src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_VdwNone_GeomW3W3_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_single
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            None
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwNone_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     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122
 123     /* Setup water-specific parameters */
 124     inr              = nlist->iinr[0];
 125     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 126     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 127     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 128
 129     jq0              = _mm_set1_ps(charge[inr+0]);
 130     jq1              = _mm_set1_ps(charge[inr+1]);
 131     jq2              = _mm_set1_ps(charge[inr+2]);
 132     qq00             = _mm_mul_ps(iq0,jq0);
 133     qq01             = _mm_mul_ps(iq0,jq1);
 134     qq02             = _mm_mul_ps(iq0,jq2);
 135     qq10             = _mm_mul_ps(iq1,jq0);
 136     qq11             = _mm_mul_ps(iq1,jq1);
 137     qq12             = _mm_mul_ps(iq1,jq2);
 138     qq20             = _mm_mul_ps(iq2,jq0);
 139     qq21             = _mm_mul_ps(iq2,jq1);
 140     qq22             = _mm_mul_ps(iq2,jq2);
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = jnrC = jnrD = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146     j_coord_offsetC = 0;
 147     j_coord_offsetD = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     for(iidx=0;iidx<4*DIM;iidx++)
 153     {
 154         scratch[iidx] = 0.0;
 155     }
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 173                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 174
 175         fix0             = _mm_setzero_ps();
 176         fiy0             = _mm_setzero_ps();
 177         fiz0             = _mm_setzero_ps();
 178         fix1             = _mm_setzero_ps();
 179         fiy1             = _mm_setzero_ps();
 180         fiz1             = _mm_setzero_ps();
 181         fix2             = _mm_setzero_ps();
 182         fiy2             = _mm_setzero_ps();
 183         fiz2             = _mm_setzero_ps();
 184
 185         /* Reset potential sums */
 186         velecsum         = _mm_setzero_ps();
 187
 188         /* Start inner kernel loop */
 189         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 190         {
 191
 192             /* Get j neighbor index, and coordinate index */
 193             jnrA             = jjnr[jidx];
 194             jnrB             = jjnr[jidx+1];
 195             jnrC             = jjnr[jidx+2];
 196             jnrD             = jjnr[jidx+3];
 197             j_coord_offsetA  = DIM*jnrA;
 198             j_coord_offsetB  = DIM*jnrB;
 199             j_coord_offsetC  = DIM*jnrC;
 200             j_coord_offsetD  = DIM*jnrD;
 201
 202             /* load j atom coordinates */
 203             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 204                                               x+j_coord_offsetC,x+j_coord_offsetD,
 205                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 206
 207             /* Calculate displacement vector */
 208             dx00             = _mm_sub_ps(ix0,jx0);
 209             dy00             = _mm_sub_ps(iy0,jy0);
 210             dz00             = _mm_sub_ps(iz0,jz0);
 211             dx01             = _mm_sub_ps(ix0,jx1);
 212             dy01             = _mm_sub_ps(iy0,jy1);
 213             dz01             = _mm_sub_ps(iz0,jz1);
 214             dx02             = _mm_sub_ps(ix0,jx2);
 215             dy02             = _mm_sub_ps(iy0,jy2);
 216             dz02             = _mm_sub_ps(iz0,jz2);
 217             dx10             = _mm_sub_ps(ix1,jx0);
 218             dy10             = _mm_sub_ps(iy1,jy0);
 219             dz10             = _mm_sub_ps(iz1,jz0);
 220             dx11             = _mm_sub_ps(ix1,jx1);
 221             dy11             = _mm_sub_ps(iy1,jy1);
 222             dz11             = _mm_sub_ps(iz1,jz1);
 223             dx12             = _mm_sub_ps(ix1,jx2);
 224             dy12             = _mm_sub_ps(iy1,jy2);
 225             dz12             = _mm_sub_ps(iz1,jz2);
 226             dx20             = _mm_sub_ps(ix2,jx0);
 227             dy20             = _mm_sub_ps(iy2,jy0);
 228             dz20             = _mm_sub_ps(iz2,jz0);
 229             dx21             = _mm_sub_ps(ix2,jx1);
 230             dy21             = _mm_sub_ps(iy2,jy1);
 231             dz21             = _mm_sub_ps(iz2,jz1);
 232             dx22             = _mm_sub_ps(ix2,jx2);
 233             dy22             = _mm_sub_ps(iy2,jy2);
 234             dz22             = _mm_sub_ps(iz2,jz2);
 235
 236             /* Calculate squared distance and things based on it */
 237             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 238             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 239             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 240             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 241             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 242             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 243             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 244             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 245             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 246
 247             rinv00           = avx128fma_invsqrt_f(rsq00);
 248             rinv01           = avx128fma_invsqrt_f(rsq01);
 249             rinv02           = avx128fma_invsqrt_f(rsq02);
 250             rinv10           = avx128fma_invsqrt_f(rsq10);
 251             rinv11           = avx128fma_invsqrt_f(rsq11);
 252             rinv12           = avx128fma_invsqrt_f(rsq12);
 253             rinv20           = avx128fma_invsqrt_f(rsq20);
 254             rinv21           = avx128fma_invsqrt_f(rsq21);
 255             rinv22           = avx128fma_invsqrt_f(rsq22);
 256
 257             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 258             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 259             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 260             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 261             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 262             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 263             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 264             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 265             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 266
 267             fjx0             = _mm_setzero_ps();
 268             fjy0             = _mm_setzero_ps();
 269             fjz0             = _mm_setzero_ps();
 270             fjx1             = _mm_setzero_ps();
 271             fjy1             = _mm_setzero_ps();
 272             fjz1             = _mm_setzero_ps();
 273             fjx2             = _mm_setzero_ps();
 274             fjy2             = _mm_setzero_ps();
 275             fjz2             = _mm_setzero_ps();
 276
 277             /**************************
 278              * CALCULATE INTERACTIONS *
 279              **************************/
 280
 281             /* COULOMB ELECTROSTATICS */
 282             velec            = _mm_mul_ps(qq00,rinv00);
 283             felec            = _mm_mul_ps(velec,rinvsq00);
 284
 285             /* Update potential sum for this i atom from the interaction with this j atom. */
 286             velecsum         = _mm_add_ps(velecsum,velec);
 287
 288             fscal            = felec;
 289
 290              /* Update vectorial force */
 291             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 292             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 293             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 294
 295             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 296             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 297             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 298
 299             /**************************
 300              * CALCULATE INTERACTIONS *
 301              **************************/
 302
 303             /* COULOMB ELECTROSTATICS */
 304             velec            = _mm_mul_ps(qq01,rinv01);
 305             felec            = _mm_mul_ps(velec,rinvsq01);
 306
 307             /* Update potential sum for this i atom from the interaction with this j atom. */
 308             velecsum         = _mm_add_ps(velecsum,velec);
 309
 310             fscal            = felec;
 311
 312              /* Update vectorial force */
 313             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 314             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 315             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 316
 317             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 318             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 319             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             /* COULOMB ELECTROSTATICS */
 326             velec            = _mm_mul_ps(qq02,rinv02);
 327             felec            = _mm_mul_ps(velec,rinvsq02);
 328
 329             /* Update potential sum for this i atom from the interaction with this j atom. */
 330             velecsum         = _mm_add_ps(velecsum,velec);
 331
 332             fscal            = felec;
 333
 334              /* Update vectorial force */
 335             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 336             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 337             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 338
 339             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 340             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 341             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 342
 343             /**************************
 344              * CALCULATE INTERACTIONS *
 345              **************************/
 346
 347             /* COULOMB ELECTROSTATICS */
 348             velec            = _mm_mul_ps(qq10,rinv10);
 349             felec            = _mm_mul_ps(velec,rinvsq10);
 350
 351             /* Update potential sum for this i atom from the interaction with this j atom. */
 352             velecsum         = _mm_add_ps(velecsum,velec);
 353
 354             fscal            = felec;
 355
 356              /* Update vectorial force */
 357             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 358             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 359             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 360
 361             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 362             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 363             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             /* COULOMB ELECTROSTATICS */
 370             velec            = _mm_mul_ps(qq11,rinv11);
 371             felec            = _mm_mul_ps(velec,rinvsq11);
 372
 373             /* Update potential sum for this i atom from the interaction with this j atom. */
 374             velecsum         = _mm_add_ps(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378              /* Update vectorial force */
 379             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 380             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 381             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 382
 383             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 384             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 385             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 386
 387             /**************************
 388              * CALCULATE INTERACTIONS *
 389              **************************/
 390
 391             /* COULOMB ELECTROSTATICS */
 392             velec            = _mm_mul_ps(qq12,rinv12);
 393             felec            = _mm_mul_ps(velec,rinvsq12);
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velecsum         = _mm_add_ps(velecsum,velec);
 397
 398             fscal            = felec;
 399
 400              /* Update vectorial force */
 401             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 402             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 403             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 404
 405             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 406             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 407             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 408
 409             /**************************
 410              * CALCULATE INTERACTIONS *
 411              **************************/
 412
 413             /* COULOMB ELECTROSTATICS */
 414             velec            = _mm_mul_ps(qq20,rinv20);
 415             felec            = _mm_mul_ps(velec,rinvsq20);
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velecsum         = _mm_add_ps(velecsum,velec);
 419
 420             fscal            = felec;
 421
 422              /* Update vectorial force */
 423             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 424             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 425             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 426
 427             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 428             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 429             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             /* COULOMB ELECTROSTATICS */
 436             velec            = _mm_mul_ps(qq21,rinv21);
 437             felec            = _mm_mul_ps(velec,rinvsq21);
 438
 439             /* Update potential sum for this i atom from the interaction with this j atom. */
 440             velecsum         = _mm_add_ps(velecsum,velec);
 441
 442             fscal            = felec;
 443
 444              /* Update vectorial force */
 445             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 446             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 447             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 448
 449             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 450             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 451             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 452
 453             /**************************
 454              * CALCULATE INTERACTIONS *
 455              **************************/
 456
 457             /* COULOMB ELECTROSTATICS */
 458             velec            = _mm_mul_ps(qq22,rinv22);
 459             felec            = _mm_mul_ps(velec,rinvsq22);
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             velecsum         = _mm_add_ps(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466              /* Update vectorial force */
 467             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 468             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 469             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 470
 471             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 472             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 473             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 474
 475             fjptrA             = f+j_coord_offsetA;
 476             fjptrB             = f+j_coord_offsetB;
 477             fjptrC             = f+j_coord_offsetC;
 478             fjptrD             = f+j_coord_offsetD;
 479
 480             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 481                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 482
 483             /* Inner loop uses 279 flops */
 484         }
 485
 486         if(jidx<j_index_end)
 487         {
 488
 489             /* Get j neighbor index, and coordinate index */
 490             jnrlistA         = jjnr[jidx];
 491             jnrlistB         = jjnr[jidx+1];
 492             jnrlistC         = jjnr[jidx+2];
 493             jnrlistD         = jjnr[jidx+3];
 494             /* Sign of each element will be negative for non-real atoms.
 495              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 496              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 497              */
 498             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 499             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 500             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 501             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 502             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 503             j_coord_offsetA  = DIM*jnrA;
 504             j_coord_offsetB  = DIM*jnrB;
 505             j_coord_offsetC  = DIM*jnrC;
 506             j_coord_offsetD  = DIM*jnrD;
 507
 508             /* load j atom coordinates */
 509             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 510                                               x+j_coord_offsetC,x+j_coord_offsetD,
 511                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 512
 513             /* Calculate displacement vector */
 514             dx00             = _mm_sub_ps(ix0,jx0);
 515             dy00             = _mm_sub_ps(iy0,jy0);
 516             dz00             = _mm_sub_ps(iz0,jz0);
 517             dx01             = _mm_sub_ps(ix0,jx1);
 518             dy01             = _mm_sub_ps(iy0,jy1);
 519             dz01             = _mm_sub_ps(iz0,jz1);
 520             dx02             = _mm_sub_ps(ix0,jx2);
 521             dy02             = _mm_sub_ps(iy0,jy2);
 522             dz02             = _mm_sub_ps(iz0,jz2);
 523             dx10             = _mm_sub_ps(ix1,jx0);
 524             dy10             = _mm_sub_ps(iy1,jy0);
 525             dz10             = _mm_sub_ps(iz1,jz0);
 526             dx11             = _mm_sub_ps(ix1,jx1);
 527             dy11             = _mm_sub_ps(iy1,jy1);
 528             dz11             = _mm_sub_ps(iz1,jz1);
 529             dx12             = _mm_sub_ps(ix1,jx2);
 530             dy12             = _mm_sub_ps(iy1,jy2);
 531             dz12             = _mm_sub_ps(iz1,jz2);
 532             dx20             = _mm_sub_ps(ix2,jx0);
 533             dy20             = _mm_sub_ps(iy2,jy0);
 534             dz20             = _mm_sub_ps(iz2,jz0);
 535             dx21             = _mm_sub_ps(ix2,jx1);
 536             dy21             = _mm_sub_ps(iy2,jy1);
 537             dz21             = _mm_sub_ps(iz2,jz1);
 538             dx22             = _mm_sub_ps(ix2,jx2);
 539             dy22             = _mm_sub_ps(iy2,jy2);
 540             dz22             = _mm_sub_ps(iz2,jz2);
 541
 542             /* Calculate squared distance and things based on it */
 543             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 544             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 545             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 546             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 547             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 548             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 549             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 550             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 551             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 552
 553             rinv00           = avx128fma_invsqrt_f(rsq00);
 554             rinv01           = avx128fma_invsqrt_f(rsq01);
 555             rinv02           = avx128fma_invsqrt_f(rsq02);
 556             rinv10           = avx128fma_invsqrt_f(rsq10);
 557             rinv11           = avx128fma_invsqrt_f(rsq11);
 558             rinv12           = avx128fma_invsqrt_f(rsq12);
 559             rinv20           = avx128fma_invsqrt_f(rsq20);
 560             rinv21           = avx128fma_invsqrt_f(rsq21);
 561             rinv22           = avx128fma_invsqrt_f(rsq22);
 562
 563             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 564             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 565             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 566             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 567             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 568             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 569             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 570             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 571             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 572
 573             fjx0             = _mm_setzero_ps();
 574             fjy0             = _mm_setzero_ps();
 575             fjz0             = _mm_setzero_ps();
 576             fjx1             = _mm_setzero_ps();
 577             fjy1             = _mm_setzero_ps();
 578             fjz1             = _mm_setzero_ps();
 579             fjx2             = _mm_setzero_ps();
 580             fjy2             = _mm_setzero_ps();
 581             fjz2             = _mm_setzero_ps();
 582
 583             /**************************
 584              * CALCULATE INTERACTIONS *
 585              **************************/
 586
 587             /* COULOMB ELECTROSTATICS */
 588             velec            = _mm_mul_ps(qq00,rinv00);
 589             felec            = _mm_mul_ps(velec,rinvsq00);
 590
 591             /* Update potential sum for this i atom from the interaction with this j atom. */
 592             velec            = _mm_andnot_ps(dummy_mask,velec);
 593             velecsum         = _mm_add_ps(velecsum,velec);
 594
 595             fscal            = felec;
 596
 597             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 598
 599              /* Update vectorial force */
 600             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 601             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 602             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 603
 604             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 605             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 606             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 607
 608             /**************************
 609              * CALCULATE INTERACTIONS *
 610              **************************/
 611
 612             /* COULOMB ELECTROSTATICS */
 613             velec            = _mm_mul_ps(qq01,rinv01);
 614             felec            = _mm_mul_ps(velec,rinvsq01);
 615
 616             /* Update potential sum for this i atom from the interaction with this j atom. */
 617             velec            = _mm_andnot_ps(dummy_mask,velec);
 618             velecsum         = _mm_add_ps(velecsum,velec);
 619
 620             fscal            = felec;
 621
 622             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 623
 624              /* Update vectorial force */
 625             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 626             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 627             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 628
 629             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 630             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 631             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 632
 633             /**************************
 634              * CALCULATE INTERACTIONS *
 635              **************************/
 636
 637             /* COULOMB ELECTROSTATICS */
 638             velec            = _mm_mul_ps(qq02,rinv02);
 639             felec            = _mm_mul_ps(velec,rinvsq02);
 640
 641             /* Update potential sum for this i atom from the interaction with this j atom. */
 642             velec            = _mm_andnot_ps(dummy_mask,velec);
 643             velecsum         = _mm_add_ps(velecsum,velec);
 644
 645             fscal            = felec;
 646
 647             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 648
 649              /* Update vectorial force */
 650             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 651             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 652             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 653
 654             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 655             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 656             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 657
 658             /**************************
 659              * CALCULATE INTERACTIONS *
 660              **************************/
 661
 662             /* COULOMB ELECTROSTATICS */
 663             velec            = _mm_mul_ps(qq10,rinv10);
 664             felec            = _mm_mul_ps(velec,rinvsq10);
 665
 666             /* Update potential sum for this i atom from the interaction with this j atom. */
 667             velec            = _mm_andnot_ps(dummy_mask,velec);
 668             velecsum         = _mm_add_ps(velecsum,velec);
 669
 670             fscal            = felec;
 671
 672             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 673
 674              /* Update vectorial force */
 675             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 676             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 677             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 678
 679             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 680             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 681             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 682
 683             /**************************
 684              * CALCULATE INTERACTIONS *
 685              **************************/
 686
 687             /* COULOMB ELECTROSTATICS */
 688             velec            = _mm_mul_ps(qq11,rinv11);
 689             felec            = _mm_mul_ps(velec,rinvsq11);
 690
 691             /* Update potential sum for this i atom from the interaction with this j atom. */
 692             velec            = _mm_andnot_ps(dummy_mask,velec);
 693             velecsum         = _mm_add_ps(velecsum,velec);
 694
 695             fscal            = felec;
 696
 697             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 698
 699              /* Update vectorial force */
 700             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 701             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 702             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 703
 704             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 705             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 706             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 707
 708             /**************************
 709              * CALCULATE INTERACTIONS *
 710              **************************/
 711
 712             /* COULOMB ELECTROSTATICS */
 713             velec            = _mm_mul_ps(qq12,rinv12);
 714             felec            = _mm_mul_ps(velec,rinvsq12);
 715
 716             /* Update potential sum for this i atom from the interaction with this j atom. */
 717             velec            = _mm_andnot_ps(dummy_mask,velec);
 718             velecsum         = _mm_add_ps(velecsum,velec);
 719
 720             fscal            = felec;
 721
 722             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 723
 724              /* Update vectorial force */
 725             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 726             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 727             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 728
 729             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 730             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 731             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 732
 733             /**************************
 734              * CALCULATE INTERACTIONS *
 735              **************************/
 736
 737             /* COULOMB ELECTROSTATICS */
 738             velec            = _mm_mul_ps(qq20,rinv20);
 739             felec            = _mm_mul_ps(velec,rinvsq20);
 740
 741             /* Update potential sum for this i atom from the interaction with this j atom. */
 742             velec            = _mm_andnot_ps(dummy_mask,velec);
 743             velecsum         = _mm_add_ps(velecsum,velec);
 744
 745             fscal            = felec;
 746
 747             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 748
 749              /* Update vectorial force */
 750             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 751             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 752             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 753
 754             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 755             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 756             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             /* COULOMB ELECTROSTATICS */
 763             velec            = _mm_mul_ps(qq21,rinv21);
 764             felec            = _mm_mul_ps(velec,rinvsq21);
 765
 766             /* Update potential sum for this i atom from the interaction with this j atom. */
 767             velec            = _mm_andnot_ps(dummy_mask,velec);
 768             velecsum         = _mm_add_ps(velecsum,velec);
 769
 770             fscal            = felec;
 771
 772             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 773
 774              /* Update vectorial force */
 775             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 776             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 777             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 778
 779             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 780             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 781             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 782
 783             /**************************
 784              * CALCULATE INTERACTIONS *
 785              **************************/
 786
 787             /* COULOMB ELECTROSTATICS */
 788             velec            = _mm_mul_ps(qq22,rinv22);
 789             felec            = _mm_mul_ps(velec,rinvsq22);
 790
 791             /* Update potential sum for this i atom from the interaction with this j atom. */
 792             velec            = _mm_andnot_ps(dummy_mask,velec);
 793             velecsum         = _mm_add_ps(velecsum,velec);
 794
 795             fscal            = felec;
 796
 797             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 798
 799              /* Update vectorial force */
 800             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 801             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 802             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 803
 804             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 805             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 806             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 807
 808             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 809             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 810             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 811             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 812
 813             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 814                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 815
 816             /* Inner loop uses 279 flops */
 817         }
 818
 819         /* End of innermost loop */
 820
 821         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 822                                               f+i_coord_offset,fshift+i_shift_offset);
 823
 824         ggid                        = gid[iidx];
 825         /* Update potential energies */
 826         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 827
 828         /* Increment number of inner iterations */
 829         inneriter                  += j_index_end - j_index_start;
 830
 831         /* Outer loop uses 19 flops */
 832     }
 833
 834     /* Increment number of outer iterations */
 835     outeriter        += nri;
 836
 837     /* Update outer/inner flops */
 838
 839     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*279);
 840 }
 841 /*
 842  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_single
 843  * Electrostatics interaction: Coulomb
 844  * VdW interaction:            None
 845  * Geometry:                   Water3-Water3
 846  * Calculate force/pot:        Force
 847  */
 848 void
 849 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_single
 850                     (t_nblist                    * gmx_restrict       nlist,
 851                      rvec                        * gmx_restrict          xx,
 852                      rvec                        * gmx_restrict          ff,
 853                      struct t_forcerec           * gmx_restrict          fr,
 854                      t_mdatoms                   * gmx_restrict     mdatoms,
 855                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 856                      t_nrnb                      * gmx_restrict        nrnb)
 857 {
 858     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 859      * just 0 for non-waters.
 860      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 861      * jnr indices corresponding to data put in the four positions in the SIMD register.
 862      */
 863     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 864     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 865     int              jnrA,jnrB,jnrC,jnrD;
 866     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 867     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 868     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 869     real             rcutoff_scalar;
 870     real             *shiftvec,*fshift,*x,*f;
 871     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 872     real             scratch[4*DIM];
 873     __m128           fscal,rcutoff,rcutoff2,jidxall;
 874     int              vdwioffset0;
 875     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 876     int              vdwioffset1;
 877     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 878     int              vdwioffset2;
 879     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 880     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 881     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 882     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 883     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 884     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 885     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 886     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 887     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 888     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 889     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 890     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 891     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 892     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 893     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 894     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 895     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 896     real             *charge;
 897     __m128           dummy_mask,cutoff_mask;
 898     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 899     __m128           one     = _mm_set1_ps(1.0);
 900     __m128           two     = _mm_set1_ps(2.0);
 901     x                = xx[0];
 902     f                = ff[0];
 903
 904     nri              = nlist->nri;
 905     iinr             = nlist->iinr;
 906     jindex           = nlist->jindex;
 907     jjnr             = nlist->jjnr;
 908     shiftidx         = nlist->shift;
 909     gid              = nlist->gid;
 910     shiftvec         = fr->shift_vec[0];
 911     fshift           = fr->fshift[0];
 912     facel            = _mm_set1_ps(fr->ic->epsfac);
 913     charge           = mdatoms->chargeA;
 914
 915     /* Setup water-specific parameters */
 916     inr              = nlist->iinr[0];
 917     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 918     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 919     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 920
 921     jq0              = _mm_set1_ps(charge[inr+0]);
 922     jq1              = _mm_set1_ps(charge[inr+1]);
 923     jq2              = _mm_set1_ps(charge[inr+2]);
 924     qq00             = _mm_mul_ps(iq0,jq0);
 925     qq01             = _mm_mul_ps(iq0,jq1);
 926     qq02             = _mm_mul_ps(iq0,jq2);
 927     qq10             = _mm_mul_ps(iq1,jq0);
 928     qq11             = _mm_mul_ps(iq1,jq1);
 929     qq12             = _mm_mul_ps(iq1,jq2);
 930     qq20             = _mm_mul_ps(iq2,jq0);
 931     qq21             = _mm_mul_ps(iq2,jq1);
 932     qq22             = _mm_mul_ps(iq2,jq2);
 933
 934     /* Avoid stupid compiler warnings */
 935     jnrA = jnrB = jnrC = jnrD = 0;
 936     j_coord_offsetA = 0;
 937     j_coord_offsetB = 0;
 938     j_coord_offsetC = 0;
 939     j_coord_offsetD = 0;
 940
 941     outeriter        = 0;
 942     inneriter        = 0;
 943
 944     for(iidx=0;iidx<4*DIM;iidx++)
 945     {
 946         scratch[iidx] = 0.0;
 947     }
 948
 949     /* Start outer loop over neighborlists */
 950     for(iidx=0; iidx<nri; iidx++)
 951     {
 952         /* Load shift vector for this list */
 953         i_shift_offset   = DIM*shiftidx[iidx];
 954
 955         /* Load limits for loop over neighbors */
 956         j_index_start    = jindex[iidx];
 957         j_index_end      = jindex[iidx+1];
 958
 959         /* Get outer coordinate index */
 960         inr              = iinr[iidx];
 961         i_coord_offset   = DIM*inr;
 962
 963         /* Load i particle coords and add shift vector */
 964         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 965                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 966
 967         fix0             = _mm_setzero_ps();
 968         fiy0             = _mm_setzero_ps();
 969         fiz0             = _mm_setzero_ps();
 970         fix1             = _mm_setzero_ps();
 971         fiy1             = _mm_setzero_ps();
 972         fiz1             = _mm_setzero_ps();
 973         fix2             = _mm_setzero_ps();
 974         fiy2             = _mm_setzero_ps();
 975         fiz2             = _mm_setzero_ps();
 976
 977         /* Start inner kernel loop */
 978         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 979         {
 980
 981             /* Get j neighbor index, and coordinate index */
 982             jnrA             = jjnr[jidx];
 983             jnrB             = jjnr[jidx+1];
 984             jnrC             = jjnr[jidx+2];
 985             jnrD             = jjnr[jidx+3];
 986             j_coord_offsetA  = DIM*jnrA;
 987             j_coord_offsetB  = DIM*jnrB;
 988             j_coord_offsetC  = DIM*jnrC;
 989             j_coord_offsetD  = DIM*jnrD;
 990
 991             /* load j atom coordinates */
 992             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 993                                               x+j_coord_offsetC,x+j_coord_offsetD,
 994                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 995
 996             /* Calculate displacement vector */
 997             dx00             = _mm_sub_ps(ix0,jx0);
 998             dy00             = _mm_sub_ps(iy0,jy0);
 999             dz00             = _mm_sub_ps(iz0,jz0);
1000             dx01             = _mm_sub_ps(ix0,jx1);
1001             dy01             = _mm_sub_ps(iy0,jy1);
1002             dz01             = _mm_sub_ps(iz0,jz1);
1003             dx02             = _mm_sub_ps(ix0,jx2);
1004             dy02             = _mm_sub_ps(iy0,jy2);
1005             dz02             = _mm_sub_ps(iz0,jz2);
1006             dx10             = _mm_sub_ps(ix1,jx0);
1007             dy10             = _mm_sub_ps(iy1,jy0);
1008             dz10             = _mm_sub_ps(iz1,jz0);
1009             dx11             = _mm_sub_ps(ix1,jx1);
1010             dy11             = _mm_sub_ps(iy1,jy1);
1011             dz11             = _mm_sub_ps(iz1,jz1);
1012             dx12             = _mm_sub_ps(ix1,jx2);
1013             dy12             = _mm_sub_ps(iy1,jy2);
1014             dz12             = _mm_sub_ps(iz1,jz2);
1015             dx20             = _mm_sub_ps(ix2,jx0);
1016             dy20             = _mm_sub_ps(iy2,jy0);
1017             dz20             = _mm_sub_ps(iz2,jz0);
1018             dx21             = _mm_sub_ps(ix2,jx1);
1019             dy21             = _mm_sub_ps(iy2,jy1);
1020             dz21             = _mm_sub_ps(iz2,jz1);
1021             dx22             = _mm_sub_ps(ix2,jx2);
1022             dy22             = _mm_sub_ps(iy2,jy2);
1023             dz22             = _mm_sub_ps(iz2,jz2);
1024
1025             /* Calculate squared distance and things based on it */
1026             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1027             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1028             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1029             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1030             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1031             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1032             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1033             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1034             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1035
1036             rinv00           = avx128fma_invsqrt_f(rsq00);
1037             rinv01           = avx128fma_invsqrt_f(rsq01);
1038             rinv02           = avx128fma_invsqrt_f(rsq02);
1039             rinv10           = avx128fma_invsqrt_f(rsq10);
1040             rinv11           = avx128fma_invsqrt_f(rsq11);
1041             rinv12           = avx128fma_invsqrt_f(rsq12);
1042             rinv20           = avx128fma_invsqrt_f(rsq20);
1043             rinv21           = avx128fma_invsqrt_f(rsq21);
1044             rinv22           = avx128fma_invsqrt_f(rsq22);
1045
1046             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1047             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1048             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1049             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1050             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1051             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1052             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1053             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1054             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1055
1056             fjx0             = _mm_setzero_ps();
1057             fjy0             = _mm_setzero_ps();
1058             fjz0             = _mm_setzero_ps();
1059             fjx1             = _mm_setzero_ps();
1060             fjy1             = _mm_setzero_ps();
1061             fjz1             = _mm_setzero_ps();
1062             fjx2             = _mm_setzero_ps();
1063             fjy2             = _mm_setzero_ps();
1064             fjz2             = _mm_setzero_ps();
1065
1066             /**************************
1067              * CALCULATE INTERACTIONS *
1068              **************************/
1069
1070             /* COULOMB ELECTROSTATICS */
1071             velec            = _mm_mul_ps(qq00,rinv00);
1072             felec            = _mm_mul_ps(velec,rinvsq00);
1073
1074             fscal            = felec;
1075
1076              /* Update vectorial force */
1077             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1078             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1079             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1080
1081             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1082             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1083             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1084
1085             /**************************
1086              * CALCULATE INTERACTIONS *
1087              **************************/
1088
1089             /* COULOMB ELECTROSTATICS */
1090             velec            = _mm_mul_ps(qq01,rinv01);
1091             felec            = _mm_mul_ps(velec,rinvsq01);
1092
1093             fscal            = felec;
1094
1095              /* Update vectorial force */
1096             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1097             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1098             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1099
1100             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1101             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1102             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1103
1104             /**************************
1105              * CALCULATE INTERACTIONS *
1106              **************************/
1107
1108             /* COULOMB ELECTROSTATICS */
1109             velec            = _mm_mul_ps(qq02,rinv02);
1110             felec            = _mm_mul_ps(velec,rinvsq02);
1111
1112             fscal            = felec;
1113
1114              /* Update vectorial force */
1115             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1116             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1117             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1118
1119             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1120             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1121             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1122
1123             /**************************
1124              * CALCULATE INTERACTIONS *
1125              **************************/
1126
1127             /* COULOMB ELECTROSTATICS */
1128             velec            = _mm_mul_ps(qq10,rinv10);
1129             felec            = _mm_mul_ps(velec,rinvsq10);
1130
1131             fscal            = felec;
1132
1133              /* Update vectorial force */
1134             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1135             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1136             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1137
1138             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1139             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1140             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1141
1142             /**************************
1143              * CALCULATE INTERACTIONS *
1144              **************************/
1145
1146             /* COULOMB ELECTROSTATICS */
1147             velec            = _mm_mul_ps(qq11,rinv11);
1148             felec            = _mm_mul_ps(velec,rinvsq11);
1149
1150             fscal            = felec;
1151
1152              /* Update vectorial force */
1153             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1154             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1155             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1156
1157             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1158             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1159             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1160
1161             /**************************
1162              * CALCULATE INTERACTIONS *
1163              **************************/
1164
1165             /* COULOMB ELECTROSTATICS */
1166             velec            = _mm_mul_ps(qq12,rinv12);
1167             felec            = _mm_mul_ps(velec,rinvsq12);
1168
1169             fscal            = felec;
1170
1171              /* Update vectorial force */
1172             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1173             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1174             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1175
1176             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1177             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1178             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1179
1180             /**************************
1181              * CALCULATE INTERACTIONS *
1182              **************************/
1183
1184             /* COULOMB ELECTROSTATICS */
1185             velec            = _mm_mul_ps(qq20,rinv20);
1186             felec            = _mm_mul_ps(velec,rinvsq20);
1187
1188             fscal            = felec;
1189
1190              /* Update vectorial force */
1191             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1192             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1193             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1194
1195             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1196             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1197             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1198
1199             /**************************
1200              * CALCULATE INTERACTIONS *
1201              **************************/
1202
1203             /* COULOMB ELECTROSTATICS */
1204             velec            = _mm_mul_ps(qq21,rinv21);
1205             felec            = _mm_mul_ps(velec,rinvsq21);
1206
1207             fscal            = felec;
1208
1209              /* Update vectorial force */
1210             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1211             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1212             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1213
1214             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1215             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1216             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1217
1218             /**************************
1219              * CALCULATE INTERACTIONS *
1220              **************************/
1221
1222             /* COULOMB ELECTROSTATICS */
1223             velec            = _mm_mul_ps(qq22,rinv22);
1224             felec            = _mm_mul_ps(velec,rinvsq22);
1225
1226             fscal            = felec;
1227
1228              /* Update vectorial force */
1229             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1230             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1231             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1232
1233             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1234             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1235             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1236
1237             fjptrA             = f+j_coord_offsetA;
1238             fjptrB             = f+j_coord_offsetB;
1239             fjptrC             = f+j_coord_offsetC;
1240             fjptrD             = f+j_coord_offsetD;
1241
1242             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1243                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1244
1245             /* Inner loop uses 270 flops */
1246         }
1247
1248         if(jidx<j_index_end)
1249         {
1250
1251             /* Get j neighbor index, and coordinate index */
1252             jnrlistA         = jjnr[jidx];
1253             jnrlistB         = jjnr[jidx+1];
1254             jnrlistC         = jjnr[jidx+2];
1255             jnrlistD         = jjnr[jidx+3];
1256             /* Sign of each element will be negative for non-real atoms.
1257              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1258              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1259              */
1260             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1261             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1262             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1263             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1264             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1265             j_coord_offsetA  = DIM*jnrA;
1266             j_coord_offsetB  = DIM*jnrB;
1267             j_coord_offsetC  = DIM*jnrC;
1268             j_coord_offsetD  = DIM*jnrD;
1269
1270             /* load j atom coordinates */
1271             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1272                                               x+j_coord_offsetC,x+j_coord_offsetD,
1273                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1274
1275             /* Calculate displacement vector */
1276             dx00             = _mm_sub_ps(ix0,jx0);
1277             dy00             = _mm_sub_ps(iy0,jy0);
1278             dz00             = _mm_sub_ps(iz0,jz0);
1279             dx01             = _mm_sub_ps(ix0,jx1);
1280             dy01             = _mm_sub_ps(iy0,jy1);
1281             dz01             = _mm_sub_ps(iz0,jz1);
1282             dx02             = _mm_sub_ps(ix0,jx2);
1283             dy02             = _mm_sub_ps(iy0,jy2);
1284             dz02             = _mm_sub_ps(iz0,jz2);
1285             dx10             = _mm_sub_ps(ix1,jx0);
1286             dy10             = _mm_sub_ps(iy1,jy0);
1287             dz10             = _mm_sub_ps(iz1,jz0);
1288             dx11             = _mm_sub_ps(ix1,jx1);
1289             dy11             = _mm_sub_ps(iy1,jy1);
1290             dz11             = _mm_sub_ps(iz1,jz1);
1291             dx12             = _mm_sub_ps(ix1,jx2);
1292             dy12             = _mm_sub_ps(iy1,jy2);
1293             dz12             = _mm_sub_ps(iz1,jz2);
1294             dx20             = _mm_sub_ps(ix2,jx0);
1295             dy20             = _mm_sub_ps(iy2,jy0);
1296             dz20             = _mm_sub_ps(iz2,jz0);
1297             dx21             = _mm_sub_ps(ix2,jx1);
1298             dy21             = _mm_sub_ps(iy2,jy1);
1299             dz21             = _mm_sub_ps(iz2,jz1);
1300             dx22             = _mm_sub_ps(ix2,jx2);
1301             dy22             = _mm_sub_ps(iy2,jy2);
1302             dz22             = _mm_sub_ps(iz2,jz2);
1303
1304             /* Calculate squared distance and things based on it */
1305             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1306             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1307             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1308             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1309             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1310             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1311             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1312             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1313             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1314
1315             rinv00           = avx128fma_invsqrt_f(rsq00);
1316             rinv01           = avx128fma_invsqrt_f(rsq01);
1317             rinv02           = avx128fma_invsqrt_f(rsq02);
1318             rinv10           = avx128fma_invsqrt_f(rsq10);
1319             rinv11           = avx128fma_invsqrt_f(rsq11);
1320             rinv12           = avx128fma_invsqrt_f(rsq12);
1321             rinv20           = avx128fma_invsqrt_f(rsq20);
1322             rinv21           = avx128fma_invsqrt_f(rsq21);
1323             rinv22           = avx128fma_invsqrt_f(rsq22);
1324
1325             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1326             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1327             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1328             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1329             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1330             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1331             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1332             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1333             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1334
1335             fjx0             = _mm_setzero_ps();
1336             fjy0             = _mm_setzero_ps();
1337             fjz0             = _mm_setzero_ps();
1338             fjx1             = _mm_setzero_ps();
1339             fjy1             = _mm_setzero_ps();
1340             fjz1             = _mm_setzero_ps();
1341             fjx2             = _mm_setzero_ps();
1342             fjy2             = _mm_setzero_ps();
1343             fjz2             = _mm_setzero_ps();
1344
1345             /**************************
1346              * CALCULATE INTERACTIONS *
1347              **************************/
1348
1349             /* COULOMB ELECTROSTATICS */
1350             velec            = _mm_mul_ps(qq00,rinv00);
1351             felec            = _mm_mul_ps(velec,rinvsq00);
1352
1353             fscal            = felec;
1354
1355             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1356
1357              /* Update vectorial force */
1358             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1359             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1360             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1361
1362             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1363             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1364             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* COULOMB ELECTROSTATICS */
1371             velec            = _mm_mul_ps(qq01,rinv01);
1372             felec            = _mm_mul_ps(velec,rinvsq01);
1373
1374             fscal            = felec;
1375
1376             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1377
1378              /* Update vectorial force */
1379             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1380             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1381             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1382
1383             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1384             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1385             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             /* COULOMB ELECTROSTATICS */
1392             velec            = _mm_mul_ps(qq02,rinv02);
1393             felec            = _mm_mul_ps(velec,rinvsq02);
1394
1395             fscal            = felec;
1396
1397             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1398
1399              /* Update vectorial force */
1400             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1401             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1402             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1403
1404             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1405             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1406             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1407
1408             /**************************
1409              * CALCULATE INTERACTIONS *
1410              **************************/
1411
1412             /* COULOMB ELECTROSTATICS */
1413             velec            = _mm_mul_ps(qq10,rinv10);
1414             felec            = _mm_mul_ps(velec,rinvsq10);
1415
1416             fscal            = felec;
1417
1418             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1419
1420              /* Update vectorial force */
1421             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1422             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1423             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1424
1425             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1426             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1427             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1428
1429             /**************************
1430              * CALCULATE INTERACTIONS *
1431              **************************/
1432
1433             /* COULOMB ELECTROSTATICS */
1434             velec            = _mm_mul_ps(qq11,rinv11);
1435             felec            = _mm_mul_ps(velec,rinvsq11);
1436
1437             fscal            = felec;
1438
1439             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1440
1441              /* Update vectorial force */
1442             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1443             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1444             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1445
1446             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1447             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1448             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             /* COULOMB ELECTROSTATICS */
1455             velec            = _mm_mul_ps(qq12,rinv12);
1456             felec            = _mm_mul_ps(velec,rinvsq12);
1457
1458             fscal            = felec;
1459
1460             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1461
1462              /* Update vectorial force */
1463             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1464             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1465             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1466
1467             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1468             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1469             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1470
1471             /**************************
1472              * CALCULATE INTERACTIONS *
1473              **************************/
1474
1475             /* COULOMB ELECTROSTATICS */
1476             velec            = _mm_mul_ps(qq20,rinv20);
1477             felec            = _mm_mul_ps(velec,rinvsq20);
1478
1479             fscal            = felec;
1480
1481             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1482
1483              /* Update vectorial force */
1484             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1485             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1486             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1487
1488             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1489             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1490             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1491
1492             /**************************
1493              * CALCULATE INTERACTIONS *
1494              **************************/
1495
1496             /* COULOMB ELECTROSTATICS */
1497             velec            = _mm_mul_ps(qq21,rinv21);
1498             felec            = _mm_mul_ps(velec,rinvsq21);
1499
1500             fscal            = felec;
1501
1502             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1503
1504              /* Update vectorial force */
1505             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1506             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1507             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1508
1509             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1510             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1511             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1512
1513             /**************************
1514              * CALCULATE INTERACTIONS *
1515              **************************/
1516
1517             /* COULOMB ELECTROSTATICS */
1518             velec            = _mm_mul_ps(qq22,rinv22);
1519             felec            = _mm_mul_ps(velec,rinvsq22);
1520
1521             fscal            = felec;
1522
1523             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1524
1525              /* Update vectorial force */
1526             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1527             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1528             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1529
1530             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1531             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1532             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1533
1534             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1535             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1536             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1537             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1538
1539             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1540                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1541
1542             /* Inner loop uses 270 flops */
1543         }
1544
1545         /* End of innermost loop */
1546
1547         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1548                                               f+i_coord_offset,fshift+i_shift_offset);
1549
1550         /* Increment number of inner iterations */
1551         inneriter                  += j_index_end - j_index_start;
1552
1553         /* Outer loop uses 18 flops */
1554     }
1555
1556     /* Increment number of outer iterations */
1557     outeriter        += nri;
1558
1559     /* Update outer/inner flops */
1560
1561     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
1562 }