src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSh_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_VdwLJSh_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_VdwLJSh_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           dummy_mask,cutoff_mask;
 104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 105     __m128           one     = _mm_set1_ps(1.0);
 106     __m128           two     = _mm_set1_ps(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_ps(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     krf              = _mm_set1_ps(fr->ic->k_rf);
 121     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 122     crf              = _mm_set1_ps(fr->ic->c_rf);
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 132     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 133
 134     jq1              = _mm_set1_ps(charge[inr+1]);
 135     jq2              = _mm_set1_ps(charge[inr+2]);
 136     jq3              = _mm_set1_ps(charge[inr+3]);
 137     vdwjidx0A        = 2*vdwtype[inr+0];
 138     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 139     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 140     qq11             = _mm_mul_ps(iq1,jq1);
 141     qq12             = _mm_mul_ps(iq1,jq2);
 142     qq13             = _mm_mul_ps(iq1,jq3);
 143     qq21             = _mm_mul_ps(iq2,jq1);
 144     qq22             = _mm_mul_ps(iq2,jq2);
 145     qq23             = _mm_mul_ps(iq2,jq3);
 146     qq31             = _mm_mul_ps(iq3,jq1);
 147     qq32             = _mm_mul_ps(iq3,jq2);
 148     qq33             = _mm_mul_ps(iq3,jq3);
 149
 150     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 151     rcutoff_scalar   = fr->rcoulomb;
 152     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 153     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 154
 155     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 156     rvdw             = _mm_set1_ps(fr->rvdw);
 157
 158     /* Avoid stupid compiler warnings */
 159     jnrA = jnrB = jnrC = jnrD = 0;
 160     j_coord_offsetA = 0;
 161     j_coord_offsetB = 0;
 162     j_coord_offsetC = 0;
 163     j_coord_offsetD = 0;
 164
 165     outeriter        = 0;
 166     inneriter        = 0;
 167
 168     for(iidx=0;iidx<4*DIM;iidx++)
 169     {
 170         scratch[iidx] = 0.0;
 171     }
 172
 173     /* Start outer loop over neighborlists */
 174     for(iidx=0; iidx<nri; iidx++)
 175     {
 176         /* Load shift vector for this list */
 177         i_shift_offset   = DIM*shiftidx[iidx];
 178
 179         /* Load limits for loop over neighbors */
 180         j_index_start    = jindex[iidx];
 181         j_index_end      = jindex[iidx+1];
 182
 183         /* Get outer coordinate index */
 184         inr              = iinr[iidx];
 185         i_coord_offset   = DIM*inr;
 186
 187         /* Load i particle coords and add shift vector */
 188         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 189                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 190
 191         fix0             = _mm_setzero_ps();
 192         fiy0             = _mm_setzero_ps();
 193         fiz0             = _mm_setzero_ps();
 194         fix1             = _mm_setzero_ps();
 195         fiy1             = _mm_setzero_ps();
 196         fiz1             = _mm_setzero_ps();
 197         fix2             = _mm_setzero_ps();
 198         fiy2             = _mm_setzero_ps();
 199         fiz2             = _mm_setzero_ps();
 200         fix3             = _mm_setzero_ps();
 201         fiy3             = _mm_setzero_ps();
 202         fiz3             = _mm_setzero_ps();
 203
 204         /* Reset potential sums */
 205         velecsum         = _mm_setzero_ps();
 206         vvdwsum          = _mm_setzero_ps();
 207
 208         /* Start inner kernel loop */
 209         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 210         {
 211
 212             /* Get j neighbor index, and coordinate index */
 213             jnrA             = jjnr[jidx];
 214             jnrB             = jjnr[jidx+1];
 215             jnrC             = jjnr[jidx+2];
 216             jnrD             = jjnr[jidx+3];
 217             j_coord_offsetA  = DIM*jnrA;
 218             j_coord_offsetB  = DIM*jnrB;
 219             j_coord_offsetC  = DIM*jnrC;
 220             j_coord_offsetD  = DIM*jnrD;
 221
 222             /* load j atom coordinates */
 223             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 224                                               x+j_coord_offsetC,x+j_coord_offsetD,
 225                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 226                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 227
 228             /* Calculate displacement vector */
 229             dx00             = _mm_sub_ps(ix0,jx0);
 230             dy00             = _mm_sub_ps(iy0,jy0);
 231             dz00             = _mm_sub_ps(iz0,jz0);
 232             dx11             = _mm_sub_ps(ix1,jx1);
 233             dy11             = _mm_sub_ps(iy1,jy1);
 234             dz11             = _mm_sub_ps(iz1,jz1);
 235             dx12             = _mm_sub_ps(ix1,jx2);
 236             dy12             = _mm_sub_ps(iy1,jy2);
 237             dz12             = _mm_sub_ps(iz1,jz2);
 238             dx13             = _mm_sub_ps(ix1,jx3);
 239             dy13             = _mm_sub_ps(iy1,jy3);
 240             dz13             = _mm_sub_ps(iz1,jz3);
 241             dx21             = _mm_sub_ps(ix2,jx1);
 242             dy21             = _mm_sub_ps(iy2,jy1);
 243             dz21             = _mm_sub_ps(iz2,jz1);
 244             dx22             = _mm_sub_ps(ix2,jx2);
 245             dy22             = _mm_sub_ps(iy2,jy2);
 246             dz22             = _mm_sub_ps(iz2,jz2);
 247             dx23             = _mm_sub_ps(ix2,jx3);
 248             dy23             = _mm_sub_ps(iy2,jy3);
 249             dz23             = _mm_sub_ps(iz2,jz3);
 250             dx31             = _mm_sub_ps(ix3,jx1);
 251             dy31             = _mm_sub_ps(iy3,jy1);
 252             dz31             = _mm_sub_ps(iz3,jz1);
 253             dx32             = _mm_sub_ps(ix3,jx2);
 254             dy32             = _mm_sub_ps(iy3,jy2);
 255             dz32             = _mm_sub_ps(iz3,jz2);
 256             dx33             = _mm_sub_ps(ix3,jx3);
 257             dy33             = _mm_sub_ps(iy3,jy3);
 258             dz33             = _mm_sub_ps(iz3,jz3);
 259
 260             /* Calculate squared distance and things based on it */
 261             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 262             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 263             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 264             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 265             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 266             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 267             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 268             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 269             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 270             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 271
 272             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 273             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 274             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 275             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 276             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 277             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 278             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 279             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 280             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 281
 282             rinvsq00         = gmx_mm_inv_ps(rsq00);
 283             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 284             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 285             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 286             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 287             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 288             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 289             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 290             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 291             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 292
 293             fjx0             = _mm_setzero_ps();
 294             fjy0             = _mm_setzero_ps();
 295             fjz0             = _mm_setzero_ps();
 296             fjx1             = _mm_setzero_ps();
 297             fjy1             = _mm_setzero_ps();
 298             fjz1             = _mm_setzero_ps();
 299             fjx2             = _mm_setzero_ps();
 300             fjy2             = _mm_setzero_ps();
 301             fjz2             = _mm_setzero_ps();
 302             fjx3             = _mm_setzero_ps();
 303             fjy3             = _mm_setzero_ps();
 304             fjz3             = _mm_setzero_ps();
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             if (gmx_mm_any_lt(rsq00,rcutoff2))
 311             {
 312
 313             /* LENNARD-JONES DISPERSION/REPULSION */
 314
 315             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 316             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 317             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 318             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 319                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 320             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 321
 322             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 323
 324             /* Update potential sum for this i atom from the interaction with this j atom. */
 325             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 326             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 327
 328             fscal            = fvdw;
 329
 330             fscal            = _mm_and_ps(fscal,cutoff_mask);
 331
 332              /* Update vectorial force */
 333             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 334             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 335             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 336
 337             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 338             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 339             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 340
 341             }
 342
 343             /**************************
 344              * CALCULATE INTERACTIONS *
 345              **************************/
 346
 347             if (gmx_mm_any_lt(rsq11,rcutoff2))
 348             {
 349
 350             /* REACTION-FIELD ELECTROSTATICS */
 351             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 352             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 353
 354             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 355
 356             /* Update potential sum for this i atom from the interaction with this j atom. */
 357             velec            = _mm_and_ps(velec,cutoff_mask);
 358             velecsum         = _mm_add_ps(velecsum,velec);
 359
 360             fscal            = felec;
 361
 362             fscal            = _mm_and_ps(fscal,cutoff_mask);
 363
 364              /* Update vectorial force */
 365             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 366             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 367             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 368
 369             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 370             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 371             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 372
 373             }
 374
 375             /**************************
 376              * CALCULATE INTERACTIONS *
 377              **************************/
 378
 379             if (gmx_mm_any_lt(rsq12,rcutoff2))
 380             {
 381
 382             /* REACTION-FIELD ELECTROSTATICS */
 383             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 384             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 385
 386             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 387
 388             /* Update potential sum for this i atom from the interaction with this j atom. */
 389             velec            = _mm_and_ps(velec,cutoff_mask);
 390             velecsum         = _mm_add_ps(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             fscal            = _mm_and_ps(fscal,cutoff_mask);
 395
 396              /* Update vectorial force */
 397             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 398             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 399             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 400
 401             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 402             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 403             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 404
 405             }
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             if (gmx_mm_any_lt(rsq13,rcutoff2))
 412             {
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 416             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 417
 418             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 419
 420             /* Update potential sum for this i atom from the interaction with this j atom. */
 421             velec            = _mm_and_ps(velec,cutoff_mask);
 422             velecsum         = _mm_add_ps(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             fscal            = _mm_and_ps(fscal,cutoff_mask);
 427
 428              /* Update vectorial force */
 429             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 430             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 431             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 432
 433             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 434             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 435             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 436
 437             }
 438
 439             /**************************
 440              * CALCULATE INTERACTIONS *
 441              **************************/
 442
 443             if (gmx_mm_any_lt(rsq21,rcutoff2))
 444             {
 445
 446             /* REACTION-FIELD ELECTROSTATICS */
 447             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 448             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 449
 450             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 451
 452             /* Update potential sum for this i atom from the interaction with this j atom. */
 453             velec            = _mm_and_ps(velec,cutoff_mask);
 454             velecsum         = _mm_add_ps(velecsum,velec);
 455
 456             fscal            = felec;
 457
 458             fscal            = _mm_and_ps(fscal,cutoff_mask);
 459
 460              /* Update vectorial force */
 461             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 462             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 463             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 464
 465             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 466             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 467             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 468
 469             }
 470
 471             /**************************
 472              * CALCULATE INTERACTIONS *
 473              **************************/
 474
 475             if (gmx_mm_any_lt(rsq22,rcutoff2))
 476             {
 477
 478             /* REACTION-FIELD ELECTROSTATICS */
 479             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 480             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 481
 482             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velec            = _mm_and_ps(velec,cutoff_mask);
 486             velecsum         = _mm_add_ps(velecsum,velec);
 487
 488             fscal            = felec;
 489
 490             fscal            = _mm_and_ps(fscal,cutoff_mask);
 491
 492              /* Update vectorial force */
 493             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 494             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 495             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 496
 497             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 498             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 499             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 500
 501             }
 502
 503             /**************************
 504              * CALCULATE INTERACTIONS *
 505              **************************/
 506
 507             if (gmx_mm_any_lt(rsq23,rcutoff2))
 508             {
 509
 510             /* REACTION-FIELD ELECTROSTATICS */
 511             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 512             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 513
 514             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 515
 516             /* Update potential sum for this i atom from the interaction with this j atom. */
 517             velec            = _mm_and_ps(velec,cutoff_mask);
 518             velecsum         = _mm_add_ps(velecsum,velec);
 519
 520             fscal            = felec;
 521
 522             fscal            = _mm_and_ps(fscal,cutoff_mask);
 523
 524              /* Update vectorial force */
 525             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 526             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 527             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 528
 529             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 530             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 531             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 532
 533             }
 534
 535             /**************************
 536              * CALCULATE INTERACTIONS *
 537              **************************/
 538
 539             if (gmx_mm_any_lt(rsq31,rcutoff2))
 540             {
 541
 542             /* REACTION-FIELD ELECTROSTATICS */
 543             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 544             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 545
 546             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 547
 548             /* Update potential sum for this i atom from the interaction with this j atom. */
 549             velec            = _mm_and_ps(velec,cutoff_mask);
 550             velecsum         = _mm_add_ps(velecsum,velec);
 551
 552             fscal            = felec;
 553
 554             fscal            = _mm_and_ps(fscal,cutoff_mask);
 555
 556              /* Update vectorial force */
 557             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 558             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 559             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 560
 561             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 562             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 563             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 564
 565             }
 566
 567             /**************************
 568              * CALCULATE INTERACTIONS *
 569              **************************/
 570
 571             if (gmx_mm_any_lt(rsq32,rcutoff2))
 572             {
 573
 574             /* REACTION-FIELD ELECTROSTATICS */
 575             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 576             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 577
 578             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 579
 580             /* Update potential sum for this i atom from the interaction with this j atom. */
 581             velec            = _mm_and_ps(velec,cutoff_mask);
 582             velecsum         = _mm_add_ps(velecsum,velec);
 583
 584             fscal            = felec;
 585
 586             fscal            = _mm_and_ps(fscal,cutoff_mask);
 587
 588              /* Update vectorial force */
 589             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 590             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 591             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 592
 593             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 594             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 595             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 596
 597             }
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             if (gmx_mm_any_lt(rsq33,rcutoff2))
 604             {
 605
 606             /* REACTION-FIELD ELECTROSTATICS */
 607             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 608             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 609
 610             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 611
 612             /* Update potential sum for this i atom from the interaction with this j atom. */
 613             velec            = _mm_and_ps(velec,cutoff_mask);
 614             velecsum         = _mm_add_ps(velecsum,velec);
 615
 616             fscal            = felec;
 617
 618             fscal            = _mm_and_ps(fscal,cutoff_mask);
 619
 620              /* Update vectorial force */
 621             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 622             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 623             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 624
 625             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 626             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 627             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 628
 629             }
 630
 631             fjptrA             = f+j_coord_offsetA;
 632             fjptrB             = f+j_coord_offsetB;
 633             fjptrC             = f+j_coord_offsetC;
 634             fjptrD             = f+j_coord_offsetD;
 635
 636             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 637                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 638                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 639
 640             /* Inner loop uses 398 flops */
 641         }
 642
 643         if(jidx<j_index_end)
 644         {
 645
 646             /* Get j neighbor index, and coordinate index */
 647             jnrlistA         = jjnr[jidx];
 648             jnrlistB         = jjnr[jidx+1];
 649             jnrlistC         = jjnr[jidx+2];
 650             jnrlistD         = jjnr[jidx+3];
 651             /* Sign of each element will be negative for non-real atoms.
 652              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 653              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 654              */
 655             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 656             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 657             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 658             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 659             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 660             j_coord_offsetA  = DIM*jnrA;
 661             j_coord_offsetB  = DIM*jnrB;
 662             j_coord_offsetC  = DIM*jnrC;
 663             j_coord_offsetD  = DIM*jnrD;
 664
 665             /* load j atom coordinates */
 666             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 667                                               x+j_coord_offsetC,x+j_coord_offsetD,
 668                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 669                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 670
 671             /* Calculate displacement vector */
 672             dx00             = _mm_sub_ps(ix0,jx0);
 673             dy00             = _mm_sub_ps(iy0,jy0);
 674             dz00             = _mm_sub_ps(iz0,jz0);
 675             dx11             = _mm_sub_ps(ix1,jx1);
 676             dy11             = _mm_sub_ps(iy1,jy1);
 677             dz11             = _mm_sub_ps(iz1,jz1);
 678             dx12             = _mm_sub_ps(ix1,jx2);
 679             dy12             = _mm_sub_ps(iy1,jy2);
 680             dz12             = _mm_sub_ps(iz1,jz2);
 681             dx13             = _mm_sub_ps(ix1,jx3);
 682             dy13             = _mm_sub_ps(iy1,jy3);
 683             dz13             = _mm_sub_ps(iz1,jz3);
 684             dx21             = _mm_sub_ps(ix2,jx1);
 685             dy21             = _mm_sub_ps(iy2,jy1);
 686             dz21             = _mm_sub_ps(iz2,jz1);
 687             dx22             = _mm_sub_ps(ix2,jx2);
 688             dy22             = _mm_sub_ps(iy2,jy2);
 689             dz22             = _mm_sub_ps(iz2,jz2);
 690             dx23             = _mm_sub_ps(ix2,jx3);
 691             dy23             = _mm_sub_ps(iy2,jy3);
 692             dz23             = _mm_sub_ps(iz2,jz3);
 693             dx31             = _mm_sub_ps(ix3,jx1);
 694             dy31             = _mm_sub_ps(iy3,jy1);
 695             dz31             = _mm_sub_ps(iz3,jz1);
 696             dx32             = _mm_sub_ps(ix3,jx2);
 697             dy32             = _mm_sub_ps(iy3,jy2);
 698             dz32             = _mm_sub_ps(iz3,jz2);
 699             dx33             = _mm_sub_ps(ix3,jx3);
 700             dy33             = _mm_sub_ps(iy3,jy3);
 701             dz33             = _mm_sub_ps(iz3,jz3);
 702
 703             /* Calculate squared distance and things based on it */
 704             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 705             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 706             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 707             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 708             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 709             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 710             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 711             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 712             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 713             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 714
 715             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 716             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 717             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 718             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 719             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 720             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 721             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 722             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 723             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 724
 725             rinvsq00         = gmx_mm_inv_ps(rsq00);
 726             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 727             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 728             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 729             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 730             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 731             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 732             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 733             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 734             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 735
 736             fjx0             = _mm_setzero_ps();
 737             fjy0             = _mm_setzero_ps();
 738             fjz0             = _mm_setzero_ps();
 739             fjx1             = _mm_setzero_ps();
 740             fjy1             = _mm_setzero_ps();
 741             fjz1             = _mm_setzero_ps();
 742             fjx2             = _mm_setzero_ps();
 743             fjy2             = _mm_setzero_ps();
 744             fjz2             = _mm_setzero_ps();
 745             fjx3             = _mm_setzero_ps();
 746             fjy3             = _mm_setzero_ps();
 747             fjz3             = _mm_setzero_ps();
 748
 749             /**************************
 750              * CALCULATE INTERACTIONS *
 751              **************************/
 752
 753             if (gmx_mm_any_lt(rsq00,rcutoff2))
 754             {
 755
 756             /* LENNARD-JONES DISPERSION/REPULSION */
 757
 758             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 759             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 760             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 761             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 762                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 763             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 764
 765             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 766
 767             /* Update potential sum for this i atom from the interaction with this j atom. */
 768             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 769             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 770             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 771
 772             fscal            = fvdw;
 773
 774             fscal            = _mm_and_ps(fscal,cutoff_mask);
 775
 776             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 777
 778              /* Update vectorial force */
 779             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 780             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 781             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 782
 783             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 784             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 785             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 786
 787             }
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             if (gmx_mm_any_lt(rsq11,rcutoff2))
 794             {
 795
 796             /* REACTION-FIELD ELECTROSTATICS */
 797             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 798             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 799
 800             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_and_ps(velec,cutoff_mask);
 804             velec            = _mm_andnot_ps(dummy_mask,velec);
 805             velecsum         = _mm_add_ps(velecsum,velec);
 806
 807             fscal            = felec;
 808
 809             fscal            = _mm_and_ps(fscal,cutoff_mask);
 810
 811             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 812
 813              /* Update vectorial force */
 814             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 815             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 816             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 817
 818             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 819             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 820             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 821
 822             }
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (gmx_mm_any_lt(rsq12,rcutoff2))
 829             {
 830
 831             /* REACTION-FIELD ELECTROSTATICS */
 832             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 833             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 834
 835             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_and_ps(velec,cutoff_mask);
 839             velec            = _mm_andnot_ps(dummy_mask,velec);
 840             velecsum         = _mm_add_ps(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_and_ps(fscal,cutoff_mask);
 845
 846             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 847
 848              /* Update vectorial force */
 849             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 850             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 851             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 852
 853             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 854             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 855             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 856
 857             }
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             if (gmx_mm_any_lt(rsq13,rcutoff2))
 864             {
 865
 866             /* REACTION-FIELD ELECTROSTATICS */
 867             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 868             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 869
 870             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 871
 872             /* Update potential sum for this i atom from the interaction with this j atom. */
 873             velec            = _mm_and_ps(velec,cutoff_mask);
 874             velec            = _mm_andnot_ps(dummy_mask,velec);
 875             velecsum         = _mm_add_ps(velecsum,velec);
 876
 877             fscal            = felec;
 878
 879             fscal            = _mm_and_ps(fscal,cutoff_mask);
 880
 881             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 882
 883              /* Update vectorial force */
 884             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 885             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 886             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 887
 888             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 889             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 890             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 891
 892             }
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             if (gmx_mm_any_lt(rsq21,rcutoff2))
 899             {
 900
 901             /* REACTION-FIELD ELECTROSTATICS */
 902             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 903             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 904
 905             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 906
 907             /* Update potential sum for this i atom from the interaction with this j atom. */
 908             velec            = _mm_and_ps(velec,cutoff_mask);
 909             velec            = _mm_andnot_ps(dummy_mask,velec);
 910             velecsum         = _mm_add_ps(velecsum,velec);
 911
 912             fscal            = felec;
 913
 914             fscal            = _mm_and_ps(fscal,cutoff_mask);
 915
 916             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 917
 918              /* Update vectorial force */
 919             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 920             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 921             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 922
 923             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 924             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 925             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 926
 927             }
 928
 929             /**************************
 930              * CALCULATE INTERACTIONS *
 931              **************************/
 932
 933             if (gmx_mm_any_lt(rsq22,rcutoff2))
 934             {
 935
 936             /* REACTION-FIELD ELECTROSTATICS */
 937             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 938             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 939
 940             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 941
 942             /* Update potential sum for this i atom from the interaction with this j atom. */
 943             velec            = _mm_and_ps(velec,cutoff_mask);
 944             velec            = _mm_andnot_ps(dummy_mask,velec);
 945             velecsum         = _mm_add_ps(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_and_ps(fscal,cutoff_mask);
 950
 951             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 952
 953              /* Update vectorial force */
 954             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 955             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 956             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 957
 958             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 959             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 960             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 961
 962             }
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             if (gmx_mm_any_lt(rsq23,rcutoff2))
 969             {
 970
 971             /* REACTION-FIELD ELECTROSTATICS */
 972             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 973             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 974
 975             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 976
 977             /* Update potential sum for this i atom from the interaction with this j atom. */
 978             velec            = _mm_and_ps(velec,cutoff_mask);
 979             velec            = _mm_andnot_ps(dummy_mask,velec);
 980             velecsum         = _mm_add_ps(velecsum,velec);
 981
 982             fscal            = felec;
 983
 984             fscal            = _mm_and_ps(fscal,cutoff_mask);
 985
 986             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 987
 988              /* Update vectorial force */
 989             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 990             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 991             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 992
 993             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 994             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 995             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 996
 997             }
 998
 999             /**************************
1000              * CALCULATE INTERACTIONS *
1001              **************************/
1002
1003             if (gmx_mm_any_lt(rsq31,rcutoff2))
1004             {
1005
1006             /* REACTION-FIELD ELECTROSTATICS */
1007             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
1008             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1009
1010             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1011
1012             /* Update potential sum for this i atom from the interaction with this j atom. */
1013             velec            = _mm_and_ps(velec,cutoff_mask);
1014             velec            = _mm_andnot_ps(dummy_mask,velec);
1015             velecsum         = _mm_add_ps(velecsum,velec);
1016
1017             fscal            = felec;
1018
1019             fscal            = _mm_and_ps(fscal,cutoff_mask);
1020
1021             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1022
1023              /* Update vectorial force */
1024             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1025             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1026             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1027
1028             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1029             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1030             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1031
1032             }
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             if (gmx_mm_any_lt(rsq32,rcutoff2))
1039             {
1040
1041             /* REACTION-FIELD ELECTROSTATICS */
1042             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
1043             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1044
1045             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1046
1047             /* Update potential sum for this i atom from the interaction with this j atom. */
1048             velec            = _mm_and_ps(velec,cutoff_mask);
1049             velec            = _mm_andnot_ps(dummy_mask,velec);
1050             velecsum         = _mm_add_ps(velecsum,velec);
1051
1052             fscal            = felec;
1053
1054             fscal            = _mm_and_ps(fscal,cutoff_mask);
1055
1056             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1057
1058              /* Update vectorial force */
1059             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1060             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1061             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1062
1063             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1064             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1065             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1066
1067             }
1068
1069             /**************************
1070              * CALCULATE INTERACTIONS *
1071              **************************/
1072
1073             if (gmx_mm_any_lt(rsq33,rcutoff2))
1074             {
1075
1076             /* REACTION-FIELD ELECTROSTATICS */
1077             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
1078             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1079
1080             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1081
1082             /* Update potential sum for this i atom from the interaction with this j atom. */
1083             velec            = _mm_and_ps(velec,cutoff_mask);
1084             velec            = _mm_andnot_ps(dummy_mask,velec);
1085             velecsum         = _mm_add_ps(velecsum,velec);
1086
1087             fscal            = felec;
1088
1089             fscal            = _mm_and_ps(fscal,cutoff_mask);
1090
1091             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1092
1093              /* Update vectorial force */
1094             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1095             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1096             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1097
1098             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1099             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1100             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1101
1102             }
1103
1104             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1105             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1106             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1107             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1108
1109             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1110                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1111                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1112
1113             /* Inner loop uses 398 flops */
1114         }
1115
1116         /* End of innermost loop */
1117
1118         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1119                                               f+i_coord_offset,fshift+i_shift_offset);
1120
1121         ggid                        = gid[iidx];
1122         /* Update potential energies */
1123         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1124         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1125
1126         /* Increment number of inner iterations */
1127         inneriter                  += j_index_end - j_index_start;
1128
1129         /* Outer loop uses 26 flops */
1130     }
1131
1132     /* Increment number of outer iterations */
1133     outeriter        += nri;
1134
1135     /* Update outer/inner flops */
1136
1137     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*398);
1138 }
1139 /*
1140  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_single
1141  * Electrostatics interaction: ReactionField
1142  * VdW interaction:            LennardJones
1143  * Geometry:                   Water4-Water4
1144  * Calculate force/pot:        Force
1145  */
1146 void
1147 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_single
1148                     (t_nblist * gmx_restrict                nlist,
1149                      rvec * gmx_restrict                    xx,
1150                      rvec * gmx_restrict                    ff,
1151                      t_forcerec * gmx_restrict              fr,
1152                      t_mdatoms * gmx_restrict               mdatoms,
1153                      nb_kernel_data_t * gmx_restrict        kernel_data,
1154                      t_nrnb * gmx_restrict                  nrnb)
1155 {
1156     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1157      * just 0 for non-waters.
1158      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1159      * jnr indices corresponding to data put in the four positions in the SIMD register.
1160      */
1161     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1162     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1163     int              jnrA,jnrB,jnrC,jnrD;
1164     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1165     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1166     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1167     real             rcutoff_scalar;
1168     real             *shiftvec,*fshift,*x,*f;
1169     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1170     real             scratch[4*DIM];
1171     __m128           fscal,rcutoff,rcutoff2,jidxall;
1172     int              vdwioffset0;
1173     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1174     int              vdwioffset1;
1175     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1176     int              vdwioffset2;
1177     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1178     int              vdwioffset3;
1179     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1180     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1181     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1182     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1183     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1184     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1185     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1186     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1187     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1188     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1189     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1190     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1191     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1192     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1193     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1194     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1195     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1196     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1197     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1198     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1199     real             *charge;
1200     int              nvdwtype;
1201     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1202     int              *vdwtype;
1203     real             *vdwparam;
1204     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1205     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1206     __m128           dummy_mask,cutoff_mask;
1207     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1208     __m128           one     = _mm_set1_ps(1.0);
1209     __m128           two     = _mm_set1_ps(2.0);
1210     x                = xx[0];
1211     f                = ff[0];
1212
1213     nri              = nlist->nri;
1214     iinr             = nlist->iinr;
1215     jindex           = nlist->jindex;
1216     jjnr             = nlist->jjnr;
1217     shiftidx         = nlist->shift;
1218     gid              = nlist->gid;
1219     shiftvec         = fr->shift_vec[0];
1220     fshift           = fr->fshift[0];
1221     facel            = _mm_set1_ps(fr->epsfac);
1222     charge           = mdatoms->chargeA;
1223     krf              = _mm_set1_ps(fr->ic->k_rf);
1224     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1225     crf              = _mm_set1_ps(fr->ic->c_rf);
1226     nvdwtype         = fr->ntype;
1227     vdwparam         = fr->nbfp;
1228     vdwtype          = mdatoms->typeA;
1229
1230     /* Setup water-specific parameters */
1231     inr              = nlist->iinr[0];
1232     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1233     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1234     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1235     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1236
1237     jq1              = _mm_set1_ps(charge[inr+1]);
1238     jq2              = _mm_set1_ps(charge[inr+2]);
1239     jq3              = _mm_set1_ps(charge[inr+3]);
1240     vdwjidx0A        = 2*vdwtype[inr+0];
1241     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1242     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1243     qq11             = _mm_mul_ps(iq1,jq1);
1244     qq12             = _mm_mul_ps(iq1,jq2);
1245     qq13             = _mm_mul_ps(iq1,jq3);
1246     qq21             = _mm_mul_ps(iq2,jq1);
1247     qq22             = _mm_mul_ps(iq2,jq2);
1248     qq23             = _mm_mul_ps(iq2,jq3);
1249     qq31             = _mm_mul_ps(iq3,jq1);
1250     qq32             = _mm_mul_ps(iq3,jq2);
1251     qq33             = _mm_mul_ps(iq3,jq3);
1252
1253     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1254     rcutoff_scalar   = fr->rcoulomb;
1255     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1256     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1257
1258     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1259     rvdw             = _mm_set1_ps(fr->rvdw);
1260
1261     /* Avoid stupid compiler warnings */
1262     jnrA = jnrB = jnrC = jnrD = 0;
1263     j_coord_offsetA = 0;
1264     j_coord_offsetB = 0;
1265     j_coord_offsetC = 0;
1266     j_coord_offsetD = 0;
1267
1268     outeriter        = 0;
1269     inneriter        = 0;
1270
1271     for(iidx=0;iidx<4*DIM;iidx++)
1272     {
1273         scratch[iidx] = 0.0;
1274     }
1275
1276     /* Start outer loop over neighborlists */
1277     for(iidx=0; iidx<nri; iidx++)
1278     {
1279         /* Load shift vector for this list */
1280         i_shift_offset   = DIM*shiftidx[iidx];
1281
1282         /* Load limits for loop over neighbors */
1283         j_index_start    = jindex[iidx];
1284         j_index_end      = jindex[iidx+1];
1285
1286         /* Get outer coordinate index */
1287         inr              = iinr[iidx];
1288         i_coord_offset   = DIM*inr;
1289
1290         /* Load i particle coords and add shift vector */
1291         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1292                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1293
1294         fix0             = _mm_setzero_ps();
1295         fiy0             = _mm_setzero_ps();
1296         fiz0             = _mm_setzero_ps();
1297         fix1             = _mm_setzero_ps();
1298         fiy1             = _mm_setzero_ps();
1299         fiz1             = _mm_setzero_ps();
1300         fix2             = _mm_setzero_ps();
1301         fiy2             = _mm_setzero_ps();
1302         fiz2             = _mm_setzero_ps();
1303         fix3             = _mm_setzero_ps();
1304         fiy3             = _mm_setzero_ps();
1305         fiz3             = _mm_setzero_ps();
1306
1307         /* Start inner kernel loop */
1308         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1309         {
1310
1311             /* Get j neighbor index, and coordinate index */
1312             jnrA             = jjnr[jidx];
1313             jnrB             = jjnr[jidx+1];
1314             jnrC             = jjnr[jidx+2];
1315             jnrD             = jjnr[jidx+3];
1316             j_coord_offsetA  = DIM*jnrA;
1317             j_coord_offsetB  = DIM*jnrB;
1318             j_coord_offsetC  = DIM*jnrC;
1319             j_coord_offsetD  = DIM*jnrD;
1320
1321             /* load j atom coordinates */
1322             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1323                                               x+j_coord_offsetC,x+j_coord_offsetD,
1324                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1325                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1326
1327             /* Calculate displacement vector */
1328             dx00             = _mm_sub_ps(ix0,jx0);
1329             dy00             = _mm_sub_ps(iy0,jy0);
1330             dz00             = _mm_sub_ps(iz0,jz0);
1331             dx11             = _mm_sub_ps(ix1,jx1);
1332             dy11             = _mm_sub_ps(iy1,jy1);
1333             dz11             = _mm_sub_ps(iz1,jz1);
1334             dx12             = _mm_sub_ps(ix1,jx2);
1335             dy12             = _mm_sub_ps(iy1,jy2);
1336             dz12             = _mm_sub_ps(iz1,jz2);
1337             dx13             = _mm_sub_ps(ix1,jx3);
1338             dy13             = _mm_sub_ps(iy1,jy3);
1339             dz13             = _mm_sub_ps(iz1,jz3);
1340             dx21             = _mm_sub_ps(ix2,jx1);
1341             dy21             = _mm_sub_ps(iy2,jy1);
1342             dz21             = _mm_sub_ps(iz2,jz1);
1343             dx22             = _mm_sub_ps(ix2,jx2);
1344             dy22             = _mm_sub_ps(iy2,jy2);
1345             dz22             = _mm_sub_ps(iz2,jz2);
1346             dx23             = _mm_sub_ps(ix2,jx3);
1347             dy23             = _mm_sub_ps(iy2,jy3);
1348             dz23             = _mm_sub_ps(iz2,jz3);
1349             dx31             = _mm_sub_ps(ix3,jx1);
1350             dy31             = _mm_sub_ps(iy3,jy1);
1351             dz31             = _mm_sub_ps(iz3,jz1);
1352             dx32             = _mm_sub_ps(ix3,jx2);
1353             dy32             = _mm_sub_ps(iy3,jy2);
1354             dz32             = _mm_sub_ps(iz3,jz2);
1355             dx33             = _mm_sub_ps(ix3,jx3);
1356             dy33             = _mm_sub_ps(iy3,jy3);
1357             dz33             = _mm_sub_ps(iz3,jz3);
1358
1359             /* Calculate squared distance and things based on it */
1360             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1361             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1362             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1363             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1364             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1365             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1366             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1367             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1368             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1369             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1370
1371             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1372             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1373             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1374             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1375             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1376             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1377             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1378             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1379             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1380
1381             rinvsq00         = gmx_mm_inv_ps(rsq00);
1382             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1383             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1384             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1385             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1386             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1387             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1388             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1389             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1390             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1391
1392             fjx0             = _mm_setzero_ps();
1393             fjy0             = _mm_setzero_ps();
1394             fjz0             = _mm_setzero_ps();
1395             fjx1             = _mm_setzero_ps();
1396             fjy1             = _mm_setzero_ps();
1397             fjz1             = _mm_setzero_ps();
1398             fjx2             = _mm_setzero_ps();
1399             fjy2             = _mm_setzero_ps();
1400             fjz2             = _mm_setzero_ps();
1401             fjx3             = _mm_setzero_ps();
1402             fjy3             = _mm_setzero_ps();
1403             fjz3             = _mm_setzero_ps();
1404
1405             /**************************
1406              * CALCULATE INTERACTIONS *
1407              **************************/
1408
1409             if (gmx_mm_any_lt(rsq00,rcutoff2))
1410             {
1411
1412             /* LENNARD-JONES DISPERSION/REPULSION */
1413
1414             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1415             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1416
1417             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1418
1419             fscal            = fvdw;
1420
1421             fscal            = _mm_and_ps(fscal,cutoff_mask);
1422
1423              /* Update vectorial force */
1424             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1425             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1426             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1427
1428             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1429             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1430             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1431
1432             }
1433
1434             /**************************
1435              * CALCULATE INTERACTIONS *
1436              **************************/
1437
1438             if (gmx_mm_any_lt(rsq11,rcutoff2))
1439             {
1440
1441             /* REACTION-FIELD ELECTROSTATICS */
1442             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1443
1444             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1445
1446             fscal            = felec;
1447
1448             fscal            = _mm_and_ps(fscal,cutoff_mask);
1449
1450              /* Update vectorial force */
1451             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1452             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1453             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1454
1455             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1456             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1457             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1458
1459             }
1460
1461             /**************************
1462              * CALCULATE INTERACTIONS *
1463              **************************/
1464
1465             if (gmx_mm_any_lt(rsq12,rcutoff2))
1466             {
1467
1468             /* REACTION-FIELD ELECTROSTATICS */
1469             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1470
1471             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1472
1473             fscal            = felec;
1474
1475             fscal            = _mm_and_ps(fscal,cutoff_mask);
1476
1477              /* Update vectorial force */
1478             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1479             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1480             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1481
1482             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1483             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1484             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1485
1486             }
1487
1488             /**************************
1489              * CALCULATE INTERACTIONS *
1490              **************************/
1491
1492             if (gmx_mm_any_lt(rsq13,rcutoff2))
1493             {
1494
1495             /* REACTION-FIELD ELECTROSTATICS */
1496             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1497
1498             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1499
1500             fscal            = felec;
1501
1502             fscal            = _mm_and_ps(fscal,cutoff_mask);
1503
1504              /* Update vectorial force */
1505             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1506             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1507             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1508
1509             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1510             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1511             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1512
1513             }
1514
1515             /**************************
1516              * CALCULATE INTERACTIONS *
1517              **************************/
1518
1519             if (gmx_mm_any_lt(rsq21,rcutoff2))
1520             {
1521
1522             /* REACTION-FIELD ELECTROSTATICS */
1523             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1524
1525             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1526
1527             fscal            = felec;
1528
1529             fscal            = _mm_and_ps(fscal,cutoff_mask);
1530
1531              /* Update vectorial force */
1532             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1533             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1534             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1535
1536             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1537             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1538             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1539
1540             }
1541
1542             /**************************
1543              * CALCULATE INTERACTIONS *
1544              **************************/
1545
1546             if (gmx_mm_any_lt(rsq22,rcutoff2))
1547             {
1548
1549             /* REACTION-FIELD ELECTROSTATICS */
1550             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1551
1552             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1553
1554             fscal            = felec;
1555
1556             fscal            = _mm_and_ps(fscal,cutoff_mask);
1557
1558              /* Update vectorial force */
1559             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1560             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1561             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1562
1563             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1564             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1565             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1566
1567             }
1568
1569             /**************************
1570              * CALCULATE INTERACTIONS *
1571              **************************/
1572
1573             if (gmx_mm_any_lt(rsq23,rcutoff2))
1574             {
1575
1576             /* REACTION-FIELD ELECTROSTATICS */
1577             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1578
1579             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1580
1581             fscal            = felec;
1582
1583             fscal            = _mm_and_ps(fscal,cutoff_mask);
1584
1585              /* Update vectorial force */
1586             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1587             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1588             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1589
1590             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1591             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1592             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1593
1594             }
1595
1596             /**************************
1597              * CALCULATE INTERACTIONS *
1598              **************************/
1599
1600             if (gmx_mm_any_lt(rsq31,rcutoff2))
1601             {
1602
1603             /* REACTION-FIELD ELECTROSTATICS */
1604             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1605
1606             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1607
1608             fscal            = felec;
1609
1610             fscal            = _mm_and_ps(fscal,cutoff_mask);
1611
1612              /* Update vectorial force */
1613             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1614             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1615             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1616
1617             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1618             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1619             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1620
1621             }
1622
1623             /**************************
1624              * CALCULATE INTERACTIONS *
1625              **************************/
1626
1627             if (gmx_mm_any_lt(rsq32,rcutoff2))
1628             {
1629
1630             /* REACTION-FIELD ELECTROSTATICS */
1631             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1632
1633             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1634
1635             fscal            = felec;
1636
1637             fscal            = _mm_and_ps(fscal,cutoff_mask);
1638
1639              /* Update vectorial force */
1640             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1641             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1642             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1643
1644             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1645             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1646             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1647
1648             }
1649
1650             /**************************
1651              * CALCULATE INTERACTIONS *
1652              **************************/
1653
1654             if (gmx_mm_any_lt(rsq33,rcutoff2))
1655             {
1656
1657             /* REACTION-FIELD ELECTROSTATICS */
1658             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1659
1660             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1661
1662             fscal            = felec;
1663
1664             fscal            = _mm_and_ps(fscal,cutoff_mask);
1665
1666              /* Update vectorial force */
1667             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1668             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1669             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1670
1671             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1672             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1673             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1674
1675             }
1676
1677             fjptrA             = f+j_coord_offsetA;
1678             fjptrB             = f+j_coord_offsetB;
1679             fjptrC             = f+j_coord_offsetC;
1680             fjptrD             = f+j_coord_offsetD;
1681
1682             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1683                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1684                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1685
1686             /* Inner loop uses 333 flops */
1687         }
1688
1689         if(jidx<j_index_end)
1690         {
1691
1692             /* Get j neighbor index, and coordinate index */
1693             jnrlistA         = jjnr[jidx];
1694             jnrlistB         = jjnr[jidx+1];
1695             jnrlistC         = jjnr[jidx+2];
1696             jnrlistD         = jjnr[jidx+3];
1697             /* Sign of each element will be negative for non-real atoms.
1698              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1699              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1700              */
1701             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1702             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1703             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1704             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1705             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1706             j_coord_offsetA  = DIM*jnrA;
1707             j_coord_offsetB  = DIM*jnrB;
1708             j_coord_offsetC  = DIM*jnrC;
1709             j_coord_offsetD  = DIM*jnrD;
1710
1711             /* load j atom coordinates */
1712             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1713                                               x+j_coord_offsetC,x+j_coord_offsetD,
1714                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1715                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1716
1717             /* Calculate displacement vector */
1718             dx00             = _mm_sub_ps(ix0,jx0);
1719             dy00             = _mm_sub_ps(iy0,jy0);
1720             dz00             = _mm_sub_ps(iz0,jz0);
1721             dx11             = _mm_sub_ps(ix1,jx1);
1722             dy11             = _mm_sub_ps(iy1,jy1);
1723             dz11             = _mm_sub_ps(iz1,jz1);
1724             dx12             = _mm_sub_ps(ix1,jx2);
1725             dy12             = _mm_sub_ps(iy1,jy2);
1726             dz12             = _mm_sub_ps(iz1,jz2);
1727             dx13             = _mm_sub_ps(ix1,jx3);
1728             dy13             = _mm_sub_ps(iy1,jy3);
1729             dz13             = _mm_sub_ps(iz1,jz3);
1730             dx21             = _mm_sub_ps(ix2,jx1);
1731             dy21             = _mm_sub_ps(iy2,jy1);
1732             dz21             = _mm_sub_ps(iz2,jz1);
1733             dx22             = _mm_sub_ps(ix2,jx2);
1734             dy22             = _mm_sub_ps(iy2,jy2);
1735             dz22             = _mm_sub_ps(iz2,jz2);
1736             dx23             = _mm_sub_ps(ix2,jx3);
1737             dy23             = _mm_sub_ps(iy2,jy3);
1738             dz23             = _mm_sub_ps(iz2,jz3);
1739             dx31             = _mm_sub_ps(ix3,jx1);
1740             dy31             = _mm_sub_ps(iy3,jy1);
1741             dz31             = _mm_sub_ps(iz3,jz1);
1742             dx32             = _mm_sub_ps(ix3,jx2);
1743             dy32             = _mm_sub_ps(iy3,jy2);
1744             dz32             = _mm_sub_ps(iz3,jz2);
1745             dx33             = _mm_sub_ps(ix3,jx3);
1746             dy33             = _mm_sub_ps(iy3,jy3);
1747             dz33             = _mm_sub_ps(iz3,jz3);
1748
1749             /* Calculate squared distance and things based on it */
1750             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1751             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1752             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1753             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1754             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1755             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1756             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1757             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1758             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1759             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1760
1761             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1762             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1763             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1764             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1765             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1766             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1767             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1768             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1769             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1770
1771             rinvsq00         = gmx_mm_inv_ps(rsq00);
1772             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1773             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1774             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1775             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1776             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1777             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1778             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1779             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1780             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1781
1782             fjx0             = _mm_setzero_ps();
1783             fjy0             = _mm_setzero_ps();
1784             fjz0             = _mm_setzero_ps();
1785             fjx1             = _mm_setzero_ps();
1786             fjy1             = _mm_setzero_ps();
1787             fjz1             = _mm_setzero_ps();
1788             fjx2             = _mm_setzero_ps();
1789             fjy2             = _mm_setzero_ps();
1790             fjz2             = _mm_setzero_ps();
1791             fjx3             = _mm_setzero_ps();
1792             fjy3             = _mm_setzero_ps();
1793             fjz3             = _mm_setzero_ps();
1794
1795             /**************************
1796              * CALCULATE INTERACTIONS *
1797              **************************/
1798
1799             if (gmx_mm_any_lt(rsq00,rcutoff2))
1800             {
1801
1802             /* LENNARD-JONES DISPERSION/REPULSION */
1803
1804             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1805             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1806
1807             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1808
1809             fscal            = fvdw;
1810
1811             fscal            = _mm_and_ps(fscal,cutoff_mask);
1812
1813             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1814
1815              /* Update vectorial force */
1816             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1817             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1818             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1819
1820             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1821             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1822             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1823
1824             }
1825
1826             /**************************
1827              * CALCULATE INTERACTIONS *
1828              **************************/
1829
1830             if (gmx_mm_any_lt(rsq11,rcutoff2))
1831             {
1832
1833             /* REACTION-FIELD ELECTROSTATICS */
1834             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1835
1836             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1837
1838             fscal            = felec;
1839
1840             fscal            = _mm_and_ps(fscal,cutoff_mask);
1841
1842             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1843
1844              /* Update vectorial force */
1845             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1846             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1847             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1848
1849             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1850             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1851             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1852
1853             }
1854
1855             /**************************
1856              * CALCULATE INTERACTIONS *
1857              **************************/
1858
1859             if (gmx_mm_any_lt(rsq12,rcutoff2))
1860             {
1861
1862             /* REACTION-FIELD ELECTROSTATICS */
1863             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1864
1865             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1866
1867             fscal            = felec;
1868
1869             fscal            = _mm_and_ps(fscal,cutoff_mask);
1870
1871             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1872
1873              /* Update vectorial force */
1874             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1875             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1876             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1877
1878             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1879             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1880             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1881
1882             }
1883
1884             /**************************
1885              * CALCULATE INTERACTIONS *
1886              **************************/
1887
1888             if (gmx_mm_any_lt(rsq13,rcutoff2))
1889             {
1890
1891             /* REACTION-FIELD ELECTROSTATICS */
1892             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1893
1894             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1895
1896             fscal            = felec;
1897
1898             fscal            = _mm_and_ps(fscal,cutoff_mask);
1899
1900             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1901
1902              /* Update vectorial force */
1903             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1904             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1905             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1906
1907             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1908             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1909             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1910
1911             }
1912
1913             /**************************
1914              * CALCULATE INTERACTIONS *
1915              **************************/
1916
1917             if (gmx_mm_any_lt(rsq21,rcutoff2))
1918             {
1919
1920             /* REACTION-FIELD ELECTROSTATICS */
1921             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1922
1923             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1924
1925             fscal            = felec;
1926
1927             fscal            = _mm_and_ps(fscal,cutoff_mask);
1928
1929             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1930
1931              /* Update vectorial force */
1932             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1933             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1934             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1935
1936             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1937             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1938             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1939
1940             }
1941
1942             /**************************
1943              * CALCULATE INTERACTIONS *
1944              **************************/
1945
1946             if (gmx_mm_any_lt(rsq22,rcutoff2))
1947             {
1948
1949             /* REACTION-FIELD ELECTROSTATICS */
1950             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1951
1952             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1953
1954             fscal            = felec;
1955
1956             fscal            = _mm_and_ps(fscal,cutoff_mask);
1957
1958             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1959
1960              /* Update vectorial force */
1961             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1962             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1963             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1964
1965             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1966             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1967             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1968
1969             }
1970
1971             /**************************
1972              * CALCULATE INTERACTIONS *
1973              **************************/
1974
1975             if (gmx_mm_any_lt(rsq23,rcutoff2))
1976             {
1977
1978             /* REACTION-FIELD ELECTROSTATICS */
1979             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1980
1981             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1982
1983             fscal            = felec;
1984
1985             fscal            = _mm_and_ps(fscal,cutoff_mask);
1986
1987             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1988
1989              /* Update vectorial force */
1990             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1991             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1992             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1993
1994             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1995             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1996             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1997
1998             }
1999
2000             /**************************
2001              * CALCULATE INTERACTIONS *
2002              **************************/
2003
2004             if (gmx_mm_any_lt(rsq31,rcutoff2))
2005             {
2006
2007             /* REACTION-FIELD ELECTROSTATICS */
2008             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
2009
2010             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2011
2012             fscal            = felec;
2013
2014             fscal            = _mm_and_ps(fscal,cutoff_mask);
2015
2016             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2017
2018              /* Update vectorial force */
2019             fix3             = _mm_macc_ps(dx31,fscal,fix3);
2020             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
2021             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
2022
2023             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
2024             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
2025             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
2026
2027             }
2028
2029             /**************************
2030              * CALCULATE INTERACTIONS *
2031              **************************/
2032
2033             if (gmx_mm_any_lt(rsq32,rcutoff2))
2034             {
2035
2036             /* REACTION-FIELD ELECTROSTATICS */
2037             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
2038
2039             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2040
2041             fscal            = felec;
2042
2043             fscal            = _mm_and_ps(fscal,cutoff_mask);
2044
2045             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2046
2047              /* Update vectorial force */
2048             fix3             = _mm_macc_ps(dx32,fscal,fix3);
2049             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
2050             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
2051
2052             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
2053             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
2054             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
2055
2056             }
2057
2058             /**************************
2059              * CALCULATE INTERACTIONS *
2060              **************************/
2061
2062             if (gmx_mm_any_lt(rsq33,rcutoff2))
2063             {
2064
2065             /* REACTION-FIELD ELECTROSTATICS */
2066             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
2067
2068             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2069
2070             fscal            = felec;
2071
2072             fscal            = _mm_and_ps(fscal,cutoff_mask);
2073
2074             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2075
2076              /* Update vectorial force */
2077             fix3             = _mm_macc_ps(dx33,fscal,fix3);
2078             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
2079             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
2080
2081             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
2082             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
2083             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
2084
2085             }
2086
2087             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2088             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2089             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2090             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2091
2092             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2093                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2094                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2095
2096             /* Inner loop uses 333 flops */
2097         }
2098
2099         /* End of innermost loop */
2100
2101         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2102                                               f+i_coord_offset,fshift+i_shift_offset);
2103
2104         /* Increment number of inner iterations */
2105         inneriter                  += j_index_end - j_index_start;
2106
2107         /* Outer loop uses 24 flops */
2108     }
2109
2110     /* Increment number of outer iterations */
2111     outeriter        += nri;
2112
2113     /* Update outer/inner flops */
2114
2115     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*333);
2116 }