src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwNone_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW3W3_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            None
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEw_VdwNone_GeomW3W3_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     __m128i          ewitab;
 109     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 110     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 111     real             *ewtab;
 112     __m128           dummy_mask,cutoff_mask;
 113     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 114     __m128           one     = _mm_set1_ps(1.0);
 115     __m128           two     = _mm_set1_ps(2.0);
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = _mm_set1_ps(fr->epsfac);
 128     charge           = mdatoms->chargeA;
 129
 130     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 131     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 132     beta2            = _mm_mul_ps(beta,beta);
 133     beta3            = _mm_mul_ps(beta,beta2);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 136     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 141     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 142     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 143
 144     jq0              = _mm_set1_ps(charge[inr+0]);
 145     jq1              = _mm_set1_ps(charge[inr+1]);
 146     jq2              = _mm_set1_ps(charge[inr+2]);
 147     qq00             = _mm_mul_ps(iq0,jq0);
 148     qq01             = _mm_mul_ps(iq0,jq1);
 149     qq02             = _mm_mul_ps(iq0,jq2);
 150     qq10             = _mm_mul_ps(iq1,jq0);
 151     qq11             = _mm_mul_ps(iq1,jq1);
 152     qq12             = _mm_mul_ps(iq1,jq2);
 153     qq20             = _mm_mul_ps(iq2,jq0);
 154     qq21             = _mm_mul_ps(iq2,jq1);
 155     qq22             = _mm_mul_ps(iq2,jq2);
 156
 157     /* Avoid stupid compiler warnings */
 158     jnrA = jnrB = jnrC = jnrD = 0;
 159     j_coord_offsetA = 0;
 160     j_coord_offsetB = 0;
 161     j_coord_offsetC = 0;
 162     j_coord_offsetD = 0;
 163
 164     outeriter        = 0;
 165     inneriter        = 0;
 166
 167     for(iidx=0;iidx<4*DIM;iidx++)
 168     {
 169         scratch[iidx] = 0.0;
 170     }
 171
 172     /* Start outer loop over neighborlists */
 173     for(iidx=0; iidx<nri; iidx++)
 174     {
 175         /* Load shift vector for this list */
 176         i_shift_offset   = DIM*shiftidx[iidx];
 177
 178         /* Load limits for loop over neighbors */
 179         j_index_start    = jindex[iidx];
 180         j_index_end      = jindex[iidx+1];
 181
 182         /* Get outer coordinate index */
 183         inr              = iinr[iidx];
 184         i_coord_offset   = DIM*inr;
 185
 186         /* Load i particle coords and add shift vector */
 187         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 188                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 189
 190         fix0             = _mm_setzero_ps();
 191         fiy0             = _mm_setzero_ps();
 192         fiz0             = _mm_setzero_ps();
 193         fix1             = _mm_setzero_ps();
 194         fiy1             = _mm_setzero_ps();
 195         fiz1             = _mm_setzero_ps();
 196         fix2             = _mm_setzero_ps();
 197         fiy2             = _mm_setzero_ps();
 198         fiz2             = _mm_setzero_ps();
 199
 200         /* Reset potential sums */
 201         velecsum         = _mm_setzero_ps();
 202
 203         /* Start inner kernel loop */
 204         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 205         {
 206
 207             /* Get j neighbor index, and coordinate index */
 208             jnrA             = jjnr[jidx];
 209             jnrB             = jjnr[jidx+1];
 210             jnrC             = jjnr[jidx+2];
 211             jnrD             = jjnr[jidx+3];
 212             j_coord_offsetA  = DIM*jnrA;
 213             j_coord_offsetB  = DIM*jnrB;
 214             j_coord_offsetC  = DIM*jnrC;
 215             j_coord_offsetD  = DIM*jnrD;
 216
 217             /* load j atom coordinates */
 218             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 219                                               x+j_coord_offsetC,x+j_coord_offsetD,
 220                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 221
 222             /* Calculate displacement vector */
 223             dx00             = _mm_sub_ps(ix0,jx0);
 224             dy00             = _mm_sub_ps(iy0,jy0);
 225             dz00             = _mm_sub_ps(iz0,jz0);
 226             dx01             = _mm_sub_ps(ix0,jx1);
 227             dy01             = _mm_sub_ps(iy0,jy1);
 228             dz01             = _mm_sub_ps(iz0,jz1);
 229             dx02             = _mm_sub_ps(ix0,jx2);
 230             dy02             = _mm_sub_ps(iy0,jy2);
 231             dz02             = _mm_sub_ps(iz0,jz2);
 232             dx10             = _mm_sub_ps(ix1,jx0);
 233             dy10             = _mm_sub_ps(iy1,jy0);
 234             dz10             = _mm_sub_ps(iz1,jz0);
 235             dx11             = _mm_sub_ps(ix1,jx1);
 236             dy11             = _mm_sub_ps(iy1,jy1);
 237             dz11             = _mm_sub_ps(iz1,jz1);
 238             dx12             = _mm_sub_ps(ix1,jx2);
 239             dy12             = _mm_sub_ps(iy1,jy2);
 240             dz12             = _mm_sub_ps(iz1,jz2);
 241             dx20             = _mm_sub_ps(ix2,jx0);
 242             dy20             = _mm_sub_ps(iy2,jy0);
 243             dz20             = _mm_sub_ps(iz2,jz0);
 244             dx21             = _mm_sub_ps(ix2,jx1);
 245             dy21             = _mm_sub_ps(iy2,jy1);
 246             dz21             = _mm_sub_ps(iz2,jz1);
 247             dx22             = _mm_sub_ps(ix2,jx2);
 248             dy22             = _mm_sub_ps(iy2,jy2);
 249             dz22             = _mm_sub_ps(iz2,jz2);
 250
 251             /* Calculate squared distance and things based on it */
 252             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 253             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 254             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 255             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 256             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 257             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 258             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 259             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 260             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 261
 262             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 263             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 264             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 265             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 266             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 267             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 268             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 269             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 270             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 271
 272             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 273             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 274             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 275             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 276             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 277             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 278             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 279             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 280             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 281
 282             fjx0             = _mm_setzero_ps();
 283             fjy0             = _mm_setzero_ps();
 284             fjz0             = _mm_setzero_ps();
 285             fjx1             = _mm_setzero_ps();
 286             fjy1             = _mm_setzero_ps();
 287             fjz1             = _mm_setzero_ps();
 288             fjx2             = _mm_setzero_ps();
 289             fjy2             = _mm_setzero_ps();
 290             fjz2             = _mm_setzero_ps();
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             r00              = _mm_mul_ps(rsq00,rinv00);
 297
 298             /* EWALD ELECTROSTATICS */
 299
 300             /* Analytical PME correction */
 301             zeta2            = _mm_mul_ps(beta2,rsq00);
 302             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 303             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 304             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 305             felec            = _mm_mul_ps(qq00,felec);
 306             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 307             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 308             velec            = _mm_mul_ps(qq00,velec);
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_ps(velecsum,velec);
 312
 313             fscal            = felec;
 314
 315              /* Update vectorial force */
 316             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 317             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 318             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 319
 320             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 321             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 322             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             r01              = _mm_mul_ps(rsq01,rinv01);
 329
 330             /* EWALD ELECTROSTATICS */
 331
 332             /* Analytical PME correction */
 333             zeta2            = _mm_mul_ps(beta2,rsq01);
 334             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 335             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 336             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 337             felec            = _mm_mul_ps(qq01,felec);
 338             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 339             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 340             velec            = _mm_mul_ps(qq01,velec);
 341
 342             /* Update potential sum for this i atom from the interaction with this j atom. */
 343             velecsum         = _mm_add_ps(velecsum,velec);
 344
 345             fscal            = felec;
 346
 347              /* Update vectorial force */
 348             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 349             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 350             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 351
 352             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 353             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 354             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             r02              = _mm_mul_ps(rsq02,rinv02);
 361
 362             /* EWALD ELECTROSTATICS */
 363
 364             /* Analytical PME correction */
 365             zeta2            = _mm_mul_ps(beta2,rsq02);
 366             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 367             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 368             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 369             felec            = _mm_mul_ps(qq02,felec);
 370             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 371             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 372             velec            = _mm_mul_ps(qq02,velec);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _mm_add_ps(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379              /* Update vectorial force */
 380             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 381             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 382             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 383
 384             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 385             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 386             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             r10              = _mm_mul_ps(rsq10,rinv10);
 393
 394             /* EWALD ELECTROSTATICS */
 395
 396             /* Analytical PME correction */
 397             zeta2            = _mm_mul_ps(beta2,rsq10);
 398             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 399             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 400             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 401             felec            = _mm_mul_ps(qq10,felec);
 402             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 403             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 404             velec            = _mm_mul_ps(qq10,velec);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velecsum         = _mm_add_ps(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411              /* Update vectorial force */
 412             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 413             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 414             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 415
 416             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 417             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 418             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             r11              = _mm_mul_ps(rsq11,rinv11);
 425
 426             /* EWALD ELECTROSTATICS */
 427
 428             /* Analytical PME correction */
 429             zeta2            = _mm_mul_ps(beta2,rsq11);
 430             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 431             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 432             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 433             felec            = _mm_mul_ps(qq11,felec);
 434             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 435             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 436             velec            = _mm_mul_ps(qq11,velec);
 437
 438             /* Update potential sum for this i atom from the interaction with this j atom. */
 439             velecsum         = _mm_add_ps(velecsum,velec);
 440
 441             fscal            = felec;
 442
 443              /* Update vectorial force */
 444             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 445             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 446             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 447
 448             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 449             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 450             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 451
 452             /**************************
 453              * CALCULATE INTERACTIONS *
 454              **************************/
 455
 456             r12              = _mm_mul_ps(rsq12,rinv12);
 457
 458             /* EWALD ELECTROSTATICS */
 459
 460             /* Analytical PME correction */
 461             zeta2            = _mm_mul_ps(beta2,rsq12);
 462             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 463             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 464             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 465             felec            = _mm_mul_ps(qq12,felec);
 466             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 467             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 468             velec            = _mm_mul_ps(qq12,velec);
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velecsum         = _mm_add_ps(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475              /* Update vectorial force */
 476             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 477             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 478             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 479
 480             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 481             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 482             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 483
 484             /**************************
 485              * CALCULATE INTERACTIONS *
 486              **************************/
 487
 488             r20              = _mm_mul_ps(rsq20,rinv20);
 489
 490             /* EWALD ELECTROSTATICS */
 491
 492             /* Analytical PME correction */
 493             zeta2            = _mm_mul_ps(beta2,rsq20);
 494             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 495             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 496             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 497             felec            = _mm_mul_ps(qq20,felec);
 498             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 499             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 500             velec            = _mm_mul_ps(qq20,velec);
 501
 502             /* Update potential sum for this i atom from the interaction with this j atom. */
 503             velecsum         = _mm_add_ps(velecsum,velec);
 504
 505             fscal            = felec;
 506
 507              /* Update vectorial force */
 508             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 509             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 510             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 511
 512             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 513             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 514             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 515
 516             /**************************
 517              * CALCULATE INTERACTIONS *
 518              **************************/
 519
 520             r21              = _mm_mul_ps(rsq21,rinv21);
 521
 522             /* EWALD ELECTROSTATICS */
 523
 524             /* Analytical PME correction */
 525             zeta2            = _mm_mul_ps(beta2,rsq21);
 526             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 527             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 528             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 529             felec            = _mm_mul_ps(qq21,felec);
 530             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 531             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 532             velec            = _mm_mul_ps(qq21,velec);
 533
 534             /* Update potential sum for this i atom from the interaction with this j atom. */
 535             velecsum         = _mm_add_ps(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539              /* Update vectorial force */
 540             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 541             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 542             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 543
 544             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 545             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 546             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 547
 548             /**************************
 549              * CALCULATE INTERACTIONS *
 550              **************************/
 551
 552             r22              = _mm_mul_ps(rsq22,rinv22);
 553
 554             /* EWALD ELECTROSTATICS */
 555
 556             /* Analytical PME correction */
 557             zeta2            = _mm_mul_ps(beta2,rsq22);
 558             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 559             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 560             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 561             felec            = _mm_mul_ps(qq22,felec);
 562             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 563             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 564             velec            = _mm_mul_ps(qq22,velec);
 565
 566             /* Update potential sum for this i atom from the interaction with this j atom. */
 567             velecsum         = _mm_add_ps(velecsum,velec);
 568
 569             fscal            = felec;
 570
 571              /* Update vectorial force */
 572             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 573             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 574             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 575
 576             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 577             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 578             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 579
 580             fjptrA             = f+j_coord_offsetA;
 581             fjptrB             = f+j_coord_offsetB;
 582             fjptrC             = f+j_coord_offsetC;
 583             fjptrD             = f+j_coord_offsetD;
 584
 585             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 586                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 587
 588             /* Inner loop uses 261 flops */
 589         }
 590
 591         if(jidx<j_index_end)
 592         {
 593
 594             /* Get j neighbor index, and coordinate index */
 595             jnrlistA         = jjnr[jidx];
 596             jnrlistB         = jjnr[jidx+1];
 597             jnrlistC         = jjnr[jidx+2];
 598             jnrlistD         = jjnr[jidx+3];
 599             /* Sign of each element will be negative for non-real atoms.
 600              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 601              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 602              */
 603             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 604             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 605             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 606             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 607             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 608             j_coord_offsetA  = DIM*jnrA;
 609             j_coord_offsetB  = DIM*jnrB;
 610             j_coord_offsetC  = DIM*jnrC;
 611             j_coord_offsetD  = DIM*jnrD;
 612
 613             /* load j atom coordinates */
 614             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 615                                               x+j_coord_offsetC,x+j_coord_offsetD,
 616                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 617
 618             /* Calculate displacement vector */
 619             dx00             = _mm_sub_ps(ix0,jx0);
 620             dy00             = _mm_sub_ps(iy0,jy0);
 621             dz00             = _mm_sub_ps(iz0,jz0);
 622             dx01             = _mm_sub_ps(ix0,jx1);
 623             dy01             = _mm_sub_ps(iy0,jy1);
 624             dz01             = _mm_sub_ps(iz0,jz1);
 625             dx02             = _mm_sub_ps(ix0,jx2);
 626             dy02             = _mm_sub_ps(iy0,jy2);
 627             dz02             = _mm_sub_ps(iz0,jz2);
 628             dx10             = _mm_sub_ps(ix1,jx0);
 629             dy10             = _mm_sub_ps(iy1,jy0);
 630             dz10             = _mm_sub_ps(iz1,jz0);
 631             dx11             = _mm_sub_ps(ix1,jx1);
 632             dy11             = _mm_sub_ps(iy1,jy1);
 633             dz11             = _mm_sub_ps(iz1,jz1);
 634             dx12             = _mm_sub_ps(ix1,jx2);
 635             dy12             = _mm_sub_ps(iy1,jy2);
 636             dz12             = _mm_sub_ps(iz1,jz2);
 637             dx20             = _mm_sub_ps(ix2,jx0);
 638             dy20             = _mm_sub_ps(iy2,jy0);
 639             dz20             = _mm_sub_ps(iz2,jz0);
 640             dx21             = _mm_sub_ps(ix2,jx1);
 641             dy21             = _mm_sub_ps(iy2,jy1);
 642             dz21             = _mm_sub_ps(iz2,jz1);
 643             dx22             = _mm_sub_ps(ix2,jx2);
 644             dy22             = _mm_sub_ps(iy2,jy2);
 645             dz22             = _mm_sub_ps(iz2,jz2);
 646
 647             /* Calculate squared distance and things based on it */
 648             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 649             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 650             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 651             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 652             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 653             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 654             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 655             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 656             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 657
 658             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 659             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 660             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 661             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 662             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 663             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 664             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 665             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 666             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 667
 668             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 669             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 670             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 671             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 672             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 673             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 674             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 675             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 676             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 677
 678             fjx0             = _mm_setzero_ps();
 679             fjy0             = _mm_setzero_ps();
 680             fjz0             = _mm_setzero_ps();
 681             fjx1             = _mm_setzero_ps();
 682             fjy1             = _mm_setzero_ps();
 683             fjz1             = _mm_setzero_ps();
 684             fjx2             = _mm_setzero_ps();
 685             fjy2             = _mm_setzero_ps();
 686             fjz2             = _mm_setzero_ps();
 687
 688             /**************************
 689              * CALCULATE INTERACTIONS *
 690              **************************/
 691
 692             r00              = _mm_mul_ps(rsq00,rinv00);
 693             r00              = _mm_andnot_ps(dummy_mask,r00);
 694
 695             /* EWALD ELECTROSTATICS */
 696
 697             /* Analytical PME correction */
 698             zeta2            = _mm_mul_ps(beta2,rsq00);
 699             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 700             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 701             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 702             felec            = _mm_mul_ps(qq00,felec);
 703             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 704             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 705             velec            = _mm_mul_ps(qq00,velec);
 706
 707             /* Update potential sum for this i atom from the interaction with this j atom. */
 708             velec            = _mm_andnot_ps(dummy_mask,velec);
 709             velecsum         = _mm_add_ps(velecsum,velec);
 710
 711             fscal            = felec;
 712
 713             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 714
 715              /* Update vectorial force */
 716             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 717             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 718             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 719
 720             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 721             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 722             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             r01              = _mm_mul_ps(rsq01,rinv01);
 729             r01              = _mm_andnot_ps(dummy_mask,r01);
 730
 731             /* EWALD ELECTROSTATICS */
 732
 733             /* Analytical PME correction */
 734             zeta2            = _mm_mul_ps(beta2,rsq01);
 735             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 736             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 737             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 738             felec            = _mm_mul_ps(qq01,felec);
 739             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 740             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 741             velec            = _mm_mul_ps(qq01,velec);
 742
 743             /* Update potential sum for this i atom from the interaction with this j atom. */
 744             velec            = _mm_andnot_ps(dummy_mask,velec);
 745             velecsum         = _mm_add_ps(velecsum,velec);
 746
 747             fscal            = felec;
 748
 749             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 750
 751              /* Update vectorial force */
 752             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 753             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 754             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 755
 756             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 757             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 758             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 759
 760             /**************************
 761              * CALCULATE INTERACTIONS *
 762              **************************/
 763
 764             r02              = _mm_mul_ps(rsq02,rinv02);
 765             r02              = _mm_andnot_ps(dummy_mask,r02);
 766
 767             /* EWALD ELECTROSTATICS */
 768
 769             /* Analytical PME correction */
 770             zeta2            = _mm_mul_ps(beta2,rsq02);
 771             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 772             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 773             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 774             felec            = _mm_mul_ps(qq02,felec);
 775             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 776             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 777             velec            = _mm_mul_ps(qq02,velec);
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             velec            = _mm_andnot_ps(dummy_mask,velec);
 781             velecsum         = _mm_add_ps(velecsum,velec);
 782
 783             fscal            = felec;
 784
 785             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 786
 787              /* Update vectorial force */
 788             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 789             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 790             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 791
 792             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 793             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 794             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             r10              = _mm_mul_ps(rsq10,rinv10);
 801             r10              = _mm_andnot_ps(dummy_mask,r10);
 802
 803             /* EWALD ELECTROSTATICS */
 804
 805             /* Analytical PME correction */
 806             zeta2            = _mm_mul_ps(beta2,rsq10);
 807             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 808             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 809             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 810             felec            = _mm_mul_ps(qq10,felec);
 811             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 812             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 813             velec            = _mm_mul_ps(qq10,velec);
 814
 815             /* Update potential sum for this i atom from the interaction with this j atom. */
 816             velec            = _mm_andnot_ps(dummy_mask,velec);
 817             velecsum         = _mm_add_ps(velecsum,velec);
 818
 819             fscal            = felec;
 820
 821             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 822
 823              /* Update vectorial force */
 824             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 825             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 826             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 827
 828             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 829             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 830             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 831
 832             /**************************
 833              * CALCULATE INTERACTIONS *
 834              **************************/
 835
 836             r11              = _mm_mul_ps(rsq11,rinv11);
 837             r11              = _mm_andnot_ps(dummy_mask,r11);
 838
 839             /* EWALD ELECTROSTATICS */
 840
 841             /* Analytical PME correction */
 842             zeta2            = _mm_mul_ps(beta2,rsq11);
 843             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 844             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 845             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 846             felec            = _mm_mul_ps(qq11,felec);
 847             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 848             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 849             velec            = _mm_mul_ps(qq11,velec);
 850
 851             /* Update potential sum for this i atom from the interaction with this j atom. */
 852             velec            = _mm_andnot_ps(dummy_mask,velec);
 853             velecsum         = _mm_add_ps(velecsum,velec);
 854
 855             fscal            = felec;
 856
 857             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 858
 859              /* Update vectorial force */
 860             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 861             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 862             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 863
 864             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 865             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 866             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 867
 868             /**************************
 869              * CALCULATE INTERACTIONS *
 870              **************************/
 871
 872             r12              = _mm_mul_ps(rsq12,rinv12);
 873             r12              = _mm_andnot_ps(dummy_mask,r12);
 874
 875             /* EWALD ELECTROSTATICS */
 876
 877             /* Analytical PME correction */
 878             zeta2            = _mm_mul_ps(beta2,rsq12);
 879             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 880             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 881             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 882             felec            = _mm_mul_ps(qq12,felec);
 883             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 884             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 885             velec            = _mm_mul_ps(qq12,velec);
 886
 887             /* Update potential sum for this i atom from the interaction with this j atom. */
 888             velec            = _mm_andnot_ps(dummy_mask,velec);
 889             velecsum         = _mm_add_ps(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 894
 895              /* Update vectorial force */
 896             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 897             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 898             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 899
 900             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 901             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 902             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             r20              = _mm_mul_ps(rsq20,rinv20);
 909             r20              = _mm_andnot_ps(dummy_mask,r20);
 910
 911             /* EWALD ELECTROSTATICS */
 912
 913             /* Analytical PME correction */
 914             zeta2            = _mm_mul_ps(beta2,rsq20);
 915             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 916             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 917             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 918             felec            = _mm_mul_ps(qq20,felec);
 919             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 920             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 921             velec            = _mm_mul_ps(qq20,velec);
 922
 923             /* Update potential sum for this i atom from the interaction with this j atom. */
 924             velec            = _mm_andnot_ps(dummy_mask,velec);
 925             velecsum         = _mm_add_ps(velecsum,velec);
 926
 927             fscal            = felec;
 928
 929             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 930
 931              /* Update vectorial force */
 932             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 933             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 934             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 935
 936             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 937             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 938             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 939
 940             /**************************
 941              * CALCULATE INTERACTIONS *
 942              **************************/
 943
 944             r21              = _mm_mul_ps(rsq21,rinv21);
 945             r21              = _mm_andnot_ps(dummy_mask,r21);
 946
 947             /* EWALD ELECTROSTATICS */
 948
 949             /* Analytical PME correction */
 950             zeta2            = _mm_mul_ps(beta2,rsq21);
 951             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 952             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 953             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 954             felec            = _mm_mul_ps(qq21,felec);
 955             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 956             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 957             velec            = _mm_mul_ps(qq21,velec);
 958
 959             /* Update potential sum for this i atom from the interaction with this j atom. */
 960             velec            = _mm_andnot_ps(dummy_mask,velec);
 961             velecsum         = _mm_add_ps(velecsum,velec);
 962
 963             fscal            = felec;
 964
 965             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 966
 967              /* Update vectorial force */
 968             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 969             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 970             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 971
 972             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 973             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 974             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 975
 976             /**************************
 977              * CALCULATE INTERACTIONS *
 978              **************************/
 979
 980             r22              = _mm_mul_ps(rsq22,rinv22);
 981             r22              = _mm_andnot_ps(dummy_mask,r22);
 982
 983             /* EWALD ELECTROSTATICS */
 984
 985             /* Analytical PME correction */
 986             zeta2            = _mm_mul_ps(beta2,rsq22);
 987             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 988             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 989             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 990             felec            = _mm_mul_ps(qq22,felec);
 991             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 992             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 993             velec            = _mm_mul_ps(qq22,velec);
 994
 995             /* Update potential sum for this i atom from the interaction with this j atom. */
 996             velec            = _mm_andnot_ps(dummy_mask,velec);
 997             velecsum         = _mm_add_ps(velecsum,velec);
 998
 999             fscal            = felec;
1000
1001             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1002
1003              /* Update vectorial force */
1004             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1005             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1006             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1007
1008             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1009             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1010             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1011
1012             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1013             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1014             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1015             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1016
1017             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1018                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1019
1020             /* Inner loop uses 270 flops */
1021         }
1022
1023         /* End of innermost loop */
1024
1025         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1026                                               f+i_coord_offset,fshift+i_shift_offset);
1027
1028         ggid                        = gid[iidx];
1029         /* Update potential energies */
1030         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1031
1032         /* Increment number of inner iterations */
1033         inneriter                  += j_index_end - j_index_start;
1034
1035         /* Outer loop uses 19 flops */
1036     }
1037
1038     /* Increment number of outer iterations */
1039     outeriter        += nri;
1040
1041     /* Update outer/inner flops */
1042
1043     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*270);
1044 }
1045 /*
1046  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW3W3_F_avx_128_fma_single
1047  * Electrostatics interaction: Ewald
1048  * VdW interaction:            None
1049  * Geometry:                   Water3-Water3
1050  * Calculate force/pot:        Force
1051  */
1052 void
1053 nb_kernel_ElecEw_VdwNone_GeomW3W3_F_avx_128_fma_single
1054                     (t_nblist                    * gmx_restrict       nlist,
1055                      rvec                        * gmx_restrict          xx,
1056                      rvec                        * gmx_restrict          ff,
1057                      t_forcerec                  * gmx_restrict          fr,
1058                      t_mdatoms                   * gmx_restrict     mdatoms,
1059                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1060                      t_nrnb                      * gmx_restrict        nrnb)
1061 {
1062     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1063      * just 0 for non-waters.
1064      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1065      * jnr indices corresponding to data put in the four positions in the SIMD register.
1066      */
1067     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1068     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1069     int              jnrA,jnrB,jnrC,jnrD;
1070     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1071     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1072     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1073     real             rcutoff_scalar;
1074     real             *shiftvec,*fshift,*x,*f;
1075     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1076     real             scratch[4*DIM];
1077     __m128           fscal,rcutoff,rcutoff2,jidxall;
1078     int              vdwioffset0;
1079     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1080     int              vdwioffset1;
1081     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1082     int              vdwioffset2;
1083     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1084     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1085     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1086     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1087     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1088     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1089     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1090     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1091     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1092     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1093     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1094     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1095     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1096     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1097     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1098     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1099     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1100     real             *charge;
1101     __m128i          ewitab;
1102     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1103     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1104     real             *ewtab;
1105     __m128           dummy_mask,cutoff_mask;
1106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1107     __m128           one     = _mm_set1_ps(1.0);
1108     __m128           two     = _mm_set1_ps(2.0);
1109     x                = xx[0];
1110     f                = ff[0];
1111
1112     nri              = nlist->nri;
1113     iinr             = nlist->iinr;
1114     jindex           = nlist->jindex;
1115     jjnr             = nlist->jjnr;
1116     shiftidx         = nlist->shift;
1117     gid              = nlist->gid;
1118     shiftvec         = fr->shift_vec[0];
1119     fshift           = fr->fshift[0];
1120     facel            = _mm_set1_ps(fr->epsfac);
1121     charge           = mdatoms->chargeA;
1122
1123     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1124     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
1125     beta2            = _mm_mul_ps(beta,beta);
1126     beta3            = _mm_mul_ps(beta,beta2);
1127     ewtab            = fr->ic->tabq_coul_F;
1128     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1129     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1130
1131     /* Setup water-specific parameters */
1132     inr              = nlist->iinr[0];
1133     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1134     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1135     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1136
1137     jq0              = _mm_set1_ps(charge[inr+0]);
1138     jq1              = _mm_set1_ps(charge[inr+1]);
1139     jq2              = _mm_set1_ps(charge[inr+2]);
1140     qq00             = _mm_mul_ps(iq0,jq0);
1141     qq01             = _mm_mul_ps(iq0,jq1);
1142     qq02             = _mm_mul_ps(iq0,jq2);
1143     qq10             = _mm_mul_ps(iq1,jq0);
1144     qq11             = _mm_mul_ps(iq1,jq1);
1145     qq12             = _mm_mul_ps(iq1,jq2);
1146     qq20             = _mm_mul_ps(iq2,jq0);
1147     qq21             = _mm_mul_ps(iq2,jq1);
1148     qq22             = _mm_mul_ps(iq2,jq2);
1149
1150     /* Avoid stupid compiler warnings */
1151     jnrA = jnrB = jnrC = jnrD = 0;
1152     j_coord_offsetA = 0;
1153     j_coord_offsetB = 0;
1154     j_coord_offsetC = 0;
1155     j_coord_offsetD = 0;
1156
1157     outeriter        = 0;
1158     inneriter        = 0;
1159
1160     for(iidx=0;iidx<4*DIM;iidx++)
1161     {
1162         scratch[iidx] = 0.0;
1163     }
1164
1165     /* Start outer loop over neighborlists */
1166     for(iidx=0; iidx<nri; iidx++)
1167     {
1168         /* Load shift vector for this list */
1169         i_shift_offset   = DIM*shiftidx[iidx];
1170
1171         /* Load limits for loop over neighbors */
1172         j_index_start    = jindex[iidx];
1173         j_index_end      = jindex[iidx+1];
1174
1175         /* Get outer coordinate index */
1176         inr              = iinr[iidx];
1177         i_coord_offset   = DIM*inr;
1178
1179         /* Load i particle coords and add shift vector */
1180         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1181                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1182
1183         fix0             = _mm_setzero_ps();
1184         fiy0             = _mm_setzero_ps();
1185         fiz0             = _mm_setzero_ps();
1186         fix1             = _mm_setzero_ps();
1187         fiy1             = _mm_setzero_ps();
1188         fiz1             = _mm_setzero_ps();
1189         fix2             = _mm_setzero_ps();
1190         fiy2             = _mm_setzero_ps();
1191         fiz2             = _mm_setzero_ps();
1192
1193         /* Start inner kernel loop */
1194         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1195         {
1196
1197             /* Get j neighbor index, and coordinate index */
1198             jnrA             = jjnr[jidx];
1199             jnrB             = jjnr[jidx+1];
1200             jnrC             = jjnr[jidx+2];
1201             jnrD             = jjnr[jidx+3];
1202             j_coord_offsetA  = DIM*jnrA;
1203             j_coord_offsetB  = DIM*jnrB;
1204             j_coord_offsetC  = DIM*jnrC;
1205             j_coord_offsetD  = DIM*jnrD;
1206
1207             /* load j atom coordinates */
1208             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1209                                               x+j_coord_offsetC,x+j_coord_offsetD,
1210                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1211
1212             /* Calculate displacement vector */
1213             dx00             = _mm_sub_ps(ix0,jx0);
1214             dy00             = _mm_sub_ps(iy0,jy0);
1215             dz00             = _mm_sub_ps(iz0,jz0);
1216             dx01             = _mm_sub_ps(ix0,jx1);
1217             dy01             = _mm_sub_ps(iy0,jy1);
1218             dz01             = _mm_sub_ps(iz0,jz1);
1219             dx02             = _mm_sub_ps(ix0,jx2);
1220             dy02             = _mm_sub_ps(iy0,jy2);
1221             dz02             = _mm_sub_ps(iz0,jz2);
1222             dx10             = _mm_sub_ps(ix1,jx0);
1223             dy10             = _mm_sub_ps(iy1,jy0);
1224             dz10             = _mm_sub_ps(iz1,jz0);
1225             dx11             = _mm_sub_ps(ix1,jx1);
1226             dy11             = _mm_sub_ps(iy1,jy1);
1227             dz11             = _mm_sub_ps(iz1,jz1);
1228             dx12             = _mm_sub_ps(ix1,jx2);
1229             dy12             = _mm_sub_ps(iy1,jy2);
1230             dz12             = _mm_sub_ps(iz1,jz2);
1231             dx20             = _mm_sub_ps(ix2,jx0);
1232             dy20             = _mm_sub_ps(iy2,jy0);
1233             dz20             = _mm_sub_ps(iz2,jz0);
1234             dx21             = _mm_sub_ps(ix2,jx1);
1235             dy21             = _mm_sub_ps(iy2,jy1);
1236             dz21             = _mm_sub_ps(iz2,jz1);
1237             dx22             = _mm_sub_ps(ix2,jx2);
1238             dy22             = _mm_sub_ps(iy2,jy2);
1239             dz22             = _mm_sub_ps(iz2,jz2);
1240
1241             /* Calculate squared distance and things based on it */
1242             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1243             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1244             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1245             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1246             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1247             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1248             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1249             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1250             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1251
1252             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1253             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1254             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1255             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1256             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1257             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1258             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1259             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1260             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1261
1262             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1263             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1264             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1265             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1266             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1267             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1268             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1269             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1270             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1271
1272             fjx0             = _mm_setzero_ps();
1273             fjy0             = _mm_setzero_ps();
1274             fjz0             = _mm_setzero_ps();
1275             fjx1             = _mm_setzero_ps();
1276             fjy1             = _mm_setzero_ps();
1277             fjz1             = _mm_setzero_ps();
1278             fjx2             = _mm_setzero_ps();
1279             fjy2             = _mm_setzero_ps();
1280             fjz2             = _mm_setzero_ps();
1281
1282             /**************************
1283              * CALCULATE INTERACTIONS *
1284              **************************/
1285
1286             r00              = _mm_mul_ps(rsq00,rinv00);
1287
1288             /* EWALD ELECTROSTATICS */
1289
1290             /* Analytical PME correction */
1291             zeta2            = _mm_mul_ps(beta2,rsq00);
1292             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1293             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1294             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1295             felec            = _mm_mul_ps(qq00,felec);
1296
1297             fscal            = felec;
1298
1299              /* Update vectorial force */
1300             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1301             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1302             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1303
1304             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1305             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1306             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1307
1308             /**************************
1309              * CALCULATE INTERACTIONS *
1310              **************************/
1311
1312             r01              = _mm_mul_ps(rsq01,rinv01);
1313
1314             /* EWALD ELECTROSTATICS */
1315
1316             /* Analytical PME correction */
1317             zeta2            = _mm_mul_ps(beta2,rsq01);
1318             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1319             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1320             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1321             felec            = _mm_mul_ps(qq01,felec);
1322
1323             fscal            = felec;
1324
1325              /* Update vectorial force */
1326             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1327             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1328             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1329
1330             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1331             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1332             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1333
1334             /**************************
1335              * CALCULATE INTERACTIONS *
1336              **************************/
1337
1338             r02              = _mm_mul_ps(rsq02,rinv02);
1339
1340             /* EWALD ELECTROSTATICS */
1341
1342             /* Analytical PME correction */
1343             zeta2            = _mm_mul_ps(beta2,rsq02);
1344             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1345             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1346             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1347             felec            = _mm_mul_ps(qq02,felec);
1348
1349             fscal            = felec;
1350
1351              /* Update vectorial force */
1352             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1353             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1354             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1355
1356             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1357             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1358             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1359
1360             /**************************
1361              * CALCULATE INTERACTIONS *
1362              **************************/
1363
1364             r10              = _mm_mul_ps(rsq10,rinv10);
1365
1366             /* EWALD ELECTROSTATICS */
1367
1368             /* Analytical PME correction */
1369             zeta2            = _mm_mul_ps(beta2,rsq10);
1370             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1371             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1372             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1373             felec            = _mm_mul_ps(qq10,felec);
1374
1375             fscal            = felec;
1376
1377              /* Update vectorial force */
1378             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1379             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1380             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1381
1382             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1383             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1384             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             r11              = _mm_mul_ps(rsq11,rinv11);
1391
1392             /* EWALD ELECTROSTATICS */
1393
1394             /* Analytical PME correction */
1395             zeta2            = _mm_mul_ps(beta2,rsq11);
1396             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1397             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1398             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1399             felec            = _mm_mul_ps(qq11,felec);
1400
1401             fscal            = felec;
1402
1403              /* Update vectorial force */
1404             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1405             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1406             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1407
1408             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1409             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1410             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             r12              = _mm_mul_ps(rsq12,rinv12);
1417
1418             /* EWALD ELECTROSTATICS */
1419
1420             /* Analytical PME correction */
1421             zeta2            = _mm_mul_ps(beta2,rsq12);
1422             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1423             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1424             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1425             felec            = _mm_mul_ps(qq12,felec);
1426
1427             fscal            = felec;
1428
1429              /* Update vectorial force */
1430             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1431             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1432             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1433
1434             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1435             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1436             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             r20              = _mm_mul_ps(rsq20,rinv20);
1443
1444             /* EWALD ELECTROSTATICS */
1445
1446             /* Analytical PME correction */
1447             zeta2            = _mm_mul_ps(beta2,rsq20);
1448             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1449             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1450             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1451             felec            = _mm_mul_ps(qq20,felec);
1452
1453             fscal            = felec;
1454
1455              /* Update vectorial force */
1456             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1457             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1458             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1459
1460             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1461             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1462             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             r21              = _mm_mul_ps(rsq21,rinv21);
1469
1470             /* EWALD ELECTROSTATICS */
1471
1472             /* Analytical PME correction */
1473             zeta2            = _mm_mul_ps(beta2,rsq21);
1474             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1475             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1476             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1477             felec            = _mm_mul_ps(qq21,felec);
1478
1479             fscal            = felec;
1480
1481              /* Update vectorial force */
1482             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1483             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1484             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1485
1486             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1487             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1488             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1489
1490             /**************************
1491              * CALCULATE INTERACTIONS *
1492              **************************/
1493
1494             r22              = _mm_mul_ps(rsq22,rinv22);
1495
1496             /* EWALD ELECTROSTATICS */
1497
1498             /* Analytical PME correction */
1499             zeta2            = _mm_mul_ps(beta2,rsq22);
1500             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1501             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1502             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1503             felec            = _mm_mul_ps(qq22,felec);
1504
1505             fscal            = felec;
1506
1507              /* Update vectorial force */
1508             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1509             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1510             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1511
1512             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1513             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1514             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1515
1516             fjptrA             = f+j_coord_offsetA;
1517             fjptrB             = f+j_coord_offsetB;
1518             fjptrC             = f+j_coord_offsetC;
1519             fjptrD             = f+j_coord_offsetD;
1520
1521             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1522                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1523
1524             /* Inner loop uses 252 flops */
1525         }
1526
1527         if(jidx<j_index_end)
1528         {
1529
1530             /* Get j neighbor index, and coordinate index */
1531             jnrlistA         = jjnr[jidx];
1532             jnrlistB         = jjnr[jidx+1];
1533             jnrlistC         = jjnr[jidx+2];
1534             jnrlistD         = jjnr[jidx+3];
1535             /* Sign of each element will be negative for non-real atoms.
1536              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1537              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1538              */
1539             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1540             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1541             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1542             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1543             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1544             j_coord_offsetA  = DIM*jnrA;
1545             j_coord_offsetB  = DIM*jnrB;
1546             j_coord_offsetC  = DIM*jnrC;
1547             j_coord_offsetD  = DIM*jnrD;
1548
1549             /* load j atom coordinates */
1550             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1551                                               x+j_coord_offsetC,x+j_coord_offsetD,
1552                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1553
1554             /* Calculate displacement vector */
1555             dx00             = _mm_sub_ps(ix0,jx0);
1556             dy00             = _mm_sub_ps(iy0,jy0);
1557             dz00             = _mm_sub_ps(iz0,jz0);
1558             dx01             = _mm_sub_ps(ix0,jx1);
1559             dy01             = _mm_sub_ps(iy0,jy1);
1560             dz01             = _mm_sub_ps(iz0,jz1);
1561             dx02             = _mm_sub_ps(ix0,jx2);
1562             dy02             = _mm_sub_ps(iy0,jy2);
1563             dz02             = _mm_sub_ps(iz0,jz2);
1564             dx10             = _mm_sub_ps(ix1,jx0);
1565             dy10             = _mm_sub_ps(iy1,jy0);
1566             dz10             = _mm_sub_ps(iz1,jz0);
1567             dx11             = _mm_sub_ps(ix1,jx1);
1568             dy11             = _mm_sub_ps(iy1,jy1);
1569             dz11             = _mm_sub_ps(iz1,jz1);
1570             dx12             = _mm_sub_ps(ix1,jx2);
1571             dy12             = _mm_sub_ps(iy1,jy2);
1572             dz12             = _mm_sub_ps(iz1,jz2);
1573             dx20             = _mm_sub_ps(ix2,jx0);
1574             dy20             = _mm_sub_ps(iy2,jy0);
1575             dz20             = _mm_sub_ps(iz2,jz0);
1576             dx21             = _mm_sub_ps(ix2,jx1);
1577             dy21             = _mm_sub_ps(iy2,jy1);
1578             dz21             = _mm_sub_ps(iz2,jz1);
1579             dx22             = _mm_sub_ps(ix2,jx2);
1580             dy22             = _mm_sub_ps(iy2,jy2);
1581             dz22             = _mm_sub_ps(iz2,jz2);
1582
1583             /* Calculate squared distance and things based on it */
1584             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1585             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1586             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1587             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1588             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1589             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1590             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1591             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1592             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1593
1594             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1595             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1596             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1597             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1598             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1599             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1600             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1601             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1602             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1603
1604             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1605             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1606             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1607             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1608             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1609             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1610             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1611             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1612             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1613
1614             fjx0             = _mm_setzero_ps();
1615             fjy0             = _mm_setzero_ps();
1616             fjz0             = _mm_setzero_ps();
1617             fjx1             = _mm_setzero_ps();
1618             fjy1             = _mm_setzero_ps();
1619             fjz1             = _mm_setzero_ps();
1620             fjx2             = _mm_setzero_ps();
1621             fjy2             = _mm_setzero_ps();
1622             fjz2             = _mm_setzero_ps();
1623
1624             /**************************
1625              * CALCULATE INTERACTIONS *
1626              **************************/
1627
1628             r00              = _mm_mul_ps(rsq00,rinv00);
1629             r00              = _mm_andnot_ps(dummy_mask,r00);
1630
1631             /* EWALD ELECTROSTATICS */
1632
1633             /* Analytical PME correction */
1634             zeta2            = _mm_mul_ps(beta2,rsq00);
1635             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1636             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1637             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1638             felec            = _mm_mul_ps(qq00,felec);
1639
1640             fscal            = felec;
1641
1642             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1643
1644              /* Update vectorial force */
1645             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1646             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1647             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1648
1649             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1650             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1651             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             r01              = _mm_mul_ps(rsq01,rinv01);
1658             r01              = _mm_andnot_ps(dummy_mask,r01);
1659
1660             /* EWALD ELECTROSTATICS */
1661
1662             /* Analytical PME correction */
1663             zeta2            = _mm_mul_ps(beta2,rsq01);
1664             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1665             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1666             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1667             felec            = _mm_mul_ps(qq01,felec);
1668
1669             fscal            = felec;
1670
1671             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1672
1673              /* Update vectorial force */
1674             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1675             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1676             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1677
1678             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1679             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1680             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1681
1682             /**************************
1683              * CALCULATE INTERACTIONS *
1684              **************************/
1685
1686             r02              = _mm_mul_ps(rsq02,rinv02);
1687             r02              = _mm_andnot_ps(dummy_mask,r02);
1688
1689             /* EWALD ELECTROSTATICS */
1690
1691             /* Analytical PME correction */
1692             zeta2            = _mm_mul_ps(beta2,rsq02);
1693             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1694             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1695             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1696             felec            = _mm_mul_ps(qq02,felec);
1697
1698             fscal            = felec;
1699
1700             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1701
1702              /* Update vectorial force */
1703             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1704             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1705             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1706
1707             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1708             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1709             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1710
1711             /**************************
1712              * CALCULATE INTERACTIONS *
1713              **************************/
1714
1715             r10              = _mm_mul_ps(rsq10,rinv10);
1716             r10              = _mm_andnot_ps(dummy_mask,r10);
1717
1718             /* EWALD ELECTROSTATICS */
1719
1720             /* Analytical PME correction */
1721             zeta2            = _mm_mul_ps(beta2,rsq10);
1722             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1723             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1724             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1725             felec            = _mm_mul_ps(qq10,felec);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1730
1731              /* Update vectorial force */
1732             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1733             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1734             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1735
1736             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1737             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1738             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1739
1740             /**************************
1741              * CALCULATE INTERACTIONS *
1742              **************************/
1743
1744             r11              = _mm_mul_ps(rsq11,rinv11);
1745             r11              = _mm_andnot_ps(dummy_mask,r11);
1746
1747             /* EWALD ELECTROSTATICS */
1748
1749             /* Analytical PME correction */
1750             zeta2            = _mm_mul_ps(beta2,rsq11);
1751             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1752             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1753             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1754             felec            = _mm_mul_ps(qq11,felec);
1755
1756             fscal            = felec;
1757
1758             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1759
1760              /* Update vectorial force */
1761             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1762             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1763             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1764
1765             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1766             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1767             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1768
1769             /**************************
1770              * CALCULATE INTERACTIONS *
1771              **************************/
1772
1773             r12              = _mm_mul_ps(rsq12,rinv12);
1774             r12              = _mm_andnot_ps(dummy_mask,r12);
1775
1776             /* EWALD ELECTROSTATICS */
1777
1778             /* Analytical PME correction */
1779             zeta2            = _mm_mul_ps(beta2,rsq12);
1780             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1781             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1782             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1783             felec            = _mm_mul_ps(qq12,felec);
1784
1785             fscal            = felec;
1786
1787             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1788
1789              /* Update vectorial force */
1790             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1791             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1792             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1793
1794             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1795             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1796             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1797
1798             /**************************
1799              * CALCULATE INTERACTIONS *
1800              **************************/
1801
1802             r20              = _mm_mul_ps(rsq20,rinv20);
1803             r20              = _mm_andnot_ps(dummy_mask,r20);
1804
1805             /* EWALD ELECTROSTATICS */
1806
1807             /* Analytical PME correction */
1808             zeta2            = _mm_mul_ps(beta2,rsq20);
1809             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1810             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1811             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1812             felec            = _mm_mul_ps(qq20,felec);
1813
1814             fscal            = felec;
1815
1816             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1817
1818              /* Update vectorial force */
1819             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1820             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1821             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1822
1823             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1824             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1825             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1826
1827             /**************************
1828              * CALCULATE INTERACTIONS *
1829              **************************/
1830
1831             r21              = _mm_mul_ps(rsq21,rinv21);
1832             r21              = _mm_andnot_ps(dummy_mask,r21);
1833
1834             /* EWALD ELECTROSTATICS */
1835
1836             /* Analytical PME correction */
1837             zeta2            = _mm_mul_ps(beta2,rsq21);
1838             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1839             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1840             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1841             felec            = _mm_mul_ps(qq21,felec);
1842
1843             fscal            = felec;
1844
1845             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1846
1847              /* Update vectorial force */
1848             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1849             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1850             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1851
1852             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1853             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1854             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1855
1856             /**************************
1857              * CALCULATE INTERACTIONS *
1858              **************************/
1859
1860             r22              = _mm_mul_ps(rsq22,rinv22);
1861             r22              = _mm_andnot_ps(dummy_mask,r22);
1862
1863             /* EWALD ELECTROSTATICS */
1864
1865             /* Analytical PME correction */
1866             zeta2            = _mm_mul_ps(beta2,rsq22);
1867             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1868             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1869             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1870             felec            = _mm_mul_ps(qq22,felec);
1871
1872             fscal            = felec;
1873
1874             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1875
1876              /* Update vectorial force */
1877             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1878             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1879             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1880
1881             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1882             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1883             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1884
1885             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1886             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1887             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1888             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1889
1890             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1891                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1892
1893             /* Inner loop uses 261 flops */
1894         }
1895
1896         /* End of innermost loop */
1897
1898         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1899                                               f+i_coord_offset,fshift+i_shift_offset);
1900
1901         /* Increment number of inner iterations */
1902         inneriter                  += j_index_end - j_index_start;
1903
1904         /* Outer loop uses 18 flops */
1905     }
1906
1907     /* Increment number of outer iterations */
1908     outeriter        += nri;
1909
1910     /* Update outer/inner flops */
1911
1912     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*261);
1913 }