src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  91     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  93     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     int              nvdwtype;
 106     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 107     int              *vdwtype;
 108     real             *vdwparam;
 109     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 110     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 111     __m128i          vfitab;
 112     __m128i          ifour       = _mm_set1_epi32(4);
 113     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 114     real             *vftab;
 115     __m128           dummy_mask,cutoff_mask;
 116     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 117     __m128           one     = _mm_set1_ps(1.0);
 118     __m128           two     = _mm_set1_ps(2.0);
 119     x                = xx[0];
 120     f                = ff[0];
 121
 122     nri              = nlist->nri;
 123     iinr             = nlist->iinr;
 124     jindex           = nlist->jindex;
 125     jjnr             = nlist->jjnr;
 126     shiftidx         = nlist->shift;
 127     gid              = nlist->gid;
 128     shiftvec         = fr->shift_vec[0];
 129     fshift           = fr->fshift[0];
 130     facel            = _mm_set1_ps(fr->ic->epsfac);
 131     charge           = mdatoms->chargeA;
 132     krf              = _mm_set1_ps(fr->ic->k_rf);
 133     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 134     crf              = _mm_set1_ps(fr->ic->c_rf);
 135     nvdwtype         = fr->ntype;
 136     vdwparam         = fr->nbfp;
 137     vdwtype          = mdatoms->typeA;
 138
 139     vftab            = kernel_data->table_vdw->data;
 140     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 141
 142     /* Setup water-specific parameters */
 143     inr              = nlist->iinr[0];
 144     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 145     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 146     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 147     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 148
 149     jq0              = _mm_set1_ps(charge[inr+0]);
 150     jq1              = _mm_set1_ps(charge[inr+1]);
 151     jq2              = _mm_set1_ps(charge[inr+2]);
 152     vdwjidx0A        = 2*vdwtype[inr+0];
 153     qq00             = _mm_mul_ps(iq0,jq0);
 154     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 155     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 156     qq01             = _mm_mul_ps(iq0,jq1);
 157     qq02             = _mm_mul_ps(iq0,jq2);
 158     qq10             = _mm_mul_ps(iq1,jq0);
 159     qq11             = _mm_mul_ps(iq1,jq1);
 160     qq12             = _mm_mul_ps(iq1,jq2);
 161     qq20             = _mm_mul_ps(iq2,jq0);
 162     qq21             = _mm_mul_ps(iq2,jq1);
 163     qq22             = _mm_mul_ps(iq2,jq2);
 164
 165     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 166     rcutoff_scalar   = fr->ic->rcoulomb;
 167     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 168     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 169
 170     /* Avoid stupid compiler warnings */
 171     jnrA = jnrB = jnrC = jnrD = 0;
 172     j_coord_offsetA = 0;
 173     j_coord_offsetB = 0;
 174     j_coord_offsetC = 0;
 175     j_coord_offsetD = 0;
 176
 177     outeriter        = 0;
 178     inneriter        = 0;
 179
 180     for(iidx=0;iidx<4*DIM;iidx++)
 181     {
 182         scratch[iidx] = 0.0;
 183     }
 184
 185     /* Start outer loop over neighborlists */
 186     for(iidx=0; iidx<nri; iidx++)
 187     {
 188         /* Load shift vector for this list */
 189         i_shift_offset   = DIM*shiftidx[iidx];
 190
 191         /* Load limits for loop over neighbors */
 192         j_index_start    = jindex[iidx];
 193         j_index_end      = jindex[iidx+1];
 194
 195         /* Get outer coordinate index */
 196         inr              = iinr[iidx];
 197         i_coord_offset   = DIM*inr;
 198
 199         /* Load i particle coords and add shift vector */
 200         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 201                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 202
 203         fix0             = _mm_setzero_ps();
 204         fiy0             = _mm_setzero_ps();
 205         fiz0             = _mm_setzero_ps();
 206         fix1             = _mm_setzero_ps();
 207         fiy1             = _mm_setzero_ps();
 208         fiz1             = _mm_setzero_ps();
 209         fix2             = _mm_setzero_ps();
 210         fiy2             = _mm_setzero_ps();
 211         fiz2             = _mm_setzero_ps();
 212
 213         /* Reset potential sums */
 214         velecsum         = _mm_setzero_ps();
 215         vvdwsum          = _mm_setzero_ps();
 216
 217         /* Start inner kernel loop */
 218         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 219         {
 220
 221             /* Get j neighbor index, and coordinate index */
 222             jnrA             = jjnr[jidx];
 223             jnrB             = jjnr[jidx+1];
 224             jnrC             = jjnr[jidx+2];
 225             jnrD             = jjnr[jidx+3];
 226             j_coord_offsetA  = DIM*jnrA;
 227             j_coord_offsetB  = DIM*jnrB;
 228             j_coord_offsetC  = DIM*jnrC;
 229             j_coord_offsetD  = DIM*jnrD;
 230
 231             /* load j atom coordinates */
 232             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 233                                               x+j_coord_offsetC,x+j_coord_offsetD,
 234                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 235
 236             /* Calculate displacement vector */
 237             dx00             = _mm_sub_ps(ix0,jx0);
 238             dy00             = _mm_sub_ps(iy0,jy0);
 239             dz00             = _mm_sub_ps(iz0,jz0);
 240             dx01             = _mm_sub_ps(ix0,jx1);
 241             dy01             = _mm_sub_ps(iy0,jy1);
 242             dz01             = _mm_sub_ps(iz0,jz1);
 243             dx02             = _mm_sub_ps(ix0,jx2);
 244             dy02             = _mm_sub_ps(iy0,jy2);
 245             dz02             = _mm_sub_ps(iz0,jz2);
 246             dx10             = _mm_sub_ps(ix1,jx0);
 247             dy10             = _mm_sub_ps(iy1,jy0);
 248             dz10             = _mm_sub_ps(iz1,jz0);
 249             dx11             = _mm_sub_ps(ix1,jx1);
 250             dy11             = _mm_sub_ps(iy1,jy1);
 251             dz11             = _mm_sub_ps(iz1,jz1);
 252             dx12             = _mm_sub_ps(ix1,jx2);
 253             dy12             = _mm_sub_ps(iy1,jy2);
 254             dz12             = _mm_sub_ps(iz1,jz2);
 255             dx20             = _mm_sub_ps(ix2,jx0);
 256             dy20             = _mm_sub_ps(iy2,jy0);
 257             dz20             = _mm_sub_ps(iz2,jz0);
 258             dx21             = _mm_sub_ps(ix2,jx1);
 259             dy21             = _mm_sub_ps(iy2,jy1);
 260             dz21             = _mm_sub_ps(iz2,jz1);
 261             dx22             = _mm_sub_ps(ix2,jx2);
 262             dy22             = _mm_sub_ps(iy2,jy2);
 263             dz22             = _mm_sub_ps(iz2,jz2);
 264
 265             /* Calculate squared distance and things based on it */
 266             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 267             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 268             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 269             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 270             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 271             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 272             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 273             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 274             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 275
 276             rinv00           = avx128fma_invsqrt_f(rsq00);
 277             rinv01           = avx128fma_invsqrt_f(rsq01);
 278             rinv02           = avx128fma_invsqrt_f(rsq02);
 279             rinv10           = avx128fma_invsqrt_f(rsq10);
 280             rinv11           = avx128fma_invsqrt_f(rsq11);
 281             rinv12           = avx128fma_invsqrt_f(rsq12);
 282             rinv20           = avx128fma_invsqrt_f(rsq20);
 283             rinv21           = avx128fma_invsqrt_f(rsq21);
 284             rinv22           = avx128fma_invsqrt_f(rsq22);
 285
 286             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 287             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 288             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 289             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 290             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 291             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 292             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 293             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 294             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 295
 296             fjx0             = _mm_setzero_ps();
 297             fjy0             = _mm_setzero_ps();
 298             fjz0             = _mm_setzero_ps();
 299             fjx1             = _mm_setzero_ps();
 300             fjy1             = _mm_setzero_ps();
 301             fjz1             = _mm_setzero_ps();
 302             fjx2             = _mm_setzero_ps();
 303             fjy2             = _mm_setzero_ps();
 304             fjz2             = _mm_setzero_ps();
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             if (gmx_mm_any_lt(rsq00,rcutoff2))
 311             {
 312
 313             r00              = _mm_mul_ps(rsq00,rinv00);
 314
 315             /* Calculate table index by multiplying r with table scale and truncate to integer */
 316             rt               = _mm_mul_ps(r00,vftabscale);
 317             vfitab           = _mm_cvttps_epi32(rt);
 318 #ifdef __XOP__
 319             vfeps            = _mm_frcz_ps(rt);
 320 #else
 321             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 322 #endif
 323             twovfeps         = _mm_add_ps(vfeps,vfeps);
 324             vfitab           = _mm_slli_epi32(vfitab,3);
 325
 326             /* REACTION-FIELD ELECTROSTATICS */
 327             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 328             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 329
 330             /* CUBIC SPLINE TABLE DISPERSION */
 331             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 332             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 333             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 334             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 335             _MM_TRANSPOSE4_PS(Y,F,G,H);
 336             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 337             VV               = _mm_macc_ps(vfeps,Fp,Y);
 338             vvdw6            = _mm_mul_ps(c6_00,VV);
 339             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 340             fvdw6            = _mm_mul_ps(c6_00,FF);
 341
 342             /* CUBIC SPLINE TABLE REPULSION */
 343             vfitab           = _mm_add_epi32(vfitab,ifour);
 344             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 345             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 346             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 347             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 348             _MM_TRANSPOSE4_PS(Y,F,G,H);
 349             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 350             VV               = _mm_macc_ps(vfeps,Fp,Y);
 351             vvdw12           = _mm_mul_ps(c12_00,VV);
 352             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 353             fvdw12           = _mm_mul_ps(c12_00,FF);
 354             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 355             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 356
 357             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velec            = _mm_and_ps(velec,cutoff_mask);
 361             velecsum         = _mm_add_ps(velecsum,velec);
 362             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 363             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 364
 365             fscal            = _mm_add_ps(felec,fvdw);
 366
 367             fscal            = _mm_and_ps(fscal,cutoff_mask);
 368
 369              /* Update vectorial force */
 370             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 371             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 372             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 373
 374             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 375             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 376             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 377
 378             }
 379
 380             /**************************
 381              * CALCULATE INTERACTIONS *
 382              **************************/
 383
 384             if (gmx_mm_any_lt(rsq01,rcutoff2))
 385             {
 386
 387             /* REACTION-FIELD ELECTROSTATICS */
 388             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 389             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 390
 391             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 392
 393             /* Update potential sum for this i atom from the interaction with this j atom. */
 394             velec            = _mm_and_ps(velec,cutoff_mask);
 395             velecsum         = _mm_add_ps(velecsum,velec);
 396
 397             fscal            = felec;
 398
 399             fscal            = _mm_and_ps(fscal,cutoff_mask);
 400
 401              /* Update vectorial force */
 402             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 403             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 404             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 405
 406             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 407             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 408             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 409
 410             }
 411
 412             /**************************
 413              * CALCULATE INTERACTIONS *
 414              **************************/
 415
 416             if (gmx_mm_any_lt(rsq02,rcutoff2))
 417             {
 418
 419             /* REACTION-FIELD ELECTROSTATICS */
 420             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 421             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 422
 423             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 424
 425             /* Update potential sum for this i atom from the interaction with this j atom. */
 426             velec            = _mm_and_ps(velec,cutoff_mask);
 427             velecsum         = _mm_add_ps(velecsum,velec);
 428
 429             fscal            = felec;
 430
 431             fscal            = _mm_and_ps(fscal,cutoff_mask);
 432
 433              /* Update vectorial force */
 434             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 435             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 436             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 437
 438             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 439             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 440             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 441
 442             }
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             if (gmx_mm_any_lt(rsq10,rcutoff2))
 449             {
 450
 451             /* REACTION-FIELD ELECTROSTATICS */
 452             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 453             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 454
 455             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velec            = _mm_and_ps(velec,cutoff_mask);
 459             velecsum         = _mm_add_ps(velecsum,velec);
 460
 461             fscal            = felec;
 462
 463             fscal            = _mm_and_ps(fscal,cutoff_mask);
 464
 465              /* Update vectorial force */
 466             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 467             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 468             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 469
 470             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 471             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 472             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 473
 474             }
 475
 476             /**************************
 477              * CALCULATE INTERACTIONS *
 478              **************************/
 479
 480             if (gmx_mm_any_lt(rsq11,rcutoff2))
 481             {
 482
 483             /* REACTION-FIELD ELECTROSTATICS */
 484             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 485             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 486
 487             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 488
 489             /* Update potential sum for this i atom from the interaction with this j atom. */
 490             velec            = _mm_and_ps(velec,cutoff_mask);
 491             velecsum         = _mm_add_ps(velecsum,velec);
 492
 493             fscal            = felec;
 494
 495             fscal            = _mm_and_ps(fscal,cutoff_mask);
 496
 497              /* Update vectorial force */
 498             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 499             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 500             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 501
 502             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 503             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 504             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 505
 506             }
 507
 508             /**************************
 509              * CALCULATE INTERACTIONS *
 510              **************************/
 511
 512             if (gmx_mm_any_lt(rsq12,rcutoff2))
 513             {
 514
 515             /* REACTION-FIELD ELECTROSTATICS */
 516             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 517             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 518
 519             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 520
 521             /* Update potential sum for this i atom from the interaction with this j atom. */
 522             velec            = _mm_and_ps(velec,cutoff_mask);
 523             velecsum         = _mm_add_ps(velecsum,velec);
 524
 525             fscal            = felec;
 526
 527             fscal            = _mm_and_ps(fscal,cutoff_mask);
 528
 529              /* Update vectorial force */
 530             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 531             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 532             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 533
 534             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 535             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 536             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 537
 538             }
 539
 540             /**************************
 541              * CALCULATE INTERACTIONS *
 542              **************************/
 543
 544             if (gmx_mm_any_lt(rsq20,rcutoff2))
 545             {
 546
 547             /* REACTION-FIELD ELECTROSTATICS */
 548             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 549             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 550
 551             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 552
 553             /* Update potential sum for this i atom from the interaction with this j atom. */
 554             velec            = _mm_and_ps(velec,cutoff_mask);
 555             velecsum         = _mm_add_ps(velecsum,velec);
 556
 557             fscal            = felec;
 558
 559             fscal            = _mm_and_ps(fscal,cutoff_mask);
 560
 561              /* Update vectorial force */
 562             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 563             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 564             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 565
 566             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 567             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 568             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 569
 570             }
 571
 572             /**************************
 573              * CALCULATE INTERACTIONS *
 574              **************************/
 575
 576             if (gmx_mm_any_lt(rsq21,rcutoff2))
 577             {
 578
 579             /* REACTION-FIELD ELECTROSTATICS */
 580             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 581             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 582
 583             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 584
 585             /* Update potential sum for this i atom from the interaction with this j atom. */
 586             velec            = _mm_and_ps(velec,cutoff_mask);
 587             velecsum         = _mm_add_ps(velecsum,velec);
 588
 589             fscal            = felec;
 590
 591             fscal            = _mm_and_ps(fscal,cutoff_mask);
 592
 593              /* Update vectorial force */
 594             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 595             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 596             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 597
 598             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 599             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 600             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 601
 602             }
 603
 604             /**************************
 605              * CALCULATE INTERACTIONS *
 606              **************************/
 607
 608             if (gmx_mm_any_lt(rsq22,rcutoff2))
 609             {
 610
 611             /* REACTION-FIELD ELECTROSTATICS */
 612             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 613             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 614
 615             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 616
 617             /* Update potential sum for this i atom from the interaction with this j atom. */
 618             velec            = _mm_and_ps(velec,cutoff_mask);
 619             velecsum         = _mm_add_ps(velecsum,velec);
 620
 621             fscal            = felec;
 622
 623             fscal            = _mm_and_ps(fscal,cutoff_mask);
 624
 625              /* Update vectorial force */
 626             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 627             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 628             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 629
 630             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 631             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 632             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 633
 634             }
 635
 636             fjptrA             = f+j_coord_offsetA;
 637             fjptrB             = f+j_coord_offsetB;
 638             fjptrC             = f+j_coord_offsetC;
 639             fjptrD             = f+j_coord_offsetD;
 640
 641             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 642                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 643
 644             /* Inner loop uses 387 flops */
 645         }
 646
 647         if(jidx<j_index_end)
 648         {
 649
 650             /* Get j neighbor index, and coordinate index */
 651             jnrlistA         = jjnr[jidx];
 652             jnrlistB         = jjnr[jidx+1];
 653             jnrlistC         = jjnr[jidx+2];
 654             jnrlistD         = jjnr[jidx+3];
 655             /* Sign of each element will be negative for non-real atoms.
 656              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 657              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 658              */
 659             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 660             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 661             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 662             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 663             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 664             j_coord_offsetA  = DIM*jnrA;
 665             j_coord_offsetB  = DIM*jnrB;
 666             j_coord_offsetC  = DIM*jnrC;
 667             j_coord_offsetD  = DIM*jnrD;
 668
 669             /* load j atom coordinates */
 670             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 671                                               x+j_coord_offsetC,x+j_coord_offsetD,
 672                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 673
 674             /* Calculate displacement vector */
 675             dx00             = _mm_sub_ps(ix0,jx0);
 676             dy00             = _mm_sub_ps(iy0,jy0);
 677             dz00             = _mm_sub_ps(iz0,jz0);
 678             dx01             = _mm_sub_ps(ix0,jx1);
 679             dy01             = _mm_sub_ps(iy0,jy1);
 680             dz01             = _mm_sub_ps(iz0,jz1);
 681             dx02             = _mm_sub_ps(ix0,jx2);
 682             dy02             = _mm_sub_ps(iy0,jy2);
 683             dz02             = _mm_sub_ps(iz0,jz2);
 684             dx10             = _mm_sub_ps(ix1,jx0);
 685             dy10             = _mm_sub_ps(iy1,jy0);
 686             dz10             = _mm_sub_ps(iz1,jz0);
 687             dx11             = _mm_sub_ps(ix1,jx1);
 688             dy11             = _mm_sub_ps(iy1,jy1);
 689             dz11             = _mm_sub_ps(iz1,jz1);
 690             dx12             = _mm_sub_ps(ix1,jx2);
 691             dy12             = _mm_sub_ps(iy1,jy2);
 692             dz12             = _mm_sub_ps(iz1,jz2);
 693             dx20             = _mm_sub_ps(ix2,jx0);
 694             dy20             = _mm_sub_ps(iy2,jy0);
 695             dz20             = _mm_sub_ps(iz2,jz0);
 696             dx21             = _mm_sub_ps(ix2,jx1);
 697             dy21             = _mm_sub_ps(iy2,jy1);
 698             dz21             = _mm_sub_ps(iz2,jz1);
 699             dx22             = _mm_sub_ps(ix2,jx2);
 700             dy22             = _mm_sub_ps(iy2,jy2);
 701             dz22             = _mm_sub_ps(iz2,jz2);
 702
 703             /* Calculate squared distance and things based on it */
 704             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 705             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 706             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 707             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 708             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 709             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 710             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 711             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 712             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 713
 714             rinv00           = avx128fma_invsqrt_f(rsq00);
 715             rinv01           = avx128fma_invsqrt_f(rsq01);
 716             rinv02           = avx128fma_invsqrt_f(rsq02);
 717             rinv10           = avx128fma_invsqrt_f(rsq10);
 718             rinv11           = avx128fma_invsqrt_f(rsq11);
 719             rinv12           = avx128fma_invsqrt_f(rsq12);
 720             rinv20           = avx128fma_invsqrt_f(rsq20);
 721             rinv21           = avx128fma_invsqrt_f(rsq21);
 722             rinv22           = avx128fma_invsqrt_f(rsq22);
 723
 724             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 725             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 726             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 727             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 728             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 729             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 730             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 731             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 732             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 733
 734             fjx0             = _mm_setzero_ps();
 735             fjy0             = _mm_setzero_ps();
 736             fjz0             = _mm_setzero_ps();
 737             fjx1             = _mm_setzero_ps();
 738             fjy1             = _mm_setzero_ps();
 739             fjz1             = _mm_setzero_ps();
 740             fjx2             = _mm_setzero_ps();
 741             fjy2             = _mm_setzero_ps();
 742             fjz2             = _mm_setzero_ps();
 743
 744             /**************************
 745              * CALCULATE INTERACTIONS *
 746              **************************/
 747
 748             if (gmx_mm_any_lt(rsq00,rcutoff2))
 749             {
 750
 751             r00              = _mm_mul_ps(rsq00,rinv00);
 752             r00              = _mm_andnot_ps(dummy_mask,r00);
 753
 754             /* Calculate table index by multiplying r with table scale and truncate to integer */
 755             rt               = _mm_mul_ps(r00,vftabscale);
 756             vfitab           = _mm_cvttps_epi32(rt);
 757 #ifdef __XOP__
 758             vfeps            = _mm_frcz_ps(rt);
 759 #else
 760             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 761 #endif
 762             twovfeps         = _mm_add_ps(vfeps,vfeps);
 763             vfitab           = _mm_slli_epi32(vfitab,3);
 764
 765             /* REACTION-FIELD ELECTROSTATICS */
 766             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 767             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 768
 769             /* CUBIC SPLINE TABLE DISPERSION */
 770             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 771             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 772             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 773             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 774             _MM_TRANSPOSE4_PS(Y,F,G,H);
 775             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 776             VV               = _mm_macc_ps(vfeps,Fp,Y);
 777             vvdw6            = _mm_mul_ps(c6_00,VV);
 778             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 779             fvdw6            = _mm_mul_ps(c6_00,FF);
 780
 781             /* CUBIC SPLINE TABLE REPULSION */
 782             vfitab           = _mm_add_epi32(vfitab,ifour);
 783             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 784             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 785             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 786             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 787             _MM_TRANSPOSE4_PS(Y,F,G,H);
 788             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 789             VV               = _mm_macc_ps(vfeps,Fp,Y);
 790             vvdw12           = _mm_mul_ps(c12_00,VV);
 791             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 792             fvdw12           = _mm_mul_ps(c12_00,FF);
 793             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 794             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 795
 796             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 797
 798             /* Update potential sum for this i atom from the interaction with this j atom. */
 799             velec            = _mm_and_ps(velec,cutoff_mask);
 800             velec            = _mm_andnot_ps(dummy_mask,velec);
 801             velecsum         = _mm_add_ps(velecsum,velec);
 802             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 803             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 804             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 805
 806             fscal            = _mm_add_ps(felec,fvdw);
 807
 808             fscal            = _mm_and_ps(fscal,cutoff_mask);
 809
 810             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 811
 812              /* Update vectorial force */
 813             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 814             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 815             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 816
 817             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 818             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 819             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 820
 821             }
 822
 823             /**************************
 824              * CALCULATE INTERACTIONS *
 825              **************************/
 826
 827             if (gmx_mm_any_lt(rsq01,rcutoff2))
 828             {
 829
 830             /* REACTION-FIELD ELECTROSTATICS */
 831             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 832             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 833
 834             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 835
 836             /* Update potential sum for this i atom from the interaction with this j atom. */
 837             velec            = _mm_and_ps(velec,cutoff_mask);
 838             velec            = _mm_andnot_ps(dummy_mask,velec);
 839             velecsum         = _mm_add_ps(velecsum,velec);
 840
 841             fscal            = felec;
 842
 843             fscal            = _mm_and_ps(fscal,cutoff_mask);
 844
 845             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 846
 847              /* Update vectorial force */
 848             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 849             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 850             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 851
 852             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 853             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 854             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 855
 856             }
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             if (gmx_mm_any_lt(rsq02,rcutoff2))
 863             {
 864
 865             /* REACTION-FIELD ELECTROSTATICS */
 866             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 867             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 868
 869             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 870
 871             /* Update potential sum for this i atom from the interaction with this j atom. */
 872             velec            = _mm_and_ps(velec,cutoff_mask);
 873             velec            = _mm_andnot_ps(dummy_mask,velec);
 874             velecsum         = _mm_add_ps(velecsum,velec);
 875
 876             fscal            = felec;
 877
 878             fscal            = _mm_and_ps(fscal,cutoff_mask);
 879
 880             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 881
 882              /* Update vectorial force */
 883             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 884             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 885             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 886
 887             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 888             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 889             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 890
 891             }
 892
 893             /**************************
 894              * CALCULATE INTERACTIONS *
 895              **************************/
 896
 897             if (gmx_mm_any_lt(rsq10,rcutoff2))
 898             {
 899
 900             /* REACTION-FIELD ELECTROSTATICS */
 901             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 902             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 903
 904             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 905
 906             /* Update potential sum for this i atom from the interaction with this j atom. */
 907             velec            = _mm_and_ps(velec,cutoff_mask);
 908             velec            = _mm_andnot_ps(dummy_mask,velec);
 909             velecsum         = _mm_add_ps(velecsum,velec);
 910
 911             fscal            = felec;
 912
 913             fscal            = _mm_and_ps(fscal,cutoff_mask);
 914
 915             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 916
 917              /* Update vectorial force */
 918             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 919             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 920             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 921
 922             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 923             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 924             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 925
 926             }
 927
 928             /**************************
 929              * CALCULATE INTERACTIONS *
 930              **************************/
 931
 932             if (gmx_mm_any_lt(rsq11,rcutoff2))
 933             {
 934
 935             /* REACTION-FIELD ELECTROSTATICS */
 936             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 937             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 938
 939             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 940
 941             /* Update potential sum for this i atom from the interaction with this j atom. */
 942             velec            = _mm_and_ps(velec,cutoff_mask);
 943             velec            = _mm_andnot_ps(dummy_mask,velec);
 944             velecsum         = _mm_add_ps(velecsum,velec);
 945
 946             fscal            = felec;
 947
 948             fscal            = _mm_and_ps(fscal,cutoff_mask);
 949
 950             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 951
 952              /* Update vectorial force */
 953             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 954             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 955             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 956
 957             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 958             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 959             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 960
 961             }
 962
 963             /**************************
 964              * CALCULATE INTERACTIONS *
 965              **************************/
 966
 967             if (gmx_mm_any_lt(rsq12,rcutoff2))
 968             {
 969
 970             /* REACTION-FIELD ELECTROSTATICS */
 971             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 972             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 973
 974             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 975
 976             /* Update potential sum for this i atom from the interaction with this j atom. */
 977             velec            = _mm_and_ps(velec,cutoff_mask);
 978             velec            = _mm_andnot_ps(dummy_mask,velec);
 979             velecsum         = _mm_add_ps(velecsum,velec);
 980
 981             fscal            = felec;
 982
 983             fscal            = _mm_and_ps(fscal,cutoff_mask);
 984
 985             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 986
 987              /* Update vectorial force */
 988             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 989             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 990             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 991
 992             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 993             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 994             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 995
 996             }
 997
 998             /**************************
 999              * CALCULATE INTERACTIONS *
1000              **************************/
1001
1002             if (gmx_mm_any_lt(rsq20,rcutoff2))
1003             {
1004
1005             /* REACTION-FIELD ELECTROSTATICS */
1006             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
1007             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1008
1009             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1010
1011             /* Update potential sum for this i atom from the interaction with this j atom. */
1012             velec            = _mm_and_ps(velec,cutoff_mask);
1013             velec            = _mm_andnot_ps(dummy_mask,velec);
1014             velecsum         = _mm_add_ps(velecsum,velec);
1015
1016             fscal            = felec;
1017
1018             fscal            = _mm_and_ps(fscal,cutoff_mask);
1019
1020             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1021
1022              /* Update vectorial force */
1023             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1024             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1025             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1026
1027             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1028             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1029             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1030
1031             }
1032
1033             /**************************
1034              * CALCULATE INTERACTIONS *
1035              **************************/
1036
1037             if (gmx_mm_any_lt(rsq21,rcutoff2))
1038             {
1039
1040             /* REACTION-FIELD ELECTROSTATICS */
1041             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
1042             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1043
1044             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1045
1046             /* Update potential sum for this i atom from the interaction with this j atom. */
1047             velec            = _mm_and_ps(velec,cutoff_mask);
1048             velec            = _mm_andnot_ps(dummy_mask,velec);
1049             velecsum         = _mm_add_ps(velecsum,velec);
1050
1051             fscal            = felec;
1052
1053             fscal            = _mm_and_ps(fscal,cutoff_mask);
1054
1055             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1056
1057              /* Update vectorial force */
1058             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1059             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1060             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1061
1062             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1063             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1064             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1065
1066             }
1067
1068             /**************************
1069              * CALCULATE INTERACTIONS *
1070              **************************/
1071
1072             if (gmx_mm_any_lt(rsq22,rcutoff2))
1073             {
1074
1075             /* REACTION-FIELD ELECTROSTATICS */
1076             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
1077             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1078
1079             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1080
1081             /* Update potential sum for this i atom from the interaction with this j atom. */
1082             velec            = _mm_and_ps(velec,cutoff_mask);
1083             velec            = _mm_andnot_ps(dummy_mask,velec);
1084             velecsum         = _mm_add_ps(velecsum,velec);
1085
1086             fscal            = felec;
1087
1088             fscal            = _mm_and_ps(fscal,cutoff_mask);
1089
1090             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1091
1092              /* Update vectorial force */
1093             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1094             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1095             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1096
1097             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1098             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1099             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1100
1101             }
1102
1103             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1104             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1105             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1106             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1107
1108             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1109                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1110
1111             /* Inner loop uses 388 flops */
1112         }
1113
1114         /* End of innermost loop */
1115
1116         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1117                                               f+i_coord_offset,fshift+i_shift_offset);
1118
1119         ggid                        = gid[iidx];
1120         /* Update potential energies */
1121         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1122         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1123
1124         /* Increment number of inner iterations */
1125         inneriter                  += j_index_end - j_index_start;
1126
1127         /* Outer loop uses 20 flops */
1128     }
1129
1130     /* Increment number of outer iterations */
1131     outeriter        += nri;
1132
1133     /* Update outer/inner flops */
1134
1135     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*388);
1136 }
1137 /*
1138  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_avx_128_fma_single
1139  * Electrostatics interaction: ReactionField
1140  * VdW interaction:            CubicSplineTable
1141  * Geometry:                   Water3-Water3
1142  * Calculate force/pot:        Force
1143  */
1144 void
1145 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_avx_128_fma_single
1146                     (t_nblist                    * gmx_restrict       nlist,
1147                      rvec                        * gmx_restrict          xx,
1148                      rvec                        * gmx_restrict          ff,
1149                      struct t_forcerec           * gmx_restrict          fr,
1150                      t_mdatoms                   * gmx_restrict     mdatoms,
1151                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1152                      t_nrnb                      * gmx_restrict        nrnb)
1153 {
1154     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1155      * just 0 for non-waters.
1156      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1157      * jnr indices corresponding to data put in the four positions in the SIMD register.
1158      */
1159     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1160     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1161     int              jnrA,jnrB,jnrC,jnrD;
1162     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1163     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1164     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1165     real             rcutoff_scalar;
1166     real             *shiftvec,*fshift,*x,*f;
1167     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1168     real             scratch[4*DIM];
1169     __m128           fscal,rcutoff,rcutoff2,jidxall;
1170     int              vdwioffset0;
1171     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1172     int              vdwioffset1;
1173     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1174     int              vdwioffset2;
1175     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1176     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1177     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1178     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1179     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1180     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1181     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1182     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1183     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1184     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1185     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1186     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1187     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1188     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1189     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1190     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1191     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1192     real             *charge;
1193     int              nvdwtype;
1194     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1195     int              *vdwtype;
1196     real             *vdwparam;
1197     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1198     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1199     __m128i          vfitab;
1200     __m128i          ifour       = _mm_set1_epi32(4);
1201     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1202     real             *vftab;
1203     __m128           dummy_mask,cutoff_mask;
1204     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1205     __m128           one     = _mm_set1_ps(1.0);
1206     __m128           two     = _mm_set1_ps(2.0);
1207     x                = xx[0];
1208     f                = ff[0];
1209
1210     nri              = nlist->nri;
1211     iinr             = nlist->iinr;
1212     jindex           = nlist->jindex;
1213     jjnr             = nlist->jjnr;
1214     shiftidx         = nlist->shift;
1215     gid              = nlist->gid;
1216     shiftvec         = fr->shift_vec[0];
1217     fshift           = fr->fshift[0];
1218     facel            = _mm_set1_ps(fr->ic->epsfac);
1219     charge           = mdatoms->chargeA;
1220     krf              = _mm_set1_ps(fr->ic->k_rf);
1221     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1222     crf              = _mm_set1_ps(fr->ic->c_rf);
1223     nvdwtype         = fr->ntype;
1224     vdwparam         = fr->nbfp;
1225     vdwtype          = mdatoms->typeA;
1226
1227     vftab            = kernel_data->table_vdw->data;
1228     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1229
1230     /* Setup water-specific parameters */
1231     inr              = nlist->iinr[0];
1232     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1233     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1234     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1235     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1236
1237     jq0              = _mm_set1_ps(charge[inr+0]);
1238     jq1              = _mm_set1_ps(charge[inr+1]);
1239     jq2              = _mm_set1_ps(charge[inr+2]);
1240     vdwjidx0A        = 2*vdwtype[inr+0];
1241     qq00             = _mm_mul_ps(iq0,jq0);
1242     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1243     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1244     qq01             = _mm_mul_ps(iq0,jq1);
1245     qq02             = _mm_mul_ps(iq0,jq2);
1246     qq10             = _mm_mul_ps(iq1,jq0);
1247     qq11             = _mm_mul_ps(iq1,jq1);
1248     qq12             = _mm_mul_ps(iq1,jq2);
1249     qq20             = _mm_mul_ps(iq2,jq0);
1250     qq21             = _mm_mul_ps(iq2,jq1);
1251     qq22             = _mm_mul_ps(iq2,jq2);
1252
1253     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1254     rcutoff_scalar   = fr->ic->rcoulomb;
1255     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1256     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1257
1258     /* Avoid stupid compiler warnings */
1259     jnrA = jnrB = jnrC = jnrD = 0;
1260     j_coord_offsetA = 0;
1261     j_coord_offsetB = 0;
1262     j_coord_offsetC = 0;
1263     j_coord_offsetD = 0;
1264
1265     outeriter        = 0;
1266     inneriter        = 0;
1267
1268     for(iidx=0;iidx<4*DIM;iidx++)
1269     {
1270         scratch[iidx] = 0.0;
1271     }
1272
1273     /* Start outer loop over neighborlists */
1274     for(iidx=0; iidx<nri; iidx++)
1275     {
1276         /* Load shift vector for this list */
1277         i_shift_offset   = DIM*shiftidx[iidx];
1278
1279         /* Load limits for loop over neighbors */
1280         j_index_start    = jindex[iidx];
1281         j_index_end      = jindex[iidx+1];
1282
1283         /* Get outer coordinate index */
1284         inr              = iinr[iidx];
1285         i_coord_offset   = DIM*inr;
1286
1287         /* Load i particle coords and add shift vector */
1288         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1289                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1290
1291         fix0             = _mm_setzero_ps();
1292         fiy0             = _mm_setzero_ps();
1293         fiz0             = _mm_setzero_ps();
1294         fix1             = _mm_setzero_ps();
1295         fiy1             = _mm_setzero_ps();
1296         fiz1             = _mm_setzero_ps();
1297         fix2             = _mm_setzero_ps();
1298         fiy2             = _mm_setzero_ps();
1299         fiz2             = _mm_setzero_ps();
1300
1301         /* Start inner kernel loop */
1302         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1303         {
1304
1305             /* Get j neighbor index, and coordinate index */
1306             jnrA             = jjnr[jidx];
1307             jnrB             = jjnr[jidx+1];
1308             jnrC             = jjnr[jidx+2];
1309             jnrD             = jjnr[jidx+3];
1310             j_coord_offsetA  = DIM*jnrA;
1311             j_coord_offsetB  = DIM*jnrB;
1312             j_coord_offsetC  = DIM*jnrC;
1313             j_coord_offsetD  = DIM*jnrD;
1314
1315             /* load j atom coordinates */
1316             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1317                                               x+j_coord_offsetC,x+j_coord_offsetD,
1318                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1319
1320             /* Calculate displacement vector */
1321             dx00             = _mm_sub_ps(ix0,jx0);
1322             dy00             = _mm_sub_ps(iy0,jy0);
1323             dz00             = _mm_sub_ps(iz0,jz0);
1324             dx01             = _mm_sub_ps(ix0,jx1);
1325             dy01             = _mm_sub_ps(iy0,jy1);
1326             dz01             = _mm_sub_ps(iz0,jz1);
1327             dx02             = _mm_sub_ps(ix0,jx2);
1328             dy02             = _mm_sub_ps(iy0,jy2);
1329             dz02             = _mm_sub_ps(iz0,jz2);
1330             dx10             = _mm_sub_ps(ix1,jx0);
1331             dy10             = _mm_sub_ps(iy1,jy0);
1332             dz10             = _mm_sub_ps(iz1,jz0);
1333             dx11             = _mm_sub_ps(ix1,jx1);
1334             dy11             = _mm_sub_ps(iy1,jy1);
1335             dz11             = _mm_sub_ps(iz1,jz1);
1336             dx12             = _mm_sub_ps(ix1,jx2);
1337             dy12             = _mm_sub_ps(iy1,jy2);
1338             dz12             = _mm_sub_ps(iz1,jz2);
1339             dx20             = _mm_sub_ps(ix2,jx0);
1340             dy20             = _mm_sub_ps(iy2,jy0);
1341             dz20             = _mm_sub_ps(iz2,jz0);
1342             dx21             = _mm_sub_ps(ix2,jx1);
1343             dy21             = _mm_sub_ps(iy2,jy1);
1344             dz21             = _mm_sub_ps(iz2,jz1);
1345             dx22             = _mm_sub_ps(ix2,jx2);
1346             dy22             = _mm_sub_ps(iy2,jy2);
1347             dz22             = _mm_sub_ps(iz2,jz2);
1348
1349             /* Calculate squared distance and things based on it */
1350             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1351             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1352             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1353             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1354             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1355             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1356             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1357             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1358             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1359
1360             rinv00           = avx128fma_invsqrt_f(rsq00);
1361             rinv01           = avx128fma_invsqrt_f(rsq01);
1362             rinv02           = avx128fma_invsqrt_f(rsq02);
1363             rinv10           = avx128fma_invsqrt_f(rsq10);
1364             rinv11           = avx128fma_invsqrt_f(rsq11);
1365             rinv12           = avx128fma_invsqrt_f(rsq12);
1366             rinv20           = avx128fma_invsqrt_f(rsq20);
1367             rinv21           = avx128fma_invsqrt_f(rsq21);
1368             rinv22           = avx128fma_invsqrt_f(rsq22);
1369
1370             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1371             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1372             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1373             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1374             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1375             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1376             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1377             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1378             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1379
1380             fjx0             = _mm_setzero_ps();
1381             fjy0             = _mm_setzero_ps();
1382             fjz0             = _mm_setzero_ps();
1383             fjx1             = _mm_setzero_ps();
1384             fjy1             = _mm_setzero_ps();
1385             fjz1             = _mm_setzero_ps();
1386             fjx2             = _mm_setzero_ps();
1387             fjy2             = _mm_setzero_ps();
1388             fjz2             = _mm_setzero_ps();
1389
1390             /**************************
1391              * CALCULATE INTERACTIONS *
1392              **************************/
1393
1394             if (gmx_mm_any_lt(rsq00,rcutoff2))
1395             {
1396
1397             r00              = _mm_mul_ps(rsq00,rinv00);
1398
1399             /* Calculate table index by multiplying r with table scale and truncate to integer */
1400             rt               = _mm_mul_ps(r00,vftabscale);
1401             vfitab           = _mm_cvttps_epi32(rt);
1402 #ifdef __XOP__
1403             vfeps            = _mm_frcz_ps(rt);
1404 #else
1405             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1406 #endif
1407             twovfeps         = _mm_add_ps(vfeps,vfeps);
1408             vfitab           = _mm_slli_epi32(vfitab,3);
1409
1410             /* REACTION-FIELD ELECTROSTATICS */
1411             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1412
1413             /* CUBIC SPLINE TABLE DISPERSION */
1414             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1415             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1416             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1417             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1418             _MM_TRANSPOSE4_PS(Y,F,G,H);
1419             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1420             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1421             fvdw6            = _mm_mul_ps(c6_00,FF);
1422
1423             /* CUBIC SPLINE TABLE REPULSION */
1424             vfitab           = _mm_add_epi32(vfitab,ifour);
1425             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1426             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1427             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1428             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1429             _MM_TRANSPOSE4_PS(Y,F,G,H);
1430             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1431             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1432             fvdw12           = _mm_mul_ps(c12_00,FF);
1433             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1434
1435             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1436
1437             fscal            = _mm_add_ps(felec,fvdw);
1438
1439             fscal            = _mm_and_ps(fscal,cutoff_mask);
1440
1441              /* Update vectorial force */
1442             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1443             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1444             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1445
1446             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1447             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1448             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1449
1450             }
1451
1452             /**************************
1453              * CALCULATE INTERACTIONS *
1454              **************************/
1455
1456             if (gmx_mm_any_lt(rsq01,rcutoff2))
1457             {
1458
1459             /* REACTION-FIELD ELECTROSTATICS */
1460             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1461
1462             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1463
1464             fscal            = felec;
1465
1466             fscal            = _mm_and_ps(fscal,cutoff_mask);
1467
1468              /* Update vectorial force */
1469             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1470             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1471             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1472
1473             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1474             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1475             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1476
1477             }
1478
1479             /**************************
1480              * CALCULATE INTERACTIONS *
1481              **************************/
1482
1483             if (gmx_mm_any_lt(rsq02,rcutoff2))
1484             {
1485
1486             /* REACTION-FIELD ELECTROSTATICS */
1487             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1488
1489             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1490
1491             fscal            = felec;
1492
1493             fscal            = _mm_and_ps(fscal,cutoff_mask);
1494
1495              /* Update vectorial force */
1496             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1497             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1498             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1499
1500             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1501             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1502             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1503
1504             }
1505
1506             /**************************
1507              * CALCULATE INTERACTIONS *
1508              **************************/
1509
1510             if (gmx_mm_any_lt(rsq10,rcutoff2))
1511             {
1512
1513             /* REACTION-FIELD ELECTROSTATICS */
1514             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1515
1516             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1517
1518             fscal            = felec;
1519
1520             fscal            = _mm_and_ps(fscal,cutoff_mask);
1521
1522              /* Update vectorial force */
1523             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1524             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1525             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1526
1527             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1528             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1529             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1530
1531             }
1532
1533             /**************************
1534              * CALCULATE INTERACTIONS *
1535              **************************/
1536
1537             if (gmx_mm_any_lt(rsq11,rcutoff2))
1538             {
1539
1540             /* REACTION-FIELD ELECTROSTATICS */
1541             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1542
1543             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1544
1545             fscal            = felec;
1546
1547             fscal            = _mm_and_ps(fscal,cutoff_mask);
1548
1549              /* Update vectorial force */
1550             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1551             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1552             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1553
1554             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1555             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1556             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1557
1558             }
1559
1560             /**************************
1561              * CALCULATE INTERACTIONS *
1562              **************************/
1563
1564             if (gmx_mm_any_lt(rsq12,rcutoff2))
1565             {
1566
1567             /* REACTION-FIELD ELECTROSTATICS */
1568             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1569
1570             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1571
1572             fscal            = felec;
1573
1574             fscal            = _mm_and_ps(fscal,cutoff_mask);
1575
1576              /* Update vectorial force */
1577             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1578             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1579             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1580
1581             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1582             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1583             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1584
1585             }
1586
1587             /**************************
1588              * CALCULATE INTERACTIONS *
1589              **************************/
1590
1591             if (gmx_mm_any_lt(rsq20,rcutoff2))
1592             {
1593
1594             /* REACTION-FIELD ELECTROSTATICS */
1595             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1596
1597             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1598
1599             fscal            = felec;
1600
1601             fscal            = _mm_and_ps(fscal,cutoff_mask);
1602
1603              /* Update vectorial force */
1604             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1605             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1606             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1607
1608             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1609             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1610             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1611
1612             }
1613
1614             /**************************
1615              * CALCULATE INTERACTIONS *
1616              **************************/
1617
1618             if (gmx_mm_any_lt(rsq21,rcutoff2))
1619             {
1620
1621             /* REACTION-FIELD ELECTROSTATICS */
1622             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1623
1624             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1625
1626             fscal            = felec;
1627
1628             fscal            = _mm_and_ps(fscal,cutoff_mask);
1629
1630              /* Update vectorial force */
1631             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1632             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1633             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1634
1635             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1636             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1637             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1638
1639             }
1640
1641             /**************************
1642              * CALCULATE INTERACTIONS *
1643              **************************/
1644
1645             if (gmx_mm_any_lt(rsq22,rcutoff2))
1646             {
1647
1648             /* REACTION-FIELD ELECTROSTATICS */
1649             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1650
1651             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1652
1653             fscal            = felec;
1654
1655             fscal            = _mm_and_ps(fscal,cutoff_mask);
1656
1657              /* Update vectorial force */
1658             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1659             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1660             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1661
1662             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1663             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1664             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1665
1666             }
1667
1668             fjptrA             = f+j_coord_offsetA;
1669             fjptrB             = f+j_coord_offsetB;
1670             fjptrC             = f+j_coord_offsetC;
1671             fjptrD             = f+j_coord_offsetD;
1672
1673             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1674                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1675
1676             /* Inner loop uses 324 flops */
1677         }
1678
1679         if(jidx<j_index_end)
1680         {
1681
1682             /* Get j neighbor index, and coordinate index */
1683             jnrlistA         = jjnr[jidx];
1684             jnrlistB         = jjnr[jidx+1];
1685             jnrlistC         = jjnr[jidx+2];
1686             jnrlistD         = jjnr[jidx+3];
1687             /* Sign of each element will be negative for non-real atoms.
1688              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1689              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1690              */
1691             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1692             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1693             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1694             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1695             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1696             j_coord_offsetA  = DIM*jnrA;
1697             j_coord_offsetB  = DIM*jnrB;
1698             j_coord_offsetC  = DIM*jnrC;
1699             j_coord_offsetD  = DIM*jnrD;
1700
1701             /* load j atom coordinates */
1702             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1703                                               x+j_coord_offsetC,x+j_coord_offsetD,
1704                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1705
1706             /* Calculate displacement vector */
1707             dx00             = _mm_sub_ps(ix0,jx0);
1708             dy00             = _mm_sub_ps(iy0,jy0);
1709             dz00             = _mm_sub_ps(iz0,jz0);
1710             dx01             = _mm_sub_ps(ix0,jx1);
1711             dy01             = _mm_sub_ps(iy0,jy1);
1712             dz01             = _mm_sub_ps(iz0,jz1);
1713             dx02             = _mm_sub_ps(ix0,jx2);
1714             dy02             = _mm_sub_ps(iy0,jy2);
1715             dz02             = _mm_sub_ps(iz0,jz2);
1716             dx10             = _mm_sub_ps(ix1,jx0);
1717             dy10             = _mm_sub_ps(iy1,jy0);
1718             dz10             = _mm_sub_ps(iz1,jz0);
1719             dx11             = _mm_sub_ps(ix1,jx1);
1720             dy11             = _mm_sub_ps(iy1,jy1);
1721             dz11             = _mm_sub_ps(iz1,jz1);
1722             dx12             = _mm_sub_ps(ix1,jx2);
1723             dy12             = _mm_sub_ps(iy1,jy2);
1724             dz12             = _mm_sub_ps(iz1,jz2);
1725             dx20             = _mm_sub_ps(ix2,jx0);
1726             dy20             = _mm_sub_ps(iy2,jy0);
1727             dz20             = _mm_sub_ps(iz2,jz0);
1728             dx21             = _mm_sub_ps(ix2,jx1);
1729             dy21             = _mm_sub_ps(iy2,jy1);
1730             dz21             = _mm_sub_ps(iz2,jz1);
1731             dx22             = _mm_sub_ps(ix2,jx2);
1732             dy22             = _mm_sub_ps(iy2,jy2);
1733             dz22             = _mm_sub_ps(iz2,jz2);
1734
1735             /* Calculate squared distance and things based on it */
1736             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1737             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1738             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1739             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1740             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1741             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1742             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1743             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1744             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1745
1746             rinv00           = avx128fma_invsqrt_f(rsq00);
1747             rinv01           = avx128fma_invsqrt_f(rsq01);
1748             rinv02           = avx128fma_invsqrt_f(rsq02);
1749             rinv10           = avx128fma_invsqrt_f(rsq10);
1750             rinv11           = avx128fma_invsqrt_f(rsq11);
1751             rinv12           = avx128fma_invsqrt_f(rsq12);
1752             rinv20           = avx128fma_invsqrt_f(rsq20);
1753             rinv21           = avx128fma_invsqrt_f(rsq21);
1754             rinv22           = avx128fma_invsqrt_f(rsq22);
1755
1756             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1757             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1758             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1759             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1760             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1761             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1762             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1763             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1764             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1765
1766             fjx0             = _mm_setzero_ps();
1767             fjy0             = _mm_setzero_ps();
1768             fjz0             = _mm_setzero_ps();
1769             fjx1             = _mm_setzero_ps();
1770             fjy1             = _mm_setzero_ps();
1771             fjz1             = _mm_setzero_ps();
1772             fjx2             = _mm_setzero_ps();
1773             fjy2             = _mm_setzero_ps();
1774             fjz2             = _mm_setzero_ps();
1775
1776             /**************************
1777              * CALCULATE INTERACTIONS *
1778              **************************/
1779
1780             if (gmx_mm_any_lt(rsq00,rcutoff2))
1781             {
1782
1783             r00              = _mm_mul_ps(rsq00,rinv00);
1784             r00              = _mm_andnot_ps(dummy_mask,r00);
1785
1786             /* Calculate table index by multiplying r with table scale and truncate to integer */
1787             rt               = _mm_mul_ps(r00,vftabscale);
1788             vfitab           = _mm_cvttps_epi32(rt);
1789 #ifdef __XOP__
1790             vfeps            = _mm_frcz_ps(rt);
1791 #else
1792             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1793 #endif
1794             twovfeps         = _mm_add_ps(vfeps,vfeps);
1795             vfitab           = _mm_slli_epi32(vfitab,3);
1796
1797             /* REACTION-FIELD ELECTROSTATICS */
1798             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1799
1800             /* CUBIC SPLINE TABLE DISPERSION */
1801             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1802             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1803             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1804             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1805             _MM_TRANSPOSE4_PS(Y,F,G,H);
1806             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1807             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1808             fvdw6            = _mm_mul_ps(c6_00,FF);
1809
1810             /* CUBIC SPLINE TABLE REPULSION */
1811             vfitab           = _mm_add_epi32(vfitab,ifour);
1812             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1813             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1814             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1815             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1816             _MM_TRANSPOSE4_PS(Y,F,G,H);
1817             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1818             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1819             fvdw12           = _mm_mul_ps(c12_00,FF);
1820             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1821
1822             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1823
1824             fscal            = _mm_add_ps(felec,fvdw);
1825
1826             fscal            = _mm_and_ps(fscal,cutoff_mask);
1827
1828             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1829
1830              /* Update vectorial force */
1831             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1832             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1833             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1834
1835             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1836             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1837             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1838
1839             }
1840
1841             /**************************
1842              * CALCULATE INTERACTIONS *
1843              **************************/
1844
1845             if (gmx_mm_any_lt(rsq01,rcutoff2))
1846             {
1847
1848             /* REACTION-FIELD ELECTROSTATICS */
1849             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1850
1851             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1852
1853             fscal            = felec;
1854
1855             fscal            = _mm_and_ps(fscal,cutoff_mask);
1856
1857             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1858
1859              /* Update vectorial force */
1860             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1861             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1862             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1863
1864             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1865             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1866             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1867
1868             }
1869
1870             /**************************
1871              * CALCULATE INTERACTIONS *
1872              **************************/
1873
1874             if (gmx_mm_any_lt(rsq02,rcutoff2))
1875             {
1876
1877             /* REACTION-FIELD ELECTROSTATICS */
1878             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1879
1880             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1881
1882             fscal            = felec;
1883
1884             fscal            = _mm_and_ps(fscal,cutoff_mask);
1885
1886             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1887
1888              /* Update vectorial force */
1889             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1890             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1891             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1892
1893             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1894             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1895             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1896
1897             }
1898
1899             /**************************
1900              * CALCULATE INTERACTIONS *
1901              **************************/
1902
1903             if (gmx_mm_any_lt(rsq10,rcutoff2))
1904             {
1905
1906             /* REACTION-FIELD ELECTROSTATICS */
1907             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1908
1909             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1910
1911             fscal            = felec;
1912
1913             fscal            = _mm_and_ps(fscal,cutoff_mask);
1914
1915             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1916
1917              /* Update vectorial force */
1918             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1919             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1920             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1921
1922             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1923             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1924             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1925
1926             }
1927
1928             /**************************
1929              * CALCULATE INTERACTIONS *
1930              **************************/
1931
1932             if (gmx_mm_any_lt(rsq11,rcutoff2))
1933             {
1934
1935             /* REACTION-FIELD ELECTROSTATICS */
1936             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1937
1938             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1939
1940             fscal            = felec;
1941
1942             fscal            = _mm_and_ps(fscal,cutoff_mask);
1943
1944             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1945
1946              /* Update vectorial force */
1947             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1948             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1949             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1950
1951             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1952             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1953             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1954
1955             }
1956
1957             /**************************
1958              * CALCULATE INTERACTIONS *
1959              **************************/
1960
1961             if (gmx_mm_any_lt(rsq12,rcutoff2))
1962             {
1963
1964             /* REACTION-FIELD ELECTROSTATICS */
1965             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1966
1967             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1968
1969             fscal            = felec;
1970
1971             fscal            = _mm_and_ps(fscal,cutoff_mask);
1972
1973             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1974
1975              /* Update vectorial force */
1976             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1977             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1978             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1979
1980             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1981             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1982             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1983
1984             }
1985
1986             /**************************
1987              * CALCULATE INTERACTIONS *
1988              **************************/
1989
1990             if (gmx_mm_any_lt(rsq20,rcutoff2))
1991             {
1992
1993             /* REACTION-FIELD ELECTROSTATICS */
1994             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1995
1996             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1997
1998             fscal            = felec;
1999
2000             fscal            = _mm_and_ps(fscal,cutoff_mask);
2001
2002             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2003
2004              /* Update vectorial force */
2005             fix2             = _mm_macc_ps(dx20,fscal,fix2);
2006             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
2007             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
2008
2009             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
2010             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
2011             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
2012
2013             }
2014
2015             /**************************
2016              * CALCULATE INTERACTIONS *
2017              **************************/
2018
2019             if (gmx_mm_any_lt(rsq21,rcutoff2))
2020             {
2021
2022             /* REACTION-FIELD ELECTROSTATICS */
2023             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
2024
2025             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2026
2027             fscal            = felec;
2028
2029             fscal            = _mm_and_ps(fscal,cutoff_mask);
2030
2031             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2032
2033              /* Update vectorial force */
2034             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2035             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2036             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2037
2038             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2039             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2040             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2041
2042             }
2043
2044             /**************************
2045              * CALCULATE INTERACTIONS *
2046              **************************/
2047
2048             if (gmx_mm_any_lt(rsq22,rcutoff2))
2049             {
2050
2051             /* REACTION-FIELD ELECTROSTATICS */
2052             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
2053
2054             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2055
2056             fscal            = felec;
2057
2058             fscal            = _mm_and_ps(fscal,cutoff_mask);
2059
2060             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2061
2062              /* Update vectorial force */
2063             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2064             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2065             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2066
2067             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2068             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2069             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2070
2071             }
2072
2073             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2074             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2075             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2076             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2077
2078             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2079                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2080
2081             /* Inner loop uses 325 flops */
2082         }
2083
2084         /* End of innermost loop */
2085
2086         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2087                                               f+i_coord_offset,fshift+i_shift_offset);
2088
2089         /* Increment number of inner iterations */
2090         inneriter                  += j_index_end - j_index_start;
2091
2092         /* Outer loop uses 18 flops */
2093     }
2094
2095     /* Increment number of outer iterations */
2096     outeriter        += nri;
2097
2098     /* Update outer/inner flops */
2099
2100     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*325);
2101 }