src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_avx_128_fma_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 avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_VF_avx_128_fma_single
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  97     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  98     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     int              nvdwtype;
 107     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 108     int              *vdwtype;
 109     real             *vdwparam;
 110     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 111     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 112     __m128i          ewitab;
 113     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 114     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 115     real             *ewtab;
 116     __m128           dummy_mask,cutoff_mask;
 117     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 118     __m128           one     = _mm_set1_ps(1.0);
 119     __m128           two     = _mm_set1_ps(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_ps(fr->epsfac);
 132     charge           = mdatoms->chargeA;
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 138     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 139     beta2            = _mm_mul_ps(beta,beta);
 140     beta3            = _mm_mul_ps(beta,beta2);
 141     ewtab            = fr->ic->tabq_coul_FDV0;
 142     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 143     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 144
 145     /* Setup water-specific parameters */
 146     inr              = nlist->iinr[0];
 147     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 148     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 149     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 150     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 151
 152     jq0              = _mm_set1_ps(charge[inr+0]);
 153     jq1              = _mm_set1_ps(charge[inr+1]);
 154     jq2              = _mm_set1_ps(charge[inr+2]);
 155     vdwjidx0A        = 2*vdwtype[inr+0];
 156     qq00             = _mm_mul_ps(iq0,jq0);
 157     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 158     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 159     qq01             = _mm_mul_ps(iq0,jq1);
 160     qq02             = _mm_mul_ps(iq0,jq2);
 161     qq10             = _mm_mul_ps(iq1,jq0);
 162     qq11             = _mm_mul_ps(iq1,jq1);
 163     qq12             = _mm_mul_ps(iq1,jq2);
 164     qq20             = _mm_mul_ps(iq2,jq0);
 165     qq21             = _mm_mul_ps(iq2,jq1);
 166     qq22             = _mm_mul_ps(iq2,jq2);
 167
 168     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 169     rcutoff_scalar   = fr->rcoulomb;
 170     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 171     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 172
 173     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 174     rvdw             = _mm_set1_ps(fr->rvdw);
 175
 176     /* Avoid stupid compiler warnings */
 177     jnrA = jnrB = jnrC = jnrD = 0;
 178     j_coord_offsetA = 0;
 179     j_coord_offsetB = 0;
 180     j_coord_offsetC = 0;
 181     j_coord_offsetD = 0;
 182
 183     outeriter        = 0;
 184     inneriter        = 0;
 185
 186     for(iidx=0;iidx<4*DIM;iidx++)
 187     {
 188         scratch[iidx] = 0.0;
 189     }
 190
 191     /* Start outer loop over neighborlists */
 192     for(iidx=0; iidx<nri; iidx++)
 193     {
 194         /* Load shift vector for this list */
 195         i_shift_offset   = DIM*shiftidx[iidx];
 196
 197         /* Load limits for loop over neighbors */
 198         j_index_start    = jindex[iidx];
 199         j_index_end      = jindex[iidx+1];
 200
 201         /* Get outer coordinate index */
 202         inr              = iinr[iidx];
 203         i_coord_offset   = DIM*inr;
 204
 205         /* Load i particle coords and add shift vector */
 206         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 207                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 208
 209         fix0             = _mm_setzero_ps();
 210         fiy0             = _mm_setzero_ps();
 211         fiz0             = _mm_setzero_ps();
 212         fix1             = _mm_setzero_ps();
 213         fiy1             = _mm_setzero_ps();
 214         fiz1             = _mm_setzero_ps();
 215         fix2             = _mm_setzero_ps();
 216         fiy2             = _mm_setzero_ps();
 217         fiz2             = _mm_setzero_ps();
 218
 219         /* Reset potential sums */
 220         velecsum         = _mm_setzero_ps();
 221         vvdwsum          = _mm_setzero_ps();
 222
 223         /* Start inner kernel loop */
 224         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 225         {
 226
 227             /* Get j neighbor index, and coordinate index */
 228             jnrA             = jjnr[jidx];
 229             jnrB             = jjnr[jidx+1];
 230             jnrC             = jjnr[jidx+2];
 231             jnrD             = jjnr[jidx+3];
 232             j_coord_offsetA  = DIM*jnrA;
 233             j_coord_offsetB  = DIM*jnrB;
 234             j_coord_offsetC  = DIM*jnrC;
 235             j_coord_offsetD  = DIM*jnrD;
 236
 237             /* load j atom coordinates */
 238             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 239                                               x+j_coord_offsetC,x+j_coord_offsetD,
 240                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 241
 242             /* Calculate displacement vector */
 243             dx00             = _mm_sub_ps(ix0,jx0);
 244             dy00             = _mm_sub_ps(iy0,jy0);
 245             dz00             = _mm_sub_ps(iz0,jz0);
 246             dx01             = _mm_sub_ps(ix0,jx1);
 247             dy01             = _mm_sub_ps(iy0,jy1);
 248             dz01             = _mm_sub_ps(iz0,jz1);
 249             dx02             = _mm_sub_ps(ix0,jx2);
 250             dy02             = _mm_sub_ps(iy0,jy2);
 251             dz02             = _mm_sub_ps(iz0,jz2);
 252             dx10             = _mm_sub_ps(ix1,jx0);
 253             dy10             = _mm_sub_ps(iy1,jy0);
 254             dz10             = _mm_sub_ps(iz1,jz0);
 255             dx11             = _mm_sub_ps(ix1,jx1);
 256             dy11             = _mm_sub_ps(iy1,jy1);
 257             dz11             = _mm_sub_ps(iz1,jz1);
 258             dx12             = _mm_sub_ps(ix1,jx2);
 259             dy12             = _mm_sub_ps(iy1,jy2);
 260             dz12             = _mm_sub_ps(iz1,jz2);
 261             dx20             = _mm_sub_ps(ix2,jx0);
 262             dy20             = _mm_sub_ps(iy2,jy0);
 263             dz20             = _mm_sub_ps(iz2,jz0);
 264             dx21             = _mm_sub_ps(ix2,jx1);
 265             dy21             = _mm_sub_ps(iy2,jy1);
 266             dz21             = _mm_sub_ps(iz2,jz1);
 267             dx22             = _mm_sub_ps(ix2,jx2);
 268             dy22             = _mm_sub_ps(iy2,jy2);
 269             dz22             = _mm_sub_ps(iz2,jz2);
 270
 271             /* Calculate squared distance and things based on it */
 272             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 273             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 274             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 275             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 276             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 277             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 278             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 279             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 280             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 281
 282             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 283             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 284             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 285             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 286             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 287             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 288             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 289             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 290             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 291
 292             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 293             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 294             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 295             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 296             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 297             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 298             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 299             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 300             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 301
 302             fjx0             = _mm_setzero_ps();
 303             fjy0             = _mm_setzero_ps();
 304             fjz0             = _mm_setzero_ps();
 305             fjx1             = _mm_setzero_ps();
 306             fjy1             = _mm_setzero_ps();
 307             fjz1             = _mm_setzero_ps();
 308             fjx2             = _mm_setzero_ps();
 309             fjy2             = _mm_setzero_ps();
 310             fjz2             = _mm_setzero_ps();
 311
 312             /**************************
 313              * CALCULATE INTERACTIONS *
 314              **************************/
 315
 316             if (gmx_mm_any_lt(rsq00,rcutoff2))
 317             {
 318
 319             r00              = _mm_mul_ps(rsq00,rinv00);
 320
 321             /* EWALD ELECTROSTATICS */
 322
 323             /* Analytical PME correction */
 324             zeta2            = _mm_mul_ps(beta2,rsq00);
 325             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 326             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 327             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 328             felec            = _mm_mul_ps(qq00,felec);
 329             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 330             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 331             velec            = _mm_mul_ps(qq00,velec);
 332
 333             /* LENNARD-JONES DISPERSION/REPULSION */
 334
 335             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 336             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 337             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 338             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 339                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 340             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 341
 342             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _mm_and_ps(velec,cutoff_mask);
 346             velecsum         = _mm_add_ps(velecsum,velec);
 347             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 348             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 349
 350             fscal            = _mm_add_ps(felec,fvdw);
 351
 352             fscal            = _mm_and_ps(fscal,cutoff_mask);
 353
 354              /* Update vectorial force */
 355             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 356             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 357             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 358
 359             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 360             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 361             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 362
 363             }
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             if (gmx_mm_any_lt(rsq01,rcutoff2))
 370             {
 371
 372             r01              = _mm_mul_ps(rsq01,rinv01);
 373
 374             /* EWALD ELECTROSTATICS */
 375
 376             /* Analytical PME correction */
 377             zeta2            = _mm_mul_ps(beta2,rsq01);
 378             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 379             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 380             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 381             felec            = _mm_mul_ps(qq01,felec);
 382             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 383             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 384             velec            = _mm_mul_ps(qq01,velec);
 385
 386             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 387
 388             /* Update potential sum for this i atom from the interaction with this j atom. */
 389             velec            = _mm_and_ps(velec,cutoff_mask);
 390             velecsum         = _mm_add_ps(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             fscal            = _mm_and_ps(fscal,cutoff_mask);
 395
 396              /* Update vectorial force */
 397             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 398             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 399             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 400
 401             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 402             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 403             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 404
 405             }
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             if (gmx_mm_any_lt(rsq02,rcutoff2))
 412             {
 413
 414             r02              = _mm_mul_ps(rsq02,rinv02);
 415
 416             /* EWALD ELECTROSTATICS */
 417
 418             /* Analytical PME correction */
 419             zeta2            = _mm_mul_ps(beta2,rsq02);
 420             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 421             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 422             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 423             felec            = _mm_mul_ps(qq02,felec);
 424             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 425             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 426             velec            = _mm_mul_ps(qq02,velec);
 427
 428             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velec            = _mm_and_ps(velec,cutoff_mask);
 432             velecsum         = _mm_add_ps(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436             fscal            = _mm_and_ps(fscal,cutoff_mask);
 437
 438              /* Update vectorial force */
 439             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 440             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 441             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 442
 443             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 444             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 445             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 446
 447             }
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             if (gmx_mm_any_lt(rsq10,rcutoff2))
 454             {
 455
 456             r10              = _mm_mul_ps(rsq10,rinv10);
 457
 458             /* EWALD ELECTROSTATICS */
 459
 460             /* Analytical PME correction */
 461             zeta2            = _mm_mul_ps(beta2,rsq10);
 462             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 463             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 464             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 465             felec            = _mm_mul_ps(qq10,felec);
 466             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 467             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 468             velec            = _mm_mul_ps(qq10,velec);
 469
 470             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velec            = _mm_and_ps(velec,cutoff_mask);
 474             velecsum         = _mm_add_ps(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478             fscal            = _mm_and_ps(fscal,cutoff_mask);
 479
 480              /* Update vectorial force */
 481             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 482             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 483             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 484
 485             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 486             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 487             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 488
 489             }
 490
 491             /**************************
 492              * CALCULATE INTERACTIONS *
 493              **************************/
 494
 495             if (gmx_mm_any_lt(rsq11,rcutoff2))
 496             {
 497
 498             r11              = _mm_mul_ps(rsq11,rinv11);
 499
 500             /* EWALD ELECTROSTATICS */
 501
 502             /* Analytical PME correction */
 503             zeta2            = _mm_mul_ps(beta2,rsq11);
 504             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 505             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 506             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 507             felec            = _mm_mul_ps(qq11,felec);
 508             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 509             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
 510             velec            = _mm_mul_ps(qq11,velec);
 511
 512             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 513
 514             /* Update potential sum for this i atom from the interaction with this j atom. */
 515             velec            = _mm_and_ps(velec,cutoff_mask);
 516             velecsum         = _mm_add_ps(velecsum,velec);
 517
 518             fscal            = felec;
 519
 520             fscal            = _mm_and_ps(fscal,cutoff_mask);
 521
 522              /* Update vectorial force */
 523             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 524             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 525             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 526
 527             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 528             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 529             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 530
 531             }
 532
 533             /**************************
 534              * CALCULATE INTERACTIONS *
 535              **************************/
 536
 537             if (gmx_mm_any_lt(rsq12,rcutoff2))
 538             {
 539
 540             r12              = _mm_mul_ps(rsq12,rinv12);
 541
 542             /* EWALD ELECTROSTATICS */
 543
 544             /* Analytical PME correction */
 545             zeta2            = _mm_mul_ps(beta2,rsq12);
 546             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 547             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 548             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 549             felec            = _mm_mul_ps(qq12,felec);
 550             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 551             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
 552             velec            = _mm_mul_ps(qq12,velec);
 553
 554             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 555
 556             /* Update potential sum for this i atom from the interaction with this j atom. */
 557             velec            = _mm_and_ps(velec,cutoff_mask);
 558             velecsum         = _mm_add_ps(velecsum,velec);
 559
 560             fscal            = felec;
 561
 562             fscal            = _mm_and_ps(fscal,cutoff_mask);
 563
 564              /* Update vectorial force */
 565             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 566             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 567             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 568
 569             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 570             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 571             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 572
 573             }
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             if (gmx_mm_any_lt(rsq20,rcutoff2))
 580             {
 581
 582             r20              = _mm_mul_ps(rsq20,rinv20);
 583
 584             /* EWALD ELECTROSTATICS */
 585
 586             /* Analytical PME correction */
 587             zeta2            = _mm_mul_ps(beta2,rsq20);
 588             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 589             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 590             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 591             felec            = _mm_mul_ps(qq20,felec);
 592             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 593             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 594             velec            = _mm_mul_ps(qq20,velec);
 595
 596             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 597
 598             /* Update potential sum for this i atom from the interaction with this j atom. */
 599             velec            = _mm_and_ps(velec,cutoff_mask);
 600             velecsum         = _mm_add_ps(velecsum,velec);
 601
 602             fscal            = felec;
 603
 604             fscal            = _mm_and_ps(fscal,cutoff_mask);
 605
 606              /* Update vectorial force */
 607             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 608             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 609             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 610
 611             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 612             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 613             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 614
 615             }
 616
 617             /**************************
 618              * CALCULATE INTERACTIONS *
 619              **************************/
 620
 621             if (gmx_mm_any_lt(rsq21,rcutoff2))
 622             {
 623
 624             r21              = _mm_mul_ps(rsq21,rinv21);
 625
 626             /* EWALD ELECTROSTATICS */
 627
 628             /* Analytical PME correction */
 629             zeta2            = _mm_mul_ps(beta2,rsq21);
 630             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 631             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 632             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 633             felec            = _mm_mul_ps(qq21,felec);
 634             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 635             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
 636             velec            = _mm_mul_ps(qq21,velec);
 637
 638             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 639
 640             /* Update potential sum for this i atom from the interaction with this j atom. */
 641             velec            = _mm_and_ps(velec,cutoff_mask);
 642             velecsum         = _mm_add_ps(velecsum,velec);
 643
 644             fscal            = felec;
 645
 646             fscal            = _mm_and_ps(fscal,cutoff_mask);
 647
 648              /* Update vectorial force */
 649             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 650             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 651             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 652
 653             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 654             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 655             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 656
 657             }
 658
 659             /**************************
 660              * CALCULATE INTERACTIONS *
 661              **************************/
 662
 663             if (gmx_mm_any_lt(rsq22,rcutoff2))
 664             {
 665
 666             r22              = _mm_mul_ps(rsq22,rinv22);
 667
 668             /* EWALD ELECTROSTATICS */
 669
 670             /* Analytical PME correction */
 671             zeta2            = _mm_mul_ps(beta2,rsq22);
 672             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 673             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 674             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 675             felec            = _mm_mul_ps(qq22,felec);
 676             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 677             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
 678             velec            = _mm_mul_ps(qq22,velec);
 679
 680             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 681
 682             /* Update potential sum for this i atom from the interaction with this j atom. */
 683             velec            = _mm_and_ps(velec,cutoff_mask);
 684             velecsum         = _mm_add_ps(velecsum,velec);
 685
 686             fscal            = felec;
 687
 688             fscal            = _mm_and_ps(fscal,cutoff_mask);
 689
 690              /* Update vectorial force */
 691             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 692             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 693             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 694
 695             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 696             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 697             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 698
 699             }
 700
 701             fjptrA             = f+j_coord_offsetA;
 702             fjptrB             = f+j_coord_offsetB;
 703             fjptrC             = f+j_coord_offsetC;
 704             fjptrD             = f+j_coord_offsetD;
 705
 706             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 707                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 708
 709             /* Inner loop uses 315 flops */
 710         }
 711
 712         if(jidx<j_index_end)
 713         {
 714
 715             /* Get j neighbor index, and coordinate index */
 716             jnrlistA         = jjnr[jidx];
 717             jnrlistB         = jjnr[jidx+1];
 718             jnrlistC         = jjnr[jidx+2];
 719             jnrlistD         = jjnr[jidx+3];
 720             /* Sign of each element will be negative for non-real atoms.
 721              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 722              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 723              */
 724             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 725             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 726             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 727             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 728             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 729             j_coord_offsetA  = DIM*jnrA;
 730             j_coord_offsetB  = DIM*jnrB;
 731             j_coord_offsetC  = DIM*jnrC;
 732             j_coord_offsetD  = DIM*jnrD;
 733
 734             /* load j atom coordinates */
 735             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 736                                               x+j_coord_offsetC,x+j_coord_offsetD,
 737                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 738
 739             /* Calculate displacement vector */
 740             dx00             = _mm_sub_ps(ix0,jx0);
 741             dy00             = _mm_sub_ps(iy0,jy0);
 742             dz00             = _mm_sub_ps(iz0,jz0);
 743             dx01             = _mm_sub_ps(ix0,jx1);
 744             dy01             = _mm_sub_ps(iy0,jy1);
 745             dz01             = _mm_sub_ps(iz0,jz1);
 746             dx02             = _mm_sub_ps(ix0,jx2);
 747             dy02             = _mm_sub_ps(iy0,jy2);
 748             dz02             = _mm_sub_ps(iz0,jz2);
 749             dx10             = _mm_sub_ps(ix1,jx0);
 750             dy10             = _mm_sub_ps(iy1,jy0);
 751             dz10             = _mm_sub_ps(iz1,jz0);
 752             dx11             = _mm_sub_ps(ix1,jx1);
 753             dy11             = _mm_sub_ps(iy1,jy1);
 754             dz11             = _mm_sub_ps(iz1,jz1);
 755             dx12             = _mm_sub_ps(ix1,jx2);
 756             dy12             = _mm_sub_ps(iy1,jy2);
 757             dz12             = _mm_sub_ps(iz1,jz2);
 758             dx20             = _mm_sub_ps(ix2,jx0);
 759             dy20             = _mm_sub_ps(iy2,jy0);
 760             dz20             = _mm_sub_ps(iz2,jz0);
 761             dx21             = _mm_sub_ps(ix2,jx1);
 762             dy21             = _mm_sub_ps(iy2,jy1);
 763             dz21             = _mm_sub_ps(iz2,jz1);
 764             dx22             = _mm_sub_ps(ix2,jx2);
 765             dy22             = _mm_sub_ps(iy2,jy2);
 766             dz22             = _mm_sub_ps(iz2,jz2);
 767
 768             /* Calculate squared distance and things based on it */
 769             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 770             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 771             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 772             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 773             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 774             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 775             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 776             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 777             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 778
 779             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 780             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 781             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 782             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 783             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 784             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 785             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 786             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 787             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 788
 789             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 790             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 791             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 792             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 793             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 794             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 795             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 796             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 797             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 798
 799             fjx0             = _mm_setzero_ps();
 800             fjy0             = _mm_setzero_ps();
 801             fjz0             = _mm_setzero_ps();
 802             fjx1             = _mm_setzero_ps();
 803             fjy1             = _mm_setzero_ps();
 804             fjz1             = _mm_setzero_ps();
 805             fjx2             = _mm_setzero_ps();
 806             fjy2             = _mm_setzero_ps();
 807             fjz2             = _mm_setzero_ps();
 808
 809             /**************************
 810              * CALCULATE INTERACTIONS *
 811              **************************/
 812
 813             if (gmx_mm_any_lt(rsq00,rcutoff2))
 814             {
 815
 816             r00              = _mm_mul_ps(rsq00,rinv00);
 817             r00              = _mm_andnot_ps(dummy_mask,r00);
 818
 819             /* EWALD ELECTROSTATICS */
 820
 821             /* Analytical PME correction */
 822             zeta2            = _mm_mul_ps(beta2,rsq00);
 823             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 824             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 825             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 826             felec            = _mm_mul_ps(qq00,felec);
 827             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 828             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 829             velec            = _mm_mul_ps(qq00,velec);
 830
 831             /* LENNARD-JONES DISPERSION/REPULSION */
 832
 833             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 834             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 835             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 836             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 837                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 838             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 839
 840             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 841
 842             /* Update potential sum for this i atom from the interaction with this j atom. */
 843             velec            = _mm_and_ps(velec,cutoff_mask);
 844             velec            = _mm_andnot_ps(dummy_mask,velec);
 845             velecsum         = _mm_add_ps(velecsum,velec);
 846             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 847             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 848             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 849
 850             fscal            = _mm_add_ps(felec,fvdw);
 851
 852             fscal            = _mm_and_ps(fscal,cutoff_mask);
 853
 854             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 855
 856              /* Update vectorial force */
 857             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 858             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 859             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 860
 861             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 862             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 863             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 864
 865             }
 866
 867             /**************************
 868              * CALCULATE INTERACTIONS *
 869              **************************/
 870
 871             if (gmx_mm_any_lt(rsq01,rcutoff2))
 872             {
 873
 874             r01              = _mm_mul_ps(rsq01,rinv01);
 875             r01              = _mm_andnot_ps(dummy_mask,r01);
 876
 877             /* EWALD ELECTROSTATICS */
 878
 879             /* Analytical PME correction */
 880             zeta2            = _mm_mul_ps(beta2,rsq01);
 881             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 882             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 883             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 884             felec            = _mm_mul_ps(qq01,felec);
 885             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 886             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 887             velec            = _mm_mul_ps(qq01,velec);
 888
 889             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 890
 891             /* Update potential sum for this i atom from the interaction with this j atom. */
 892             velec            = _mm_and_ps(velec,cutoff_mask);
 893             velec            = _mm_andnot_ps(dummy_mask,velec);
 894             velecsum         = _mm_add_ps(velecsum,velec);
 895
 896             fscal            = felec;
 897
 898             fscal            = _mm_and_ps(fscal,cutoff_mask);
 899
 900             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 901
 902              /* Update vectorial force */
 903             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 904             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 905             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 906
 907             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 908             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 909             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 910
 911             }
 912
 913             /**************************
 914              * CALCULATE INTERACTIONS *
 915              **************************/
 916
 917             if (gmx_mm_any_lt(rsq02,rcutoff2))
 918             {
 919
 920             r02              = _mm_mul_ps(rsq02,rinv02);
 921             r02              = _mm_andnot_ps(dummy_mask,r02);
 922
 923             /* EWALD ELECTROSTATICS */
 924
 925             /* Analytical PME correction */
 926             zeta2            = _mm_mul_ps(beta2,rsq02);
 927             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 928             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 929             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 930             felec            = _mm_mul_ps(qq02,felec);
 931             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 932             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 933             velec            = _mm_mul_ps(qq02,velec);
 934
 935             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 936
 937             /* Update potential sum for this i atom from the interaction with this j atom. */
 938             velec            = _mm_and_ps(velec,cutoff_mask);
 939             velec            = _mm_andnot_ps(dummy_mask,velec);
 940             velecsum         = _mm_add_ps(velecsum,velec);
 941
 942             fscal            = felec;
 943
 944             fscal            = _mm_and_ps(fscal,cutoff_mask);
 945
 946             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 947
 948              /* Update vectorial force */
 949             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 950             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 951             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 952
 953             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 954             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 955             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 956
 957             }
 958
 959             /**************************
 960              * CALCULATE INTERACTIONS *
 961              **************************/
 962
 963             if (gmx_mm_any_lt(rsq10,rcutoff2))
 964             {
 965
 966             r10              = _mm_mul_ps(rsq10,rinv10);
 967             r10              = _mm_andnot_ps(dummy_mask,r10);
 968
 969             /* EWALD ELECTROSTATICS */
 970
 971             /* Analytical PME correction */
 972             zeta2            = _mm_mul_ps(beta2,rsq10);
 973             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 974             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 975             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 976             felec            = _mm_mul_ps(qq10,felec);
 977             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 978             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 979             velec            = _mm_mul_ps(qq10,velec);
 980
 981             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 982
 983             /* Update potential sum for this i atom from the interaction with this j atom. */
 984             velec            = _mm_and_ps(velec,cutoff_mask);
 985             velec            = _mm_andnot_ps(dummy_mask,velec);
 986             velecsum         = _mm_add_ps(velecsum,velec);
 987
 988             fscal            = felec;
 989
 990             fscal            = _mm_and_ps(fscal,cutoff_mask);
 991
 992             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 993
 994              /* Update vectorial force */
 995             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 996             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 997             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 998
 999             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1000             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1001             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1002
1003             }
1004
1005             /**************************
1006              * CALCULATE INTERACTIONS *
1007              **************************/
1008
1009             if (gmx_mm_any_lt(rsq11,rcutoff2))
1010             {
1011
1012             r11              = _mm_mul_ps(rsq11,rinv11);
1013             r11              = _mm_andnot_ps(dummy_mask,r11);
1014
1015             /* EWALD ELECTROSTATICS */
1016
1017             /* Analytical PME correction */
1018             zeta2            = _mm_mul_ps(beta2,rsq11);
1019             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1020             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1021             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1022             felec            = _mm_mul_ps(qq11,felec);
1023             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1024             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
1025             velec            = _mm_mul_ps(qq11,velec);
1026
1027             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1028
1029             /* Update potential sum for this i atom from the interaction with this j atom. */
1030             velec            = _mm_and_ps(velec,cutoff_mask);
1031             velec            = _mm_andnot_ps(dummy_mask,velec);
1032             velecsum         = _mm_add_ps(velecsum,velec);
1033
1034             fscal            = felec;
1035
1036             fscal            = _mm_and_ps(fscal,cutoff_mask);
1037
1038             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1039
1040              /* Update vectorial force */
1041             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1042             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1043             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1044
1045             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1046             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1047             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1048
1049             }
1050
1051             /**************************
1052              * CALCULATE INTERACTIONS *
1053              **************************/
1054
1055             if (gmx_mm_any_lt(rsq12,rcutoff2))
1056             {
1057
1058             r12              = _mm_mul_ps(rsq12,rinv12);
1059             r12              = _mm_andnot_ps(dummy_mask,r12);
1060
1061             /* EWALD ELECTROSTATICS */
1062
1063             /* Analytical PME correction */
1064             zeta2            = _mm_mul_ps(beta2,rsq12);
1065             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1066             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1067             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1068             felec            = _mm_mul_ps(qq12,felec);
1069             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1070             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
1071             velec            = _mm_mul_ps(qq12,velec);
1072
1073             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1074
1075             /* Update potential sum for this i atom from the interaction with this j atom. */
1076             velec            = _mm_and_ps(velec,cutoff_mask);
1077             velec            = _mm_andnot_ps(dummy_mask,velec);
1078             velecsum         = _mm_add_ps(velecsum,velec);
1079
1080             fscal            = felec;
1081
1082             fscal            = _mm_and_ps(fscal,cutoff_mask);
1083
1084             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1085
1086              /* Update vectorial force */
1087             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1088             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1089             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1090
1091             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1092             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1093             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1094
1095             }
1096
1097             /**************************
1098              * CALCULATE INTERACTIONS *
1099              **************************/
1100
1101             if (gmx_mm_any_lt(rsq20,rcutoff2))
1102             {
1103
1104             r20              = _mm_mul_ps(rsq20,rinv20);
1105             r20              = _mm_andnot_ps(dummy_mask,r20);
1106
1107             /* EWALD ELECTROSTATICS */
1108
1109             /* Analytical PME correction */
1110             zeta2            = _mm_mul_ps(beta2,rsq20);
1111             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1112             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1113             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1114             felec            = _mm_mul_ps(qq20,felec);
1115             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1116             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
1117             velec            = _mm_mul_ps(qq20,velec);
1118
1119             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1120
1121             /* Update potential sum for this i atom from the interaction with this j atom. */
1122             velec            = _mm_and_ps(velec,cutoff_mask);
1123             velec            = _mm_andnot_ps(dummy_mask,velec);
1124             velecsum         = _mm_add_ps(velecsum,velec);
1125
1126             fscal            = felec;
1127
1128             fscal            = _mm_and_ps(fscal,cutoff_mask);
1129
1130             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1131
1132              /* Update vectorial force */
1133             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1134             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1135             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1136
1137             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1138             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1139             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1140
1141             }
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             if (gmx_mm_any_lt(rsq21,rcutoff2))
1148             {
1149
1150             r21              = _mm_mul_ps(rsq21,rinv21);
1151             r21              = _mm_andnot_ps(dummy_mask,r21);
1152
1153             /* EWALD ELECTROSTATICS */
1154
1155             /* Analytical PME correction */
1156             zeta2            = _mm_mul_ps(beta2,rsq21);
1157             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1158             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1159             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1160             felec            = _mm_mul_ps(qq21,felec);
1161             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1162             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
1163             velec            = _mm_mul_ps(qq21,velec);
1164
1165             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1166
1167             /* Update potential sum for this i atom from the interaction with this j atom. */
1168             velec            = _mm_and_ps(velec,cutoff_mask);
1169             velec            = _mm_andnot_ps(dummy_mask,velec);
1170             velecsum         = _mm_add_ps(velecsum,velec);
1171
1172             fscal            = felec;
1173
1174             fscal            = _mm_and_ps(fscal,cutoff_mask);
1175
1176             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1177
1178              /* Update vectorial force */
1179             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1180             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1181             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1182
1183             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1184             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1185             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1186
1187             }
1188
1189             /**************************
1190              * CALCULATE INTERACTIONS *
1191              **************************/
1192
1193             if (gmx_mm_any_lt(rsq22,rcutoff2))
1194             {
1195
1196             r22              = _mm_mul_ps(rsq22,rinv22);
1197             r22              = _mm_andnot_ps(dummy_mask,r22);
1198
1199             /* EWALD ELECTROSTATICS */
1200
1201             /* Analytical PME correction */
1202             zeta2            = _mm_mul_ps(beta2,rsq22);
1203             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1204             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1205             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1206             felec            = _mm_mul_ps(qq22,felec);
1207             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1208             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
1209             velec            = _mm_mul_ps(qq22,velec);
1210
1211             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1212
1213             /* Update potential sum for this i atom from the interaction with this j atom. */
1214             velec            = _mm_and_ps(velec,cutoff_mask);
1215             velec            = _mm_andnot_ps(dummy_mask,velec);
1216             velecsum         = _mm_add_ps(velecsum,velec);
1217
1218             fscal            = felec;
1219
1220             fscal            = _mm_and_ps(fscal,cutoff_mask);
1221
1222             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1223
1224              /* Update vectorial force */
1225             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1226             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1227             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1228
1229             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1230             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1231             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1232
1233             }
1234
1235             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1236             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1237             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1238             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1239
1240             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1241                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1242
1243             /* Inner loop uses 324 flops */
1244         }
1245
1246         /* End of innermost loop */
1247
1248         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1249                                               f+i_coord_offset,fshift+i_shift_offset);
1250
1251         ggid                        = gid[iidx];
1252         /* Update potential energies */
1253         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1254         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1255
1256         /* Increment number of inner iterations */
1257         inneriter                  += j_index_end - j_index_start;
1258
1259         /* Outer loop uses 20 flops */
1260     }
1261
1262     /* Increment number of outer iterations */
1263     outeriter        += nri;
1264
1265     /* Update outer/inner flops */
1266
1267     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*324);
1268 }
1269 /*
1270  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_F_avx_128_fma_single
1271  * Electrostatics interaction: Ewald
1272  * VdW interaction:            LennardJones
1273  * Geometry:                   Water3-Water3
1274  * Calculate force/pot:        Force
1275  */
1276 void
1277 nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_F_avx_128_fma_single
1278                     (t_nblist                    * gmx_restrict       nlist,
1279                      rvec                        * gmx_restrict          xx,
1280                      rvec                        * gmx_restrict          ff,
1281                      t_forcerec                  * gmx_restrict          fr,
1282                      t_mdatoms                   * gmx_restrict     mdatoms,
1283                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1284                      t_nrnb                      * gmx_restrict        nrnb)
1285 {
1286     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1287      * just 0 for non-waters.
1288      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1289      * jnr indices corresponding to data put in the four positions in the SIMD register.
1290      */
1291     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1292     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1293     int              jnrA,jnrB,jnrC,jnrD;
1294     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1295     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1296     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1297     real             rcutoff_scalar;
1298     real             *shiftvec,*fshift,*x,*f;
1299     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1300     real             scratch[4*DIM];
1301     __m128           fscal,rcutoff,rcutoff2,jidxall;
1302     int              vdwioffset0;
1303     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1304     int              vdwioffset1;
1305     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1306     int              vdwioffset2;
1307     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1308     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1309     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1310     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1311     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1312     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1313     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1314     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1315     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1316     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1317     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1318     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1319     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1320     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1321     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1322     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1323     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1324     real             *charge;
1325     int              nvdwtype;
1326     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1327     int              *vdwtype;
1328     real             *vdwparam;
1329     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1330     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1331     __m128i          ewitab;
1332     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1333     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1334     real             *ewtab;
1335     __m128           dummy_mask,cutoff_mask;
1336     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1337     __m128           one     = _mm_set1_ps(1.0);
1338     __m128           two     = _mm_set1_ps(2.0);
1339     x                = xx[0];
1340     f                = ff[0];
1341
1342     nri              = nlist->nri;
1343     iinr             = nlist->iinr;
1344     jindex           = nlist->jindex;
1345     jjnr             = nlist->jjnr;
1346     shiftidx         = nlist->shift;
1347     gid              = nlist->gid;
1348     shiftvec         = fr->shift_vec[0];
1349     fshift           = fr->fshift[0];
1350     facel            = _mm_set1_ps(fr->epsfac);
1351     charge           = mdatoms->chargeA;
1352     nvdwtype         = fr->ntype;
1353     vdwparam         = fr->nbfp;
1354     vdwtype          = mdatoms->typeA;
1355
1356     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1357     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
1358     beta2            = _mm_mul_ps(beta,beta);
1359     beta3            = _mm_mul_ps(beta,beta2);
1360     ewtab            = fr->ic->tabq_coul_F;
1361     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1362     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1363
1364     /* Setup water-specific parameters */
1365     inr              = nlist->iinr[0];
1366     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1367     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1368     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1369     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1370
1371     jq0              = _mm_set1_ps(charge[inr+0]);
1372     jq1              = _mm_set1_ps(charge[inr+1]);
1373     jq2              = _mm_set1_ps(charge[inr+2]);
1374     vdwjidx0A        = 2*vdwtype[inr+0];
1375     qq00             = _mm_mul_ps(iq0,jq0);
1376     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1377     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1378     qq01             = _mm_mul_ps(iq0,jq1);
1379     qq02             = _mm_mul_ps(iq0,jq2);
1380     qq10             = _mm_mul_ps(iq1,jq0);
1381     qq11             = _mm_mul_ps(iq1,jq1);
1382     qq12             = _mm_mul_ps(iq1,jq2);
1383     qq20             = _mm_mul_ps(iq2,jq0);
1384     qq21             = _mm_mul_ps(iq2,jq1);
1385     qq22             = _mm_mul_ps(iq2,jq2);
1386
1387     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1388     rcutoff_scalar   = fr->rcoulomb;
1389     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1390     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1391
1392     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1393     rvdw             = _mm_set1_ps(fr->rvdw);
1394
1395     /* Avoid stupid compiler warnings */
1396     jnrA = jnrB = jnrC = jnrD = 0;
1397     j_coord_offsetA = 0;
1398     j_coord_offsetB = 0;
1399     j_coord_offsetC = 0;
1400     j_coord_offsetD = 0;
1401
1402     outeriter        = 0;
1403     inneriter        = 0;
1404
1405     for(iidx=0;iidx<4*DIM;iidx++)
1406     {
1407         scratch[iidx] = 0.0;
1408     }
1409
1410     /* Start outer loop over neighborlists */
1411     for(iidx=0; iidx<nri; iidx++)
1412     {
1413         /* Load shift vector for this list */
1414         i_shift_offset   = DIM*shiftidx[iidx];
1415
1416         /* Load limits for loop over neighbors */
1417         j_index_start    = jindex[iidx];
1418         j_index_end      = jindex[iidx+1];
1419
1420         /* Get outer coordinate index */
1421         inr              = iinr[iidx];
1422         i_coord_offset   = DIM*inr;
1423
1424         /* Load i particle coords and add shift vector */
1425         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1426                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1427
1428         fix0             = _mm_setzero_ps();
1429         fiy0             = _mm_setzero_ps();
1430         fiz0             = _mm_setzero_ps();
1431         fix1             = _mm_setzero_ps();
1432         fiy1             = _mm_setzero_ps();
1433         fiz1             = _mm_setzero_ps();
1434         fix2             = _mm_setzero_ps();
1435         fiy2             = _mm_setzero_ps();
1436         fiz2             = _mm_setzero_ps();
1437
1438         /* Start inner kernel loop */
1439         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1440         {
1441
1442             /* Get j neighbor index, and coordinate index */
1443             jnrA             = jjnr[jidx];
1444             jnrB             = jjnr[jidx+1];
1445             jnrC             = jjnr[jidx+2];
1446             jnrD             = jjnr[jidx+3];
1447             j_coord_offsetA  = DIM*jnrA;
1448             j_coord_offsetB  = DIM*jnrB;
1449             j_coord_offsetC  = DIM*jnrC;
1450             j_coord_offsetD  = DIM*jnrD;
1451
1452             /* load j atom coordinates */
1453             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1454                                               x+j_coord_offsetC,x+j_coord_offsetD,
1455                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1456
1457             /* Calculate displacement vector */
1458             dx00             = _mm_sub_ps(ix0,jx0);
1459             dy00             = _mm_sub_ps(iy0,jy0);
1460             dz00             = _mm_sub_ps(iz0,jz0);
1461             dx01             = _mm_sub_ps(ix0,jx1);
1462             dy01             = _mm_sub_ps(iy0,jy1);
1463             dz01             = _mm_sub_ps(iz0,jz1);
1464             dx02             = _mm_sub_ps(ix0,jx2);
1465             dy02             = _mm_sub_ps(iy0,jy2);
1466             dz02             = _mm_sub_ps(iz0,jz2);
1467             dx10             = _mm_sub_ps(ix1,jx0);
1468             dy10             = _mm_sub_ps(iy1,jy0);
1469             dz10             = _mm_sub_ps(iz1,jz0);
1470             dx11             = _mm_sub_ps(ix1,jx1);
1471             dy11             = _mm_sub_ps(iy1,jy1);
1472             dz11             = _mm_sub_ps(iz1,jz1);
1473             dx12             = _mm_sub_ps(ix1,jx2);
1474             dy12             = _mm_sub_ps(iy1,jy2);
1475             dz12             = _mm_sub_ps(iz1,jz2);
1476             dx20             = _mm_sub_ps(ix2,jx0);
1477             dy20             = _mm_sub_ps(iy2,jy0);
1478             dz20             = _mm_sub_ps(iz2,jz0);
1479             dx21             = _mm_sub_ps(ix2,jx1);
1480             dy21             = _mm_sub_ps(iy2,jy1);
1481             dz21             = _mm_sub_ps(iz2,jz1);
1482             dx22             = _mm_sub_ps(ix2,jx2);
1483             dy22             = _mm_sub_ps(iy2,jy2);
1484             dz22             = _mm_sub_ps(iz2,jz2);
1485
1486             /* Calculate squared distance and things based on it */
1487             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1488             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1489             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1490             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1491             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1492             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1493             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1494             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1495             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1496
1497             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1498             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1499             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1500             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1501             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1502             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1503             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1504             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1505             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1506
1507             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1508             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1509             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1510             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1511             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1512             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1513             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1514             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1515             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1516
1517             fjx0             = _mm_setzero_ps();
1518             fjy0             = _mm_setzero_ps();
1519             fjz0             = _mm_setzero_ps();
1520             fjx1             = _mm_setzero_ps();
1521             fjy1             = _mm_setzero_ps();
1522             fjz1             = _mm_setzero_ps();
1523             fjx2             = _mm_setzero_ps();
1524             fjy2             = _mm_setzero_ps();
1525             fjz2             = _mm_setzero_ps();
1526
1527             /**************************
1528              * CALCULATE INTERACTIONS *
1529              **************************/
1530
1531             if (gmx_mm_any_lt(rsq00,rcutoff2))
1532             {
1533
1534             r00              = _mm_mul_ps(rsq00,rinv00);
1535
1536             /* EWALD ELECTROSTATICS */
1537
1538             /* Analytical PME correction */
1539             zeta2            = _mm_mul_ps(beta2,rsq00);
1540             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1541             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1542             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1543             felec            = _mm_mul_ps(qq00,felec);
1544
1545             /* LENNARD-JONES DISPERSION/REPULSION */
1546
1547             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1548             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1549
1550             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1551
1552             fscal            = _mm_add_ps(felec,fvdw);
1553
1554             fscal            = _mm_and_ps(fscal,cutoff_mask);
1555
1556              /* Update vectorial force */
1557             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1558             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1559             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1560
1561             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1562             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1563             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1564
1565             }
1566
1567             /**************************
1568              * CALCULATE INTERACTIONS *
1569              **************************/
1570
1571             if (gmx_mm_any_lt(rsq01,rcutoff2))
1572             {
1573
1574             r01              = _mm_mul_ps(rsq01,rinv01);
1575
1576             /* EWALD ELECTROSTATICS */
1577
1578             /* Analytical PME correction */
1579             zeta2            = _mm_mul_ps(beta2,rsq01);
1580             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1581             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1582             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1583             felec            = _mm_mul_ps(qq01,felec);
1584
1585             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1586
1587             fscal            = felec;
1588
1589             fscal            = _mm_and_ps(fscal,cutoff_mask);
1590
1591              /* Update vectorial force */
1592             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1593             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1594             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1595
1596             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1597             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1598             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1599
1600             }
1601
1602             /**************************
1603              * CALCULATE INTERACTIONS *
1604              **************************/
1605
1606             if (gmx_mm_any_lt(rsq02,rcutoff2))
1607             {
1608
1609             r02              = _mm_mul_ps(rsq02,rinv02);
1610
1611             /* EWALD ELECTROSTATICS */
1612
1613             /* Analytical PME correction */
1614             zeta2            = _mm_mul_ps(beta2,rsq02);
1615             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1616             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1617             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1618             felec            = _mm_mul_ps(qq02,felec);
1619
1620             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1621
1622             fscal            = felec;
1623
1624             fscal            = _mm_and_ps(fscal,cutoff_mask);
1625
1626              /* Update vectorial force */
1627             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1628             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1629             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1630
1631             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1632             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1633             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1634
1635             }
1636
1637             /**************************
1638              * CALCULATE INTERACTIONS *
1639              **************************/
1640
1641             if (gmx_mm_any_lt(rsq10,rcutoff2))
1642             {
1643
1644             r10              = _mm_mul_ps(rsq10,rinv10);
1645
1646             /* EWALD ELECTROSTATICS */
1647
1648             /* Analytical PME correction */
1649             zeta2            = _mm_mul_ps(beta2,rsq10);
1650             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1651             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1652             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1653             felec            = _mm_mul_ps(qq10,felec);
1654
1655             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1656
1657             fscal            = felec;
1658
1659             fscal            = _mm_and_ps(fscal,cutoff_mask);
1660
1661              /* Update vectorial force */
1662             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1663             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1664             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1665
1666             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1667             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1668             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1669
1670             }
1671
1672             /**************************
1673              * CALCULATE INTERACTIONS *
1674              **************************/
1675
1676             if (gmx_mm_any_lt(rsq11,rcutoff2))
1677             {
1678
1679             r11              = _mm_mul_ps(rsq11,rinv11);
1680
1681             /* EWALD ELECTROSTATICS */
1682
1683             /* Analytical PME correction */
1684             zeta2            = _mm_mul_ps(beta2,rsq11);
1685             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1686             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1687             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1688             felec            = _mm_mul_ps(qq11,felec);
1689
1690             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1691
1692             fscal            = felec;
1693
1694             fscal            = _mm_and_ps(fscal,cutoff_mask);
1695
1696              /* Update vectorial force */
1697             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1698             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1699             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1700
1701             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1702             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1703             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1704
1705             }
1706
1707             /**************************
1708              * CALCULATE INTERACTIONS *
1709              **************************/
1710
1711             if (gmx_mm_any_lt(rsq12,rcutoff2))
1712             {
1713
1714             r12              = _mm_mul_ps(rsq12,rinv12);
1715
1716             /* EWALD ELECTROSTATICS */
1717
1718             /* Analytical PME correction */
1719             zeta2            = _mm_mul_ps(beta2,rsq12);
1720             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1721             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1722             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1723             felec            = _mm_mul_ps(qq12,felec);
1724
1725             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_and_ps(fscal,cutoff_mask);
1730
1731              /* Update vectorial force */
1732             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1733             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1734             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1735
1736             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1737             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1738             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1739
1740             }
1741
1742             /**************************
1743              * CALCULATE INTERACTIONS *
1744              **************************/
1745
1746             if (gmx_mm_any_lt(rsq20,rcutoff2))
1747             {
1748
1749             r20              = _mm_mul_ps(rsq20,rinv20);
1750
1751             /* EWALD ELECTROSTATICS */
1752
1753             /* Analytical PME correction */
1754             zeta2            = _mm_mul_ps(beta2,rsq20);
1755             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1756             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1757             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1758             felec            = _mm_mul_ps(qq20,felec);
1759
1760             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1761
1762             fscal            = felec;
1763
1764             fscal            = _mm_and_ps(fscal,cutoff_mask);
1765
1766              /* Update vectorial force */
1767             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1768             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1769             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1770
1771             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1772             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1773             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1774
1775             }
1776
1777             /**************************
1778              * CALCULATE INTERACTIONS *
1779              **************************/
1780
1781             if (gmx_mm_any_lt(rsq21,rcutoff2))
1782             {
1783
1784             r21              = _mm_mul_ps(rsq21,rinv21);
1785
1786             /* EWALD ELECTROSTATICS */
1787
1788             /* Analytical PME correction */
1789             zeta2            = _mm_mul_ps(beta2,rsq21);
1790             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1791             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1792             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1793             felec            = _mm_mul_ps(qq21,felec);
1794
1795             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1796
1797             fscal            = felec;
1798
1799             fscal            = _mm_and_ps(fscal,cutoff_mask);
1800
1801              /* Update vectorial force */
1802             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1803             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1804             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1805
1806             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1807             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1808             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1809
1810             }
1811
1812             /**************************
1813              * CALCULATE INTERACTIONS *
1814              **************************/
1815
1816             if (gmx_mm_any_lt(rsq22,rcutoff2))
1817             {
1818
1819             r22              = _mm_mul_ps(rsq22,rinv22);
1820
1821             /* EWALD ELECTROSTATICS */
1822
1823             /* Analytical PME correction */
1824             zeta2            = _mm_mul_ps(beta2,rsq22);
1825             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1826             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1827             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1828             felec            = _mm_mul_ps(qq22,felec);
1829
1830             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1831
1832             fscal            = felec;
1833
1834             fscal            = _mm_and_ps(fscal,cutoff_mask);
1835
1836              /* Update vectorial force */
1837             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1838             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1839             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1840
1841             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1842             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1843             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1844
1845             }
1846
1847             fjptrA             = f+j_coord_offsetA;
1848             fjptrB             = f+j_coord_offsetB;
1849             fjptrC             = f+j_coord_offsetC;
1850             fjptrD             = f+j_coord_offsetD;
1851
1852             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1853                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1854
1855             /* Inner loop uses 286 flops */
1856         }
1857
1858         if(jidx<j_index_end)
1859         {
1860
1861             /* Get j neighbor index, and coordinate index */
1862             jnrlistA         = jjnr[jidx];
1863             jnrlistB         = jjnr[jidx+1];
1864             jnrlistC         = jjnr[jidx+2];
1865             jnrlistD         = jjnr[jidx+3];
1866             /* Sign of each element will be negative for non-real atoms.
1867              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1868              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1869              */
1870             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1871             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1872             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1873             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1874             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1875             j_coord_offsetA  = DIM*jnrA;
1876             j_coord_offsetB  = DIM*jnrB;
1877             j_coord_offsetC  = DIM*jnrC;
1878             j_coord_offsetD  = DIM*jnrD;
1879
1880             /* load j atom coordinates */
1881             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1882                                               x+j_coord_offsetC,x+j_coord_offsetD,
1883                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1884
1885             /* Calculate displacement vector */
1886             dx00             = _mm_sub_ps(ix0,jx0);
1887             dy00             = _mm_sub_ps(iy0,jy0);
1888             dz00             = _mm_sub_ps(iz0,jz0);
1889             dx01             = _mm_sub_ps(ix0,jx1);
1890             dy01             = _mm_sub_ps(iy0,jy1);
1891             dz01             = _mm_sub_ps(iz0,jz1);
1892             dx02             = _mm_sub_ps(ix0,jx2);
1893             dy02             = _mm_sub_ps(iy0,jy2);
1894             dz02             = _mm_sub_ps(iz0,jz2);
1895             dx10             = _mm_sub_ps(ix1,jx0);
1896             dy10             = _mm_sub_ps(iy1,jy0);
1897             dz10             = _mm_sub_ps(iz1,jz0);
1898             dx11             = _mm_sub_ps(ix1,jx1);
1899             dy11             = _mm_sub_ps(iy1,jy1);
1900             dz11             = _mm_sub_ps(iz1,jz1);
1901             dx12             = _mm_sub_ps(ix1,jx2);
1902             dy12             = _mm_sub_ps(iy1,jy2);
1903             dz12             = _mm_sub_ps(iz1,jz2);
1904             dx20             = _mm_sub_ps(ix2,jx0);
1905             dy20             = _mm_sub_ps(iy2,jy0);
1906             dz20             = _mm_sub_ps(iz2,jz0);
1907             dx21             = _mm_sub_ps(ix2,jx1);
1908             dy21             = _mm_sub_ps(iy2,jy1);
1909             dz21             = _mm_sub_ps(iz2,jz1);
1910             dx22             = _mm_sub_ps(ix2,jx2);
1911             dy22             = _mm_sub_ps(iy2,jy2);
1912             dz22             = _mm_sub_ps(iz2,jz2);
1913
1914             /* Calculate squared distance and things based on it */
1915             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1916             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1917             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1918             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1919             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1920             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1921             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1922             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1923             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1924
1925             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1926             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1927             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1928             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1929             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1930             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1931             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1932             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1933             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1934
1935             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1936             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1937             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1938             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1939             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1940             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1941             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1942             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1943             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1944
1945             fjx0             = _mm_setzero_ps();
1946             fjy0             = _mm_setzero_ps();
1947             fjz0             = _mm_setzero_ps();
1948             fjx1             = _mm_setzero_ps();
1949             fjy1             = _mm_setzero_ps();
1950             fjz1             = _mm_setzero_ps();
1951             fjx2             = _mm_setzero_ps();
1952             fjy2             = _mm_setzero_ps();
1953             fjz2             = _mm_setzero_ps();
1954
1955             /**************************
1956              * CALCULATE INTERACTIONS *
1957              **************************/
1958
1959             if (gmx_mm_any_lt(rsq00,rcutoff2))
1960             {
1961
1962             r00              = _mm_mul_ps(rsq00,rinv00);
1963             r00              = _mm_andnot_ps(dummy_mask,r00);
1964
1965             /* EWALD ELECTROSTATICS */
1966
1967             /* Analytical PME correction */
1968             zeta2            = _mm_mul_ps(beta2,rsq00);
1969             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1970             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1971             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1972             felec            = _mm_mul_ps(qq00,felec);
1973
1974             /* LENNARD-JONES DISPERSION/REPULSION */
1975
1976             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1977             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1978
1979             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1980
1981             fscal            = _mm_add_ps(felec,fvdw);
1982
1983             fscal            = _mm_and_ps(fscal,cutoff_mask);
1984
1985             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1986
1987              /* Update vectorial force */
1988             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1989             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1990             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1991
1992             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1993             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1994             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1995
1996             }
1997
1998             /**************************
1999              * CALCULATE INTERACTIONS *
2000              **************************/
2001
2002             if (gmx_mm_any_lt(rsq01,rcutoff2))
2003             {
2004
2005             r01              = _mm_mul_ps(rsq01,rinv01);
2006             r01              = _mm_andnot_ps(dummy_mask,r01);
2007
2008             /* EWALD ELECTROSTATICS */
2009
2010             /* Analytical PME correction */
2011             zeta2            = _mm_mul_ps(beta2,rsq01);
2012             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
2013             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2014             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2015             felec            = _mm_mul_ps(qq01,felec);
2016
2017             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
2018
2019             fscal            = felec;
2020
2021             fscal            = _mm_and_ps(fscal,cutoff_mask);
2022
2023             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2024
2025              /* Update vectorial force */
2026             fix0             = _mm_macc_ps(dx01,fscal,fix0);
2027             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
2028             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
2029
2030             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
2031             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
2032             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
2033
2034             }
2035
2036             /**************************
2037              * CALCULATE INTERACTIONS *
2038              **************************/
2039
2040             if (gmx_mm_any_lt(rsq02,rcutoff2))
2041             {
2042
2043             r02              = _mm_mul_ps(rsq02,rinv02);
2044             r02              = _mm_andnot_ps(dummy_mask,r02);
2045
2046             /* EWALD ELECTROSTATICS */
2047
2048             /* Analytical PME correction */
2049             zeta2            = _mm_mul_ps(beta2,rsq02);
2050             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
2051             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2052             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2053             felec            = _mm_mul_ps(qq02,felec);
2054
2055             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
2056
2057             fscal            = felec;
2058
2059             fscal            = _mm_and_ps(fscal,cutoff_mask);
2060
2061             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2062
2063              /* Update vectorial force */
2064             fix0             = _mm_macc_ps(dx02,fscal,fix0);
2065             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
2066             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
2067
2068             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
2069             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
2070             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
2071
2072             }
2073
2074             /**************************
2075              * CALCULATE INTERACTIONS *
2076              **************************/
2077
2078             if (gmx_mm_any_lt(rsq10,rcutoff2))
2079             {
2080
2081             r10              = _mm_mul_ps(rsq10,rinv10);
2082             r10              = _mm_andnot_ps(dummy_mask,r10);
2083
2084             /* EWALD ELECTROSTATICS */
2085
2086             /* Analytical PME correction */
2087             zeta2            = _mm_mul_ps(beta2,rsq10);
2088             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
2089             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2090             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2091             felec            = _mm_mul_ps(qq10,felec);
2092
2093             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
2094
2095             fscal            = felec;
2096
2097             fscal            = _mm_and_ps(fscal,cutoff_mask);
2098
2099             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2100
2101              /* Update vectorial force */
2102             fix1             = _mm_macc_ps(dx10,fscal,fix1);
2103             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
2104             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
2105
2106             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
2107             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
2108             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
2109
2110             }
2111
2112             /**************************
2113              * CALCULATE INTERACTIONS *
2114              **************************/
2115
2116             if (gmx_mm_any_lt(rsq11,rcutoff2))
2117             {
2118
2119             r11              = _mm_mul_ps(rsq11,rinv11);
2120             r11              = _mm_andnot_ps(dummy_mask,r11);
2121
2122             /* EWALD ELECTROSTATICS */
2123
2124             /* Analytical PME correction */
2125             zeta2            = _mm_mul_ps(beta2,rsq11);
2126             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
2127             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2128             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2129             felec            = _mm_mul_ps(qq11,felec);
2130
2131             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2132
2133             fscal            = felec;
2134
2135             fscal            = _mm_and_ps(fscal,cutoff_mask);
2136
2137             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2138
2139              /* Update vectorial force */
2140             fix1             = _mm_macc_ps(dx11,fscal,fix1);
2141             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
2142             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
2143
2144             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
2145             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
2146             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
2147
2148             }
2149
2150             /**************************
2151              * CALCULATE INTERACTIONS *
2152              **************************/
2153
2154             if (gmx_mm_any_lt(rsq12,rcutoff2))
2155             {
2156
2157             r12              = _mm_mul_ps(rsq12,rinv12);
2158             r12              = _mm_andnot_ps(dummy_mask,r12);
2159
2160             /* EWALD ELECTROSTATICS */
2161
2162             /* Analytical PME correction */
2163             zeta2            = _mm_mul_ps(beta2,rsq12);
2164             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
2165             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2166             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2167             felec            = _mm_mul_ps(qq12,felec);
2168
2169             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2170
2171             fscal            = felec;
2172
2173             fscal            = _mm_and_ps(fscal,cutoff_mask);
2174
2175             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2176
2177              /* Update vectorial force */
2178             fix1             = _mm_macc_ps(dx12,fscal,fix1);
2179             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
2180             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
2181
2182             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
2183             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
2184             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
2185
2186             }
2187
2188             /**************************
2189              * CALCULATE INTERACTIONS *
2190              **************************/
2191
2192             if (gmx_mm_any_lt(rsq20,rcutoff2))
2193             {
2194
2195             r20              = _mm_mul_ps(rsq20,rinv20);
2196             r20              = _mm_andnot_ps(dummy_mask,r20);
2197
2198             /* EWALD ELECTROSTATICS */
2199
2200             /* Analytical PME correction */
2201             zeta2            = _mm_mul_ps(beta2,rsq20);
2202             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
2203             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2204             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2205             felec            = _mm_mul_ps(qq20,felec);
2206
2207             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2208
2209             fscal            = felec;
2210
2211             fscal            = _mm_and_ps(fscal,cutoff_mask);
2212
2213             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2214
2215              /* Update vectorial force */
2216             fix2             = _mm_macc_ps(dx20,fscal,fix2);
2217             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
2218             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
2219
2220             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
2221             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
2222             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
2223
2224             }
2225
2226             /**************************
2227              * CALCULATE INTERACTIONS *
2228              **************************/
2229
2230             if (gmx_mm_any_lt(rsq21,rcutoff2))
2231             {
2232
2233             r21              = _mm_mul_ps(rsq21,rinv21);
2234             r21              = _mm_andnot_ps(dummy_mask,r21);
2235
2236             /* EWALD ELECTROSTATICS */
2237
2238             /* Analytical PME correction */
2239             zeta2            = _mm_mul_ps(beta2,rsq21);
2240             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
2241             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2242             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2243             felec            = _mm_mul_ps(qq21,felec);
2244
2245             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2246
2247             fscal            = felec;
2248
2249             fscal            = _mm_and_ps(fscal,cutoff_mask);
2250
2251             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2252
2253              /* Update vectorial force */
2254             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2255             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2256             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2257
2258             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2259             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2260             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2261
2262             }
2263
2264             /**************************
2265              * CALCULATE INTERACTIONS *
2266              **************************/
2267
2268             if (gmx_mm_any_lt(rsq22,rcutoff2))
2269             {
2270
2271             r22              = _mm_mul_ps(rsq22,rinv22);
2272             r22              = _mm_andnot_ps(dummy_mask,r22);
2273
2274             /* EWALD ELECTROSTATICS */
2275
2276             /* Analytical PME correction */
2277             zeta2            = _mm_mul_ps(beta2,rsq22);
2278             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
2279             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2280             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2281             felec            = _mm_mul_ps(qq22,felec);
2282
2283             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2284
2285             fscal            = felec;
2286
2287             fscal            = _mm_and_ps(fscal,cutoff_mask);
2288
2289             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2290
2291              /* Update vectorial force */
2292             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2293             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2294             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2295
2296             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2297             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2298             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2299
2300             }
2301
2302             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2303             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2304             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2305             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2306
2307             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2308                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2309
2310             /* Inner loop uses 295 flops */
2311         }
2312
2313         /* End of innermost loop */
2314
2315         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2316                                               f+i_coord_offset,fshift+i_shift_offset);
2317
2318         /* Increment number of inner iterations */
2319         inneriter                  += j_index_end - j_index_start;
2320
2321         /* Outer loop uses 18 flops */
2322     }
2323
2324     /* Increment number of outer iterations */
2325     outeriter        += nri;
2326
2327     /* Update outer/inner flops */
2328
2329     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*295);
2330 }