src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_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_GeomW4W4_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_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              vdwioffset3;
  76     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  77     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  78     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  79     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  80     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  81     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  82     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  83     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  84     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  85     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  86     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  87     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  88     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  89     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  90     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  91     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  92     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  93     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  94     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  95     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 102     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 103     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 104     real             rswitch_scalar,d_scalar;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_ps(fr->ic->k_rf);
 123     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_ps(fr->ic->c_rf);
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 132     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 133     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 134     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 135
 136     jq1              = _mm_set1_ps(charge[inr+1]);
 137     jq2              = _mm_set1_ps(charge[inr+2]);
 138     jq3              = _mm_set1_ps(charge[inr+3]);
 139     vdwjidx0A        = 2*vdwtype[inr+0];
 140     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 141     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 142     qq11             = _mm_mul_ps(iq1,jq1);
 143     qq12             = _mm_mul_ps(iq1,jq2);
 144     qq13             = _mm_mul_ps(iq1,jq3);
 145     qq21             = _mm_mul_ps(iq2,jq1);
 146     qq22             = _mm_mul_ps(iq2,jq2);
 147     qq23             = _mm_mul_ps(iq2,jq3);
 148     qq31             = _mm_mul_ps(iq3,jq1);
 149     qq32             = _mm_mul_ps(iq3,jq2);
 150     qq33             = _mm_mul_ps(iq3,jq3);
 151
 152     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 153     rcutoff_scalar   = fr->rcoulomb;
 154     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 155     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 156
 157     rswitch_scalar   = fr->rvdw_switch;
 158     rswitch          = _mm_set1_ps(rswitch_scalar);
 159     /* Setup switch parameters */
 160     d_scalar         = rcutoff_scalar-rswitch_scalar;
 161     d                = _mm_set1_ps(d_scalar);
 162     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 163     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 164     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 165     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 166     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 167     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 168
 169     /* Avoid stupid compiler warnings */
 170     jnrA = jnrB = jnrC = jnrD = 0;
 171     j_coord_offsetA = 0;
 172     j_coord_offsetB = 0;
 173     j_coord_offsetC = 0;
 174     j_coord_offsetD = 0;
 175
 176     outeriter        = 0;
 177     inneriter        = 0;
 178
 179     for(iidx=0;iidx<4*DIM;iidx++)
 180     {
 181         scratch[iidx] = 0.0;
 182     }
 183
 184     /* Start outer loop over neighborlists */
 185     for(iidx=0; iidx<nri; iidx++)
 186     {
 187         /* Load shift vector for this list */
 188         i_shift_offset   = DIM*shiftidx[iidx];
 189
 190         /* Load limits for loop over neighbors */
 191         j_index_start    = jindex[iidx];
 192         j_index_end      = jindex[iidx+1];
 193
 194         /* Get outer coordinate index */
 195         inr              = iinr[iidx];
 196         i_coord_offset   = DIM*inr;
 197
 198         /* Load i particle coords and add shift vector */
 199         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 200                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 201
 202         fix0             = _mm_setzero_ps();
 203         fiy0             = _mm_setzero_ps();
 204         fiz0             = _mm_setzero_ps();
 205         fix1             = _mm_setzero_ps();
 206         fiy1             = _mm_setzero_ps();
 207         fiz1             = _mm_setzero_ps();
 208         fix2             = _mm_setzero_ps();
 209         fiy2             = _mm_setzero_ps();
 210         fiz2             = _mm_setzero_ps();
 211         fix3             = _mm_setzero_ps();
 212         fiy3             = _mm_setzero_ps();
 213         fiz3             = _mm_setzero_ps();
 214
 215         /* Reset potential sums */
 216         velecsum         = _mm_setzero_ps();
 217         vvdwsum          = _mm_setzero_ps();
 218
 219         /* Start inner kernel loop */
 220         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 221         {
 222
 223             /* Get j neighbor index, and coordinate index */
 224             jnrA             = jjnr[jidx];
 225             jnrB             = jjnr[jidx+1];
 226             jnrC             = jjnr[jidx+2];
 227             jnrD             = jjnr[jidx+3];
 228             j_coord_offsetA  = DIM*jnrA;
 229             j_coord_offsetB  = DIM*jnrB;
 230             j_coord_offsetC  = DIM*jnrC;
 231             j_coord_offsetD  = DIM*jnrD;
 232
 233             /* load j atom coordinates */
 234             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 235                                               x+j_coord_offsetC,x+j_coord_offsetD,
 236                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 237                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 238
 239             /* Calculate displacement vector */
 240             dx00             = _mm_sub_ps(ix0,jx0);
 241             dy00             = _mm_sub_ps(iy0,jy0);
 242             dz00             = _mm_sub_ps(iz0,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             dx13             = _mm_sub_ps(ix1,jx3);
 250             dy13             = _mm_sub_ps(iy1,jy3);
 251             dz13             = _mm_sub_ps(iz1,jz3);
 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             dx23             = _mm_sub_ps(ix2,jx3);
 259             dy23             = _mm_sub_ps(iy2,jy3);
 260             dz23             = _mm_sub_ps(iz2,jz3);
 261             dx31             = _mm_sub_ps(ix3,jx1);
 262             dy31             = _mm_sub_ps(iy3,jy1);
 263             dz31             = _mm_sub_ps(iz3,jz1);
 264             dx32             = _mm_sub_ps(ix3,jx2);
 265             dy32             = _mm_sub_ps(iy3,jy2);
 266             dz32             = _mm_sub_ps(iz3,jz2);
 267             dx33             = _mm_sub_ps(ix3,jx3);
 268             dy33             = _mm_sub_ps(iy3,jy3);
 269             dz33             = _mm_sub_ps(iz3,jz3);
 270
 271             /* Calculate squared distance and things based on it */
 272             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 273             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 274             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 275             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 276             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 277             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 278             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 279             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 280             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 281             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 282
 283             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 284             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 285             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 286             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 287             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 288             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 289             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 290             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 291             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 292             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 293
 294             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 295             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 296             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 297             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 298             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 299             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 300             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 301             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 302             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 303             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 304
 305             fjx0             = _mm_setzero_ps();
 306             fjy0             = _mm_setzero_ps();
 307             fjz0             = _mm_setzero_ps();
 308             fjx1             = _mm_setzero_ps();
 309             fjy1             = _mm_setzero_ps();
 310             fjz1             = _mm_setzero_ps();
 311             fjx2             = _mm_setzero_ps();
 312             fjy2             = _mm_setzero_ps();
 313             fjz2             = _mm_setzero_ps();
 314             fjx3             = _mm_setzero_ps();
 315             fjy3             = _mm_setzero_ps();
 316             fjz3             = _mm_setzero_ps();
 317
 318             /**************************
 319              * CALCULATE INTERACTIONS *
 320              **************************/
 321
 322             if (gmx_mm_any_lt(rsq00,rcutoff2))
 323             {
 324
 325             r00              = _mm_mul_ps(rsq00,rinv00);
 326
 327             /* LENNARD-JONES DISPERSION/REPULSION */
 328
 329             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 330             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 331             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 332             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 333             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 334
 335             d                = _mm_sub_ps(r00,rswitch);
 336             d                = _mm_max_ps(d,_mm_setzero_ps());
 337             d2               = _mm_mul_ps(d,d);
 338             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 339
 340             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 341
 342             /* Evaluate switch function */
 343             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 344             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 345             vvdw             = _mm_mul_ps(vvdw,sw);
 346             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 347
 348             /* Update potential sum for this i atom from the interaction with this j atom. */
 349             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 350             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 351
 352             fscal            = fvdw;
 353
 354             fscal            = _mm_and_ps(fscal,cutoff_mask);
 355
 356              /* Update vectorial force */
 357             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 358             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 359             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 360
 361             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 362             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 363             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 364
 365             }
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             if (gmx_mm_any_lt(rsq11,rcutoff2))
 372             {
 373
 374             /* REACTION-FIELD ELECTROSTATICS */
 375             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 376             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 377
 378             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 379
 380             /* Update potential sum for this i atom from the interaction with this j atom. */
 381             velec            = _mm_and_ps(velec,cutoff_mask);
 382             velecsum         = _mm_add_ps(velecsum,velec);
 383
 384             fscal            = felec;
 385
 386             fscal            = _mm_and_ps(fscal,cutoff_mask);
 387
 388              /* Update vectorial force */
 389             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 390             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 391             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 392
 393             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 394             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 395             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 396
 397             }
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             if (gmx_mm_any_lt(rsq12,rcutoff2))
 404             {
 405
 406             /* REACTION-FIELD ELECTROSTATICS */
 407             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 408             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 409
 410             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 411
 412             /* Update potential sum for this i atom from the interaction with this j atom. */
 413             velec            = _mm_and_ps(velec,cutoff_mask);
 414             velecsum         = _mm_add_ps(velecsum,velec);
 415
 416             fscal            = felec;
 417
 418             fscal            = _mm_and_ps(fscal,cutoff_mask);
 419
 420              /* Update vectorial force */
 421             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 422             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 423             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 424
 425             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 426             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 427             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 428
 429             }
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             if (gmx_mm_any_lt(rsq13,rcutoff2))
 436             {
 437
 438             /* REACTION-FIELD ELECTROSTATICS */
 439             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 440             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 441
 442             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 443
 444             /* Update potential sum for this i atom from the interaction with this j atom. */
 445             velec            = _mm_and_ps(velec,cutoff_mask);
 446             velecsum         = _mm_add_ps(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             fscal            = _mm_and_ps(fscal,cutoff_mask);
 451
 452              /* Update vectorial force */
 453             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 454             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 455             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 456
 457             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 458             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 459             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 460
 461             }
 462
 463             /**************************
 464              * CALCULATE INTERACTIONS *
 465              **************************/
 466
 467             if (gmx_mm_any_lt(rsq21,rcutoff2))
 468             {
 469
 470             /* REACTION-FIELD ELECTROSTATICS */
 471             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 472             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 473
 474             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 475
 476             /* Update potential sum for this i atom from the interaction with this j atom. */
 477             velec            = _mm_and_ps(velec,cutoff_mask);
 478             velecsum         = _mm_add_ps(velecsum,velec);
 479
 480             fscal            = felec;
 481
 482             fscal            = _mm_and_ps(fscal,cutoff_mask);
 483
 484              /* Update vectorial force */
 485             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 486             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 487             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 488
 489             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 490             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 491             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 492
 493             }
 494
 495             /**************************
 496              * CALCULATE INTERACTIONS *
 497              **************************/
 498
 499             if (gmx_mm_any_lt(rsq22,rcutoff2))
 500             {
 501
 502             /* REACTION-FIELD ELECTROSTATICS */
 503             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 504             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 505
 506             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 507
 508             /* Update potential sum for this i atom from the interaction with this j atom. */
 509             velec            = _mm_and_ps(velec,cutoff_mask);
 510             velecsum         = _mm_add_ps(velecsum,velec);
 511
 512             fscal            = felec;
 513
 514             fscal            = _mm_and_ps(fscal,cutoff_mask);
 515
 516              /* Update vectorial force */
 517             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 518             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 519             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 520
 521             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 522             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 523             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 524
 525             }
 526
 527             /**************************
 528              * CALCULATE INTERACTIONS *
 529              **************************/
 530
 531             if (gmx_mm_any_lt(rsq23,rcutoff2))
 532             {
 533
 534             /* REACTION-FIELD ELECTROSTATICS */
 535             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 536             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 537
 538             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 539
 540             /* Update potential sum for this i atom from the interaction with this j atom. */
 541             velec            = _mm_and_ps(velec,cutoff_mask);
 542             velecsum         = _mm_add_ps(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546             fscal            = _mm_and_ps(fscal,cutoff_mask);
 547
 548              /* Update vectorial force */
 549             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 550             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 551             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 552
 553             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 554             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 555             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 556
 557             }
 558
 559             /**************************
 560              * CALCULATE INTERACTIONS *
 561              **************************/
 562
 563             if (gmx_mm_any_lt(rsq31,rcutoff2))
 564             {
 565
 566             /* REACTION-FIELD ELECTROSTATICS */
 567             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 568             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 569
 570             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 571
 572             /* Update potential sum for this i atom from the interaction with this j atom. */
 573             velec            = _mm_and_ps(velec,cutoff_mask);
 574             velecsum         = _mm_add_ps(velecsum,velec);
 575
 576             fscal            = felec;
 577
 578             fscal            = _mm_and_ps(fscal,cutoff_mask);
 579
 580              /* Update vectorial force */
 581             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 582             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 583             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 584
 585             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 586             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 587             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 588
 589             }
 590
 591             /**************************
 592              * CALCULATE INTERACTIONS *
 593              **************************/
 594
 595             if (gmx_mm_any_lt(rsq32,rcutoff2))
 596             {
 597
 598             /* REACTION-FIELD ELECTROSTATICS */
 599             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 600             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 601
 602             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 603
 604             /* Update potential sum for this i atom from the interaction with this j atom. */
 605             velec            = _mm_and_ps(velec,cutoff_mask);
 606             velecsum         = _mm_add_ps(velecsum,velec);
 607
 608             fscal            = felec;
 609
 610             fscal            = _mm_and_ps(fscal,cutoff_mask);
 611
 612              /* Update vectorial force */
 613             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 614             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 615             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 616
 617             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 618             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 619             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 620
 621             }
 622
 623             /**************************
 624              * CALCULATE INTERACTIONS *
 625              **************************/
 626
 627             if (gmx_mm_any_lt(rsq33,rcutoff2))
 628             {
 629
 630             /* REACTION-FIELD ELECTROSTATICS */
 631             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 632             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 633
 634             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 635
 636             /* Update potential sum for this i atom from the interaction with this j atom. */
 637             velec            = _mm_and_ps(velec,cutoff_mask);
 638             velecsum         = _mm_add_ps(velecsum,velec);
 639
 640             fscal            = felec;
 641
 642             fscal            = _mm_and_ps(fscal,cutoff_mask);
 643
 644              /* Update vectorial force */
 645             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 646             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 647             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 648
 649             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 650             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 651             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 652
 653             }
 654
 655             fjptrA             = f+j_coord_offsetA;
 656             fjptrB             = f+j_coord_offsetB;
 657             fjptrC             = f+j_coord_offsetC;
 658             fjptrD             = f+j_coord_offsetD;
 659
 660             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 661                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 662                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 663
 664             /* Inner loop uses 416 flops */
 665         }
 666
 667         if(jidx<j_index_end)
 668         {
 669
 670             /* Get j neighbor index, and coordinate index */
 671             jnrlistA         = jjnr[jidx];
 672             jnrlistB         = jjnr[jidx+1];
 673             jnrlistC         = jjnr[jidx+2];
 674             jnrlistD         = jjnr[jidx+3];
 675             /* Sign of each element will be negative for non-real atoms.
 676              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 677              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 678              */
 679             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 680             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 681             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 682             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 683             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 684             j_coord_offsetA  = DIM*jnrA;
 685             j_coord_offsetB  = DIM*jnrB;
 686             j_coord_offsetC  = DIM*jnrC;
 687             j_coord_offsetD  = DIM*jnrD;
 688
 689             /* load j atom coordinates */
 690             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 691                                               x+j_coord_offsetC,x+j_coord_offsetD,
 692                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 693                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 694
 695             /* Calculate displacement vector */
 696             dx00             = _mm_sub_ps(ix0,jx0);
 697             dy00             = _mm_sub_ps(iy0,jy0);
 698             dz00             = _mm_sub_ps(iz0,jz0);
 699             dx11             = _mm_sub_ps(ix1,jx1);
 700             dy11             = _mm_sub_ps(iy1,jy1);
 701             dz11             = _mm_sub_ps(iz1,jz1);
 702             dx12             = _mm_sub_ps(ix1,jx2);
 703             dy12             = _mm_sub_ps(iy1,jy2);
 704             dz12             = _mm_sub_ps(iz1,jz2);
 705             dx13             = _mm_sub_ps(ix1,jx3);
 706             dy13             = _mm_sub_ps(iy1,jy3);
 707             dz13             = _mm_sub_ps(iz1,jz3);
 708             dx21             = _mm_sub_ps(ix2,jx1);
 709             dy21             = _mm_sub_ps(iy2,jy1);
 710             dz21             = _mm_sub_ps(iz2,jz1);
 711             dx22             = _mm_sub_ps(ix2,jx2);
 712             dy22             = _mm_sub_ps(iy2,jy2);
 713             dz22             = _mm_sub_ps(iz2,jz2);
 714             dx23             = _mm_sub_ps(ix2,jx3);
 715             dy23             = _mm_sub_ps(iy2,jy3);
 716             dz23             = _mm_sub_ps(iz2,jz3);
 717             dx31             = _mm_sub_ps(ix3,jx1);
 718             dy31             = _mm_sub_ps(iy3,jy1);
 719             dz31             = _mm_sub_ps(iz3,jz1);
 720             dx32             = _mm_sub_ps(ix3,jx2);
 721             dy32             = _mm_sub_ps(iy3,jy2);
 722             dz32             = _mm_sub_ps(iz3,jz2);
 723             dx33             = _mm_sub_ps(ix3,jx3);
 724             dy33             = _mm_sub_ps(iy3,jy3);
 725             dz33             = _mm_sub_ps(iz3,jz3);
 726
 727             /* Calculate squared distance and things based on it */
 728             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 729             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 730             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 731             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 732             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 733             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 734             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 735             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 736             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 737             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 738
 739             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 740             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 741             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 742             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 743             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 744             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 745             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 746             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 747             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 748             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 749
 750             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 751             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 752             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 753             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 754             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 755             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 756             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 757             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 758             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 759             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 760
 761             fjx0             = _mm_setzero_ps();
 762             fjy0             = _mm_setzero_ps();
 763             fjz0             = _mm_setzero_ps();
 764             fjx1             = _mm_setzero_ps();
 765             fjy1             = _mm_setzero_ps();
 766             fjz1             = _mm_setzero_ps();
 767             fjx2             = _mm_setzero_ps();
 768             fjy2             = _mm_setzero_ps();
 769             fjz2             = _mm_setzero_ps();
 770             fjx3             = _mm_setzero_ps();
 771             fjy3             = _mm_setzero_ps();
 772             fjz3             = _mm_setzero_ps();
 773
 774             /**************************
 775              * CALCULATE INTERACTIONS *
 776              **************************/
 777
 778             if (gmx_mm_any_lt(rsq00,rcutoff2))
 779             {
 780
 781             r00              = _mm_mul_ps(rsq00,rinv00);
 782             r00              = _mm_andnot_ps(dummy_mask,r00);
 783
 784             /* LENNARD-JONES DISPERSION/REPULSION */
 785
 786             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 787             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 788             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 789             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 790             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 791
 792             d                = _mm_sub_ps(r00,rswitch);
 793             d                = _mm_max_ps(d,_mm_setzero_ps());
 794             d2               = _mm_mul_ps(d,d);
 795             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 796
 797             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 798
 799             /* Evaluate switch function */
 800             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 801             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 802             vvdw             = _mm_mul_ps(vvdw,sw);
 803             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 804
 805             /* Update potential sum for this i atom from the interaction with this j atom. */
 806             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 807             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 808             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 809
 810             fscal            = fvdw;
 811
 812             fscal            = _mm_and_ps(fscal,cutoff_mask);
 813
 814             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 815
 816              /* Update vectorial force */
 817             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 818             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 819             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 820
 821             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 822             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 823             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 824
 825             }
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             if (gmx_mm_any_lt(rsq11,rcutoff2))
 832             {
 833
 834             /* REACTION-FIELD ELECTROSTATICS */
 835             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 836             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 837
 838             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 839
 840             /* Update potential sum for this i atom from the interaction with this j atom. */
 841             velec            = _mm_and_ps(velec,cutoff_mask);
 842             velec            = _mm_andnot_ps(dummy_mask,velec);
 843             velecsum         = _mm_add_ps(velecsum,velec);
 844
 845             fscal            = felec;
 846
 847             fscal            = _mm_and_ps(fscal,cutoff_mask);
 848
 849             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 850
 851              /* Update vectorial force */
 852             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 853             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 854             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 855
 856             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 857             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 858             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 859
 860             }
 861
 862             /**************************
 863              * CALCULATE INTERACTIONS *
 864              **************************/
 865
 866             if (gmx_mm_any_lt(rsq12,rcutoff2))
 867             {
 868
 869             /* REACTION-FIELD ELECTROSTATICS */
 870             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 871             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 872
 873             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 874
 875             /* Update potential sum for this i atom from the interaction with this j atom. */
 876             velec            = _mm_and_ps(velec,cutoff_mask);
 877             velec            = _mm_andnot_ps(dummy_mask,velec);
 878             velecsum         = _mm_add_ps(velecsum,velec);
 879
 880             fscal            = felec;
 881
 882             fscal            = _mm_and_ps(fscal,cutoff_mask);
 883
 884             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 885
 886              /* Update vectorial force */
 887             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 888             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 889             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 890
 891             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 892             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 893             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 894
 895             }
 896
 897             /**************************
 898              * CALCULATE INTERACTIONS *
 899              **************************/
 900
 901             if (gmx_mm_any_lt(rsq13,rcutoff2))
 902             {
 903
 904             /* REACTION-FIELD ELECTROSTATICS */
 905             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 906             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 907
 908             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 909
 910             /* Update potential sum for this i atom from the interaction with this j atom. */
 911             velec            = _mm_and_ps(velec,cutoff_mask);
 912             velec            = _mm_andnot_ps(dummy_mask,velec);
 913             velecsum         = _mm_add_ps(velecsum,velec);
 914
 915             fscal            = felec;
 916
 917             fscal            = _mm_and_ps(fscal,cutoff_mask);
 918
 919             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 920
 921              /* Update vectorial force */
 922             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 923             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 924             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 925
 926             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 927             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 928             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 929
 930             }
 931
 932             /**************************
 933              * CALCULATE INTERACTIONS *
 934              **************************/
 935
 936             if (gmx_mm_any_lt(rsq21,rcutoff2))
 937             {
 938
 939             /* REACTION-FIELD ELECTROSTATICS */
 940             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 941             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 942
 943             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 944
 945             /* Update potential sum for this i atom from the interaction with this j atom. */
 946             velec            = _mm_and_ps(velec,cutoff_mask);
 947             velec            = _mm_andnot_ps(dummy_mask,velec);
 948             velecsum         = _mm_add_ps(velecsum,velec);
 949
 950             fscal            = felec;
 951
 952             fscal            = _mm_and_ps(fscal,cutoff_mask);
 953
 954             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 955
 956              /* Update vectorial force */
 957             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 958             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 959             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 960
 961             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 962             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 963             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 964
 965             }
 966
 967             /**************************
 968              * CALCULATE INTERACTIONS *
 969              **************************/
 970
 971             if (gmx_mm_any_lt(rsq22,rcutoff2))
 972             {
 973
 974             /* REACTION-FIELD ELECTROSTATICS */
 975             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 976             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 977
 978             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 979
 980             /* Update potential sum for this i atom from the interaction with this j atom. */
 981             velec            = _mm_and_ps(velec,cutoff_mask);
 982             velec            = _mm_andnot_ps(dummy_mask,velec);
 983             velecsum         = _mm_add_ps(velecsum,velec);
 984
 985             fscal            = felec;
 986
 987             fscal            = _mm_and_ps(fscal,cutoff_mask);
 988
 989             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 990
 991              /* Update vectorial force */
 992             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 993             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 994             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 995
 996             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 997             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 998             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 999
1000             }
1001
1002             /**************************
1003              * CALCULATE INTERACTIONS *
1004              **************************/
1005
1006             if (gmx_mm_any_lt(rsq23,rcutoff2))
1007             {
1008
1009             /* REACTION-FIELD ELECTROSTATICS */
1010             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
1011             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1012
1013             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1014
1015             /* Update potential sum for this i atom from the interaction with this j atom. */
1016             velec            = _mm_and_ps(velec,cutoff_mask);
1017             velec            = _mm_andnot_ps(dummy_mask,velec);
1018             velecsum         = _mm_add_ps(velecsum,velec);
1019
1020             fscal            = felec;
1021
1022             fscal            = _mm_and_ps(fscal,cutoff_mask);
1023
1024             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1025
1026              /* Update vectorial force */
1027             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1028             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1029             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1030
1031             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1032             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1033             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1034
1035             }
1036
1037             /**************************
1038              * CALCULATE INTERACTIONS *
1039              **************************/
1040
1041             if (gmx_mm_any_lt(rsq31,rcutoff2))
1042             {
1043
1044             /* REACTION-FIELD ELECTROSTATICS */
1045             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
1046             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1047
1048             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1049
1050             /* Update potential sum for this i atom from the interaction with this j atom. */
1051             velec            = _mm_and_ps(velec,cutoff_mask);
1052             velec            = _mm_andnot_ps(dummy_mask,velec);
1053             velecsum         = _mm_add_ps(velecsum,velec);
1054
1055             fscal            = felec;
1056
1057             fscal            = _mm_and_ps(fscal,cutoff_mask);
1058
1059             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1060
1061              /* Update vectorial force */
1062             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1063             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1064             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1065
1066             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1067             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1068             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1069
1070             }
1071
1072             /**************************
1073              * CALCULATE INTERACTIONS *
1074              **************************/
1075
1076             if (gmx_mm_any_lt(rsq32,rcutoff2))
1077             {
1078
1079             /* REACTION-FIELD ELECTROSTATICS */
1080             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
1081             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1082
1083             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1084
1085             /* Update potential sum for this i atom from the interaction with this j atom. */
1086             velec            = _mm_and_ps(velec,cutoff_mask);
1087             velec            = _mm_andnot_ps(dummy_mask,velec);
1088             velecsum         = _mm_add_ps(velecsum,velec);
1089
1090             fscal            = felec;
1091
1092             fscal            = _mm_and_ps(fscal,cutoff_mask);
1093
1094             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1095
1096              /* Update vectorial force */
1097             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1098             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1099             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1100
1101             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1102             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1103             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1104
1105             }
1106
1107             /**************************
1108              * CALCULATE INTERACTIONS *
1109              **************************/
1110
1111             if (gmx_mm_any_lt(rsq33,rcutoff2))
1112             {
1113
1114             /* REACTION-FIELD ELECTROSTATICS */
1115             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
1116             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1117
1118             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1119
1120             /* Update potential sum for this i atom from the interaction with this j atom. */
1121             velec            = _mm_and_ps(velec,cutoff_mask);
1122             velec            = _mm_andnot_ps(dummy_mask,velec);
1123             velecsum         = _mm_add_ps(velecsum,velec);
1124
1125             fscal            = felec;
1126
1127             fscal            = _mm_and_ps(fscal,cutoff_mask);
1128
1129             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1130
1131              /* Update vectorial force */
1132             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1133             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1134             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1135
1136             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1137             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1138             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1139
1140             }
1141
1142             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1143             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1144             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1145             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1146
1147             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1148                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1149                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1150
1151             /* Inner loop uses 417 flops */
1152         }
1153
1154         /* End of innermost loop */
1155
1156         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1157                                               f+i_coord_offset,fshift+i_shift_offset);
1158
1159         ggid                        = gid[iidx];
1160         /* Update potential energies */
1161         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1162         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1163
1164         /* Increment number of inner iterations */
1165         inneriter                  += j_index_end - j_index_start;
1166
1167         /* Outer loop uses 26 flops */
1168     }
1169
1170     /* Increment number of outer iterations */
1171     outeriter        += nri;
1172
1173     /* Update outer/inner flops */
1174
1175     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*417);
1176 }
1177 /*
1178  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_single
1179  * Electrostatics interaction: ReactionField
1180  * VdW interaction:            LennardJones
1181  * Geometry:                   Water4-Water4
1182  * Calculate force/pot:        Force
1183  */
1184 void
1185 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_single
1186                     (t_nblist * gmx_restrict                nlist,
1187                      rvec * gmx_restrict                    xx,
1188                      rvec * gmx_restrict                    ff,
1189                      t_forcerec * gmx_restrict              fr,
1190                      t_mdatoms * gmx_restrict               mdatoms,
1191                      nb_kernel_data_t * gmx_restrict        kernel_data,
1192                      t_nrnb * gmx_restrict                  nrnb)
1193 {
1194     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1195      * just 0 for non-waters.
1196      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1197      * jnr indices corresponding to data put in the four positions in the SIMD register.
1198      */
1199     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1200     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1201     int              jnrA,jnrB,jnrC,jnrD;
1202     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1203     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1204     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1205     real             rcutoff_scalar;
1206     real             *shiftvec,*fshift,*x,*f;
1207     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1208     real             scratch[4*DIM];
1209     __m128           fscal,rcutoff,rcutoff2,jidxall;
1210     int              vdwioffset0;
1211     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1212     int              vdwioffset1;
1213     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1214     int              vdwioffset2;
1215     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1216     int              vdwioffset3;
1217     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1218     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1219     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1220     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1221     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1222     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1223     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1224     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1225     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1226     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1227     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1228     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1229     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1230     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1231     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1232     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1233     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1234     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1235     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1236     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1237     real             *charge;
1238     int              nvdwtype;
1239     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1240     int              *vdwtype;
1241     real             *vdwparam;
1242     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1243     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1244     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1245     real             rswitch_scalar,d_scalar;
1246     __m128           dummy_mask,cutoff_mask;
1247     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1248     __m128           one     = _mm_set1_ps(1.0);
1249     __m128           two     = _mm_set1_ps(2.0);
1250     x                = xx[0];
1251     f                = ff[0];
1252
1253     nri              = nlist->nri;
1254     iinr             = nlist->iinr;
1255     jindex           = nlist->jindex;
1256     jjnr             = nlist->jjnr;
1257     shiftidx         = nlist->shift;
1258     gid              = nlist->gid;
1259     shiftvec         = fr->shift_vec[0];
1260     fshift           = fr->fshift[0];
1261     facel            = _mm_set1_ps(fr->epsfac);
1262     charge           = mdatoms->chargeA;
1263     krf              = _mm_set1_ps(fr->ic->k_rf);
1264     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1265     crf              = _mm_set1_ps(fr->ic->c_rf);
1266     nvdwtype         = fr->ntype;
1267     vdwparam         = fr->nbfp;
1268     vdwtype          = mdatoms->typeA;
1269
1270     /* Setup water-specific parameters */
1271     inr              = nlist->iinr[0];
1272     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1273     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1274     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1275     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1276
1277     jq1              = _mm_set1_ps(charge[inr+1]);
1278     jq2              = _mm_set1_ps(charge[inr+2]);
1279     jq3              = _mm_set1_ps(charge[inr+3]);
1280     vdwjidx0A        = 2*vdwtype[inr+0];
1281     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1282     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1283     qq11             = _mm_mul_ps(iq1,jq1);
1284     qq12             = _mm_mul_ps(iq1,jq2);
1285     qq13             = _mm_mul_ps(iq1,jq3);
1286     qq21             = _mm_mul_ps(iq2,jq1);
1287     qq22             = _mm_mul_ps(iq2,jq2);
1288     qq23             = _mm_mul_ps(iq2,jq3);
1289     qq31             = _mm_mul_ps(iq3,jq1);
1290     qq32             = _mm_mul_ps(iq3,jq2);
1291     qq33             = _mm_mul_ps(iq3,jq3);
1292
1293     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1294     rcutoff_scalar   = fr->rcoulomb;
1295     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1296     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1297
1298     rswitch_scalar   = fr->rvdw_switch;
1299     rswitch          = _mm_set1_ps(rswitch_scalar);
1300     /* Setup switch parameters */
1301     d_scalar         = rcutoff_scalar-rswitch_scalar;
1302     d                = _mm_set1_ps(d_scalar);
1303     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
1304     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1305     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1306     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
1307     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1308     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1309
1310     /* Avoid stupid compiler warnings */
1311     jnrA = jnrB = jnrC = jnrD = 0;
1312     j_coord_offsetA = 0;
1313     j_coord_offsetB = 0;
1314     j_coord_offsetC = 0;
1315     j_coord_offsetD = 0;
1316
1317     outeriter        = 0;
1318     inneriter        = 0;
1319
1320     for(iidx=0;iidx<4*DIM;iidx++)
1321     {
1322         scratch[iidx] = 0.0;
1323     }
1324
1325     /* Start outer loop over neighborlists */
1326     for(iidx=0; iidx<nri; iidx++)
1327     {
1328         /* Load shift vector for this list */
1329         i_shift_offset   = DIM*shiftidx[iidx];
1330
1331         /* Load limits for loop over neighbors */
1332         j_index_start    = jindex[iidx];
1333         j_index_end      = jindex[iidx+1];
1334
1335         /* Get outer coordinate index */
1336         inr              = iinr[iidx];
1337         i_coord_offset   = DIM*inr;
1338
1339         /* Load i particle coords and add shift vector */
1340         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1341                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1342
1343         fix0             = _mm_setzero_ps();
1344         fiy0             = _mm_setzero_ps();
1345         fiz0             = _mm_setzero_ps();
1346         fix1             = _mm_setzero_ps();
1347         fiy1             = _mm_setzero_ps();
1348         fiz1             = _mm_setzero_ps();
1349         fix2             = _mm_setzero_ps();
1350         fiy2             = _mm_setzero_ps();
1351         fiz2             = _mm_setzero_ps();
1352         fix3             = _mm_setzero_ps();
1353         fiy3             = _mm_setzero_ps();
1354         fiz3             = _mm_setzero_ps();
1355
1356         /* Start inner kernel loop */
1357         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1358         {
1359
1360             /* Get j neighbor index, and coordinate index */
1361             jnrA             = jjnr[jidx];
1362             jnrB             = jjnr[jidx+1];
1363             jnrC             = jjnr[jidx+2];
1364             jnrD             = jjnr[jidx+3];
1365             j_coord_offsetA  = DIM*jnrA;
1366             j_coord_offsetB  = DIM*jnrB;
1367             j_coord_offsetC  = DIM*jnrC;
1368             j_coord_offsetD  = DIM*jnrD;
1369
1370             /* load j atom coordinates */
1371             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1372                                               x+j_coord_offsetC,x+j_coord_offsetD,
1373                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1374                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1375
1376             /* Calculate displacement vector */
1377             dx00             = _mm_sub_ps(ix0,jx0);
1378             dy00             = _mm_sub_ps(iy0,jy0);
1379             dz00             = _mm_sub_ps(iz0,jz0);
1380             dx11             = _mm_sub_ps(ix1,jx1);
1381             dy11             = _mm_sub_ps(iy1,jy1);
1382             dz11             = _mm_sub_ps(iz1,jz1);
1383             dx12             = _mm_sub_ps(ix1,jx2);
1384             dy12             = _mm_sub_ps(iy1,jy2);
1385             dz12             = _mm_sub_ps(iz1,jz2);
1386             dx13             = _mm_sub_ps(ix1,jx3);
1387             dy13             = _mm_sub_ps(iy1,jy3);
1388             dz13             = _mm_sub_ps(iz1,jz3);
1389             dx21             = _mm_sub_ps(ix2,jx1);
1390             dy21             = _mm_sub_ps(iy2,jy1);
1391             dz21             = _mm_sub_ps(iz2,jz1);
1392             dx22             = _mm_sub_ps(ix2,jx2);
1393             dy22             = _mm_sub_ps(iy2,jy2);
1394             dz22             = _mm_sub_ps(iz2,jz2);
1395             dx23             = _mm_sub_ps(ix2,jx3);
1396             dy23             = _mm_sub_ps(iy2,jy3);
1397             dz23             = _mm_sub_ps(iz2,jz3);
1398             dx31             = _mm_sub_ps(ix3,jx1);
1399             dy31             = _mm_sub_ps(iy3,jy1);
1400             dz31             = _mm_sub_ps(iz3,jz1);
1401             dx32             = _mm_sub_ps(ix3,jx2);
1402             dy32             = _mm_sub_ps(iy3,jy2);
1403             dz32             = _mm_sub_ps(iz3,jz2);
1404             dx33             = _mm_sub_ps(ix3,jx3);
1405             dy33             = _mm_sub_ps(iy3,jy3);
1406             dz33             = _mm_sub_ps(iz3,jz3);
1407
1408             /* Calculate squared distance and things based on it */
1409             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1410             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1411             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1412             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1413             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1414             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1415             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1416             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1417             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1418             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1419
1420             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1421             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1422             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1423             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1424             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1425             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1426             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1427             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1428             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1429             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1430
1431             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1432             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1433             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1434             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1435             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1436             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1437             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1438             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1439             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1440             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1441
1442             fjx0             = _mm_setzero_ps();
1443             fjy0             = _mm_setzero_ps();
1444             fjz0             = _mm_setzero_ps();
1445             fjx1             = _mm_setzero_ps();
1446             fjy1             = _mm_setzero_ps();
1447             fjz1             = _mm_setzero_ps();
1448             fjx2             = _mm_setzero_ps();
1449             fjy2             = _mm_setzero_ps();
1450             fjz2             = _mm_setzero_ps();
1451             fjx3             = _mm_setzero_ps();
1452             fjy3             = _mm_setzero_ps();
1453             fjz3             = _mm_setzero_ps();
1454
1455             /**************************
1456              * CALCULATE INTERACTIONS *
1457              **************************/
1458
1459             if (gmx_mm_any_lt(rsq00,rcutoff2))
1460             {
1461
1462             r00              = _mm_mul_ps(rsq00,rinv00);
1463
1464             /* LENNARD-JONES DISPERSION/REPULSION */
1465
1466             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1467             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1468             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1469             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1470             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1471
1472             d                = _mm_sub_ps(r00,rswitch);
1473             d                = _mm_max_ps(d,_mm_setzero_ps());
1474             d2               = _mm_mul_ps(d,d);
1475             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1476
1477             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1478
1479             /* Evaluate switch function */
1480             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1481             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1482             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1483
1484             fscal            = fvdw;
1485
1486             fscal            = _mm_and_ps(fscal,cutoff_mask);
1487
1488              /* Update vectorial force */
1489             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1490             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1491             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1492
1493             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1494             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1495             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1496
1497             }
1498
1499             /**************************
1500              * CALCULATE INTERACTIONS *
1501              **************************/
1502
1503             if (gmx_mm_any_lt(rsq11,rcutoff2))
1504             {
1505
1506             /* REACTION-FIELD ELECTROSTATICS */
1507             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1508
1509             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1510
1511             fscal            = felec;
1512
1513             fscal            = _mm_and_ps(fscal,cutoff_mask);
1514
1515              /* Update vectorial force */
1516             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1517             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1518             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1519
1520             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1521             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1522             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1523
1524             }
1525
1526             /**************************
1527              * CALCULATE INTERACTIONS *
1528              **************************/
1529
1530             if (gmx_mm_any_lt(rsq12,rcutoff2))
1531             {
1532
1533             /* REACTION-FIELD ELECTROSTATICS */
1534             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1535
1536             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1537
1538             fscal            = felec;
1539
1540             fscal            = _mm_and_ps(fscal,cutoff_mask);
1541
1542              /* Update vectorial force */
1543             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1544             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1545             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1546
1547             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1548             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1549             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1550
1551             }
1552
1553             /**************************
1554              * CALCULATE INTERACTIONS *
1555              **************************/
1556
1557             if (gmx_mm_any_lt(rsq13,rcutoff2))
1558             {
1559
1560             /* REACTION-FIELD ELECTROSTATICS */
1561             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1562
1563             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1564
1565             fscal            = felec;
1566
1567             fscal            = _mm_and_ps(fscal,cutoff_mask);
1568
1569              /* Update vectorial force */
1570             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1571             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1572             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1573
1574             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1575             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1576             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1577
1578             }
1579
1580             /**************************
1581              * CALCULATE INTERACTIONS *
1582              **************************/
1583
1584             if (gmx_mm_any_lt(rsq21,rcutoff2))
1585             {
1586
1587             /* REACTION-FIELD ELECTROSTATICS */
1588             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1589
1590             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1591
1592             fscal            = felec;
1593
1594             fscal            = _mm_and_ps(fscal,cutoff_mask);
1595
1596              /* Update vectorial force */
1597             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1598             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1599             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1600
1601             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1602             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1603             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1604
1605             }
1606
1607             /**************************
1608              * CALCULATE INTERACTIONS *
1609              **************************/
1610
1611             if (gmx_mm_any_lt(rsq22,rcutoff2))
1612             {
1613
1614             /* REACTION-FIELD ELECTROSTATICS */
1615             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1616
1617             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1618
1619             fscal            = felec;
1620
1621             fscal            = _mm_and_ps(fscal,cutoff_mask);
1622
1623              /* Update vectorial force */
1624             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1625             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1626             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1627
1628             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1629             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1630             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1631
1632             }
1633
1634             /**************************
1635              * CALCULATE INTERACTIONS *
1636              **************************/
1637
1638             if (gmx_mm_any_lt(rsq23,rcutoff2))
1639             {
1640
1641             /* REACTION-FIELD ELECTROSTATICS */
1642             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1643
1644             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1645
1646             fscal            = felec;
1647
1648             fscal            = _mm_and_ps(fscal,cutoff_mask);
1649
1650              /* Update vectorial force */
1651             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1652             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1653             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1654
1655             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1656             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1657             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1658
1659             }
1660
1661             /**************************
1662              * CALCULATE INTERACTIONS *
1663              **************************/
1664
1665             if (gmx_mm_any_lt(rsq31,rcutoff2))
1666             {
1667
1668             /* REACTION-FIELD ELECTROSTATICS */
1669             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1670
1671             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1672
1673             fscal            = felec;
1674
1675             fscal            = _mm_and_ps(fscal,cutoff_mask);
1676
1677              /* Update vectorial force */
1678             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1679             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1680             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1681
1682             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1683             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1684             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1685
1686             }
1687
1688             /**************************
1689              * CALCULATE INTERACTIONS *
1690              **************************/
1691
1692             if (gmx_mm_any_lt(rsq32,rcutoff2))
1693             {
1694
1695             /* REACTION-FIELD ELECTROSTATICS */
1696             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1697
1698             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1699
1700             fscal            = felec;
1701
1702             fscal            = _mm_and_ps(fscal,cutoff_mask);
1703
1704              /* Update vectorial force */
1705             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1706             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1707             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1708
1709             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1710             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1711             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1712
1713             }
1714
1715             /**************************
1716              * CALCULATE INTERACTIONS *
1717              **************************/
1718
1719             if (gmx_mm_any_lt(rsq33,rcutoff2))
1720             {
1721
1722             /* REACTION-FIELD ELECTROSTATICS */
1723             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1724
1725             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_and_ps(fscal,cutoff_mask);
1730
1731              /* Update vectorial force */
1732             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1733             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1734             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1735
1736             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1737             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1738             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1739
1740             }
1741
1742             fjptrA             = f+j_coord_offsetA;
1743             fjptrB             = f+j_coord_offsetB;
1744             fjptrC             = f+j_coord_offsetC;
1745             fjptrD             = f+j_coord_offsetD;
1746
1747             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1748                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1749                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1750
1751             /* Inner loop uses 359 flops */
1752         }
1753
1754         if(jidx<j_index_end)
1755         {
1756
1757             /* Get j neighbor index, and coordinate index */
1758             jnrlistA         = jjnr[jidx];
1759             jnrlistB         = jjnr[jidx+1];
1760             jnrlistC         = jjnr[jidx+2];
1761             jnrlistD         = jjnr[jidx+3];
1762             /* Sign of each element will be negative for non-real atoms.
1763              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1764              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1765              */
1766             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1767             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1768             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1769             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1770             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1771             j_coord_offsetA  = DIM*jnrA;
1772             j_coord_offsetB  = DIM*jnrB;
1773             j_coord_offsetC  = DIM*jnrC;
1774             j_coord_offsetD  = DIM*jnrD;
1775
1776             /* load j atom coordinates */
1777             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1778                                               x+j_coord_offsetC,x+j_coord_offsetD,
1779                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1780                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1781
1782             /* Calculate displacement vector */
1783             dx00             = _mm_sub_ps(ix0,jx0);
1784             dy00             = _mm_sub_ps(iy0,jy0);
1785             dz00             = _mm_sub_ps(iz0,jz0);
1786             dx11             = _mm_sub_ps(ix1,jx1);
1787             dy11             = _mm_sub_ps(iy1,jy1);
1788             dz11             = _mm_sub_ps(iz1,jz1);
1789             dx12             = _mm_sub_ps(ix1,jx2);
1790             dy12             = _mm_sub_ps(iy1,jy2);
1791             dz12             = _mm_sub_ps(iz1,jz2);
1792             dx13             = _mm_sub_ps(ix1,jx3);
1793             dy13             = _mm_sub_ps(iy1,jy3);
1794             dz13             = _mm_sub_ps(iz1,jz3);
1795             dx21             = _mm_sub_ps(ix2,jx1);
1796             dy21             = _mm_sub_ps(iy2,jy1);
1797             dz21             = _mm_sub_ps(iz2,jz1);
1798             dx22             = _mm_sub_ps(ix2,jx2);
1799             dy22             = _mm_sub_ps(iy2,jy2);
1800             dz22             = _mm_sub_ps(iz2,jz2);
1801             dx23             = _mm_sub_ps(ix2,jx3);
1802             dy23             = _mm_sub_ps(iy2,jy3);
1803             dz23             = _mm_sub_ps(iz2,jz3);
1804             dx31             = _mm_sub_ps(ix3,jx1);
1805             dy31             = _mm_sub_ps(iy3,jy1);
1806             dz31             = _mm_sub_ps(iz3,jz1);
1807             dx32             = _mm_sub_ps(ix3,jx2);
1808             dy32             = _mm_sub_ps(iy3,jy2);
1809             dz32             = _mm_sub_ps(iz3,jz2);
1810             dx33             = _mm_sub_ps(ix3,jx3);
1811             dy33             = _mm_sub_ps(iy3,jy3);
1812             dz33             = _mm_sub_ps(iz3,jz3);
1813
1814             /* Calculate squared distance and things based on it */
1815             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1816             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1817             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1818             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1819             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1820             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1821             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1822             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1823             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1824             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1825
1826             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1827             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1828             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1829             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1830             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1831             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1832             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1833             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1834             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1835             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1836
1837             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1838             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1839             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1840             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1841             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1842             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1843             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1844             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1845             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1846             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1847
1848             fjx0             = _mm_setzero_ps();
1849             fjy0             = _mm_setzero_ps();
1850             fjz0             = _mm_setzero_ps();
1851             fjx1             = _mm_setzero_ps();
1852             fjy1             = _mm_setzero_ps();
1853             fjz1             = _mm_setzero_ps();
1854             fjx2             = _mm_setzero_ps();
1855             fjy2             = _mm_setzero_ps();
1856             fjz2             = _mm_setzero_ps();
1857             fjx3             = _mm_setzero_ps();
1858             fjy3             = _mm_setzero_ps();
1859             fjz3             = _mm_setzero_ps();
1860
1861             /**************************
1862              * CALCULATE INTERACTIONS *
1863              **************************/
1864
1865             if (gmx_mm_any_lt(rsq00,rcutoff2))
1866             {
1867
1868             r00              = _mm_mul_ps(rsq00,rinv00);
1869             r00              = _mm_andnot_ps(dummy_mask,r00);
1870
1871             /* LENNARD-JONES DISPERSION/REPULSION */
1872
1873             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1874             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1875             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1876             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1877             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1878
1879             d                = _mm_sub_ps(r00,rswitch);
1880             d                = _mm_max_ps(d,_mm_setzero_ps());
1881             d2               = _mm_mul_ps(d,d);
1882             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1883
1884             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1885
1886             /* Evaluate switch function */
1887             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1888             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1889             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1890
1891             fscal            = fvdw;
1892
1893             fscal            = _mm_and_ps(fscal,cutoff_mask);
1894
1895             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1896
1897              /* Update vectorial force */
1898             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1899             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1900             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1901
1902             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1903             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1904             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1905
1906             }
1907
1908             /**************************
1909              * CALCULATE INTERACTIONS *
1910              **************************/
1911
1912             if (gmx_mm_any_lt(rsq11,rcutoff2))
1913             {
1914
1915             /* REACTION-FIELD ELECTROSTATICS */
1916             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1917
1918             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1919
1920             fscal            = felec;
1921
1922             fscal            = _mm_and_ps(fscal,cutoff_mask);
1923
1924             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1925
1926              /* Update vectorial force */
1927             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1928             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1929             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1930
1931             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1932             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1933             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1934
1935             }
1936
1937             /**************************
1938              * CALCULATE INTERACTIONS *
1939              **************************/
1940
1941             if (gmx_mm_any_lt(rsq12,rcutoff2))
1942             {
1943
1944             /* REACTION-FIELD ELECTROSTATICS */
1945             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1946
1947             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1948
1949             fscal            = felec;
1950
1951             fscal            = _mm_and_ps(fscal,cutoff_mask);
1952
1953             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1954
1955              /* Update vectorial force */
1956             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1957             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1958             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1959
1960             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1961             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1962             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1963
1964             }
1965
1966             /**************************
1967              * CALCULATE INTERACTIONS *
1968              **************************/
1969
1970             if (gmx_mm_any_lt(rsq13,rcutoff2))
1971             {
1972
1973             /* REACTION-FIELD ELECTROSTATICS */
1974             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1975
1976             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1977
1978             fscal            = felec;
1979
1980             fscal            = _mm_and_ps(fscal,cutoff_mask);
1981
1982             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1983
1984              /* Update vectorial force */
1985             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1986             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1987             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1988
1989             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1990             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1991             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1992
1993             }
1994
1995             /**************************
1996              * CALCULATE INTERACTIONS *
1997              **************************/
1998
1999             if (gmx_mm_any_lt(rsq21,rcutoff2))
2000             {
2001
2002             /* REACTION-FIELD ELECTROSTATICS */
2003             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
2004
2005             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2006
2007             fscal            = felec;
2008
2009             fscal            = _mm_and_ps(fscal,cutoff_mask);
2010
2011             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2012
2013              /* Update vectorial force */
2014             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2015             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2016             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2017
2018             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2019             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2020             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2021
2022             }
2023
2024             /**************************
2025              * CALCULATE INTERACTIONS *
2026              **************************/
2027
2028             if (gmx_mm_any_lt(rsq22,rcutoff2))
2029             {
2030
2031             /* REACTION-FIELD ELECTROSTATICS */
2032             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
2033
2034             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2035
2036             fscal            = felec;
2037
2038             fscal            = _mm_and_ps(fscal,cutoff_mask);
2039
2040             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2041
2042              /* Update vectorial force */
2043             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2044             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2045             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2046
2047             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2048             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2049             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2050
2051             }
2052
2053             /**************************
2054              * CALCULATE INTERACTIONS *
2055              **************************/
2056
2057             if (gmx_mm_any_lt(rsq23,rcutoff2))
2058             {
2059
2060             /* REACTION-FIELD ELECTROSTATICS */
2061             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
2062
2063             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2064
2065             fscal            = felec;
2066
2067             fscal            = _mm_and_ps(fscal,cutoff_mask);
2068
2069             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2070
2071              /* Update vectorial force */
2072             fix2             = _mm_macc_ps(dx23,fscal,fix2);
2073             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
2074             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
2075
2076             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
2077             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
2078             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
2079
2080             }
2081
2082             /**************************
2083              * CALCULATE INTERACTIONS *
2084              **************************/
2085
2086             if (gmx_mm_any_lt(rsq31,rcutoff2))
2087             {
2088
2089             /* REACTION-FIELD ELECTROSTATICS */
2090             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
2091
2092             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2093
2094             fscal            = felec;
2095
2096             fscal            = _mm_and_ps(fscal,cutoff_mask);
2097
2098             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2099
2100              /* Update vectorial force */
2101             fix3             = _mm_macc_ps(dx31,fscal,fix3);
2102             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
2103             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
2104
2105             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
2106             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
2107             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
2108
2109             }
2110
2111             /**************************
2112              * CALCULATE INTERACTIONS *
2113              **************************/
2114
2115             if (gmx_mm_any_lt(rsq32,rcutoff2))
2116             {
2117
2118             /* REACTION-FIELD ELECTROSTATICS */
2119             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
2120
2121             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2122
2123             fscal            = felec;
2124
2125             fscal            = _mm_and_ps(fscal,cutoff_mask);
2126
2127             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2128
2129              /* Update vectorial force */
2130             fix3             = _mm_macc_ps(dx32,fscal,fix3);
2131             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
2132             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
2133
2134             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
2135             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
2136             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
2137
2138             }
2139
2140             /**************************
2141              * CALCULATE INTERACTIONS *
2142              **************************/
2143
2144             if (gmx_mm_any_lt(rsq33,rcutoff2))
2145             {
2146
2147             /* REACTION-FIELD ELECTROSTATICS */
2148             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
2149
2150             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2151
2152             fscal            = felec;
2153
2154             fscal            = _mm_and_ps(fscal,cutoff_mask);
2155
2156             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2157
2158              /* Update vectorial force */
2159             fix3             = _mm_macc_ps(dx33,fscal,fix3);
2160             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
2161             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
2162
2163             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
2164             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
2165             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
2166
2167             }
2168
2169             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2170             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2171             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2172             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2173
2174             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2175                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2176                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2177
2178             /* Inner loop uses 360 flops */
2179         }
2180
2181         /* End of innermost loop */
2182
2183         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2184                                               f+i_coord_offset,fshift+i_shift_offset);
2185
2186         /* Increment number of inner iterations */
2187         inneriter                  += j_index_end - j_index_start;
2188
2189         /* Outer loop uses 24 flops */
2190     }
2191
2192     /* Increment number of outer iterations */
2193     outeriter        += nri;
2194
2195     /* Update outer/inner flops */
2196
2197     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*360);
2198 }