
 229             dy01             = _mm_sub_ps(iy0,jy1);
 230             dz01             = _mm_sub_ps(iz0,jz1);
 231             dx02             = _mm_sub_ps(ix0,jx2);
 232             dy02             = _mm_sub_ps(iy0,jy2);
 233             dz02             = _mm_sub_ps(iz0,jz2);
 234             dx10             = _mm_sub_ps(ix1,jx0);
 235             dy10             = _mm_sub_ps(iy1,jy0);
 236             dz10             = _mm_sub_ps(iz1,jz0);
 237             dx11             = _mm_sub_ps(ix1,jx1);
 238             dy11             = _mm_sub_ps(iy1,jy1);
 239             dz11             = _mm_sub_ps(iz1,jz1);
 240             dx12             = _mm_sub_ps(ix1,jx2);
 241             dy12             = _mm_sub_ps(iy1,jy2);
 242             dz12             = _mm_sub_ps(iz1,jz2);
 243             dx20             = _mm_sub_ps(ix2,jx0);
 244             dy20             = _mm_sub_ps(iy2,jy0);
 245             dz20             = _mm_sub_ps(iz2,jz0);
 246             dx21             = _mm_sub_ps(ix2,jx1);
 247             dy21             = _mm_sub_ps(iy2,jy1);
 248             dz21             = _mm_sub_ps(iz2,jz1);
 249             dx22             = _mm_sub_ps(ix2,jx2);
 250             dy22             = _mm_sub_ps(iy2,jy2);
 251             dz22             = _mm_sub_ps(iz2,jz2);
 252
 253             /* Calculate squared distance and things based on it */
 254             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 255             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 256             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 257             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 258             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 259             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 260             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 261             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 262             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 263
 264             rinv00           = sse41_invsqrt_f(rsq00);
 265             rinv01           = sse41_invsqrt_f(rsq01);
 266             rinv02           = sse41_invsqrt_f(rsq02);
 267             rinv10           = sse41_invsqrt_f(rsq10);
 268             rinv11           = sse41_invsqrt_f(rsq11);
 269             rinv12           = sse41_invsqrt_f(rsq12);
 270             rinv20           = sse41_invsqrt_f(rsq20);
 271             rinv21           = sse41_invsqrt_f(rsq21);
 272             rinv22           = sse41_invsqrt_f(rsq22);
 273
 274             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 275             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 276             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 277             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 278             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 279             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 280             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 281             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 282             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 283
 284             fjx0             = _mm_setzero_ps();
 285             fjy0             = _mm_setzero_ps();
 286             fjz0             = _mm_setzero_ps();
 287             fjx1             = _mm_setzero_ps();
 288             fjy1             = _mm_setzero_ps();
 289             fjz1             = _mm_setzero_ps();
 290             fjx2             = _mm_setzero_ps();
 291             fjy2             = _mm_setzero_ps();
 292             fjz2             = _mm_setzero_ps();
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             /* REACTION-FIELD ELECTROSTATICS */
 299             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 300             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 301
 302             /* LENNARD-JONES DISPERSION/REPULSION */
 303
 304             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 305             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 306             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 307             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 308             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_ps(velecsum,velec);
 312             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 313
 314             fscal            = _mm_add_ps(felec,fvdw);
 315
 316             /* Calculate temporary vectorial force */
 317             tx               = _mm_mul_ps(fscal,dx00);
 318             ty               = _mm_mul_ps(fscal,dy00);
 319             tz               = _mm_mul_ps(fscal,dz00);
 320
 321             /* Update vectorial force */
 322             fix0             = _mm_add_ps(fix0,tx);
 323             fiy0             = _mm_add_ps(fiy0,ty);
 324             fiz0             = _mm_add_ps(fiz0,tz);
 325
 326             fjx0             = _mm_add_ps(fjx0,tx);
 327             fjy0             = _mm_add_ps(fjy0,ty);
 328             fjz0             = _mm_add_ps(fjz0,tz);
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             /* REACTION-FIELD ELECTROSTATICS */
 335             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 336             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 337
 338             /* Update potential sum for this i atom from the interaction with this j atom. */
 339             velecsum         = _mm_add_ps(velecsum,velec);
 340
 341             fscal            = felec;
 342
 343             /* Calculate temporary vectorial force */
 344             tx               = _mm_mul_ps(fscal,dx01);
 345             ty               = _mm_mul_ps(fscal,dy01);
 346             tz               = _mm_mul_ps(fscal,dz01);
 347
 348             /* Update vectorial force */
 349             fix0             = _mm_add_ps(fix0,tx);
 350             fiy0             = _mm_add_ps(fiy0,ty);
 351             fiz0             = _mm_add_ps(fiz0,tz);
 352
 353             fjx1             = _mm_add_ps(fjx1,tx);
 354             fjy1             = _mm_add_ps(fjy1,ty);
 355             fjz1             = _mm_add_ps(fjz1,tz);
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             /* REACTION-FIELD ELECTROSTATICS */
 362             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 363             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             velecsum         = _mm_add_ps(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370             /* Calculate temporary vectorial force */
 371             tx               = _mm_mul_ps(fscal,dx02);
 372             ty               = _mm_mul_ps(fscal,dy02);
 373             tz               = _mm_mul_ps(fscal,dz02);
 374
 375             /* Update vectorial force */
 376             fix0             = _mm_add_ps(fix0,tx);
 377             fiy0             = _mm_add_ps(fiy0,ty);
 378             fiz0             = _mm_add_ps(fiz0,tz);
 379
 380             fjx2             = _mm_add_ps(fjx2,tx);
 381             fjy2             = _mm_add_ps(fjy2,ty);
 382             fjz2             = _mm_add_ps(fjz2,tz);
 383
 384             /**************************
 385              * CALCULATE INTERACTIONS *
 386              **************************/
 387
 388             /* REACTION-FIELD ELECTROSTATICS */
 389             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 390             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 391
 392             /* Update potential sum for this i atom from the interaction with this j atom. */
 393             velecsum         = _mm_add_ps(velecsum,velec);
 394
 395             fscal            = felec;
 396
 397             /* Calculate temporary vectorial force */
 398             tx               = _mm_mul_ps(fscal,dx10);
 399             ty               = _mm_mul_ps(fscal,dy10);
 400             tz               = _mm_mul_ps(fscal,dz10);
 401
 402             /* Update vectorial force */
 403             fix1             = _mm_add_ps(fix1,tx);
 404             fiy1             = _mm_add_ps(fiy1,ty);
 405             fiz1             = _mm_add_ps(fiz1,tz);
 406
 407             fjx0             = _mm_add_ps(fjx0,tx);
 408             fjy0             = _mm_add_ps(fjy0,ty);
 409             fjz0             = _mm_add_ps(fjz0,tz);
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             /* REACTION-FIELD ELECTROSTATICS */
 416             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 417             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velecsum         = _mm_add_ps(velecsum,velec);
 421
 422             fscal            = felec;
 423
 424             /* Calculate temporary vectorial force */
 425             tx               = _mm_mul_ps(fscal,dx11);
 426             ty               = _mm_mul_ps(fscal,dy11);
 427             tz               = _mm_mul_ps(fscal,dz11);
 428
 429             /* Update vectorial force */
 430             fix1             = _mm_add_ps(fix1,tx);
 431             fiy1             = _mm_add_ps(fiy1,ty);
 432             fiz1             = _mm_add_ps(fiz1,tz);
 433
 434             fjx1             = _mm_add_ps(fjx1,tx);
 435             fjy1             = _mm_add_ps(fjy1,ty);
 436             fjz1             = _mm_add_ps(fjz1,tz);
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             /* REACTION-FIELD ELECTROSTATICS */
 443             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 444             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velecsum         = _mm_add_ps(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             /* Calculate temporary vectorial force */
 452             tx               = _mm_mul_ps(fscal,dx12);
 453             ty               = _mm_mul_ps(fscal,dy12);
 454             tz               = _mm_mul_ps(fscal,dz12);
 455
 456             /* Update vectorial force */
 457             fix1             = _mm_add_ps(fix1,tx);
 458             fiy1             = _mm_add_ps(fiy1,ty);
 459             fiz1             = _mm_add_ps(fiz1,tz);
 460
 461             fjx2             = _mm_add_ps(fjx2,tx);
 462             fjy2             = _mm_add_ps(fjy2,ty);
 463             fjz2             = _mm_add_ps(fjz2,tz);
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             /* REACTION-FIELD ELECTROSTATICS */
 470             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 471             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 472
 473             /* Update potential sum for this i atom from the interaction with this j atom. */
 474             velecsum         = _mm_add_ps(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478             /* Calculate temporary vectorial force */
 479             tx               = _mm_mul_ps(fscal,dx20);
 480             ty               = _mm_mul_ps(fscal,dy20);
 481             tz               = _mm_mul_ps(fscal,dz20);
 482
 483             /* Update vectorial force */
 484             fix2             = _mm_add_ps(fix2,tx);
 485             fiy2             = _mm_add_ps(fiy2,ty);
 486             fiz2             = _mm_add_ps(fiz2,tz);
 487
 488             fjx0             = _mm_add_ps(fjx0,tx);
 489             fjy0             = _mm_add_ps(fjy0,ty);
 490             fjz0             = _mm_add_ps(fjz0,tz);
 491
 492             /**************************
 493              * CALCULATE INTERACTIONS *
 494              **************************/
 495
 496             /* REACTION-FIELD ELECTROSTATICS */
 497             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 498             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 499
 500             /* Update potential sum for this i atom from the interaction with this j atom. */
 501             velecsum         = _mm_add_ps(velecsum,velec);
 502
 503             fscal            = felec;
 504
 505             /* Calculate temporary vectorial force */
 506             tx               = _mm_mul_ps(fscal,dx21);
 507             ty               = _mm_mul_ps(fscal,dy21);
 508             tz               = _mm_mul_ps(fscal,dz21);
 509
 510             /* Update vectorial force */
 511             fix2             = _mm_add_ps(fix2,tx);
 512             fiy2             = _mm_add_ps(fiy2,ty);
 513             fiz2             = _mm_add_ps(fiz2,tz);
 514
 515             fjx1             = _mm_add_ps(fjx1,tx);
 516             fjy1             = _mm_add_ps(fjy1,ty);
 517             fjz1             = _mm_add_ps(fjz1,tz);
 518
 519             /**************************
 520              * CALCULATE INTERACTIONS *
 521              **************************/
 522
 523             /* REACTION-FIELD ELECTROSTATICS */
 524             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 525             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 526
 527             /* Update potential sum for this i atom from the interaction with this j atom. */
 528             velecsum         = _mm_add_ps(velecsum,velec);
 529
 530             fscal            = felec;
 531
 532             /* Calculate temporary vectorial force */
 533             tx               = _mm_mul_ps(fscal,dx22);
 534             ty               = _mm_mul_ps(fscal,dy22);
 535             tz               = _mm_mul_ps(fscal,dz22);
 536
 537             /* Update vectorial force */
 538             fix2             = _mm_add_ps(fix2,tx);
 539             fiy2             = _mm_add_ps(fiy2,ty);
 540             fiz2             = _mm_add_ps(fiz2,tz);
 541
 542             fjx2             = _mm_add_ps(fjx2,tx);
 543             fjy2             = _mm_add_ps(fjy2,ty);
 544             fjz2             = _mm_add_ps(fjz2,tz);
 545
 546             fjptrA             = f+j_coord_offsetA;
 547             fjptrB             = f+j_coord_offsetB;
 548             fjptrC             = f+j_coord_offsetC;
 549             fjptrD             = f+j_coord_offsetD;
 550
 551             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 552                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 553
 554             /* Inner loop uses 300 flops */
 555         }
 556
 557         if(jidx<j_index_end)
 558         {
 559
 560             /* Get j neighbor index, and coordinate index */
 561             jnrlistA         = jjnr[jidx];
 562             jnrlistB         = jjnr[jidx+1];
 563             jnrlistC         = jjnr[jidx+2];
 564             jnrlistD         = jjnr[jidx+3];
 565             /* Sign of each element will be negative for non-real atoms.
 566              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 567              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 568              */
 569             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 570             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 571             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 572             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 573             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 574             j_coord_offsetA  = DIM*jnrA;
 575             j_coord_offsetB  = DIM*jnrB;
 576             j_coord_offsetC  = DIM*jnrC;
 577             j_coord_offsetD  = DIM*jnrD;
 578
 579             /* load j atom coordinates */
 580             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 581                                               x+j_coord_offsetC,x+j_coord_offsetD,
 582                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 583
 584             /* Calculate displacement vector */
 585             dx00             = _mm_sub_ps(ix0,jx0);
 586             dy00             = _mm_sub_ps(iy0,jy0);
 587             dz00             = _mm_sub_ps(iz0,jz0);
 588             dx01             = _mm_sub_ps(ix0,jx1);
 589             dy01             = _mm_sub_ps(iy0,jy1);
 590             dz01             = _mm_sub_ps(iz0,jz1);
 591             dx02             = _mm_sub_ps(ix0,jx2);
 592             dy02             = _mm_sub_ps(iy0,jy2);
 593             dz02             = _mm_sub_ps(iz0,jz2);
 594             dx10             = _mm_sub_ps(ix1,jx0);
 595             dy10             = _mm_sub_ps(iy1,jy0);
 596             dz10             = _mm_sub_ps(iz1,jz0);
 597             dx11             = _mm_sub_ps(ix1,jx1);
 598             dy11             = _mm_sub_ps(iy1,jy1);
 599             dz11             = _mm_sub_ps(iz1,jz1);
 600             dx12             = _mm_sub_ps(ix1,jx2);
 601             dy12             = _mm_sub_ps(iy1,jy2);
 602             dz12             = _mm_sub_ps(iz1,jz2);
 603             dx20             = _mm_sub_ps(ix2,jx0);
 604             dy20             = _mm_sub_ps(iy2,jy0);
 605             dz20             = _mm_sub_ps(iz2,jz0);
 606             dx21             = _mm_sub_ps(ix2,jx1);
 607             dy21             = _mm_sub_ps(iy2,jy1);
 608             dz21             = _mm_sub_ps(iz2,jz1);
 609             dx22             = _mm_sub_ps(ix2,jx2);
 610             dy22             = _mm_sub_ps(iy2,jy2);
 611             dz22             = _mm_sub_ps(iz2,jz2);
 612
 613             /* Calculate squared distance and things based on it */
 614             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 615             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 616             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 617             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 618             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 619             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 620             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 621             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 622             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 623
 624             rinv00           = sse41_invsqrt_f(rsq00);
 625             rinv01           = sse41_invsqrt_f(rsq01);
 626             rinv02           = sse41_invsqrt_f(rsq02);
 627             rinv10           = sse41_invsqrt_f(rsq10);
 628             rinv11           = sse41_invsqrt_f(rsq11);
 629             rinv12           = sse41_invsqrt_f(rsq12);
 630             rinv20           = sse41_invsqrt_f(rsq20);
 631             rinv21           = sse41_invsqrt_f(rsq21);
 632             rinv22           = sse41_invsqrt_f(rsq22);
 633
 634             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 635             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 636             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 637             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 638             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 639             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 640             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 641             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 642             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 643
 644             fjx0             = _mm_setzero_ps();
 645             fjy0             = _mm_setzero_ps();
 646             fjz0             = _mm_setzero_ps();
 647             fjx1             = _mm_setzero_ps();
 648             fjy1             = _mm_setzero_ps();
 649             fjz1             = _mm_setzero_ps();
 650             fjx2             = _mm_setzero_ps();
 651             fjy2             = _mm_setzero_ps();
 652             fjz2             = _mm_setzero_ps();
 653
 654             /**************************
 655              * CALCULATE INTERACTIONS *
 656              **************************/
 657
 658             /* REACTION-FIELD ELECTROSTATICS */
 659             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 660             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 661
 662             /* LENNARD-JONES DISPERSION/REPULSION */
 663
 664             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 665             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 666             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 667             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 668             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 669
 670             /* Update potential sum for this i atom from the interaction with this j atom. */
 671             velec            = _mm_andnot_ps(dummy_mask,velec);
 672             velecsum         = _mm_add_ps(velecsum,velec);
 673             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 674             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 675
 676             fscal            = _mm_add_ps(felec,fvdw);
 677
 678             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 679
 680             /* Calculate temporary vectorial force */
 681             tx               = _mm_mul_ps(fscal,dx00);
 682             ty               = _mm_mul_ps(fscal,dy00);
 683             tz               = _mm_mul_ps(fscal,dz00);
 684
 685             /* Update vectorial force */
 686             fix0             = _mm_add_ps(fix0,tx);
 687             fiy0             = _mm_add_ps(fiy0,ty);
 688             fiz0             = _mm_add_ps(fiz0,tz);
 689
 690             fjx0             = _mm_add_ps(fjx0,tx);
 691             fjy0             = _mm_add_ps(fjy0,ty);
 692             fjz0             = _mm_add_ps(fjz0,tz);
 693
 694             /**************************
 695              * CALCULATE INTERACTIONS *
 696              **************************/
 697
 698             /* REACTION-FIELD ELECTROSTATICS */
 699             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 700             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 701
 702             /* Update potential sum for this i atom from the interaction with this j atom. */
 703             velec            = _mm_andnot_ps(dummy_mask,velec);
 704             velecsum         = _mm_add_ps(velecsum,velec);
 705
 706             fscal            = felec;
 707
 708             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 709
 710             /* Calculate temporary vectorial force */
 711             tx               = _mm_mul_ps(fscal,dx01);
 712             ty               = _mm_mul_ps(fscal,dy01);
 713             tz               = _mm_mul_ps(fscal,dz01);
 714
 715             /* Update vectorial force */
 716             fix0             = _mm_add_ps(fix0,tx);
 717             fiy0             = _mm_add_ps(fiy0,ty);
 718             fiz0             = _mm_add_ps(fiz0,tz);
 719
 720             fjx1             = _mm_add_ps(fjx1,tx);
 721             fjy1             = _mm_add_ps(fjy1,ty);
 722             fjz1             = _mm_add_ps(fjz1,tz);
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             /* REACTION-FIELD ELECTROSTATICS */
 729             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 730             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 731
 732             /* Update potential sum for this i atom from the interaction with this j atom. */
 733             velec            = _mm_andnot_ps(dummy_mask,velec);
 734             velecsum         = _mm_add_ps(velecsum,velec);
 735
 736             fscal            = felec;
 737
 738             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 739
 740             /* Calculate temporary vectorial force */
 741             tx               = _mm_mul_ps(fscal,dx02);
 742             ty               = _mm_mul_ps(fscal,dy02);
 743             tz               = _mm_mul_ps(fscal,dz02);
 744
 745             /* Update vectorial force */
 746             fix0             = _mm_add_ps(fix0,tx);
 747             fiy0             = _mm_add_ps(fiy0,ty);
 748             fiz0             = _mm_add_ps(fiz0,tz);
 749
 750             fjx2             = _mm_add_ps(fjx2,tx);
 751             fjy2             = _mm_add_ps(fjy2,ty);
 752             fjz2             = _mm_add_ps(fjz2,tz);
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             /* REACTION-FIELD ELECTROSTATICS */
 759             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 760             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 761
 762             /* Update potential sum for this i atom from the interaction with this j atom. */
 763             velec            = _mm_andnot_ps(dummy_mask,velec);
 764             velecsum         = _mm_add_ps(velecsum,velec);
 765
 766             fscal            = felec;
 767
 768             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 769
 770             /* Calculate temporary vectorial force */
 771             tx               = _mm_mul_ps(fscal,dx10);
 772             ty               = _mm_mul_ps(fscal,dy10);
 773             tz               = _mm_mul_ps(fscal,dz10);
 774
 775             /* Update vectorial force */
 776             fix1             = _mm_add_ps(fix1,tx);
 777             fiy1             = _mm_add_ps(fiy1,ty);
 778             fiz1             = _mm_add_ps(fiz1,tz);
 779
 780             fjx0             = _mm_add_ps(fjx0,tx);
 781             fjy0             = _mm_add_ps(fjy0,ty);
 782             fjz0             = _mm_add_ps(fjz0,tz);
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             /* REACTION-FIELD ELECTROSTATICS */
 789             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 790             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 791
 792             /* Update potential sum for this i atom from the interaction with this j atom. */
 793             velec            = _mm_andnot_ps(dummy_mask,velec);
 794             velecsum         = _mm_add_ps(velecsum,velec);
 795
 796             fscal            = felec;
 797
 798             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 799
 800             /* Calculate temporary vectorial force */
 801             tx               = _mm_mul_ps(fscal,dx11);
 802             ty               = _mm_mul_ps(fscal,dy11);
 803             tz               = _mm_mul_ps(fscal,dz11);
 804
 805             /* Update vectorial force */
 806             fix1             = _mm_add_ps(fix1,tx);
 807             fiy1             = _mm_add_ps(fiy1,ty);
 808             fiz1             = _mm_add_ps(fiz1,tz);
 809
 810             fjx1             = _mm_add_ps(fjx1,tx);
 811             fjy1             = _mm_add_ps(fjy1,ty);
 812             fjz1             = _mm_add_ps(fjz1,tz);
 813
 814             /**************************
 815              * CALCULATE INTERACTIONS *
 816              **************************/
 817
 818             /* REACTION-FIELD ELECTROSTATICS */
 819             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 820             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 821
 822             /* Update potential sum for this i atom from the interaction with this j atom. */
 823             velec            = _mm_andnot_ps(dummy_mask,velec);
 824             velecsum         = _mm_add_ps(velecsum,velec);
 825
 826             fscal            = felec;
 827
 828             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 829
 830             /* Calculate temporary vectorial force */
 831             tx               = _mm_mul_ps(fscal,dx12);
 832             ty               = _mm_mul_ps(fscal,dy12);
 833             tz               = _mm_mul_ps(fscal,dz12);
 834
 835             /* Update vectorial force */
 836             fix1             = _mm_add_ps(fix1,tx);
 837             fiy1             = _mm_add_ps(fiy1,ty);
 838             fiz1             = _mm_add_ps(fiz1,tz);
 839
 840             fjx2             = _mm_add_ps(fjx2,tx);
 841             fjy2             = _mm_add_ps(fjy2,ty);
 842             fjz2             = _mm_add_ps(fjz2,tz);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* REACTION-FIELD ELECTROSTATICS */
 849             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 850             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 851
 852             /* Update potential sum for this i atom from the interaction with this j atom. */
 853             velec            = _mm_andnot_ps(dummy_mask,velec);
 854             velecsum         = _mm_add_ps(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 859
 860             /* Calculate temporary vectorial force */
 861             tx               = _mm_mul_ps(fscal,dx20);
 862             ty               = _mm_mul_ps(fscal,dy20);
 863             tz               = _mm_mul_ps(fscal,dz20);
 864
 865             /* Update vectorial force */
 866             fix2             = _mm_add_ps(fix2,tx);
 867             fiy2             = _mm_add_ps(fiy2,ty);
 868             fiz2             = _mm_add_ps(fiz2,tz);
 869
 870             fjx0             = _mm_add_ps(fjx0,tx);
 871             fjy0             = _mm_add_ps(fjy0,ty);
 872             fjz0             = _mm_add_ps(fjz0,tz);
 873
 874             /**************************
 875              * CALCULATE INTERACTIONS *
 876              **************************/
 877
 878             /* REACTION-FIELD ELECTROSTATICS */
 879             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 880             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 881
 882             /* Update potential sum for this i atom from the interaction with this j atom. */
 883             velec            = _mm_andnot_ps(dummy_mask,velec);
 884             velecsum         = _mm_add_ps(velecsum,velec);
 885
 886             fscal            = felec;
 887
 888             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 889
 890             /* Calculate temporary vectorial force */
 891             tx               = _mm_mul_ps(fscal,dx21);
 892             ty               = _mm_mul_ps(fscal,dy21);
 893             tz               = _mm_mul_ps(fscal,dz21);
 894
 895             /* Update vectorial force */
 896             fix2             = _mm_add_ps(fix2,tx);
 897             fiy2             = _mm_add_ps(fiy2,ty);
 898             fiz2             = _mm_add_ps(fiz2,tz);
 899
 900             fjx1             = _mm_add_ps(fjx1,tx);
 901             fjy1             = _mm_add_ps(fjy1,ty);
 902             fjz1             = _mm_add_ps(fjz1,tz);
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             /* REACTION-FIELD ELECTROSTATICS */
 909             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 910             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 911
 912             /* Update potential sum for this i atom from the interaction with this j atom. */
 913             velec            = _mm_andnot_ps(dummy_mask,velec);
 914             velecsum         = _mm_add_ps(velecsum,velec);
 915
 916             fscal            = felec;
 917
 918             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 919
 920             /* Calculate temporary vectorial force */
 921             tx               = _mm_mul_ps(fscal,dx22);
 922             ty               = _mm_mul_ps(fscal,dy22);
 923             tz               = _mm_mul_ps(fscal,dz22);
 924
 925             /* Update vectorial force */
 926             fix2             = _mm_add_ps(fix2,tx);
 927             fiy2             = _mm_add_ps(fiy2,ty);
 928             fiz2             = _mm_add_ps(fiz2,tz);
 929
 930             fjx2             = _mm_add_ps(fjx2,tx);
 931             fjy2             = _mm_add_ps(fjy2,ty);
 932             fjz2             = _mm_add_ps(fjz2,tz);
 933
 934             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 935             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 936             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 937             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 938
 939             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 940                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 941
 942             /* Inner loop uses 300 flops */
 943         }
 944
 945         /* End of innermost loop */
 946
 947         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 948                                               f+i_coord_offset,fshift+i_shift_offset);
 949
 950         ggid                        = gid[iidx];
 951         /* Update potential energies */
 952         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 953         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 954
 955         /* Increment number of inner iterations */
 956         inneriter                  += j_index_end - j_index_start;
 957
 958         /* Outer loop uses 20 flops */
 959     }
 960
 961     /* Increment number of outer iterations */
 962     outeriter        += nri;
 963
 964     /* Update outer/inner flops */
 965
 966     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*300);
 967 }
 968 /*
 969  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse4_1_single
 970  * Electrostatics interaction: ReactionField
 971  * VdW interaction:            LennardJones
 972  * Geometry:                   Water3-Water3
 973  * Calculate force/pot:        Force
 974  */
 975 void
 976 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse4_1_single
 977                     (t_nblist                    * gmx_restrict       nlist,
 978                      rvec                        * gmx_restrict          xx,
 979                      rvec                        * gmx_restrict          ff,
 980                      struct t_forcerec           * gmx_restrict          fr,
 981                      t_mdatoms                   * gmx_restrict     mdatoms,
 982                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 983                      t_nrnb                      * gmx_restrict        nrnb)
 984 {
 985     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 986      * just 0 for non-waters.
 987      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 988      * jnr indices corresponding to data put in the four positions in the SIMD register.
 989      */
 990     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 991     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 992     int              jnrA,jnrB,jnrC,jnrD;
 993     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 994     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 995     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 996     real             rcutoff_scalar;
 997     real             *shiftvec,*fshift,*x,*f;
 998     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 999     real             scratch[4*DIM];
1000     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1001     int              vdwioffset0;
1002     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1003     int              vdwioffset1;
1004     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1005     int              vdwioffset2;
1006     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1007     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1008     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1009     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1010     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1011     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1012     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1013     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1014     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1015     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1016     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1017     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1018     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1019     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1020     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1021     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1022     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1023     real             *charge;
1024     int              nvdwtype;
1025     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1026     int              *vdwtype;
1027     real             *vdwparam;
1028     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1029     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1030     __m128           dummy_mask,cutoff_mask;
1031     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1032     __m128           one     = _mm_set1_ps(1.0);
1033     __m128           two     = _mm_set1_ps(2.0);
1034     x                = xx[0];
1035     f                = ff[0];
1036
1037     nri              = nlist->nri;
1038     iinr             = nlist->iinr;
1039     jindex           = nlist->jindex;
1040     jjnr             = nlist->jjnr;
1041     shiftidx         = nlist->shift;
1042     gid              = nlist->gid;
1043     shiftvec         = fr->shift_vec[0];
1044     fshift           = fr->fshift[0];
1045     facel            = _mm_set1_ps(fr->ic->epsfac);
1046     charge           = mdatoms->chargeA;
1047     krf              = _mm_set1_ps(fr->ic->k_rf);
1048     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1049     crf              = _mm_set1_ps(fr->ic->c_rf);
1050     nvdwtype         = fr->ntype;
1051     vdwparam         = fr->nbfp;
1052     vdwtype          = mdatoms->typeA;
1053
1054     /* Setup water-specific parameters */
1055     inr              = nlist->iinr[0];
1056     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1057     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1058     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1059     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1060
1061     jq0              = _mm_set1_ps(charge[inr+0]);
1062     jq1              = _mm_set1_ps(charge[inr+1]);
1063     jq2              = _mm_set1_ps(charge[inr+2]);
1064     vdwjidx0A        = 2*vdwtype[inr+0];
1065     qq00             = _mm_mul_ps(iq0,jq0);
1066     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1067     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1068     qq01             = _mm_mul_ps(iq0,jq1);
1069     qq02             = _mm_mul_ps(iq0,jq2);
1070     qq10             = _mm_mul_ps(iq1,jq0);
1071     qq11             = _mm_mul_ps(iq1,jq1);
1072     qq12             = _mm_mul_ps(iq1,jq2);
1073     qq20             = _mm_mul_ps(iq2,jq0);
1074     qq21             = _mm_mul_ps(iq2,jq1);
1075     qq22             = _mm_mul_ps(iq2,jq2);
1076
1077     /* Avoid stupid compiler warnings */
1078     jnrA = jnrB = jnrC = jnrD = 0;
1079     j_coord_offsetA = 0;
1080     j_coord_offsetB = 0;
1081     j_coord_offsetC = 0;
1082     j_coord_offsetD = 0;
1083
1084     outeriter        = 0;
1085     inneriter        = 0;
1086
1087     for(iidx=0;iidx<4*DIM;iidx++)
1088     {
1089         scratch[iidx] = 0.0;
1090     }
1091
1092     /* Start outer loop over neighborlists */
1093     for(iidx=0; iidx<nri; iidx++)
1094     {
1095         /* Load shift vector for this list */
1096         i_shift_offset   = DIM*shiftidx[iidx];
1097
1098         /* Load limits for loop over neighbors */
1099         j_index_start    = jindex[iidx];
1100         j_index_end      = jindex[iidx+1];
1101
1102         /* Get outer coordinate index */
1103         inr              = iinr[iidx];
1104         i_coord_offset   = DIM*inr;
1105
1106         /* Load i particle coords and add shift vector */
1107         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1108                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1109
1110         fix0             = _mm_setzero_ps();
1111         fiy0             = _mm_setzero_ps();
1112         fiz0             = _mm_setzero_ps();
1113         fix1             = _mm_setzero_ps();
1114         fiy1             = _mm_setzero_ps();
1115         fiz1             = _mm_setzero_ps();
1116         fix2             = _mm_setzero_ps();
1117         fiy2             = _mm_setzero_ps();
1118         fiz2             = _mm_setzero_ps();
1119
1120         /* Start inner kernel loop */
1121         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1122         {
1123
1124             /* Get j neighbor index, and coordinate index */
1125             jnrA             = jjnr[jidx];
1126             jnrB             = jjnr[jidx+1];
1127             jnrC             = jjnr[jidx+2];
1128             jnrD             = jjnr[jidx+3];
1129             j_coord_offsetA  = DIM*jnrA;
1130             j_coord_offsetB  = DIM*jnrB;
1131             j_coord_offsetC  = DIM*jnrC;
1132             j_coord_offsetD  = DIM*jnrD;
1133
1134             /* load j atom coordinates */
1135             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1136                                               x+j_coord_offsetC,x+j_coord_offsetD,
1137                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1138
1139             /* Calculate displacement vector */
1140             dx00             = _mm_sub_ps(ix0,jx0);
1141             dy00             = _mm_sub_ps(iy0,jy0);
1142             dz00             = _mm_sub_ps(iz0,jz0);
1143             dx01             = _mm_sub_ps(ix0,jx1);
1144             dy01             = _mm_sub_ps(iy0,jy1);
1145             dz01             = _mm_sub_ps(iz0,jz1);
1146             dx02             = _mm_sub_ps(ix0,jx2);
1147             dy02             = _mm_sub_ps(iy0,jy2);
1148             dz02             = _mm_sub_ps(iz0,jz2);
1149             dx10             = _mm_sub_ps(ix1,jx0);
1150             dy10             = _mm_sub_ps(iy1,jy0);
1151             dz10             = _mm_sub_ps(iz1,jz0);
1152             dx11             = _mm_sub_ps(ix1,jx1);
1153             dy11             = _mm_sub_ps(iy1,jy1);
1154             dz11             = _mm_sub_ps(iz1,jz1);
1155             dx12             = _mm_sub_ps(ix1,jx2);
1156             dy12             = _mm_sub_ps(iy1,jy2);
1157             dz12             = _mm_sub_ps(iz1,jz2);
1158             dx20             = _mm_sub_ps(ix2,jx0);
1159             dy20             = _mm_sub_ps(iy2,jy0);
1160             dz20             = _mm_sub_ps(iz2,jz0);
1161             dx21             = _mm_sub_ps(ix2,jx1);
1162             dy21             = _mm_sub_ps(iy2,jy1);
1163             dz21             = _mm_sub_ps(iz2,jz1);
1164             dx22             = _mm_sub_ps(ix2,jx2);
1165             dy22             = _mm_sub_ps(iy2,jy2);
1166             dz22             = _mm_sub_ps(iz2,jz2);
1167
1168             /* Calculate squared distance and things based on it */
1169             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1170             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1171             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1172             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1173             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1174             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1175             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1176             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1177             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1178
1179             rinv00           = sse41_invsqrt_f(rsq00);
1180             rinv01           = sse41_invsqrt_f(rsq01);
1181             rinv02           = sse41_invsqrt_f(rsq02);
1182             rinv10           = sse41_invsqrt_f(rsq10);
1183             rinv11           = sse41_invsqrt_f(rsq11);
1184             rinv12           = sse41_invsqrt_f(rsq12);
1185             rinv20           = sse41_invsqrt_f(rsq20);
1186             rinv21           = sse41_invsqrt_f(rsq21);
1187             rinv22           = sse41_invsqrt_f(rsq22);
1188
1189             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1190             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1191             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1192             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1193             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1194             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1195             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1196             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1197             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1198
1199             fjx0             = _mm_setzero_ps();
1200             fjy0             = _mm_setzero_ps();
1201             fjz0             = _mm_setzero_ps();
1202             fjx1             = _mm_setzero_ps();
1203             fjy1             = _mm_setzero_ps();
1204             fjz1             = _mm_setzero_ps();
1205             fjx2             = _mm_setzero_ps();
1206             fjy2             = _mm_setzero_ps();
1207             fjz2             = _mm_setzero_ps();
1208
1209             /**************************
1210              * CALCULATE INTERACTIONS *
1211              **************************/
1212
1213             /* REACTION-FIELD ELECTROSTATICS */
1214             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1215
1216             /* LENNARD-JONES DISPERSION/REPULSION */
1217
1218             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1219             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1220
1221             fscal            = _mm_add_ps(felec,fvdw);
1222
1223             /* Calculate temporary vectorial force */
1224             tx               = _mm_mul_ps(fscal,dx00);
1225             ty               = _mm_mul_ps(fscal,dy00);
1226             tz               = _mm_mul_ps(fscal,dz00);
1227
1228             /* Update vectorial force */
1229             fix0             = _mm_add_ps(fix0,tx);
1230             fiy0             = _mm_add_ps(fiy0,ty);
1231             fiz0             = _mm_add_ps(fiz0,tz);
1232
1233             fjx0             = _mm_add_ps(fjx0,tx);
1234             fjy0             = _mm_add_ps(fjy0,ty);
1235             fjz0             = _mm_add_ps(fjz0,tz);
1236
1237             /**************************
1238              * CALCULATE INTERACTIONS *
1239              **************************/
1240
1241             /* REACTION-FIELD ELECTROSTATICS */
1242             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1243
1244             fscal            = felec;
1245
1246             /* Calculate temporary vectorial force */
1247             tx               = _mm_mul_ps(fscal,dx01);
1248             ty               = _mm_mul_ps(fscal,dy01);
1249             tz               = _mm_mul_ps(fscal,dz01);
1250
1251             /* Update vectorial force */
1252             fix0             = _mm_add_ps(fix0,tx);
1253             fiy0             = _mm_add_ps(fiy0,ty);
1254             fiz0             = _mm_add_ps(fiz0,tz);
1255
1256             fjx1             = _mm_add_ps(fjx1,tx);
1257             fjy1             = _mm_add_ps(fjy1,ty);
1258             fjz1             = _mm_add_ps(fjz1,tz);
1259
1260             /**************************
1261              * CALCULATE INTERACTIONS *
1262              **************************/
1263
1264             /* REACTION-FIELD ELECTROSTATICS */
1265             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1266
1267             fscal            = felec;
1268
1269             /* Calculate temporary vectorial force */
1270             tx               = _mm_mul_ps(fscal,dx02);
1271             ty               = _mm_mul_ps(fscal,dy02);
1272             tz               = _mm_mul_ps(fscal,dz02);
1273
1274             /* Update vectorial force */
1275             fix0             = _mm_add_ps(fix0,tx);
1276             fiy0             = _mm_add_ps(fiy0,ty);
1277             fiz0             = _mm_add_ps(fiz0,tz);
1278
1279             fjx2             = _mm_add_ps(fjx2,tx);
1280             fjy2             = _mm_add_ps(fjy2,ty);
1281             fjz2             = _mm_add_ps(fjz2,tz);
1282
1283             /**************************
1284              * CALCULATE INTERACTIONS *
1285              **************************/
1286
1287             /* REACTION-FIELD ELECTROSTATICS */
1288             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1289
1290             fscal            = felec;
1291
1292             /* Calculate temporary vectorial force */
1293             tx               = _mm_mul_ps(fscal,dx10);
1294             ty               = _mm_mul_ps(fscal,dy10);
1295             tz               = _mm_mul_ps(fscal,dz10);
1296
1297             /* Update vectorial force */
1298             fix1             = _mm_add_ps(fix1,tx);
1299             fiy1             = _mm_add_ps(fiy1,ty);
1300             fiz1             = _mm_add_ps(fiz1,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             /* REACTION-FIELD ELECTROSTATICS */
1311             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1312
1313             fscal            = felec;
1314
1315             /* Calculate temporary vectorial force */
1316             tx               = _mm_mul_ps(fscal,dx11);
1317             ty               = _mm_mul_ps(fscal,dy11);
1318             tz               = _mm_mul_ps(fscal,dz11);
1319
1320             /* Update vectorial force */
1321             fix1             = _mm_add_ps(fix1,tx);
1322             fiy1             = _mm_add_ps(fiy1,ty);
1323             fiz1             = _mm_add_ps(fiz1,tz);
1324
1325             fjx1             = _mm_add_ps(fjx1,tx);
1326             fjy1             = _mm_add_ps(fjy1,ty);
1327             fjz1             = _mm_add_ps(fjz1,tz);
1328
1329             /**************************
1330              * CALCULATE INTERACTIONS *
1331              **************************/
1332
1333             /* REACTION-FIELD ELECTROSTATICS */
1334             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1335
1336             fscal            = felec;
1337
1338             /* Calculate temporary vectorial force */
1339             tx               = _mm_mul_ps(fscal,dx12);
1340             ty               = _mm_mul_ps(fscal,dy12);
1341             tz               = _mm_mul_ps(fscal,dz12);
1342
1343             /* Update vectorial force */
1344             fix1             = _mm_add_ps(fix1,tx);
1345             fiy1             = _mm_add_ps(fiy1,ty);
1346             fiz1             = _mm_add_ps(fiz1,tz);
1347
1348             fjx2             = _mm_add_ps(fjx2,tx);
1349             fjy2             = _mm_add_ps(fjy2,ty);
1350             fjz2             = _mm_add_ps(fjz2,tz);
1351
1352             /**************************
1353              * CALCULATE INTERACTIONS *
1354              **************************/
1355
1356             /* REACTION-FIELD ELECTROSTATICS */
1357             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1358
1359             fscal            = felec;
1360
1361             /* Calculate temporary vectorial force */
1362             tx               = _mm_mul_ps(fscal,dx20);
1363             ty               = _mm_mul_ps(fscal,dy20);
1364             tz               = _mm_mul_ps(fscal,dz20);
1365
1366             /* Update vectorial force */
1367             fix2             = _mm_add_ps(fix2,tx);
1368             fiy2             = _mm_add_ps(fiy2,ty);
1369             fiz2             = _mm_add_ps(fiz2,tz);
1370
1371             fjx0             = _mm_add_ps(fjx0,tx);
1372             fjy0             = _mm_add_ps(fjy0,ty);
1373             fjz0             = _mm_add_ps(fjz0,tz);
1374
1375             /**************************
1376              * CALCULATE INTERACTIONS *
1377              **************************/
1378
1379             /* REACTION-FIELD ELECTROSTATICS */
1380             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1381
1382             fscal            = felec;
1383
1384             /* Calculate temporary vectorial force */
1385             tx               = _mm_mul_ps(fscal,dx21);
1386             ty               = _mm_mul_ps(fscal,dy21);
1387             tz               = _mm_mul_ps(fscal,dz21);
1388
1389             /* Update vectorial force */
1390             fix2             = _mm_add_ps(fix2,tx);
1391             fiy2             = _mm_add_ps(fiy2,ty);
1392             fiz2             = _mm_add_ps(fiz2,tz);
1393
1394             fjx1             = _mm_add_ps(fjx1,tx);
1395             fjy1             = _mm_add_ps(fjy1,ty);
1396             fjz1             = _mm_add_ps(fjz1,tz);
1397
1398             /**************************
1399              * CALCULATE INTERACTIONS *
1400              **************************/
1401
1402             /* REACTION-FIELD ELECTROSTATICS */
1403             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1404
1405             fscal            = felec;
1406
1407             /* Calculate temporary vectorial force */
1408             tx               = _mm_mul_ps(fscal,dx22);
1409             ty               = _mm_mul_ps(fscal,dy22);
1410             tz               = _mm_mul_ps(fscal,dz22);
1411
1412             /* Update vectorial force */
1413             fix2             = _mm_add_ps(fix2,tx);
1414             fiy2             = _mm_add_ps(fiy2,ty);
1415             fiz2             = _mm_add_ps(fiz2,tz);
1416
1417             fjx2             = _mm_add_ps(fjx2,tx);
1418             fjy2             = _mm_add_ps(fjy2,ty);
1419             fjz2             = _mm_add_ps(fjz2,tz);
1420
1421             fjptrA             = f+j_coord_offsetA;
1422             fjptrB             = f+j_coord_offsetB;
1423             fjptrC             = f+j_coord_offsetC;
1424             fjptrD             = f+j_coord_offsetD;
1425
1426             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1427                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1428
1429             /* Inner loop uses 250 flops */
1430         }
1431
1432         if(jidx<j_index_end)
1433         {
1434
1435             /* Get j neighbor index, and coordinate index */
1436             jnrlistA         = jjnr[jidx];
1437             jnrlistB         = jjnr[jidx+1];
1438             jnrlistC         = jjnr[jidx+2];
1439             jnrlistD         = jjnr[jidx+3];
1440             /* Sign of each element will be negative for non-real atoms.
1441              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1442              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1443              */
1444             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1445             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1446             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1447             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1448             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1449             j_coord_offsetA  = DIM*jnrA;
1450             j_coord_offsetB  = DIM*jnrB;
1451             j_coord_offsetC  = DIM*jnrC;
1452             j_coord_offsetD  = DIM*jnrD;
1453
1454             /* load j atom coordinates */
1455             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1456                                               x+j_coord_offsetC,x+j_coord_offsetD,
1457                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1458
1459             /* Calculate displacement vector */
1460             dx00             = _mm_sub_ps(ix0,jx0);
1461             dy00             = _mm_sub_ps(iy0,jy0);
1462             dz00             = _mm_sub_ps(iz0,jz0);
1463             dx01             = _mm_sub_ps(ix0,jx1);
1464             dy01             = _mm_sub_ps(iy0,jy1);
1465             dz01             = _mm_sub_ps(iz0,jz1);
1466             dx02             = _mm_sub_ps(ix0,jx2);
1467             dy02             = _mm_sub_ps(iy0,jy2);
1468             dz02             = _mm_sub_ps(iz0,jz2);
1469             dx10             = _mm_sub_ps(ix1,jx0);
1470             dy10             = _mm_sub_ps(iy1,jy0);
1471             dz10             = _mm_sub_ps(iz1,jz0);
1472             dx11             = _mm_sub_ps(ix1,jx1);
1473             dy11             = _mm_sub_ps(iy1,jy1);
1474             dz11             = _mm_sub_ps(iz1,jz1);
1475             dx12             = _mm_sub_ps(ix1,jx2);
1476             dy12             = _mm_sub_ps(iy1,jy2);
1477             dz12             = _mm_sub_ps(iz1,jz2);
1478             dx20             = _mm_sub_ps(ix2,jx0);
1479             dy20             = _mm_sub_ps(iy2,jy0);
1480             dz20             = _mm_sub_ps(iz2,jz0);
1481             dx21             = _mm_sub_ps(ix2,jx1);
1482             dy21             = _mm_sub_ps(iy2,jy1);
1483             dz21             = _mm_sub_ps(iz2,jz1);
1484             dx22             = _mm_sub_ps(ix2,jx2);
1485             dy22             = _mm_sub_ps(iy2,jy2);
1486             dz22             = _mm_sub_ps(iz2,jz2);
1487
1488             /* Calculate squared distance and things based on it */
1489             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1490             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1491             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1492             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1493             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1494             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1495             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1496             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1497             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1498
1499             rinv00           = sse41_invsqrt_f(rsq00);
1500             rinv01           = sse41_invsqrt_f(rsq01);
1501             rinv02           = sse41_invsqrt_f(rsq02);
1502             rinv10           = sse41_invsqrt_f(rsq10);
1503             rinv11           = sse41_invsqrt_f(rsq11);
1504             rinv12           = sse41_invsqrt_f(rsq12);
1505             rinv20           = sse41_invsqrt_f(rsq20);
1506             rinv21           = sse41_invsqrt_f(rsq21);
1507             rinv22           = sse41_invsqrt_f(rsq22);
1508
1509             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1510             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1511             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1512             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1513             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1514             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1515             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1516             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1517             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1518
1519             fjx0             = _mm_setzero_ps();
1520             fjy0             = _mm_setzero_ps();
1521             fjz0             = _mm_setzero_ps();
1522             fjx1             = _mm_setzero_ps();
1523             fjy1             = _mm_setzero_ps();
1524             fjz1             = _mm_setzero_ps();
1525             fjx2             = _mm_setzero_ps();
1526             fjy2             = _mm_setzero_ps();
1527             fjz2             = _mm_setzero_ps();
1528
1529             /**************************
1530              * CALCULATE INTERACTIONS *
1531              **************************/
1532
1533             /* REACTION-FIELD ELECTROSTATICS */
1534             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1535
1536             /* LENNARD-JONES DISPERSION/REPULSION */
1537
1538             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1539             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1540
1541             fscal            = _mm_add_ps(felec,fvdw);
1542
1543             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1544
1545             /* Calculate temporary vectorial force */
1546             tx               = _mm_mul_ps(fscal,dx00);
1547             ty               = _mm_mul_ps(fscal,dy00);
1548             tz               = _mm_mul_ps(fscal,dz00);
1549
1550             /* Update vectorial force */
1551             fix0             = _mm_add_ps(fix0,tx);
1552             fiy0             = _mm_add_ps(fiy0,ty);
1553             fiz0             = _mm_add_ps(fiz0,tz);
1554
1555             fjx0             = _mm_add_ps(fjx0,tx);
1556             fjy0             = _mm_add_ps(fjy0,ty);
1557             fjz0             = _mm_add_ps(fjz0,tz);
1558
1559             /**************************
1560              * CALCULATE INTERACTIONS *
1561              **************************/
1562
1563             /* REACTION-FIELD ELECTROSTATICS */
1564             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1565
1566             fscal            = felec;
1567
1568             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1569
1570             /* Calculate temporary vectorial force */
1571             tx               = _mm_mul_ps(fscal,dx01);
1572             ty               = _mm_mul_ps(fscal,dy01);
1573             tz               = _mm_mul_ps(fscal,dz01);
1574
1575             /* Update vectorial force */
1576             fix0             = _mm_add_ps(fix0,tx);
1577             fiy0             = _mm_add_ps(fiy0,ty);
1578             fiz0             = _mm_add_ps(fiz0,tz);
1579
1580             fjx1             = _mm_add_ps(fjx1,tx);
1581             fjy1             = _mm_add_ps(fjy1,ty);
1582             fjz1             = _mm_add_ps(fjz1,tz);
1583
1584             /**************************
1585              * CALCULATE INTERACTIONS *
1586              **************************/
1587
1588             /* REACTION-FIELD ELECTROSTATICS */
1589             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1590
1591             fscal            = felec;
1592
1593             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1594
1595             /* Calculate temporary vectorial force */
1596             tx               = _mm_mul_ps(fscal,dx02);
1597             ty               = _mm_mul_ps(fscal,dy02);
1598             tz               = _mm_mul_ps(fscal,dz02);
1599
1600             /* Update vectorial force */
1601             fix0             = _mm_add_ps(fix0,tx);
1602             fiy0             = _mm_add_ps(fiy0,ty);
1603             fiz0             = _mm_add_ps(fiz0,tz);
1604
1605             fjx2             = _mm_add_ps(fjx2,tx);
1606             fjy2             = _mm_add_ps(fjy2,ty);
1607             fjz2             = _mm_add_ps(fjz2,tz);
1608
1609             /**************************
1610              * CALCULATE INTERACTIONS *
1611              **************************/
1612
1613             /* REACTION-FIELD ELECTROSTATICS */
1614             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1615
1616             fscal            = felec;
1617
1618             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1619
1620             /* Calculate temporary vectorial force */
1621             tx               = _mm_mul_ps(fscal,dx10);
1622             ty               = _mm_mul_ps(fscal,dy10);
1623             tz               = _mm_mul_ps(fscal,dz10);
1624
1625             /* Update vectorial force */
1626             fix1             = _mm_add_ps(fix1,tx);
1627             fiy1             = _mm_add_ps(fiy1,ty);
1628             fiz1             = _mm_add_ps(fiz1,tz);
1629
1630             fjx0             = _mm_add_ps(fjx0,tx);
1631             fjy0             = _mm_add_ps(fjy0,ty);
1632             fjz0             = _mm_add_ps(fjz0,tz);
1633
1634             /**************************
1635              * CALCULATE INTERACTIONS *
1636              **************************/
1637
1638             /* REACTION-FIELD ELECTROSTATICS */
1639             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1640
1641             fscal            = felec;
1642
1643             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1644
1645             /* Calculate temporary vectorial force */
1646             tx               = _mm_mul_ps(fscal,dx11);
1647             ty               = _mm_mul_ps(fscal,dy11);
1648             tz               = _mm_mul_ps(fscal,dz11);
1649
1650             /* Update vectorial force */
1651             fix1             = _mm_add_ps(fix1,tx);
1652             fiy1             = _mm_add_ps(fiy1,ty);
1653             fiz1             = _mm_add_ps(fiz1,tz);
1654
1655             fjx1             = _mm_add_ps(fjx1,tx);
1656             fjy1             = _mm_add_ps(fjy1,ty);
1657             fjz1             = _mm_add_ps(fjz1,tz);
1658
1659             /**************************
1660              * CALCULATE INTERACTIONS *
1661              **************************/
1662
1663             /* REACTION-FIELD ELECTROSTATICS */
1664             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1665
1666             fscal            = felec;
1667
1668             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1669
1670             /* Calculate temporary vectorial force */
1671             tx               = _mm_mul_ps(fscal,dx12);
1672             ty               = _mm_mul_ps(fscal,dy12);
1673             tz               = _mm_mul_ps(fscal,dz12);
1674
1675             /* Update vectorial force */
1676             fix1             = _mm_add_ps(fix1,tx);
1677             fiy1             = _mm_add_ps(fiy1,ty);
1678             fiz1             = _mm_add_ps(fiz1,tz);
1679
1680             fjx2             = _mm_add_ps(fjx2,tx);
1681             fjy2             = _mm_add_ps(fjy2,ty);
1682             fjz2             = _mm_add_ps(fjz2,tz);
1683
1684             /**************************
1685              * CALCULATE INTERACTIONS *
1686              **************************/
1687
1688             /* REACTION-FIELD ELECTROSTATICS */
1689             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1690
1691             fscal            = felec;
1692
1693             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1694
1695             /* Calculate temporary vectorial force */
1696             tx               = _mm_mul_ps(fscal,dx20);
1697             ty               = _mm_mul_ps(fscal,dy20);
1698             tz               = _mm_mul_ps(fscal,dz20);
1699
1700             /* Update vectorial force */
1701             fix2             = _mm_add_ps(fix2,tx);
1702             fiy2             = _mm_add_ps(fiy2,ty);
1703             fiz2             = _mm_add_ps(fiz2,tz);
1704
1705             fjx0             = _mm_add_ps(fjx0,tx);
1706             fjy0             = _mm_add_ps(fjy0,ty);
1707             fjz0             = _mm_add_ps(fjz0,tz);
1708
1709             /**************************
1710              * CALCULATE INTERACTIONS *
1711              **************************/
1712
1713             /* REACTION-FIELD ELECTROSTATICS */
1714             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1715
1716             fscal            = felec;
1717
1718             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1719
1720             /* Calculate temporary vectorial force */
1721             tx               = _mm_mul_ps(fscal,dx21);
1722             ty               = _mm_mul_ps(fscal,dy21);
1723             tz               = _mm_mul_ps(fscal,dz21);
1724
1725             /* Update vectorial force */
1726             fix2             = _mm_add_ps(fix2,tx);
1727             fiy2             = _mm_add_ps(fiy2,ty);
1728             fiz2             = _mm_add_ps(fiz2,tz);
1729
1730             fjx1             = _mm_add_ps(fjx1,tx);
1731             fjy1             = _mm_add_ps(fjy1,ty);
1732             fjz1             = _mm_add_ps(fjz1,tz);
1733
1734             /**************************
1735              * CALCULATE INTERACTIONS *
1736              **************************/
1737
1738             /* REACTION-FIELD ELECTROSTATICS */
1739             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1740
1741             fscal            = felec;
1742
1743             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1744
1745             /* Calculate temporary vectorial force */
1746             tx               = _mm_mul_ps(fscal,dx22);
1747             ty               = _mm_mul_ps(fscal,dy22);
1748             tz               = _mm_mul_ps(fscal,dz22);
1749
1750             /* Update vectorial force */
1751             fix2             = _mm_add_ps(fix2,tx);
1752             fiy2             = _mm_add_ps(fiy2,ty);
1753             fiz2             = _mm_add_ps(fiz2,tz);
1754
1755             fjx2             = _mm_add_ps(fjx2,tx);
1756             fjy2             = _mm_add_ps(fjy2,ty);
1757             fjz2             = _mm_add_ps(fjz2,tz);
1758
1759             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1760             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1761             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1762             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1763
1764             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1765                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1766
1767             /* Inner loop uses 250 flops */
1768         }
1769
1770         /* End of innermost loop */
1771
1772         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1773                                               f+i_coord_offset,fshift+i_shift_offset);
1774
1775         /* Increment number of inner iterations */
1776         inneriter                  += j_index_end - j_index_start;
1777
1778         /* Outer loop uses 18 flops */
1779     }
1780
1781     /* Increment number of outer iterations */
1782     outeriter        += nri;
1783
1784     /* Update outer/inner flops */
1785
1786     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1787 }