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) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_single
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  98     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  99     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 100     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 101     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 102     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 103     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 104     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 105     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 106     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 107     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 108     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 109     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 110     real             *charge;
 111     int              nvdwtype;
 112     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 113     int              *vdwtype;
 114     real             *vdwparam;
 115     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 116     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 117     __m128i          vfitab;
 118     __m128i          ifour       = _mm_set1_epi32(4);
 119     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 120     real             *vftab;
 121     __m128           dummy_mask,cutoff_mask;
 122     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 123     __m128           one     = _mm_set1_ps(1.0);
 124     __m128           two     = _mm_set1_ps(2.0);
 125     x                = xx[0];
 126     f                = ff[0];
 127
 128     nri              = nlist->nri;
 129     iinr             = nlist->iinr;
 130     jindex           = nlist->jindex;
 131     jjnr             = nlist->jjnr;
 132     shiftidx         = nlist->shift;
 133     gid              = nlist->gid;
 134     shiftvec         = fr->shift_vec[0];
 135     fshift           = fr->fshift[0];
 136     facel            = _mm_set1_ps(fr->epsfac);
 137     charge           = mdatoms->chargeA;
 138     krf              = _mm_set1_ps(fr->ic->k_rf);
 139     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 140     crf              = _mm_set1_ps(fr->ic->c_rf);
 141     nvdwtype         = fr->ntype;
 142     vdwparam         = fr->nbfp;
 143     vdwtype          = mdatoms->typeA;
 144
 145     vftab            = kernel_data->table_vdw->data;
 146     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 147
 148     /* Setup water-specific parameters */
 149     inr              = nlist->iinr[0];
 150     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 151     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 152     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 153     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 154
 155     jq1              = _mm_set1_ps(charge[inr+1]);
 156     jq2              = _mm_set1_ps(charge[inr+2]);
 157     jq3              = _mm_set1_ps(charge[inr+3]);
 158     vdwjidx0A        = 2*vdwtype[inr+0];
 159     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 160     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 161     qq11             = _mm_mul_ps(iq1,jq1);
 162     qq12             = _mm_mul_ps(iq1,jq2);
 163     qq13             = _mm_mul_ps(iq1,jq3);
 164     qq21             = _mm_mul_ps(iq2,jq1);
 165     qq22             = _mm_mul_ps(iq2,jq2);
 166     qq23             = _mm_mul_ps(iq2,jq3);
 167     qq31             = _mm_mul_ps(iq3,jq1);
 168     qq32             = _mm_mul_ps(iq3,jq2);
 169     qq33             = _mm_mul_ps(iq3,jq3);
 170
 171     /* Avoid stupid compiler warnings */
 172     jnrA = jnrB = jnrC = jnrD = 0;
 173     j_coord_offsetA = 0;
 174     j_coord_offsetB = 0;
 175     j_coord_offsetC = 0;
 176     j_coord_offsetD = 0;
 177
 178     outeriter        = 0;
 179     inneriter        = 0;
 180
 181     for(iidx=0;iidx<4*DIM;iidx++)
 182     {
 183         scratch[iidx] = 0.0;
 184     }
 185
 186     /* Start outer loop over neighborlists */
 187     for(iidx=0; iidx<nri; iidx++)
 188     {
 189         /* Load shift vector for this list */
 190         i_shift_offset   = DIM*shiftidx[iidx];
 191
 192         /* Load limits for loop over neighbors */
 193         j_index_start    = jindex[iidx];
 194         j_index_end      = jindex[iidx+1];
 195
 196         /* Get outer coordinate index */
 197         inr              = iinr[iidx];
 198         i_coord_offset   = DIM*inr;
 199
 200         /* Load i particle coords and add shift vector */
 201         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 202                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 203
 204         fix0             = _mm_setzero_ps();
 205         fiy0             = _mm_setzero_ps();
 206         fiz0             = _mm_setzero_ps();
 207         fix1             = _mm_setzero_ps();
 208         fiy1             = _mm_setzero_ps();
 209         fiz1             = _mm_setzero_ps();
 210         fix2             = _mm_setzero_ps();
 211         fiy2             = _mm_setzero_ps();
 212         fiz2             = _mm_setzero_ps();
 213         fix3             = _mm_setzero_ps();
 214         fiy3             = _mm_setzero_ps();
 215         fiz3             = _mm_setzero_ps();
 216
 217         /* Reset potential sums */
 218         velecsum         = _mm_setzero_ps();
 219         vvdwsum          = _mm_setzero_ps();
 220
 221         /* Start inner kernel loop */
 222         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 223         {
 224
 225             /* Get j neighbor index, and coordinate index */
 226             jnrA             = jjnr[jidx];
 227             jnrB             = jjnr[jidx+1];
 228             jnrC             = jjnr[jidx+2];
 229             jnrD             = jjnr[jidx+3];
 230             j_coord_offsetA  = DIM*jnrA;
 231             j_coord_offsetB  = DIM*jnrB;
 232             j_coord_offsetC  = DIM*jnrC;
 233             j_coord_offsetD  = DIM*jnrD;
 234
 235             /* load j atom coordinates */
 236             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 237                                               x+j_coord_offsetC,x+j_coord_offsetD,
 238                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 239                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 240
 241             /* Calculate displacement vector */
 242             dx00             = _mm_sub_ps(ix0,jx0);
 243             dy00             = _mm_sub_ps(iy0,jy0);
 244             dz00             = _mm_sub_ps(iz0,jz0);
 245             dx11             = _mm_sub_ps(ix1,jx1);
 246             dy11             = _mm_sub_ps(iy1,jy1);
 247             dz11             = _mm_sub_ps(iz1,jz1);
 248             dx12             = _mm_sub_ps(ix1,jx2);
 249             dy12             = _mm_sub_ps(iy1,jy2);
 250             dz12             = _mm_sub_ps(iz1,jz2);
 251             dx13             = _mm_sub_ps(ix1,jx3);
 252             dy13             = _mm_sub_ps(iy1,jy3);
 253             dz13             = _mm_sub_ps(iz1,jz3);
 254             dx21             = _mm_sub_ps(ix2,jx1);
 255             dy21             = _mm_sub_ps(iy2,jy1);
 256             dz21             = _mm_sub_ps(iz2,jz1);
 257             dx22             = _mm_sub_ps(ix2,jx2);
 258             dy22             = _mm_sub_ps(iy2,jy2);
 259             dz22             = _mm_sub_ps(iz2,jz2);
 260             dx23             = _mm_sub_ps(ix2,jx3);
 261             dy23             = _mm_sub_ps(iy2,jy3);
 262             dz23             = _mm_sub_ps(iz2,jz3);
 263             dx31             = _mm_sub_ps(ix3,jx1);
 264             dy31             = _mm_sub_ps(iy3,jy1);
 265             dz31             = _mm_sub_ps(iz3,jz1);
 266             dx32             = _mm_sub_ps(ix3,jx2);
 267             dy32             = _mm_sub_ps(iy3,jy2);
 268             dz32             = _mm_sub_ps(iz3,jz2);
 269             dx33             = _mm_sub_ps(ix3,jx3);
 270             dy33             = _mm_sub_ps(iy3,jy3);
 271             dz33             = _mm_sub_ps(iz3,jz3);
 272
 273             /* Calculate squared distance and things based on it */
 274             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 275             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 276             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 277             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 278             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 279             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 280             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 281             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 282             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 283             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 284
 285             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 286             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 287             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 288             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 289             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 290             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 291             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 292             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 293             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 294             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 295
 296             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 297             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 298             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 299             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 300             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 301             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 302             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 303             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 304             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 305
 306             fjx0             = _mm_setzero_ps();
 307             fjy0             = _mm_setzero_ps();
 308             fjz0             = _mm_setzero_ps();
 309             fjx1             = _mm_setzero_ps();
 310             fjy1             = _mm_setzero_ps();
 311             fjz1             = _mm_setzero_ps();
 312             fjx2             = _mm_setzero_ps();
 313             fjy2             = _mm_setzero_ps();
 314             fjz2             = _mm_setzero_ps();
 315             fjx3             = _mm_setzero_ps();
 316             fjy3             = _mm_setzero_ps();
 317             fjz3             = _mm_setzero_ps();
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             r00              = _mm_mul_ps(rsq00,rinv00);
 324
 325             /* Calculate table index by multiplying r with table scale and truncate to integer */
 326             rt               = _mm_mul_ps(r00,vftabscale);
 327             vfitab           = _mm_cvttps_epi32(rt);
 328 #ifdef __XOP__
 329             vfeps            = _mm_frcz_ps(rt);
 330 #else
 331             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 332 #endif
 333             twovfeps         = _mm_add_ps(vfeps,vfeps);
 334             vfitab           = _mm_slli_epi32(vfitab,3);
 335
 336             /* CUBIC SPLINE TABLE DISPERSION */
 337             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 338             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 339             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 340             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 341             _MM_TRANSPOSE4_PS(Y,F,G,H);
 342             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 343             VV               = _mm_macc_ps(vfeps,Fp,Y);
 344             vvdw6            = _mm_mul_ps(c6_00,VV);
 345             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 346             fvdw6            = _mm_mul_ps(c6_00,FF);
 347
 348             /* CUBIC SPLINE TABLE REPULSION */
 349             vfitab           = _mm_add_epi32(vfitab,ifour);
 350             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 351             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 352             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 353             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 354             _MM_TRANSPOSE4_PS(Y,F,G,H);
 355             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 356             VV               = _mm_macc_ps(vfeps,Fp,Y);
 357             vvdw12           = _mm_mul_ps(c12_00,VV);
 358             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 359             fvdw12           = _mm_mul_ps(c12_00,FF);
 360             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 361             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 365
 366             fscal            = fvdw;
 367
 368              /* Update vectorial force */
 369             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 370             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 371             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 372
 373             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 374             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 375             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 376
 377             /**************************
 378              * CALCULATE INTERACTIONS *
 379              **************************/
 380
 381             /* REACTION-FIELD ELECTROSTATICS */
 382             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 383             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velecsum         = _mm_add_ps(velecsum,velec);
 387
 388             fscal            = felec;
 389
 390              /* Update vectorial force */
 391             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 392             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 393             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 394
 395             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 396             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 397             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             /* REACTION-FIELD ELECTROSTATICS */
 404             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 405             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velecsum         = _mm_add_ps(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412              /* Update vectorial force */
 413             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 414             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 415             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 416
 417             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 418             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 419             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             /* REACTION-FIELD ELECTROSTATICS */
 426             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 427             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 428
 429             /* Update potential sum for this i atom from the interaction with this j atom. */
 430             velecsum         = _mm_add_ps(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434              /* Update vectorial force */
 435             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 436             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 437             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 438
 439             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 440             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 441             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             /* REACTION-FIELD ELECTROSTATICS */
 448             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 449             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 450
 451             /* Update potential sum for this i atom from the interaction with this j atom. */
 452             velecsum         = _mm_add_ps(velecsum,velec);
 453
 454             fscal            = felec;
 455
 456              /* Update vectorial force */
 457             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 458             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 459             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 460
 461             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 462             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 463             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             /* REACTION-FIELD ELECTROSTATICS */
 470             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 471             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 472
 473             /* Update potential sum for this i atom from the interaction with this j atom. */
 474             velecsum         = _mm_add_ps(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478              /* Update vectorial force */
 479             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 480             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 481             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 482
 483             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 484             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 485             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 486
 487             /**************************
 488              * CALCULATE INTERACTIONS *
 489              **************************/
 490
 491             /* REACTION-FIELD ELECTROSTATICS */
 492             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 493             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 494
 495             /* Update potential sum for this i atom from the interaction with this j atom. */
 496             velecsum         = _mm_add_ps(velecsum,velec);
 497
 498             fscal            = felec;
 499
 500              /* Update vectorial force */
 501             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 502             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 503             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 504
 505             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 506             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 507             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 508
 509             /**************************
 510              * CALCULATE INTERACTIONS *
 511              **************************/
 512
 513             /* REACTION-FIELD ELECTROSTATICS */
 514             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 515             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 516
 517             /* Update potential sum for this i atom from the interaction with this j atom. */
 518             velecsum         = _mm_add_ps(velecsum,velec);
 519
 520             fscal            = felec;
 521
 522              /* Update vectorial force */
 523             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 524             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 525             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 526
 527             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 528             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 529             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             /* REACTION-FIELD ELECTROSTATICS */
 536             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 537             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 538
 539             /* Update potential sum for this i atom from the interaction with this j atom. */
 540             velecsum         = _mm_add_ps(velecsum,velec);
 541
 542             fscal            = felec;
 543
 544              /* Update vectorial force */
 545             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 546             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 547             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 548
 549             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 550             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 551             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 552
 553             /**************************
 554              * CALCULATE INTERACTIONS *
 555              **************************/
 556
 557             /* REACTION-FIELD ELECTROSTATICS */
 558             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 559             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 560
 561             /* Update potential sum for this i atom from the interaction with this j atom. */
 562             velecsum         = _mm_add_ps(velecsum,velec);
 563
 564             fscal            = felec;
 565
 566              /* Update vectorial force */
 567             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 568             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 569             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 570
 571             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 572             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 573             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 574
 575             fjptrA             = f+j_coord_offsetA;
 576             fjptrB             = f+j_coord_offsetB;
 577             fjptrC             = f+j_coord_offsetC;
 578             fjptrD             = f+j_coord_offsetD;
 579
 580             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 581                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 582                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 583
 584             /* Inner loop uses 377 flops */
 585         }
 586
 587         if(jidx<j_index_end)
 588         {
 589
 590             /* Get j neighbor index, and coordinate index */
 591             jnrlistA         = jjnr[jidx];
 592             jnrlistB         = jjnr[jidx+1];
 593             jnrlistC         = jjnr[jidx+2];
 594             jnrlistD         = jjnr[jidx+3];
 595             /* Sign of each element will be negative for non-real atoms.
 596              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 597              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 598              */
 599             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 600             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 601             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 602             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 603             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 604             j_coord_offsetA  = DIM*jnrA;
 605             j_coord_offsetB  = DIM*jnrB;
 606             j_coord_offsetC  = DIM*jnrC;
 607             j_coord_offsetD  = DIM*jnrD;
 608
 609             /* load j atom coordinates */
 610             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 611                                               x+j_coord_offsetC,x+j_coord_offsetD,
 612                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 613                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 614
 615             /* Calculate displacement vector */
 616             dx00             = _mm_sub_ps(ix0,jx0);
 617             dy00             = _mm_sub_ps(iy0,jy0);
 618             dz00             = _mm_sub_ps(iz0,jz0);
 619             dx11             = _mm_sub_ps(ix1,jx1);
 620             dy11             = _mm_sub_ps(iy1,jy1);
 621             dz11             = _mm_sub_ps(iz1,jz1);
 622             dx12             = _mm_sub_ps(ix1,jx2);
 623             dy12             = _mm_sub_ps(iy1,jy2);
 624             dz12             = _mm_sub_ps(iz1,jz2);
 625             dx13             = _mm_sub_ps(ix1,jx3);
 626             dy13             = _mm_sub_ps(iy1,jy3);
 627             dz13             = _mm_sub_ps(iz1,jz3);
 628             dx21             = _mm_sub_ps(ix2,jx1);
 629             dy21             = _mm_sub_ps(iy2,jy1);
 630             dz21             = _mm_sub_ps(iz2,jz1);
 631             dx22             = _mm_sub_ps(ix2,jx2);
 632             dy22             = _mm_sub_ps(iy2,jy2);
 633             dz22             = _mm_sub_ps(iz2,jz2);
 634             dx23             = _mm_sub_ps(ix2,jx3);
 635             dy23             = _mm_sub_ps(iy2,jy3);
 636             dz23             = _mm_sub_ps(iz2,jz3);
 637             dx31             = _mm_sub_ps(ix3,jx1);
 638             dy31             = _mm_sub_ps(iy3,jy1);
 639             dz31             = _mm_sub_ps(iz3,jz1);
 640             dx32             = _mm_sub_ps(ix3,jx2);
 641             dy32             = _mm_sub_ps(iy3,jy2);
 642             dz32             = _mm_sub_ps(iz3,jz2);
 643             dx33             = _mm_sub_ps(ix3,jx3);
 644             dy33             = _mm_sub_ps(iy3,jy3);
 645             dz33             = _mm_sub_ps(iz3,jz3);
 646
 647             /* Calculate squared distance and things based on it */
 648             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 649             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 650             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 651             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 652             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 653             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 654             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 655             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 656             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 657             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 658
 659             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 660             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 661             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 662             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 663             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 664             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 665             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 666             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 667             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 668             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 669
 670             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 671             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 672             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 673             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 674             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 675             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 676             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 677             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 678             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 679
 680             fjx0             = _mm_setzero_ps();
 681             fjy0             = _mm_setzero_ps();
 682             fjz0             = _mm_setzero_ps();
 683             fjx1             = _mm_setzero_ps();
 684             fjy1             = _mm_setzero_ps();
 685             fjz1             = _mm_setzero_ps();
 686             fjx2             = _mm_setzero_ps();
 687             fjy2             = _mm_setzero_ps();
 688             fjz2             = _mm_setzero_ps();
 689             fjx3             = _mm_setzero_ps();
 690             fjy3             = _mm_setzero_ps();
 691             fjz3             = _mm_setzero_ps();
 692
 693             /**************************
 694              * CALCULATE INTERACTIONS *
 695              **************************/
 696
 697             r00              = _mm_mul_ps(rsq00,rinv00);
 698             r00              = _mm_andnot_ps(dummy_mask,r00);
 699
 700             /* Calculate table index by multiplying r with table scale and truncate to integer */
 701             rt               = _mm_mul_ps(r00,vftabscale);
 702             vfitab           = _mm_cvttps_epi32(rt);
 703 #ifdef __XOP__
 704             vfeps            = _mm_frcz_ps(rt);
 705 #else
 706             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 707 #endif
 708             twovfeps         = _mm_add_ps(vfeps,vfeps);
 709             vfitab           = _mm_slli_epi32(vfitab,3);
 710
 711             /* CUBIC SPLINE TABLE DISPERSION */
 712             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 713             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 714             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 715             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 716             _MM_TRANSPOSE4_PS(Y,F,G,H);
 717             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 718             VV               = _mm_macc_ps(vfeps,Fp,Y);
 719             vvdw6            = _mm_mul_ps(c6_00,VV);
 720             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 721             fvdw6            = _mm_mul_ps(c6_00,FF);
 722
 723             /* CUBIC SPLINE TABLE REPULSION */
 724             vfitab           = _mm_add_epi32(vfitab,ifour);
 725             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 726             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 727             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 728             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 729             _MM_TRANSPOSE4_PS(Y,F,G,H);
 730             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 731             VV               = _mm_macc_ps(vfeps,Fp,Y);
 732             vvdw12           = _mm_mul_ps(c12_00,VV);
 733             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 734             fvdw12           = _mm_mul_ps(c12_00,FF);
 735             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 736             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 737
 738             /* Update potential sum for this i atom from the interaction with this j atom. */
 739             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 740             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 741
 742             fscal            = fvdw;
 743
 744             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 745
 746              /* Update vectorial force */
 747             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 748             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 749             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 750
 751             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 752             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 753             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             /* REACTION-FIELD ELECTROSTATICS */
 760             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 761             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _mm_andnot_ps(dummy_mask,velec);
 765             velecsum         = _mm_add_ps(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 770
 771              /* Update vectorial force */
 772             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 773             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 774             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 775
 776             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 777             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 778             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 779
 780             /**************************
 781              * CALCULATE INTERACTIONS *
 782              **************************/
 783
 784             /* REACTION-FIELD ELECTROSTATICS */
 785             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 786             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 787
 788             /* Update potential sum for this i atom from the interaction with this j atom. */
 789             velec            = _mm_andnot_ps(dummy_mask,velec);
 790             velecsum         = _mm_add_ps(velecsum,velec);
 791
 792             fscal            = felec;
 793
 794             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 795
 796              /* Update vectorial force */
 797             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 798             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 799             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 800
 801             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 802             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 803             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 804
 805             /**************************
 806              * CALCULATE INTERACTIONS *
 807              **************************/
 808
 809             /* REACTION-FIELD ELECTROSTATICS */
 810             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 811             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _mm_andnot_ps(dummy_mask,velec);
 815             velecsum         = _mm_add_ps(velecsum,velec);
 816
 817             fscal            = felec;
 818
 819             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 820
 821              /* Update vectorial force */
 822             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 823             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 824             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 825
 826             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 827             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 828             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             /* REACTION-FIELD ELECTROSTATICS */
 835             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 836             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 837
 838             /* Update potential sum for this i atom from the interaction with this j atom. */
 839             velec            = _mm_andnot_ps(dummy_mask,velec);
 840             velecsum         = _mm_add_ps(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 845
 846              /* Update vectorial force */
 847             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 848             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 849             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 850
 851             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 852             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 853             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 854
 855             /**************************
 856              * CALCULATE INTERACTIONS *
 857              **************************/
 858
 859             /* REACTION-FIELD ELECTROSTATICS */
 860             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 861             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 862
 863             /* Update potential sum for this i atom from the interaction with this j atom. */
 864             velec            = _mm_andnot_ps(dummy_mask,velec);
 865             velecsum         = _mm_add_ps(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 870
 871              /* Update vectorial force */
 872             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 873             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 874             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 875
 876             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 877             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 878             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 879
 880             /**************************
 881              * CALCULATE INTERACTIONS *
 882              **************************/
 883
 884             /* REACTION-FIELD ELECTROSTATICS */
 885             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 886             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 887
 888             /* Update potential sum for this i atom from the interaction with this j atom. */
 889             velec            = _mm_andnot_ps(dummy_mask,velec);
 890             velecsum         = _mm_add_ps(velecsum,velec);
 891
 892             fscal            = felec;
 893
 894             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 895
 896              /* Update vectorial force */
 897             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 898             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 899             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 900
 901             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 902             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 903             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             /* REACTION-FIELD ELECTROSTATICS */
 910             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 911             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 912
 913             /* Update potential sum for this i atom from the interaction with this j atom. */
 914             velec            = _mm_andnot_ps(dummy_mask,velec);
 915             velecsum         = _mm_add_ps(velecsum,velec);
 916
 917             fscal            = felec;
 918
 919             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 920
 921              /* Update vectorial force */
 922             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 923             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 924             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 925
 926             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 927             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 928             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 929
 930             /**************************
 931              * CALCULATE INTERACTIONS *
 932              **************************/
 933
 934             /* REACTION-FIELD ELECTROSTATICS */
 935             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 936             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 937
 938             /* Update potential sum for this i atom from the interaction with this j atom. */
 939             velec            = _mm_andnot_ps(dummy_mask,velec);
 940             velecsum         = _mm_add_ps(velecsum,velec);
 941
 942             fscal            = felec;
 943
 944             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 945
 946              /* Update vectorial force */
 947             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 948             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 949             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 950
 951             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 952             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 953             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 954
 955             /**************************
 956              * CALCULATE INTERACTIONS *
 957              **************************/
 958
 959             /* REACTION-FIELD ELECTROSTATICS */
 960             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 961             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 962
 963             /* Update potential sum for this i atom from the interaction with this j atom. */
 964             velec            = _mm_andnot_ps(dummy_mask,velec);
 965             velecsum         = _mm_add_ps(velecsum,velec);
 966
 967             fscal            = felec;
 968
 969             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 970
 971              /* Update vectorial force */
 972             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 973             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 974             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 975
 976             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 977             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 978             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 979
 980             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 981             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 982             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 983             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 984
 985             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 986                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 987                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 988
 989             /* Inner loop uses 378 flops */
 990         }
 991
 992         /* End of innermost loop */
 993
 994         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 995                                               f+i_coord_offset,fshift+i_shift_offset);
 996
 997         ggid                        = gid[iidx];
 998         /* Update potential energies */
 999         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1000         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1001
1002         /* Increment number of inner iterations */
1003         inneriter                  += j_index_end - j_index_start;
1004
1005         /* Outer loop uses 26 flops */
1006     }
1007
1008     /* Increment number of outer iterations */
1009     outeriter        += nri;
1010
1011     /* Update outer/inner flops */
1012
1013     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*378);
1014 }
1015 /*
1016  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1017  * Electrostatics interaction: ReactionField
1018  * VdW interaction:            CubicSplineTable
1019  * Geometry:                   Water4-Water4
1020  * Calculate force/pot:        Force
1021  */
1022 void
1023 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1024                     (t_nblist                    * gmx_restrict       nlist,
1025                      rvec                        * gmx_restrict          xx,
1026                      rvec                        * gmx_restrict          ff,
1027                      t_forcerec                  * gmx_restrict          fr,
1028                      t_mdatoms                   * gmx_restrict     mdatoms,
1029                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1030                      t_nrnb                      * gmx_restrict        nrnb)
1031 {
1032     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1033      * just 0 for non-waters.
1034      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1035      * jnr indices corresponding to data put in the four positions in the SIMD register.
1036      */
1037     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1038     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1039     int              jnrA,jnrB,jnrC,jnrD;
1040     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1041     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1042     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1043     real             rcutoff_scalar;
1044     real             *shiftvec,*fshift,*x,*f;
1045     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1046     real             scratch[4*DIM];
1047     __m128           fscal,rcutoff,rcutoff2,jidxall;
1048     int              vdwioffset0;
1049     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1050     int              vdwioffset1;
1051     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1052     int              vdwioffset2;
1053     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1054     int              vdwioffset3;
1055     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1056     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1057     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1058     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1059     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1060     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1061     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1062     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1063     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1064     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1065     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1066     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1067     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1068     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1069     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1070     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1071     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1072     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1073     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1074     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1075     real             *charge;
1076     int              nvdwtype;
1077     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1078     int              *vdwtype;
1079     real             *vdwparam;
1080     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1081     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1082     __m128i          vfitab;
1083     __m128i          ifour       = _mm_set1_epi32(4);
1084     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1085     real             *vftab;
1086     __m128           dummy_mask,cutoff_mask;
1087     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1088     __m128           one     = _mm_set1_ps(1.0);
1089     __m128           two     = _mm_set1_ps(2.0);
1090     x                = xx[0];
1091     f                = ff[0];
1092
1093     nri              = nlist->nri;
1094     iinr             = nlist->iinr;
1095     jindex           = nlist->jindex;
1096     jjnr             = nlist->jjnr;
1097     shiftidx         = nlist->shift;
1098     gid              = nlist->gid;
1099     shiftvec         = fr->shift_vec[0];
1100     fshift           = fr->fshift[0];
1101     facel            = _mm_set1_ps(fr->epsfac);
1102     charge           = mdatoms->chargeA;
1103     krf              = _mm_set1_ps(fr->ic->k_rf);
1104     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1105     crf              = _mm_set1_ps(fr->ic->c_rf);
1106     nvdwtype         = fr->ntype;
1107     vdwparam         = fr->nbfp;
1108     vdwtype          = mdatoms->typeA;
1109
1110     vftab            = kernel_data->table_vdw->data;
1111     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1112
1113     /* Setup water-specific parameters */
1114     inr              = nlist->iinr[0];
1115     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1116     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1117     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1118     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1119
1120     jq1              = _mm_set1_ps(charge[inr+1]);
1121     jq2              = _mm_set1_ps(charge[inr+2]);
1122     jq3              = _mm_set1_ps(charge[inr+3]);
1123     vdwjidx0A        = 2*vdwtype[inr+0];
1124     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1125     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1126     qq11             = _mm_mul_ps(iq1,jq1);
1127     qq12             = _mm_mul_ps(iq1,jq2);
1128     qq13             = _mm_mul_ps(iq1,jq3);
1129     qq21             = _mm_mul_ps(iq2,jq1);
1130     qq22             = _mm_mul_ps(iq2,jq2);
1131     qq23             = _mm_mul_ps(iq2,jq3);
1132     qq31             = _mm_mul_ps(iq3,jq1);
1133     qq32             = _mm_mul_ps(iq3,jq2);
1134     qq33             = _mm_mul_ps(iq3,jq3);
1135
1136     /* Avoid stupid compiler warnings */
1137     jnrA = jnrB = jnrC = jnrD = 0;
1138     j_coord_offsetA = 0;
1139     j_coord_offsetB = 0;
1140     j_coord_offsetC = 0;
1141     j_coord_offsetD = 0;
1142
1143     outeriter        = 0;
1144     inneriter        = 0;
1145
1146     for(iidx=0;iidx<4*DIM;iidx++)
1147     {
1148         scratch[iidx] = 0.0;
1149     }
1150
1151     /* Start outer loop over neighborlists */
1152     for(iidx=0; iidx<nri; iidx++)
1153     {
1154         /* Load shift vector for this list */
1155         i_shift_offset   = DIM*shiftidx[iidx];
1156
1157         /* Load limits for loop over neighbors */
1158         j_index_start    = jindex[iidx];
1159         j_index_end      = jindex[iidx+1];
1160
1161         /* Get outer coordinate index */
1162         inr              = iinr[iidx];
1163         i_coord_offset   = DIM*inr;
1164
1165         /* Load i particle coords and add shift vector */
1166         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1167                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1168
1169         fix0             = _mm_setzero_ps();
1170         fiy0             = _mm_setzero_ps();
1171         fiz0             = _mm_setzero_ps();
1172         fix1             = _mm_setzero_ps();
1173         fiy1             = _mm_setzero_ps();
1174         fiz1             = _mm_setzero_ps();
1175         fix2             = _mm_setzero_ps();
1176         fiy2             = _mm_setzero_ps();
1177         fiz2             = _mm_setzero_ps();
1178         fix3             = _mm_setzero_ps();
1179         fiy3             = _mm_setzero_ps();
1180         fiz3             = _mm_setzero_ps();
1181
1182         /* Start inner kernel loop */
1183         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1184         {
1185
1186             /* Get j neighbor index, and coordinate index */
1187             jnrA             = jjnr[jidx];
1188             jnrB             = jjnr[jidx+1];
1189             jnrC             = jjnr[jidx+2];
1190             jnrD             = jjnr[jidx+3];
1191             j_coord_offsetA  = DIM*jnrA;
1192             j_coord_offsetB  = DIM*jnrB;
1193             j_coord_offsetC  = DIM*jnrC;
1194             j_coord_offsetD  = DIM*jnrD;
1195
1196             /* load j atom coordinates */
1197             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1198                                               x+j_coord_offsetC,x+j_coord_offsetD,
1199                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1200                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1201
1202             /* Calculate displacement vector */
1203             dx00             = _mm_sub_ps(ix0,jx0);
1204             dy00             = _mm_sub_ps(iy0,jy0);
1205             dz00             = _mm_sub_ps(iz0,jz0);
1206             dx11             = _mm_sub_ps(ix1,jx1);
1207             dy11             = _mm_sub_ps(iy1,jy1);
1208             dz11             = _mm_sub_ps(iz1,jz1);
1209             dx12             = _mm_sub_ps(ix1,jx2);
1210             dy12             = _mm_sub_ps(iy1,jy2);
1211             dz12             = _mm_sub_ps(iz1,jz2);
1212             dx13             = _mm_sub_ps(ix1,jx3);
1213             dy13             = _mm_sub_ps(iy1,jy3);
1214             dz13             = _mm_sub_ps(iz1,jz3);
1215             dx21             = _mm_sub_ps(ix2,jx1);
1216             dy21             = _mm_sub_ps(iy2,jy1);
1217             dz21             = _mm_sub_ps(iz2,jz1);
1218             dx22             = _mm_sub_ps(ix2,jx2);
1219             dy22             = _mm_sub_ps(iy2,jy2);
1220             dz22             = _mm_sub_ps(iz2,jz2);
1221             dx23             = _mm_sub_ps(ix2,jx3);
1222             dy23             = _mm_sub_ps(iy2,jy3);
1223             dz23             = _mm_sub_ps(iz2,jz3);
1224             dx31             = _mm_sub_ps(ix3,jx1);
1225             dy31             = _mm_sub_ps(iy3,jy1);
1226             dz31             = _mm_sub_ps(iz3,jz1);
1227             dx32             = _mm_sub_ps(ix3,jx2);
1228             dy32             = _mm_sub_ps(iy3,jy2);
1229             dz32             = _mm_sub_ps(iz3,jz2);
1230             dx33             = _mm_sub_ps(ix3,jx3);
1231             dy33             = _mm_sub_ps(iy3,jy3);
1232             dz33             = _mm_sub_ps(iz3,jz3);
1233
1234             /* Calculate squared distance and things based on it */
1235             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1236             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1237             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1238             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1239             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1240             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1241             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1242             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1243             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1244             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1245
1246             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1247             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1248             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1249             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1250             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1251             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1252             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1253             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1254             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1255             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1256
1257             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1258             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1259             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1260             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1261             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1262             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1263             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1264             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1265             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1266
1267             fjx0             = _mm_setzero_ps();
1268             fjy0             = _mm_setzero_ps();
1269             fjz0             = _mm_setzero_ps();
1270             fjx1             = _mm_setzero_ps();
1271             fjy1             = _mm_setzero_ps();
1272             fjz1             = _mm_setzero_ps();
1273             fjx2             = _mm_setzero_ps();
1274             fjy2             = _mm_setzero_ps();
1275             fjz2             = _mm_setzero_ps();
1276             fjx3             = _mm_setzero_ps();
1277             fjy3             = _mm_setzero_ps();
1278             fjz3             = _mm_setzero_ps();
1279
1280             /**************************
1281              * CALCULATE INTERACTIONS *
1282              **************************/
1283
1284             r00              = _mm_mul_ps(rsq00,rinv00);
1285
1286             /* Calculate table index by multiplying r with table scale and truncate to integer */
1287             rt               = _mm_mul_ps(r00,vftabscale);
1288             vfitab           = _mm_cvttps_epi32(rt);
1289 #ifdef __XOP__
1290             vfeps            = _mm_frcz_ps(rt);
1291 #else
1292             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1293 #endif
1294             twovfeps         = _mm_add_ps(vfeps,vfeps);
1295             vfitab           = _mm_slli_epi32(vfitab,3);
1296
1297             /* CUBIC SPLINE TABLE DISPERSION */
1298             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1299             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1300             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1301             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1302             _MM_TRANSPOSE4_PS(Y,F,G,H);
1303             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1304             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1305             fvdw6            = _mm_mul_ps(c6_00,FF);
1306
1307             /* CUBIC SPLINE TABLE REPULSION */
1308             vfitab           = _mm_add_epi32(vfitab,ifour);
1309             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1310             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1311             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1312             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1313             _MM_TRANSPOSE4_PS(Y,F,G,H);
1314             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1315             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1316             fvdw12           = _mm_mul_ps(c12_00,FF);
1317             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1318
1319             fscal            = fvdw;
1320
1321              /* Update vectorial force */
1322             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1323             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1324             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1325
1326             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1327             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1328             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1329
1330             /**************************
1331              * CALCULATE INTERACTIONS *
1332              **************************/
1333
1334             /* REACTION-FIELD ELECTROSTATICS */
1335             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1336
1337             fscal            = felec;
1338
1339              /* Update vectorial force */
1340             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1341             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1342             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1343
1344             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1345             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1346             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* REACTION-FIELD ELECTROSTATICS */
1353             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1354
1355             fscal            = felec;
1356
1357              /* Update vectorial force */
1358             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1359             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1360             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1361
1362             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1363             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1364             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* REACTION-FIELD ELECTROSTATICS */
1371             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1372
1373             fscal            = felec;
1374
1375              /* Update vectorial force */
1376             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1377             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1378             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1379
1380             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1381             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1382             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             /* REACTION-FIELD ELECTROSTATICS */
1389             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1390
1391             fscal            = felec;
1392
1393              /* Update vectorial force */
1394             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1395             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1396             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1397
1398             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1399             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1400             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1401
1402             /**************************
1403              * CALCULATE INTERACTIONS *
1404              **************************/
1405
1406             /* REACTION-FIELD ELECTROSTATICS */
1407             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1408
1409             fscal            = felec;
1410
1411              /* Update vectorial force */
1412             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1413             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1414             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1415
1416             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1417             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1418             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1419
1420             /**************************
1421              * CALCULATE INTERACTIONS *
1422              **************************/
1423
1424             /* REACTION-FIELD ELECTROSTATICS */
1425             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1426
1427             fscal            = felec;
1428
1429              /* Update vectorial force */
1430             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1431             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1432             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1433
1434             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1435             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1436             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             /* REACTION-FIELD ELECTROSTATICS */
1443             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1444
1445             fscal            = felec;
1446
1447              /* Update vectorial force */
1448             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1449             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1450             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1451
1452             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1453             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1454             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1455
1456             /**************************
1457              * CALCULATE INTERACTIONS *
1458              **************************/
1459
1460             /* REACTION-FIELD ELECTROSTATICS */
1461             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1462
1463             fscal            = felec;
1464
1465              /* Update vectorial force */
1466             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1467             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1468             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1469
1470             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1471             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1472             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1473
1474             /**************************
1475              * CALCULATE INTERACTIONS *
1476              **************************/
1477
1478             /* REACTION-FIELD ELECTROSTATICS */
1479             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1480
1481             fscal            = felec;
1482
1483              /* Update vectorial force */
1484             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1485             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1486             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1487
1488             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1489             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1490             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1491
1492             fjptrA             = f+j_coord_offsetA;
1493             fjptrB             = f+j_coord_offsetB;
1494             fjptrC             = f+j_coord_offsetC;
1495             fjptrD             = f+j_coord_offsetD;
1496
1497             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1498                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1499                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1500
1501             /* Inner loop uses 324 flops */
1502         }
1503
1504         if(jidx<j_index_end)
1505         {
1506
1507             /* Get j neighbor index, and coordinate index */
1508             jnrlistA         = jjnr[jidx];
1509             jnrlistB         = jjnr[jidx+1];
1510             jnrlistC         = jjnr[jidx+2];
1511             jnrlistD         = jjnr[jidx+3];
1512             /* Sign of each element will be negative for non-real atoms.
1513              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1514              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1515              */
1516             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1517             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1518             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1519             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1520             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1521             j_coord_offsetA  = DIM*jnrA;
1522             j_coord_offsetB  = DIM*jnrB;
1523             j_coord_offsetC  = DIM*jnrC;
1524             j_coord_offsetD  = DIM*jnrD;
1525
1526             /* load j atom coordinates */
1527             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1528                                               x+j_coord_offsetC,x+j_coord_offsetD,
1529                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1530                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1531
1532             /* Calculate displacement vector */
1533             dx00             = _mm_sub_ps(ix0,jx0);
1534             dy00             = _mm_sub_ps(iy0,jy0);
1535             dz00             = _mm_sub_ps(iz0,jz0);
1536             dx11             = _mm_sub_ps(ix1,jx1);
1537             dy11             = _mm_sub_ps(iy1,jy1);
1538             dz11             = _mm_sub_ps(iz1,jz1);
1539             dx12             = _mm_sub_ps(ix1,jx2);
1540             dy12             = _mm_sub_ps(iy1,jy2);
1541             dz12             = _mm_sub_ps(iz1,jz2);
1542             dx13             = _mm_sub_ps(ix1,jx3);
1543             dy13             = _mm_sub_ps(iy1,jy3);
1544             dz13             = _mm_sub_ps(iz1,jz3);
1545             dx21             = _mm_sub_ps(ix2,jx1);
1546             dy21             = _mm_sub_ps(iy2,jy1);
1547             dz21             = _mm_sub_ps(iz2,jz1);
1548             dx22             = _mm_sub_ps(ix2,jx2);
1549             dy22             = _mm_sub_ps(iy2,jy2);
1550             dz22             = _mm_sub_ps(iz2,jz2);
1551             dx23             = _mm_sub_ps(ix2,jx3);
1552             dy23             = _mm_sub_ps(iy2,jy3);
1553             dz23             = _mm_sub_ps(iz2,jz3);
1554             dx31             = _mm_sub_ps(ix3,jx1);
1555             dy31             = _mm_sub_ps(iy3,jy1);
1556             dz31             = _mm_sub_ps(iz3,jz1);
1557             dx32             = _mm_sub_ps(ix3,jx2);
1558             dy32             = _mm_sub_ps(iy3,jy2);
1559             dz32             = _mm_sub_ps(iz3,jz2);
1560             dx33             = _mm_sub_ps(ix3,jx3);
1561             dy33             = _mm_sub_ps(iy3,jy3);
1562             dz33             = _mm_sub_ps(iz3,jz3);
1563
1564             /* Calculate squared distance and things based on it */
1565             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1566             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1567             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1568             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1569             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1570             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1571             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1572             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1573             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1574             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1575
1576             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1577             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1578             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1579             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1580             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1581             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1582             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1583             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1584             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1585             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1586
1587             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1588             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1589             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1590             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1591             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1592             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1593             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1594             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1595             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1596
1597             fjx0             = _mm_setzero_ps();
1598             fjy0             = _mm_setzero_ps();
1599             fjz0             = _mm_setzero_ps();
1600             fjx1             = _mm_setzero_ps();
1601             fjy1             = _mm_setzero_ps();
1602             fjz1             = _mm_setzero_ps();
1603             fjx2             = _mm_setzero_ps();
1604             fjy2             = _mm_setzero_ps();
1605             fjz2             = _mm_setzero_ps();
1606             fjx3             = _mm_setzero_ps();
1607             fjy3             = _mm_setzero_ps();
1608             fjz3             = _mm_setzero_ps();
1609
1610             /**************************
1611              * CALCULATE INTERACTIONS *
1612              **************************/
1613
1614             r00              = _mm_mul_ps(rsq00,rinv00);
1615             r00              = _mm_andnot_ps(dummy_mask,r00);
1616
1617             /* Calculate table index by multiplying r with table scale and truncate to integer */
1618             rt               = _mm_mul_ps(r00,vftabscale);
1619             vfitab           = _mm_cvttps_epi32(rt);
1620 #ifdef __XOP__
1621             vfeps            = _mm_frcz_ps(rt);
1622 #else
1623             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1624 #endif
1625             twovfeps         = _mm_add_ps(vfeps,vfeps);
1626             vfitab           = _mm_slli_epi32(vfitab,3);
1627
1628             /* CUBIC SPLINE TABLE DISPERSION */
1629             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1630             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1631             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1632             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1633             _MM_TRANSPOSE4_PS(Y,F,G,H);
1634             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1635             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1636             fvdw6            = _mm_mul_ps(c6_00,FF);
1637
1638             /* CUBIC SPLINE TABLE REPULSION */
1639             vfitab           = _mm_add_epi32(vfitab,ifour);
1640             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1641             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1642             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1643             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1644             _MM_TRANSPOSE4_PS(Y,F,G,H);
1645             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1646             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1647             fvdw12           = _mm_mul_ps(c12_00,FF);
1648             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1649
1650             fscal            = fvdw;
1651
1652             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1653
1654              /* Update vectorial force */
1655             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1656             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1657             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1658
1659             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1660             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1661             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1662
1663             /**************************
1664              * CALCULATE INTERACTIONS *
1665              **************************/
1666
1667             /* REACTION-FIELD ELECTROSTATICS */
1668             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1669
1670             fscal            = felec;
1671
1672             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1673
1674              /* Update vectorial force */
1675             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1676             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1677             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1678
1679             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1680             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1681             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1682
1683             /**************************
1684              * CALCULATE INTERACTIONS *
1685              **************************/
1686
1687             /* REACTION-FIELD ELECTROSTATICS */
1688             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1689
1690             fscal            = felec;
1691
1692             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1693
1694              /* Update vectorial force */
1695             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1696             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1697             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1698
1699             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1700             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1701             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1702
1703             /**************************
1704              * CALCULATE INTERACTIONS *
1705              **************************/
1706
1707             /* REACTION-FIELD ELECTROSTATICS */
1708             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1709
1710             fscal            = felec;
1711
1712             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1713
1714              /* Update vectorial force */
1715             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1716             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1717             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1718
1719             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1720             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1721             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1722
1723             /**************************
1724              * CALCULATE INTERACTIONS *
1725              **************************/
1726
1727             /* REACTION-FIELD ELECTROSTATICS */
1728             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1729
1730             fscal            = felec;
1731
1732             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1733
1734              /* Update vectorial force */
1735             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1736             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1737             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1738
1739             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1740             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1741             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1742
1743             /**************************
1744              * CALCULATE INTERACTIONS *
1745              **************************/
1746
1747             /* REACTION-FIELD ELECTROSTATICS */
1748             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1749
1750             fscal            = felec;
1751
1752             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1753
1754              /* Update vectorial force */
1755             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1756             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1757             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1758
1759             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1760             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1761             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1762
1763             /**************************
1764              * CALCULATE INTERACTIONS *
1765              **************************/
1766
1767             /* REACTION-FIELD ELECTROSTATICS */
1768             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1769
1770             fscal            = felec;
1771
1772             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1773
1774              /* Update vectorial force */
1775             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1776             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1777             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1778
1779             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1780             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1781             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1782
1783             /**************************
1784              * CALCULATE INTERACTIONS *
1785              **************************/
1786
1787             /* REACTION-FIELD ELECTROSTATICS */
1788             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1789
1790             fscal            = felec;
1791
1792             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1793
1794              /* Update vectorial force */
1795             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1796             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1797             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1798
1799             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1800             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1801             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1802
1803             /**************************
1804              * CALCULATE INTERACTIONS *
1805              **************************/
1806
1807             /* REACTION-FIELD ELECTROSTATICS */
1808             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1809
1810             fscal            = felec;
1811
1812             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1813
1814              /* Update vectorial force */
1815             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1816             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1817             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1818
1819             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1820             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1821             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1822
1823             /**************************
1824              * CALCULATE INTERACTIONS *
1825              **************************/
1826
1827             /* REACTION-FIELD ELECTROSTATICS */
1828             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1829
1830             fscal            = felec;
1831
1832             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1833
1834              /* Update vectorial force */
1835             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1836             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1837             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1838
1839             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1840             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1841             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1842
1843             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1844             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1845             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1846             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1847
1848             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1849                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1850                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1851
1852             /* Inner loop uses 325 flops */
1853         }
1854
1855         /* End of innermost loop */
1856
1857         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1858                                               f+i_coord_offset,fshift+i_shift_offset);
1859
1860         /* Increment number of inner iterations */
1861         inneriter                  += j_index_end - j_index_start;
1862
1863         /* Outer loop uses 24 flops */
1864     }
1865
1866     /* Increment number of outer iterations */
1867     outeriter        += nri;
1868
1869     /* Update outer/inner flops */
1870
1871     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*325);
1872 }