src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_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_ElecRFCut_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_ElecRFCut_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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 172     rcutoff_scalar   = fr->rcoulomb;
 173     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 174     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 175
 176     /* Avoid stupid compiler warnings */
 177     jnrA = jnrB = jnrC = jnrD = 0;
 178     j_coord_offsetA = 0;
 179     j_coord_offsetB = 0;
 180     j_coord_offsetC = 0;
 181     j_coord_offsetD = 0;
 182
 183     outeriter        = 0;
 184     inneriter        = 0;
 185
 186     for(iidx=0;iidx<4*DIM;iidx++)
 187     {
 188         scratch[iidx] = 0.0;
 189     }
 190
 191     /* Start outer loop over neighborlists */
 192     for(iidx=0; iidx<nri; iidx++)
 193     {
 194         /* Load shift vector for this list */
 195         i_shift_offset   = DIM*shiftidx[iidx];
 196
 197         /* Load limits for loop over neighbors */
 198         j_index_start    = jindex[iidx];
 199         j_index_end      = jindex[iidx+1];
 200
 201         /* Get outer coordinate index */
 202         inr              = iinr[iidx];
 203         i_coord_offset   = DIM*inr;
 204
 205         /* Load i particle coords and add shift vector */
 206         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 207                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 208
 209         fix0             = _mm_setzero_ps();
 210         fiy0             = _mm_setzero_ps();
 211         fiz0             = _mm_setzero_ps();
 212         fix1             = _mm_setzero_ps();
 213         fiy1             = _mm_setzero_ps();
 214         fiz1             = _mm_setzero_ps();
 215         fix2             = _mm_setzero_ps();
 216         fiy2             = _mm_setzero_ps();
 217         fiz2             = _mm_setzero_ps();
 218         fix3             = _mm_setzero_ps();
 219         fiy3             = _mm_setzero_ps();
 220         fiz3             = _mm_setzero_ps();
 221
 222         /* Reset potential sums */
 223         velecsum         = _mm_setzero_ps();
 224         vvdwsum          = _mm_setzero_ps();
 225
 226         /* Start inner kernel loop */
 227         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 228         {
 229
 230             /* Get j neighbor index, and coordinate index */
 231             jnrA             = jjnr[jidx];
 232             jnrB             = jjnr[jidx+1];
 233             jnrC             = jjnr[jidx+2];
 234             jnrD             = jjnr[jidx+3];
 235             j_coord_offsetA  = DIM*jnrA;
 236             j_coord_offsetB  = DIM*jnrB;
 237             j_coord_offsetC  = DIM*jnrC;
 238             j_coord_offsetD  = DIM*jnrD;
 239
 240             /* load j atom coordinates */
 241             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 242                                               x+j_coord_offsetC,x+j_coord_offsetD,
 243                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 244                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 245
 246             /* Calculate displacement vector */
 247             dx00             = _mm_sub_ps(ix0,jx0);
 248             dy00             = _mm_sub_ps(iy0,jy0);
 249             dz00             = _mm_sub_ps(iz0,jz0);
 250             dx11             = _mm_sub_ps(ix1,jx1);
 251             dy11             = _mm_sub_ps(iy1,jy1);
 252             dz11             = _mm_sub_ps(iz1,jz1);
 253             dx12             = _mm_sub_ps(ix1,jx2);
 254             dy12             = _mm_sub_ps(iy1,jy2);
 255             dz12             = _mm_sub_ps(iz1,jz2);
 256             dx13             = _mm_sub_ps(ix1,jx3);
 257             dy13             = _mm_sub_ps(iy1,jy3);
 258             dz13             = _mm_sub_ps(iz1,jz3);
 259             dx21             = _mm_sub_ps(ix2,jx1);
 260             dy21             = _mm_sub_ps(iy2,jy1);
 261             dz21             = _mm_sub_ps(iz2,jz1);
 262             dx22             = _mm_sub_ps(ix2,jx2);
 263             dy22             = _mm_sub_ps(iy2,jy2);
 264             dz22             = _mm_sub_ps(iz2,jz2);
 265             dx23             = _mm_sub_ps(ix2,jx3);
 266             dy23             = _mm_sub_ps(iy2,jy3);
 267             dz23             = _mm_sub_ps(iz2,jz3);
 268             dx31             = _mm_sub_ps(ix3,jx1);
 269             dy31             = _mm_sub_ps(iy3,jy1);
 270             dz31             = _mm_sub_ps(iz3,jz1);
 271             dx32             = _mm_sub_ps(ix3,jx2);
 272             dy32             = _mm_sub_ps(iy3,jy2);
 273             dz32             = _mm_sub_ps(iz3,jz2);
 274             dx33             = _mm_sub_ps(ix3,jx3);
 275             dy33             = _mm_sub_ps(iy3,jy3);
 276             dz33             = _mm_sub_ps(iz3,jz3);
 277
 278             /* Calculate squared distance and things based on it */
 279             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 280             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 281             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 282             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 283             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 284             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 285             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 286             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 287             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 288             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 289
 290             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 291             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 292             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 293             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 294             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 295             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 296             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 297             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 298             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 299             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 300
 301             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 302             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 303             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 304             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 305             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 306             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 307             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 308             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 309             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 310
 311             fjx0             = _mm_setzero_ps();
 312             fjy0             = _mm_setzero_ps();
 313             fjz0             = _mm_setzero_ps();
 314             fjx1             = _mm_setzero_ps();
 315             fjy1             = _mm_setzero_ps();
 316             fjz1             = _mm_setzero_ps();
 317             fjx2             = _mm_setzero_ps();
 318             fjy2             = _mm_setzero_ps();
 319             fjz2             = _mm_setzero_ps();
 320             fjx3             = _mm_setzero_ps();
 321             fjy3             = _mm_setzero_ps();
 322             fjz3             = _mm_setzero_ps();
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             if (gmx_mm_any_lt(rsq00,rcutoff2))
 329             {
 330
 331             r00              = _mm_mul_ps(rsq00,rinv00);
 332
 333             /* Calculate table index by multiplying r with table scale and truncate to integer */
 334             rt               = _mm_mul_ps(r00,vftabscale);
 335             vfitab           = _mm_cvttps_epi32(rt);
 336 #ifdef __XOP__
 337             vfeps            = _mm_frcz_ps(rt);
 338 #else
 339             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 340 #endif
 341             twovfeps         = _mm_add_ps(vfeps,vfeps);
 342             vfitab           = _mm_slli_epi32(vfitab,3);
 343
 344             /* CUBIC SPLINE TABLE DISPERSION */
 345             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 346             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 347             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 348             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 349             _MM_TRANSPOSE4_PS(Y,F,G,H);
 350             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 351             VV               = _mm_macc_ps(vfeps,Fp,Y);
 352             vvdw6            = _mm_mul_ps(c6_00,VV);
 353             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 354             fvdw6            = _mm_mul_ps(c6_00,FF);
 355
 356             /* CUBIC SPLINE TABLE REPULSION */
 357             vfitab           = _mm_add_epi32(vfitab,ifour);
 358             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 359             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 360             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 361             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 362             _MM_TRANSPOSE4_PS(Y,F,G,H);
 363             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 364             VV               = _mm_macc_ps(vfeps,Fp,Y);
 365             vvdw12           = _mm_mul_ps(c12_00,VV);
 366             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 367             fvdw12           = _mm_mul_ps(c12_00,FF);
 368             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 369             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 370
 371             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 372
 373             /* Update potential sum for this i atom from the interaction with this j atom. */
 374             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 375             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 376
 377             fscal            = fvdw;
 378
 379             fscal            = _mm_and_ps(fscal,cutoff_mask);
 380
 381              /* Update vectorial force */
 382             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 383             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 384             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 385
 386             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 387             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 388             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 389
 390             }
 391
 392             /**************************
 393              * CALCULATE INTERACTIONS *
 394              **************************/
 395
 396             if (gmx_mm_any_lt(rsq11,rcutoff2))
 397             {
 398
 399             /* REACTION-FIELD ELECTROSTATICS */
 400             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 401             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 402
 403             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 404
 405             /* Update potential sum for this i atom from the interaction with this j atom. */
 406             velec            = _mm_and_ps(velec,cutoff_mask);
 407             velecsum         = _mm_add_ps(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411             fscal            = _mm_and_ps(fscal,cutoff_mask);
 412
 413              /* Update vectorial force */
 414             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 415             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 416             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 417
 418             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 419             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 420             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 421
 422             }
 423
 424             /**************************
 425              * CALCULATE INTERACTIONS *
 426              **************************/
 427
 428             if (gmx_mm_any_lt(rsq12,rcutoff2))
 429             {
 430
 431             /* REACTION-FIELD ELECTROSTATICS */
 432             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 433             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 434
 435             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 436
 437             /* Update potential sum for this i atom from the interaction with this j atom. */
 438             velec            = _mm_and_ps(velec,cutoff_mask);
 439             velecsum         = _mm_add_ps(velecsum,velec);
 440
 441             fscal            = felec;
 442
 443             fscal            = _mm_and_ps(fscal,cutoff_mask);
 444
 445              /* Update vectorial force */
 446             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 447             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 448             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 449
 450             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 451             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 452             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 453
 454             }
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             if (gmx_mm_any_lt(rsq13,rcutoff2))
 461             {
 462
 463             /* REACTION-FIELD ELECTROSTATICS */
 464             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 465             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 466
 467             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 468
 469             /* Update potential sum for this i atom from the interaction with this j atom. */
 470             velec            = _mm_and_ps(velec,cutoff_mask);
 471             velecsum         = _mm_add_ps(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475             fscal            = _mm_and_ps(fscal,cutoff_mask);
 476
 477              /* Update vectorial force */
 478             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 479             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 480             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 481
 482             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 483             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 484             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 485
 486             }
 487
 488             /**************************
 489              * CALCULATE INTERACTIONS *
 490              **************************/
 491
 492             if (gmx_mm_any_lt(rsq21,rcutoff2))
 493             {
 494
 495             /* REACTION-FIELD ELECTROSTATICS */
 496             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 497             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 498
 499             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 500
 501             /* Update potential sum for this i atom from the interaction with this j atom. */
 502             velec            = _mm_and_ps(velec,cutoff_mask);
 503             velecsum         = _mm_add_ps(velecsum,velec);
 504
 505             fscal            = felec;
 506
 507             fscal            = _mm_and_ps(fscal,cutoff_mask);
 508
 509              /* Update vectorial force */
 510             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 511             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 512             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 513
 514             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 515             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 516             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 517
 518             }
 519
 520             /**************************
 521              * CALCULATE INTERACTIONS *
 522              **************************/
 523
 524             if (gmx_mm_any_lt(rsq22,rcutoff2))
 525             {
 526
 527             /* REACTION-FIELD ELECTROSTATICS */
 528             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 529             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 530
 531             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 532
 533             /* Update potential sum for this i atom from the interaction with this j atom. */
 534             velec            = _mm_and_ps(velec,cutoff_mask);
 535             velecsum         = _mm_add_ps(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539             fscal            = _mm_and_ps(fscal,cutoff_mask);
 540
 541              /* Update vectorial force */
 542             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 543             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 544             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 545
 546             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 547             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 548             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 549
 550             }
 551
 552             /**************************
 553              * CALCULATE INTERACTIONS *
 554              **************************/
 555
 556             if (gmx_mm_any_lt(rsq23,rcutoff2))
 557             {
 558
 559             /* REACTION-FIELD ELECTROSTATICS */
 560             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 561             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 562
 563             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 564
 565             /* Update potential sum for this i atom from the interaction with this j atom. */
 566             velec            = _mm_and_ps(velec,cutoff_mask);
 567             velecsum         = _mm_add_ps(velecsum,velec);
 568
 569             fscal            = felec;
 570
 571             fscal            = _mm_and_ps(fscal,cutoff_mask);
 572
 573              /* Update vectorial force */
 574             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 575             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 576             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 577
 578             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 579             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 580             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 581
 582             }
 583
 584             /**************************
 585              * CALCULATE INTERACTIONS *
 586              **************************/
 587
 588             if (gmx_mm_any_lt(rsq31,rcutoff2))
 589             {
 590
 591             /* REACTION-FIELD ELECTROSTATICS */
 592             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 593             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 594
 595             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 596
 597             /* Update potential sum for this i atom from the interaction with this j atom. */
 598             velec            = _mm_and_ps(velec,cutoff_mask);
 599             velecsum         = _mm_add_ps(velecsum,velec);
 600
 601             fscal            = felec;
 602
 603             fscal            = _mm_and_ps(fscal,cutoff_mask);
 604
 605              /* Update vectorial force */
 606             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 607             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 608             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 609
 610             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 611             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 612             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 613
 614             }
 615
 616             /**************************
 617              * CALCULATE INTERACTIONS *
 618              **************************/
 619
 620             if (gmx_mm_any_lt(rsq32,rcutoff2))
 621             {
 622
 623             /* REACTION-FIELD ELECTROSTATICS */
 624             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 625             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 626
 627             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 628
 629             /* Update potential sum for this i atom from the interaction with this j atom. */
 630             velec            = _mm_and_ps(velec,cutoff_mask);
 631             velecsum         = _mm_add_ps(velecsum,velec);
 632
 633             fscal            = felec;
 634
 635             fscal            = _mm_and_ps(fscal,cutoff_mask);
 636
 637              /* Update vectorial force */
 638             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 639             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 640             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 641
 642             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 643             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 644             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 645
 646             }
 647
 648             /**************************
 649              * CALCULATE INTERACTIONS *
 650              **************************/
 651
 652             if (gmx_mm_any_lt(rsq33,rcutoff2))
 653             {
 654
 655             /* REACTION-FIELD ELECTROSTATICS */
 656             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 657             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 658
 659             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 660
 661             /* Update potential sum for this i atom from the interaction with this j atom. */
 662             velec            = _mm_and_ps(velec,cutoff_mask);
 663             velecsum         = _mm_add_ps(velecsum,velec);
 664
 665             fscal            = felec;
 666
 667             fscal            = _mm_and_ps(fscal,cutoff_mask);
 668
 669              /* Update vectorial force */
 670             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 671             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 672             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 673
 674             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 675             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 676             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 677
 678             }
 679
 680             fjptrA             = f+j_coord_offsetA;
 681             fjptrB             = f+j_coord_offsetB;
 682             fjptrC             = f+j_coord_offsetC;
 683             fjptrD             = f+j_coord_offsetD;
 684
 685             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 686                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 687                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 688
 689             /* Inner loop uses 417 flops */
 690         }
 691
 692         if(jidx<j_index_end)
 693         {
 694
 695             /* Get j neighbor index, and coordinate index */
 696             jnrlistA         = jjnr[jidx];
 697             jnrlistB         = jjnr[jidx+1];
 698             jnrlistC         = jjnr[jidx+2];
 699             jnrlistD         = jjnr[jidx+3];
 700             /* Sign of each element will be negative for non-real atoms.
 701              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 702              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 703              */
 704             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 705             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 706             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 707             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 708             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 709             j_coord_offsetA  = DIM*jnrA;
 710             j_coord_offsetB  = DIM*jnrB;
 711             j_coord_offsetC  = DIM*jnrC;
 712             j_coord_offsetD  = DIM*jnrD;
 713
 714             /* load j atom coordinates */
 715             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 716                                               x+j_coord_offsetC,x+j_coord_offsetD,
 717                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 718                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 719
 720             /* Calculate displacement vector */
 721             dx00             = _mm_sub_ps(ix0,jx0);
 722             dy00             = _mm_sub_ps(iy0,jy0);
 723             dz00             = _mm_sub_ps(iz0,jz0);
 724             dx11             = _mm_sub_ps(ix1,jx1);
 725             dy11             = _mm_sub_ps(iy1,jy1);
 726             dz11             = _mm_sub_ps(iz1,jz1);
 727             dx12             = _mm_sub_ps(ix1,jx2);
 728             dy12             = _mm_sub_ps(iy1,jy2);
 729             dz12             = _mm_sub_ps(iz1,jz2);
 730             dx13             = _mm_sub_ps(ix1,jx3);
 731             dy13             = _mm_sub_ps(iy1,jy3);
 732             dz13             = _mm_sub_ps(iz1,jz3);
 733             dx21             = _mm_sub_ps(ix2,jx1);
 734             dy21             = _mm_sub_ps(iy2,jy1);
 735             dz21             = _mm_sub_ps(iz2,jz1);
 736             dx22             = _mm_sub_ps(ix2,jx2);
 737             dy22             = _mm_sub_ps(iy2,jy2);
 738             dz22             = _mm_sub_ps(iz2,jz2);
 739             dx23             = _mm_sub_ps(ix2,jx3);
 740             dy23             = _mm_sub_ps(iy2,jy3);
 741             dz23             = _mm_sub_ps(iz2,jz3);
 742             dx31             = _mm_sub_ps(ix3,jx1);
 743             dy31             = _mm_sub_ps(iy3,jy1);
 744             dz31             = _mm_sub_ps(iz3,jz1);
 745             dx32             = _mm_sub_ps(ix3,jx2);
 746             dy32             = _mm_sub_ps(iy3,jy2);
 747             dz32             = _mm_sub_ps(iz3,jz2);
 748             dx33             = _mm_sub_ps(ix3,jx3);
 749             dy33             = _mm_sub_ps(iy3,jy3);
 750             dz33             = _mm_sub_ps(iz3,jz3);
 751
 752             /* Calculate squared distance and things based on it */
 753             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 754             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 755             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 756             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 757             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 758             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 759             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 760             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 761             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 762             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 763
 764             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 765             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 766             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 767             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 768             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 769             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 770             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 771             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 772             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 773             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 774
 775             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 776             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 777             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 778             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 779             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 780             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 781             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 782             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 783             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 784
 785             fjx0             = _mm_setzero_ps();
 786             fjy0             = _mm_setzero_ps();
 787             fjz0             = _mm_setzero_ps();
 788             fjx1             = _mm_setzero_ps();
 789             fjy1             = _mm_setzero_ps();
 790             fjz1             = _mm_setzero_ps();
 791             fjx2             = _mm_setzero_ps();
 792             fjy2             = _mm_setzero_ps();
 793             fjz2             = _mm_setzero_ps();
 794             fjx3             = _mm_setzero_ps();
 795             fjy3             = _mm_setzero_ps();
 796             fjz3             = _mm_setzero_ps();
 797
 798             /**************************
 799              * CALCULATE INTERACTIONS *
 800              **************************/
 801
 802             if (gmx_mm_any_lt(rsq00,rcutoff2))
 803             {
 804
 805             r00              = _mm_mul_ps(rsq00,rinv00);
 806             r00              = _mm_andnot_ps(dummy_mask,r00);
 807
 808             /* Calculate table index by multiplying r with table scale and truncate to integer */
 809             rt               = _mm_mul_ps(r00,vftabscale);
 810             vfitab           = _mm_cvttps_epi32(rt);
 811 #ifdef __XOP__
 812             vfeps            = _mm_frcz_ps(rt);
 813 #else
 814             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 815 #endif
 816             twovfeps         = _mm_add_ps(vfeps,vfeps);
 817             vfitab           = _mm_slli_epi32(vfitab,3);
 818
 819             /* CUBIC SPLINE TABLE DISPERSION */
 820             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 821             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 822             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 823             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 824             _MM_TRANSPOSE4_PS(Y,F,G,H);
 825             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 826             VV               = _mm_macc_ps(vfeps,Fp,Y);
 827             vvdw6            = _mm_mul_ps(c6_00,VV);
 828             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 829             fvdw6            = _mm_mul_ps(c6_00,FF);
 830
 831             /* CUBIC SPLINE TABLE REPULSION */
 832             vfitab           = _mm_add_epi32(vfitab,ifour);
 833             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 834             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 835             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 836             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 837             _MM_TRANSPOSE4_PS(Y,F,G,H);
 838             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 839             VV               = _mm_macc_ps(vfeps,Fp,Y);
 840             vvdw12           = _mm_mul_ps(c12_00,VV);
 841             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 842             fvdw12           = _mm_mul_ps(c12_00,FF);
 843             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 844             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 845
 846             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 847
 848             /* Update potential sum for this i atom from the interaction with this j atom. */
 849             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 850             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 851             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 852
 853             fscal            = fvdw;
 854
 855             fscal            = _mm_and_ps(fscal,cutoff_mask);
 856
 857             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 858
 859              /* Update vectorial force */
 860             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 861             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 862             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 863
 864             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 865             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 866             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 867
 868             }
 869
 870             /**************************
 871              * CALCULATE INTERACTIONS *
 872              **************************/
 873
 874             if (gmx_mm_any_lt(rsq11,rcutoff2))
 875             {
 876
 877             /* REACTION-FIELD ELECTROSTATICS */
 878             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 879             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 880
 881             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 882
 883             /* Update potential sum for this i atom from the interaction with this j atom. */
 884             velec            = _mm_and_ps(velec,cutoff_mask);
 885             velec            = _mm_andnot_ps(dummy_mask,velec);
 886             velecsum         = _mm_add_ps(velecsum,velec);
 887
 888             fscal            = felec;
 889
 890             fscal            = _mm_and_ps(fscal,cutoff_mask);
 891
 892             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 893
 894              /* Update vectorial force */
 895             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 896             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 897             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 898
 899             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 900             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 901             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 902
 903             }
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             if (gmx_mm_any_lt(rsq12,rcutoff2))
 910             {
 911
 912             /* REACTION-FIELD ELECTROSTATICS */
 913             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 914             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 915
 916             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 917
 918             /* Update potential sum for this i atom from the interaction with this j atom. */
 919             velec            = _mm_and_ps(velec,cutoff_mask);
 920             velec            = _mm_andnot_ps(dummy_mask,velec);
 921             velecsum         = _mm_add_ps(velecsum,velec);
 922
 923             fscal            = felec;
 924
 925             fscal            = _mm_and_ps(fscal,cutoff_mask);
 926
 927             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 928
 929              /* Update vectorial force */
 930             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 931             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 932             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 933
 934             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 935             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 936             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 937
 938             }
 939
 940             /**************************
 941              * CALCULATE INTERACTIONS *
 942              **************************/
 943
 944             if (gmx_mm_any_lt(rsq13,rcutoff2))
 945             {
 946
 947             /* REACTION-FIELD ELECTROSTATICS */
 948             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 949             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 950
 951             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 952
 953             /* Update potential sum for this i atom from the interaction with this j atom. */
 954             velec            = _mm_and_ps(velec,cutoff_mask);
 955             velec            = _mm_andnot_ps(dummy_mask,velec);
 956             velecsum         = _mm_add_ps(velecsum,velec);
 957
 958             fscal            = felec;
 959
 960             fscal            = _mm_and_ps(fscal,cutoff_mask);
 961
 962             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 963
 964              /* Update vectorial force */
 965             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 966             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 967             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 968
 969             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 970             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 971             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 972
 973             }
 974
 975             /**************************
 976              * CALCULATE INTERACTIONS *
 977              **************************/
 978
 979             if (gmx_mm_any_lt(rsq21,rcutoff2))
 980             {
 981
 982             /* REACTION-FIELD ELECTROSTATICS */
 983             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 984             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 985
 986             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 987
 988             /* Update potential sum for this i atom from the interaction with this j atom. */
 989             velec            = _mm_and_ps(velec,cutoff_mask);
 990             velec            = _mm_andnot_ps(dummy_mask,velec);
 991             velecsum         = _mm_add_ps(velecsum,velec);
 992
 993             fscal            = felec;
 994
 995             fscal            = _mm_and_ps(fscal,cutoff_mask);
 996
 997             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 998
 999              /* Update vectorial force */
1000             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1001             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1002             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1003
1004             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1005             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1006             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1007
1008             }
1009
1010             /**************************
1011              * CALCULATE INTERACTIONS *
1012              **************************/
1013
1014             if (gmx_mm_any_lt(rsq22,rcutoff2))
1015             {
1016
1017             /* REACTION-FIELD ELECTROSTATICS */
1018             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
1019             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1020
1021             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1022
1023             /* Update potential sum for this i atom from the interaction with this j atom. */
1024             velec            = _mm_and_ps(velec,cutoff_mask);
1025             velec            = _mm_andnot_ps(dummy_mask,velec);
1026             velecsum         = _mm_add_ps(velecsum,velec);
1027
1028             fscal            = felec;
1029
1030             fscal            = _mm_and_ps(fscal,cutoff_mask);
1031
1032             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1033
1034              /* Update vectorial force */
1035             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1036             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1037             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1038
1039             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1040             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1041             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1042
1043             }
1044
1045             /**************************
1046              * CALCULATE INTERACTIONS *
1047              **************************/
1048
1049             if (gmx_mm_any_lt(rsq23,rcutoff2))
1050             {
1051
1052             /* REACTION-FIELD ELECTROSTATICS */
1053             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
1054             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1055
1056             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1057
1058             /* Update potential sum for this i atom from the interaction with this j atom. */
1059             velec            = _mm_and_ps(velec,cutoff_mask);
1060             velec            = _mm_andnot_ps(dummy_mask,velec);
1061             velecsum         = _mm_add_ps(velecsum,velec);
1062
1063             fscal            = felec;
1064
1065             fscal            = _mm_and_ps(fscal,cutoff_mask);
1066
1067             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1068
1069              /* Update vectorial force */
1070             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1071             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1072             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1073
1074             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1075             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1076             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1077
1078             }
1079
1080             /**************************
1081              * CALCULATE INTERACTIONS *
1082              **************************/
1083
1084             if (gmx_mm_any_lt(rsq31,rcutoff2))
1085             {
1086
1087             /* REACTION-FIELD ELECTROSTATICS */
1088             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
1089             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1090
1091             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1092
1093             /* Update potential sum for this i atom from the interaction with this j atom. */
1094             velec            = _mm_and_ps(velec,cutoff_mask);
1095             velec            = _mm_andnot_ps(dummy_mask,velec);
1096             velecsum         = _mm_add_ps(velecsum,velec);
1097
1098             fscal            = felec;
1099
1100             fscal            = _mm_and_ps(fscal,cutoff_mask);
1101
1102             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1103
1104              /* Update vectorial force */
1105             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1106             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1107             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1108
1109             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1110             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1111             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1112
1113             }
1114
1115             /**************************
1116              * CALCULATE INTERACTIONS *
1117              **************************/
1118
1119             if (gmx_mm_any_lt(rsq32,rcutoff2))
1120             {
1121
1122             /* REACTION-FIELD ELECTROSTATICS */
1123             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
1124             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1125
1126             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1127
1128             /* Update potential sum for this i atom from the interaction with this j atom. */
1129             velec            = _mm_and_ps(velec,cutoff_mask);
1130             velec            = _mm_andnot_ps(dummy_mask,velec);
1131             velecsum         = _mm_add_ps(velecsum,velec);
1132
1133             fscal            = felec;
1134
1135             fscal            = _mm_and_ps(fscal,cutoff_mask);
1136
1137             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1138
1139              /* Update vectorial force */
1140             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1141             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1142             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1143
1144             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1145             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1146             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1147
1148             }
1149
1150             /**************************
1151              * CALCULATE INTERACTIONS *
1152              **************************/
1153
1154             if (gmx_mm_any_lt(rsq33,rcutoff2))
1155             {
1156
1157             /* REACTION-FIELD ELECTROSTATICS */
1158             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
1159             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1160
1161             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1162
1163             /* Update potential sum for this i atom from the interaction with this j atom. */
1164             velec            = _mm_and_ps(velec,cutoff_mask);
1165             velec            = _mm_andnot_ps(dummy_mask,velec);
1166             velecsum         = _mm_add_ps(velecsum,velec);
1167
1168             fscal            = felec;
1169
1170             fscal            = _mm_and_ps(fscal,cutoff_mask);
1171
1172             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1173
1174              /* Update vectorial force */
1175             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1176             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1177             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1178
1179             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1180             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1181             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1182
1183             }
1184
1185             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1186             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1187             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1188             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1189
1190             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1191                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1192                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1193
1194             /* Inner loop uses 418 flops */
1195         }
1196
1197         /* End of innermost loop */
1198
1199         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1200                                               f+i_coord_offset,fshift+i_shift_offset);
1201
1202         ggid                        = gid[iidx];
1203         /* Update potential energies */
1204         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1205         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1206
1207         /* Increment number of inner iterations */
1208         inneriter                  += j_index_end - j_index_start;
1209
1210         /* Outer loop uses 26 flops */
1211     }
1212
1213     /* Increment number of outer iterations */
1214     outeriter        += nri;
1215
1216     /* Update outer/inner flops */
1217
1218     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*418);
1219 }
1220 /*
1221  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1222  * Electrostatics interaction: ReactionField
1223  * VdW interaction:            CubicSplineTable
1224  * Geometry:                   Water4-Water4
1225  * Calculate force/pot:        Force
1226  */
1227 void
1228 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1229                     (t_nblist                    * gmx_restrict       nlist,
1230                      rvec                        * gmx_restrict          xx,
1231                      rvec                        * gmx_restrict          ff,
1232                      t_forcerec                  * gmx_restrict          fr,
1233                      t_mdatoms                   * gmx_restrict     mdatoms,
1234                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1235                      t_nrnb                      * gmx_restrict        nrnb)
1236 {
1237     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1238      * just 0 for non-waters.
1239      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1240      * jnr indices corresponding to data put in the four positions in the SIMD register.
1241      */
1242     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1243     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1244     int              jnrA,jnrB,jnrC,jnrD;
1245     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1246     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1247     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1248     real             rcutoff_scalar;
1249     real             *shiftvec,*fshift,*x,*f;
1250     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1251     real             scratch[4*DIM];
1252     __m128           fscal,rcutoff,rcutoff2,jidxall;
1253     int              vdwioffset0;
1254     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1255     int              vdwioffset1;
1256     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1257     int              vdwioffset2;
1258     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1259     int              vdwioffset3;
1260     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1261     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1262     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1263     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1264     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1265     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1266     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1267     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1268     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1269     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1270     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1271     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1272     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1273     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1274     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1275     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1276     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1277     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1278     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1279     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1280     real             *charge;
1281     int              nvdwtype;
1282     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1283     int              *vdwtype;
1284     real             *vdwparam;
1285     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1286     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1287     __m128i          vfitab;
1288     __m128i          ifour       = _mm_set1_epi32(4);
1289     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1290     real             *vftab;
1291     __m128           dummy_mask,cutoff_mask;
1292     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1293     __m128           one     = _mm_set1_ps(1.0);
1294     __m128           two     = _mm_set1_ps(2.0);
1295     x                = xx[0];
1296     f                = ff[0];
1297
1298     nri              = nlist->nri;
1299     iinr             = nlist->iinr;
1300     jindex           = nlist->jindex;
1301     jjnr             = nlist->jjnr;
1302     shiftidx         = nlist->shift;
1303     gid              = nlist->gid;
1304     shiftvec         = fr->shift_vec[0];
1305     fshift           = fr->fshift[0];
1306     facel            = _mm_set1_ps(fr->epsfac);
1307     charge           = mdatoms->chargeA;
1308     krf              = _mm_set1_ps(fr->ic->k_rf);
1309     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1310     crf              = _mm_set1_ps(fr->ic->c_rf);
1311     nvdwtype         = fr->ntype;
1312     vdwparam         = fr->nbfp;
1313     vdwtype          = mdatoms->typeA;
1314
1315     vftab            = kernel_data->table_vdw->data;
1316     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1317
1318     /* Setup water-specific parameters */
1319     inr              = nlist->iinr[0];
1320     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1321     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1322     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1323     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1324
1325     jq1              = _mm_set1_ps(charge[inr+1]);
1326     jq2              = _mm_set1_ps(charge[inr+2]);
1327     jq3              = _mm_set1_ps(charge[inr+3]);
1328     vdwjidx0A        = 2*vdwtype[inr+0];
1329     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1330     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1331     qq11             = _mm_mul_ps(iq1,jq1);
1332     qq12             = _mm_mul_ps(iq1,jq2);
1333     qq13             = _mm_mul_ps(iq1,jq3);
1334     qq21             = _mm_mul_ps(iq2,jq1);
1335     qq22             = _mm_mul_ps(iq2,jq2);
1336     qq23             = _mm_mul_ps(iq2,jq3);
1337     qq31             = _mm_mul_ps(iq3,jq1);
1338     qq32             = _mm_mul_ps(iq3,jq2);
1339     qq33             = _mm_mul_ps(iq3,jq3);
1340
1341     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1342     rcutoff_scalar   = fr->rcoulomb;
1343     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1344     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1345
1346     /* Avoid stupid compiler warnings */
1347     jnrA = jnrB = jnrC = jnrD = 0;
1348     j_coord_offsetA = 0;
1349     j_coord_offsetB = 0;
1350     j_coord_offsetC = 0;
1351     j_coord_offsetD = 0;
1352
1353     outeriter        = 0;
1354     inneriter        = 0;
1355
1356     for(iidx=0;iidx<4*DIM;iidx++)
1357     {
1358         scratch[iidx] = 0.0;
1359     }
1360
1361     /* Start outer loop over neighborlists */
1362     for(iidx=0; iidx<nri; iidx++)
1363     {
1364         /* Load shift vector for this list */
1365         i_shift_offset   = DIM*shiftidx[iidx];
1366
1367         /* Load limits for loop over neighbors */
1368         j_index_start    = jindex[iidx];
1369         j_index_end      = jindex[iidx+1];
1370
1371         /* Get outer coordinate index */
1372         inr              = iinr[iidx];
1373         i_coord_offset   = DIM*inr;
1374
1375         /* Load i particle coords and add shift vector */
1376         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1377                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1378
1379         fix0             = _mm_setzero_ps();
1380         fiy0             = _mm_setzero_ps();
1381         fiz0             = _mm_setzero_ps();
1382         fix1             = _mm_setzero_ps();
1383         fiy1             = _mm_setzero_ps();
1384         fiz1             = _mm_setzero_ps();
1385         fix2             = _mm_setzero_ps();
1386         fiy2             = _mm_setzero_ps();
1387         fiz2             = _mm_setzero_ps();
1388         fix3             = _mm_setzero_ps();
1389         fiy3             = _mm_setzero_ps();
1390         fiz3             = _mm_setzero_ps();
1391
1392         /* Start inner kernel loop */
1393         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1394         {
1395
1396             /* Get j neighbor index, and coordinate index */
1397             jnrA             = jjnr[jidx];
1398             jnrB             = jjnr[jidx+1];
1399             jnrC             = jjnr[jidx+2];
1400             jnrD             = jjnr[jidx+3];
1401             j_coord_offsetA  = DIM*jnrA;
1402             j_coord_offsetB  = DIM*jnrB;
1403             j_coord_offsetC  = DIM*jnrC;
1404             j_coord_offsetD  = DIM*jnrD;
1405
1406             /* load j atom coordinates */
1407             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1408                                               x+j_coord_offsetC,x+j_coord_offsetD,
1409                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1410                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1411
1412             /* Calculate displacement vector */
1413             dx00             = _mm_sub_ps(ix0,jx0);
1414             dy00             = _mm_sub_ps(iy0,jy0);
1415             dz00             = _mm_sub_ps(iz0,jz0);
1416             dx11             = _mm_sub_ps(ix1,jx1);
1417             dy11             = _mm_sub_ps(iy1,jy1);
1418             dz11             = _mm_sub_ps(iz1,jz1);
1419             dx12             = _mm_sub_ps(ix1,jx2);
1420             dy12             = _mm_sub_ps(iy1,jy2);
1421             dz12             = _mm_sub_ps(iz1,jz2);
1422             dx13             = _mm_sub_ps(ix1,jx3);
1423             dy13             = _mm_sub_ps(iy1,jy3);
1424             dz13             = _mm_sub_ps(iz1,jz3);
1425             dx21             = _mm_sub_ps(ix2,jx1);
1426             dy21             = _mm_sub_ps(iy2,jy1);
1427             dz21             = _mm_sub_ps(iz2,jz1);
1428             dx22             = _mm_sub_ps(ix2,jx2);
1429             dy22             = _mm_sub_ps(iy2,jy2);
1430             dz22             = _mm_sub_ps(iz2,jz2);
1431             dx23             = _mm_sub_ps(ix2,jx3);
1432             dy23             = _mm_sub_ps(iy2,jy3);
1433             dz23             = _mm_sub_ps(iz2,jz3);
1434             dx31             = _mm_sub_ps(ix3,jx1);
1435             dy31             = _mm_sub_ps(iy3,jy1);
1436             dz31             = _mm_sub_ps(iz3,jz1);
1437             dx32             = _mm_sub_ps(ix3,jx2);
1438             dy32             = _mm_sub_ps(iy3,jy2);
1439             dz32             = _mm_sub_ps(iz3,jz2);
1440             dx33             = _mm_sub_ps(ix3,jx3);
1441             dy33             = _mm_sub_ps(iy3,jy3);
1442             dz33             = _mm_sub_ps(iz3,jz3);
1443
1444             /* Calculate squared distance and things based on it */
1445             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1446             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1447             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1448             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1449             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1450             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1451             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1452             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1453             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1454             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1455
1456             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1457             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1458             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1459             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1460             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1461             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1462             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1463             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1464             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1465             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1466
1467             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1468             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1469             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1470             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1471             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1472             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1473             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1474             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1475             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1476
1477             fjx0             = _mm_setzero_ps();
1478             fjy0             = _mm_setzero_ps();
1479             fjz0             = _mm_setzero_ps();
1480             fjx1             = _mm_setzero_ps();
1481             fjy1             = _mm_setzero_ps();
1482             fjz1             = _mm_setzero_ps();
1483             fjx2             = _mm_setzero_ps();
1484             fjy2             = _mm_setzero_ps();
1485             fjz2             = _mm_setzero_ps();
1486             fjx3             = _mm_setzero_ps();
1487             fjy3             = _mm_setzero_ps();
1488             fjz3             = _mm_setzero_ps();
1489
1490             /**************************
1491              * CALCULATE INTERACTIONS *
1492              **************************/
1493
1494             if (gmx_mm_any_lt(rsq00,rcutoff2))
1495             {
1496
1497             r00              = _mm_mul_ps(rsq00,rinv00);
1498
1499             /* Calculate table index by multiplying r with table scale and truncate to integer */
1500             rt               = _mm_mul_ps(r00,vftabscale);
1501             vfitab           = _mm_cvttps_epi32(rt);
1502 #ifdef __XOP__
1503             vfeps            = _mm_frcz_ps(rt);
1504 #else
1505             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1506 #endif
1507             twovfeps         = _mm_add_ps(vfeps,vfeps);
1508             vfitab           = _mm_slli_epi32(vfitab,3);
1509
1510             /* CUBIC SPLINE TABLE DISPERSION */
1511             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1512             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1513             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1514             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1515             _MM_TRANSPOSE4_PS(Y,F,G,H);
1516             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1517             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1518             fvdw6            = _mm_mul_ps(c6_00,FF);
1519
1520             /* CUBIC SPLINE TABLE REPULSION */
1521             vfitab           = _mm_add_epi32(vfitab,ifour);
1522             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1523             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1524             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1525             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1526             _MM_TRANSPOSE4_PS(Y,F,G,H);
1527             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1528             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1529             fvdw12           = _mm_mul_ps(c12_00,FF);
1530             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1531
1532             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1533
1534             fscal            = fvdw;
1535
1536             fscal            = _mm_and_ps(fscal,cutoff_mask);
1537
1538              /* Update vectorial force */
1539             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1540             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1541             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1542
1543             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1544             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1545             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1546
1547             }
1548
1549             /**************************
1550              * CALCULATE INTERACTIONS *
1551              **************************/
1552
1553             if (gmx_mm_any_lt(rsq11,rcutoff2))
1554             {
1555
1556             /* REACTION-FIELD ELECTROSTATICS */
1557             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1558
1559             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1560
1561             fscal            = felec;
1562
1563             fscal            = _mm_and_ps(fscal,cutoff_mask);
1564
1565              /* Update vectorial force */
1566             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1567             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1568             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1569
1570             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1571             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1572             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1573
1574             }
1575
1576             /**************************
1577              * CALCULATE INTERACTIONS *
1578              **************************/
1579
1580             if (gmx_mm_any_lt(rsq12,rcutoff2))
1581             {
1582
1583             /* REACTION-FIELD ELECTROSTATICS */
1584             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1585
1586             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1587
1588             fscal            = felec;
1589
1590             fscal            = _mm_and_ps(fscal,cutoff_mask);
1591
1592              /* Update vectorial force */
1593             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1594             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1595             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1596
1597             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1598             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1599             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1600
1601             }
1602
1603             /**************************
1604              * CALCULATE INTERACTIONS *
1605              **************************/
1606
1607             if (gmx_mm_any_lt(rsq13,rcutoff2))
1608             {
1609
1610             /* REACTION-FIELD ELECTROSTATICS */
1611             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1612
1613             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1614
1615             fscal            = felec;
1616
1617             fscal            = _mm_and_ps(fscal,cutoff_mask);
1618
1619              /* Update vectorial force */
1620             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1621             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1622             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1623
1624             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1625             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1626             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1627
1628             }
1629
1630             /**************************
1631              * CALCULATE INTERACTIONS *
1632              **************************/
1633
1634             if (gmx_mm_any_lt(rsq21,rcutoff2))
1635             {
1636
1637             /* REACTION-FIELD ELECTROSTATICS */
1638             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1639
1640             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1641
1642             fscal            = felec;
1643
1644             fscal            = _mm_and_ps(fscal,cutoff_mask);
1645
1646              /* Update vectorial force */
1647             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1648             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1649             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1650
1651             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1652             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1653             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1654
1655             }
1656
1657             /**************************
1658              * CALCULATE INTERACTIONS *
1659              **************************/
1660
1661             if (gmx_mm_any_lt(rsq22,rcutoff2))
1662             {
1663
1664             /* REACTION-FIELD ELECTROSTATICS */
1665             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1666
1667             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1668
1669             fscal            = felec;
1670
1671             fscal            = _mm_and_ps(fscal,cutoff_mask);
1672
1673              /* Update vectorial force */
1674             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1675             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1676             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1677
1678             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1679             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1680             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1681
1682             }
1683
1684             /**************************
1685              * CALCULATE INTERACTIONS *
1686              **************************/
1687
1688             if (gmx_mm_any_lt(rsq23,rcutoff2))
1689             {
1690
1691             /* REACTION-FIELD ELECTROSTATICS */
1692             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1693
1694             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1695
1696             fscal            = felec;
1697
1698             fscal            = _mm_and_ps(fscal,cutoff_mask);
1699
1700              /* Update vectorial force */
1701             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1702             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1703             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1704
1705             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1706             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1707             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1708
1709             }
1710
1711             /**************************
1712              * CALCULATE INTERACTIONS *
1713              **************************/
1714
1715             if (gmx_mm_any_lt(rsq31,rcutoff2))
1716             {
1717
1718             /* REACTION-FIELD ELECTROSTATICS */
1719             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1720
1721             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1722
1723             fscal            = felec;
1724
1725             fscal            = _mm_and_ps(fscal,cutoff_mask);
1726
1727              /* Update vectorial force */
1728             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1729             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1730             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1731
1732             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1733             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1734             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1735
1736             }
1737
1738             /**************************
1739              * CALCULATE INTERACTIONS *
1740              **************************/
1741
1742             if (gmx_mm_any_lt(rsq32,rcutoff2))
1743             {
1744
1745             /* REACTION-FIELD ELECTROSTATICS */
1746             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1747
1748             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1749
1750             fscal            = felec;
1751
1752             fscal            = _mm_and_ps(fscal,cutoff_mask);
1753
1754              /* Update vectorial force */
1755             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1756             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1757             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1758
1759             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1760             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1761             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1762
1763             }
1764
1765             /**************************
1766              * CALCULATE INTERACTIONS *
1767              **************************/
1768
1769             if (gmx_mm_any_lt(rsq33,rcutoff2))
1770             {
1771
1772             /* REACTION-FIELD ELECTROSTATICS */
1773             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1774
1775             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1776
1777             fscal            = felec;
1778
1779             fscal            = _mm_and_ps(fscal,cutoff_mask);
1780
1781              /* Update vectorial force */
1782             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1783             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1784             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1785
1786             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1787             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1788             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1789
1790             }
1791
1792             fjptrA             = f+j_coord_offsetA;
1793             fjptrB             = f+j_coord_offsetB;
1794             fjptrC             = f+j_coord_offsetC;
1795             fjptrD             = f+j_coord_offsetD;
1796
1797             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1798                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1799                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1800
1801             /* Inner loop uses 354 flops */
1802         }
1803
1804         if(jidx<j_index_end)
1805         {
1806
1807             /* Get j neighbor index, and coordinate index */
1808             jnrlistA         = jjnr[jidx];
1809             jnrlistB         = jjnr[jidx+1];
1810             jnrlistC         = jjnr[jidx+2];
1811             jnrlistD         = jjnr[jidx+3];
1812             /* Sign of each element will be negative for non-real atoms.
1813              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1814              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1815              */
1816             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1817             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1818             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1819             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1820             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1821             j_coord_offsetA  = DIM*jnrA;
1822             j_coord_offsetB  = DIM*jnrB;
1823             j_coord_offsetC  = DIM*jnrC;
1824             j_coord_offsetD  = DIM*jnrD;
1825
1826             /* load j atom coordinates */
1827             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1828                                               x+j_coord_offsetC,x+j_coord_offsetD,
1829                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1830                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1831
1832             /* Calculate displacement vector */
1833             dx00             = _mm_sub_ps(ix0,jx0);
1834             dy00             = _mm_sub_ps(iy0,jy0);
1835             dz00             = _mm_sub_ps(iz0,jz0);
1836             dx11             = _mm_sub_ps(ix1,jx1);
1837             dy11             = _mm_sub_ps(iy1,jy1);
1838             dz11             = _mm_sub_ps(iz1,jz1);
1839             dx12             = _mm_sub_ps(ix1,jx2);
1840             dy12             = _mm_sub_ps(iy1,jy2);
1841             dz12             = _mm_sub_ps(iz1,jz2);
1842             dx13             = _mm_sub_ps(ix1,jx3);
1843             dy13             = _mm_sub_ps(iy1,jy3);
1844             dz13             = _mm_sub_ps(iz1,jz3);
1845             dx21             = _mm_sub_ps(ix2,jx1);
1846             dy21             = _mm_sub_ps(iy2,jy1);
1847             dz21             = _mm_sub_ps(iz2,jz1);
1848             dx22             = _mm_sub_ps(ix2,jx2);
1849             dy22             = _mm_sub_ps(iy2,jy2);
1850             dz22             = _mm_sub_ps(iz2,jz2);
1851             dx23             = _mm_sub_ps(ix2,jx3);
1852             dy23             = _mm_sub_ps(iy2,jy3);
1853             dz23             = _mm_sub_ps(iz2,jz3);
1854             dx31             = _mm_sub_ps(ix3,jx1);
1855             dy31             = _mm_sub_ps(iy3,jy1);
1856             dz31             = _mm_sub_ps(iz3,jz1);
1857             dx32             = _mm_sub_ps(ix3,jx2);
1858             dy32             = _mm_sub_ps(iy3,jy2);
1859             dz32             = _mm_sub_ps(iz3,jz2);
1860             dx33             = _mm_sub_ps(ix3,jx3);
1861             dy33             = _mm_sub_ps(iy3,jy3);
1862             dz33             = _mm_sub_ps(iz3,jz3);
1863
1864             /* Calculate squared distance and things based on it */
1865             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1866             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1867             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1868             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1869             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1870             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1871             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1872             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1873             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1874             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1875
1876             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1877             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1878             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1879             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1880             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1881             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1882             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1883             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1884             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1885             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1886
1887             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1888             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1889             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1890             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1891             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1892             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1893             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1894             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1895             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1896
1897             fjx0             = _mm_setzero_ps();
1898             fjy0             = _mm_setzero_ps();
1899             fjz0             = _mm_setzero_ps();
1900             fjx1             = _mm_setzero_ps();
1901             fjy1             = _mm_setzero_ps();
1902             fjz1             = _mm_setzero_ps();
1903             fjx2             = _mm_setzero_ps();
1904             fjy2             = _mm_setzero_ps();
1905             fjz2             = _mm_setzero_ps();
1906             fjx3             = _mm_setzero_ps();
1907             fjy3             = _mm_setzero_ps();
1908             fjz3             = _mm_setzero_ps();
1909
1910             /**************************
1911              * CALCULATE INTERACTIONS *
1912              **************************/
1913
1914             if (gmx_mm_any_lt(rsq00,rcutoff2))
1915             {
1916
1917             r00              = _mm_mul_ps(rsq00,rinv00);
1918             r00              = _mm_andnot_ps(dummy_mask,r00);
1919
1920             /* Calculate table index by multiplying r with table scale and truncate to integer */
1921             rt               = _mm_mul_ps(r00,vftabscale);
1922             vfitab           = _mm_cvttps_epi32(rt);
1923 #ifdef __XOP__
1924             vfeps            = _mm_frcz_ps(rt);
1925 #else
1926             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1927 #endif
1928             twovfeps         = _mm_add_ps(vfeps,vfeps);
1929             vfitab           = _mm_slli_epi32(vfitab,3);
1930
1931             /* CUBIC SPLINE TABLE DISPERSION */
1932             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1933             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1934             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1935             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1936             _MM_TRANSPOSE4_PS(Y,F,G,H);
1937             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1938             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1939             fvdw6            = _mm_mul_ps(c6_00,FF);
1940
1941             /* CUBIC SPLINE TABLE REPULSION */
1942             vfitab           = _mm_add_epi32(vfitab,ifour);
1943             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1944             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1945             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1946             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1947             _MM_TRANSPOSE4_PS(Y,F,G,H);
1948             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1949             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1950             fvdw12           = _mm_mul_ps(c12_00,FF);
1951             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1952
1953             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1954
1955             fscal            = fvdw;
1956
1957             fscal            = _mm_and_ps(fscal,cutoff_mask);
1958
1959             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1960
1961              /* Update vectorial force */
1962             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1963             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1964             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1965
1966             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1967             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1968             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1969
1970             }
1971
1972             /**************************
1973              * CALCULATE INTERACTIONS *
1974              **************************/
1975
1976             if (gmx_mm_any_lt(rsq11,rcutoff2))
1977             {
1978
1979             /* REACTION-FIELD ELECTROSTATICS */
1980             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1981
1982             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1983
1984             fscal            = felec;
1985
1986             fscal            = _mm_and_ps(fscal,cutoff_mask);
1987
1988             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1989
1990              /* Update vectorial force */
1991             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1992             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1993             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1994
1995             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1996             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1997             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1998
1999             }
2000
2001             /**************************
2002              * CALCULATE INTERACTIONS *
2003              **************************/
2004
2005             if (gmx_mm_any_lt(rsq12,rcutoff2))
2006             {
2007
2008             /* REACTION-FIELD ELECTROSTATICS */
2009             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
2010
2011             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2012
2013             fscal            = felec;
2014
2015             fscal            = _mm_and_ps(fscal,cutoff_mask);
2016
2017             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2018
2019              /* Update vectorial force */
2020             fix1             = _mm_macc_ps(dx12,fscal,fix1);
2021             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
2022             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
2023
2024             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
2025             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
2026             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
2027
2028             }
2029
2030             /**************************
2031              * CALCULATE INTERACTIONS *
2032              **************************/
2033
2034             if (gmx_mm_any_lt(rsq13,rcutoff2))
2035             {
2036
2037             /* REACTION-FIELD ELECTROSTATICS */
2038             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
2039
2040             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
2041
2042             fscal            = felec;
2043
2044             fscal            = _mm_and_ps(fscal,cutoff_mask);
2045
2046             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2047
2048              /* Update vectorial force */
2049             fix1             = _mm_macc_ps(dx13,fscal,fix1);
2050             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
2051             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
2052
2053             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
2054             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
2055             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
2056
2057             }
2058
2059             /**************************
2060              * CALCULATE INTERACTIONS *
2061              **************************/
2062
2063             if (gmx_mm_any_lt(rsq21,rcutoff2))
2064             {
2065
2066             /* REACTION-FIELD ELECTROSTATICS */
2067             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
2068
2069             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2070
2071             fscal            = felec;
2072
2073             fscal            = _mm_and_ps(fscal,cutoff_mask);
2074
2075             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2076
2077              /* Update vectorial force */
2078             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2079             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2080             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2081
2082             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2083             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2084             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2085
2086             }
2087
2088             /**************************
2089              * CALCULATE INTERACTIONS *
2090              **************************/
2091
2092             if (gmx_mm_any_lt(rsq22,rcutoff2))
2093             {
2094
2095             /* REACTION-FIELD ELECTROSTATICS */
2096             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
2097
2098             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2099
2100             fscal            = felec;
2101
2102             fscal            = _mm_and_ps(fscal,cutoff_mask);
2103
2104             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2105
2106              /* Update vectorial force */
2107             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2108             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2109             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2110
2111             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2112             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2113             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2114
2115             }
2116
2117             /**************************
2118              * CALCULATE INTERACTIONS *
2119              **************************/
2120
2121             if (gmx_mm_any_lt(rsq23,rcutoff2))
2122             {
2123
2124             /* REACTION-FIELD ELECTROSTATICS */
2125             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
2126
2127             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2128
2129             fscal            = felec;
2130
2131             fscal            = _mm_and_ps(fscal,cutoff_mask);
2132
2133             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2134
2135              /* Update vectorial force */
2136             fix2             = _mm_macc_ps(dx23,fscal,fix2);
2137             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
2138             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
2139
2140             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
2141             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
2142             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
2143
2144             }
2145
2146             /**************************
2147              * CALCULATE INTERACTIONS *
2148              **************************/
2149
2150             if (gmx_mm_any_lt(rsq31,rcutoff2))
2151             {
2152
2153             /* REACTION-FIELD ELECTROSTATICS */
2154             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
2155
2156             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2157
2158             fscal            = felec;
2159
2160             fscal            = _mm_and_ps(fscal,cutoff_mask);
2161
2162             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2163
2164              /* Update vectorial force */
2165             fix3             = _mm_macc_ps(dx31,fscal,fix3);
2166             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
2167             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
2168
2169             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
2170             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
2171             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
2172
2173             }
2174
2175             /**************************
2176              * CALCULATE INTERACTIONS *
2177              **************************/
2178
2179             if (gmx_mm_any_lt(rsq32,rcutoff2))
2180             {
2181
2182             /* REACTION-FIELD ELECTROSTATICS */
2183             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
2184
2185             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2186
2187             fscal            = felec;
2188
2189             fscal            = _mm_and_ps(fscal,cutoff_mask);
2190
2191             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2192
2193              /* Update vectorial force */
2194             fix3             = _mm_macc_ps(dx32,fscal,fix3);
2195             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
2196             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
2197
2198             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
2199             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
2200             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
2201
2202             }
2203
2204             /**************************
2205              * CALCULATE INTERACTIONS *
2206              **************************/
2207
2208             if (gmx_mm_any_lt(rsq33,rcutoff2))
2209             {
2210
2211             /* REACTION-FIELD ELECTROSTATICS */
2212             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
2213
2214             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2215
2216             fscal            = felec;
2217
2218             fscal            = _mm_and_ps(fscal,cutoff_mask);
2219
2220             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2221
2222              /* Update vectorial force */
2223             fix3             = _mm_macc_ps(dx33,fscal,fix3);
2224             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
2225             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
2226
2227             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
2228             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
2229             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
2230
2231             }
2232
2233             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2234             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2235             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2236             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2237
2238             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2239                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2240                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2241
2242             /* Inner loop uses 355 flops */
2243         }
2244
2245         /* End of innermost loop */
2246
2247         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2248                                               f+i_coord_offset,fshift+i_shift_offset);
2249
2250         /* Increment number of inner iterations */
2251         inneriter                  += j_index_end - j_index_start;
2252
2253         /* Outer loop uses 24 flops */
2254     }
2255
2256     /* Increment number of outer iterations */
2257     outeriter        += nri;
2258
2259     /* Update outer/inner flops */
2260
2261     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*355);
2262 }