src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_single
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  97     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  98     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     int              nvdwtype;
 107     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 108     int              *vdwtype;
 109     real             *vdwparam;
 110     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 111     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 112     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 113     real             rswitch_scalar,d_scalar;
 114     __m128           dummy_mask,cutoff_mask;
 115     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 116     __m128           one     = _mm_set1_ps(1.0);
 117     __m128           two     = _mm_set1_ps(2.0);
 118     x                = xx[0];
 119     f                = ff[0];
 120
 121     nri              = nlist->nri;
 122     iinr             = nlist->iinr;
 123     jindex           = nlist->jindex;
 124     jjnr             = nlist->jjnr;
 125     shiftidx         = nlist->shift;
 126     gid              = nlist->gid;
 127     shiftvec         = fr->shift_vec[0];
 128     fshift           = fr->fshift[0];
 129     facel            = _mm_set1_ps(fr->epsfac);
 130     charge           = mdatoms->chargeA;
 131     krf              = _mm_set1_ps(fr->ic->k_rf);
 132     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 133     crf              = _mm_set1_ps(fr->ic->c_rf);
 134     nvdwtype         = fr->ntype;
 135     vdwparam         = fr->nbfp;
 136     vdwtype          = mdatoms->typeA;
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 141     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 142     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     jq0              = _mm_set1_ps(charge[inr+0]);
 146     jq1              = _mm_set1_ps(charge[inr+1]);
 147     jq2              = _mm_set1_ps(charge[inr+2]);
 148     vdwjidx0A        = 2*vdwtype[inr+0];
 149     qq00             = _mm_mul_ps(iq0,jq0);
 150     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 151     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 152     qq01             = _mm_mul_ps(iq0,jq1);
 153     qq02             = _mm_mul_ps(iq0,jq2);
 154     qq10             = _mm_mul_ps(iq1,jq0);
 155     qq11             = _mm_mul_ps(iq1,jq1);
 156     qq12             = _mm_mul_ps(iq1,jq2);
 157     qq20             = _mm_mul_ps(iq2,jq0);
 158     qq21             = _mm_mul_ps(iq2,jq1);
 159     qq22             = _mm_mul_ps(iq2,jq2);
 160
 161     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 162     rcutoff_scalar   = fr->rcoulomb;
 163     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 164     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 165
 166     rswitch_scalar   = fr->rvdw_switch;
 167     rswitch          = _mm_set1_ps(rswitch_scalar);
 168     /* Setup switch parameters */
 169     d_scalar         = rcutoff_scalar-rswitch_scalar;
 170     d                = _mm_set1_ps(d_scalar);
 171     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 172     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 173     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 174     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 175     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 176     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 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             /* REACTION-FIELD ELECTROSTATICS */
 324             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 325             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 326
 327             /* LENNARD-JONES DISPERSION/REPULSION */
 328
 329             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 330             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 331             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 332             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 333             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 334
 335             d                = _mm_sub_ps(r00,rswitch);
 336             d                = _mm_max_ps(d,_mm_setzero_ps());
 337             d2               = _mm_mul_ps(d,d);
 338             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 339
 340             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 341
 342             /* Evaluate switch function */
 343             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 344             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 345             vvdw             = _mm_mul_ps(vvdw,sw);
 346             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 347
 348             /* Update potential sum for this i atom from the interaction with this j atom. */
 349             velec            = _mm_and_ps(velec,cutoff_mask);
 350             velecsum         = _mm_add_ps(velecsum,velec);
 351             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 352             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 353
 354             fscal            = _mm_add_ps(felec,fvdw);
 355
 356             fscal            = _mm_and_ps(fscal,cutoff_mask);
 357
 358              /* Update vectorial force */
 359             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 360             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 361             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 362
 363             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 364             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 365             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 366
 367             }
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             if (gmx_mm_any_lt(rsq01,rcutoff2))
 374             {
 375
 376             /* REACTION-FIELD ELECTROSTATICS */
 377             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 378             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 379
 380             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velec            = _mm_and_ps(velec,cutoff_mask);
 384             velecsum         = _mm_add_ps(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             fscal            = _mm_and_ps(fscal,cutoff_mask);
 389
 390              /* Update vectorial force */
 391             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 392             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 393             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 394
 395             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 396             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 397             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 398
 399             }
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             if (gmx_mm_any_lt(rsq02,rcutoff2))
 406             {
 407
 408             /* REACTION-FIELD ELECTROSTATICS */
 409             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 410             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 411
 412             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velec            = _mm_and_ps(velec,cutoff_mask);
 416             velecsum         = _mm_add_ps(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             fscal            = _mm_and_ps(fscal,cutoff_mask);
 421
 422              /* Update vectorial force */
 423             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 424             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 425             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 426
 427             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 428             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 429             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 430
 431             }
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             if (gmx_mm_any_lt(rsq10,rcutoff2))
 438             {
 439
 440             /* REACTION-FIELD ELECTROSTATICS */
 441             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 442             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 443
 444             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velec            = _mm_and_ps(velec,cutoff_mask);
 448             velecsum         = _mm_add_ps(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             fscal            = _mm_and_ps(fscal,cutoff_mask);
 453
 454              /* Update vectorial force */
 455             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 456             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 457             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 458
 459             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 460             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 461             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 462
 463             }
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             if (gmx_mm_any_lt(rsq11,rcutoff2))
 470             {
 471
 472             /* REACTION-FIELD ELECTROSTATICS */
 473             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 474             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 475
 476             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 477
 478             /* Update potential sum for this i atom from the interaction with this j atom. */
 479             velec            = _mm_and_ps(velec,cutoff_mask);
 480             velecsum         = _mm_add_ps(velecsum,velec);
 481
 482             fscal            = felec;
 483
 484             fscal            = _mm_and_ps(fscal,cutoff_mask);
 485
 486              /* Update vectorial force */
 487             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 488             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 489             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 490
 491             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 492             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 493             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 494
 495             }
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             if (gmx_mm_any_lt(rsq12,rcutoff2))
 502             {
 503
 504             /* REACTION-FIELD ELECTROSTATICS */
 505             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 506             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 507
 508             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 509
 510             /* Update potential sum for this i atom from the interaction with this j atom. */
 511             velec            = _mm_and_ps(velec,cutoff_mask);
 512             velecsum         = _mm_add_ps(velecsum,velec);
 513
 514             fscal            = felec;
 515
 516             fscal            = _mm_and_ps(fscal,cutoff_mask);
 517
 518              /* Update vectorial force */
 519             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 520             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 521             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 522
 523             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 524             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 525             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 526
 527             }
 528
 529             /**************************
 530              * CALCULATE INTERACTIONS *
 531              **************************/
 532
 533             if (gmx_mm_any_lt(rsq20,rcutoff2))
 534             {
 535
 536             /* REACTION-FIELD ELECTROSTATICS */
 537             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 538             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 539
 540             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 541
 542             /* Update potential sum for this i atom from the interaction with this j atom. */
 543             velec            = _mm_and_ps(velec,cutoff_mask);
 544             velecsum         = _mm_add_ps(velecsum,velec);
 545
 546             fscal            = felec;
 547
 548             fscal            = _mm_and_ps(fscal,cutoff_mask);
 549
 550              /* Update vectorial force */
 551             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 552             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 553             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 554
 555             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 556             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 557             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 558
 559             }
 560
 561             /**************************
 562              * CALCULATE INTERACTIONS *
 563              **************************/
 564
 565             if (gmx_mm_any_lt(rsq21,rcutoff2))
 566             {
 567
 568             /* REACTION-FIELD ELECTROSTATICS */
 569             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 570             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 571
 572             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 573
 574             /* Update potential sum for this i atom from the interaction with this j atom. */
 575             velec            = _mm_and_ps(velec,cutoff_mask);
 576             velecsum         = _mm_add_ps(velecsum,velec);
 577
 578             fscal            = felec;
 579
 580             fscal            = _mm_and_ps(fscal,cutoff_mask);
 581
 582              /* Update vectorial force */
 583             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 584             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 585             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 586
 587             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 588             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 589             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 590
 591             }
 592
 593             /**************************
 594              * CALCULATE INTERACTIONS *
 595              **************************/
 596
 597             if (gmx_mm_any_lt(rsq22,rcutoff2))
 598             {
 599
 600             /* REACTION-FIELD ELECTROSTATICS */
 601             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 602             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 603
 604             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 605
 606             /* Update potential sum for this i atom from the interaction with this j atom. */
 607             velec            = _mm_and_ps(velec,cutoff_mask);
 608             velecsum         = _mm_add_ps(velecsum,velec);
 609
 610             fscal            = felec;
 611
 612             fscal            = _mm_and_ps(fscal,cutoff_mask);
 613
 614              /* Update vectorial force */
 615             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 616             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 617             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 618
 619             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 620             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 621             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 622
 623             }
 624
 625             fjptrA             = f+j_coord_offsetA;
 626             fjptrB             = f+j_coord_offsetB;
 627             fjptrC             = f+j_coord_offsetC;
 628             fjptrD             = f+j_coord_offsetD;
 629
 630             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 631                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 632
 633             /* Inner loop uses 385 flops */
 634         }
 635
 636         if(jidx<j_index_end)
 637         {
 638
 639             /* Get j neighbor index, and coordinate index */
 640             jnrlistA         = jjnr[jidx];
 641             jnrlistB         = jjnr[jidx+1];
 642             jnrlistC         = jjnr[jidx+2];
 643             jnrlistD         = jjnr[jidx+3];
 644             /* Sign of each element will be negative for non-real atoms.
 645              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 646              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 647              */
 648             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 649             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 650             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 651             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 652             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 653             j_coord_offsetA  = DIM*jnrA;
 654             j_coord_offsetB  = DIM*jnrB;
 655             j_coord_offsetC  = DIM*jnrC;
 656             j_coord_offsetD  = DIM*jnrD;
 657
 658             /* load j atom coordinates */
 659             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 660                                               x+j_coord_offsetC,x+j_coord_offsetD,
 661                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 662
 663             /* Calculate displacement vector */
 664             dx00             = _mm_sub_ps(ix0,jx0);
 665             dy00             = _mm_sub_ps(iy0,jy0);
 666             dz00             = _mm_sub_ps(iz0,jz0);
 667             dx01             = _mm_sub_ps(ix0,jx1);
 668             dy01             = _mm_sub_ps(iy0,jy1);
 669             dz01             = _mm_sub_ps(iz0,jz1);
 670             dx02             = _mm_sub_ps(ix0,jx2);
 671             dy02             = _mm_sub_ps(iy0,jy2);
 672             dz02             = _mm_sub_ps(iz0,jz2);
 673             dx10             = _mm_sub_ps(ix1,jx0);
 674             dy10             = _mm_sub_ps(iy1,jy0);
 675             dz10             = _mm_sub_ps(iz1,jz0);
 676             dx11             = _mm_sub_ps(ix1,jx1);
 677             dy11             = _mm_sub_ps(iy1,jy1);
 678             dz11             = _mm_sub_ps(iz1,jz1);
 679             dx12             = _mm_sub_ps(ix1,jx2);
 680             dy12             = _mm_sub_ps(iy1,jy2);
 681             dz12             = _mm_sub_ps(iz1,jz2);
 682             dx20             = _mm_sub_ps(ix2,jx0);
 683             dy20             = _mm_sub_ps(iy2,jy0);
 684             dz20             = _mm_sub_ps(iz2,jz0);
 685             dx21             = _mm_sub_ps(ix2,jx1);
 686             dy21             = _mm_sub_ps(iy2,jy1);
 687             dz21             = _mm_sub_ps(iz2,jz1);
 688             dx22             = _mm_sub_ps(ix2,jx2);
 689             dy22             = _mm_sub_ps(iy2,jy2);
 690             dz22             = _mm_sub_ps(iz2,jz2);
 691
 692             /* Calculate squared distance and things based on it */
 693             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 694             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 695             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 696             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 697             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 698             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 699             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 700             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 701             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 702
 703             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 704             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 705             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 706             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 707             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 708             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 709             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 710             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 711             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 712
 713             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 714             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 715             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 716             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 717             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 718             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 719             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 720             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 721             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 722
 723             fjx0             = _mm_setzero_ps();
 724             fjy0             = _mm_setzero_ps();
 725             fjz0             = _mm_setzero_ps();
 726             fjx1             = _mm_setzero_ps();
 727             fjy1             = _mm_setzero_ps();
 728             fjz1             = _mm_setzero_ps();
 729             fjx2             = _mm_setzero_ps();
 730             fjy2             = _mm_setzero_ps();
 731             fjz2             = _mm_setzero_ps();
 732
 733             /**************************
 734              * CALCULATE INTERACTIONS *
 735              **************************/
 736
 737             if (gmx_mm_any_lt(rsq00,rcutoff2))
 738             {
 739
 740             r00              = _mm_mul_ps(rsq00,rinv00);
 741             r00              = _mm_andnot_ps(dummy_mask,r00);
 742
 743             /* REACTION-FIELD ELECTROSTATICS */
 744             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 745             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 746
 747             /* LENNARD-JONES DISPERSION/REPULSION */
 748
 749             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 750             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 751             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 752             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 753             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 754
 755             d                = _mm_sub_ps(r00,rswitch);
 756             d                = _mm_max_ps(d,_mm_setzero_ps());
 757             d2               = _mm_mul_ps(d,d);
 758             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 759
 760             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 761
 762             /* Evaluate switch function */
 763             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 764             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 765             vvdw             = _mm_mul_ps(vvdw,sw);
 766             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 767
 768             /* Update potential sum for this i atom from the interaction with this j atom. */
 769             velec            = _mm_and_ps(velec,cutoff_mask);
 770             velec            = _mm_andnot_ps(dummy_mask,velec);
 771             velecsum         = _mm_add_ps(velecsum,velec);
 772             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 773             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 774             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 775
 776             fscal            = _mm_add_ps(felec,fvdw);
 777
 778             fscal            = _mm_and_ps(fscal,cutoff_mask);
 779
 780             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 781
 782              /* Update vectorial force */
 783             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 784             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 785             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 786
 787             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 788             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 789             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 790
 791             }
 792
 793             /**************************
 794              * CALCULATE INTERACTIONS *
 795              **************************/
 796
 797             if (gmx_mm_any_lt(rsq01,rcutoff2))
 798             {
 799
 800             /* REACTION-FIELD ELECTROSTATICS */
 801             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 802             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 803
 804             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 805
 806             /* Update potential sum for this i atom from the interaction with this j atom. */
 807             velec            = _mm_and_ps(velec,cutoff_mask);
 808             velec            = _mm_andnot_ps(dummy_mask,velec);
 809             velecsum         = _mm_add_ps(velecsum,velec);
 810
 811             fscal            = felec;
 812
 813             fscal            = _mm_and_ps(fscal,cutoff_mask);
 814
 815             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 816
 817              /* Update vectorial force */
 818             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 819             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 820             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 821
 822             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 823             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 824             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 825
 826             }
 827
 828             /**************************
 829              * CALCULATE INTERACTIONS *
 830              **************************/
 831
 832             if (gmx_mm_any_lt(rsq02,rcutoff2))
 833             {
 834
 835             /* REACTION-FIELD ELECTROSTATICS */
 836             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 837             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 838
 839             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 840
 841             /* Update potential sum for this i atom from the interaction with this j atom. */
 842             velec            = _mm_and_ps(velec,cutoff_mask);
 843             velec            = _mm_andnot_ps(dummy_mask,velec);
 844             velecsum         = _mm_add_ps(velecsum,velec);
 845
 846             fscal            = felec;
 847
 848             fscal            = _mm_and_ps(fscal,cutoff_mask);
 849
 850             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 851
 852              /* Update vectorial force */
 853             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 854             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 855             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 856
 857             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 858             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 859             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 860
 861             }
 862
 863             /**************************
 864              * CALCULATE INTERACTIONS *
 865              **************************/
 866
 867             if (gmx_mm_any_lt(rsq10,rcutoff2))
 868             {
 869
 870             /* REACTION-FIELD ELECTROSTATICS */
 871             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 872             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 873
 874             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 875
 876             /* Update potential sum for this i atom from the interaction with this j atom. */
 877             velec            = _mm_and_ps(velec,cutoff_mask);
 878             velec            = _mm_andnot_ps(dummy_mask,velec);
 879             velecsum         = _mm_add_ps(velecsum,velec);
 880
 881             fscal            = felec;
 882
 883             fscal            = _mm_and_ps(fscal,cutoff_mask);
 884
 885             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 886
 887              /* Update vectorial force */
 888             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 889             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 890             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 891
 892             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 893             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 894             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 895
 896             }
 897
 898             /**************************
 899              * CALCULATE INTERACTIONS *
 900              **************************/
 901
 902             if (gmx_mm_any_lt(rsq11,rcutoff2))
 903             {
 904
 905             /* REACTION-FIELD ELECTROSTATICS */
 906             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 907             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 908
 909             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 910
 911             /* Update potential sum for this i atom from the interaction with this j atom. */
 912             velec            = _mm_and_ps(velec,cutoff_mask);
 913             velec            = _mm_andnot_ps(dummy_mask,velec);
 914             velecsum         = _mm_add_ps(velecsum,velec);
 915
 916             fscal            = felec;
 917
 918             fscal            = _mm_and_ps(fscal,cutoff_mask);
 919
 920             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 921
 922              /* Update vectorial force */
 923             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 924             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 925             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 926
 927             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 928             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 929             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 930
 931             }
 932
 933             /**************************
 934              * CALCULATE INTERACTIONS *
 935              **************************/
 936
 937             if (gmx_mm_any_lt(rsq12,rcutoff2))
 938             {
 939
 940             /* REACTION-FIELD ELECTROSTATICS */
 941             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 942             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 943
 944             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 945
 946             /* Update potential sum for this i atom from the interaction with this j atom. */
 947             velec            = _mm_and_ps(velec,cutoff_mask);
 948             velec            = _mm_andnot_ps(dummy_mask,velec);
 949             velecsum         = _mm_add_ps(velecsum,velec);
 950
 951             fscal            = felec;
 952
 953             fscal            = _mm_and_ps(fscal,cutoff_mask);
 954
 955             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 956
 957              /* Update vectorial force */
 958             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 959             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 960             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 961
 962             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 963             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 964             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 965
 966             }
 967
 968             /**************************
 969              * CALCULATE INTERACTIONS *
 970              **************************/
 971
 972             if (gmx_mm_any_lt(rsq20,rcutoff2))
 973             {
 974
 975             /* REACTION-FIELD ELECTROSTATICS */
 976             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 977             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 978
 979             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 980
 981             /* Update potential sum for this i atom from the interaction with this j atom. */
 982             velec            = _mm_and_ps(velec,cutoff_mask);
 983             velec            = _mm_andnot_ps(dummy_mask,velec);
 984             velecsum         = _mm_add_ps(velecsum,velec);
 985
 986             fscal            = felec;
 987
 988             fscal            = _mm_and_ps(fscal,cutoff_mask);
 989
 990             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 991
 992              /* Update vectorial force */
 993             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 994             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 995             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 996
 997             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 998             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 999             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1000
1001             }
1002
1003             /**************************
1004              * CALCULATE INTERACTIONS *
1005              **************************/
1006
1007             if (gmx_mm_any_lt(rsq21,rcutoff2))
1008             {
1009
1010             /* REACTION-FIELD ELECTROSTATICS */
1011             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
1012             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1013
1014             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1015
1016             /* Update potential sum for this i atom from the interaction with this j atom. */
1017             velec            = _mm_and_ps(velec,cutoff_mask);
1018             velec            = _mm_andnot_ps(dummy_mask,velec);
1019             velecsum         = _mm_add_ps(velecsum,velec);
1020
1021             fscal            = felec;
1022
1023             fscal            = _mm_and_ps(fscal,cutoff_mask);
1024
1025             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1026
1027              /* Update vectorial force */
1028             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1029             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1030             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1031
1032             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1033             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1034             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1035
1036             }
1037
1038             /**************************
1039              * CALCULATE INTERACTIONS *
1040              **************************/
1041
1042             if (gmx_mm_any_lt(rsq22,rcutoff2))
1043             {
1044
1045             /* REACTION-FIELD ELECTROSTATICS */
1046             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
1047             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1048
1049             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1050
1051             /* Update potential sum for this i atom from the interaction with this j atom. */
1052             velec            = _mm_and_ps(velec,cutoff_mask);
1053             velec            = _mm_andnot_ps(dummy_mask,velec);
1054             velecsum         = _mm_add_ps(velecsum,velec);
1055
1056             fscal            = felec;
1057
1058             fscal            = _mm_and_ps(fscal,cutoff_mask);
1059
1060             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1061
1062              /* Update vectorial force */
1063             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1064             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1065             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1066
1067             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1068             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1069             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1070
1071             }
1072
1073             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1074             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1075             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1076             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1077
1078             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1079                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1080
1081             /* Inner loop uses 386 flops */
1082         }
1083
1084         /* End of innermost loop */
1085
1086         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1087                                               f+i_coord_offset,fshift+i_shift_offset);
1088
1089         ggid                        = gid[iidx];
1090         /* Update potential energies */
1091         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1092         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1093
1094         /* Increment number of inner iterations */
1095         inneriter                  += j_index_end - j_index_start;
1096
1097         /* Outer loop uses 20 flops */
1098     }
1099
1100     /* Increment number of outer iterations */
1101     outeriter        += nri;
1102
1103     /* Update outer/inner flops */
1104
1105     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*386);
1106 }
1107 /*
1108  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_single
1109  * Electrostatics interaction: ReactionField
1110  * VdW interaction:            LennardJones
1111  * Geometry:                   Water3-Water3
1112  * Calculate force/pot:        Force
1113  */
1114 void
1115 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_single
1116                     (t_nblist                    * gmx_restrict       nlist,
1117                      rvec                        * gmx_restrict          xx,
1118                      rvec                        * gmx_restrict          ff,
1119                      t_forcerec                  * gmx_restrict          fr,
1120                      t_mdatoms                   * gmx_restrict     mdatoms,
1121                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1122                      t_nrnb                      * gmx_restrict        nrnb)
1123 {
1124     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1125      * just 0 for non-waters.
1126      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1127      * jnr indices corresponding to data put in the four positions in the SIMD register.
1128      */
1129     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1130     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1131     int              jnrA,jnrB,jnrC,jnrD;
1132     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1133     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1134     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1135     real             rcutoff_scalar;
1136     real             *shiftvec,*fshift,*x,*f;
1137     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1138     real             scratch[4*DIM];
1139     __m128           fscal,rcutoff,rcutoff2,jidxall;
1140     int              vdwioffset0;
1141     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1142     int              vdwioffset1;
1143     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1144     int              vdwioffset2;
1145     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1146     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1147     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1148     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1149     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1150     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1151     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1152     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1153     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1154     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1155     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1156     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1157     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1158     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1159     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1160     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1161     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1162     real             *charge;
1163     int              nvdwtype;
1164     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1165     int              *vdwtype;
1166     real             *vdwparam;
1167     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1168     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1169     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1170     real             rswitch_scalar,d_scalar;
1171     __m128           dummy_mask,cutoff_mask;
1172     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1173     __m128           one     = _mm_set1_ps(1.0);
1174     __m128           two     = _mm_set1_ps(2.0);
1175     x                = xx[0];
1176     f                = ff[0];
1177
1178     nri              = nlist->nri;
1179     iinr             = nlist->iinr;
1180     jindex           = nlist->jindex;
1181     jjnr             = nlist->jjnr;
1182     shiftidx         = nlist->shift;
1183     gid              = nlist->gid;
1184     shiftvec         = fr->shift_vec[0];
1185     fshift           = fr->fshift[0];
1186     facel            = _mm_set1_ps(fr->epsfac);
1187     charge           = mdatoms->chargeA;
1188     krf              = _mm_set1_ps(fr->ic->k_rf);
1189     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1190     crf              = _mm_set1_ps(fr->ic->c_rf);
1191     nvdwtype         = fr->ntype;
1192     vdwparam         = fr->nbfp;
1193     vdwtype          = mdatoms->typeA;
1194
1195     /* Setup water-specific parameters */
1196     inr              = nlist->iinr[0];
1197     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1198     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1199     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1200     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1201
1202     jq0              = _mm_set1_ps(charge[inr+0]);
1203     jq1              = _mm_set1_ps(charge[inr+1]);
1204     jq2              = _mm_set1_ps(charge[inr+2]);
1205     vdwjidx0A        = 2*vdwtype[inr+0];
1206     qq00             = _mm_mul_ps(iq0,jq0);
1207     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1208     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1209     qq01             = _mm_mul_ps(iq0,jq1);
1210     qq02             = _mm_mul_ps(iq0,jq2);
1211     qq10             = _mm_mul_ps(iq1,jq0);
1212     qq11             = _mm_mul_ps(iq1,jq1);
1213     qq12             = _mm_mul_ps(iq1,jq2);
1214     qq20             = _mm_mul_ps(iq2,jq0);
1215     qq21             = _mm_mul_ps(iq2,jq1);
1216     qq22             = _mm_mul_ps(iq2,jq2);
1217
1218     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1219     rcutoff_scalar   = fr->rcoulomb;
1220     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1221     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1222
1223     rswitch_scalar   = fr->rvdw_switch;
1224     rswitch          = _mm_set1_ps(rswitch_scalar);
1225     /* Setup switch parameters */
1226     d_scalar         = rcutoff_scalar-rswitch_scalar;
1227     d                = _mm_set1_ps(d_scalar);
1228     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
1229     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1230     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1231     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
1232     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1233     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1234
1235     /* Avoid stupid compiler warnings */
1236     jnrA = jnrB = jnrC = jnrD = 0;
1237     j_coord_offsetA = 0;
1238     j_coord_offsetB = 0;
1239     j_coord_offsetC = 0;
1240     j_coord_offsetD = 0;
1241
1242     outeriter        = 0;
1243     inneriter        = 0;
1244
1245     for(iidx=0;iidx<4*DIM;iidx++)
1246     {
1247         scratch[iidx] = 0.0;
1248     }
1249
1250     /* Start outer loop over neighborlists */
1251     for(iidx=0; iidx<nri; iidx++)
1252     {
1253         /* Load shift vector for this list */
1254         i_shift_offset   = DIM*shiftidx[iidx];
1255
1256         /* Load limits for loop over neighbors */
1257         j_index_start    = jindex[iidx];
1258         j_index_end      = jindex[iidx+1];
1259
1260         /* Get outer coordinate index */
1261         inr              = iinr[iidx];
1262         i_coord_offset   = DIM*inr;
1263
1264         /* Load i particle coords and add shift vector */
1265         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1266                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1267
1268         fix0             = _mm_setzero_ps();
1269         fiy0             = _mm_setzero_ps();
1270         fiz0             = _mm_setzero_ps();
1271         fix1             = _mm_setzero_ps();
1272         fiy1             = _mm_setzero_ps();
1273         fiz1             = _mm_setzero_ps();
1274         fix2             = _mm_setzero_ps();
1275         fiy2             = _mm_setzero_ps();
1276         fiz2             = _mm_setzero_ps();
1277
1278         /* Start inner kernel loop */
1279         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1280         {
1281
1282             /* Get j neighbor index, and coordinate index */
1283             jnrA             = jjnr[jidx];
1284             jnrB             = jjnr[jidx+1];
1285             jnrC             = jjnr[jidx+2];
1286             jnrD             = jjnr[jidx+3];
1287             j_coord_offsetA  = DIM*jnrA;
1288             j_coord_offsetB  = DIM*jnrB;
1289             j_coord_offsetC  = DIM*jnrC;
1290             j_coord_offsetD  = DIM*jnrD;
1291
1292             /* load j atom coordinates */
1293             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1294                                               x+j_coord_offsetC,x+j_coord_offsetD,
1295                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1296
1297             /* Calculate displacement vector */
1298             dx00             = _mm_sub_ps(ix0,jx0);
1299             dy00             = _mm_sub_ps(iy0,jy0);
1300             dz00             = _mm_sub_ps(iz0,jz0);
1301             dx01             = _mm_sub_ps(ix0,jx1);
1302             dy01             = _mm_sub_ps(iy0,jy1);
1303             dz01             = _mm_sub_ps(iz0,jz1);
1304             dx02             = _mm_sub_ps(ix0,jx2);
1305             dy02             = _mm_sub_ps(iy0,jy2);
1306             dz02             = _mm_sub_ps(iz0,jz2);
1307             dx10             = _mm_sub_ps(ix1,jx0);
1308             dy10             = _mm_sub_ps(iy1,jy0);
1309             dz10             = _mm_sub_ps(iz1,jz0);
1310             dx11             = _mm_sub_ps(ix1,jx1);
1311             dy11             = _mm_sub_ps(iy1,jy1);
1312             dz11             = _mm_sub_ps(iz1,jz1);
1313             dx12             = _mm_sub_ps(ix1,jx2);
1314             dy12             = _mm_sub_ps(iy1,jy2);
1315             dz12             = _mm_sub_ps(iz1,jz2);
1316             dx20             = _mm_sub_ps(ix2,jx0);
1317             dy20             = _mm_sub_ps(iy2,jy0);
1318             dz20             = _mm_sub_ps(iz2,jz0);
1319             dx21             = _mm_sub_ps(ix2,jx1);
1320             dy21             = _mm_sub_ps(iy2,jy1);
1321             dz21             = _mm_sub_ps(iz2,jz1);
1322             dx22             = _mm_sub_ps(ix2,jx2);
1323             dy22             = _mm_sub_ps(iy2,jy2);
1324             dz22             = _mm_sub_ps(iz2,jz2);
1325
1326             /* Calculate squared distance and things based on it */
1327             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1328             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1329             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1330             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1331             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1332             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1333             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1334             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1335             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1336
1337             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1338             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1339             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1340             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1341             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1342             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1343             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1344             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1345             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1346
1347             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1348             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1349             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1350             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1351             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1352             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1353             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1354             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1355             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1356
1357             fjx0             = _mm_setzero_ps();
1358             fjy0             = _mm_setzero_ps();
1359             fjz0             = _mm_setzero_ps();
1360             fjx1             = _mm_setzero_ps();
1361             fjy1             = _mm_setzero_ps();
1362             fjz1             = _mm_setzero_ps();
1363             fjx2             = _mm_setzero_ps();
1364             fjy2             = _mm_setzero_ps();
1365             fjz2             = _mm_setzero_ps();
1366
1367             /**************************
1368              * CALCULATE INTERACTIONS *
1369              **************************/
1370
1371             if (gmx_mm_any_lt(rsq00,rcutoff2))
1372             {
1373
1374             r00              = _mm_mul_ps(rsq00,rinv00);
1375
1376             /* REACTION-FIELD ELECTROSTATICS */
1377             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1378
1379             /* LENNARD-JONES DISPERSION/REPULSION */
1380
1381             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1382             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1383             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1384             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1385             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1386
1387             d                = _mm_sub_ps(r00,rswitch);
1388             d                = _mm_max_ps(d,_mm_setzero_ps());
1389             d2               = _mm_mul_ps(d,d);
1390             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1391
1392             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1393
1394             /* Evaluate switch function */
1395             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1396             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1397             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1398
1399             fscal            = _mm_add_ps(felec,fvdw);
1400
1401             fscal            = _mm_and_ps(fscal,cutoff_mask);
1402
1403              /* Update vectorial force */
1404             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1405             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1406             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1407
1408             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1409             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1410             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1411
1412             }
1413
1414             /**************************
1415              * CALCULATE INTERACTIONS *
1416              **************************/
1417
1418             if (gmx_mm_any_lt(rsq01,rcutoff2))
1419             {
1420
1421             /* REACTION-FIELD ELECTROSTATICS */
1422             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1423
1424             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1425
1426             fscal            = felec;
1427
1428             fscal            = _mm_and_ps(fscal,cutoff_mask);
1429
1430              /* Update vectorial force */
1431             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1432             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1433             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1434
1435             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1436             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1437             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1438
1439             }
1440
1441             /**************************
1442              * CALCULATE INTERACTIONS *
1443              **************************/
1444
1445             if (gmx_mm_any_lt(rsq02,rcutoff2))
1446             {
1447
1448             /* REACTION-FIELD ELECTROSTATICS */
1449             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1450
1451             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1452
1453             fscal            = felec;
1454
1455             fscal            = _mm_and_ps(fscal,cutoff_mask);
1456
1457              /* Update vectorial force */
1458             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1459             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1460             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1461
1462             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1463             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1464             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1465
1466             }
1467
1468             /**************************
1469              * CALCULATE INTERACTIONS *
1470              **************************/
1471
1472             if (gmx_mm_any_lt(rsq10,rcutoff2))
1473             {
1474
1475             /* REACTION-FIELD ELECTROSTATICS */
1476             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1477
1478             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1479
1480             fscal            = felec;
1481
1482             fscal            = _mm_and_ps(fscal,cutoff_mask);
1483
1484              /* Update vectorial force */
1485             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1486             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1487             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1488
1489             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1490             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1491             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1492
1493             }
1494
1495             /**************************
1496              * CALCULATE INTERACTIONS *
1497              **************************/
1498
1499             if (gmx_mm_any_lt(rsq11,rcutoff2))
1500             {
1501
1502             /* REACTION-FIELD ELECTROSTATICS */
1503             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1504
1505             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1506
1507             fscal            = felec;
1508
1509             fscal            = _mm_and_ps(fscal,cutoff_mask);
1510
1511              /* Update vectorial force */
1512             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1513             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1514             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1515
1516             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1517             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1518             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1519
1520             }
1521
1522             /**************************
1523              * CALCULATE INTERACTIONS *
1524              **************************/
1525
1526             if (gmx_mm_any_lt(rsq12,rcutoff2))
1527             {
1528
1529             /* REACTION-FIELD ELECTROSTATICS */
1530             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1531
1532             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1533
1534             fscal            = felec;
1535
1536             fscal            = _mm_and_ps(fscal,cutoff_mask);
1537
1538              /* Update vectorial force */
1539             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1540             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1541             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1542
1543             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1544             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1545             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1546
1547             }
1548
1549             /**************************
1550              * CALCULATE INTERACTIONS *
1551              **************************/
1552
1553             if (gmx_mm_any_lt(rsq20,rcutoff2))
1554             {
1555
1556             /* REACTION-FIELD ELECTROSTATICS */
1557             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1558
1559             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1560
1561             fscal            = felec;
1562
1563             fscal            = _mm_and_ps(fscal,cutoff_mask);
1564
1565              /* Update vectorial force */
1566             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1567             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1568             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1569
1570             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1571             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1572             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1573
1574             }
1575
1576             /**************************
1577              * CALCULATE INTERACTIONS *
1578              **************************/
1579
1580             if (gmx_mm_any_lt(rsq21,rcutoff2))
1581             {
1582
1583             /* REACTION-FIELD ELECTROSTATICS */
1584             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1585
1586             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1587
1588             fscal            = felec;
1589
1590             fscal            = _mm_and_ps(fscal,cutoff_mask);
1591
1592              /* Update vectorial force */
1593             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1594             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1595             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1596
1597             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1598             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1599             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1600
1601             }
1602
1603             /**************************
1604              * CALCULATE INTERACTIONS *
1605              **************************/
1606
1607             if (gmx_mm_any_lt(rsq22,rcutoff2))
1608             {
1609
1610             /* REACTION-FIELD ELECTROSTATICS */
1611             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1612
1613             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1614
1615             fscal            = felec;
1616
1617             fscal            = _mm_and_ps(fscal,cutoff_mask);
1618
1619              /* Update vectorial force */
1620             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1621             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1622             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1623
1624             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1625             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1626             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1627
1628             }
1629
1630             fjptrA             = f+j_coord_offsetA;
1631             fjptrB             = f+j_coord_offsetB;
1632             fjptrC             = f+j_coord_offsetC;
1633             fjptrD             = f+j_coord_offsetD;
1634
1635             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1636                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1637
1638             /* Inner loop uses 328 flops */
1639         }
1640
1641         if(jidx<j_index_end)
1642         {
1643
1644             /* Get j neighbor index, and coordinate index */
1645             jnrlistA         = jjnr[jidx];
1646             jnrlistB         = jjnr[jidx+1];
1647             jnrlistC         = jjnr[jidx+2];
1648             jnrlistD         = jjnr[jidx+3];
1649             /* Sign of each element will be negative for non-real atoms.
1650              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1651              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1652              */
1653             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1654             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1655             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1656             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1657             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1658             j_coord_offsetA  = DIM*jnrA;
1659             j_coord_offsetB  = DIM*jnrB;
1660             j_coord_offsetC  = DIM*jnrC;
1661             j_coord_offsetD  = DIM*jnrD;
1662
1663             /* load j atom coordinates */
1664             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1665                                               x+j_coord_offsetC,x+j_coord_offsetD,
1666                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1667
1668             /* Calculate displacement vector */
1669             dx00             = _mm_sub_ps(ix0,jx0);
1670             dy00             = _mm_sub_ps(iy0,jy0);
1671             dz00             = _mm_sub_ps(iz0,jz0);
1672             dx01             = _mm_sub_ps(ix0,jx1);
1673             dy01             = _mm_sub_ps(iy0,jy1);
1674             dz01             = _mm_sub_ps(iz0,jz1);
1675             dx02             = _mm_sub_ps(ix0,jx2);
1676             dy02             = _mm_sub_ps(iy0,jy2);
1677             dz02             = _mm_sub_ps(iz0,jz2);
1678             dx10             = _mm_sub_ps(ix1,jx0);
1679             dy10             = _mm_sub_ps(iy1,jy0);
1680             dz10             = _mm_sub_ps(iz1,jz0);
1681             dx11             = _mm_sub_ps(ix1,jx1);
1682             dy11             = _mm_sub_ps(iy1,jy1);
1683             dz11             = _mm_sub_ps(iz1,jz1);
1684             dx12             = _mm_sub_ps(ix1,jx2);
1685             dy12             = _mm_sub_ps(iy1,jy2);
1686             dz12             = _mm_sub_ps(iz1,jz2);
1687             dx20             = _mm_sub_ps(ix2,jx0);
1688             dy20             = _mm_sub_ps(iy2,jy0);
1689             dz20             = _mm_sub_ps(iz2,jz0);
1690             dx21             = _mm_sub_ps(ix2,jx1);
1691             dy21             = _mm_sub_ps(iy2,jy1);
1692             dz21             = _mm_sub_ps(iz2,jz1);
1693             dx22             = _mm_sub_ps(ix2,jx2);
1694             dy22             = _mm_sub_ps(iy2,jy2);
1695             dz22             = _mm_sub_ps(iz2,jz2);
1696
1697             /* Calculate squared distance and things based on it */
1698             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1699             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1700             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1701             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1702             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1703             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1704             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1705             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1706             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1707
1708             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1709             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1710             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1711             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1712             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1713             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1714             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1715             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1716             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1717
1718             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1719             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1720             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1721             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1722             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1723             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1724             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1725             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1726             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1727
1728             fjx0             = _mm_setzero_ps();
1729             fjy0             = _mm_setzero_ps();
1730             fjz0             = _mm_setzero_ps();
1731             fjx1             = _mm_setzero_ps();
1732             fjy1             = _mm_setzero_ps();
1733             fjz1             = _mm_setzero_ps();
1734             fjx2             = _mm_setzero_ps();
1735             fjy2             = _mm_setzero_ps();
1736             fjz2             = _mm_setzero_ps();
1737
1738             /**************************
1739              * CALCULATE INTERACTIONS *
1740              **************************/
1741
1742             if (gmx_mm_any_lt(rsq00,rcutoff2))
1743             {
1744
1745             r00              = _mm_mul_ps(rsq00,rinv00);
1746             r00              = _mm_andnot_ps(dummy_mask,r00);
1747
1748             /* REACTION-FIELD ELECTROSTATICS */
1749             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1750
1751             /* LENNARD-JONES DISPERSION/REPULSION */
1752
1753             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1754             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1755             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1756             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1757             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1758
1759             d                = _mm_sub_ps(r00,rswitch);
1760             d                = _mm_max_ps(d,_mm_setzero_ps());
1761             d2               = _mm_mul_ps(d,d);
1762             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1763
1764             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1765
1766             /* Evaluate switch function */
1767             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1768             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1769             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1770
1771             fscal            = _mm_add_ps(felec,fvdw);
1772
1773             fscal            = _mm_and_ps(fscal,cutoff_mask);
1774
1775             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1776
1777              /* Update vectorial force */
1778             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1779             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1780             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1781
1782             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1783             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1784             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1785
1786             }
1787
1788             /**************************
1789              * CALCULATE INTERACTIONS *
1790              **************************/
1791
1792             if (gmx_mm_any_lt(rsq01,rcutoff2))
1793             {
1794
1795             /* REACTION-FIELD ELECTROSTATICS */
1796             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1797
1798             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1799
1800             fscal            = felec;
1801
1802             fscal            = _mm_and_ps(fscal,cutoff_mask);
1803
1804             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1805
1806              /* Update vectorial force */
1807             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1808             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1809             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1810
1811             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1812             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1813             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1814
1815             }
1816
1817             /**************************
1818              * CALCULATE INTERACTIONS *
1819              **************************/
1820
1821             if (gmx_mm_any_lt(rsq02,rcutoff2))
1822             {
1823
1824             /* REACTION-FIELD ELECTROSTATICS */
1825             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1826
1827             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1828
1829             fscal            = felec;
1830
1831             fscal            = _mm_and_ps(fscal,cutoff_mask);
1832
1833             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1834
1835              /* Update vectorial force */
1836             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1837             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1838             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1839
1840             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1841             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1842             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1843
1844             }
1845
1846             /**************************
1847              * CALCULATE INTERACTIONS *
1848              **************************/
1849
1850             if (gmx_mm_any_lt(rsq10,rcutoff2))
1851             {
1852
1853             /* REACTION-FIELD ELECTROSTATICS */
1854             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1855
1856             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1857
1858             fscal            = felec;
1859
1860             fscal            = _mm_and_ps(fscal,cutoff_mask);
1861
1862             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1863
1864              /* Update vectorial force */
1865             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1866             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1867             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1868
1869             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1870             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1871             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1872
1873             }
1874
1875             /**************************
1876              * CALCULATE INTERACTIONS *
1877              **************************/
1878
1879             if (gmx_mm_any_lt(rsq11,rcutoff2))
1880             {
1881
1882             /* REACTION-FIELD ELECTROSTATICS */
1883             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1884
1885             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1886
1887             fscal            = felec;
1888
1889             fscal            = _mm_and_ps(fscal,cutoff_mask);
1890
1891             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1892
1893              /* Update vectorial force */
1894             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1895             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1896             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1897
1898             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1899             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1900             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1901
1902             }
1903
1904             /**************************
1905              * CALCULATE INTERACTIONS *
1906              **************************/
1907
1908             if (gmx_mm_any_lt(rsq12,rcutoff2))
1909             {
1910
1911             /* REACTION-FIELD ELECTROSTATICS */
1912             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1913
1914             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1915
1916             fscal            = felec;
1917
1918             fscal            = _mm_and_ps(fscal,cutoff_mask);
1919
1920             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1921
1922              /* Update vectorial force */
1923             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1924             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1925             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1926
1927             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1928             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1929             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1930
1931             }
1932
1933             /**************************
1934              * CALCULATE INTERACTIONS *
1935              **************************/
1936
1937             if (gmx_mm_any_lt(rsq20,rcutoff2))
1938             {
1939
1940             /* REACTION-FIELD ELECTROSTATICS */
1941             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1942
1943             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1944
1945             fscal            = felec;
1946
1947             fscal            = _mm_and_ps(fscal,cutoff_mask);
1948
1949             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1950
1951              /* Update vectorial force */
1952             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1953             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1954             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1955
1956             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1957             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1958             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1959
1960             }
1961
1962             /**************************
1963              * CALCULATE INTERACTIONS *
1964              **************************/
1965
1966             if (gmx_mm_any_lt(rsq21,rcutoff2))
1967             {
1968
1969             /* REACTION-FIELD ELECTROSTATICS */
1970             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1971
1972             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1973
1974             fscal            = felec;
1975
1976             fscal            = _mm_and_ps(fscal,cutoff_mask);
1977
1978             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1979
1980              /* Update vectorial force */
1981             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1982             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1983             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1984
1985             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1986             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1987             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1988
1989             }
1990
1991             /**************************
1992              * CALCULATE INTERACTIONS *
1993              **************************/
1994
1995             if (gmx_mm_any_lt(rsq22,rcutoff2))
1996             {
1997
1998             /* REACTION-FIELD ELECTROSTATICS */
1999             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
2000
2001             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2002
2003             fscal            = felec;
2004
2005             fscal            = _mm_and_ps(fscal,cutoff_mask);
2006
2007             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2008
2009              /* Update vectorial force */
2010             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2011             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2012             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2013
2014             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2015             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2016             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2017
2018             }
2019
2020             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2021             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2022             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2023             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2024
2025             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2026                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2027
2028             /* Inner loop uses 329 flops */
2029         }
2030
2031         /* End of innermost loop */
2032
2033         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2034                                               f+i_coord_offset,fshift+i_shift_offset);
2035
2036         /* Increment number of inner iterations */
2037         inneriter                  += j_index_end - j_index_start;
2038
2039         /* Outer loop uses 18 flops */
2040     }
2041
2042     /* Increment number of outer iterations */
2043     outeriter        += nri;
2044
2045     /* Update outer/inner flops */
2046
2047     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*329);
2048 }