src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwLJ_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128i          ewitab;
  99     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 100     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 101     real             *ewtab;
 102     __m128           dummy_mask,cutoff_mask;
 103     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 104     __m128           one     = _mm_set1_ps(1.0);
 105     __m128           two     = _mm_set1_ps(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_ps(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 124     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 125     beta2            = _mm_mul_ps(beta,beta);
 126     beta3            = _mm_mul_ps(beta,beta2);
 127     ewtab            = fr->ic->tabq_coul_FDV0;
 128     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 129     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 130
 131     /* Setup water-specific parameters */
 132     inr              = nlist->iinr[0];
 133     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 134     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 135     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 136     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 137
 138     jq0              = _mm_set1_ps(charge[inr+0]);
 139     jq1              = _mm_set1_ps(charge[inr+1]);
 140     jq2              = _mm_set1_ps(charge[inr+2]);
 141     vdwjidx0A        = 2*vdwtype[inr+0];
 142     qq00             = _mm_mul_ps(iq0,jq0);
 143     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 144     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 145     qq01             = _mm_mul_ps(iq0,jq1);
 146     qq02             = _mm_mul_ps(iq0,jq2);
 147     qq10             = _mm_mul_ps(iq1,jq0);
 148     qq11             = _mm_mul_ps(iq1,jq1);
 149     qq12             = _mm_mul_ps(iq1,jq2);
 150     qq20             = _mm_mul_ps(iq2,jq0);
 151     qq21             = _mm_mul_ps(iq2,jq1);
 152     qq22             = _mm_mul_ps(iq2,jq2);
 153
 154     /* Avoid stupid compiler warnings */
 155     jnrA = jnrB = jnrC = jnrD = 0;
 156     j_coord_offsetA = 0;
 157     j_coord_offsetB = 0;
 158     j_coord_offsetC = 0;
 159     j_coord_offsetD = 0;
 160
 161     outeriter        = 0;
 162     inneriter        = 0;
 163
 164     for(iidx=0;iidx<4*DIM;iidx++)
 165     {
 166         scratch[iidx] = 0.0;
 167     }
 168
 169     /* Start outer loop over neighborlists */
 170     for(iidx=0; iidx<nri; iidx++)
 171     {
 172         /* Load shift vector for this list */
 173         i_shift_offset   = DIM*shiftidx[iidx];
 174
 175         /* Load limits for loop over neighbors */
 176         j_index_start    = jindex[iidx];
 177         j_index_end      = jindex[iidx+1];
 178
 179         /* Get outer coordinate index */
 180         inr              = iinr[iidx];
 181         i_coord_offset   = DIM*inr;
 182
 183         /* Load i particle coords and add shift vector */
 184         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 185                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 186
 187         fix0             = _mm_setzero_ps();
 188         fiy0             = _mm_setzero_ps();
 189         fiz0             = _mm_setzero_ps();
 190         fix1             = _mm_setzero_ps();
 191         fiy1             = _mm_setzero_ps();
 192         fiz1             = _mm_setzero_ps();
 193         fix2             = _mm_setzero_ps();
 194         fiy2             = _mm_setzero_ps();
 195         fiz2             = _mm_setzero_ps();
 196
 197         /* Reset potential sums */
 198         velecsum         = _mm_setzero_ps();
 199         vvdwsum          = _mm_setzero_ps();
 200
 201         /* Start inner kernel loop */
 202         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 203         {
 204
 205             /* Get j neighbor index, and coordinate index */
 206             jnrA             = jjnr[jidx];
 207             jnrB             = jjnr[jidx+1];
 208             jnrC             = jjnr[jidx+2];
 209             jnrD             = jjnr[jidx+3];
 210             j_coord_offsetA  = DIM*jnrA;
 211             j_coord_offsetB  = DIM*jnrB;
 212             j_coord_offsetC  = DIM*jnrC;
 213             j_coord_offsetD  = DIM*jnrD;
 214
 215             /* load j atom coordinates */
 216             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 217                                               x+j_coord_offsetC,x+j_coord_offsetD,
 218                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 219
 220             /* Calculate displacement vector */
 221             dx00             = _mm_sub_ps(ix0,jx0);
 222             dy00             = _mm_sub_ps(iy0,jy0);
 223             dz00             = _mm_sub_ps(iz0,jz0);
 224             dx01             = _mm_sub_ps(ix0,jx1);
 225             dy01             = _mm_sub_ps(iy0,jy1);
 226             dz01             = _mm_sub_ps(iz0,jz1);
 227             dx02             = _mm_sub_ps(ix0,jx2);
 228             dy02             = _mm_sub_ps(iy0,jy2);
 229             dz02             = _mm_sub_ps(iz0,jz2);
 230             dx10             = _mm_sub_ps(ix1,jx0);
 231             dy10             = _mm_sub_ps(iy1,jy0);
 232             dz10             = _mm_sub_ps(iz1,jz0);
 233             dx11             = _mm_sub_ps(ix1,jx1);
 234             dy11             = _mm_sub_ps(iy1,jy1);
 235             dz11             = _mm_sub_ps(iz1,jz1);
 236             dx12             = _mm_sub_ps(ix1,jx2);
 237             dy12             = _mm_sub_ps(iy1,jy2);
 238             dz12             = _mm_sub_ps(iz1,jz2);
 239             dx20             = _mm_sub_ps(ix2,jx0);
 240             dy20             = _mm_sub_ps(iy2,jy0);
 241             dz20             = _mm_sub_ps(iz2,jz0);
 242             dx21             = _mm_sub_ps(ix2,jx1);
 243             dy21             = _mm_sub_ps(iy2,jy1);
 244             dz21             = _mm_sub_ps(iz2,jz1);
 245             dx22             = _mm_sub_ps(ix2,jx2);
 246             dy22             = _mm_sub_ps(iy2,jy2);
 247             dz22             = _mm_sub_ps(iz2,jz2);
 248
 249             /* Calculate squared distance and things based on it */
 250             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 251             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 252             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 253             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 254             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 255             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 256             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 257             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 258             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 259
 260             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 261             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 262             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 263             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 264             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 265             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 266             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 267             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 268             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 269
 270             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 271             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 272             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 273             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 274             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 275             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 276             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 277             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 278             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 279
 280             fjx0             = _mm_setzero_ps();
 281             fjy0             = _mm_setzero_ps();
 282             fjz0             = _mm_setzero_ps();
 283             fjx1             = _mm_setzero_ps();
 284             fjy1             = _mm_setzero_ps();
 285             fjz1             = _mm_setzero_ps();
 286             fjx2             = _mm_setzero_ps();
 287             fjy2             = _mm_setzero_ps();
 288             fjz2             = _mm_setzero_ps();
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             r00              = _mm_mul_ps(rsq00,rinv00);
 295
 296             /* EWALD ELECTROSTATICS */
 297
 298             /* Analytical PME correction */
 299             zeta2            = _mm_mul_ps(beta2,rsq00);
 300             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 301             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 302             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 303             felec            = _mm_mul_ps(qq00,felec);
 304             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 305             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 306             velec            = _mm_mul_ps(qq00,velec);
 307
 308             /* LENNARD-JONES DISPERSION/REPULSION */
 309
 310             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 311             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 312             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 313             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 314             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 315
 316             /* Update potential sum for this i atom from the interaction with this j atom. */
 317             velecsum         = _mm_add_ps(velecsum,velec);
 318             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 319
 320             fscal            = _mm_add_ps(felec,fvdw);
 321
 322              /* Update vectorial force */
 323             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 324             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 325             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 326
 327             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 328             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 329             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 330
 331             /**************************
 332              * CALCULATE INTERACTIONS *
 333              **************************/
 334
 335             r01              = _mm_mul_ps(rsq01,rinv01);
 336
 337             /* EWALD ELECTROSTATICS */
 338
 339             /* Analytical PME correction */
 340             zeta2            = _mm_mul_ps(beta2,rsq01);
 341             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 342             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 343             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 344             felec            = _mm_mul_ps(qq01,felec);
 345             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 346             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 347             velec            = _mm_mul_ps(qq01,velec);
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             velecsum         = _mm_add_ps(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354              /* Update vectorial force */
 355             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 356             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 357             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 358
 359             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 360             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 361             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             r02              = _mm_mul_ps(rsq02,rinv02);
 368
 369             /* EWALD ELECTROSTATICS */
 370
 371             /* Analytical PME correction */
 372             zeta2            = _mm_mul_ps(beta2,rsq02);
 373             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 374             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 375             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 376             felec            = _mm_mul_ps(qq02,felec);
 377             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 378             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 379             velec            = _mm_mul_ps(qq02,velec);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velecsum         = _mm_add_ps(velecsum,velec);
 383
 384             fscal            = felec;
 385
 386              /* Update vectorial force */
 387             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 388             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 389             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 390
 391             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 392             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 393             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             r10              = _mm_mul_ps(rsq10,rinv10);
 400
 401             /* EWALD ELECTROSTATICS */
 402
 403             /* Analytical PME correction */
 404             zeta2            = _mm_mul_ps(beta2,rsq10);
 405             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 406             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 407             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 408             felec            = _mm_mul_ps(qq10,felec);
 409             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 410             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 411             velec            = _mm_mul_ps(qq10,velec);
 412
 413             /* Update potential sum for this i atom from the interaction with this j atom. */
 414             velecsum         = _mm_add_ps(velecsum,velec);
 415
 416             fscal            = felec;
 417
 418              /* Update vectorial force */
 419             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 420             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 421             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 422
 423             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 424             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 425             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 426
 427             /**************************
 428              * CALCULATE INTERACTIONS *
 429              **************************/
 430
 431             r11              = _mm_mul_ps(rsq11,rinv11);
 432
 433             /* EWALD ELECTROSTATICS */
 434
 435             /* Analytical PME correction */
 436             zeta2            = _mm_mul_ps(beta2,rsq11);
 437             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 438             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 439             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 440             felec            = _mm_mul_ps(qq11,felec);
 441             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 442             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 443             velec            = _mm_mul_ps(qq11,velec);
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velecsum         = _mm_add_ps(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450              /* Update vectorial force */
 451             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 452             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 453             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 454
 455             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 456             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 457             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 458
 459             /**************************
 460              * CALCULATE INTERACTIONS *
 461              **************************/
 462
 463             r12              = _mm_mul_ps(rsq12,rinv12);
 464
 465             /* EWALD ELECTROSTATICS */
 466
 467             /* Analytical PME correction */
 468             zeta2            = _mm_mul_ps(beta2,rsq12);
 469             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 470             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 471             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 472             felec            = _mm_mul_ps(qq12,felec);
 473             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 474             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 475             velec            = _mm_mul_ps(qq12,velec);
 476
 477             /* Update potential sum for this i atom from the interaction with this j atom. */
 478             velecsum         = _mm_add_ps(velecsum,velec);
 479
 480             fscal            = felec;
 481
 482              /* Update vectorial force */
 483             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 484             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 485             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 486
 487             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 488             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 489             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 490
 491             /**************************
 492              * CALCULATE INTERACTIONS *
 493              **************************/
 494
 495             r20              = _mm_mul_ps(rsq20,rinv20);
 496
 497             /* EWALD ELECTROSTATICS */
 498
 499             /* Analytical PME correction */
 500             zeta2            = _mm_mul_ps(beta2,rsq20);
 501             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 502             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 503             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 504             felec            = _mm_mul_ps(qq20,felec);
 505             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 506             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 507             velec            = _mm_mul_ps(qq20,velec);
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velecsum         = _mm_add_ps(velecsum,velec);
 511
 512             fscal            = felec;
 513
 514              /* Update vectorial force */
 515             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 516             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 517             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 518
 519             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 520             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 521             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 522
 523             /**************************
 524              * CALCULATE INTERACTIONS *
 525              **************************/
 526
 527             r21              = _mm_mul_ps(rsq21,rinv21);
 528
 529             /* EWALD ELECTROSTATICS */
 530
 531             /* Analytical PME correction */
 532             zeta2            = _mm_mul_ps(beta2,rsq21);
 533             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 534             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 535             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 536             felec            = _mm_mul_ps(qq21,felec);
 537             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 538             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 539             velec            = _mm_mul_ps(qq21,velec);
 540
 541             /* Update potential sum for this i atom from the interaction with this j atom. */
 542             velecsum         = _mm_add_ps(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546              /* Update vectorial force */
 547             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 548             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 549             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 550
 551             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 552             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 553             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             r22              = _mm_mul_ps(rsq22,rinv22);
 560
 561             /* EWALD ELECTROSTATICS */
 562
 563             /* Analytical PME correction */
 564             zeta2            = _mm_mul_ps(beta2,rsq22);
 565             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 566             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 567             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 568             felec            = _mm_mul_ps(qq22,felec);
 569             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 570             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 571             velec            = _mm_mul_ps(qq22,velec);
 572
 573             /* Update potential sum for this i atom from the interaction with this j atom. */
 574             velecsum         = _mm_add_ps(velecsum,velec);
 575
 576             fscal            = felec;
 577
 578              /* Update vectorial force */
 579             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 580             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 581             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 582
 583             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 584             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 585             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 586
 587             fjptrA             = f+j_coord_offsetA;
 588             fjptrB             = f+j_coord_offsetB;
 589             fjptrC             = f+j_coord_offsetC;
 590             fjptrD             = f+j_coord_offsetD;
 591
 592             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 593                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 594
 595             /* Inner loop uses 273 flops */
 596         }
 597
 598         if(jidx<j_index_end)
 599         {
 600
 601             /* Get j neighbor index, and coordinate index */
 602             jnrlistA         = jjnr[jidx];
 603             jnrlistB         = jjnr[jidx+1];
 604             jnrlistC         = jjnr[jidx+2];
 605             jnrlistD         = jjnr[jidx+3];
 606             /* Sign of each element will be negative for non-real atoms.
 607              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 608              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 609              */
 610             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 611             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 612             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 613             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 614             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 615             j_coord_offsetA  = DIM*jnrA;
 616             j_coord_offsetB  = DIM*jnrB;
 617             j_coord_offsetC  = DIM*jnrC;
 618             j_coord_offsetD  = DIM*jnrD;
 619
 620             /* load j atom coordinates */
 621             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 622                                               x+j_coord_offsetC,x+j_coord_offsetD,
 623                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 624
 625             /* Calculate displacement vector */
 626             dx00             = _mm_sub_ps(ix0,jx0);
 627             dy00             = _mm_sub_ps(iy0,jy0);
 628             dz00             = _mm_sub_ps(iz0,jz0);
 629             dx01             = _mm_sub_ps(ix0,jx1);
 630             dy01             = _mm_sub_ps(iy0,jy1);
 631             dz01             = _mm_sub_ps(iz0,jz1);
 632             dx02             = _mm_sub_ps(ix0,jx2);
 633             dy02             = _mm_sub_ps(iy0,jy2);
 634             dz02             = _mm_sub_ps(iz0,jz2);
 635             dx10             = _mm_sub_ps(ix1,jx0);
 636             dy10             = _mm_sub_ps(iy1,jy0);
 637             dz10             = _mm_sub_ps(iz1,jz0);
 638             dx11             = _mm_sub_ps(ix1,jx1);
 639             dy11             = _mm_sub_ps(iy1,jy1);
 640             dz11             = _mm_sub_ps(iz1,jz1);
 641             dx12             = _mm_sub_ps(ix1,jx2);
 642             dy12             = _mm_sub_ps(iy1,jy2);
 643             dz12             = _mm_sub_ps(iz1,jz2);
 644             dx20             = _mm_sub_ps(ix2,jx0);
 645             dy20             = _mm_sub_ps(iy2,jy0);
 646             dz20             = _mm_sub_ps(iz2,jz0);
 647             dx21             = _mm_sub_ps(ix2,jx1);
 648             dy21             = _mm_sub_ps(iy2,jy1);
 649             dz21             = _mm_sub_ps(iz2,jz1);
 650             dx22             = _mm_sub_ps(ix2,jx2);
 651             dy22             = _mm_sub_ps(iy2,jy2);
 652             dz22             = _mm_sub_ps(iz2,jz2);
 653
 654             /* Calculate squared distance and things based on it */
 655             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 656             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 657             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 658             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 659             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 660             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 661             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 662             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 663             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 664
 665             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 666             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 667             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 668             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 669             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 670             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 671             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 672             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 673             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 674
 675             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 676             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 677             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 678             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 679             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 680             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 681             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 682             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 683             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 684
 685             fjx0             = _mm_setzero_ps();
 686             fjy0             = _mm_setzero_ps();
 687             fjz0             = _mm_setzero_ps();
 688             fjx1             = _mm_setzero_ps();
 689             fjy1             = _mm_setzero_ps();
 690             fjz1             = _mm_setzero_ps();
 691             fjx2             = _mm_setzero_ps();
 692             fjy2             = _mm_setzero_ps();
 693             fjz2             = _mm_setzero_ps();
 694
 695             /**************************
 696              * CALCULATE INTERACTIONS *
 697              **************************/
 698
 699             r00              = _mm_mul_ps(rsq00,rinv00);
 700             r00              = _mm_andnot_ps(dummy_mask,r00);
 701
 702             /* EWALD ELECTROSTATICS */
 703
 704             /* Analytical PME correction */
 705             zeta2            = _mm_mul_ps(beta2,rsq00);
 706             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 707             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 708             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 709             felec            = _mm_mul_ps(qq00,felec);
 710             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 711             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 712             velec            = _mm_mul_ps(qq00,velec);
 713
 714             /* LENNARD-JONES DISPERSION/REPULSION */
 715
 716             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 717             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 718             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 719             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 720             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 721
 722             /* Update potential sum for this i atom from the interaction with this j atom. */
 723             velec            = _mm_andnot_ps(dummy_mask,velec);
 724             velecsum         = _mm_add_ps(velecsum,velec);
 725             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 726             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 727
 728             fscal            = _mm_add_ps(felec,fvdw);
 729
 730             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 731
 732              /* Update vectorial force */
 733             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 734             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 735             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 736
 737             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 738             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 739             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 740
 741             /**************************
 742              * CALCULATE INTERACTIONS *
 743              **************************/
 744
 745             r01              = _mm_mul_ps(rsq01,rinv01);
 746             r01              = _mm_andnot_ps(dummy_mask,r01);
 747
 748             /* EWALD ELECTROSTATICS */
 749
 750             /* Analytical PME correction */
 751             zeta2            = _mm_mul_ps(beta2,rsq01);
 752             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 753             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 754             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 755             felec            = _mm_mul_ps(qq01,felec);
 756             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 757             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 758             velec            = _mm_mul_ps(qq01,velec);
 759
 760             /* Update potential sum for this i atom from the interaction with this j atom. */
 761             velec            = _mm_andnot_ps(dummy_mask,velec);
 762             velecsum         = _mm_add_ps(velecsum,velec);
 763
 764             fscal            = felec;
 765
 766             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 767
 768              /* Update vectorial force */
 769             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 770             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 771             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 772
 773             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 774             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 775             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 776
 777             /**************************
 778              * CALCULATE INTERACTIONS *
 779              **************************/
 780
 781             r02              = _mm_mul_ps(rsq02,rinv02);
 782             r02              = _mm_andnot_ps(dummy_mask,r02);
 783
 784             /* EWALD ELECTROSTATICS */
 785
 786             /* Analytical PME correction */
 787             zeta2            = _mm_mul_ps(beta2,rsq02);
 788             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 789             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 790             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 791             felec            = _mm_mul_ps(qq02,felec);
 792             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 793             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 794             velec            = _mm_mul_ps(qq02,velec);
 795
 796             /* Update potential sum for this i atom from the interaction with this j atom. */
 797             velec            = _mm_andnot_ps(dummy_mask,velec);
 798             velecsum         = _mm_add_ps(velecsum,velec);
 799
 800             fscal            = felec;
 801
 802             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 803
 804              /* Update vectorial force */
 805             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 806             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 807             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 808
 809             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 810             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 811             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 812
 813             /**************************
 814              * CALCULATE INTERACTIONS *
 815              **************************/
 816
 817             r10              = _mm_mul_ps(rsq10,rinv10);
 818             r10              = _mm_andnot_ps(dummy_mask,r10);
 819
 820             /* EWALD ELECTROSTATICS */
 821
 822             /* Analytical PME correction */
 823             zeta2            = _mm_mul_ps(beta2,rsq10);
 824             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 825             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 826             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 827             felec            = _mm_mul_ps(qq10,felec);
 828             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 829             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 830             velec            = _mm_mul_ps(qq10,velec);
 831
 832             /* Update potential sum for this i atom from the interaction with this j atom. */
 833             velec            = _mm_andnot_ps(dummy_mask,velec);
 834             velecsum         = _mm_add_ps(velecsum,velec);
 835
 836             fscal            = felec;
 837
 838             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 839
 840              /* Update vectorial force */
 841             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 842             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 843             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 844
 845             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 846             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 847             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 848
 849             /**************************
 850              * CALCULATE INTERACTIONS *
 851              **************************/
 852
 853             r11              = _mm_mul_ps(rsq11,rinv11);
 854             r11              = _mm_andnot_ps(dummy_mask,r11);
 855
 856             /* EWALD ELECTROSTATICS */
 857
 858             /* Analytical PME correction */
 859             zeta2            = _mm_mul_ps(beta2,rsq11);
 860             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 861             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 862             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 863             felec            = _mm_mul_ps(qq11,felec);
 864             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 865             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 866             velec            = _mm_mul_ps(qq11,velec);
 867
 868             /* Update potential sum for this i atom from the interaction with this j atom. */
 869             velec            = _mm_andnot_ps(dummy_mask,velec);
 870             velecsum         = _mm_add_ps(velecsum,velec);
 871
 872             fscal            = felec;
 873
 874             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 875
 876              /* Update vectorial force */
 877             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 878             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 879             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 880
 881             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 882             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 883             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 884
 885             /**************************
 886              * CALCULATE INTERACTIONS *
 887              **************************/
 888
 889             r12              = _mm_mul_ps(rsq12,rinv12);
 890             r12              = _mm_andnot_ps(dummy_mask,r12);
 891
 892             /* EWALD ELECTROSTATICS */
 893
 894             /* Analytical PME correction */
 895             zeta2            = _mm_mul_ps(beta2,rsq12);
 896             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 897             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 898             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 899             felec            = _mm_mul_ps(qq12,felec);
 900             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 901             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 902             velec            = _mm_mul_ps(qq12,velec);
 903
 904             /* Update potential sum for this i atom from the interaction with this j atom. */
 905             velec            = _mm_andnot_ps(dummy_mask,velec);
 906             velecsum         = _mm_add_ps(velecsum,velec);
 907
 908             fscal            = felec;
 909
 910             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 911
 912              /* Update vectorial force */
 913             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 914             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 915             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 916
 917             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 918             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 919             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 920
 921             /**************************
 922              * CALCULATE INTERACTIONS *
 923              **************************/
 924
 925             r20              = _mm_mul_ps(rsq20,rinv20);
 926             r20              = _mm_andnot_ps(dummy_mask,r20);
 927
 928             /* EWALD ELECTROSTATICS */
 929
 930             /* Analytical PME correction */
 931             zeta2            = _mm_mul_ps(beta2,rsq20);
 932             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 933             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 934             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 935             felec            = _mm_mul_ps(qq20,felec);
 936             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 937             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 938             velec            = _mm_mul_ps(qq20,velec);
 939
 940             /* Update potential sum for this i atom from the interaction with this j atom. */
 941             velec            = _mm_andnot_ps(dummy_mask,velec);
 942             velecsum         = _mm_add_ps(velecsum,velec);
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 947
 948              /* Update vectorial force */
 949             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 950             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 951             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 952
 953             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 954             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 955             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 956
 957             /**************************
 958              * CALCULATE INTERACTIONS *
 959              **************************/
 960
 961             r21              = _mm_mul_ps(rsq21,rinv21);
 962             r21              = _mm_andnot_ps(dummy_mask,r21);
 963
 964             /* EWALD ELECTROSTATICS */
 965
 966             /* Analytical PME correction */
 967             zeta2            = _mm_mul_ps(beta2,rsq21);
 968             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 969             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 970             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 971             felec            = _mm_mul_ps(qq21,felec);
 972             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 973             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 974             velec            = _mm_mul_ps(qq21,velec);
 975
 976             /* Update potential sum for this i atom from the interaction with this j atom. */
 977             velec            = _mm_andnot_ps(dummy_mask,velec);
 978             velecsum         = _mm_add_ps(velecsum,velec);
 979
 980             fscal            = felec;
 981
 982             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 983
 984              /* Update vectorial force */
 985             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 986             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 987             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 988
 989             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 990             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 991             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 992
 993             /**************************
 994              * CALCULATE INTERACTIONS *
 995              **************************/
 996
 997             r22              = _mm_mul_ps(rsq22,rinv22);
 998             r22              = _mm_andnot_ps(dummy_mask,r22);
 999
1000             /* EWALD ELECTROSTATICS */
1001
1002             /* Analytical PME correction */
1003             zeta2            = _mm_mul_ps(beta2,rsq22);
1004             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1005             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1006             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1007             felec            = _mm_mul_ps(qq22,felec);
1008             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1009             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
1010             velec            = _mm_mul_ps(qq22,velec);
1011
1012             /* Update potential sum for this i atom from the interaction with this j atom. */
1013             velec            = _mm_andnot_ps(dummy_mask,velec);
1014             velecsum         = _mm_add_ps(velecsum,velec);
1015
1016             fscal            = felec;
1017
1018             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1019
1020              /* Update vectorial force */
1021             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1022             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1023             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1024
1025             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1026             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1027             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1028
1029             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1030             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1031             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1032             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1033
1034             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1035                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1036
1037             /* Inner loop uses 282 flops */
1038         }
1039
1040         /* End of innermost loop */
1041
1042         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1043                                               f+i_coord_offset,fshift+i_shift_offset);
1044
1045         ggid                        = gid[iidx];
1046         /* Update potential energies */
1047         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1048         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1049
1050         /* Increment number of inner iterations */
1051         inneriter                  += j_index_end - j_index_start;
1052
1053         /* Outer loop uses 20 flops */
1054     }
1055
1056     /* Increment number of outer iterations */
1057     outeriter        += nri;
1058
1059     /* Update outer/inner flops */
1060
1061     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*282);
1062 }
1063 /*
1064  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_avx_128_fma_single
1065  * Electrostatics interaction: Ewald
1066  * VdW interaction:            LennardJones
1067  * Geometry:                   Water3-Water3
1068  * Calculate force/pot:        Force
1069  */
1070 void
1071 nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_avx_128_fma_single
1072                     (t_nblist * gmx_restrict                nlist,
1073                      rvec * gmx_restrict                    xx,
1074                      rvec * gmx_restrict                    ff,
1075                      t_forcerec * gmx_restrict              fr,
1076                      t_mdatoms * gmx_restrict               mdatoms,
1077                      nb_kernel_data_t * gmx_restrict        kernel_data,
1078                      t_nrnb * gmx_restrict                  nrnb)
1079 {
1080     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1081      * just 0 for non-waters.
1082      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1083      * jnr indices corresponding to data put in the four positions in the SIMD register.
1084      */
1085     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1086     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1087     int              jnrA,jnrB,jnrC,jnrD;
1088     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1089     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1090     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1091     real             rcutoff_scalar;
1092     real             *shiftvec,*fshift,*x,*f;
1093     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1094     real             scratch[4*DIM];
1095     __m128           fscal,rcutoff,rcutoff2,jidxall;
1096     int              vdwioffset0;
1097     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1098     int              vdwioffset1;
1099     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1100     int              vdwioffset2;
1101     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1102     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1103     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1104     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1105     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1106     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1107     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1108     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1109     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1110     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1111     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1112     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1113     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1114     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1115     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1116     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1117     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1118     real             *charge;
1119     int              nvdwtype;
1120     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1121     int              *vdwtype;
1122     real             *vdwparam;
1123     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1124     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1125     __m128i          ewitab;
1126     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1127     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1128     real             *ewtab;
1129     __m128           dummy_mask,cutoff_mask;
1130     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1131     __m128           one     = _mm_set1_ps(1.0);
1132     __m128           two     = _mm_set1_ps(2.0);
1133     x                = xx[0];
1134     f                = ff[0];
1135
1136     nri              = nlist->nri;
1137     iinr             = nlist->iinr;
1138     jindex           = nlist->jindex;
1139     jjnr             = nlist->jjnr;
1140     shiftidx         = nlist->shift;
1141     gid              = nlist->gid;
1142     shiftvec         = fr->shift_vec[0];
1143     fshift           = fr->fshift[0];
1144     facel            = _mm_set1_ps(fr->epsfac);
1145     charge           = mdatoms->chargeA;
1146     nvdwtype         = fr->ntype;
1147     vdwparam         = fr->nbfp;
1148     vdwtype          = mdatoms->typeA;
1149
1150     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1151     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
1152     beta2            = _mm_mul_ps(beta,beta);
1153     beta3            = _mm_mul_ps(beta,beta2);
1154     ewtab            = fr->ic->tabq_coul_F;
1155     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1156     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1157
1158     /* Setup water-specific parameters */
1159     inr              = nlist->iinr[0];
1160     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1161     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1162     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1163     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1164
1165     jq0              = _mm_set1_ps(charge[inr+0]);
1166     jq1              = _mm_set1_ps(charge[inr+1]);
1167     jq2              = _mm_set1_ps(charge[inr+2]);
1168     vdwjidx0A        = 2*vdwtype[inr+0];
1169     qq00             = _mm_mul_ps(iq0,jq0);
1170     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1171     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1172     qq01             = _mm_mul_ps(iq0,jq1);
1173     qq02             = _mm_mul_ps(iq0,jq2);
1174     qq10             = _mm_mul_ps(iq1,jq0);
1175     qq11             = _mm_mul_ps(iq1,jq1);
1176     qq12             = _mm_mul_ps(iq1,jq2);
1177     qq20             = _mm_mul_ps(iq2,jq0);
1178     qq21             = _mm_mul_ps(iq2,jq1);
1179     qq22             = _mm_mul_ps(iq2,jq2);
1180
1181     /* Avoid stupid compiler warnings */
1182     jnrA = jnrB = jnrC = jnrD = 0;
1183     j_coord_offsetA = 0;
1184     j_coord_offsetB = 0;
1185     j_coord_offsetC = 0;
1186     j_coord_offsetD = 0;
1187
1188     outeriter        = 0;
1189     inneriter        = 0;
1190
1191     for(iidx=0;iidx<4*DIM;iidx++)
1192     {
1193         scratch[iidx] = 0.0;
1194     }
1195
1196     /* Start outer loop over neighborlists */
1197     for(iidx=0; iidx<nri; iidx++)
1198     {
1199         /* Load shift vector for this list */
1200         i_shift_offset   = DIM*shiftidx[iidx];
1201
1202         /* Load limits for loop over neighbors */
1203         j_index_start    = jindex[iidx];
1204         j_index_end      = jindex[iidx+1];
1205
1206         /* Get outer coordinate index */
1207         inr              = iinr[iidx];
1208         i_coord_offset   = DIM*inr;
1209
1210         /* Load i particle coords and add shift vector */
1211         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1212                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1213
1214         fix0             = _mm_setzero_ps();
1215         fiy0             = _mm_setzero_ps();
1216         fiz0             = _mm_setzero_ps();
1217         fix1             = _mm_setzero_ps();
1218         fiy1             = _mm_setzero_ps();
1219         fiz1             = _mm_setzero_ps();
1220         fix2             = _mm_setzero_ps();
1221         fiy2             = _mm_setzero_ps();
1222         fiz2             = _mm_setzero_ps();
1223
1224         /* Start inner kernel loop */
1225         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1226         {
1227
1228             /* Get j neighbor index, and coordinate index */
1229             jnrA             = jjnr[jidx];
1230             jnrB             = jjnr[jidx+1];
1231             jnrC             = jjnr[jidx+2];
1232             jnrD             = jjnr[jidx+3];
1233             j_coord_offsetA  = DIM*jnrA;
1234             j_coord_offsetB  = DIM*jnrB;
1235             j_coord_offsetC  = DIM*jnrC;
1236             j_coord_offsetD  = DIM*jnrD;
1237
1238             /* load j atom coordinates */
1239             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1240                                               x+j_coord_offsetC,x+j_coord_offsetD,
1241                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1242
1243             /* Calculate displacement vector */
1244             dx00             = _mm_sub_ps(ix0,jx0);
1245             dy00             = _mm_sub_ps(iy0,jy0);
1246             dz00             = _mm_sub_ps(iz0,jz0);
1247             dx01             = _mm_sub_ps(ix0,jx1);
1248             dy01             = _mm_sub_ps(iy0,jy1);
1249             dz01             = _mm_sub_ps(iz0,jz1);
1250             dx02             = _mm_sub_ps(ix0,jx2);
1251             dy02             = _mm_sub_ps(iy0,jy2);
1252             dz02             = _mm_sub_ps(iz0,jz2);
1253             dx10             = _mm_sub_ps(ix1,jx0);
1254             dy10             = _mm_sub_ps(iy1,jy0);
1255             dz10             = _mm_sub_ps(iz1,jz0);
1256             dx11             = _mm_sub_ps(ix1,jx1);
1257             dy11             = _mm_sub_ps(iy1,jy1);
1258             dz11             = _mm_sub_ps(iz1,jz1);
1259             dx12             = _mm_sub_ps(ix1,jx2);
1260             dy12             = _mm_sub_ps(iy1,jy2);
1261             dz12             = _mm_sub_ps(iz1,jz2);
1262             dx20             = _mm_sub_ps(ix2,jx0);
1263             dy20             = _mm_sub_ps(iy2,jy0);
1264             dz20             = _mm_sub_ps(iz2,jz0);
1265             dx21             = _mm_sub_ps(ix2,jx1);
1266             dy21             = _mm_sub_ps(iy2,jy1);
1267             dz21             = _mm_sub_ps(iz2,jz1);
1268             dx22             = _mm_sub_ps(ix2,jx2);
1269             dy22             = _mm_sub_ps(iy2,jy2);
1270             dz22             = _mm_sub_ps(iz2,jz2);
1271
1272             /* Calculate squared distance and things based on it */
1273             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1274             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1275             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1276             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1277             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1278             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1279             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1280             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1281             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1282
1283             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1284             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1285             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1286             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1287             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1288             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1289             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1290             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1291             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1292
1293             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1294             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1295             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1296             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1297             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1298             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1299             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1300             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1301             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1302
1303             fjx0             = _mm_setzero_ps();
1304             fjy0             = _mm_setzero_ps();
1305             fjz0             = _mm_setzero_ps();
1306             fjx1             = _mm_setzero_ps();
1307             fjy1             = _mm_setzero_ps();
1308             fjz1             = _mm_setzero_ps();
1309             fjx2             = _mm_setzero_ps();
1310             fjy2             = _mm_setzero_ps();
1311             fjz2             = _mm_setzero_ps();
1312
1313             /**************************
1314              * CALCULATE INTERACTIONS *
1315              **************************/
1316
1317             r00              = _mm_mul_ps(rsq00,rinv00);
1318
1319             /* EWALD ELECTROSTATICS */
1320
1321             /* Analytical PME correction */
1322             zeta2            = _mm_mul_ps(beta2,rsq00);
1323             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1324             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1325             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1326             felec            = _mm_mul_ps(qq00,felec);
1327
1328             /* LENNARD-JONES DISPERSION/REPULSION */
1329
1330             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1331             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1332
1333             fscal            = _mm_add_ps(felec,fvdw);
1334
1335              /* Update vectorial force */
1336             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1337             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1338             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1339
1340             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1341             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1342             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             r01              = _mm_mul_ps(rsq01,rinv01);
1349
1350             /* EWALD ELECTROSTATICS */
1351
1352             /* Analytical PME correction */
1353             zeta2            = _mm_mul_ps(beta2,rsq01);
1354             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1355             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1356             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1357             felec            = _mm_mul_ps(qq01,felec);
1358
1359             fscal            = felec;
1360
1361              /* Update vectorial force */
1362             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1363             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1364             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1365
1366             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1367             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1368             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1369
1370             /**************************
1371              * CALCULATE INTERACTIONS *
1372              **************************/
1373
1374             r02              = _mm_mul_ps(rsq02,rinv02);
1375
1376             /* EWALD ELECTROSTATICS */
1377
1378             /* Analytical PME correction */
1379             zeta2            = _mm_mul_ps(beta2,rsq02);
1380             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1381             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1382             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1383             felec            = _mm_mul_ps(qq02,felec);
1384
1385             fscal            = felec;
1386
1387              /* Update vectorial force */
1388             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1389             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1390             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1391
1392             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1393             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1394             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1395
1396             /**************************
1397              * CALCULATE INTERACTIONS *
1398              **************************/
1399
1400             r10              = _mm_mul_ps(rsq10,rinv10);
1401
1402             /* EWALD ELECTROSTATICS */
1403
1404             /* Analytical PME correction */
1405             zeta2            = _mm_mul_ps(beta2,rsq10);
1406             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1407             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1408             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1409             felec            = _mm_mul_ps(qq10,felec);
1410
1411             fscal            = felec;
1412
1413              /* Update vectorial force */
1414             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1415             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1416             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1417
1418             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1419             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1420             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1421
1422             /**************************
1423              * CALCULATE INTERACTIONS *
1424              **************************/
1425
1426             r11              = _mm_mul_ps(rsq11,rinv11);
1427
1428             /* EWALD ELECTROSTATICS */
1429
1430             /* Analytical PME correction */
1431             zeta2            = _mm_mul_ps(beta2,rsq11);
1432             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1433             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1434             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1435             felec            = _mm_mul_ps(qq11,felec);
1436
1437             fscal            = felec;
1438
1439              /* Update vectorial force */
1440             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1441             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1442             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1443
1444             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1445             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1446             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1447
1448             /**************************
1449              * CALCULATE INTERACTIONS *
1450              **************************/
1451
1452             r12              = _mm_mul_ps(rsq12,rinv12);
1453
1454             /* EWALD ELECTROSTATICS */
1455
1456             /* Analytical PME correction */
1457             zeta2            = _mm_mul_ps(beta2,rsq12);
1458             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1459             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1460             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1461             felec            = _mm_mul_ps(qq12,felec);
1462
1463             fscal            = felec;
1464
1465              /* Update vectorial force */
1466             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1467             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1468             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1469
1470             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1471             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1472             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1473
1474             /**************************
1475              * CALCULATE INTERACTIONS *
1476              **************************/
1477
1478             r20              = _mm_mul_ps(rsq20,rinv20);
1479
1480             /* EWALD ELECTROSTATICS */
1481
1482             /* Analytical PME correction */
1483             zeta2            = _mm_mul_ps(beta2,rsq20);
1484             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1485             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1486             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1487             felec            = _mm_mul_ps(qq20,felec);
1488
1489             fscal            = felec;
1490
1491              /* Update vectorial force */
1492             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1493             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1494             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1495
1496             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1497             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1498             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1499
1500             /**************************
1501              * CALCULATE INTERACTIONS *
1502              **************************/
1503
1504             r21              = _mm_mul_ps(rsq21,rinv21);
1505
1506             /* EWALD ELECTROSTATICS */
1507
1508             /* Analytical PME correction */
1509             zeta2            = _mm_mul_ps(beta2,rsq21);
1510             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1511             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1512             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1513             felec            = _mm_mul_ps(qq21,felec);
1514
1515             fscal            = felec;
1516
1517              /* Update vectorial force */
1518             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1519             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1520             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1521
1522             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1523             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1524             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1525
1526             /**************************
1527              * CALCULATE INTERACTIONS *
1528              **************************/
1529
1530             r22              = _mm_mul_ps(rsq22,rinv22);
1531
1532             /* EWALD ELECTROSTATICS */
1533
1534             /* Analytical PME correction */
1535             zeta2            = _mm_mul_ps(beta2,rsq22);
1536             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1537             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1538             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1539             felec            = _mm_mul_ps(qq22,felec);
1540
1541             fscal            = felec;
1542
1543              /* Update vectorial force */
1544             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1545             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1546             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1547
1548             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1549             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1550             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1551
1552             fjptrA             = f+j_coord_offsetA;
1553             fjptrB             = f+j_coord_offsetB;
1554             fjptrC             = f+j_coord_offsetC;
1555             fjptrD             = f+j_coord_offsetD;
1556
1557             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1558                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1559
1560             /* Inner loop uses 259 flops */
1561         }
1562
1563         if(jidx<j_index_end)
1564         {
1565
1566             /* Get j neighbor index, and coordinate index */
1567             jnrlistA         = jjnr[jidx];
1568             jnrlistB         = jjnr[jidx+1];
1569             jnrlistC         = jjnr[jidx+2];
1570             jnrlistD         = jjnr[jidx+3];
1571             /* Sign of each element will be negative for non-real atoms.
1572              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1573              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1574              */
1575             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1576             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1577             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1578             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1579             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1580             j_coord_offsetA  = DIM*jnrA;
1581             j_coord_offsetB  = DIM*jnrB;
1582             j_coord_offsetC  = DIM*jnrC;
1583             j_coord_offsetD  = DIM*jnrD;
1584
1585             /* load j atom coordinates */
1586             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1587                                               x+j_coord_offsetC,x+j_coord_offsetD,
1588                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1589
1590             /* Calculate displacement vector */
1591             dx00             = _mm_sub_ps(ix0,jx0);
1592             dy00             = _mm_sub_ps(iy0,jy0);
1593             dz00             = _mm_sub_ps(iz0,jz0);
1594             dx01             = _mm_sub_ps(ix0,jx1);
1595             dy01             = _mm_sub_ps(iy0,jy1);
1596             dz01             = _mm_sub_ps(iz0,jz1);
1597             dx02             = _mm_sub_ps(ix0,jx2);
1598             dy02             = _mm_sub_ps(iy0,jy2);
1599             dz02             = _mm_sub_ps(iz0,jz2);
1600             dx10             = _mm_sub_ps(ix1,jx0);
1601             dy10             = _mm_sub_ps(iy1,jy0);
1602             dz10             = _mm_sub_ps(iz1,jz0);
1603             dx11             = _mm_sub_ps(ix1,jx1);
1604             dy11             = _mm_sub_ps(iy1,jy1);
1605             dz11             = _mm_sub_ps(iz1,jz1);
1606             dx12             = _mm_sub_ps(ix1,jx2);
1607             dy12             = _mm_sub_ps(iy1,jy2);
1608             dz12             = _mm_sub_ps(iz1,jz2);
1609             dx20             = _mm_sub_ps(ix2,jx0);
1610             dy20             = _mm_sub_ps(iy2,jy0);
1611             dz20             = _mm_sub_ps(iz2,jz0);
1612             dx21             = _mm_sub_ps(ix2,jx1);
1613             dy21             = _mm_sub_ps(iy2,jy1);
1614             dz21             = _mm_sub_ps(iz2,jz1);
1615             dx22             = _mm_sub_ps(ix2,jx2);
1616             dy22             = _mm_sub_ps(iy2,jy2);
1617             dz22             = _mm_sub_ps(iz2,jz2);
1618
1619             /* Calculate squared distance and things based on it */
1620             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1621             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1622             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1623             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1624             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1625             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1626             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1627             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1628             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1629
1630             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1631             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1632             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1633             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1634             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1635             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1636             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1637             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1638             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1639
1640             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1641             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1642             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1643             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1644             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1645             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1646             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1647             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1648             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1649
1650             fjx0             = _mm_setzero_ps();
1651             fjy0             = _mm_setzero_ps();
1652             fjz0             = _mm_setzero_ps();
1653             fjx1             = _mm_setzero_ps();
1654             fjy1             = _mm_setzero_ps();
1655             fjz1             = _mm_setzero_ps();
1656             fjx2             = _mm_setzero_ps();
1657             fjy2             = _mm_setzero_ps();
1658             fjz2             = _mm_setzero_ps();
1659
1660             /**************************
1661              * CALCULATE INTERACTIONS *
1662              **************************/
1663
1664             r00              = _mm_mul_ps(rsq00,rinv00);
1665             r00              = _mm_andnot_ps(dummy_mask,r00);
1666
1667             /* EWALD ELECTROSTATICS */
1668
1669             /* Analytical PME correction */
1670             zeta2            = _mm_mul_ps(beta2,rsq00);
1671             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1672             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1673             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1674             felec            = _mm_mul_ps(qq00,felec);
1675
1676             /* LENNARD-JONES DISPERSION/REPULSION */
1677
1678             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1679             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1680
1681             fscal            = _mm_add_ps(felec,fvdw);
1682
1683             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1684
1685              /* Update vectorial force */
1686             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1687             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1688             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1689
1690             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1691             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1692             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1693
1694             /**************************
1695              * CALCULATE INTERACTIONS *
1696              **************************/
1697
1698             r01              = _mm_mul_ps(rsq01,rinv01);
1699             r01              = _mm_andnot_ps(dummy_mask,r01);
1700
1701             /* EWALD ELECTROSTATICS */
1702
1703             /* Analytical PME correction */
1704             zeta2            = _mm_mul_ps(beta2,rsq01);
1705             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1706             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1707             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1708             felec            = _mm_mul_ps(qq01,felec);
1709
1710             fscal            = felec;
1711
1712             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1713
1714              /* Update vectorial force */
1715             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1716             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1717             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1718
1719             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1720             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1721             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1722
1723             /**************************
1724              * CALCULATE INTERACTIONS *
1725              **************************/
1726
1727             r02              = _mm_mul_ps(rsq02,rinv02);
1728             r02              = _mm_andnot_ps(dummy_mask,r02);
1729
1730             /* EWALD ELECTROSTATICS */
1731
1732             /* Analytical PME correction */
1733             zeta2            = _mm_mul_ps(beta2,rsq02);
1734             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1735             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1736             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1737             felec            = _mm_mul_ps(qq02,felec);
1738
1739             fscal            = felec;
1740
1741             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1742
1743              /* Update vectorial force */
1744             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1745             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1746             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1747
1748             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1749             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1750             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1751
1752             /**************************
1753              * CALCULATE INTERACTIONS *
1754              **************************/
1755
1756             r10              = _mm_mul_ps(rsq10,rinv10);
1757             r10              = _mm_andnot_ps(dummy_mask,r10);
1758
1759             /* EWALD ELECTROSTATICS */
1760
1761             /* Analytical PME correction */
1762             zeta2            = _mm_mul_ps(beta2,rsq10);
1763             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1764             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1765             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1766             felec            = _mm_mul_ps(qq10,felec);
1767
1768             fscal            = felec;
1769
1770             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1771
1772              /* Update vectorial force */
1773             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1774             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1775             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1776
1777             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1778             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1779             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1780
1781             /**************************
1782              * CALCULATE INTERACTIONS *
1783              **************************/
1784
1785             r11              = _mm_mul_ps(rsq11,rinv11);
1786             r11              = _mm_andnot_ps(dummy_mask,r11);
1787
1788             /* EWALD ELECTROSTATICS */
1789
1790             /* Analytical PME correction */
1791             zeta2            = _mm_mul_ps(beta2,rsq11);
1792             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1793             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1794             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1795             felec            = _mm_mul_ps(qq11,felec);
1796
1797             fscal            = felec;
1798
1799             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1800
1801              /* Update vectorial force */
1802             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1803             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1804             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1805
1806             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1807             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1808             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1809
1810             /**************************
1811              * CALCULATE INTERACTIONS *
1812              **************************/
1813
1814             r12              = _mm_mul_ps(rsq12,rinv12);
1815             r12              = _mm_andnot_ps(dummy_mask,r12);
1816
1817             /* EWALD ELECTROSTATICS */
1818
1819             /* Analytical PME correction */
1820             zeta2            = _mm_mul_ps(beta2,rsq12);
1821             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1822             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1823             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1824             felec            = _mm_mul_ps(qq12,felec);
1825
1826             fscal            = felec;
1827
1828             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1829
1830              /* Update vectorial force */
1831             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1832             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1833             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1834
1835             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1836             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1837             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1838
1839             /**************************
1840              * CALCULATE INTERACTIONS *
1841              **************************/
1842
1843             r20              = _mm_mul_ps(rsq20,rinv20);
1844             r20              = _mm_andnot_ps(dummy_mask,r20);
1845
1846             /* EWALD ELECTROSTATICS */
1847
1848             /* Analytical PME correction */
1849             zeta2            = _mm_mul_ps(beta2,rsq20);
1850             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1851             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1852             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1853             felec            = _mm_mul_ps(qq20,felec);
1854
1855             fscal            = felec;
1856
1857             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1858
1859              /* Update vectorial force */
1860             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1861             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1862             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1863
1864             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1865             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1866             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1867
1868             /**************************
1869              * CALCULATE INTERACTIONS *
1870              **************************/
1871
1872             r21              = _mm_mul_ps(rsq21,rinv21);
1873             r21              = _mm_andnot_ps(dummy_mask,r21);
1874
1875             /* EWALD ELECTROSTATICS */
1876
1877             /* Analytical PME correction */
1878             zeta2            = _mm_mul_ps(beta2,rsq21);
1879             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1880             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1881             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1882             felec            = _mm_mul_ps(qq21,felec);
1883
1884             fscal            = felec;
1885
1886             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1887
1888              /* Update vectorial force */
1889             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1890             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1891             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1892
1893             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1894             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1895             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1896
1897             /**************************
1898              * CALCULATE INTERACTIONS *
1899              **************************/
1900
1901             r22              = _mm_mul_ps(rsq22,rinv22);
1902             r22              = _mm_andnot_ps(dummy_mask,r22);
1903
1904             /* EWALD ELECTROSTATICS */
1905
1906             /* Analytical PME correction */
1907             zeta2            = _mm_mul_ps(beta2,rsq22);
1908             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1909             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1910             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1911             felec            = _mm_mul_ps(qq22,felec);
1912
1913             fscal            = felec;
1914
1915             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1916
1917              /* Update vectorial force */
1918             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1919             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1920             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1921
1922             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1923             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1924             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1925
1926             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1927             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1928             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1929             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1930
1931             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1932                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1933
1934             /* Inner loop uses 268 flops */
1935         }
1936
1937         /* End of innermost loop */
1938
1939         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1940                                               f+i_coord_offset,fshift+i_shift_offset);
1941
1942         /* Increment number of inner iterations */
1943         inneriter                  += j_index_end - j_index_start;
1944
1945         /* Outer loop uses 18 flops */
1946     }
1947
1948     /* Increment number of outer iterations */
1949     outeriter        += nri;
1950
1951     /* Update outer/inner flops */
1952
1953     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*268);
1954 }