src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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 sse2_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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_single.h"
  50 #include "kernelutil_x86_sse2_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_VF_sse2_single
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_VF_sse2_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 113     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 114     __m128           dummy_mask,cutoff_mask;
 115     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 116     __m128           one     = _mm_set1_ps(1.0);
 117     __m128           two     = _mm_set1_ps(2.0);
 118     x                = xx[0];
 119     f                = ff[0];
 120
 121     nri              = nlist->nri;
 122     iinr             = nlist->iinr;
 123     jindex           = nlist->jindex;
 124     jjnr             = nlist->jjnr;
 125     shiftidx         = nlist->shift;
 126     gid              = nlist->gid;
 127     shiftvec         = fr->shift_vec[0];
 128     fshift           = fr->fshift[0];
 129     facel            = _mm_set1_ps(fr->epsfac);
 130     charge           = mdatoms->chargeA;
 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           = gmx_mm_invsqrt_ps(rsq00);
 265             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 266             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 267             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 268             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 269             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 270             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 271             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 272             rinv22           = gmx_mm_invsqrt_ps(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             /* COULOMB ELECTROSTATICS */
 299             velec            = _mm_mul_ps(qq00,rinv00);
 300             felec            = _mm_mul_ps(velec,rinvsq00);
 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             /* COULOMB ELECTROSTATICS */
 335             velec            = _mm_mul_ps(qq01,rinv01);
 336             felec            = _mm_mul_ps(velec,rinvsq01);
 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             /* COULOMB ELECTROSTATICS */
 362             velec            = _mm_mul_ps(qq02,rinv02);
 363             felec            = _mm_mul_ps(velec,rinvsq02);
 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             /* COULOMB ELECTROSTATICS */
 389             velec            = _mm_mul_ps(qq10,rinv10);
 390             felec            = _mm_mul_ps(velec,rinvsq10);
 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             /* COULOMB ELECTROSTATICS */
 416             velec            = _mm_mul_ps(qq11,rinv11);
 417             felec            = _mm_mul_ps(velec,rinvsq11);
 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             /* COULOMB ELECTROSTATICS */
 443             velec            = _mm_mul_ps(qq12,rinv12);
 444             felec            = _mm_mul_ps(velec,rinvsq12);
 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             /* COULOMB ELECTROSTATICS */
 470             velec            = _mm_mul_ps(qq20,rinv20);
 471             felec            = _mm_mul_ps(velec,rinvsq20);
 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             /* COULOMB ELECTROSTATICS */
 497             velec            = _mm_mul_ps(qq21,rinv21);
 498             felec            = _mm_mul_ps(velec,rinvsq21);
 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             /* COULOMB ELECTROSTATICS */
 524             velec            = _mm_mul_ps(qq22,rinv22);
 525             felec            = _mm_mul_ps(velec,rinvsq22);
 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 264 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           = gmx_mm_invsqrt_ps(rsq00);
 625             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 626             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 627             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 628             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 629             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 630             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 631             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 632             rinv22           = gmx_mm_invsqrt_ps(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             /* COULOMB ELECTROSTATICS */
 659             velec            = _mm_mul_ps(qq00,rinv00);
 660             felec            = _mm_mul_ps(velec,rinvsq00);
 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             /* COULOMB ELECTROSTATICS */
 699             velec            = _mm_mul_ps(qq01,rinv01);
 700             felec            = _mm_mul_ps(velec,rinvsq01);
 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             /* COULOMB ELECTROSTATICS */
 729             velec            = _mm_mul_ps(qq02,rinv02);
 730             felec            = _mm_mul_ps(velec,rinvsq02);
 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             /* COULOMB ELECTROSTATICS */
 759             velec            = _mm_mul_ps(qq10,rinv10);
 760             felec            = _mm_mul_ps(velec,rinvsq10);
 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             /* COULOMB ELECTROSTATICS */
 789             velec            = _mm_mul_ps(qq11,rinv11);
 790             felec            = _mm_mul_ps(velec,rinvsq11);
 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             /* COULOMB ELECTROSTATICS */
 819             velec            = _mm_mul_ps(qq12,rinv12);
 820             felec            = _mm_mul_ps(velec,rinvsq12);
 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             /* COULOMB ELECTROSTATICS */
 849             velec            = _mm_mul_ps(qq20,rinv20);
 850             felec            = _mm_mul_ps(velec,rinvsq20);
 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             /* COULOMB ELECTROSTATICS */
 879             velec            = _mm_mul_ps(qq21,rinv21);
 880             felec            = _mm_mul_ps(velec,rinvsq21);
 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             /* COULOMB ELECTROSTATICS */
 909             velec            = _mm_mul_ps(qq22,rinv22);
 910             felec            = _mm_mul_ps(velec,rinvsq22);
 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 264 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*264);
 967 }
 968 /*
 969  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
 970  * Electrostatics interaction: Coulomb
 971  * VdW interaction:            LennardJones
 972  * Geometry:                   Water3-Water3
 973  * Calculate force/pot:        Force
 974  */
 975 void
 976 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
 977                     (t_nblist                    * gmx_restrict       nlist,
 978                      rvec                        * gmx_restrict          xx,
 979                      rvec                        * gmx_restrict          ff,
 980                      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->epsfac);
1046     charge           = mdatoms->chargeA;
1047     nvdwtype         = fr->ntype;
1048     vdwparam         = fr->nbfp;
1049     vdwtype          = mdatoms->typeA;
1050
1051     /* Setup water-specific parameters */
1052     inr              = nlist->iinr[0];
1053     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1054     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1055     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1056     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1057
1058     jq0              = _mm_set1_ps(charge[inr+0]);
1059     jq1              = _mm_set1_ps(charge[inr+1]);
1060     jq2              = _mm_set1_ps(charge[inr+2]);
1061     vdwjidx0A        = 2*vdwtype[inr+0];
1062     qq00             = _mm_mul_ps(iq0,jq0);
1063     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1064     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1065     qq01             = _mm_mul_ps(iq0,jq1);
1066     qq02             = _mm_mul_ps(iq0,jq2);
1067     qq10             = _mm_mul_ps(iq1,jq0);
1068     qq11             = _mm_mul_ps(iq1,jq1);
1069     qq12             = _mm_mul_ps(iq1,jq2);
1070     qq20             = _mm_mul_ps(iq2,jq0);
1071     qq21             = _mm_mul_ps(iq2,jq1);
1072     qq22             = _mm_mul_ps(iq2,jq2);
1073
1074     /* Avoid stupid compiler warnings */
1075     jnrA = jnrB = jnrC = jnrD = 0;
1076     j_coord_offsetA = 0;
1077     j_coord_offsetB = 0;
1078     j_coord_offsetC = 0;
1079     j_coord_offsetD = 0;
1080
1081     outeriter        = 0;
1082     inneriter        = 0;
1083
1084     for(iidx=0;iidx<4*DIM;iidx++)
1085     {
1086         scratch[iidx] = 0.0;
1087     }
1088
1089     /* Start outer loop over neighborlists */
1090     for(iidx=0; iidx<nri; iidx++)
1091     {
1092         /* Load shift vector for this list */
1093         i_shift_offset   = DIM*shiftidx[iidx];
1094
1095         /* Load limits for loop over neighbors */
1096         j_index_start    = jindex[iidx];
1097         j_index_end      = jindex[iidx+1];
1098
1099         /* Get outer coordinate index */
1100         inr              = iinr[iidx];
1101         i_coord_offset   = DIM*inr;
1102
1103         /* Load i particle coords and add shift vector */
1104         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1105                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1106
1107         fix0             = _mm_setzero_ps();
1108         fiy0             = _mm_setzero_ps();
1109         fiz0             = _mm_setzero_ps();
1110         fix1             = _mm_setzero_ps();
1111         fiy1             = _mm_setzero_ps();
1112         fiz1             = _mm_setzero_ps();
1113         fix2             = _mm_setzero_ps();
1114         fiy2             = _mm_setzero_ps();
1115         fiz2             = _mm_setzero_ps();
1116
1117         /* Start inner kernel loop */
1118         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1119         {
1120
1121             /* Get j neighbor index, and coordinate index */
1122             jnrA             = jjnr[jidx];
1123             jnrB             = jjnr[jidx+1];
1124             jnrC             = jjnr[jidx+2];
1125             jnrD             = jjnr[jidx+3];
1126             j_coord_offsetA  = DIM*jnrA;
1127             j_coord_offsetB  = DIM*jnrB;
1128             j_coord_offsetC  = DIM*jnrC;
1129             j_coord_offsetD  = DIM*jnrD;
1130
1131             /* load j atom coordinates */
1132             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1133                                               x+j_coord_offsetC,x+j_coord_offsetD,
1134                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1135
1136             /* Calculate displacement vector */
1137             dx00             = _mm_sub_ps(ix0,jx0);
1138             dy00             = _mm_sub_ps(iy0,jy0);
1139             dz00             = _mm_sub_ps(iz0,jz0);
1140             dx01             = _mm_sub_ps(ix0,jx1);
1141             dy01             = _mm_sub_ps(iy0,jy1);
1142             dz01             = _mm_sub_ps(iz0,jz1);
1143             dx02             = _mm_sub_ps(ix0,jx2);
1144             dy02             = _mm_sub_ps(iy0,jy2);
1145             dz02             = _mm_sub_ps(iz0,jz2);
1146             dx10             = _mm_sub_ps(ix1,jx0);
1147             dy10             = _mm_sub_ps(iy1,jy0);
1148             dz10             = _mm_sub_ps(iz1,jz0);
1149             dx11             = _mm_sub_ps(ix1,jx1);
1150             dy11             = _mm_sub_ps(iy1,jy1);
1151             dz11             = _mm_sub_ps(iz1,jz1);
1152             dx12             = _mm_sub_ps(ix1,jx2);
1153             dy12             = _mm_sub_ps(iy1,jy2);
1154             dz12             = _mm_sub_ps(iz1,jz2);
1155             dx20             = _mm_sub_ps(ix2,jx0);
1156             dy20             = _mm_sub_ps(iy2,jy0);
1157             dz20             = _mm_sub_ps(iz2,jz0);
1158             dx21             = _mm_sub_ps(ix2,jx1);
1159             dy21             = _mm_sub_ps(iy2,jy1);
1160             dz21             = _mm_sub_ps(iz2,jz1);
1161             dx22             = _mm_sub_ps(ix2,jx2);
1162             dy22             = _mm_sub_ps(iy2,jy2);
1163             dz22             = _mm_sub_ps(iz2,jz2);
1164
1165             /* Calculate squared distance and things based on it */
1166             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1167             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1168             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1169             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1170             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1171             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1172             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1173             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1174             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1175
1176             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1177             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1178             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1179             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1180             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1181             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1182             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1183             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1184             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1185
1186             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1187             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1188             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1189             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1190             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1191             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1192             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1193             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1194             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1195
1196             fjx0             = _mm_setzero_ps();
1197             fjy0             = _mm_setzero_ps();
1198             fjz0             = _mm_setzero_ps();
1199             fjx1             = _mm_setzero_ps();
1200             fjy1             = _mm_setzero_ps();
1201             fjz1             = _mm_setzero_ps();
1202             fjx2             = _mm_setzero_ps();
1203             fjy2             = _mm_setzero_ps();
1204             fjz2             = _mm_setzero_ps();
1205
1206             /**************************
1207              * CALCULATE INTERACTIONS *
1208              **************************/
1209
1210             /* COULOMB ELECTROSTATICS */
1211             velec            = _mm_mul_ps(qq00,rinv00);
1212             felec            = _mm_mul_ps(velec,rinvsq00);
1213
1214             /* LENNARD-JONES DISPERSION/REPULSION */
1215
1216             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1217             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1218
1219             fscal            = _mm_add_ps(felec,fvdw);
1220
1221             /* Calculate temporary vectorial force */
1222             tx               = _mm_mul_ps(fscal,dx00);
1223             ty               = _mm_mul_ps(fscal,dy00);
1224             tz               = _mm_mul_ps(fscal,dz00);
1225
1226             /* Update vectorial force */
1227             fix0             = _mm_add_ps(fix0,tx);
1228             fiy0             = _mm_add_ps(fiy0,ty);
1229             fiz0             = _mm_add_ps(fiz0,tz);
1230
1231             fjx0             = _mm_add_ps(fjx0,tx);
1232             fjy0             = _mm_add_ps(fjy0,ty);
1233             fjz0             = _mm_add_ps(fjz0,tz);
1234
1235             /**************************
1236              * CALCULATE INTERACTIONS *
1237              **************************/
1238
1239             /* COULOMB ELECTROSTATICS */
1240             velec            = _mm_mul_ps(qq01,rinv01);
1241             felec            = _mm_mul_ps(velec,rinvsq01);
1242
1243             fscal            = felec;
1244
1245             /* Calculate temporary vectorial force */
1246             tx               = _mm_mul_ps(fscal,dx01);
1247             ty               = _mm_mul_ps(fscal,dy01);
1248             tz               = _mm_mul_ps(fscal,dz01);
1249
1250             /* Update vectorial force */
1251             fix0             = _mm_add_ps(fix0,tx);
1252             fiy0             = _mm_add_ps(fiy0,ty);
1253             fiz0             = _mm_add_ps(fiz0,tz);
1254
1255             fjx1             = _mm_add_ps(fjx1,tx);
1256             fjy1             = _mm_add_ps(fjy1,ty);
1257             fjz1             = _mm_add_ps(fjz1,tz);
1258
1259             /**************************
1260              * CALCULATE INTERACTIONS *
1261              **************************/
1262
1263             /* COULOMB ELECTROSTATICS */
1264             velec            = _mm_mul_ps(qq02,rinv02);
1265             felec            = _mm_mul_ps(velec,rinvsq02);
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             /* COULOMB ELECTROSTATICS */
1288             velec            = _mm_mul_ps(qq10,rinv10);
1289             felec            = _mm_mul_ps(velec,rinvsq10);
1290
1291             fscal            = felec;
1292
1293             /* Calculate temporary vectorial force */
1294             tx               = _mm_mul_ps(fscal,dx10);
1295             ty               = _mm_mul_ps(fscal,dy10);
1296             tz               = _mm_mul_ps(fscal,dz10);
1297
1298             /* Update vectorial force */
1299             fix1             = _mm_add_ps(fix1,tx);
1300             fiy1             = _mm_add_ps(fiy1,ty);
1301             fiz1             = _mm_add_ps(fiz1,tz);
1302
1303             fjx0             = _mm_add_ps(fjx0,tx);
1304             fjy0             = _mm_add_ps(fjy0,ty);
1305             fjz0             = _mm_add_ps(fjz0,tz);
1306
1307             /**************************
1308              * CALCULATE INTERACTIONS *
1309              **************************/
1310
1311             /* COULOMB ELECTROSTATICS */
1312             velec            = _mm_mul_ps(qq11,rinv11);
1313             felec            = _mm_mul_ps(velec,rinvsq11);
1314
1315             fscal            = felec;
1316
1317             /* Calculate temporary vectorial force */
1318             tx               = _mm_mul_ps(fscal,dx11);
1319             ty               = _mm_mul_ps(fscal,dy11);
1320             tz               = _mm_mul_ps(fscal,dz11);
1321
1322             /* Update vectorial force */
1323             fix1             = _mm_add_ps(fix1,tx);
1324             fiy1             = _mm_add_ps(fiy1,ty);
1325             fiz1             = _mm_add_ps(fiz1,tz);
1326
1327             fjx1             = _mm_add_ps(fjx1,tx);
1328             fjy1             = _mm_add_ps(fjy1,ty);
1329             fjz1             = _mm_add_ps(fjz1,tz);
1330
1331             /**************************
1332              * CALCULATE INTERACTIONS *
1333              **************************/
1334
1335             /* COULOMB ELECTROSTATICS */
1336             velec            = _mm_mul_ps(qq12,rinv12);
1337             felec            = _mm_mul_ps(velec,rinvsq12);
1338
1339             fscal            = felec;
1340
1341             /* Calculate temporary vectorial force */
1342             tx               = _mm_mul_ps(fscal,dx12);
1343             ty               = _mm_mul_ps(fscal,dy12);
1344             tz               = _mm_mul_ps(fscal,dz12);
1345
1346             /* Update vectorial force */
1347             fix1             = _mm_add_ps(fix1,tx);
1348             fiy1             = _mm_add_ps(fiy1,ty);
1349             fiz1             = _mm_add_ps(fiz1,tz);
1350
1351             fjx2             = _mm_add_ps(fjx2,tx);
1352             fjy2             = _mm_add_ps(fjy2,ty);
1353             fjz2             = _mm_add_ps(fjz2,tz);
1354
1355             /**************************
1356              * CALCULATE INTERACTIONS *
1357              **************************/
1358
1359             /* COULOMB ELECTROSTATICS */
1360             velec            = _mm_mul_ps(qq20,rinv20);
1361             felec            = _mm_mul_ps(velec,rinvsq20);
1362
1363             fscal            = felec;
1364
1365             /* Calculate temporary vectorial force */
1366             tx               = _mm_mul_ps(fscal,dx20);
1367             ty               = _mm_mul_ps(fscal,dy20);
1368             tz               = _mm_mul_ps(fscal,dz20);
1369
1370             /* Update vectorial force */
1371             fix2             = _mm_add_ps(fix2,tx);
1372             fiy2             = _mm_add_ps(fiy2,ty);
1373             fiz2             = _mm_add_ps(fiz2,tz);
1374
1375             fjx0             = _mm_add_ps(fjx0,tx);
1376             fjy0             = _mm_add_ps(fjy0,ty);
1377             fjz0             = _mm_add_ps(fjz0,tz);
1378
1379             /**************************
1380              * CALCULATE INTERACTIONS *
1381              **************************/
1382
1383             /* COULOMB ELECTROSTATICS */
1384             velec            = _mm_mul_ps(qq21,rinv21);
1385             felec            = _mm_mul_ps(velec,rinvsq21);
1386
1387             fscal            = felec;
1388
1389             /* Calculate temporary vectorial force */
1390             tx               = _mm_mul_ps(fscal,dx21);
1391             ty               = _mm_mul_ps(fscal,dy21);
1392             tz               = _mm_mul_ps(fscal,dz21);
1393
1394             /* Update vectorial force */
1395             fix2             = _mm_add_ps(fix2,tx);
1396             fiy2             = _mm_add_ps(fiy2,ty);
1397             fiz2             = _mm_add_ps(fiz2,tz);
1398
1399             fjx1             = _mm_add_ps(fjx1,tx);
1400             fjy1             = _mm_add_ps(fjy1,ty);
1401             fjz1             = _mm_add_ps(fjz1,tz);
1402
1403             /**************************
1404              * CALCULATE INTERACTIONS *
1405              **************************/
1406
1407             /* COULOMB ELECTROSTATICS */
1408             velec            = _mm_mul_ps(qq22,rinv22);
1409             felec            = _mm_mul_ps(velec,rinvsq22);
1410
1411             fscal            = felec;
1412
1413             /* Calculate temporary vectorial force */
1414             tx               = _mm_mul_ps(fscal,dx22);
1415             ty               = _mm_mul_ps(fscal,dy22);
1416             tz               = _mm_mul_ps(fscal,dz22);
1417
1418             /* Update vectorial force */
1419             fix2             = _mm_add_ps(fix2,tx);
1420             fiy2             = _mm_add_ps(fiy2,ty);
1421             fiz2             = _mm_add_ps(fiz2,tz);
1422
1423             fjx2             = _mm_add_ps(fjx2,tx);
1424             fjy2             = _mm_add_ps(fjy2,ty);
1425             fjz2             = _mm_add_ps(fjz2,tz);
1426
1427             fjptrA             = f+j_coord_offsetA;
1428             fjptrB             = f+j_coord_offsetB;
1429             fjptrC             = f+j_coord_offsetC;
1430             fjptrD             = f+j_coord_offsetD;
1431
1432             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1433                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1434
1435             /* Inner loop uses 250 flops */
1436         }
1437
1438         if(jidx<j_index_end)
1439         {
1440
1441             /* Get j neighbor index, and coordinate index */
1442             jnrlistA         = jjnr[jidx];
1443             jnrlistB         = jjnr[jidx+1];
1444             jnrlistC         = jjnr[jidx+2];
1445             jnrlistD         = jjnr[jidx+3];
1446             /* Sign of each element will be negative for non-real atoms.
1447              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1448              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1449              */
1450             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1451             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1452             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1453             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1454             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1455             j_coord_offsetA  = DIM*jnrA;
1456             j_coord_offsetB  = DIM*jnrB;
1457             j_coord_offsetC  = DIM*jnrC;
1458             j_coord_offsetD  = DIM*jnrD;
1459
1460             /* load j atom coordinates */
1461             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1462                                               x+j_coord_offsetC,x+j_coord_offsetD,
1463                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1464
1465             /* Calculate displacement vector */
1466             dx00             = _mm_sub_ps(ix0,jx0);
1467             dy00             = _mm_sub_ps(iy0,jy0);
1468             dz00             = _mm_sub_ps(iz0,jz0);
1469             dx01             = _mm_sub_ps(ix0,jx1);
1470             dy01             = _mm_sub_ps(iy0,jy1);
1471             dz01             = _mm_sub_ps(iz0,jz1);
1472             dx02             = _mm_sub_ps(ix0,jx2);
1473             dy02             = _mm_sub_ps(iy0,jy2);
1474             dz02             = _mm_sub_ps(iz0,jz2);
1475             dx10             = _mm_sub_ps(ix1,jx0);
1476             dy10             = _mm_sub_ps(iy1,jy0);
1477             dz10             = _mm_sub_ps(iz1,jz0);
1478             dx11             = _mm_sub_ps(ix1,jx1);
1479             dy11             = _mm_sub_ps(iy1,jy1);
1480             dz11             = _mm_sub_ps(iz1,jz1);
1481             dx12             = _mm_sub_ps(ix1,jx2);
1482             dy12             = _mm_sub_ps(iy1,jy2);
1483             dz12             = _mm_sub_ps(iz1,jz2);
1484             dx20             = _mm_sub_ps(ix2,jx0);
1485             dy20             = _mm_sub_ps(iy2,jy0);
1486             dz20             = _mm_sub_ps(iz2,jz0);
1487             dx21             = _mm_sub_ps(ix2,jx1);
1488             dy21             = _mm_sub_ps(iy2,jy1);
1489             dz21             = _mm_sub_ps(iz2,jz1);
1490             dx22             = _mm_sub_ps(ix2,jx2);
1491             dy22             = _mm_sub_ps(iy2,jy2);
1492             dz22             = _mm_sub_ps(iz2,jz2);
1493
1494             /* Calculate squared distance and things based on it */
1495             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1496             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1497             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1498             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1499             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1500             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1501             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1502             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1503             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1504
1505             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1506             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1507             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1508             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1509             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1510             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1511             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1512             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1513             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1514
1515             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1516             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1517             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1518             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1519             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1520             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1521             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1522             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1523             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1524
1525             fjx0             = _mm_setzero_ps();
1526             fjy0             = _mm_setzero_ps();
1527             fjz0             = _mm_setzero_ps();
1528             fjx1             = _mm_setzero_ps();
1529             fjy1             = _mm_setzero_ps();
1530             fjz1             = _mm_setzero_ps();
1531             fjx2             = _mm_setzero_ps();
1532             fjy2             = _mm_setzero_ps();
1533             fjz2             = _mm_setzero_ps();
1534
1535             /**************************
1536              * CALCULATE INTERACTIONS *
1537              **************************/
1538
1539             /* COULOMB ELECTROSTATICS */
1540             velec            = _mm_mul_ps(qq00,rinv00);
1541             felec            = _mm_mul_ps(velec,rinvsq00);
1542
1543             /* LENNARD-JONES DISPERSION/REPULSION */
1544
1545             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1546             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1547
1548             fscal            = _mm_add_ps(felec,fvdw);
1549
1550             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1551
1552             /* Calculate temporary vectorial force */
1553             tx               = _mm_mul_ps(fscal,dx00);
1554             ty               = _mm_mul_ps(fscal,dy00);
1555             tz               = _mm_mul_ps(fscal,dz00);
1556
1557             /* Update vectorial force */
1558             fix0             = _mm_add_ps(fix0,tx);
1559             fiy0             = _mm_add_ps(fiy0,ty);
1560             fiz0             = _mm_add_ps(fiz0,tz);
1561
1562             fjx0             = _mm_add_ps(fjx0,tx);
1563             fjy0             = _mm_add_ps(fjy0,ty);
1564             fjz0             = _mm_add_ps(fjz0,tz);
1565
1566             /**************************
1567              * CALCULATE INTERACTIONS *
1568              **************************/
1569
1570             /* COULOMB ELECTROSTATICS */
1571             velec            = _mm_mul_ps(qq01,rinv01);
1572             felec            = _mm_mul_ps(velec,rinvsq01);
1573
1574             fscal            = felec;
1575
1576             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1577
1578             /* Calculate temporary vectorial force */
1579             tx               = _mm_mul_ps(fscal,dx01);
1580             ty               = _mm_mul_ps(fscal,dy01);
1581             tz               = _mm_mul_ps(fscal,dz01);
1582
1583             /* Update vectorial force */
1584             fix0             = _mm_add_ps(fix0,tx);
1585             fiy0             = _mm_add_ps(fiy0,ty);
1586             fiz0             = _mm_add_ps(fiz0,tz);
1587
1588             fjx1             = _mm_add_ps(fjx1,tx);
1589             fjy1             = _mm_add_ps(fjy1,ty);
1590             fjz1             = _mm_add_ps(fjz1,tz);
1591
1592             /**************************
1593              * CALCULATE INTERACTIONS *
1594              **************************/
1595
1596             /* COULOMB ELECTROSTATICS */
1597             velec            = _mm_mul_ps(qq02,rinv02);
1598             felec            = _mm_mul_ps(velec,rinvsq02);
1599
1600             fscal            = felec;
1601
1602             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1603
1604             /* Calculate temporary vectorial force */
1605             tx               = _mm_mul_ps(fscal,dx02);
1606             ty               = _mm_mul_ps(fscal,dy02);
1607             tz               = _mm_mul_ps(fscal,dz02);
1608
1609             /* Update vectorial force */
1610             fix0             = _mm_add_ps(fix0,tx);
1611             fiy0             = _mm_add_ps(fiy0,ty);
1612             fiz0             = _mm_add_ps(fiz0,tz);
1613
1614             fjx2             = _mm_add_ps(fjx2,tx);
1615             fjy2             = _mm_add_ps(fjy2,ty);
1616             fjz2             = _mm_add_ps(fjz2,tz);
1617
1618             /**************************
1619              * CALCULATE INTERACTIONS *
1620              **************************/
1621
1622             /* COULOMB ELECTROSTATICS */
1623             velec            = _mm_mul_ps(qq10,rinv10);
1624             felec            = _mm_mul_ps(velec,rinvsq10);
1625
1626             fscal            = felec;
1627
1628             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1629
1630             /* Calculate temporary vectorial force */
1631             tx               = _mm_mul_ps(fscal,dx10);
1632             ty               = _mm_mul_ps(fscal,dy10);
1633             tz               = _mm_mul_ps(fscal,dz10);
1634
1635             /* Update vectorial force */
1636             fix1             = _mm_add_ps(fix1,tx);
1637             fiy1             = _mm_add_ps(fiy1,ty);
1638             fiz1             = _mm_add_ps(fiz1,tz);
1639
1640             fjx0             = _mm_add_ps(fjx0,tx);
1641             fjy0             = _mm_add_ps(fjy0,ty);
1642             fjz0             = _mm_add_ps(fjz0,tz);
1643
1644             /**************************
1645              * CALCULATE INTERACTIONS *
1646              **************************/
1647
1648             /* COULOMB ELECTROSTATICS */
1649             velec            = _mm_mul_ps(qq11,rinv11);
1650             felec            = _mm_mul_ps(velec,rinvsq11);
1651
1652             fscal            = felec;
1653
1654             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1655
1656             /* Calculate temporary vectorial force */
1657             tx               = _mm_mul_ps(fscal,dx11);
1658             ty               = _mm_mul_ps(fscal,dy11);
1659             tz               = _mm_mul_ps(fscal,dz11);
1660
1661             /* Update vectorial force */
1662             fix1             = _mm_add_ps(fix1,tx);
1663             fiy1             = _mm_add_ps(fiy1,ty);
1664             fiz1             = _mm_add_ps(fiz1,tz);
1665
1666             fjx1             = _mm_add_ps(fjx1,tx);
1667             fjy1             = _mm_add_ps(fjy1,ty);
1668             fjz1             = _mm_add_ps(fjz1,tz);
1669
1670             /**************************
1671              * CALCULATE INTERACTIONS *
1672              **************************/
1673
1674             /* COULOMB ELECTROSTATICS */
1675             velec            = _mm_mul_ps(qq12,rinv12);
1676             felec            = _mm_mul_ps(velec,rinvsq12);
1677
1678             fscal            = felec;
1679
1680             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1681
1682             /* Calculate temporary vectorial force */
1683             tx               = _mm_mul_ps(fscal,dx12);
1684             ty               = _mm_mul_ps(fscal,dy12);
1685             tz               = _mm_mul_ps(fscal,dz12);
1686
1687             /* Update vectorial force */
1688             fix1             = _mm_add_ps(fix1,tx);
1689             fiy1             = _mm_add_ps(fiy1,ty);
1690             fiz1             = _mm_add_ps(fiz1,tz);
1691
1692             fjx2             = _mm_add_ps(fjx2,tx);
1693             fjy2             = _mm_add_ps(fjy2,ty);
1694             fjz2             = _mm_add_ps(fjz2,tz);
1695
1696             /**************************
1697              * CALCULATE INTERACTIONS *
1698              **************************/
1699
1700             /* COULOMB ELECTROSTATICS */
1701             velec            = _mm_mul_ps(qq20,rinv20);
1702             felec            = _mm_mul_ps(velec,rinvsq20);
1703
1704             fscal            = felec;
1705
1706             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1707
1708             /* Calculate temporary vectorial force */
1709             tx               = _mm_mul_ps(fscal,dx20);
1710             ty               = _mm_mul_ps(fscal,dy20);
1711             tz               = _mm_mul_ps(fscal,dz20);
1712
1713             /* Update vectorial force */
1714             fix2             = _mm_add_ps(fix2,tx);
1715             fiy2             = _mm_add_ps(fiy2,ty);
1716             fiz2             = _mm_add_ps(fiz2,tz);
1717
1718             fjx0             = _mm_add_ps(fjx0,tx);
1719             fjy0             = _mm_add_ps(fjy0,ty);
1720             fjz0             = _mm_add_ps(fjz0,tz);
1721
1722             /**************************
1723              * CALCULATE INTERACTIONS *
1724              **************************/
1725
1726             /* COULOMB ELECTROSTATICS */
1727             velec            = _mm_mul_ps(qq21,rinv21);
1728             felec            = _mm_mul_ps(velec,rinvsq21);
1729
1730             fscal            = felec;
1731
1732             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1733
1734             /* Calculate temporary vectorial force */
1735             tx               = _mm_mul_ps(fscal,dx21);
1736             ty               = _mm_mul_ps(fscal,dy21);
1737             tz               = _mm_mul_ps(fscal,dz21);
1738
1739             /* Update vectorial force */
1740             fix2             = _mm_add_ps(fix2,tx);
1741             fiy2             = _mm_add_ps(fiy2,ty);
1742             fiz2             = _mm_add_ps(fiz2,tz);
1743
1744             fjx1             = _mm_add_ps(fjx1,tx);
1745             fjy1             = _mm_add_ps(fjy1,ty);
1746             fjz1             = _mm_add_ps(fjz1,tz);
1747
1748             /**************************
1749              * CALCULATE INTERACTIONS *
1750              **************************/
1751
1752             /* COULOMB ELECTROSTATICS */
1753             velec            = _mm_mul_ps(qq22,rinv22);
1754             felec            = _mm_mul_ps(velec,rinvsq22);
1755
1756             fscal            = felec;
1757
1758             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1759
1760             /* Calculate temporary vectorial force */
1761             tx               = _mm_mul_ps(fscal,dx22);
1762             ty               = _mm_mul_ps(fscal,dy22);
1763             tz               = _mm_mul_ps(fscal,dz22);
1764
1765             /* Update vectorial force */
1766             fix2             = _mm_add_ps(fix2,tx);
1767             fiy2             = _mm_add_ps(fiy2,ty);
1768             fiz2             = _mm_add_ps(fiz2,tz);
1769
1770             fjx2             = _mm_add_ps(fjx2,tx);
1771             fjy2             = _mm_add_ps(fjy2,ty);
1772             fjz2             = _mm_add_ps(fjz2,tz);
1773
1774             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1775             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1776             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1777             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1778
1779             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1780                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1781
1782             /* Inner loop uses 250 flops */
1783         }
1784
1785         /* End of innermost loop */
1786
1787         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1788                                               f+i_coord_offset,fshift+i_shift_offset);
1789
1790         /* Increment number of inner iterations */
1791         inneriter                  += j_index_end - j_index_start;
1792
1793         /* Outer loop uses 18 flops */
1794     }
1795
1796     /* Increment number of outer iterations */
1797     outeriter        += nri;
1798
1799     /* Update outer/inner flops */
1800
1801     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1802 }