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, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 113     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 114     __m128i          ewitab;
 115     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 116     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 117     real             *ewtab;
 118     __m128           dummy_mask,cutoff_mask;
 119     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 120     __m128           one     = _mm_set1_ps(1.0);
 121     __m128           two     = _mm_set1_ps(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_ps(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     nvdwtype         = fr->ntype;
 136     vdwparam         = fr->nbfp;
 137     vdwtype          = mdatoms->typeA;
 138
 139     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 140     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 141     beta2            = _mm_mul_ps(beta,beta);
 142     beta3            = _mm_mul_ps(beta,beta2);
 143     ewtab            = fr->ic->tabq_coul_FDV0;
 144     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 145     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 146
 147     /* Setup water-specific parameters */
 148     inr              = nlist->iinr[0];
 149     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 150     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 151     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 152     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 153
 154     jq0              = _mm_set1_ps(charge[inr+0]);
 155     jq1              = _mm_set1_ps(charge[inr+1]);
 156     jq2              = _mm_set1_ps(charge[inr+2]);
 157     vdwjidx0A        = 2*vdwtype[inr+0];
 158     qq00             = _mm_mul_ps(iq0,jq0);
 159     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 160     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 161     qq01             = _mm_mul_ps(iq0,jq1);
 162     qq02             = _mm_mul_ps(iq0,jq2);
 163     qq10             = _mm_mul_ps(iq1,jq0);
 164     qq11             = _mm_mul_ps(iq1,jq1);
 165     qq12             = _mm_mul_ps(iq1,jq2);
 166     qq20             = _mm_mul_ps(iq2,jq0);
 167     qq21             = _mm_mul_ps(iq2,jq1);
 168     qq22             = _mm_mul_ps(iq2,jq2);
 169
 170     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 171     rcutoff_scalar   = fr->rcoulomb;
 172     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 173     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 174
 175     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 176     rvdw             = _mm_set1_ps(fr->rvdw);
 177
 178     /* Avoid stupid compiler warnings */
 179     jnrA = jnrB = jnrC = jnrD = 0;
 180     j_coord_offsetA = 0;
 181     j_coord_offsetB = 0;
 182     j_coord_offsetC = 0;
 183     j_coord_offsetD = 0;
 184
 185     outeriter        = 0;
 186     inneriter        = 0;
 187
 188     for(iidx=0;iidx<4*DIM;iidx++)
 189     {
 190         scratch[iidx] = 0.0;
 191     }
 192
 193     /* Start outer loop over neighborlists */
 194     for(iidx=0; iidx<nri; iidx++)
 195     {
 196         /* Load shift vector for this list */
 197         i_shift_offset   = DIM*shiftidx[iidx];
 198
 199         /* Load limits for loop over neighbors */
 200         j_index_start    = jindex[iidx];
 201         j_index_end      = jindex[iidx+1];
 202
 203         /* Get outer coordinate index */
 204         inr              = iinr[iidx];
 205         i_coord_offset   = DIM*inr;
 206
 207         /* Load i particle coords and add shift vector */
 208         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 209                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 210
 211         fix0             = _mm_setzero_ps();
 212         fiy0             = _mm_setzero_ps();
 213         fiz0             = _mm_setzero_ps();
 214         fix1             = _mm_setzero_ps();
 215         fiy1             = _mm_setzero_ps();
 216         fiz1             = _mm_setzero_ps();
 217         fix2             = _mm_setzero_ps();
 218         fiy2             = _mm_setzero_ps();
 219         fiz2             = _mm_setzero_ps();
 220
 221         /* Reset potential sums */
 222         velecsum         = _mm_setzero_ps();
 223         vvdwsum          = _mm_setzero_ps();
 224
 225         /* Start inner kernel loop */
 226         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 227         {
 228
 229             /* Get j neighbor index, and coordinate index */
 230             jnrA             = jjnr[jidx];
 231             jnrB             = jjnr[jidx+1];
 232             jnrC             = jjnr[jidx+2];
 233             jnrD             = jjnr[jidx+3];
 234             j_coord_offsetA  = DIM*jnrA;
 235             j_coord_offsetB  = DIM*jnrB;
 236             j_coord_offsetC  = DIM*jnrC;
 237             j_coord_offsetD  = DIM*jnrD;
 238
 239             /* load j atom coordinates */
 240             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 241                                               x+j_coord_offsetC,x+j_coord_offsetD,
 242                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 243
 244             /* Calculate displacement vector */
 245             dx00             = _mm_sub_ps(ix0,jx0);
 246             dy00             = _mm_sub_ps(iy0,jy0);
 247             dz00             = _mm_sub_ps(iz0,jz0);
 248             dx01             = _mm_sub_ps(ix0,jx1);
 249             dy01             = _mm_sub_ps(iy0,jy1);
 250             dz01             = _mm_sub_ps(iz0,jz1);
 251             dx02             = _mm_sub_ps(ix0,jx2);
 252             dy02             = _mm_sub_ps(iy0,jy2);
 253             dz02             = _mm_sub_ps(iz0,jz2);
 254             dx10             = _mm_sub_ps(ix1,jx0);
 255             dy10             = _mm_sub_ps(iy1,jy0);
 256             dz10             = _mm_sub_ps(iz1,jz0);
 257             dx11             = _mm_sub_ps(ix1,jx1);
 258             dy11             = _mm_sub_ps(iy1,jy1);
 259             dz11             = _mm_sub_ps(iz1,jz1);
 260             dx12             = _mm_sub_ps(ix1,jx2);
 261             dy12             = _mm_sub_ps(iy1,jy2);
 262             dz12             = _mm_sub_ps(iz1,jz2);
 263             dx20             = _mm_sub_ps(ix2,jx0);
 264             dy20             = _mm_sub_ps(iy2,jy0);
 265             dz20             = _mm_sub_ps(iz2,jz0);
 266             dx21             = _mm_sub_ps(ix2,jx1);
 267             dy21             = _mm_sub_ps(iy2,jy1);
 268             dz21             = _mm_sub_ps(iz2,jz1);
 269             dx22             = _mm_sub_ps(ix2,jx2);
 270             dy22             = _mm_sub_ps(iy2,jy2);
 271             dz22             = _mm_sub_ps(iz2,jz2);
 272
 273             /* Calculate squared distance and things based on it */
 274             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 275             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 276             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 277             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 278             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 279             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 280             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 281             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 282             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 283
 284             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 285             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 286             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 287             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 288             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 289             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 290             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 291             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 292             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 293
 294             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 295             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 296             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 297             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 298             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 299             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 300             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 301             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 302             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 303
 304             fjx0             = _mm_setzero_ps();
 305             fjy0             = _mm_setzero_ps();
 306             fjz0             = _mm_setzero_ps();
 307             fjx1             = _mm_setzero_ps();
 308             fjy1             = _mm_setzero_ps();
 309             fjz1             = _mm_setzero_ps();
 310             fjx2             = _mm_setzero_ps();
 311             fjy2             = _mm_setzero_ps();
 312             fjz2             = _mm_setzero_ps();
 313
 314             /**************************
 315              * CALCULATE INTERACTIONS *
 316              **************************/
 317
 318             if (gmx_mm_any_lt(rsq00,rcutoff2))
 319             {
 320
 321             r00              = _mm_mul_ps(rsq00,rinv00);
 322
 323             /* EWALD ELECTROSTATICS */
 324
 325             /* Analytical PME correction */
 326             zeta2            = _mm_mul_ps(beta2,rsq00);
 327             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 328             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 329             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 330             felec            = _mm_mul_ps(qq00,felec);
 331             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 332             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 333             velec            = _mm_mul_ps(qq00,velec);
 334
 335             /* LENNARD-JONES DISPERSION/REPULSION */
 336
 337             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 338             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 339             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 340             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 341                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 342             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 343
 344             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velec            = _mm_and_ps(velec,cutoff_mask);
 348             velecsum         = _mm_add_ps(velecsum,velec);
 349             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 350             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 351
 352             fscal            = _mm_add_ps(felec,fvdw);
 353
 354             fscal            = _mm_and_ps(fscal,cutoff_mask);
 355
 356              /* Update vectorial force */
 357             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 358             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 359             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 360
 361             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 362             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 363             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 364
 365             }
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             if (gmx_mm_any_lt(rsq01,rcutoff2))
 372             {
 373
 374             r01              = _mm_mul_ps(rsq01,rinv01);
 375
 376             /* EWALD ELECTROSTATICS */
 377
 378             /* Analytical PME correction */
 379             zeta2            = _mm_mul_ps(beta2,rsq01);
 380             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 381             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 382             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 383             felec            = _mm_mul_ps(qq01,felec);
 384             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 385             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 386             velec            = _mm_mul_ps(qq01,velec);
 387
 388             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 389
 390             /* Update potential sum for this i atom from the interaction with this j atom. */
 391             velec            = _mm_and_ps(velec,cutoff_mask);
 392             velecsum         = _mm_add_ps(velecsum,velec);
 393
 394             fscal            = felec;
 395
 396             fscal            = _mm_and_ps(fscal,cutoff_mask);
 397
 398              /* Update vectorial force */
 399             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 400             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 401             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 402
 403             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 404             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 405             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 406
 407             }
 408
 409             /**************************
 410              * CALCULATE INTERACTIONS *
 411              **************************/
 412
 413             if (gmx_mm_any_lt(rsq02,rcutoff2))
 414             {
 415
 416             r02              = _mm_mul_ps(rsq02,rinv02);
 417
 418             /* EWALD ELECTROSTATICS */
 419
 420             /* Analytical PME correction */
 421             zeta2            = _mm_mul_ps(beta2,rsq02);
 422             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 423             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 424             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 425             felec            = _mm_mul_ps(qq02,felec);
 426             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 427             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 428             velec            = _mm_mul_ps(qq02,velec);
 429
 430             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velec            = _mm_and_ps(velec,cutoff_mask);
 434             velecsum         = _mm_add_ps(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             fscal            = _mm_and_ps(fscal,cutoff_mask);
 439
 440              /* Update vectorial force */
 441             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 442             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 443             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 444
 445             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 446             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 447             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 448
 449             }
 450
 451             /**************************
 452              * CALCULATE INTERACTIONS *
 453              **************************/
 454
 455             if (gmx_mm_any_lt(rsq10,rcutoff2))
 456             {
 457
 458             r10              = _mm_mul_ps(rsq10,rinv10);
 459
 460             /* EWALD ELECTROSTATICS */
 461
 462             /* Analytical PME correction */
 463             zeta2            = _mm_mul_ps(beta2,rsq10);
 464             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 465             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 466             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 467             felec            = _mm_mul_ps(qq10,felec);
 468             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 469             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 470             velec            = _mm_mul_ps(qq10,velec);
 471
 472             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 473
 474             /* Update potential sum for this i atom from the interaction with this j atom. */
 475             velec            = _mm_and_ps(velec,cutoff_mask);
 476             velecsum         = _mm_add_ps(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480             fscal            = _mm_and_ps(fscal,cutoff_mask);
 481
 482              /* Update vectorial force */
 483             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 484             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 485             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 486
 487             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 488             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 489             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 490
 491             }
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             if (gmx_mm_any_lt(rsq11,rcutoff2))
 498             {
 499
 500             r11              = _mm_mul_ps(rsq11,rinv11);
 501
 502             /* EWALD ELECTROSTATICS */
 503
 504             /* Analytical PME correction */
 505             zeta2            = _mm_mul_ps(beta2,rsq11);
 506             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 507             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 508             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 509             felec            = _mm_mul_ps(qq11,felec);
 510             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 511             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
 512             velec            = _mm_mul_ps(qq11,velec);
 513
 514             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 515
 516             /* Update potential sum for this i atom from the interaction with this j atom. */
 517             velec            = _mm_and_ps(velec,cutoff_mask);
 518             velecsum         = _mm_add_ps(velecsum,velec);
 519
 520             fscal            = felec;
 521
 522             fscal            = _mm_and_ps(fscal,cutoff_mask);
 523
 524              /* Update vectorial force */
 525             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 526             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 527             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 528
 529             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 530             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 531             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 532
 533             }
 534
 535             /**************************
 536              * CALCULATE INTERACTIONS *
 537              **************************/
 538
 539             if (gmx_mm_any_lt(rsq12,rcutoff2))
 540             {
 541
 542             r12              = _mm_mul_ps(rsq12,rinv12);
 543
 544             /* EWALD ELECTROSTATICS */
 545
 546             /* Analytical PME correction */
 547             zeta2            = _mm_mul_ps(beta2,rsq12);
 548             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 549             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 550             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 551             felec            = _mm_mul_ps(qq12,felec);
 552             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 553             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
 554             velec            = _mm_mul_ps(qq12,velec);
 555
 556             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 557
 558             /* Update potential sum for this i atom from the interaction with this j atom. */
 559             velec            = _mm_and_ps(velec,cutoff_mask);
 560             velecsum         = _mm_add_ps(velecsum,velec);
 561
 562             fscal            = felec;
 563
 564             fscal            = _mm_and_ps(fscal,cutoff_mask);
 565
 566              /* Update vectorial force */
 567             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 568             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 569             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 570
 571             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 572             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 573             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 574
 575             }
 576
 577             /**************************
 578              * CALCULATE INTERACTIONS *
 579              **************************/
 580
 581             if (gmx_mm_any_lt(rsq20,rcutoff2))
 582             {
 583
 584             r20              = _mm_mul_ps(rsq20,rinv20);
 585
 586             /* EWALD ELECTROSTATICS */
 587
 588             /* Analytical PME correction */
 589             zeta2            = _mm_mul_ps(beta2,rsq20);
 590             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 591             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 592             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 593             felec            = _mm_mul_ps(qq20,felec);
 594             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 595             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 596             velec            = _mm_mul_ps(qq20,velec);
 597
 598             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 599
 600             /* Update potential sum for this i atom from the interaction with this j atom. */
 601             velec            = _mm_and_ps(velec,cutoff_mask);
 602             velecsum         = _mm_add_ps(velecsum,velec);
 603
 604             fscal            = felec;
 605
 606             fscal            = _mm_and_ps(fscal,cutoff_mask);
 607
 608              /* Update vectorial force */
 609             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 610             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 611             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 612
 613             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 614             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 615             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 616
 617             }
 618
 619             /**************************
 620              * CALCULATE INTERACTIONS *
 621              **************************/
 622
 623             if (gmx_mm_any_lt(rsq21,rcutoff2))
 624             {
 625
 626             r21              = _mm_mul_ps(rsq21,rinv21);
 627
 628             /* EWALD ELECTROSTATICS */
 629
 630             /* Analytical PME correction */
 631             zeta2            = _mm_mul_ps(beta2,rsq21);
 632             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 633             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 634             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 635             felec            = _mm_mul_ps(qq21,felec);
 636             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 637             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
 638             velec            = _mm_mul_ps(qq21,velec);
 639
 640             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 641
 642             /* Update potential sum for this i atom from the interaction with this j atom. */
 643             velec            = _mm_and_ps(velec,cutoff_mask);
 644             velecsum         = _mm_add_ps(velecsum,velec);
 645
 646             fscal            = felec;
 647
 648             fscal            = _mm_and_ps(fscal,cutoff_mask);
 649
 650              /* Update vectorial force */
 651             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 652             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 653             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 654
 655             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 656             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 657             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 658
 659             }
 660
 661             /**************************
 662              * CALCULATE INTERACTIONS *
 663              **************************/
 664
 665             if (gmx_mm_any_lt(rsq22,rcutoff2))
 666             {
 667
 668             r22              = _mm_mul_ps(rsq22,rinv22);
 669
 670             /* EWALD ELECTROSTATICS */
 671
 672             /* Analytical PME correction */
 673             zeta2            = _mm_mul_ps(beta2,rsq22);
 674             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 675             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 676             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 677             felec            = _mm_mul_ps(qq22,felec);
 678             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 679             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
 680             velec            = _mm_mul_ps(qq22,velec);
 681
 682             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 683
 684             /* Update potential sum for this i atom from the interaction with this j atom. */
 685             velec            = _mm_and_ps(velec,cutoff_mask);
 686             velecsum         = _mm_add_ps(velecsum,velec);
 687
 688             fscal            = felec;
 689
 690             fscal            = _mm_and_ps(fscal,cutoff_mask);
 691
 692              /* Update vectorial force */
 693             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 694             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 695             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 696
 697             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 698             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 699             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 700
 701             }
 702
 703             fjptrA             = f+j_coord_offsetA;
 704             fjptrB             = f+j_coord_offsetB;
 705             fjptrC             = f+j_coord_offsetC;
 706             fjptrD             = f+j_coord_offsetD;
 707
 708             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 709                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 710
 711             /* Inner loop uses 315 flops */
 712         }
 713
 714         if(jidx<j_index_end)
 715         {
 716
 717             /* Get j neighbor index, and coordinate index */
 718             jnrlistA         = jjnr[jidx];
 719             jnrlistB         = jjnr[jidx+1];
 720             jnrlistC         = jjnr[jidx+2];
 721             jnrlistD         = jjnr[jidx+3];
 722             /* Sign of each element will be negative for non-real atoms.
 723              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 724              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 725              */
 726             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 727             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 728             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 729             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 730             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 731             j_coord_offsetA  = DIM*jnrA;
 732             j_coord_offsetB  = DIM*jnrB;
 733             j_coord_offsetC  = DIM*jnrC;
 734             j_coord_offsetD  = DIM*jnrD;
 735
 736             /* load j atom coordinates */
 737             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 738                                               x+j_coord_offsetC,x+j_coord_offsetD,
 739                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 740
 741             /* Calculate displacement vector */
 742             dx00             = _mm_sub_ps(ix0,jx0);
 743             dy00             = _mm_sub_ps(iy0,jy0);
 744             dz00             = _mm_sub_ps(iz0,jz0);
 745             dx01             = _mm_sub_ps(ix0,jx1);
 746             dy01             = _mm_sub_ps(iy0,jy1);
 747             dz01             = _mm_sub_ps(iz0,jz1);
 748             dx02             = _mm_sub_ps(ix0,jx2);
 749             dy02             = _mm_sub_ps(iy0,jy2);
 750             dz02             = _mm_sub_ps(iz0,jz2);
 751             dx10             = _mm_sub_ps(ix1,jx0);
 752             dy10             = _mm_sub_ps(iy1,jy0);
 753             dz10             = _mm_sub_ps(iz1,jz0);
 754             dx11             = _mm_sub_ps(ix1,jx1);
 755             dy11             = _mm_sub_ps(iy1,jy1);
 756             dz11             = _mm_sub_ps(iz1,jz1);
 757             dx12             = _mm_sub_ps(ix1,jx2);
 758             dy12             = _mm_sub_ps(iy1,jy2);
 759             dz12             = _mm_sub_ps(iz1,jz2);
 760             dx20             = _mm_sub_ps(ix2,jx0);
 761             dy20             = _mm_sub_ps(iy2,jy0);
 762             dz20             = _mm_sub_ps(iz2,jz0);
 763             dx21             = _mm_sub_ps(ix2,jx1);
 764             dy21             = _mm_sub_ps(iy2,jy1);
 765             dz21             = _mm_sub_ps(iz2,jz1);
 766             dx22             = _mm_sub_ps(ix2,jx2);
 767             dy22             = _mm_sub_ps(iy2,jy2);
 768             dz22             = _mm_sub_ps(iz2,jz2);
 769
 770             /* Calculate squared distance and things based on it */
 771             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 772             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 773             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 774             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 775             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 776             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 777             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 778             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 779             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 780
 781             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 782             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 783             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 784             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 785             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 786             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 787             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 788             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 789             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 790
 791             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 792             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 793             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 794             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 795             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 796             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 797             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 798             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 799             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 800
 801             fjx0             = _mm_setzero_ps();
 802             fjy0             = _mm_setzero_ps();
 803             fjz0             = _mm_setzero_ps();
 804             fjx1             = _mm_setzero_ps();
 805             fjy1             = _mm_setzero_ps();
 806             fjz1             = _mm_setzero_ps();
 807             fjx2             = _mm_setzero_ps();
 808             fjy2             = _mm_setzero_ps();
 809             fjz2             = _mm_setzero_ps();
 810
 811             /**************************
 812              * CALCULATE INTERACTIONS *
 813              **************************/
 814
 815             if (gmx_mm_any_lt(rsq00,rcutoff2))
 816             {
 817
 818             r00              = _mm_mul_ps(rsq00,rinv00);
 819             r00              = _mm_andnot_ps(dummy_mask,r00);
 820
 821             /* EWALD ELECTROSTATICS */
 822
 823             /* Analytical PME correction */
 824             zeta2            = _mm_mul_ps(beta2,rsq00);
 825             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 826             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 827             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 828             felec            = _mm_mul_ps(qq00,felec);
 829             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 830             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 831             velec            = _mm_mul_ps(qq00,velec);
 832
 833             /* LENNARD-JONES DISPERSION/REPULSION */
 834
 835             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 836             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 837             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 838             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 839                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 840             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 841
 842             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 843
 844             /* Update potential sum for this i atom from the interaction with this j atom. */
 845             velec            = _mm_and_ps(velec,cutoff_mask);
 846             velec            = _mm_andnot_ps(dummy_mask,velec);
 847             velecsum         = _mm_add_ps(velecsum,velec);
 848             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 849             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 850             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 851
 852             fscal            = _mm_add_ps(felec,fvdw);
 853
 854             fscal            = _mm_and_ps(fscal,cutoff_mask);
 855
 856             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 857
 858              /* Update vectorial force */
 859             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 860             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 861             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 862
 863             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 864             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 865             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 866
 867             }
 868
 869             /**************************
 870              * CALCULATE INTERACTIONS *
 871              **************************/
 872
 873             if (gmx_mm_any_lt(rsq01,rcutoff2))
 874             {
 875
 876             r01              = _mm_mul_ps(rsq01,rinv01);
 877             r01              = _mm_andnot_ps(dummy_mask,r01);
 878
 879             /* EWALD ELECTROSTATICS */
 880
 881             /* Analytical PME correction */
 882             zeta2            = _mm_mul_ps(beta2,rsq01);
 883             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 884             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 885             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 886             felec            = _mm_mul_ps(qq01,felec);
 887             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 888             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 889             velec            = _mm_mul_ps(qq01,velec);
 890
 891             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 892
 893             /* Update potential sum for this i atom from the interaction with this j atom. */
 894             velec            = _mm_and_ps(velec,cutoff_mask);
 895             velec            = _mm_andnot_ps(dummy_mask,velec);
 896             velecsum         = _mm_add_ps(velecsum,velec);
 897
 898             fscal            = felec;
 899
 900             fscal            = _mm_and_ps(fscal,cutoff_mask);
 901
 902             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 903
 904              /* Update vectorial force */
 905             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 906             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 907             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 908
 909             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 910             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 911             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 912
 913             }
 914
 915             /**************************
 916              * CALCULATE INTERACTIONS *
 917              **************************/
 918
 919             if (gmx_mm_any_lt(rsq02,rcutoff2))
 920             {
 921
 922             r02              = _mm_mul_ps(rsq02,rinv02);
 923             r02              = _mm_andnot_ps(dummy_mask,r02);
 924
 925             /* EWALD ELECTROSTATICS */
 926
 927             /* Analytical PME correction */
 928             zeta2            = _mm_mul_ps(beta2,rsq02);
 929             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 930             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 931             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 932             felec            = _mm_mul_ps(qq02,felec);
 933             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 934             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 935             velec            = _mm_mul_ps(qq02,velec);
 936
 937             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 938
 939             /* Update potential sum for this i atom from the interaction with this j atom. */
 940             velec            = _mm_and_ps(velec,cutoff_mask);
 941             velec            = _mm_andnot_ps(dummy_mask,velec);
 942             velecsum         = _mm_add_ps(velecsum,velec);
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_and_ps(fscal,cutoff_mask);
 947
 948             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 949
 950              /* Update vectorial force */
 951             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 952             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 953             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 954
 955             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 956             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 957             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 958
 959             }
 960
 961             /**************************
 962              * CALCULATE INTERACTIONS *
 963              **************************/
 964
 965             if (gmx_mm_any_lt(rsq10,rcutoff2))
 966             {
 967
 968             r10              = _mm_mul_ps(rsq10,rinv10);
 969             r10              = _mm_andnot_ps(dummy_mask,r10);
 970
 971             /* EWALD ELECTROSTATICS */
 972
 973             /* Analytical PME correction */
 974             zeta2            = _mm_mul_ps(beta2,rsq10);
 975             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 976             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 977             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 978             felec            = _mm_mul_ps(qq10,felec);
 979             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 980             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 981             velec            = _mm_mul_ps(qq10,velec);
 982
 983             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 984
 985             /* Update potential sum for this i atom from the interaction with this j atom. */
 986             velec            = _mm_and_ps(velec,cutoff_mask);
 987             velec            = _mm_andnot_ps(dummy_mask,velec);
 988             velecsum         = _mm_add_ps(velecsum,velec);
 989
 990             fscal            = felec;
 991
 992             fscal            = _mm_and_ps(fscal,cutoff_mask);
 993
 994             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 995
 996              /* Update vectorial force */
 997             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 998             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 999             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1000
1001             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1002             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1003             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1004
1005             }
1006
1007             /**************************
1008              * CALCULATE INTERACTIONS *
1009              **************************/
1010
1011             if (gmx_mm_any_lt(rsq11,rcutoff2))
1012             {
1013
1014             r11              = _mm_mul_ps(rsq11,rinv11);
1015             r11              = _mm_andnot_ps(dummy_mask,r11);
1016
1017             /* EWALD ELECTROSTATICS */
1018
1019             /* Analytical PME correction */
1020             zeta2            = _mm_mul_ps(beta2,rsq11);
1021             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1022             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1023             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1024             felec            = _mm_mul_ps(qq11,felec);
1025             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1026             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
1027             velec            = _mm_mul_ps(qq11,velec);
1028
1029             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1030
1031             /* Update potential sum for this i atom from the interaction with this j atom. */
1032             velec            = _mm_and_ps(velec,cutoff_mask);
1033             velec            = _mm_andnot_ps(dummy_mask,velec);
1034             velecsum         = _mm_add_ps(velecsum,velec);
1035
1036             fscal            = felec;
1037
1038             fscal            = _mm_and_ps(fscal,cutoff_mask);
1039
1040             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1041
1042              /* Update vectorial force */
1043             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1044             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1045             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1046
1047             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1048             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1049             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1050
1051             }
1052
1053             /**************************
1054              * CALCULATE INTERACTIONS *
1055              **************************/
1056
1057             if (gmx_mm_any_lt(rsq12,rcutoff2))
1058             {
1059
1060             r12              = _mm_mul_ps(rsq12,rinv12);
1061             r12              = _mm_andnot_ps(dummy_mask,r12);
1062
1063             /* EWALD ELECTROSTATICS */
1064
1065             /* Analytical PME correction */
1066             zeta2            = _mm_mul_ps(beta2,rsq12);
1067             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1068             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1069             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1070             felec            = _mm_mul_ps(qq12,felec);
1071             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1072             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
1073             velec            = _mm_mul_ps(qq12,velec);
1074
1075             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1076
1077             /* Update potential sum for this i atom from the interaction with this j atom. */
1078             velec            = _mm_and_ps(velec,cutoff_mask);
1079             velec            = _mm_andnot_ps(dummy_mask,velec);
1080             velecsum         = _mm_add_ps(velecsum,velec);
1081
1082             fscal            = felec;
1083
1084             fscal            = _mm_and_ps(fscal,cutoff_mask);
1085
1086             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1087
1088              /* Update vectorial force */
1089             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1090             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1091             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1092
1093             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1094             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1095             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1096
1097             }
1098
1099             /**************************
1100              * CALCULATE INTERACTIONS *
1101              **************************/
1102
1103             if (gmx_mm_any_lt(rsq20,rcutoff2))
1104             {
1105
1106             r20              = _mm_mul_ps(rsq20,rinv20);
1107             r20              = _mm_andnot_ps(dummy_mask,r20);
1108
1109             /* EWALD ELECTROSTATICS */
1110
1111             /* Analytical PME correction */
1112             zeta2            = _mm_mul_ps(beta2,rsq20);
1113             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1114             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1115             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1116             felec            = _mm_mul_ps(qq20,felec);
1117             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1118             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
1119             velec            = _mm_mul_ps(qq20,velec);
1120
1121             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1122
1123             /* Update potential sum for this i atom from the interaction with this j atom. */
1124             velec            = _mm_and_ps(velec,cutoff_mask);
1125             velec            = _mm_andnot_ps(dummy_mask,velec);
1126             velecsum         = _mm_add_ps(velecsum,velec);
1127
1128             fscal            = felec;
1129
1130             fscal            = _mm_and_ps(fscal,cutoff_mask);
1131
1132             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1133
1134              /* Update vectorial force */
1135             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1136             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1137             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1138
1139             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1140             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1141             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1142
1143             }
1144
1145             /**************************
1146              * CALCULATE INTERACTIONS *
1147              **************************/
1148
1149             if (gmx_mm_any_lt(rsq21,rcutoff2))
1150             {
1151
1152             r21              = _mm_mul_ps(rsq21,rinv21);
1153             r21              = _mm_andnot_ps(dummy_mask,r21);
1154
1155             /* EWALD ELECTROSTATICS */
1156
1157             /* Analytical PME correction */
1158             zeta2            = _mm_mul_ps(beta2,rsq21);
1159             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1160             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1161             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1162             felec            = _mm_mul_ps(qq21,felec);
1163             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1164             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
1165             velec            = _mm_mul_ps(qq21,velec);
1166
1167             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1168
1169             /* Update potential sum for this i atom from the interaction with this j atom. */
1170             velec            = _mm_and_ps(velec,cutoff_mask);
1171             velec            = _mm_andnot_ps(dummy_mask,velec);
1172             velecsum         = _mm_add_ps(velecsum,velec);
1173
1174             fscal            = felec;
1175
1176             fscal            = _mm_and_ps(fscal,cutoff_mask);
1177
1178             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1179
1180              /* Update vectorial force */
1181             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1182             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1183             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1184
1185             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1186             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1187             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1188
1189             }
1190
1191             /**************************
1192              * CALCULATE INTERACTIONS *
1193              **************************/
1194
1195             if (gmx_mm_any_lt(rsq22,rcutoff2))
1196             {
1197
1198             r22              = _mm_mul_ps(rsq22,rinv22);
1199             r22              = _mm_andnot_ps(dummy_mask,r22);
1200
1201             /* EWALD ELECTROSTATICS */
1202
1203             /* Analytical PME correction */
1204             zeta2            = _mm_mul_ps(beta2,rsq22);
1205             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1206             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1207             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1208             felec            = _mm_mul_ps(qq22,felec);
1209             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1210             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
1211             velec            = _mm_mul_ps(qq22,velec);
1212
1213             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1214
1215             /* Update potential sum for this i atom from the interaction with this j atom. */
1216             velec            = _mm_and_ps(velec,cutoff_mask);
1217             velec            = _mm_andnot_ps(dummy_mask,velec);
1218             velecsum         = _mm_add_ps(velecsum,velec);
1219
1220             fscal            = felec;
1221
1222             fscal            = _mm_and_ps(fscal,cutoff_mask);
1223
1224             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1225
1226              /* Update vectorial force */
1227             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1228             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1229             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1230
1231             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1232             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1233             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1234
1235             }
1236
1237             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1238             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1239             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1240             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1241
1242             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1243                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1244
1245             /* Inner loop uses 324 flops */
1246         }
1247
1248         /* End of innermost loop */
1249
1250         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1251                                               f+i_coord_offset,fshift+i_shift_offset);
1252
1253         ggid                        = gid[iidx];
1254         /* Update potential energies */
1255         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1256         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1257
1258         /* Increment number of inner iterations */
1259         inneriter                  += j_index_end - j_index_start;
1260
1261         /* Outer loop uses 20 flops */
1262     }
1263
1264     /* Increment number of outer iterations */
1265     outeriter        += nri;
1266
1267     /* Update outer/inner flops */
1268
1269     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*324);
1270 }
1271 /*
1272  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_F_avx_128_fma_single
1273  * Electrostatics interaction: Ewald
1274  * VdW interaction:            LennardJones
1275  * Geometry:                   Water3-Water3
1276  * Calculate force/pot:        Force
1277  */
1278 void
1279 nb_kernel_ElecEwSh_VdwLJSh_GeomW3W3_F_avx_128_fma_single
1280                     (t_nblist                    * gmx_restrict       nlist,
1281                      rvec                        * gmx_restrict          xx,
1282                      rvec                        * gmx_restrict          ff,
1283                      t_forcerec                  * gmx_restrict          fr,
1284                      t_mdatoms                   * gmx_restrict     mdatoms,
1285                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1286                      t_nrnb                      * gmx_restrict        nrnb)
1287 {
1288     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1289      * just 0 for non-waters.
1290      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1291      * jnr indices corresponding to data put in the four positions in the SIMD register.
1292      */
1293     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1294     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1295     int              jnrA,jnrB,jnrC,jnrD;
1296     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1297     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1298     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1299     real             rcutoff_scalar;
1300     real             *shiftvec,*fshift,*x,*f;
1301     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1302     real             scratch[4*DIM];
1303     __m128           fscal,rcutoff,rcutoff2,jidxall;
1304     int              vdwioffset0;
1305     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1306     int              vdwioffset1;
1307     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1308     int              vdwioffset2;
1309     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1310     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1311     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1312     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1313     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1314     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1315     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1316     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1317     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1318     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1319     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1320     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1321     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1322     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1323     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1324     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1325     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1326     real             *charge;
1327     int              nvdwtype;
1328     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1329     int              *vdwtype;
1330     real             *vdwparam;
1331     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1332     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1333     __m128i          ewitab;
1334     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1335     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1336     real             *ewtab;
1337     __m128           dummy_mask,cutoff_mask;
1338     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1339     __m128           one     = _mm_set1_ps(1.0);
1340     __m128           two     = _mm_set1_ps(2.0);
1341     x                = xx[0];
1342     f                = ff[0];
1343
1344     nri              = nlist->nri;
1345     iinr             = nlist->iinr;
1346     jindex           = nlist->jindex;
1347     jjnr             = nlist->jjnr;
1348     shiftidx         = nlist->shift;
1349     gid              = nlist->gid;
1350     shiftvec         = fr->shift_vec[0];
1351     fshift           = fr->fshift[0];
1352     facel            = _mm_set1_ps(fr->epsfac);
1353     charge           = mdatoms->chargeA;
1354     nvdwtype         = fr->ntype;
1355     vdwparam         = fr->nbfp;
1356     vdwtype          = mdatoms->typeA;
1357
1358     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1359     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
1360     beta2            = _mm_mul_ps(beta,beta);
1361     beta3            = _mm_mul_ps(beta,beta2);
1362     ewtab            = fr->ic->tabq_coul_F;
1363     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1364     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1365
1366     /* Setup water-specific parameters */
1367     inr              = nlist->iinr[0];
1368     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1369     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1370     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1371     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1372
1373     jq0              = _mm_set1_ps(charge[inr+0]);
1374     jq1              = _mm_set1_ps(charge[inr+1]);
1375     jq2              = _mm_set1_ps(charge[inr+2]);
1376     vdwjidx0A        = 2*vdwtype[inr+0];
1377     qq00             = _mm_mul_ps(iq0,jq0);
1378     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1379     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1380     qq01             = _mm_mul_ps(iq0,jq1);
1381     qq02             = _mm_mul_ps(iq0,jq2);
1382     qq10             = _mm_mul_ps(iq1,jq0);
1383     qq11             = _mm_mul_ps(iq1,jq1);
1384     qq12             = _mm_mul_ps(iq1,jq2);
1385     qq20             = _mm_mul_ps(iq2,jq0);
1386     qq21             = _mm_mul_ps(iq2,jq1);
1387     qq22             = _mm_mul_ps(iq2,jq2);
1388
1389     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1390     rcutoff_scalar   = fr->rcoulomb;
1391     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1392     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1393
1394     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1395     rvdw             = _mm_set1_ps(fr->rvdw);
1396
1397     /* Avoid stupid compiler warnings */
1398     jnrA = jnrB = jnrC = jnrD = 0;
1399     j_coord_offsetA = 0;
1400     j_coord_offsetB = 0;
1401     j_coord_offsetC = 0;
1402     j_coord_offsetD = 0;
1403
1404     outeriter        = 0;
1405     inneriter        = 0;
1406
1407     for(iidx=0;iidx<4*DIM;iidx++)
1408     {
1409         scratch[iidx] = 0.0;
1410     }
1411
1412     /* Start outer loop over neighborlists */
1413     for(iidx=0; iidx<nri; iidx++)
1414     {
1415         /* Load shift vector for this list */
1416         i_shift_offset   = DIM*shiftidx[iidx];
1417
1418         /* Load limits for loop over neighbors */
1419         j_index_start    = jindex[iidx];
1420         j_index_end      = jindex[iidx+1];
1421
1422         /* Get outer coordinate index */
1423         inr              = iinr[iidx];
1424         i_coord_offset   = DIM*inr;
1425
1426         /* Load i particle coords and add shift vector */
1427         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1428                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1429
1430         fix0             = _mm_setzero_ps();
1431         fiy0             = _mm_setzero_ps();
1432         fiz0             = _mm_setzero_ps();
1433         fix1             = _mm_setzero_ps();
1434         fiy1             = _mm_setzero_ps();
1435         fiz1             = _mm_setzero_ps();
1436         fix2             = _mm_setzero_ps();
1437         fiy2             = _mm_setzero_ps();
1438         fiz2             = _mm_setzero_ps();
1439
1440         /* Start inner kernel loop */
1441         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1442         {
1443
1444             /* Get j neighbor index, and coordinate index */
1445             jnrA             = jjnr[jidx];
1446             jnrB             = jjnr[jidx+1];
1447             jnrC             = jjnr[jidx+2];
1448             jnrD             = jjnr[jidx+3];
1449             j_coord_offsetA  = DIM*jnrA;
1450             j_coord_offsetB  = DIM*jnrB;
1451             j_coord_offsetC  = DIM*jnrC;
1452             j_coord_offsetD  = DIM*jnrD;
1453
1454             /* load j atom coordinates */
1455             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1456                                               x+j_coord_offsetC,x+j_coord_offsetD,
1457                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1458
1459             /* Calculate displacement vector */
1460             dx00             = _mm_sub_ps(ix0,jx0);
1461             dy00             = _mm_sub_ps(iy0,jy0);
1462             dz00             = _mm_sub_ps(iz0,jz0);
1463             dx01             = _mm_sub_ps(ix0,jx1);
1464             dy01             = _mm_sub_ps(iy0,jy1);
1465             dz01             = _mm_sub_ps(iz0,jz1);
1466             dx02             = _mm_sub_ps(ix0,jx2);
1467             dy02             = _mm_sub_ps(iy0,jy2);
1468             dz02             = _mm_sub_ps(iz0,jz2);
1469             dx10             = _mm_sub_ps(ix1,jx0);
1470             dy10             = _mm_sub_ps(iy1,jy0);
1471             dz10             = _mm_sub_ps(iz1,jz0);
1472             dx11             = _mm_sub_ps(ix1,jx1);
1473             dy11             = _mm_sub_ps(iy1,jy1);
1474             dz11             = _mm_sub_ps(iz1,jz1);
1475             dx12             = _mm_sub_ps(ix1,jx2);
1476             dy12             = _mm_sub_ps(iy1,jy2);
1477             dz12             = _mm_sub_ps(iz1,jz2);
1478             dx20             = _mm_sub_ps(ix2,jx0);
1479             dy20             = _mm_sub_ps(iy2,jy0);
1480             dz20             = _mm_sub_ps(iz2,jz0);
1481             dx21             = _mm_sub_ps(ix2,jx1);
1482             dy21             = _mm_sub_ps(iy2,jy1);
1483             dz21             = _mm_sub_ps(iz2,jz1);
1484             dx22             = _mm_sub_ps(ix2,jx2);
1485             dy22             = _mm_sub_ps(iy2,jy2);
1486             dz22             = _mm_sub_ps(iz2,jz2);
1487
1488             /* Calculate squared distance and things based on it */
1489             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1490             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1491             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1492             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1493             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1494             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1495             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1496             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1497             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1498
1499             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1500             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1501             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1502             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1503             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1504             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1505             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1506             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1507             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1508
1509             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1510             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1511             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1512             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1513             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1514             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1515             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1516             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1517             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1518
1519             fjx0             = _mm_setzero_ps();
1520             fjy0             = _mm_setzero_ps();
1521             fjz0             = _mm_setzero_ps();
1522             fjx1             = _mm_setzero_ps();
1523             fjy1             = _mm_setzero_ps();
1524             fjz1             = _mm_setzero_ps();
1525             fjx2             = _mm_setzero_ps();
1526             fjy2             = _mm_setzero_ps();
1527             fjz2             = _mm_setzero_ps();
1528
1529             /**************************
1530              * CALCULATE INTERACTIONS *
1531              **************************/
1532
1533             if (gmx_mm_any_lt(rsq00,rcutoff2))
1534             {
1535
1536             r00              = _mm_mul_ps(rsq00,rinv00);
1537
1538             /* EWALD ELECTROSTATICS */
1539
1540             /* Analytical PME correction */
1541             zeta2            = _mm_mul_ps(beta2,rsq00);
1542             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1543             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1544             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1545             felec            = _mm_mul_ps(qq00,felec);
1546
1547             /* LENNARD-JONES DISPERSION/REPULSION */
1548
1549             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1550             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1551
1552             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1553
1554             fscal            = _mm_add_ps(felec,fvdw);
1555
1556             fscal            = _mm_and_ps(fscal,cutoff_mask);
1557
1558              /* Update vectorial force */
1559             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1560             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1561             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1562
1563             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1564             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1565             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1566
1567             }
1568
1569             /**************************
1570              * CALCULATE INTERACTIONS *
1571              **************************/
1572
1573             if (gmx_mm_any_lt(rsq01,rcutoff2))
1574             {
1575
1576             r01              = _mm_mul_ps(rsq01,rinv01);
1577
1578             /* EWALD ELECTROSTATICS */
1579
1580             /* Analytical PME correction */
1581             zeta2            = _mm_mul_ps(beta2,rsq01);
1582             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1583             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1584             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1585             felec            = _mm_mul_ps(qq01,felec);
1586
1587             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1588
1589             fscal            = felec;
1590
1591             fscal            = _mm_and_ps(fscal,cutoff_mask);
1592
1593              /* Update vectorial force */
1594             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1595             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1596             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1597
1598             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1599             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1600             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1601
1602             }
1603
1604             /**************************
1605              * CALCULATE INTERACTIONS *
1606              **************************/
1607
1608             if (gmx_mm_any_lt(rsq02,rcutoff2))
1609             {
1610
1611             r02              = _mm_mul_ps(rsq02,rinv02);
1612
1613             /* EWALD ELECTROSTATICS */
1614
1615             /* Analytical PME correction */
1616             zeta2            = _mm_mul_ps(beta2,rsq02);
1617             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1618             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1619             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1620             felec            = _mm_mul_ps(qq02,felec);
1621
1622             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1623
1624             fscal            = felec;
1625
1626             fscal            = _mm_and_ps(fscal,cutoff_mask);
1627
1628              /* Update vectorial force */
1629             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1630             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1631             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1632
1633             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1634             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1635             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1636
1637             }
1638
1639             /**************************
1640              * CALCULATE INTERACTIONS *
1641              **************************/
1642
1643             if (gmx_mm_any_lt(rsq10,rcutoff2))
1644             {
1645
1646             r10              = _mm_mul_ps(rsq10,rinv10);
1647
1648             /* EWALD ELECTROSTATICS */
1649
1650             /* Analytical PME correction */
1651             zeta2            = _mm_mul_ps(beta2,rsq10);
1652             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1653             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1654             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1655             felec            = _mm_mul_ps(qq10,felec);
1656
1657             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1658
1659             fscal            = felec;
1660
1661             fscal            = _mm_and_ps(fscal,cutoff_mask);
1662
1663              /* Update vectorial force */
1664             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1665             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1666             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1667
1668             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1669             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1670             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1671
1672             }
1673
1674             /**************************
1675              * CALCULATE INTERACTIONS *
1676              **************************/
1677
1678             if (gmx_mm_any_lt(rsq11,rcutoff2))
1679             {
1680
1681             r11              = _mm_mul_ps(rsq11,rinv11);
1682
1683             /* EWALD ELECTROSTATICS */
1684
1685             /* Analytical PME correction */
1686             zeta2            = _mm_mul_ps(beta2,rsq11);
1687             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1688             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1689             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1690             felec            = _mm_mul_ps(qq11,felec);
1691
1692             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1693
1694             fscal            = felec;
1695
1696             fscal            = _mm_and_ps(fscal,cutoff_mask);
1697
1698              /* Update vectorial force */
1699             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1700             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1701             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1702
1703             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1704             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1705             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1706
1707             }
1708
1709             /**************************
1710              * CALCULATE INTERACTIONS *
1711              **************************/
1712
1713             if (gmx_mm_any_lt(rsq12,rcutoff2))
1714             {
1715
1716             r12              = _mm_mul_ps(rsq12,rinv12);
1717
1718             /* EWALD ELECTROSTATICS */
1719
1720             /* Analytical PME correction */
1721             zeta2            = _mm_mul_ps(beta2,rsq12);
1722             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1723             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1724             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1725             felec            = _mm_mul_ps(qq12,felec);
1726
1727             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1728
1729             fscal            = felec;
1730
1731             fscal            = _mm_and_ps(fscal,cutoff_mask);
1732
1733              /* Update vectorial force */
1734             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1735             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1736             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1737
1738             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1739             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1740             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1741
1742             }
1743
1744             /**************************
1745              * CALCULATE INTERACTIONS *
1746              **************************/
1747
1748             if (gmx_mm_any_lt(rsq20,rcutoff2))
1749             {
1750
1751             r20              = _mm_mul_ps(rsq20,rinv20);
1752
1753             /* EWALD ELECTROSTATICS */
1754
1755             /* Analytical PME correction */
1756             zeta2            = _mm_mul_ps(beta2,rsq20);
1757             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1758             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1759             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1760             felec            = _mm_mul_ps(qq20,felec);
1761
1762             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1763
1764             fscal            = felec;
1765
1766             fscal            = _mm_and_ps(fscal,cutoff_mask);
1767
1768              /* Update vectorial force */
1769             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1770             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1771             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1772
1773             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1774             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1775             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1776
1777             }
1778
1779             /**************************
1780              * CALCULATE INTERACTIONS *
1781              **************************/
1782
1783             if (gmx_mm_any_lt(rsq21,rcutoff2))
1784             {
1785
1786             r21              = _mm_mul_ps(rsq21,rinv21);
1787
1788             /* EWALD ELECTROSTATICS */
1789
1790             /* Analytical PME correction */
1791             zeta2            = _mm_mul_ps(beta2,rsq21);
1792             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1793             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1794             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1795             felec            = _mm_mul_ps(qq21,felec);
1796
1797             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1798
1799             fscal            = felec;
1800
1801             fscal            = _mm_and_ps(fscal,cutoff_mask);
1802
1803              /* Update vectorial force */
1804             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1805             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1806             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1807
1808             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1809             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1810             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1811
1812             }
1813
1814             /**************************
1815              * CALCULATE INTERACTIONS *
1816              **************************/
1817
1818             if (gmx_mm_any_lt(rsq22,rcutoff2))
1819             {
1820
1821             r22              = _mm_mul_ps(rsq22,rinv22);
1822
1823             /* EWALD ELECTROSTATICS */
1824
1825             /* Analytical PME correction */
1826             zeta2            = _mm_mul_ps(beta2,rsq22);
1827             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1828             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1829             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1830             felec            = _mm_mul_ps(qq22,felec);
1831
1832             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1833
1834             fscal            = felec;
1835
1836             fscal            = _mm_and_ps(fscal,cutoff_mask);
1837
1838              /* Update vectorial force */
1839             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1840             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1841             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1842
1843             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1844             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1845             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1846
1847             }
1848
1849             fjptrA             = f+j_coord_offsetA;
1850             fjptrB             = f+j_coord_offsetB;
1851             fjptrC             = f+j_coord_offsetC;
1852             fjptrD             = f+j_coord_offsetD;
1853
1854             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1855                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1856
1857             /* Inner loop uses 286 flops */
1858         }
1859
1860         if(jidx<j_index_end)
1861         {
1862
1863             /* Get j neighbor index, and coordinate index */
1864             jnrlistA         = jjnr[jidx];
1865             jnrlistB         = jjnr[jidx+1];
1866             jnrlistC         = jjnr[jidx+2];
1867             jnrlistD         = jjnr[jidx+3];
1868             /* Sign of each element will be negative for non-real atoms.
1869              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1870              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1871              */
1872             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1873             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1874             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1875             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1876             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1877             j_coord_offsetA  = DIM*jnrA;
1878             j_coord_offsetB  = DIM*jnrB;
1879             j_coord_offsetC  = DIM*jnrC;
1880             j_coord_offsetD  = DIM*jnrD;
1881
1882             /* load j atom coordinates */
1883             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1884                                               x+j_coord_offsetC,x+j_coord_offsetD,
1885                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1886
1887             /* Calculate displacement vector */
1888             dx00             = _mm_sub_ps(ix0,jx0);
1889             dy00             = _mm_sub_ps(iy0,jy0);
1890             dz00             = _mm_sub_ps(iz0,jz0);
1891             dx01             = _mm_sub_ps(ix0,jx1);
1892             dy01             = _mm_sub_ps(iy0,jy1);
1893             dz01             = _mm_sub_ps(iz0,jz1);
1894             dx02             = _mm_sub_ps(ix0,jx2);
1895             dy02             = _mm_sub_ps(iy0,jy2);
1896             dz02             = _mm_sub_ps(iz0,jz2);
1897             dx10             = _mm_sub_ps(ix1,jx0);
1898             dy10             = _mm_sub_ps(iy1,jy0);
1899             dz10             = _mm_sub_ps(iz1,jz0);
1900             dx11             = _mm_sub_ps(ix1,jx1);
1901             dy11             = _mm_sub_ps(iy1,jy1);
1902             dz11             = _mm_sub_ps(iz1,jz1);
1903             dx12             = _mm_sub_ps(ix1,jx2);
1904             dy12             = _mm_sub_ps(iy1,jy2);
1905             dz12             = _mm_sub_ps(iz1,jz2);
1906             dx20             = _mm_sub_ps(ix2,jx0);
1907             dy20             = _mm_sub_ps(iy2,jy0);
1908             dz20             = _mm_sub_ps(iz2,jz0);
1909             dx21             = _mm_sub_ps(ix2,jx1);
1910             dy21             = _mm_sub_ps(iy2,jy1);
1911             dz21             = _mm_sub_ps(iz2,jz1);
1912             dx22             = _mm_sub_ps(ix2,jx2);
1913             dy22             = _mm_sub_ps(iy2,jy2);
1914             dz22             = _mm_sub_ps(iz2,jz2);
1915
1916             /* Calculate squared distance and things based on it */
1917             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1918             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1919             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1920             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1921             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1922             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1923             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1924             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1925             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1926
1927             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1928             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1929             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1930             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1931             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1932             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1933             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1934             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1935             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1936
1937             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1938             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1939             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1940             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1941             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1942             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1943             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1944             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1945             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1946
1947             fjx0             = _mm_setzero_ps();
1948             fjy0             = _mm_setzero_ps();
1949             fjz0             = _mm_setzero_ps();
1950             fjx1             = _mm_setzero_ps();
1951             fjy1             = _mm_setzero_ps();
1952             fjz1             = _mm_setzero_ps();
1953             fjx2             = _mm_setzero_ps();
1954             fjy2             = _mm_setzero_ps();
1955             fjz2             = _mm_setzero_ps();
1956
1957             /**************************
1958              * CALCULATE INTERACTIONS *
1959              **************************/
1960
1961             if (gmx_mm_any_lt(rsq00,rcutoff2))
1962             {
1963
1964             r00              = _mm_mul_ps(rsq00,rinv00);
1965             r00              = _mm_andnot_ps(dummy_mask,r00);
1966
1967             /* EWALD ELECTROSTATICS */
1968
1969             /* Analytical PME correction */
1970             zeta2            = _mm_mul_ps(beta2,rsq00);
1971             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1972             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1973             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1974             felec            = _mm_mul_ps(qq00,felec);
1975
1976             /* LENNARD-JONES DISPERSION/REPULSION */
1977
1978             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1979             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1980
1981             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1982
1983             fscal            = _mm_add_ps(felec,fvdw);
1984
1985             fscal            = _mm_and_ps(fscal,cutoff_mask);
1986
1987             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1988
1989              /* Update vectorial force */
1990             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1991             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1992             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1993
1994             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1995             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1996             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1997
1998             }
1999
2000             /**************************
2001              * CALCULATE INTERACTIONS *
2002              **************************/
2003
2004             if (gmx_mm_any_lt(rsq01,rcutoff2))
2005             {
2006
2007             r01              = _mm_mul_ps(rsq01,rinv01);
2008             r01              = _mm_andnot_ps(dummy_mask,r01);
2009
2010             /* EWALD ELECTROSTATICS */
2011
2012             /* Analytical PME correction */
2013             zeta2            = _mm_mul_ps(beta2,rsq01);
2014             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
2015             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2016             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2017             felec            = _mm_mul_ps(qq01,felec);
2018
2019             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
2020
2021             fscal            = felec;
2022
2023             fscal            = _mm_and_ps(fscal,cutoff_mask);
2024
2025             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2026
2027              /* Update vectorial force */
2028             fix0             = _mm_macc_ps(dx01,fscal,fix0);
2029             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
2030             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
2031
2032             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
2033             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
2034             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
2035
2036             }
2037
2038             /**************************
2039              * CALCULATE INTERACTIONS *
2040              **************************/
2041
2042             if (gmx_mm_any_lt(rsq02,rcutoff2))
2043             {
2044
2045             r02              = _mm_mul_ps(rsq02,rinv02);
2046             r02              = _mm_andnot_ps(dummy_mask,r02);
2047
2048             /* EWALD ELECTROSTATICS */
2049
2050             /* Analytical PME correction */
2051             zeta2            = _mm_mul_ps(beta2,rsq02);
2052             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
2053             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2054             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2055             felec            = _mm_mul_ps(qq02,felec);
2056
2057             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
2058
2059             fscal            = felec;
2060
2061             fscal            = _mm_and_ps(fscal,cutoff_mask);
2062
2063             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2064
2065              /* Update vectorial force */
2066             fix0             = _mm_macc_ps(dx02,fscal,fix0);
2067             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
2068             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
2069
2070             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
2071             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
2072             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
2073
2074             }
2075
2076             /**************************
2077              * CALCULATE INTERACTIONS *
2078              **************************/
2079
2080             if (gmx_mm_any_lt(rsq10,rcutoff2))
2081             {
2082
2083             r10              = _mm_mul_ps(rsq10,rinv10);
2084             r10              = _mm_andnot_ps(dummy_mask,r10);
2085
2086             /* EWALD ELECTROSTATICS */
2087
2088             /* Analytical PME correction */
2089             zeta2            = _mm_mul_ps(beta2,rsq10);
2090             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
2091             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2092             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2093             felec            = _mm_mul_ps(qq10,felec);
2094
2095             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
2096
2097             fscal            = felec;
2098
2099             fscal            = _mm_and_ps(fscal,cutoff_mask);
2100
2101             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2102
2103              /* Update vectorial force */
2104             fix1             = _mm_macc_ps(dx10,fscal,fix1);
2105             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
2106             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
2107
2108             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
2109             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
2110             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
2111
2112             }
2113
2114             /**************************
2115              * CALCULATE INTERACTIONS *
2116              **************************/
2117
2118             if (gmx_mm_any_lt(rsq11,rcutoff2))
2119             {
2120
2121             r11              = _mm_mul_ps(rsq11,rinv11);
2122             r11              = _mm_andnot_ps(dummy_mask,r11);
2123
2124             /* EWALD ELECTROSTATICS */
2125
2126             /* Analytical PME correction */
2127             zeta2            = _mm_mul_ps(beta2,rsq11);
2128             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
2129             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2130             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2131             felec            = _mm_mul_ps(qq11,felec);
2132
2133             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2134
2135             fscal            = felec;
2136
2137             fscal            = _mm_and_ps(fscal,cutoff_mask);
2138
2139             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2140
2141              /* Update vectorial force */
2142             fix1             = _mm_macc_ps(dx11,fscal,fix1);
2143             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
2144             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
2145
2146             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
2147             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
2148             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
2149
2150             }
2151
2152             /**************************
2153              * CALCULATE INTERACTIONS *
2154              **************************/
2155
2156             if (gmx_mm_any_lt(rsq12,rcutoff2))
2157             {
2158
2159             r12              = _mm_mul_ps(rsq12,rinv12);
2160             r12              = _mm_andnot_ps(dummy_mask,r12);
2161
2162             /* EWALD ELECTROSTATICS */
2163
2164             /* Analytical PME correction */
2165             zeta2            = _mm_mul_ps(beta2,rsq12);
2166             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
2167             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2168             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2169             felec            = _mm_mul_ps(qq12,felec);
2170
2171             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2172
2173             fscal            = felec;
2174
2175             fscal            = _mm_and_ps(fscal,cutoff_mask);
2176
2177             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2178
2179              /* Update vectorial force */
2180             fix1             = _mm_macc_ps(dx12,fscal,fix1);
2181             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
2182             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
2183
2184             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
2185             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
2186             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
2187
2188             }
2189
2190             /**************************
2191              * CALCULATE INTERACTIONS *
2192              **************************/
2193
2194             if (gmx_mm_any_lt(rsq20,rcutoff2))
2195             {
2196
2197             r20              = _mm_mul_ps(rsq20,rinv20);
2198             r20              = _mm_andnot_ps(dummy_mask,r20);
2199
2200             /* EWALD ELECTROSTATICS */
2201
2202             /* Analytical PME correction */
2203             zeta2            = _mm_mul_ps(beta2,rsq20);
2204             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
2205             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2206             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2207             felec            = _mm_mul_ps(qq20,felec);
2208
2209             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2210
2211             fscal            = felec;
2212
2213             fscal            = _mm_and_ps(fscal,cutoff_mask);
2214
2215             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2216
2217              /* Update vectorial force */
2218             fix2             = _mm_macc_ps(dx20,fscal,fix2);
2219             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
2220             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
2221
2222             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
2223             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
2224             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
2225
2226             }
2227
2228             /**************************
2229              * CALCULATE INTERACTIONS *
2230              **************************/
2231
2232             if (gmx_mm_any_lt(rsq21,rcutoff2))
2233             {
2234
2235             r21              = _mm_mul_ps(rsq21,rinv21);
2236             r21              = _mm_andnot_ps(dummy_mask,r21);
2237
2238             /* EWALD ELECTROSTATICS */
2239
2240             /* Analytical PME correction */
2241             zeta2            = _mm_mul_ps(beta2,rsq21);
2242             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
2243             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2244             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2245             felec            = _mm_mul_ps(qq21,felec);
2246
2247             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2248
2249             fscal            = felec;
2250
2251             fscal            = _mm_and_ps(fscal,cutoff_mask);
2252
2253             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2254
2255              /* Update vectorial force */
2256             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2257             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2258             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2259
2260             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2261             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2262             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2263
2264             }
2265
2266             /**************************
2267              * CALCULATE INTERACTIONS *
2268              **************************/
2269
2270             if (gmx_mm_any_lt(rsq22,rcutoff2))
2271             {
2272
2273             r22              = _mm_mul_ps(rsq22,rinv22);
2274             r22              = _mm_andnot_ps(dummy_mask,r22);
2275
2276             /* EWALD ELECTROSTATICS */
2277
2278             /* Analytical PME correction */
2279             zeta2            = _mm_mul_ps(beta2,rsq22);
2280             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
2281             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2282             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2283             felec            = _mm_mul_ps(qq22,felec);
2284
2285             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2286
2287             fscal            = felec;
2288
2289             fscal            = _mm_and_ps(fscal,cutoff_mask);
2290
2291             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2292
2293              /* Update vectorial force */
2294             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2295             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2296             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2297
2298             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2299             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2300             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2301
2302             }
2303
2304             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2305             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2306             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2307             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2308
2309             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2310                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2311
2312             /* Inner loop uses 295 flops */
2313         }
2314
2315         /* End of innermost loop */
2316
2317         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2318                                               f+i_coord_offset,fshift+i_shift_offset);
2319
2320         /* Increment number of inner iterations */
2321         inneriter                  += j_index_end - j_index_start;
2322
2323         /* Outer loop uses 18 flops */
2324     }
2325
2326     /* Increment number of outer iterations */
2327     outeriter        += nri;
2328
2329     /* Update outer/inner flops */
2330
2331     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*295);
2332 }