src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRF_VdwCSTab_GeomW4W4_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwioffset3;
  92     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  93     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  94     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  95     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  96     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  97     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  98     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  99     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 100     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 101     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 102     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 103     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 104     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 105     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 106     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 107     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 108     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 109     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 110     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 111     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 112     real             *charge;
 113     int              nvdwtype;
 114     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 115     int              *vdwtype;
 116     real             *vdwparam;
 117     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 118     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 119     __m128i          vfitab;
 120     __m128i          ifour       = _mm_set1_epi32(4);
 121     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 122     real             *vftab;
 123     __m128           dummy_mask,cutoff_mask;
 124     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 125     __m128           one     = _mm_set1_ps(1.0);
 126     __m128           two     = _mm_set1_ps(2.0);
 127     x                = xx[0];
 128     f                = ff[0];
 129
 130     nri              = nlist->nri;
 131     iinr             = nlist->iinr;
 132     jindex           = nlist->jindex;
 133     jjnr             = nlist->jjnr;
 134     shiftidx         = nlist->shift;
 135     gid              = nlist->gid;
 136     shiftvec         = fr->shift_vec[0];
 137     fshift           = fr->fshift[0];
 138     facel            = _mm_set1_ps(fr->epsfac);
 139     charge           = mdatoms->chargeA;
 140     krf              = _mm_set1_ps(fr->ic->k_rf);
 141     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 142     crf              = _mm_set1_ps(fr->ic->c_rf);
 143     nvdwtype         = fr->ntype;
 144     vdwparam         = fr->nbfp;
 145     vdwtype          = mdatoms->typeA;
 146
 147     vftab            = kernel_data->table_vdw->data;
 148     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 149
 150     /* Setup water-specific parameters */
 151     inr              = nlist->iinr[0];
 152     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 153     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 154     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 155     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 156
 157     jq1              = _mm_set1_ps(charge[inr+1]);
 158     jq2              = _mm_set1_ps(charge[inr+2]);
 159     jq3              = _mm_set1_ps(charge[inr+3]);
 160     vdwjidx0A        = 2*vdwtype[inr+0];
 161     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 162     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 163     qq11             = _mm_mul_ps(iq1,jq1);
 164     qq12             = _mm_mul_ps(iq1,jq2);
 165     qq13             = _mm_mul_ps(iq1,jq3);
 166     qq21             = _mm_mul_ps(iq2,jq1);
 167     qq22             = _mm_mul_ps(iq2,jq2);
 168     qq23             = _mm_mul_ps(iq2,jq3);
 169     qq31             = _mm_mul_ps(iq3,jq1);
 170     qq32             = _mm_mul_ps(iq3,jq2);
 171     qq33             = _mm_mul_ps(iq3,jq3);
 172
 173     /* Avoid stupid compiler warnings */
 174     jnrA = jnrB = jnrC = jnrD = 0;
 175     j_coord_offsetA = 0;
 176     j_coord_offsetB = 0;
 177     j_coord_offsetC = 0;
 178     j_coord_offsetD = 0;
 179
 180     outeriter        = 0;
 181     inneriter        = 0;
 182
 183     for(iidx=0;iidx<4*DIM;iidx++)
 184     {
 185         scratch[iidx] = 0.0;
 186     }
 187
 188     /* Start outer loop over neighborlists */
 189     for(iidx=0; iidx<nri; iidx++)
 190     {
 191         /* Load shift vector for this list */
 192         i_shift_offset   = DIM*shiftidx[iidx];
 193
 194         /* Load limits for loop over neighbors */
 195         j_index_start    = jindex[iidx];
 196         j_index_end      = jindex[iidx+1];
 197
 198         /* Get outer coordinate index */
 199         inr              = iinr[iidx];
 200         i_coord_offset   = DIM*inr;
 201
 202         /* Load i particle coords and add shift vector */
 203         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 204                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 205
 206         fix0             = _mm_setzero_ps();
 207         fiy0             = _mm_setzero_ps();
 208         fiz0             = _mm_setzero_ps();
 209         fix1             = _mm_setzero_ps();
 210         fiy1             = _mm_setzero_ps();
 211         fiz1             = _mm_setzero_ps();
 212         fix2             = _mm_setzero_ps();
 213         fiy2             = _mm_setzero_ps();
 214         fiz2             = _mm_setzero_ps();
 215         fix3             = _mm_setzero_ps();
 216         fiy3             = _mm_setzero_ps();
 217         fiz3             = _mm_setzero_ps();
 218
 219         /* Reset potential sums */
 220         velecsum         = _mm_setzero_ps();
 221         vvdwsum          = _mm_setzero_ps();
 222
 223         /* Start inner kernel loop */
 224         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 225         {
 226
 227             /* Get j neighbor index, and coordinate index */
 228             jnrA             = jjnr[jidx];
 229             jnrB             = jjnr[jidx+1];
 230             jnrC             = jjnr[jidx+2];
 231             jnrD             = jjnr[jidx+3];
 232             j_coord_offsetA  = DIM*jnrA;
 233             j_coord_offsetB  = DIM*jnrB;
 234             j_coord_offsetC  = DIM*jnrC;
 235             j_coord_offsetD  = DIM*jnrD;
 236
 237             /* load j atom coordinates */
 238             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 239                                               x+j_coord_offsetC,x+j_coord_offsetD,
 240                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 241                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 242
 243             /* Calculate displacement vector */
 244             dx00             = _mm_sub_ps(ix0,jx0);
 245             dy00             = _mm_sub_ps(iy0,jy0);
 246             dz00             = _mm_sub_ps(iz0,jz0);
 247             dx11             = _mm_sub_ps(ix1,jx1);
 248             dy11             = _mm_sub_ps(iy1,jy1);
 249             dz11             = _mm_sub_ps(iz1,jz1);
 250             dx12             = _mm_sub_ps(ix1,jx2);
 251             dy12             = _mm_sub_ps(iy1,jy2);
 252             dz12             = _mm_sub_ps(iz1,jz2);
 253             dx13             = _mm_sub_ps(ix1,jx3);
 254             dy13             = _mm_sub_ps(iy1,jy3);
 255             dz13             = _mm_sub_ps(iz1,jz3);
 256             dx21             = _mm_sub_ps(ix2,jx1);
 257             dy21             = _mm_sub_ps(iy2,jy1);
 258             dz21             = _mm_sub_ps(iz2,jz1);
 259             dx22             = _mm_sub_ps(ix2,jx2);
 260             dy22             = _mm_sub_ps(iy2,jy2);
 261             dz22             = _mm_sub_ps(iz2,jz2);
 262             dx23             = _mm_sub_ps(ix2,jx3);
 263             dy23             = _mm_sub_ps(iy2,jy3);
 264             dz23             = _mm_sub_ps(iz2,jz3);
 265             dx31             = _mm_sub_ps(ix3,jx1);
 266             dy31             = _mm_sub_ps(iy3,jy1);
 267             dz31             = _mm_sub_ps(iz3,jz1);
 268             dx32             = _mm_sub_ps(ix3,jx2);
 269             dy32             = _mm_sub_ps(iy3,jy2);
 270             dz32             = _mm_sub_ps(iz3,jz2);
 271             dx33             = _mm_sub_ps(ix3,jx3);
 272             dy33             = _mm_sub_ps(iy3,jy3);
 273             dz33             = _mm_sub_ps(iz3,jz3);
 274
 275             /* Calculate squared distance and things based on it */
 276             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 277             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 278             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 279             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 280             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 281             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 282             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 283             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 284             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 285             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 286
 287             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 288             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 289             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 290             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 291             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 292             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 293             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 294             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 295             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 296             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 297
 298             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 299             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 300             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 301             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 302             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 303             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 304             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 305             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 306             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 307
 308             fjx0             = _mm_setzero_ps();
 309             fjy0             = _mm_setzero_ps();
 310             fjz0             = _mm_setzero_ps();
 311             fjx1             = _mm_setzero_ps();
 312             fjy1             = _mm_setzero_ps();
 313             fjz1             = _mm_setzero_ps();
 314             fjx2             = _mm_setzero_ps();
 315             fjy2             = _mm_setzero_ps();
 316             fjz2             = _mm_setzero_ps();
 317             fjx3             = _mm_setzero_ps();
 318             fjy3             = _mm_setzero_ps();
 319             fjz3             = _mm_setzero_ps();
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             r00              = _mm_mul_ps(rsq00,rinv00);
 326
 327             /* Calculate table index by multiplying r with table scale and truncate to integer */
 328             rt               = _mm_mul_ps(r00,vftabscale);
 329             vfitab           = _mm_cvttps_epi32(rt);
 330 #ifdef __XOP__
 331             vfeps            = _mm_frcz_ps(rt);
 332 #else
 333             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 334 #endif
 335             twovfeps         = _mm_add_ps(vfeps,vfeps);
 336             vfitab           = _mm_slli_epi32(vfitab,3);
 337
 338             /* CUBIC SPLINE TABLE DISPERSION */
 339             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 340             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 341             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 342             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 343             _MM_TRANSPOSE4_PS(Y,F,G,H);
 344             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 345             VV               = _mm_macc_ps(vfeps,Fp,Y);
 346             vvdw6            = _mm_mul_ps(c6_00,VV);
 347             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 348             fvdw6            = _mm_mul_ps(c6_00,FF);
 349
 350             /* CUBIC SPLINE TABLE REPULSION */
 351             vfitab           = _mm_add_epi32(vfitab,ifour);
 352             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 353             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 354             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 355             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 356             _MM_TRANSPOSE4_PS(Y,F,G,H);
 357             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 358             VV               = _mm_macc_ps(vfeps,Fp,Y);
 359             vvdw12           = _mm_mul_ps(c12_00,VV);
 360             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 361             fvdw12           = _mm_mul_ps(c12_00,FF);
 362             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 363             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 367
 368             fscal            = fvdw;
 369
 370              /* Update vectorial force */
 371             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 372             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 373             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 374
 375             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 376             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 377             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* REACTION-FIELD ELECTROSTATICS */
 384             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 385             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velecsum         = _mm_add_ps(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392              /* Update vectorial force */
 393             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 394             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 395             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 396
 397             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 398             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 399             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             /* REACTION-FIELD ELECTROSTATICS */
 406             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 407             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 408
 409             /* Update potential sum for this i atom from the interaction with this j atom. */
 410             velecsum         = _mm_add_ps(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414              /* Update vectorial force */
 415             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 416             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 417             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 418
 419             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 420             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 421             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             /* REACTION-FIELD ELECTROSTATICS */
 428             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 429             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 430
 431             /* Update potential sum for this i atom from the interaction with this j atom. */
 432             velecsum         = _mm_add_ps(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436              /* Update vectorial force */
 437             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 438             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 439             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 440
 441             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 442             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 443             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* REACTION-FIELD ELECTROSTATICS */
 450             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 451             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 452
 453             /* Update potential sum for this i atom from the interaction with this j atom. */
 454             velecsum         = _mm_add_ps(velecsum,velec);
 455
 456             fscal            = felec;
 457
 458              /* Update vectorial force */
 459             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 460             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 461             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 462
 463             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 464             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 465             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 466
 467             /**************************
 468              * CALCULATE INTERACTIONS *
 469              **************************/
 470
 471             /* REACTION-FIELD ELECTROSTATICS */
 472             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 473             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velecsum         = _mm_add_ps(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480              /* Update vectorial force */
 481             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 482             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 483             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 484
 485             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 486             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 487             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 488
 489             /**************************
 490              * CALCULATE INTERACTIONS *
 491              **************************/
 492
 493             /* REACTION-FIELD ELECTROSTATICS */
 494             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 495             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velecsum         = _mm_add_ps(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502              /* Update vectorial force */
 503             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 504             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 505             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 506
 507             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 508             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 509             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 510
 511             /**************************
 512              * CALCULATE INTERACTIONS *
 513              **************************/
 514
 515             /* REACTION-FIELD ELECTROSTATICS */
 516             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 517             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 518
 519             /* Update potential sum for this i atom from the interaction with this j atom. */
 520             velecsum         = _mm_add_ps(velecsum,velec);
 521
 522             fscal            = felec;
 523
 524              /* Update vectorial force */
 525             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 526             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 527             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 528
 529             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 530             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 531             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 532
 533             /**************************
 534              * CALCULATE INTERACTIONS *
 535              **************************/
 536
 537             /* REACTION-FIELD ELECTROSTATICS */
 538             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 539             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 540
 541             /* Update potential sum for this i atom from the interaction with this j atom. */
 542             velecsum         = _mm_add_ps(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546              /* Update vectorial force */
 547             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 548             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 549             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 550
 551             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 552             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 553             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             /* REACTION-FIELD ELECTROSTATICS */
 560             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 561             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 562
 563             /* Update potential sum for this i atom from the interaction with this j atom. */
 564             velecsum         = _mm_add_ps(velecsum,velec);
 565
 566             fscal            = felec;
 567
 568              /* Update vectorial force */
 569             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 570             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 571             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 572
 573             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 574             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 575             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 576
 577             fjptrA             = f+j_coord_offsetA;
 578             fjptrB             = f+j_coord_offsetB;
 579             fjptrC             = f+j_coord_offsetC;
 580             fjptrD             = f+j_coord_offsetD;
 581
 582             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 583                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 584                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 585
 586             /* Inner loop uses 377 flops */
 587         }
 588
 589         if(jidx<j_index_end)
 590         {
 591
 592             /* Get j neighbor index, and coordinate index */
 593             jnrlistA         = jjnr[jidx];
 594             jnrlistB         = jjnr[jidx+1];
 595             jnrlistC         = jjnr[jidx+2];
 596             jnrlistD         = jjnr[jidx+3];
 597             /* Sign of each element will be negative for non-real atoms.
 598              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 599              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 600              */
 601             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 602             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 603             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 604             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 605             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 606             j_coord_offsetA  = DIM*jnrA;
 607             j_coord_offsetB  = DIM*jnrB;
 608             j_coord_offsetC  = DIM*jnrC;
 609             j_coord_offsetD  = DIM*jnrD;
 610
 611             /* load j atom coordinates */
 612             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 613                                               x+j_coord_offsetC,x+j_coord_offsetD,
 614                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 615                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 616
 617             /* Calculate displacement vector */
 618             dx00             = _mm_sub_ps(ix0,jx0);
 619             dy00             = _mm_sub_ps(iy0,jy0);
 620             dz00             = _mm_sub_ps(iz0,jz0);
 621             dx11             = _mm_sub_ps(ix1,jx1);
 622             dy11             = _mm_sub_ps(iy1,jy1);
 623             dz11             = _mm_sub_ps(iz1,jz1);
 624             dx12             = _mm_sub_ps(ix1,jx2);
 625             dy12             = _mm_sub_ps(iy1,jy2);
 626             dz12             = _mm_sub_ps(iz1,jz2);
 627             dx13             = _mm_sub_ps(ix1,jx3);
 628             dy13             = _mm_sub_ps(iy1,jy3);
 629             dz13             = _mm_sub_ps(iz1,jz3);
 630             dx21             = _mm_sub_ps(ix2,jx1);
 631             dy21             = _mm_sub_ps(iy2,jy1);
 632             dz21             = _mm_sub_ps(iz2,jz1);
 633             dx22             = _mm_sub_ps(ix2,jx2);
 634             dy22             = _mm_sub_ps(iy2,jy2);
 635             dz22             = _mm_sub_ps(iz2,jz2);
 636             dx23             = _mm_sub_ps(ix2,jx3);
 637             dy23             = _mm_sub_ps(iy2,jy3);
 638             dz23             = _mm_sub_ps(iz2,jz3);
 639             dx31             = _mm_sub_ps(ix3,jx1);
 640             dy31             = _mm_sub_ps(iy3,jy1);
 641             dz31             = _mm_sub_ps(iz3,jz1);
 642             dx32             = _mm_sub_ps(ix3,jx2);
 643             dy32             = _mm_sub_ps(iy3,jy2);
 644             dz32             = _mm_sub_ps(iz3,jz2);
 645             dx33             = _mm_sub_ps(ix3,jx3);
 646             dy33             = _mm_sub_ps(iy3,jy3);
 647             dz33             = _mm_sub_ps(iz3,jz3);
 648
 649             /* Calculate squared distance and things based on it */
 650             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 651             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 652             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 653             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 654             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 655             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 656             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 657             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 658             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 659             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 660
 661             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 662             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 663             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 664             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 665             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 666             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 667             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 668             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 669             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 670             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 671
 672             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 673             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 674             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 675             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 676             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 677             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 678             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 679             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 680             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 681
 682             fjx0             = _mm_setzero_ps();
 683             fjy0             = _mm_setzero_ps();
 684             fjz0             = _mm_setzero_ps();
 685             fjx1             = _mm_setzero_ps();
 686             fjy1             = _mm_setzero_ps();
 687             fjz1             = _mm_setzero_ps();
 688             fjx2             = _mm_setzero_ps();
 689             fjy2             = _mm_setzero_ps();
 690             fjz2             = _mm_setzero_ps();
 691             fjx3             = _mm_setzero_ps();
 692             fjy3             = _mm_setzero_ps();
 693             fjz3             = _mm_setzero_ps();
 694
 695             /**************************
 696              * CALCULATE INTERACTIONS *
 697              **************************/
 698
 699             r00              = _mm_mul_ps(rsq00,rinv00);
 700             r00              = _mm_andnot_ps(dummy_mask,r00);
 701
 702             /* Calculate table index by multiplying r with table scale and truncate to integer */
 703             rt               = _mm_mul_ps(r00,vftabscale);
 704             vfitab           = _mm_cvttps_epi32(rt);
 705 #ifdef __XOP__
 706             vfeps            = _mm_frcz_ps(rt);
 707 #else
 708             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 709 #endif
 710             twovfeps         = _mm_add_ps(vfeps,vfeps);
 711             vfitab           = _mm_slli_epi32(vfitab,3);
 712
 713             /* CUBIC SPLINE TABLE DISPERSION */
 714             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 715             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 716             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 717             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 718             _MM_TRANSPOSE4_PS(Y,F,G,H);
 719             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 720             VV               = _mm_macc_ps(vfeps,Fp,Y);
 721             vvdw6            = _mm_mul_ps(c6_00,VV);
 722             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 723             fvdw6            = _mm_mul_ps(c6_00,FF);
 724
 725             /* CUBIC SPLINE TABLE REPULSION */
 726             vfitab           = _mm_add_epi32(vfitab,ifour);
 727             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 728             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 729             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 730             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 731             _MM_TRANSPOSE4_PS(Y,F,G,H);
 732             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 733             VV               = _mm_macc_ps(vfeps,Fp,Y);
 734             vvdw12           = _mm_mul_ps(c12_00,VV);
 735             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 736             fvdw12           = _mm_mul_ps(c12_00,FF);
 737             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 738             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 739
 740             /* Update potential sum for this i atom from the interaction with this j atom. */
 741             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 742             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 743
 744             fscal            = fvdw;
 745
 746             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 747
 748              /* Update vectorial force */
 749             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 750             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 751             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 752
 753             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 754             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 755             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 756
 757             /**************************
 758              * CALCULATE INTERACTIONS *
 759              **************************/
 760
 761             /* REACTION-FIELD ELECTROSTATICS */
 762             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 763             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 764
 765             /* Update potential sum for this i atom from the interaction with this j atom. */
 766             velec            = _mm_andnot_ps(dummy_mask,velec);
 767             velecsum         = _mm_add_ps(velecsum,velec);
 768
 769             fscal            = felec;
 770
 771             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 772
 773              /* Update vectorial force */
 774             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 775             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 776             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 777
 778             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 779             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 780             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 781
 782             /**************************
 783              * CALCULATE INTERACTIONS *
 784              **************************/
 785
 786             /* REACTION-FIELD ELECTROSTATICS */
 787             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 788             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 789
 790             /* Update potential sum for this i atom from the interaction with this j atom. */
 791             velec            = _mm_andnot_ps(dummy_mask,velec);
 792             velecsum         = _mm_add_ps(velecsum,velec);
 793
 794             fscal            = felec;
 795
 796             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 797
 798              /* Update vectorial force */
 799             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 800             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 801             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 802
 803             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 804             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 805             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 806
 807             /**************************
 808              * CALCULATE INTERACTIONS *
 809              **************************/
 810
 811             /* REACTION-FIELD ELECTROSTATICS */
 812             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 813             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 814
 815             /* Update potential sum for this i atom from the interaction with this j atom. */
 816             velec            = _mm_andnot_ps(dummy_mask,velec);
 817             velecsum         = _mm_add_ps(velecsum,velec);
 818
 819             fscal            = felec;
 820
 821             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 822
 823              /* Update vectorial force */
 824             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 825             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 826             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 827
 828             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 829             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 830             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 831
 832             /**************************
 833              * CALCULATE INTERACTIONS *
 834              **************************/
 835
 836             /* REACTION-FIELD ELECTROSTATICS */
 837             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 838             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 839
 840             /* Update potential sum for this i atom from the interaction with this j atom. */
 841             velec            = _mm_andnot_ps(dummy_mask,velec);
 842             velecsum         = _mm_add_ps(velecsum,velec);
 843
 844             fscal            = felec;
 845
 846             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 847
 848              /* Update vectorial force */
 849             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 850             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 851             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 852
 853             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 854             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 855             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 856
 857             /**************************
 858              * CALCULATE INTERACTIONS *
 859              **************************/
 860
 861             /* REACTION-FIELD ELECTROSTATICS */
 862             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 863             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 864
 865             /* Update potential sum for this i atom from the interaction with this j atom. */
 866             velec            = _mm_andnot_ps(dummy_mask,velec);
 867             velecsum         = _mm_add_ps(velecsum,velec);
 868
 869             fscal            = felec;
 870
 871             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 872
 873              /* Update vectorial force */
 874             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 875             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 876             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 877
 878             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 879             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 880             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 881
 882             /**************************
 883              * CALCULATE INTERACTIONS *
 884              **************************/
 885
 886             /* REACTION-FIELD ELECTROSTATICS */
 887             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 888             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 889
 890             /* Update potential sum for this i atom from the interaction with this j atom. */
 891             velec            = _mm_andnot_ps(dummy_mask,velec);
 892             velecsum         = _mm_add_ps(velecsum,velec);
 893
 894             fscal            = felec;
 895
 896             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 897
 898              /* Update vectorial force */
 899             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 900             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 901             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 902
 903             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 904             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 905             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 906
 907             /**************************
 908              * CALCULATE INTERACTIONS *
 909              **************************/
 910
 911             /* REACTION-FIELD ELECTROSTATICS */
 912             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 913             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 914
 915             /* Update potential sum for this i atom from the interaction with this j atom. */
 916             velec            = _mm_andnot_ps(dummy_mask,velec);
 917             velecsum         = _mm_add_ps(velecsum,velec);
 918
 919             fscal            = felec;
 920
 921             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 922
 923              /* Update vectorial force */
 924             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 925             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 926             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 927
 928             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 929             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 930             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 931
 932             /**************************
 933              * CALCULATE INTERACTIONS *
 934              **************************/
 935
 936             /* REACTION-FIELD ELECTROSTATICS */
 937             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 938             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 939
 940             /* Update potential sum for this i atom from the interaction with this j atom. */
 941             velec            = _mm_andnot_ps(dummy_mask,velec);
 942             velecsum         = _mm_add_ps(velecsum,velec);
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 947
 948              /* Update vectorial force */
 949             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 950             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 951             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 952
 953             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 954             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 955             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 956
 957             /**************************
 958              * CALCULATE INTERACTIONS *
 959              **************************/
 960
 961             /* REACTION-FIELD ELECTROSTATICS */
 962             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 963             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 964
 965             /* Update potential sum for this i atom from the interaction with this j atom. */
 966             velec            = _mm_andnot_ps(dummy_mask,velec);
 967             velecsum         = _mm_add_ps(velecsum,velec);
 968
 969             fscal            = felec;
 970
 971             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 972
 973              /* Update vectorial force */
 974             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 975             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 976             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 977
 978             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 979             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 980             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 981
 982             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 983             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 984             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 985             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 986
 987             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 988                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 989                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 990
 991             /* Inner loop uses 378 flops */
 992         }
 993
 994         /* End of innermost loop */
 995
 996         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 997                                               f+i_coord_offset,fshift+i_shift_offset);
 998
 999         ggid                        = gid[iidx];
1000         /* Update potential energies */
1001         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1002         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1003
1004         /* Increment number of inner iterations */
1005         inneriter                  += j_index_end - j_index_start;
1006
1007         /* Outer loop uses 26 flops */
1008     }
1009
1010     /* Increment number of outer iterations */
1011     outeriter        += nri;
1012
1013     /* Update outer/inner flops */
1014
1015     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*378);
1016 }
1017 /*
1018  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1019  * Electrostatics interaction: ReactionField
1020  * VdW interaction:            CubicSplineTable
1021  * Geometry:                   Water4-Water4
1022  * Calculate force/pot:        Force
1023  */
1024 void
1025 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1026                     (t_nblist                    * gmx_restrict       nlist,
1027                      rvec                        * gmx_restrict          xx,
1028                      rvec                        * gmx_restrict          ff,
1029                      t_forcerec                  * gmx_restrict          fr,
1030                      t_mdatoms                   * gmx_restrict     mdatoms,
1031                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1032                      t_nrnb                      * gmx_restrict        nrnb)
1033 {
1034     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1035      * just 0 for non-waters.
1036      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1037      * jnr indices corresponding to data put in the four positions in the SIMD register.
1038      */
1039     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1040     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1041     int              jnrA,jnrB,jnrC,jnrD;
1042     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1043     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1044     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1045     real             rcutoff_scalar;
1046     real             *shiftvec,*fshift,*x,*f;
1047     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1048     real             scratch[4*DIM];
1049     __m128           fscal,rcutoff,rcutoff2,jidxall;
1050     int              vdwioffset0;
1051     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1052     int              vdwioffset1;
1053     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1054     int              vdwioffset2;
1055     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1056     int              vdwioffset3;
1057     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1058     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1059     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1060     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1061     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1062     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1063     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1064     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1065     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1066     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1067     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1068     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1069     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1070     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1071     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1072     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1073     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1074     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1075     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1076     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1077     real             *charge;
1078     int              nvdwtype;
1079     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1080     int              *vdwtype;
1081     real             *vdwparam;
1082     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1083     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1084     __m128i          vfitab;
1085     __m128i          ifour       = _mm_set1_epi32(4);
1086     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1087     real             *vftab;
1088     __m128           dummy_mask,cutoff_mask;
1089     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1090     __m128           one     = _mm_set1_ps(1.0);
1091     __m128           two     = _mm_set1_ps(2.0);
1092     x                = xx[0];
1093     f                = ff[0];
1094
1095     nri              = nlist->nri;
1096     iinr             = nlist->iinr;
1097     jindex           = nlist->jindex;
1098     jjnr             = nlist->jjnr;
1099     shiftidx         = nlist->shift;
1100     gid              = nlist->gid;
1101     shiftvec         = fr->shift_vec[0];
1102     fshift           = fr->fshift[0];
1103     facel            = _mm_set1_ps(fr->epsfac);
1104     charge           = mdatoms->chargeA;
1105     krf              = _mm_set1_ps(fr->ic->k_rf);
1106     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1107     crf              = _mm_set1_ps(fr->ic->c_rf);
1108     nvdwtype         = fr->ntype;
1109     vdwparam         = fr->nbfp;
1110     vdwtype          = mdatoms->typeA;
1111
1112     vftab            = kernel_data->table_vdw->data;
1113     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1114
1115     /* Setup water-specific parameters */
1116     inr              = nlist->iinr[0];
1117     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1118     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1119     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1120     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1121
1122     jq1              = _mm_set1_ps(charge[inr+1]);
1123     jq2              = _mm_set1_ps(charge[inr+2]);
1124     jq3              = _mm_set1_ps(charge[inr+3]);
1125     vdwjidx0A        = 2*vdwtype[inr+0];
1126     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1127     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1128     qq11             = _mm_mul_ps(iq1,jq1);
1129     qq12             = _mm_mul_ps(iq1,jq2);
1130     qq13             = _mm_mul_ps(iq1,jq3);
1131     qq21             = _mm_mul_ps(iq2,jq1);
1132     qq22             = _mm_mul_ps(iq2,jq2);
1133     qq23             = _mm_mul_ps(iq2,jq3);
1134     qq31             = _mm_mul_ps(iq3,jq1);
1135     qq32             = _mm_mul_ps(iq3,jq2);
1136     qq33             = _mm_mul_ps(iq3,jq3);
1137
1138     /* Avoid stupid compiler warnings */
1139     jnrA = jnrB = jnrC = jnrD = 0;
1140     j_coord_offsetA = 0;
1141     j_coord_offsetB = 0;
1142     j_coord_offsetC = 0;
1143     j_coord_offsetD = 0;
1144
1145     outeriter        = 0;
1146     inneriter        = 0;
1147
1148     for(iidx=0;iidx<4*DIM;iidx++)
1149     {
1150         scratch[iidx] = 0.0;
1151     }
1152
1153     /* Start outer loop over neighborlists */
1154     for(iidx=0; iidx<nri; iidx++)
1155     {
1156         /* Load shift vector for this list */
1157         i_shift_offset   = DIM*shiftidx[iidx];
1158
1159         /* Load limits for loop over neighbors */
1160         j_index_start    = jindex[iidx];
1161         j_index_end      = jindex[iidx+1];
1162
1163         /* Get outer coordinate index */
1164         inr              = iinr[iidx];
1165         i_coord_offset   = DIM*inr;
1166
1167         /* Load i particle coords and add shift vector */
1168         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1169                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1170
1171         fix0             = _mm_setzero_ps();
1172         fiy0             = _mm_setzero_ps();
1173         fiz0             = _mm_setzero_ps();
1174         fix1             = _mm_setzero_ps();
1175         fiy1             = _mm_setzero_ps();
1176         fiz1             = _mm_setzero_ps();
1177         fix2             = _mm_setzero_ps();
1178         fiy2             = _mm_setzero_ps();
1179         fiz2             = _mm_setzero_ps();
1180         fix3             = _mm_setzero_ps();
1181         fiy3             = _mm_setzero_ps();
1182         fiz3             = _mm_setzero_ps();
1183
1184         /* Start inner kernel loop */
1185         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1186         {
1187
1188             /* Get j neighbor index, and coordinate index */
1189             jnrA             = jjnr[jidx];
1190             jnrB             = jjnr[jidx+1];
1191             jnrC             = jjnr[jidx+2];
1192             jnrD             = jjnr[jidx+3];
1193             j_coord_offsetA  = DIM*jnrA;
1194             j_coord_offsetB  = DIM*jnrB;
1195             j_coord_offsetC  = DIM*jnrC;
1196             j_coord_offsetD  = DIM*jnrD;
1197
1198             /* load j atom coordinates */
1199             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1200                                               x+j_coord_offsetC,x+j_coord_offsetD,
1201                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1202                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1203
1204             /* Calculate displacement vector */
1205             dx00             = _mm_sub_ps(ix0,jx0);
1206             dy00             = _mm_sub_ps(iy0,jy0);
1207             dz00             = _mm_sub_ps(iz0,jz0);
1208             dx11             = _mm_sub_ps(ix1,jx1);
1209             dy11             = _mm_sub_ps(iy1,jy1);
1210             dz11             = _mm_sub_ps(iz1,jz1);
1211             dx12             = _mm_sub_ps(ix1,jx2);
1212             dy12             = _mm_sub_ps(iy1,jy2);
1213             dz12             = _mm_sub_ps(iz1,jz2);
1214             dx13             = _mm_sub_ps(ix1,jx3);
1215             dy13             = _mm_sub_ps(iy1,jy3);
1216             dz13             = _mm_sub_ps(iz1,jz3);
1217             dx21             = _mm_sub_ps(ix2,jx1);
1218             dy21             = _mm_sub_ps(iy2,jy1);
1219             dz21             = _mm_sub_ps(iz2,jz1);
1220             dx22             = _mm_sub_ps(ix2,jx2);
1221             dy22             = _mm_sub_ps(iy2,jy2);
1222             dz22             = _mm_sub_ps(iz2,jz2);
1223             dx23             = _mm_sub_ps(ix2,jx3);
1224             dy23             = _mm_sub_ps(iy2,jy3);
1225             dz23             = _mm_sub_ps(iz2,jz3);
1226             dx31             = _mm_sub_ps(ix3,jx1);
1227             dy31             = _mm_sub_ps(iy3,jy1);
1228             dz31             = _mm_sub_ps(iz3,jz1);
1229             dx32             = _mm_sub_ps(ix3,jx2);
1230             dy32             = _mm_sub_ps(iy3,jy2);
1231             dz32             = _mm_sub_ps(iz3,jz2);
1232             dx33             = _mm_sub_ps(ix3,jx3);
1233             dy33             = _mm_sub_ps(iy3,jy3);
1234             dz33             = _mm_sub_ps(iz3,jz3);
1235
1236             /* Calculate squared distance and things based on it */
1237             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1238             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1239             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1240             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1241             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1242             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1243             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1244             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1245             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1246             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1247
1248             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1249             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1250             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1251             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1252             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1253             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1254             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1255             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1256             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1257             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1258
1259             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1260             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1261             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1262             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1263             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1264             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1265             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1266             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1267             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1268
1269             fjx0             = _mm_setzero_ps();
1270             fjy0             = _mm_setzero_ps();
1271             fjz0             = _mm_setzero_ps();
1272             fjx1             = _mm_setzero_ps();
1273             fjy1             = _mm_setzero_ps();
1274             fjz1             = _mm_setzero_ps();
1275             fjx2             = _mm_setzero_ps();
1276             fjy2             = _mm_setzero_ps();
1277             fjz2             = _mm_setzero_ps();
1278             fjx3             = _mm_setzero_ps();
1279             fjy3             = _mm_setzero_ps();
1280             fjz3             = _mm_setzero_ps();
1281
1282             /**************************
1283              * CALCULATE INTERACTIONS *
1284              **************************/
1285
1286             r00              = _mm_mul_ps(rsq00,rinv00);
1287
1288             /* Calculate table index by multiplying r with table scale and truncate to integer */
1289             rt               = _mm_mul_ps(r00,vftabscale);
1290             vfitab           = _mm_cvttps_epi32(rt);
1291 #ifdef __XOP__
1292             vfeps            = _mm_frcz_ps(rt);
1293 #else
1294             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1295 #endif
1296             twovfeps         = _mm_add_ps(vfeps,vfeps);
1297             vfitab           = _mm_slli_epi32(vfitab,3);
1298
1299             /* CUBIC SPLINE TABLE DISPERSION */
1300             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1301             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1302             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1303             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1304             _MM_TRANSPOSE4_PS(Y,F,G,H);
1305             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1306             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1307             fvdw6            = _mm_mul_ps(c6_00,FF);
1308
1309             /* CUBIC SPLINE TABLE REPULSION */
1310             vfitab           = _mm_add_epi32(vfitab,ifour);
1311             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1312             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1313             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1314             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1315             _MM_TRANSPOSE4_PS(Y,F,G,H);
1316             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1317             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1318             fvdw12           = _mm_mul_ps(c12_00,FF);
1319             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1320
1321             fscal            = fvdw;
1322
1323              /* Update vectorial force */
1324             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1325             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1326             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1327
1328             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1329             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1330             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1331
1332             /**************************
1333              * CALCULATE INTERACTIONS *
1334              **************************/
1335
1336             /* REACTION-FIELD ELECTROSTATICS */
1337             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1338
1339             fscal            = felec;
1340
1341              /* Update vectorial force */
1342             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1343             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1344             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1345
1346             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1347             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1348             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1349
1350             /**************************
1351              * CALCULATE INTERACTIONS *
1352              **************************/
1353
1354             /* REACTION-FIELD ELECTROSTATICS */
1355             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1356
1357             fscal            = felec;
1358
1359              /* Update vectorial force */
1360             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1361             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1362             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1363
1364             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1365             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1366             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1367
1368             /**************************
1369              * CALCULATE INTERACTIONS *
1370              **************************/
1371
1372             /* REACTION-FIELD ELECTROSTATICS */
1373             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1374
1375             fscal            = felec;
1376
1377              /* Update vectorial force */
1378             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1379             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1380             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1381
1382             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1383             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1384             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             /* REACTION-FIELD ELECTROSTATICS */
1391             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1392
1393             fscal            = felec;
1394
1395              /* Update vectorial force */
1396             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1397             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1398             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1399
1400             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1401             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1402             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1403
1404             /**************************
1405              * CALCULATE INTERACTIONS *
1406              **************************/
1407
1408             /* REACTION-FIELD ELECTROSTATICS */
1409             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1410
1411             fscal            = felec;
1412
1413              /* Update vectorial force */
1414             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1415             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1416             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1417
1418             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1419             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1420             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1421
1422             /**************************
1423              * CALCULATE INTERACTIONS *
1424              **************************/
1425
1426             /* REACTION-FIELD ELECTROSTATICS */
1427             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1428
1429             fscal            = felec;
1430
1431              /* Update vectorial force */
1432             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1433             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1434             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1435
1436             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1437             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1438             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1439
1440             /**************************
1441              * CALCULATE INTERACTIONS *
1442              **************************/
1443
1444             /* REACTION-FIELD ELECTROSTATICS */
1445             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1446
1447             fscal            = felec;
1448
1449              /* Update vectorial force */
1450             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1451             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1452             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1453
1454             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1455             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1456             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1457
1458             /**************************
1459              * CALCULATE INTERACTIONS *
1460              **************************/
1461
1462             /* REACTION-FIELD ELECTROSTATICS */
1463             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1464
1465             fscal            = felec;
1466
1467              /* Update vectorial force */
1468             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1469             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1470             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1471
1472             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1473             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1474             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1475
1476             /**************************
1477              * CALCULATE INTERACTIONS *
1478              **************************/
1479
1480             /* REACTION-FIELD ELECTROSTATICS */
1481             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1482
1483             fscal            = felec;
1484
1485              /* Update vectorial force */
1486             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1487             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1488             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1489
1490             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1491             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1492             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1493
1494             fjptrA             = f+j_coord_offsetA;
1495             fjptrB             = f+j_coord_offsetB;
1496             fjptrC             = f+j_coord_offsetC;
1497             fjptrD             = f+j_coord_offsetD;
1498
1499             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1500                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1501                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1502
1503             /* Inner loop uses 324 flops */
1504         }
1505
1506         if(jidx<j_index_end)
1507         {
1508
1509             /* Get j neighbor index, and coordinate index */
1510             jnrlistA         = jjnr[jidx];
1511             jnrlistB         = jjnr[jidx+1];
1512             jnrlistC         = jjnr[jidx+2];
1513             jnrlistD         = jjnr[jidx+3];
1514             /* Sign of each element will be negative for non-real atoms.
1515              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1516              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1517              */
1518             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1519             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1520             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1521             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1522             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1523             j_coord_offsetA  = DIM*jnrA;
1524             j_coord_offsetB  = DIM*jnrB;
1525             j_coord_offsetC  = DIM*jnrC;
1526             j_coord_offsetD  = DIM*jnrD;
1527
1528             /* load j atom coordinates */
1529             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1530                                               x+j_coord_offsetC,x+j_coord_offsetD,
1531                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1532                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1533
1534             /* Calculate displacement vector */
1535             dx00             = _mm_sub_ps(ix0,jx0);
1536             dy00             = _mm_sub_ps(iy0,jy0);
1537             dz00             = _mm_sub_ps(iz0,jz0);
1538             dx11             = _mm_sub_ps(ix1,jx1);
1539             dy11             = _mm_sub_ps(iy1,jy1);
1540             dz11             = _mm_sub_ps(iz1,jz1);
1541             dx12             = _mm_sub_ps(ix1,jx2);
1542             dy12             = _mm_sub_ps(iy1,jy2);
1543             dz12             = _mm_sub_ps(iz1,jz2);
1544             dx13             = _mm_sub_ps(ix1,jx3);
1545             dy13             = _mm_sub_ps(iy1,jy3);
1546             dz13             = _mm_sub_ps(iz1,jz3);
1547             dx21             = _mm_sub_ps(ix2,jx1);
1548             dy21             = _mm_sub_ps(iy2,jy1);
1549             dz21             = _mm_sub_ps(iz2,jz1);
1550             dx22             = _mm_sub_ps(ix2,jx2);
1551             dy22             = _mm_sub_ps(iy2,jy2);
1552             dz22             = _mm_sub_ps(iz2,jz2);
1553             dx23             = _mm_sub_ps(ix2,jx3);
1554             dy23             = _mm_sub_ps(iy2,jy3);
1555             dz23             = _mm_sub_ps(iz2,jz3);
1556             dx31             = _mm_sub_ps(ix3,jx1);
1557             dy31             = _mm_sub_ps(iy3,jy1);
1558             dz31             = _mm_sub_ps(iz3,jz1);
1559             dx32             = _mm_sub_ps(ix3,jx2);
1560             dy32             = _mm_sub_ps(iy3,jy2);
1561             dz32             = _mm_sub_ps(iz3,jz2);
1562             dx33             = _mm_sub_ps(ix3,jx3);
1563             dy33             = _mm_sub_ps(iy3,jy3);
1564             dz33             = _mm_sub_ps(iz3,jz3);
1565
1566             /* Calculate squared distance and things based on it */
1567             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1568             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1569             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1570             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1571             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1572             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1573             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1574             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1575             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1576             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1577
1578             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1579             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1580             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1581             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1582             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1583             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1584             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1585             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1586             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1587             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1588
1589             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1590             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1591             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1592             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1593             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1594             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1595             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1596             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1597             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1598
1599             fjx0             = _mm_setzero_ps();
1600             fjy0             = _mm_setzero_ps();
1601             fjz0             = _mm_setzero_ps();
1602             fjx1             = _mm_setzero_ps();
1603             fjy1             = _mm_setzero_ps();
1604             fjz1             = _mm_setzero_ps();
1605             fjx2             = _mm_setzero_ps();
1606             fjy2             = _mm_setzero_ps();
1607             fjz2             = _mm_setzero_ps();
1608             fjx3             = _mm_setzero_ps();
1609             fjy3             = _mm_setzero_ps();
1610             fjz3             = _mm_setzero_ps();
1611
1612             /**************************
1613              * CALCULATE INTERACTIONS *
1614              **************************/
1615
1616             r00              = _mm_mul_ps(rsq00,rinv00);
1617             r00              = _mm_andnot_ps(dummy_mask,r00);
1618
1619             /* Calculate table index by multiplying r with table scale and truncate to integer */
1620             rt               = _mm_mul_ps(r00,vftabscale);
1621             vfitab           = _mm_cvttps_epi32(rt);
1622 #ifdef __XOP__
1623             vfeps            = _mm_frcz_ps(rt);
1624 #else
1625             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1626 #endif
1627             twovfeps         = _mm_add_ps(vfeps,vfeps);
1628             vfitab           = _mm_slli_epi32(vfitab,3);
1629
1630             /* CUBIC SPLINE TABLE DISPERSION */
1631             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1632             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1633             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1634             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1635             _MM_TRANSPOSE4_PS(Y,F,G,H);
1636             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1637             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1638             fvdw6            = _mm_mul_ps(c6_00,FF);
1639
1640             /* CUBIC SPLINE TABLE REPULSION */
1641             vfitab           = _mm_add_epi32(vfitab,ifour);
1642             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1643             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1644             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1645             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1646             _MM_TRANSPOSE4_PS(Y,F,G,H);
1647             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1648             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1649             fvdw12           = _mm_mul_ps(c12_00,FF);
1650             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1651
1652             fscal            = fvdw;
1653
1654             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1655
1656              /* Update vectorial force */
1657             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1658             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1659             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1660
1661             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1662             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1663             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1664
1665             /**************************
1666              * CALCULATE INTERACTIONS *
1667              **************************/
1668
1669             /* REACTION-FIELD ELECTROSTATICS */
1670             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1671
1672             fscal            = felec;
1673
1674             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1675
1676              /* Update vectorial force */
1677             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1678             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1679             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1680
1681             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1682             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1683             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1684
1685             /**************************
1686              * CALCULATE INTERACTIONS *
1687              **************************/
1688
1689             /* REACTION-FIELD ELECTROSTATICS */
1690             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1691
1692             fscal            = felec;
1693
1694             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1695
1696              /* Update vectorial force */
1697             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1698             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1699             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1700
1701             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1702             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1703             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1704
1705             /**************************
1706              * CALCULATE INTERACTIONS *
1707              **************************/
1708
1709             /* REACTION-FIELD ELECTROSTATICS */
1710             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1711
1712             fscal            = felec;
1713
1714             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1715
1716              /* Update vectorial force */
1717             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1718             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1719             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1720
1721             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1722             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1723             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1724
1725             /**************************
1726              * CALCULATE INTERACTIONS *
1727              **************************/
1728
1729             /* REACTION-FIELD ELECTROSTATICS */
1730             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1731
1732             fscal            = felec;
1733
1734             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1735
1736              /* Update vectorial force */
1737             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1738             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1739             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1740
1741             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1742             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1743             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1744
1745             /**************************
1746              * CALCULATE INTERACTIONS *
1747              **************************/
1748
1749             /* REACTION-FIELD ELECTROSTATICS */
1750             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1751
1752             fscal            = felec;
1753
1754             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1755
1756              /* Update vectorial force */
1757             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1758             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1759             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1760
1761             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1762             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1763             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1764
1765             /**************************
1766              * CALCULATE INTERACTIONS *
1767              **************************/
1768
1769             /* REACTION-FIELD ELECTROSTATICS */
1770             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1771
1772             fscal            = felec;
1773
1774             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1775
1776              /* Update vectorial force */
1777             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1778             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1779             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1780
1781             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1782             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1783             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1784
1785             /**************************
1786              * CALCULATE INTERACTIONS *
1787              **************************/
1788
1789             /* REACTION-FIELD ELECTROSTATICS */
1790             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1791
1792             fscal            = felec;
1793
1794             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1795
1796              /* Update vectorial force */
1797             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1798             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1799             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1800
1801             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1802             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1803             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1804
1805             /**************************
1806              * CALCULATE INTERACTIONS *
1807              **************************/
1808
1809             /* REACTION-FIELD ELECTROSTATICS */
1810             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1811
1812             fscal            = felec;
1813
1814             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1815
1816              /* Update vectorial force */
1817             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1818             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1819             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1820
1821             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1822             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1823             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1824
1825             /**************************
1826              * CALCULATE INTERACTIONS *
1827              **************************/
1828
1829             /* REACTION-FIELD ELECTROSTATICS */
1830             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1831
1832             fscal            = felec;
1833
1834             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1835
1836              /* Update vectorial force */
1837             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1838             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1839             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1840
1841             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1842             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1843             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1844
1845             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1846             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1847             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1848             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1849
1850             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1851                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1852                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1853
1854             /* Inner loop uses 325 flops */
1855         }
1856
1857         /* End of innermost loop */
1858
1859         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1860                                               f+i_coord_offset,fshift+i_shift_offset);
1861
1862         /* Increment number of inner iterations */
1863         inneriter                  += j_index_end - j_index_start;
1864
1865         /* Outer loop uses 24 flops */
1866     }
1867
1868     /* Increment number of outer iterations */
1869     outeriter        += nri;
1870
1871     /* Update outer/inner flops */
1872
1873     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*325);
1874 }