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

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  99     real             rswitch_scalar,d_scalar;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     krf              = _mm_set1_ps(fr->ic->k_rf);
 118     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 119     crf              = _mm_set1_ps(fr->ic->c_rf);
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 127     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 128     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 129     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 130
 131     jq0              = _mm_set1_ps(charge[inr+0]);
 132     jq1              = _mm_set1_ps(charge[inr+1]);
 133     jq2              = _mm_set1_ps(charge[inr+2]);
 134     vdwjidx0A        = 2*vdwtype[inr+0];
 135     qq00             = _mm_mul_ps(iq0,jq0);
 136     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 137     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 138     qq01             = _mm_mul_ps(iq0,jq1);
 139     qq02             = _mm_mul_ps(iq0,jq2);
 140     qq10             = _mm_mul_ps(iq1,jq0);
 141     qq11             = _mm_mul_ps(iq1,jq1);
 142     qq12             = _mm_mul_ps(iq1,jq2);
 143     qq20             = _mm_mul_ps(iq2,jq0);
 144     qq21             = _mm_mul_ps(iq2,jq1);
 145     qq22             = _mm_mul_ps(iq2,jq2);
 146
 147     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 148     rcutoff_scalar   = fr->rcoulomb;
 149     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 150     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 151
 152     rswitch_scalar   = fr->rvdw_switch;
 153     rswitch          = _mm_set1_ps(rswitch_scalar);
 154     /* Setup switch parameters */
 155     d_scalar         = rcutoff_scalar-rswitch_scalar;
 156     d                = _mm_set1_ps(d_scalar);
 157     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 158     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 159     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 160     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 161     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 162     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 163
 164     /* Avoid stupid compiler warnings */
 165     jnrA = jnrB = jnrC = jnrD = 0;
 166     j_coord_offsetA = 0;
 167     j_coord_offsetB = 0;
 168     j_coord_offsetC = 0;
 169     j_coord_offsetD = 0;
 170
 171     outeriter        = 0;
 172     inneriter        = 0;
 173
 174     for(iidx=0;iidx<4*DIM;iidx++)
 175     {
 176         scratch[iidx] = 0.0;
 177     }
 178
 179     /* Start outer loop over neighborlists */
 180     for(iidx=0; iidx<nri; iidx++)
 181     {
 182         /* Load shift vector for this list */
 183         i_shift_offset   = DIM*shiftidx[iidx];
 184
 185         /* Load limits for loop over neighbors */
 186         j_index_start    = jindex[iidx];
 187         j_index_end      = jindex[iidx+1];
 188
 189         /* Get outer coordinate index */
 190         inr              = iinr[iidx];
 191         i_coord_offset   = DIM*inr;
 192
 193         /* Load i particle coords and add shift vector */
 194         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 195                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 196
 197         fix0             = _mm_setzero_ps();
 198         fiy0             = _mm_setzero_ps();
 199         fiz0             = _mm_setzero_ps();
 200         fix1             = _mm_setzero_ps();
 201         fiy1             = _mm_setzero_ps();
 202         fiz1             = _mm_setzero_ps();
 203         fix2             = _mm_setzero_ps();
 204         fiy2             = _mm_setzero_ps();
 205         fiz2             = _mm_setzero_ps();
 206
 207         /* Reset potential sums */
 208         velecsum         = _mm_setzero_ps();
 209         vvdwsum          = _mm_setzero_ps();
 210
 211         /* Start inner kernel loop */
 212         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 213         {
 214
 215             /* Get j neighbor index, and coordinate index */
 216             jnrA             = jjnr[jidx];
 217             jnrB             = jjnr[jidx+1];
 218             jnrC             = jjnr[jidx+2];
 219             jnrD             = jjnr[jidx+3];
 220             j_coord_offsetA  = DIM*jnrA;
 221             j_coord_offsetB  = DIM*jnrB;
 222             j_coord_offsetC  = DIM*jnrC;
 223             j_coord_offsetD  = DIM*jnrD;
 224
 225             /* load j atom coordinates */
 226             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 227                                               x+j_coord_offsetC,x+j_coord_offsetD,
 228                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 229
 230             /* Calculate displacement vector */
 231             dx00             = _mm_sub_ps(ix0,jx0);
 232             dy00             = _mm_sub_ps(iy0,jy0);
 233             dz00             = _mm_sub_ps(iz0,jz0);
 234             dx01             = _mm_sub_ps(ix0,jx1);
 235             dy01             = _mm_sub_ps(iy0,jy1);
 236             dz01             = _mm_sub_ps(iz0,jz1);
 237             dx02             = _mm_sub_ps(ix0,jx2);
 238             dy02             = _mm_sub_ps(iy0,jy2);
 239             dz02             = _mm_sub_ps(iz0,jz2);
 240             dx10             = _mm_sub_ps(ix1,jx0);
 241             dy10             = _mm_sub_ps(iy1,jy0);
 242             dz10             = _mm_sub_ps(iz1,jz0);
 243             dx11             = _mm_sub_ps(ix1,jx1);
 244             dy11             = _mm_sub_ps(iy1,jy1);
 245             dz11             = _mm_sub_ps(iz1,jz1);
 246             dx12             = _mm_sub_ps(ix1,jx2);
 247             dy12             = _mm_sub_ps(iy1,jy2);
 248             dz12             = _mm_sub_ps(iz1,jz2);
 249             dx20             = _mm_sub_ps(ix2,jx0);
 250             dy20             = _mm_sub_ps(iy2,jy0);
 251             dz20             = _mm_sub_ps(iz2,jz0);
 252             dx21             = _mm_sub_ps(ix2,jx1);
 253             dy21             = _mm_sub_ps(iy2,jy1);
 254             dz21             = _mm_sub_ps(iz2,jz1);
 255             dx22             = _mm_sub_ps(ix2,jx2);
 256             dy22             = _mm_sub_ps(iy2,jy2);
 257             dz22             = _mm_sub_ps(iz2,jz2);
 258
 259             /* Calculate squared distance and things based on it */
 260             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 261             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 262             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 263             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 264             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 265             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 266             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 267             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 268             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 269
 270             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 271             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 272             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 273             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 274             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 275             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 276             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 277             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 278             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 279
 280             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 281             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 282             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 283             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 284             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 285             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 286             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 287             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 288             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 289
 290             fjx0             = _mm_setzero_ps();
 291             fjy0             = _mm_setzero_ps();
 292             fjz0             = _mm_setzero_ps();
 293             fjx1             = _mm_setzero_ps();
 294             fjy1             = _mm_setzero_ps();
 295             fjz1             = _mm_setzero_ps();
 296             fjx2             = _mm_setzero_ps();
 297             fjy2             = _mm_setzero_ps();
 298             fjz2             = _mm_setzero_ps();
 299
 300             /**************************
 301              * CALCULATE INTERACTIONS *
 302              **************************/
 303
 304             if (gmx_mm_any_lt(rsq00,rcutoff2))
 305             {
 306
 307             r00              = _mm_mul_ps(rsq00,rinv00);
 308
 309             /* REACTION-FIELD ELECTROSTATICS */
 310             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 311             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 312
 313             /* LENNARD-JONES DISPERSION/REPULSION */
 314
 315             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 316             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 317             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 318             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 319             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 320
 321             d                = _mm_sub_ps(r00,rswitch);
 322             d                = _mm_max_ps(d,_mm_setzero_ps());
 323             d2               = _mm_mul_ps(d,d);
 324             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 325
 326             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 327
 328             /* Evaluate switch function */
 329             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 330             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 331             vvdw             = _mm_mul_ps(vvdw,sw);
 332             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 333
 334             /* Update potential sum for this i atom from the interaction with this j atom. */
 335             velec            = _mm_and_ps(velec,cutoff_mask);
 336             velecsum         = _mm_add_ps(velecsum,velec);
 337             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 338             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 339
 340             fscal            = _mm_add_ps(felec,fvdw);
 341
 342             fscal            = _mm_and_ps(fscal,cutoff_mask);
 343
 344              /* Update vectorial force */
 345             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 346             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 347             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 348
 349             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 350             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 351             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 352
 353             }
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             if (gmx_mm_any_lt(rsq01,rcutoff2))
 360             {
 361
 362             /* REACTION-FIELD ELECTROSTATICS */
 363             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 364             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 365
 366             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velec            = _mm_and_ps(velec,cutoff_mask);
 370             velecsum         = _mm_add_ps(velecsum,velec);
 371
 372             fscal            = felec;
 373
 374             fscal            = _mm_and_ps(fscal,cutoff_mask);
 375
 376              /* Update vectorial force */
 377             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 378             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 379             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 380
 381             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 382             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 383             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 384
 385             }
 386
 387             /**************************
 388              * CALCULATE INTERACTIONS *
 389              **************************/
 390
 391             if (gmx_mm_any_lt(rsq02,rcutoff2))
 392             {
 393
 394             /* REACTION-FIELD ELECTROSTATICS */
 395             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 396             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 397
 398             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             velec            = _mm_and_ps(velec,cutoff_mask);
 402             velecsum         = _mm_add_ps(velecsum,velec);
 403
 404             fscal            = felec;
 405
 406             fscal            = _mm_and_ps(fscal,cutoff_mask);
 407
 408              /* Update vectorial force */
 409             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 410             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 411             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 412
 413             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 414             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 415             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 416
 417             }
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             if (gmx_mm_any_lt(rsq10,rcutoff2))
 424             {
 425
 426             /* REACTION-FIELD ELECTROSTATICS */
 427             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 428             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 429
 430             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velec            = _mm_and_ps(velec,cutoff_mask);
 434             velecsum         = _mm_add_ps(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             fscal            = _mm_and_ps(fscal,cutoff_mask);
 439
 440              /* Update vectorial force */
 441             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 442             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 443             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 444
 445             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 446             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 447             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 448
 449             }
 450
 451             /**************************
 452              * CALCULATE INTERACTIONS *
 453              **************************/
 454
 455             if (gmx_mm_any_lt(rsq11,rcutoff2))
 456             {
 457
 458             /* REACTION-FIELD ELECTROSTATICS */
 459             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 460             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 461
 462             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 463
 464             /* Update potential sum for this i atom from the interaction with this j atom. */
 465             velec            = _mm_and_ps(velec,cutoff_mask);
 466             velecsum         = _mm_add_ps(velecsum,velec);
 467
 468             fscal            = felec;
 469
 470             fscal            = _mm_and_ps(fscal,cutoff_mask);
 471
 472              /* Update vectorial force */
 473             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 474             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 475             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 476
 477             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 478             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 479             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 480
 481             }
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             if (gmx_mm_any_lt(rsq12,rcutoff2))
 488             {
 489
 490             /* REACTION-FIELD ELECTROSTATICS */
 491             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 492             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 493
 494             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velec            = _mm_and_ps(velec,cutoff_mask);
 498             velecsum         = _mm_add_ps(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502             fscal            = _mm_and_ps(fscal,cutoff_mask);
 503
 504              /* Update vectorial force */
 505             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 506             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 507             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 508
 509             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 510             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 511             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 512
 513             }
 514
 515             /**************************
 516              * CALCULATE INTERACTIONS *
 517              **************************/
 518
 519             if (gmx_mm_any_lt(rsq20,rcutoff2))
 520             {
 521
 522             /* REACTION-FIELD ELECTROSTATICS */
 523             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 524             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 525
 526             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 527
 528             /* Update potential sum for this i atom from the interaction with this j atom. */
 529             velec            = _mm_and_ps(velec,cutoff_mask);
 530             velecsum         = _mm_add_ps(velecsum,velec);
 531
 532             fscal            = felec;
 533
 534             fscal            = _mm_and_ps(fscal,cutoff_mask);
 535
 536              /* Update vectorial force */
 537             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 538             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 539             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 540
 541             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 542             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 543             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 544
 545             }
 546
 547             /**************************
 548              * CALCULATE INTERACTIONS *
 549              **************************/
 550
 551             if (gmx_mm_any_lt(rsq21,rcutoff2))
 552             {
 553
 554             /* REACTION-FIELD ELECTROSTATICS */
 555             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 556             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 557
 558             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 559
 560             /* Update potential sum for this i atom from the interaction with this j atom. */
 561             velec            = _mm_and_ps(velec,cutoff_mask);
 562             velecsum         = _mm_add_ps(velecsum,velec);
 563
 564             fscal            = felec;
 565
 566             fscal            = _mm_and_ps(fscal,cutoff_mask);
 567
 568              /* Update vectorial force */
 569             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 570             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 571             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 572
 573             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 574             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 575             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 576
 577             }
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             if (gmx_mm_any_lt(rsq22,rcutoff2))
 584             {
 585
 586             /* REACTION-FIELD ELECTROSTATICS */
 587             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 588             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 589
 590             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 591
 592             /* Update potential sum for this i atom from the interaction with this j atom. */
 593             velec            = _mm_and_ps(velec,cutoff_mask);
 594             velecsum         = _mm_add_ps(velecsum,velec);
 595
 596             fscal            = felec;
 597
 598             fscal            = _mm_and_ps(fscal,cutoff_mask);
 599
 600              /* Update vectorial force */
 601             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 602             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 603             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 604
 605             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 606             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 607             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 608
 609             }
 610
 611             fjptrA             = f+j_coord_offsetA;
 612             fjptrB             = f+j_coord_offsetB;
 613             fjptrC             = f+j_coord_offsetC;
 614             fjptrD             = f+j_coord_offsetD;
 615
 616             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 617                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 618
 619             /* Inner loop uses 385 flops */
 620         }
 621
 622         if(jidx<j_index_end)
 623         {
 624
 625             /* Get j neighbor index, and coordinate index */
 626             jnrlistA         = jjnr[jidx];
 627             jnrlistB         = jjnr[jidx+1];
 628             jnrlistC         = jjnr[jidx+2];
 629             jnrlistD         = jjnr[jidx+3];
 630             /* Sign of each element will be negative for non-real atoms.
 631              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 632              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 633              */
 634             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 635             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 636             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 637             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 638             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 639             j_coord_offsetA  = DIM*jnrA;
 640             j_coord_offsetB  = DIM*jnrB;
 641             j_coord_offsetC  = DIM*jnrC;
 642             j_coord_offsetD  = DIM*jnrD;
 643
 644             /* load j atom coordinates */
 645             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 646                                               x+j_coord_offsetC,x+j_coord_offsetD,
 647                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 648
 649             /* Calculate displacement vector */
 650             dx00             = _mm_sub_ps(ix0,jx0);
 651             dy00             = _mm_sub_ps(iy0,jy0);
 652             dz00             = _mm_sub_ps(iz0,jz0);
 653             dx01             = _mm_sub_ps(ix0,jx1);
 654             dy01             = _mm_sub_ps(iy0,jy1);
 655             dz01             = _mm_sub_ps(iz0,jz1);
 656             dx02             = _mm_sub_ps(ix0,jx2);
 657             dy02             = _mm_sub_ps(iy0,jy2);
 658             dz02             = _mm_sub_ps(iz0,jz2);
 659             dx10             = _mm_sub_ps(ix1,jx0);
 660             dy10             = _mm_sub_ps(iy1,jy0);
 661             dz10             = _mm_sub_ps(iz1,jz0);
 662             dx11             = _mm_sub_ps(ix1,jx1);
 663             dy11             = _mm_sub_ps(iy1,jy1);
 664             dz11             = _mm_sub_ps(iz1,jz1);
 665             dx12             = _mm_sub_ps(ix1,jx2);
 666             dy12             = _mm_sub_ps(iy1,jy2);
 667             dz12             = _mm_sub_ps(iz1,jz2);
 668             dx20             = _mm_sub_ps(ix2,jx0);
 669             dy20             = _mm_sub_ps(iy2,jy0);
 670             dz20             = _mm_sub_ps(iz2,jz0);
 671             dx21             = _mm_sub_ps(ix2,jx1);
 672             dy21             = _mm_sub_ps(iy2,jy1);
 673             dz21             = _mm_sub_ps(iz2,jz1);
 674             dx22             = _mm_sub_ps(ix2,jx2);
 675             dy22             = _mm_sub_ps(iy2,jy2);
 676             dz22             = _mm_sub_ps(iz2,jz2);
 677
 678             /* Calculate squared distance and things based on it */
 679             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 680             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 681             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 682             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 683             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 684             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 685             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 686             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 687             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 688
 689             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 690             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 691             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 692             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 693             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 694             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 695             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 696             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 697             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 698
 699             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 700             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 701             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 702             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 703             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 704             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 705             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 706             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 707             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 708
 709             fjx0             = _mm_setzero_ps();
 710             fjy0             = _mm_setzero_ps();
 711             fjz0             = _mm_setzero_ps();
 712             fjx1             = _mm_setzero_ps();
 713             fjy1             = _mm_setzero_ps();
 714             fjz1             = _mm_setzero_ps();
 715             fjx2             = _mm_setzero_ps();
 716             fjy2             = _mm_setzero_ps();
 717             fjz2             = _mm_setzero_ps();
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             if (gmx_mm_any_lt(rsq00,rcutoff2))
 724             {
 725
 726             r00              = _mm_mul_ps(rsq00,rinv00);
 727             r00              = _mm_andnot_ps(dummy_mask,r00);
 728
 729             /* REACTION-FIELD ELECTROSTATICS */
 730             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 731             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 732
 733             /* LENNARD-JONES DISPERSION/REPULSION */
 734
 735             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 736             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 737             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 738             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 739             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 740
 741             d                = _mm_sub_ps(r00,rswitch);
 742             d                = _mm_max_ps(d,_mm_setzero_ps());
 743             d2               = _mm_mul_ps(d,d);
 744             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 745
 746             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 747
 748             /* Evaluate switch function */
 749             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 750             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 751             vvdw             = _mm_mul_ps(vvdw,sw);
 752             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 753
 754             /* Update potential sum for this i atom from the interaction with this j atom. */
 755             velec            = _mm_and_ps(velec,cutoff_mask);
 756             velec            = _mm_andnot_ps(dummy_mask,velec);
 757             velecsum         = _mm_add_ps(velecsum,velec);
 758             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 759             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 760             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 761
 762             fscal            = _mm_add_ps(felec,fvdw);
 763
 764             fscal            = _mm_and_ps(fscal,cutoff_mask);
 765
 766             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 767
 768              /* Update vectorial force */
 769             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 770             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 771             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 772
 773             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 774             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 775             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 776
 777             }
 778
 779             /**************************
 780              * CALCULATE INTERACTIONS *
 781              **************************/
 782
 783             if (gmx_mm_any_lt(rsq01,rcutoff2))
 784             {
 785
 786             /* REACTION-FIELD ELECTROSTATICS */
 787             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 788             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 789
 790             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 791
 792             /* Update potential sum for this i atom from the interaction with this j atom. */
 793             velec            = _mm_and_ps(velec,cutoff_mask);
 794             velec            = _mm_andnot_ps(dummy_mask,velec);
 795             velecsum         = _mm_add_ps(velecsum,velec);
 796
 797             fscal            = felec;
 798
 799             fscal            = _mm_and_ps(fscal,cutoff_mask);
 800
 801             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 802
 803              /* Update vectorial force */
 804             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 805             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 806             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 807
 808             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 809             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 810             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 811
 812             }
 813
 814             /**************************
 815              * CALCULATE INTERACTIONS *
 816              **************************/
 817
 818             if (gmx_mm_any_lt(rsq02,rcutoff2))
 819             {
 820
 821             /* REACTION-FIELD ELECTROSTATICS */
 822             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 823             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 824
 825             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 826
 827             /* Update potential sum for this i atom from the interaction with this j atom. */
 828             velec            = _mm_and_ps(velec,cutoff_mask);
 829             velec            = _mm_andnot_ps(dummy_mask,velec);
 830             velecsum         = _mm_add_ps(velecsum,velec);
 831
 832             fscal            = felec;
 833
 834             fscal            = _mm_and_ps(fscal,cutoff_mask);
 835
 836             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 837
 838              /* Update vectorial force */
 839             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 840             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 841             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 842
 843             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 844             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 845             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 846
 847             }
 848
 849             /**************************
 850              * CALCULATE INTERACTIONS *
 851              **************************/
 852
 853             if (gmx_mm_any_lt(rsq10,rcutoff2))
 854             {
 855
 856             /* REACTION-FIELD ELECTROSTATICS */
 857             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 858             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 859
 860             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 861
 862             /* Update potential sum for this i atom from the interaction with this j atom. */
 863             velec            = _mm_and_ps(velec,cutoff_mask);
 864             velec            = _mm_andnot_ps(dummy_mask,velec);
 865             velecsum         = _mm_add_ps(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_and_ps(fscal,cutoff_mask);
 870
 871             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 872
 873              /* Update vectorial force */
 874             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 875             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 876             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 877
 878             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 879             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 880             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 881
 882             }
 883
 884             /**************************
 885              * CALCULATE INTERACTIONS *
 886              **************************/
 887
 888             if (gmx_mm_any_lt(rsq11,rcutoff2))
 889             {
 890
 891             /* REACTION-FIELD ELECTROSTATICS */
 892             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 893             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 894
 895             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 896
 897             /* Update potential sum for this i atom from the interaction with this j atom. */
 898             velec            = _mm_and_ps(velec,cutoff_mask);
 899             velec            = _mm_andnot_ps(dummy_mask,velec);
 900             velecsum         = _mm_add_ps(velecsum,velec);
 901
 902             fscal            = felec;
 903
 904             fscal            = _mm_and_ps(fscal,cutoff_mask);
 905
 906             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 907
 908              /* Update vectorial force */
 909             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 910             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 911             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 912
 913             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 914             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 915             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 916
 917             }
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             if (gmx_mm_any_lt(rsq12,rcutoff2))
 924             {
 925
 926             /* REACTION-FIELD ELECTROSTATICS */
 927             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 928             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 929
 930             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 931
 932             /* Update potential sum for this i atom from the interaction with this j atom. */
 933             velec            = _mm_and_ps(velec,cutoff_mask);
 934             velec            = _mm_andnot_ps(dummy_mask,velec);
 935             velecsum         = _mm_add_ps(velecsum,velec);
 936
 937             fscal            = felec;
 938
 939             fscal            = _mm_and_ps(fscal,cutoff_mask);
 940
 941             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 942
 943              /* Update vectorial force */
 944             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 945             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 946             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 947
 948             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 949             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 950             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 951
 952             }
 953
 954             /**************************
 955              * CALCULATE INTERACTIONS *
 956              **************************/
 957
 958             if (gmx_mm_any_lt(rsq20,rcutoff2))
 959             {
 960
 961             /* REACTION-FIELD ELECTROSTATICS */
 962             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 963             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 964
 965             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 966
 967             /* Update potential sum for this i atom from the interaction with this j atom. */
 968             velec            = _mm_and_ps(velec,cutoff_mask);
 969             velec            = _mm_andnot_ps(dummy_mask,velec);
 970             velecsum         = _mm_add_ps(velecsum,velec);
 971
 972             fscal            = felec;
 973
 974             fscal            = _mm_and_ps(fscal,cutoff_mask);
 975
 976             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 977
 978              /* Update vectorial force */
 979             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 980             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 981             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 982
 983             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 984             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 985             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 986
 987             }
 988
 989             /**************************
 990              * CALCULATE INTERACTIONS *
 991              **************************/
 992
 993             if (gmx_mm_any_lt(rsq21,rcutoff2))
 994             {
 995
 996             /* REACTION-FIELD ELECTROSTATICS */
 997             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 998             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 999
1000             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1001
1002             /* Update potential sum for this i atom from the interaction with this j atom. */
1003             velec            = _mm_and_ps(velec,cutoff_mask);
1004             velec            = _mm_andnot_ps(dummy_mask,velec);
1005             velecsum         = _mm_add_ps(velecsum,velec);
1006
1007             fscal            = felec;
1008
1009             fscal            = _mm_and_ps(fscal,cutoff_mask);
1010
1011             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1012
1013              /* Update vectorial force */
1014             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1015             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1016             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1017
1018             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1019             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1020             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1021
1022             }
1023
1024             /**************************
1025              * CALCULATE INTERACTIONS *
1026              **************************/
1027
1028             if (gmx_mm_any_lt(rsq22,rcutoff2))
1029             {
1030
1031             /* REACTION-FIELD ELECTROSTATICS */
1032             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
1033             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1034
1035             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1036
1037             /* Update potential sum for this i atom from the interaction with this j atom. */
1038             velec            = _mm_and_ps(velec,cutoff_mask);
1039             velec            = _mm_andnot_ps(dummy_mask,velec);
1040             velecsum         = _mm_add_ps(velecsum,velec);
1041
1042             fscal            = felec;
1043
1044             fscal            = _mm_and_ps(fscal,cutoff_mask);
1045
1046             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1047
1048              /* Update vectorial force */
1049             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1050             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1051             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1052
1053             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1054             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1055             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1056
1057             }
1058
1059             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1060             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1061             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1062             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1063
1064             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1065                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1066
1067             /* Inner loop uses 386 flops */
1068         }
1069
1070         /* End of innermost loop */
1071
1072         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1073                                               f+i_coord_offset,fshift+i_shift_offset);
1074
1075         ggid                        = gid[iidx];
1076         /* Update potential energies */
1077         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1078         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1079
1080         /* Increment number of inner iterations */
1081         inneriter                  += j_index_end - j_index_start;
1082
1083         /* Outer loop uses 20 flops */
1084     }
1085
1086     /* Increment number of outer iterations */
1087     outeriter        += nri;
1088
1089     /* Update outer/inner flops */
1090
1091     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*386);
1092 }
1093 /*
1094  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_single
1095  * Electrostatics interaction: ReactionField
1096  * VdW interaction:            LennardJones
1097  * Geometry:                   Water3-Water3
1098  * Calculate force/pot:        Force
1099  */
1100 void
1101 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_single
1102                     (t_nblist * gmx_restrict                nlist,
1103                      rvec * gmx_restrict                    xx,
1104                      rvec * gmx_restrict                    ff,
1105                      t_forcerec * gmx_restrict              fr,
1106                      t_mdatoms * gmx_restrict               mdatoms,
1107                      nb_kernel_data_t * gmx_restrict        kernel_data,
1108                      t_nrnb * gmx_restrict                  nrnb)
1109 {
1110     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1111      * just 0 for non-waters.
1112      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1113      * jnr indices corresponding to data put in the four positions in the SIMD register.
1114      */
1115     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1116     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1117     int              jnrA,jnrB,jnrC,jnrD;
1118     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1119     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1120     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1121     real             rcutoff_scalar;
1122     real             *shiftvec,*fshift,*x,*f;
1123     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1124     real             scratch[4*DIM];
1125     __m128           fscal,rcutoff,rcutoff2,jidxall;
1126     int              vdwioffset0;
1127     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1128     int              vdwioffset1;
1129     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1130     int              vdwioffset2;
1131     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1132     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1133     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1134     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1135     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1136     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1137     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1138     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1139     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1140     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1141     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1142     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1143     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1144     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1145     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1146     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1147     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1148     real             *charge;
1149     int              nvdwtype;
1150     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1151     int              *vdwtype;
1152     real             *vdwparam;
1153     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1154     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1155     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1156     real             rswitch_scalar,d_scalar;
1157     __m128           dummy_mask,cutoff_mask;
1158     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1159     __m128           one     = _mm_set1_ps(1.0);
1160     __m128           two     = _mm_set1_ps(2.0);
1161     x                = xx[0];
1162     f                = ff[0];
1163
1164     nri              = nlist->nri;
1165     iinr             = nlist->iinr;
1166     jindex           = nlist->jindex;
1167     jjnr             = nlist->jjnr;
1168     shiftidx         = nlist->shift;
1169     gid              = nlist->gid;
1170     shiftvec         = fr->shift_vec[0];
1171     fshift           = fr->fshift[0];
1172     facel            = _mm_set1_ps(fr->epsfac);
1173     charge           = mdatoms->chargeA;
1174     krf              = _mm_set1_ps(fr->ic->k_rf);
1175     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1176     crf              = _mm_set1_ps(fr->ic->c_rf);
1177     nvdwtype         = fr->ntype;
1178     vdwparam         = fr->nbfp;
1179     vdwtype          = mdatoms->typeA;
1180
1181     /* Setup water-specific parameters */
1182     inr              = nlist->iinr[0];
1183     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1184     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1185     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1186     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1187
1188     jq0              = _mm_set1_ps(charge[inr+0]);
1189     jq1              = _mm_set1_ps(charge[inr+1]);
1190     jq2              = _mm_set1_ps(charge[inr+2]);
1191     vdwjidx0A        = 2*vdwtype[inr+0];
1192     qq00             = _mm_mul_ps(iq0,jq0);
1193     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1194     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1195     qq01             = _mm_mul_ps(iq0,jq1);
1196     qq02             = _mm_mul_ps(iq0,jq2);
1197     qq10             = _mm_mul_ps(iq1,jq0);
1198     qq11             = _mm_mul_ps(iq1,jq1);
1199     qq12             = _mm_mul_ps(iq1,jq2);
1200     qq20             = _mm_mul_ps(iq2,jq0);
1201     qq21             = _mm_mul_ps(iq2,jq1);
1202     qq22             = _mm_mul_ps(iq2,jq2);
1203
1204     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1205     rcutoff_scalar   = fr->rcoulomb;
1206     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1207     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1208
1209     rswitch_scalar   = fr->rvdw_switch;
1210     rswitch          = _mm_set1_ps(rswitch_scalar);
1211     /* Setup switch parameters */
1212     d_scalar         = rcutoff_scalar-rswitch_scalar;
1213     d                = _mm_set1_ps(d_scalar);
1214     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
1215     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1216     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1217     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
1218     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1219     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1220
1221     /* Avoid stupid compiler warnings */
1222     jnrA = jnrB = jnrC = jnrD = 0;
1223     j_coord_offsetA = 0;
1224     j_coord_offsetB = 0;
1225     j_coord_offsetC = 0;
1226     j_coord_offsetD = 0;
1227
1228     outeriter        = 0;
1229     inneriter        = 0;
1230
1231     for(iidx=0;iidx<4*DIM;iidx++)
1232     {
1233         scratch[iidx] = 0.0;
1234     }
1235
1236     /* Start outer loop over neighborlists */
1237     for(iidx=0; iidx<nri; iidx++)
1238     {
1239         /* Load shift vector for this list */
1240         i_shift_offset   = DIM*shiftidx[iidx];
1241
1242         /* Load limits for loop over neighbors */
1243         j_index_start    = jindex[iidx];
1244         j_index_end      = jindex[iidx+1];
1245
1246         /* Get outer coordinate index */
1247         inr              = iinr[iidx];
1248         i_coord_offset   = DIM*inr;
1249
1250         /* Load i particle coords and add shift vector */
1251         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1252                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1253
1254         fix0             = _mm_setzero_ps();
1255         fiy0             = _mm_setzero_ps();
1256         fiz0             = _mm_setzero_ps();
1257         fix1             = _mm_setzero_ps();
1258         fiy1             = _mm_setzero_ps();
1259         fiz1             = _mm_setzero_ps();
1260         fix2             = _mm_setzero_ps();
1261         fiy2             = _mm_setzero_ps();
1262         fiz2             = _mm_setzero_ps();
1263
1264         /* Start inner kernel loop */
1265         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1266         {
1267
1268             /* Get j neighbor index, and coordinate index */
1269             jnrA             = jjnr[jidx];
1270             jnrB             = jjnr[jidx+1];
1271             jnrC             = jjnr[jidx+2];
1272             jnrD             = jjnr[jidx+3];
1273             j_coord_offsetA  = DIM*jnrA;
1274             j_coord_offsetB  = DIM*jnrB;
1275             j_coord_offsetC  = DIM*jnrC;
1276             j_coord_offsetD  = DIM*jnrD;
1277
1278             /* load j atom coordinates */
1279             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1280                                               x+j_coord_offsetC,x+j_coord_offsetD,
1281                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1282
1283             /* Calculate displacement vector */
1284             dx00             = _mm_sub_ps(ix0,jx0);
1285             dy00             = _mm_sub_ps(iy0,jy0);
1286             dz00             = _mm_sub_ps(iz0,jz0);
1287             dx01             = _mm_sub_ps(ix0,jx1);
1288             dy01             = _mm_sub_ps(iy0,jy1);
1289             dz01             = _mm_sub_ps(iz0,jz1);
1290             dx02             = _mm_sub_ps(ix0,jx2);
1291             dy02             = _mm_sub_ps(iy0,jy2);
1292             dz02             = _mm_sub_ps(iz0,jz2);
1293             dx10             = _mm_sub_ps(ix1,jx0);
1294             dy10             = _mm_sub_ps(iy1,jy0);
1295             dz10             = _mm_sub_ps(iz1,jz0);
1296             dx11             = _mm_sub_ps(ix1,jx1);
1297             dy11             = _mm_sub_ps(iy1,jy1);
1298             dz11             = _mm_sub_ps(iz1,jz1);
1299             dx12             = _mm_sub_ps(ix1,jx2);
1300             dy12             = _mm_sub_ps(iy1,jy2);
1301             dz12             = _mm_sub_ps(iz1,jz2);
1302             dx20             = _mm_sub_ps(ix2,jx0);
1303             dy20             = _mm_sub_ps(iy2,jy0);
1304             dz20             = _mm_sub_ps(iz2,jz0);
1305             dx21             = _mm_sub_ps(ix2,jx1);
1306             dy21             = _mm_sub_ps(iy2,jy1);
1307             dz21             = _mm_sub_ps(iz2,jz1);
1308             dx22             = _mm_sub_ps(ix2,jx2);
1309             dy22             = _mm_sub_ps(iy2,jy2);
1310             dz22             = _mm_sub_ps(iz2,jz2);
1311
1312             /* Calculate squared distance and things based on it */
1313             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1314             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1315             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1316             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1317             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1318             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1319             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1320             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1321             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1322
1323             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1324             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1325             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1326             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1327             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1328             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1329             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1330             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1331             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1332
1333             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1334             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1335             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1336             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1337             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1338             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1339             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1340             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1341             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1342
1343             fjx0             = _mm_setzero_ps();
1344             fjy0             = _mm_setzero_ps();
1345             fjz0             = _mm_setzero_ps();
1346             fjx1             = _mm_setzero_ps();
1347             fjy1             = _mm_setzero_ps();
1348             fjz1             = _mm_setzero_ps();
1349             fjx2             = _mm_setzero_ps();
1350             fjy2             = _mm_setzero_ps();
1351             fjz2             = _mm_setzero_ps();
1352
1353             /**************************
1354              * CALCULATE INTERACTIONS *
1355              **************************/
1356
1357             if (gmx_mm_any_lt(rsq00,rcutoff2))
1358             {
1359
1360             r00              = _mm_mul_ps(rsq00,rinv00);
1361
1362             /* REACTION-FIELD ELECTROSTATICS */
1363             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1364
1365             /* LENNARD-JONES DISPERSION/REPULSION */
1366
1367             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1368             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1369             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1370             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1371             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1372
1373             d                = _mm_sub_ps(r00,rswitch);
1374             d                = _mm_max_ps(d,_mm_setzero_ps());
1375             d2               = _mm_mul_ps(d,d);
1376             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1377
1378             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1379
1380             /* Evaluate switch function */
1381             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1382             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1383             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1384
1385             fscal            = _mm_add_ps(felec,fvdw);
1386
1387             fscal            = _mm_and_ps(fscal,cutoff_mask);
1388
1389              /* Update vectorial force */
1390             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1391             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1392             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1393
1394             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1395             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1396             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1397
1398             }
1399
1400             /**************************
1401              * CALCULATE INTERACTIONS *
1402              **************************/
1403
1404             if (gmx_mm_any_lt(rsq01,rcutoff2))
1405             {
1406
1407             /* REACTION-FIELD ELECTROSTATICS */
1408             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1409
1410             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1411
1412             fscal            = felec;
1413
1414             fscal            = _mm_and_ps(fscal,cutoff_mask);
1415
1416              /* Update vectorial force */
1417             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1418             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1419             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1420
1421             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1422             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1423             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1424
1425             }
1426
1427             /**************************
1428              * CALCULATE INTERACTIONS *
1429              **************************/
1430
1431             if (gmx_mm_any_lt(rsq02,rcutoff2))
1432             {
1433
1434             /* REACTION-FIELD ELECTROSTATICS */
1435             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1436
1437             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1438
1439             fscal            = felec;
1440
1441             fscal            = _mm_and_ps(fscal,cutoff_mask);
1442
1443              /* Update vectorial force */
1444             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1445             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1446             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1447
1448             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1449             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1450             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1451
1452             }
1453
1454             /**************************
1455              * CALCULATE INTERACTIONS *
1456              **************************/
1457
1458             if (gmx_mm_any_lt(rsq10,rcutoff2))
1459             {
1460
1461             /* REACTION-FIELD ELECTROSTATICS */
1462             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1463
1464             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1465
1466             fscal            = felec;
1467
1468             fscal            = _mm_and_ps(fscal,cutoff_mask);
1469
1470              /* Update vectorial force */
1471             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1472             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1473             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1474
1475             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1476             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1477             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1478
1479             }
1480
1481             /**************************
1482              * CALCULATE INTERACTIONS *
1483              **************************/
1484
1485             if (gmx_mm_any_lt(rsq11,rcutoff2))
1486             {
1487
1488             /* REACTION-FIELD ELECTROSTATICS */
1489             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1490
1491             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1492
1493             fscal            = felec;
1494
1495             fscal            = _mm_and_ps(fscal,cutoff_mask);
1496
1497              /* Update vectorial force */
1498             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1499             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1500             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1501
1502             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1503             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1504             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1505
1506             }
1507
1508             /**************************
1509              * CALCULATE INTERACTIONS *
1510              **************************/
1511
1512             if (gmx_mm_any_lt(rsq12,rcutoff2))
1513             {
1514
1515             /* REACTION-FIELD ELECTROSTATICS */
1516             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1517
1518             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1519
1520             fscal            = felec;
1521
1522             fscal            = _mm_and_ps(fscal,cutoff_mask);
1523
1524              /* Update vectorial force */
1525             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1526             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1527             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1528
1529             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1530             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1531             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1532
1533             }
1534
1535             /**************************
1536              * CALCULATE INTERACTIONS *
1537              **************************/
1538
1539             if (gmx_mm_any_lt(rsq20,rcutoff2))
1540             {
1541
1542             /* REACTION-FIELD ELECTROSTATICS */
1543             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1544
1545             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1546
1547             fscal            = felec;
1548
1549             fscal            = _mm_and_ps(fscal,cutoff_mask);
1550
1551              /* Update vectorial force */
1552             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1553             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1554             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1555
1556             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1557             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1558             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1559
1560             }
1561
1562             /**************************
1563              * CALCULATE INTERACTIONS *
1564              **************************/
1565
1566             if (gmx_mm_any_lt(rsq21,rcutoff2))
1567             {
1568
1569             /* REACTION-FIELD ELECTROSTATICS */
1570             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1571
1572             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1573
1574             fscal            = felec;
1575
1576             fscal            = _mm_and_ps(fscal,cutoff_mask);
1577
1578              /* Update vectorial force */
1579             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1580             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1581             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1582
1583             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1584             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1585             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1586
1587             }
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             if (gmx_mm_any_lt(rsq22,rcutoff2))
1594             {
1595
1596             /* REACTION-FIELD ELECTROSTATICS */
1597             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1598
1599             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1600
1601             fscal            = felec;
1602
1603             fscal            = _mm_and_ps(fscal,cutoff_mask);
1604
1605              /* Update vectorial force */
1606             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1607             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1608             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1609
1610             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1611             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1612             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1613
1614             }
1615
1616             fjptrA             = f+j_coord_offsetA;
1617             fjptrB             = f+j_coord_offsetB;
1618             fjptrC             = f+j_coord_offsetC;
1619             fjptrD             = f+j_coord_offsetD;
1620
1621             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1622                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1623
1624             /* Inner loop uses 328 flops */
1625         }
1626
1627         if(jidx<j_index_end)
1628         {
1629
1630             /* Get j neighbor index, and coordinate index */
1631             jnrlistA         = jjnr[jidx];
1632             jnrlistB         = jjnr[jidx+1];
1633             jnrlistC         = jjnr[jidx+2];
1634             jnrlistD         = jjnr[jidx+3];
1635             /* Sign of each element will be negative for non-real atoms.
1636              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1637              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1638              */
1639             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1640             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1641             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1642             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1643             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1644             j_coord_offsetA  = DIM*jnrA;
1645             j_coord_offsetB  = DIM*jnrB;
1646             j_coord_offsetC  = DIM*jnrC;
1647             j_coord_offsetD  = DIM*jnrD;
1648
1649             /* load j atom coordinates */
1650             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1651                                               x+j_coord_offsetC,x+j_coord_offsetD,
1652                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1653
1654             /* Calculate displacement vector */
1655             dx00             = _mm_sub_ps(ix0,jx0);
1656             dy00             = _mm_sub_ps(iy0,jy0);
1657             dz00             = _mm_sub_ps(iz0,jz0);
1658             dx01             = _mm_sub_ps(ix0,jx1);
1659             dy01             = _mm_sub_ps(iy0,jy1);
1660             dz01             = _mm_sub_ps(iz0,jz1);
1661             dx02             = _mm_sub_ps(ix0,jx2);
1662             dy02             = _mm_sub_ps(iy0,jy2);
1663             dz02             = _mm_sub_ps(iz0,jz2);
1664             dx10             = _mm_sub_ps(ix1,jx0);
1665             dy10             = _mm_sub_ps(iy1,jy0);
1666             dz10             = _mm_sub_ps(iz1,jz0);
1667             dx11             = _mm_sub_ps(ix1,jx1);
1668             dy11             = _mm_sub_ps(iy1,jy1);
1669             dz11             = _mm_sub_ps(iz1,jz1);
1670             dx12             = _mm_sub_ps(ix1,jx2);
1671             dy12             = _mm_sub_ps(iy1,jy2);
1672             dz12             = _mm_sub_ps(iz1,jz2);
1673             dx20             = _mm_sub_ps(ix2,jx0);
1674             dy20             = _mm_sub_ps(iy2,jy0);
1675             dz20             = _mm_sub_ps(iz2,jz0);
1676             dx21             = _mm_sub_ps(ix2,jx1);
1677             dy21             = _mm_sub_ps(iy2,jy1);
1678             dz21             = _mm_sub_ps(iz2,jz1);
1679             dx22             = _mm_sub_ps(ix2,jx2);
1680             dy22             = _mm_sub_ps(iy2,jy2);
1681             dz22             = _mm_sub_ps(iz2,jz2);
1682
1683             /* Calculate squared distance and things based on it */
1684             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1685             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1686             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1687             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1688             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1689             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1690             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1691             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1692             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1693
1694             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1695             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1696             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1697             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1698             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1699             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1700             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1701             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1702             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1703
1704             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1705             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1706             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1707             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1708             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1709             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1710             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1711             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1712             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1713
1714             fjx0             = _mm_setzero_ps();
1715             fjy0             = _mm_setzero_ps();
1716             fjz0             = _mm_setzero_ps();
1717             fjx1             = _mm_setzero_ps();
1718             fjy1             = _mm_setzero_ps();
1719             fjz1             = _mm_setzero_ps();
1720             fjx2             = _mm_setzero_ps();
1721             fjy2             = _mm_setzero_ps();
1722             fjz2             = _mm_setzero_ps();
1723
1724             /**************************
1725              * CALCULATE INTERACTIONS *
1726              **************************/
1727
1728             if (gmx_mm_any_lt(rsq00,rcutoff2))
1729             {
1730
1731             r00              = _mm_mul_ps(rsq00,rinv00);
1732             r00              = _mm_andnot_ps(dummy_mask,r00);
1733
1734             /* REACTION-FIELD ELECTROSTATICS */
1735             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1736
1737             /* LENNARD-JONES DISPERSION/REPULSION */
1738
1739             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1740             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1741             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1742             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1743             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1744
1745             d                = _mm_sub_ps(r00,rswitch);
1746             d                = _mm_max_ps(d,_mm_setzero_ps());
1747             d2               = _mm_mul_ps(d,d);
1748             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1749
1750             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1751
1752             /* Evaluate switch function */
1753             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1754             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1755             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1756
1757             fscal            = _mm_add_ps(felec,fvdw);
1758
1759             fscal            = _mm_and_ps(fscal,cutoff_mask);
1760
1761             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1762
1763              /* Update vectorial force */
1764             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1765             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1766             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1767
1768             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1769             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1770             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1771
1772             }
1773
1774             /**************************
1775              * CALCULATE INTERACTIONS *
1776              **************************/
1777
1778             if (gmx_mm_any_lt(rsq01,rcutoff2))
1779             {
1780
1781             /* REACTION-FIELD ELECTROSTATICS */
1782             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1783
1784             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1785
1786             fscal            = felec;
1787
1788             fscal            = _mm_and_ps(fscal,cutoff_mask);
1789
1790             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1791
1792              /* Update vectorial force */
1793             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1794             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1795             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1796
1797             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1798             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1799             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1800
1801             }
1802
1803             /**************************
1804              * CALCULATE INTERACTIONS *
1805              **************************/
1806
1807             if (gmx_mm_any_lt(rsq02,rcutoff2))
1808             {
1809
1810             /* REACTION-FIELD ELECTROSTATICS */
1811             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1812
1813             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1814
1815             fscal            = felec;
1816
1817             fscal            = _mm_and_ps(fscal,cutoff_mask);
1818
1819             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1820
1821              /* Update vectorial force */
1822             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1823             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1824             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1825
1826             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1827             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1828             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1829
1830             }
1831
1832             /**************************
1833              * CALCULATE INTERACTIONS *
1834              **************************/
1835
1836             if (gmx_mm_any_lt(rsq10,rcutoff2))
1837             {
1838
1839             /* REACTION-FIELD ELECTROSTATICS */
1840             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1841
1842             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1843
1844             fscal            = felec;
1845
1846             fscal            = _mm_and_ps(fscal,cutoff_mask);
1847
1848             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1849
1850              /* Update vectorial force */
1851             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1852             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1853             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1854
1855             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1856             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1857             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1858
1859             }
1860
1861             /**************************
1862              * CALCULATE INTERACTIONS *
1863              **************************/
1864
1865             if (gmx_mm_any_lt(rsq11,rcutoff2))
1866             {
1867
1868             /* REACTION-FIELD ELECTROSTATICS */
1869             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1870
1871             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1872
1873             fscal            = felec;
1874
1875             fscal            = _mm_and_ps(fscal,cutoff_mask);
1876
1877             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1878
1879              /* Update vectorial force */
1880             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1881             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1882             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1883
1884             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1885             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1886             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1887
1888             }
1889
1890             /**************************
1891              * CALCULATE INTERACTIONS *
1892              **************************/
1893
1894             if (gmx_mm_any_lt(rsq12,rcutoff2))
1895             {
1896
1897             /* REACTION-FIELD ELECTROSTATICS */
1898             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1899
1900             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1901
1902             fscal            = felec;
1903
1904             fscal            = _mm_and_ps(fscal,cutoff_mask);
1905
1906             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1907
1908              /* Update vectorial force */
1909             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1910             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1911             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1912
1913             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1914             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1915             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1916
1917             }
1918
1919             /**************************
1920              * CALCULATE INTERACTIONS *
1921              **************************/
1922
1923             if (gmx_mm_any_lt(rsq20,rcutoff2))
1924             {
1925
1926             /* REACTION-FIELD ELECTROSTATICS */
1927             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1928
1929             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1930
1931             fscal            = felec;
1932
1933             fscal            = _mm_and_ps(fscal,cutoff_mask);
1934
1935             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1936
1937              /* Update vectorial force */
1938             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1939             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1940             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1941
1942             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1943             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1944             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1945
1946             }
1947
1948             /**************************
1949              * CALCULATE INTERACTIONS *
1950              **************************/
1951
1952             if (gmx_mm_any_lt(rsq21,rcutoff2))
1953             {
1954
1955             /* REACTION-FIELD ELECTROSTATICS */
1956             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1957
1958             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1959
1960             fscal            = felec;
1961
1962             fscal            = _mm_and_ps(fscal,cutoff_mask);
1963
1964             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1965
1966              /* Update vectorial force */
1967             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1968             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1969             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1970
1971             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1972             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1973             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1974
1975             }
1976
1977             /**************************
1978              * CALCULATE INTERACTIONS *
1979              **************************/
1980
1981             if (gmx_mm_any_lt(rsq22,rcutoff2))
1982             {
1983
1984             /* REACTION-FIELD ELECTROSTATICS */
1985             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1986
1987             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1988
1989             fscal            = felec;
1990
1991             fscal            = _mm_and_ps(fscal,cutoff_mask);
1992
1993             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1994
1995              /* Update vectorial force */
1996             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1997             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1998             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1999
2000             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2001             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2002             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2003
2004             }
2005
2006             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2007             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2008             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2009             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2010
2011             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2012                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2013
2014             /* Inner loop uses 329 flops */
2015         }
2016
2017         /* End of innermost loop */
2018
2019         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2020                                               f+i_coord_offset,fshift+i_shift_offset);
2021
2022         /* Increment number of inner iterations */
2023         inneriter                  += j_index_end - j_index_start;
2024
2025         /* Outer loop uses 18 flops */
2026     }
2027
2028     /* Increment number of outer iterations */
2029     outeriter        += nri;
2030
2031     /* Update outer/inner flops */
2032
2033     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*329);
2034 }