src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwNone_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_VdwNone_GeomW3W3_VF_avx_128_fma_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            None
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_VdwNone_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     __m128i          ewitab;
  93     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  94     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  95     real             *ewtab;
  96     __m128           dummy_mask,cutoff_mask;
  97     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  98     __m128           one     = _mm_set1_ps(1.0);
  99     __m128           two     = _mm_set1_ps(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = _mm_set1_ps(fr->epsfac);
 112     charge           = mdatoms->chargeA;
 113
 114     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 115     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 116     beta2            = _mm_mul_ps(beta,beta);
 117     beta3            = _mm_mul_ps(beta,beta2);
 118     ewtab            = fr->ic->tabq_coul_FDV0;
 119     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 120     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 121
 122     /* Setup water-specific parameters */
 123     inr              = nlist->iinr[0];
 124     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 125     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 126     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 127
 128     jq0              = _mm_set1_ps(charge[inr+0]);
 129     jq1              = _mm_set1_ps(charge[inr+1]);
 130     jq2              = _mm_set1_ps(charge[inr+2]);
 131     qq00             = _mm_mul_ps(iq0,jq0);
 132     qq01             = _mm_mul_ps(iq0,jq1);
 133     qq02             = _mm_mul_ps(iq0,jq2);
 134     qq10             = _mm_mul_ps(iq1,jq0);
 135     qq11             = _mm_mul_ps(iq1,jq1);
 136     qq12             = _mm_mul_ps(iq1,jq2);
 137     qq20             = _mm_mul_ps(iq2,jq0);
 138     qq21             = _mm_mul_ps(iq2,jq1);
 139     qq22             = _mm_mul_ps(iq2,jq2);
 140
 141     /* Avoid stupid compiler warnings */
 142     jnrA = jnrB = jnrC = jnrD = 0;
 143     j_coord_offsetA = 0;
 144     j_coord_offsetB = 0;
 145     j_coord_offsetC = 0;
 146     j_coord_offsetD = 0;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     for(iidx=0;iidx<4*DIM;iidx++)
 152     {
 153         scratch[iidx] = 0.0;
 154     }
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 172                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 173
 174         fix0             = _mm_setzero_ps();
 175         fiy0             = _mm_setzero_ps();
 176         fiz0             = _mm_setzero_ps();
 177         fix1             = _mm_setzero_ps();
 178         fiy1             = _mm_setzero_ps();
 179         fiz1             = _mm_setzero_ps();
 180         fix2             = _mm_setzero_ps();
 181         fiy2             = _mm_setzero_ps();
 182         fiz2             = _mm_setzero_ps();
 183
 184         /* Reset potential sums */
 185         velecsum         = _mm_setzero_ps();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             jnrC             = jjnr[jidx+2];
 195             jnrD             = jjnr[jidx+3];
 196             j_coord_offsetA  = DIM*jnrA;
 197             j_coord_offsetB  = DIM*jnrB;
 198             j_coord_offsetC  = DIM*jnrC;
 199             j_coord_offsetD  = DIM*jnrD;
 200
 201             /* load j atom coordinates */
 202             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 203                                               x+j_coord_offsetC,x+j_coord_offsetD,
 204                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 205
 206             /* Calculate displacement vector */
 207             dx00             = _mm_sub_ps(ix0,jx0);
 208             dy00             = _mm_sub_ps(iy0,jy0);
 209             dz00             = _mm_sub_ps(iz0,jz0);
 210             dx01             = _mm_sub_ps(ix0,jx1);
 211             dy01             = _mm_sub_ps(iy0,jy1);
 212             dz01             = _mm_sub_ps(iz0,jz1);
 213             dx02             = _mm_sub_ps(ix0,jx2);
 214             dy02             = _mm_sub_ps(iy0,jy2);
 215             dz02             = _mm_sub_ps(iz0,jz2);
 216             dx10             = _mm_sub_ps(ix1,jx0);
 217             dy10             = _mm_sub_ps(iy1,jy0);
 218             dz10             = _mm_sub_ps(iz1,jz0);
 219             dx11             = _mm_sub_ps(ix1,jx1);
 220             dy11             = _mm_sub_ps(iy1,jy1);
 221             dz11             = _mm_sub_ps(iz1,jz1);
 222             dx12             = _mm_sub_ps(ix1,jx2);
 223             dy12             = _mm_sub_ps(iy1,jy2);
 224             dz12             = _mm_sub_ps(iz1,jz2);
 225             dx20             = _mm_sub_ps(ix2,jx0);
 226             dy20             = _mm_sub_ps(iy2,jy0);
 227             dz20             = _mm_sub_ps(iz2,jz0);
 228             dx21             = _mm_sub_ps(ix2,jx1);
 229             dy21             = _mm_sub_ps(iy2,jy1);
 230             dz21             = _mm_sub_ps(iz2,jz1);
 231             dx22             = _mm_sub_ps(ix2,jx2);
 232             dy22             = _mm_sub_ps(iy2,jy2);
 233             dz22             = _mm_sub_ps(iz2,jz2);
 234
 235             /* Calculate squared distance and things based on it */
 236             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 237             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 238             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 239             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 240             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 241             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 242             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 243             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 244             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 245
 246             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 247             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 248             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 249             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 250             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 251             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 252             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 253             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 254             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 255
 256             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 257             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 258             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 259             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 260             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 261             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 262             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 263             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 264             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 265
 266             fjx0             = _mm_setzero_ps();
 267             fjy0             = _mm_setzero_ps();
 268             fjz0             = _mm_setzero_ps();
 269             fjx1             = _mm_setzero_ps();
 270             fjy1             = _mm_setzero_ps();
 271             fjz1             = _mm_setzero_ps();
 272             fjx2             = _mm_setzero_ps();
 273             fjy2             = _mm_setzero_ps();
 274             fjz2             = _mm_setzero_ps();
 275
 276             /**************************
 277              * CALCULATE INTERACTIONS *
 278              **************************/
 279
 280             r00              = _mm_mul_ps(rsq00,rinv00);
 281
 282             /* EWALD ELECTROSTATICS */
 283
 284             /* Analytical PME correction */
 285             zeta2            = _mm_mul_ps(beta2,rsq00);
 286             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 287             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 288             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 289             felec            = _mm_mul_ps(qq00,felec);
 290             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 291             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 292             velec            = _mm_mul_ps(qq00,velec);
 293
 294             /* Update potential sum for this i atom from the interaction with this j atom. */
 295             velecsum         = _mm_add_ps(velecsum,velec);
 296
 297             fscal            = felec;
 298
 299              /* Update vectorial force */
 300             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 301             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 302             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 303
 304             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 305             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 306             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             r01              = _mm_mul_ps(rsq01,rinv01);
 313
 314             /* EWALD ELECTROSTATICS */
 315
 316             /* Analytical PME correction */
 317             zeta2            = _mm_mul_ps(beta2,rsq01);
 318             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 319             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 320             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 321             felec            = _mm_mul_ps(qq01,felec);
 322             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 323             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 324             velec            = _mm_mul_ps(qq01,velec);
 325
 326             /* Update potential sum for this i atom from the interaction with this j atom. */
 327             velecsum         = _mm_add_ps(velecsum,velec);
 328
 329             fscal            = felec;
 330
 331              /* Update vectorial force */
 332             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 333             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 334             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 335
 336             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 337             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 338             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 339
 340             /**************************
 341              * CALCULATE INTERACTIONS *
 342              **************************/
 343
 344             r02              = _mm_mul_ps(rsq02,rinv02);
 345
 346             /* EWALD ELECTROSTATICS */
 347
 348             /* Analytical PME correction */
 349             zeta2            = _mm_mul_ps(beta2,rsq02);
 350             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 351             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 352             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 353             felec            = _mm_mul_ps(qq02,felec);
 354             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 355             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 356             velec            = _mm_mul_ps(qq02,velec);
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velecsum         = _mm_add_ps(velecsum,velec);
 360
 361             fscal            = felec;
 362
 363              /* Update vectorial force */
 364             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 365             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 366             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 367
 368             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 369             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 370             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             r10              = _mm_mul_ps(rsq10,rinv10);
 377
 378             /* EWALD ELECTROSTATICS */
 379
 380             /* Analytical PME correction */
 381             zeta2            = _mm_mul_ps(beta2,rsq10);
 382             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 383             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 384             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 385             felec            = _mm_mul_ps(qq10,felec);
 386             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 387             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 388             velec            = _mm_mul_ps(qq10,velec);
 389
 390             /* Update potential sum for this i atom from the interaction with this j atom. */
 391             velecsum         = _mm_add_ps(velecsum,velec);
 392
 393             fscal            = felec;
 394
 395              /* Update vectorial force */
 396             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 397             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 398             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 399
 400             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 401             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 402             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 403
 404             /**************************
 405              * CALCULATE INTERACTIONS *
 406              **************************/
 407
 408             r11              = _mm_mul_ps(rsq11,rinv11);
 409
 410             /* EWALD ELECTROSTATICS */
 411
 412             /* Analytical PME correction */
 413             zeta2            = _mm_mul_ps(beta2,rsq11);
 414             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 415             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 416             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 417             felec            = _mm_mul_ps(qq11,felec);
 418             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 419             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 420             velec            = _mm_mul_ps(qq11,velec);
 421
 422             /* Update potential sum for this i atom from the interaction with this j atom. */
 423             velecsum         = _mm_add_ps(velecsum,velec);
 424
 425             fscal            = felec;
 426
 427              /* Update vectorial force */
 428             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 429             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 430             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 431
 432             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 433             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 434             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 435
 436             /**************************
 437              * CALCULATE INTERACTIONS *
 438              **************************/
 439
 440             r12              = _mm_mul_ps(rsq12,rinv12);
 441
 442             /* EWALD ELECTROSTATICS */
 443
 444             /* Analytical PME correction */
 445             zeta2            = _mm_mul_ps(beta2,rsq12);
 446             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 447             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 448             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 449             felec            = _mm_mul_ps(qq12,felec);
 450             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 451             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 452             velec            = _mm_mul_ps(qq12,velec);
 453
 454             /* Update potential sum for this i atom from the interaction with this j atom. */
 455             velecsum         = _mm_add_ps(velecsum,velec);
 456
 457             fscal            = felec;
 458
 459              /* Update vectorial force */
 460             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 461             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 462             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 463
 464             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 465             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 466             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 467
 468             /**************************
 469              * CALCULATE INTERACTIONS *
 470              **************************/
 471
 472             r20              = _mm_mul_ps(rsq20,rinv20);
 473
 474             /* EWALD ELECTROSTATICS */
 475
 476             /* Analytical PME correction */
 477             zeta2            = _mm_mul_ps(beta2,rsq20);
 478             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 479             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 480             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 481             felec            = _mm_mul_ps(qq20,felec);
 482             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 483             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 484             velec            = _mm_mul_ps(qq20,velec);
 485
 486             /* Update potential sum for this i atom from the interaction with this j atom. */
 487             velecsum         = _mm_add_ps(velecsum,velec);
 488
 489             fscal            = felec;
 490
 491              /* Update vectorial force */
 492             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 493             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 494             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 495
 496             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 497             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 498             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 499
 500             /**************************
 501              * CALCULATE INTERACTIONS *
 502              **************************/
 503
 504             r21              = _mm_mul_ps(rsq21,rinv21);
 505
 506             /* EWALD ELECTROSTATICS */
 507
 508             /* Analytical PME correction */
 509             zeta2            = _mm_mul_ps(beta2,rsq21);
 510             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 511             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 512             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 513             felec            = _mm_mul_ps(qq21,felec);
 514             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 515             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 516             velec            = _mm_mul_ps(qq21,velec);
 517
 518             /* Update potential sum for this i atom from the interaction with this j atom. */
 519             velecsum         = _mm_add_ps(velecsum,velec);
 520
 521             fscal            = felec;
 522
 523              /* Update vectorial force */
 524             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 525             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 526             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 527
 528             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 529             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 530             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 531
 532             /**************************
 533              * CALCULATE INTERACTIONS *
 534              **************************/
 535
 536             r22              = _mm_mul_ps(rsq22,rinv22);
 537
 538             /* EWALD ELECTROSTATICS */
 539
 540             /* Analytical PME correction */
 541             zeta2            = _mm_mul_ps(beta2,rsq22);
 542             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 543             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 544             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 545             felec            = _mm_mul_ps(qq22,felec);
 546             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 547             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 548             velec            = _mm_mul_ps(qq22,velec);
 549
 550             /* Update potential sum for this i atom from the interaction with this j atom. */
 551             velecsum         = _mm_add_ps(velecsum,velec);
 552
 553             fscal            = felec;
 554
 555              /* Update vectorial force */
 556             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 557             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 558             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 559
 560             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 561             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 562             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 563
 564             fjptrA             = f+j_coord_offsetA;
 565             fjptrB             = f+j_coord_offsetB;
 566             fjptrC             = f+j_coord_offsetC;
 567             fjptrD             = f+j_coord_offsetD;
 568
 569             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 570                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 571
 572             /* Inner loop uses 261 flops */
 573         }
 574
 575         if(jidx<j_index_end)
 576         {
 577
 578             /* Get j neighbor index, and coordinate index */
 579             jnrlistA         = jjnr[jidx];
 580             jnrlistB         = jjnr[jidx+1];
 581             jnrlistC         = jjnr[jidx+2];
 582             jnrlistD         = jjnr[jidx+3];
 583             /* Sign of each element will be negative for non-real atoms.
 584              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 585              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 586              */
 587             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 588             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 589             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 590             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 591             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 592             j_coord_offsetA  = DIM*jnrA;
 593             j_coord_offsetB  = DIM*jnrB;
 594             j_coord_offsetC  = DIM*jnrC;
 595             j_coord_offsetD  = DIM*jnrD;
 596
 597             /* load j atom coordinates */
 598             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 599                                               x+j_coord_offsetC,x+j_coord_offsetD,
 600                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 601
 602             /* Calculate displacement vector */
 603             dx00             = _mm_sub_ps(ix0,jx0);
 604             dy00             = _mm_sub_ps(iy0,jy0);
 605             dz00             = _mm_sub_ps(iz0,jz0);
 606             dx01             = _mm_sub_ps(ix0,jx1);
 607             dy01             = _mm_sub_ps(iy0,jy1);
 608             dz01             = _mm_sub_ps(iz0,jz1);
 609             dx02             = _mm_sub_ps(ix0,jx2);
 610             dy02             = _mm_sub_ps(iy0,jy2);
 611             dz02             = _mm_sub_ps(iz0,jz2);
 612             dx10             = _mm_sub_ps(ix1,jx0);
 613             dy10             = _mm_sub_ps(iy1,jy0);
 614             dz10             = _mm_sub_ps(iz1,jz0);
 615             dx11             = _mm_sub_ps(ix1,jx1);
 616             dy11             = _mm_sub_ps(iy1,jy1);
 617             dz11             = _mm_sub_ps(iz1,jz1);
 618             dx12             = _mm_sub_ps(ix1,jx2);
 619             dy12             = _mm_sub_ps(iy1,jy2);
 620             dz12             = _mm_sub_ps(iz1,jz2);
 621             dx20             = _mm_sub_ps(ix2,jx0);
 622             dy20             = _mm_sub_ps(iy2,jy0);
 623             dz20             = _mm_sub_ps(iz2,jz0);
 624             dx21             = _mm_sub_ps(ix2,jx1);
 625             dy21             = _mm_sub_ps(iy2,jy1);
 626             dz21             = _mm_sub_ps(iz2,jz1);
 627             dx22             = _mm_sub_ps(ix2,jx2);
 628             dy22             = _mm_sub_ps(iy2,jy2);
 629             dz22             = _mm_sub_ps(iz2,jz2);
 630
 631             /* Calculate squared distance and things based on it */
 632             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 633             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 634             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 635             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 636             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 637             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 638             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 639             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 640             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 641
 642             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 643             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 644             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 645             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 646             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 647             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 648             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 649             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 650             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 651
 652             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 653             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 654             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 655             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 656             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 657             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 658             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 659             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 660             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 661
 662             fjx0             = _mm_setzero_ps();
 663             fjy0             = _mm_setzero_ps();
 664             fjz0             = _mm_setzero_ps();
 665             fjx1             = _mm_setzero_ps();
 666             fjy1             = _mm_setzero_ps();
 667             fjz1             = _mm_setzero_ps();
 668             fjx2             = _mm_setzero_ps();
 669             fjy2             = _mm_setzero_ps();
 670             fjz2             = _mm_setzero_ps();
 671
 672             /**************************
 673              * CALCULATE INTERACTIONS *
 674              **************************/
 675
 676             r00              = _mm_mul_ps(rsq00,rinv00);
 677             r00              = _mm_andnot_ps(dummy_mask,r00);
 678
 679             /* EWALD ELECTROSTATICS */
 680
 681             /* Analytical PME correction */
 682             zeta2            = _mm_mul_ps(beta2,rsq00);
 683             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 684             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 685             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 686             felec            = _mm_mul_ps(qq00,felec);
 687             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 688             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 689             velec            = _mm_mul_ps(qq00,velec);
 690
 691             /* Update potential sum for this i atom from the interaction with this j atom. */
 692             velec            = _mm_andnot_ps(dummy_mask,velec);
 693             velecsum         = _mm_add_ps(velecsum,velec);
 694
 695             fscal            = felec;
 696
 697             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 698
 699              /* Update vectorial force */
 700             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 701             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 702             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 703
 704             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 705             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 706             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 707
 708             /**************************
 709              * CALCULATE INTERACTIONS *
 710              **************************/
 711
 712             r01              = _mm_mul_ps(rsq01,rinv01);
 713             r01              = _mm_andnot_ps(dummy_mask,r01);
 714
 715             /* EWALD ELECTROSTATICS */
 716
 717             /* Analytical PME correction */
 718             zeta2            = _mm_mul_ps(beta2,rsq01);
 719             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 720             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 721             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 722             felec            = _mm_mul_ps(qq01,felec);
 723             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 724             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 725             velec            = _mm_mul_ps(qq01,velec);
 726
 727             /* Update potential sum for this i atom from the interaction with this j atom. */
 728             velec            = _mm_andnot_ps(dummy_mask,velec);
 729             velecsum         = _mm_add_ps(velecsum,velec);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 734
 735              /* Update vectorial force */
 736             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 737             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 738             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 739
 740             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 741             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 742             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 743
 744             /**************************
 745              * CALCULATE INTERACTIONS *
 746              **************************/
 747
 748             r02              = _mm_mul_ps(rsq02,rinv02);
 749             r02              = _mm_andnot_ps(dummy_mask,r02);
 750
 751             /* EWALD ELECTROSTATICS */
 752
 753             /* Analytical PME correction */
 754             zeta2            = _mm_mul_ps(beta2,rsq02);
 755             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 756             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 757             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 758             felec            = _mm_mul_ps(qq02,felec);
 759             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 760             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 761             velec            = _mm_mul_ps(qq02,velec);
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _mm_andnot_ps(dummy_mask,velec);
 765             velecsum         = _mm_add_ps(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 770
 771              /* Update vectorial force */
 772             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 773             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 774             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 775
 776             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 777             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 778             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 779
 780             /**************************
 781              * CALCULATE INTERACTIONS *
 782              **************************/
 783
 784             r10              = _mm_mul_ps(rsq10,rinv10);
 785             r10              = _mm_andnot_ps(dummy_mask,r10);
 786
 787             /* EWALD ELECTROSTATICS */
 788
 789             /* Analytical PME correction */
 790             zeta2            = _mm_mul_ps(beta2,rsq10);
 791             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 792             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 793             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 794             felec            = _mm_mul_ps(qq10,felec);
 795             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 796             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 797             velec            = _mm_mul_ps(qq10,velec);
 798
 799             /* Update potential sum for this i atom from the interaction with this j atom. */
 800             velec            = _mm_andnot_ps(dummy_mask,velec);
 801             velecsum         = _mm_add_ps(velecsum,velec);
 802
 803             fscal            = felec;
 804
 805             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 806
 807              /* Update vectorial force */
 808             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 809             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 810             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 811
 812             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 813             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 814             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 815
 816             /**************************
 817              * CALCULATE INTERACTIONS *
 818              **************************/
 819
 820             r11              = _mm_mul_ps(rsq11,rinv11);
 821             r11              = _mm_andnot_ps(dummy_mask,r11);
 822
 823             /* EWALD ELECTROSTATICS */
 824
 825             /* Analytical PME correction */
 826             zeta2            = _mm_mul_ps(beta2,rsq11);
 827             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 828             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 829             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 830             felec            = _mm_mul_ps(qq11,felec);
 831             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 832             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 833             velec            = _mm_mul_ps(qq11,velec);
 834
 835             /* Update potential sum for this i atom from the interaction with this j atom. */
 836             velec            = _mm_andnot_ps(dummy_mask,velec);
 837             velecsum         = _mm_add_ps(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 842
 843              /* Update vectorial force */
 844             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 845             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 846             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 847
 848             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 849             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 850             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 851
 852             /**************************
 853              * CALCULATE INTERACTIONS *
 854              **************************/
 855
 856             r12              = _mm_mul_ps(rsq12,rinv12);
 857             r12              = _mm_andnot_ps(dummy_mask,r12);
 858
 859             /* EWALD ELECTROSTATICS */
 860
 861             /* Analytical PME correction */
 862             zeta2            = _mm_mul_ps(beta2,rsq12);
 863             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 864             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 865             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 866             felec            = _mm_mul_ps(qq12,felec);
 867             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 868             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 869             velec            = _mm_mul_ps(qq12,velec);
 870
 871             /* Update potential sum for this i atom from the interaction with this j atom. */
 872             velec            = _mm_andnot_ps(dummy_mask,velec);
 873             velecsum         = _mm_add_ps(velecsum,velec);
 874
 875             fscal            = felec;
 876
 877             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 878
 879              /* Update vectorial force */
 880             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 881             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 882             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 883
 884             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 885             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 886             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             r20              = _mm_mul_ps(rsq20,rinv20);
 893             r20              = _mm_andnot_ps(dummy_mask,r20);
 894
 895             /* EWALD ELECTROSTATICS */
 896
 897             /* Analytical PME correction */
 898             zeta2            = _mm_mul_ps(beta2,rsq20);
 899             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 900             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 901             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 902             felec            = _mm_mul_ps(qq20,felec);
 903             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 904             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 905             velec            = _mm_mul_ps(qq20,velec);
 906
 907             /* Update potential sum for this i atom from the interaction with this j atom. */
 908             velec            = _mm_andnot_ps(dummy_mask,velec);
 909             velecsum         = _mm_add_ps(velecsum,velec);
 910
 911             fscal            = felec;
 912
 913             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 914
 915              /* Update vectorial force */
 916             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 917             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 918             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 919
 920             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 921             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 922             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 923
 924             /**************************
 925              * CALCULATE INTERACTIONS *
 926              **************************/
 927
 928             r21              = _mm_mul_ps(rsq21,rinv21);
 929             r21              = _mm_andnot_ps(dummy_mask,r21);
 930
 931             /* EWALD ELECTROSTATICS */
 932
 933             /* Analytical PME correction */
 934             zeta2            = _mm_mul_ps(beta2,rsq21);
 935             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 936             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 937             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 938             felec            = _mm_mul_ps(qq21,felec);
 939             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 940             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 941             velec            = _mm_mul_ps(qq21,velec);
 942
 943             /* Update potential sum for this i atom from the interaction with this j atom. */
 944             velec            = _mm_andnot_ps(dummy_mask,velec);
 945             velecsum         = _mm_add_ps(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 950
 951              /* Update vectorial force */
 952             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 953             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 954             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 955
 956             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 957             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 958             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 959
 960             /**************************
 961              * CALCULATE INTERACTIONS *
 962              **************************/
 963
 964             r22              = _mm_mul_ps(rsq22,rinv22);
 965             r22              = _mm_andnot_ps(dummy_mask,r22);
 966
 967             /* EWALD ELECTROSTATICS */
 968
 969             /* Analytical PME correction */
 970             zeta2            = _mm_mul_ps(beta2,rsq22);
 971             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 972             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 973             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 974             felec            = _mm_mul_ps(qq22,felec);
 975             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 976             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 977             velec            = _mm_mul_ps(qq22,velec);
 978
 979             /* Update potential sum for this i atom from the interaction with this j atom. */
 980             velec            = _mm_andnot_ps(dummy_mask,velec);
 981             velecsum         = _mm_add_ps(velecsum,velec);
 982
 983             fscal            = felec;
 984
 985             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 986
 987              /* Update vectorial force */
 988             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 989             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 990             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 991
 992             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 993             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 994             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 995
 996             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 997             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 998             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 999             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1000
1001             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1002                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1003
1004             /* Inner loop uses 270 flops */
1005         }
1006
1007         /* End of innermost loop */
1008
1009         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1010                                               f+i_coord_offset,fshift+i_shift_offset);
1011
1012         ggid                        = gid[iidx];
1013         /* Update potential energies */
1014         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1015
1016         /* Increment number of inner iterations */
1017         inneriter                  += j_index_end - j_index_start;
1018
1019         /* Outer loop uses 19 flops */
1020     }
1021
1022     /* Increment number of outer iterations */
1023     outeriter        += nri;
1024
1025     /* Update outer/inner flops */
1026
1027     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*270);
1028 }
1029 /*
1030  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW3W3_F_avx_128_fma_single
1031  * Electrostatics interaction: Ewald
1032  * VdW interaction:            None
1033  * Geometry:                   Water3-Water3
1034  * Calculate force/pot:        Force
1035  */
1036 void
1037 nb_kernel_ElecEw_VdwNone_GeomW3W3_F_avx_128_fma_single
1038                     (t_nblist * gmx_restrict                nlist,
1039                      rvec * gmx_restrict                    xx,
1040                      rvec * gmx_restrict                    ff,
1041                      t_forcerec * gmx_restrict              fr,
1042                      t_mdatoms * gmx_restrict               mdatoms,
1043                      nb_kernel_data_t * gmx_restrict        kernel_data,
1044                      t_nrnb * gmx_restrict                  nrnb)
1045 {
1046     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1047      * just 0 for non-waters.
1048      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1049      * jnr indices corresponding to data put in the four positions in the SIMD register.
1050      */
1051     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1052     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1053     int              jnrA,jnrB,jnrC,jnrD;
1054     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1055     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1056     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1057     real             rcutoff_scalar;
1058     real             *shiftvec,*fshift,*x,*f;
1059     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1060     real             scratch[4*DIM];
1061     __m128           fscal,rcutoff,rcutoff2,jidxall;
1062     int              vdwioffset0;
1063     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1064     int              vdwioffset1;
1065     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1066     int              vdwioffset2;
1067     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1068     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1069     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1070     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1071     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1072     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1073     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1074     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1075     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1076     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1077     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1078     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1079     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1080     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1081     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1082     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1083     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1084     real             *charge;
1085     __m128i          ewitab;
1086     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1087     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1088     real             *ewtab;
1089     __m128           dummy_mask,cutoff_mask;
1090     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1091     __m128           one     = _mm_set1_ps(1.0);
1092     __m128           two     = _mm_set1_ps(2.0);
1093     x                = xx[0];
1094     f                = ff[0];
1095
1096     nri              = nlist->nri;
1097     iinr             = nlist->iinr;
1098     jindex           = nlist->jindex;
1099     jjnr             = nlist->jjnr;
1100     shiftidx         = nlist->shift;
1101     gid              = nlist->gid;
1102     shiftvec         = fr->shift_vec[0];
1103     fshift           = fr->fshift[0];
1104     facel            = _mm_set1_ps(fr->epsfac);
1105     charge           = mdatoms->chargeA;
1106
1107     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1108     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
1109     beta2            = _mm_mul_ps(beta,beta);
1110     beta3            = _mm_mul_ps(beta,beta2);
1111     ewtab            = fr->ic->tabq_coul_F;
1112     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1113     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1114
1115     /* Setup water-specific parameters */
1116     inr              = nlist->iinr[0];
1117     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1118     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1119     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1120
1121     jq0              = _mm_set1_ps(charge[inr+0]);
1122     jq1              = _mm_set1_ps(charge[inr+1]);
1123     jq2              = _mm_set1_ps(charge[inr+2]);
1124     qq00             = _mm_mul_ps(iq0,jq0);
1125     qq01             = _mm_mul_ps(iq0,jq1);
1126     qq02             = _mm_mul_ps(iq0,jq2);
1127     qq10             = _mm_mul_ps(iq1,jq0);
1128     qq11             = _mm_mul_ps(iq1,jq1);
1129     qq12             = _mm_mul_ps(iq1,jq2);
1130     qq20             = _mm_mul_ps(iq2,jq0);
1131     qq21             = _mm_mul_ps(iq2,jq1);
1132     qq22             = _mm_mul_ps(iq2,jq2);
1133
1134     /* Avoid stupid compiler warnings */
1135     jnrA = jnrB = jnrC = jnrD = 0;
1136     j_coord_offsetA = 0;
1137     j_coord_offsetB = 0;
1138     j_coord_offsetC = 0;
1139     j_coord_offsetD = 0;
1140
1141     outeriter        = 0;
1142     inneriter        = 0;
1143
1144     for(iidx=0;iidx<4*DIM;iidx++)
1145     {
1146         scratch[iidx] = 0.0;
1147     }
1148
1149     /* Start outer loop over neighborlists */
1150     for(iidx=0; iidx<nri; iidx++)
1151     {
1152         /* Load shift vector for this list */
1153         i_shift_offset   = DIM*shiftidx[iidx];
1154
1155         /* Load limits for loop over neighbors */
1156         j_index_start    = jindex[iidx];
1157         j_index_end      = jindex[iidx+1];
1158
1159         /* Get outer coordinate index */
1160         inr              = iinr[iidx];
1161         i_coord_offset   = DIM*inr;
1162
1163         /* Load i particle coords and add shift vector */
1164         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1165                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1166
1167         fix0             = _mm_setzero_ps();
1168         fiy0             = _mm_setzero_ps();
1169         fiz0             = _mm_setzero_ps();
1170         fix1             = _mm_setzero_ps();
1171         fiy1             = _mm_setzero_ps();
1172         fiz1             = _mm_setzero_ps();
1173         fix2             = _mm_setzero_ps();
1174         fiy2             = _mm_setzero_ps();
1175         fiz2             = _mm_setzero_ps();
1176
1177         /* Start inner kernel loop */
1178         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1179         {
1180
1181             /* Get j neighbor index, and coordinate index */
1182             jnrA             = jjnr[jidx];
1183             jnrB             = jjnr[jidx+1];
1184             jnrC             = jjnr[jidx+2];
1185             jnrD             = jjnr[jidx+3];
1186             j_coord_offsetA  = DIM*jnrA;
1187             j_coord_offsetB  = DIM*jnrB;
1188             j_coord_offsetC  = DIM*jnrC;
1189             j_coord_offsetD  = DIM*jnrD;
1190
1191             /* load j atom coordinates */
1192             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1193                                               x+j_coord_offsetC,x+j_coord_offsetD,
1194                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1195
1196             /* Calculate displacement vector */
1197             dx00             = _mm_sub_ps(ix0,jx0);
1198             dy00             = _mm_sub_ps(iy0,jy0);
1199             dz00             = _mm_sub_ps(iz0,jz0);
1200             dx01             = _mm_sub_ps(ix0,jx1);
1201             dy01             = _mm_sub_ps(iy0,jy1);
1202             dz01             = _mm_sub_ps(iz0,jz1);
1203             dx02             = _mm_sub_ps(ix0,jx2);
1204             dy02             = _mm_sub_ps(iy0,jy2);
1205             dz02             = _mm_sub_ps(iz0,jz2);
1206             dx10             = _mm_sub_ps(ix1,jx0);
1207             dy10             = _mm_sub_ps(iy1,jy0);
1208             dz10             = _mm_sub_ps(iz1,jz0);
1209             dx11             = _mm_sub_ps(ix1,jx1);
1210             dy11             = _mm_sub_ps(iy1,jy1);
1211             dz11             = _mm_sub_ps(iz1,jz1);
1212             dx12             = _mm_sub_ps(ix1,jx2);
1213             dy12             = _mm_sub_ps(iy1,jy2);
1214             dz12             = _mm_sub_ps(iz1,jz2);
1215             dx20             = _mm_sub_ps(ix2,jx0);
1216             dy20             = _mm_sub_ps(iy2,jy0);
1217             dz20             = _mm_sub_ps(iz2,jz0);
1218             dx21             = _mm_sub_ps(ix2,jx1);
1219             dy21             = _mm_sub_ps(iy2,jy1);
1220             dz21             = _mm_sub_ps(iz2,jz1);
1221             dx22             = _mm_sub_ps(ix2,jx2);
1222             dy22             = _mm_sub_ps(iy2,jy2);
1223             dz22             = _mm_sub_ps(iz2,jz2);
1224
1225             /* Calculate squared distance and things based on it */
1226             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1227             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1228             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1229             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1230             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1231             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1232             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1233             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1234             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1235
1236             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1237             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1238             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1239             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1240             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1241             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1242             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1243             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1244             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1245
1246             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1247             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1248             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1249             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1250             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1251             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1252             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1253             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1254             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1255
1256             fjx0             = _mm_setzero_ps();
1257             fjy0             = _mm_setzero_ps();
1258             fjz0             = _mm_setzero_ps();
1259             fjx1             = _mm_setzero_ps();
1260             fjy1             = _mm_setzero_ps();
1261             fjz1             = _mm_setzero_ps();
1262             fjx2             = _mm_setzero_ps();
1263             fjy2             = _mm_setzero_ps();
1264             fjz2             = _mm_setzero_ps();
1265
1266             /**************************
1267              * CALCULATE INTERACTIONS *
1268              **************************/
1269
1270             r00              = _mm_mul_ps(rsq00,rinv00);
1271
1272             /* EWALD ELECTROSTATICS */
1273
1274             /* Analytical PME correction */
1275             zeta2            = _mm_mul_ps(beta2,rsq00);
1276             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1277             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1278             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1279             felec            = _mm_mul_ps(qq00,felec);
1280
1281             fscal            = felec;
1282
1283              /* Update vectorial force */
1284             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1285             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1286             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1287
1288             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1289             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1290             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             r01              = _mm_mul_ps(rsq01,rinv01);
1297
1298             /* EWALD ELECTROSTATICS */
1299
1300             /* Analytical PME correction */
1301             zeta2            = _mm_mul_ps(beta2,rsq01);
1302             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1303             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1304             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1305             felec            = _mm_mul_ps(qq01,felec);
1306
1307             fscal            = felec;
1308
1309              /* Update vectorial force */
1310             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1311             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1312             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1313
1314             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1315             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1316             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1317
1318             /**************************
1319              * CALCULATE INTERACTIONS *
1320              **************************/
1321
1322             r02              = _mm_mul_ps(rsq02,rinv02);
1323
1324             /* EWALD ELECTROSTATICS */
1325
1326             /* Analytical PME correction */
1327             zeta2            = _mm_mul_ps(beta2,rsq02);
1328             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1329             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1330             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1331             felec            = _mm_mul_ps(qq02,felec);
1332
1333             fscal            = felec;
1334
1335              /* Update vectorial force */
1336             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1337             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1338             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1339
1340             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1341             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1342             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             r10              = _mm_mul_ps(rsq10,rinv10);
1349
1350             /* EWALD ELECTROSTATICS */
1351
1352             /* Analytical PME correction */
1353             zeta2            = _mm_mul_ps(beta2,rsq10);
1354             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1355             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1356             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1357             felec            = _mm_mul_ps(qq10,felec);
1358
1359             fscal            = felec;
1360
1361              /* Update vectorial force */
1362             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1363             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1364             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1365
1366             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1367             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1368             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1369
1370             /**************************
1371              * CALCULATE INTERACTIONS *
1372              **************************/
1373
1374             r11              = _mm_mul_ps(rsq11,rinv11);
1375
1376             /* EWALD ELECTROSTATICS */
1377
1378             /* Analytical PME correction */
1379             zeta2            = _mm_mul_ps(beta2,rsq11);
1380             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1381             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1382             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1383             felec            = _mm_mul_ps(qq11,felec);
1384
1385             fscal            = felec;
1386
1387              /* Update vectorial force */
1388             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1389             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1390             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1391
1392             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1393             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1394             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1395
1396             /**************************
1397              * CALCULATE INTERACTIONS *
1398              **************************/
1399
1400             r12              = _mm_mul_ps(rsq12,rinv12);
1401
1402             /* EWALD ELECTROSTATICS */
1403
1404             /* Analytical PME correction */
1405             zeta2            = _mm_mul_ps(beta2,rsq12);
1406             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1407             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1408             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1409             felec            = _mm_mul_ps(qq12,felec);
1410
1411             fscal            = felec;
1412
1413              /* Update vectorial force */
1414             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1415             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1416             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1417
1418             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1419             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1420             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1421
1422             /**************************
1423              * CALCULATE INTERACTIONS *
1424              **************************/
1425
1426             r20              = _mm_mul_ps(rsq20,rinv20);
1427
1428             /* EWALD ELECTROSTATICS */
1429
1430             /* Analytical PME correction */
1431             zeta2            = _mm_mul_ps(beta2,rsq20);
1432             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1433             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1434             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1435             felec            = _mm_mul_ps(qq20,felec);
1436
1437             fscal            = felec;
1438
1439              /* Update vectorial force */
1440             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1441             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1442             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1443
1444             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1445             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1446             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1447
1448             /**************************
1449              * CALCULATE INTERACTIONS *
1450              **************************/
1451
1452             r21              = _mm_mul_ps(rsq21,rinv21);
1453
1454             /* EWALD ELECTROSTATICS */
1455
1456             /* Analytical PME correction */
1457             zeta2            = _mm_mul_ps(beta2,rsq21);
1458             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1459             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1460             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1461             felec            = _mm_mul_ps(qq21,felec);
1462
1463             fscal            = felec;
1464
1465              /* Update vectorial force */
1466             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1467             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1468             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1469
1470             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1471             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1472             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1473
1474             /**************************
1475              * CALCULATE INTERACTIONS *
1476              **************************/
1477
1478             r22              = _mm_mul_ps(rsq22,rinv22);
1479
1480             /* EWALD ELECTROSTATICS */
1481
1482             /* Analytical PME correction */
1483             zeta2            = _mm_mul_ps(beta2,rsq22);
1484             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1485             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1486             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1487             felec            = _mm_mul_ps(qq22,felec);
1488
1489             fscal            = felec;
1490
1491              /* Update vectorial force */
1492             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1493             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1494             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1495
1496             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1497             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1498             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1499
1500             fjptrA             = f+j_coord_offsetA;
1501             fjptrB             = f+j_coord_offsetB;
1502             fjptrC             = f+j_coord_offsetC;
1503             fjptrD             = f+j_coord_offsetD;
1504
1505             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1506                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1507
1508             /* Inner loop uses 252 flops */
1509         }
1510
1511         if(jidx<j_index_end)
1512         {
1513
1514             /* Get j neighbor index, and coordinate index */
1515             jnrlistA         = jjnr[jidx];
1516             jnrlistB         = jjnr[jidx+1];
1517             jnrlistC         = jjnr[jidx+2];
1518             jnrlistD         = jjnr[jidx+3];
1519             /* Sign of each element will be negative for non-real atoms.
1520              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1521              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1522              */
1523             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1524             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1525             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1526             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1527             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1528             j_coord_offsetA  = DIM*jnrA;
1529             j_coord_offsetB  = DIM*jnrB;
1530             j_coord_offsetC  = DIM*jnrC;
1531             j_coord_offsetD  = DIM*jnrD;
1532
1533             /* load j atom coordinates */
1534             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1535                                               x+j_coord_offsetC,x+j_coord_offsetD,
1536                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1537
1538             /* Calculate displacement vector */
1539             dx00             = _mm_sub_ps(ix0,jx0);
1540             dy00             = _mm_sub_ps(iy0,jy0);
1541             dz00             = _mm_sub_ps(iz0,jz0);
1542             dx01             = _mm_sub_ps(ix0,jx1);
1543             dy01             = _mm_sub_ps(iy0,jy1);
1544             dz01             = _mm_sub_ps(iz0,jz1);
1545             dx02             = _mm_sub_ps(ix0,jx2);
1546             dy02             = _mm_sub_ps(iy0,jy2);
1547             dz02             = _mm_sub_ps(iz0,jz2);
1548             dx10             = _mm_sub_ps(ix1,jx0);
1549             dy10             = _mm_sub_ps(iy1,jy0);
1550             dz10             = _mm_sub_ps(iz1,jz0);
1551             dx11             = _mm_sub_ps(ix1,jx1);
1552             dy11             = _mm_sub_ps(iy1,jy1);
1553             dz11             = _mm_sub_ps(iz1,jz1);
1554             dx12             = _mm_sub_ps(ix1,jx2);
1555             dy12             = _mm_sub_ps(iy1,jy2);
1556             dz12             = _mm_sub_ps(iz1,jz2);
1557             dx20             = _mm_sub_ps(ix2,jx0);
1558             dy20             = _mm_sub_ps(iy2,jy0);
1559             dz20             = _mm_sub_ps(iz2,jz0);
1560             dx21             = _mm_sub_ps(ix2,jx1);
1561             dy21             = _mm_sub_ps(iy2,jy1);
1562             dz21             = _mm_sub_ps(iz2,jz1);
1563             dx22             = _mm_sub_ps(ix2,jx2);
1564             dy22             = _mm_sub_ps(iy2,jy2);
1565             dz22             = _mm_sub_ps(iz2,jz2);
1566
1567             /* Calculate squared distance and things based on it */
1568             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1569             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1570             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1571             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1572             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1573             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1574             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1575             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1576             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1577
1578             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1579             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1580             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1581             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1582             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1583             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1584             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1585             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1586             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1587
1588             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1589             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1590             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1591             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1592             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1593             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1594             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1595             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1596             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1597
1598             fjx0             = _mm_setzero_ps();
1599             fjy0             = _mm_setzero_ps();
1600             fjz0             = _mm_setzero_ps();
1601             fjx1             = _mm_setzero_ps();
1602             fjy1             = _mm_setzero_ps();
1603             fjz1             = _mm_setzero_ps();
1604             fjx2             = _mm_setzero_ps();
1605             fjy2             = _mm_setzero_ps();
1606             fjz2             = _mm_setzero_ps();
1607
1608             /**************************
1609              * CALCULATE INTERACTIONS *
1610              **************************/
1611
1612             r00              = _mm_mul_ps(rsq00,rinv00);
1613             r00              = _mm_andnot_ps(dummy_mask,r00);
1614
1615             /* EWALD ELECTROSTATICS */
1616
1617             /* Analytical PME correction */
1618             zeta2            = _mm_mul_ps(beta2,rsq00);
1619             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1620             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1621             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1622             felec            = _mm_mul_ps(qq00,felec);
1623
1624             fscal            = felec;
1625
1626             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1627
1628              /* Update vectorial force */
1629             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1630             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1631             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1632
1633             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1634             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1635             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1636
1637             /**************************
1638              * CALCULATE INTERACTIONS *
1639              **************************/
1640
1641             r01              = _mm_mul_ps(rsq01,rinv01);
1642             r01              = _mm_andnot_ps(dummy_mask,r01);
1643
1644             /* EWALD ELECTROSTATICS */
1645
1646             /* Analytical PME correction */
1647             zeta2            = _mm_mul_ps(beta2,rsq01);
1648             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1649             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1650             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1651             felec            = _mm_mul_ps(qq01,felec);
1652
1653             fscal            = felec;
1654
1655             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1656
1657              /* Update vectorial force */
1658             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1659             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1660             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1661
1662             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1663             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1664             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1665
1666             /**************************
1667              * CALCULATE INTERACTIONS *
1668              **************************/
1669
1670             r02              = _mm_mul_ps(rsq02,rinv02);
1671             r02              = _mm_andnot_ps(dummy_mask,r02);
1672
1673             /* EWALD ELECTROSTATICS */
1674
1675             /* Analytical PME correction */
1676             zeta2            = _mm_mul_ps(beta2,rsq02);
1677             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1678             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1679             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1680             felec            = _mm_mul_ps(qq02,felec);
1681
1682             fscal            = felec;
1683
1684             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1685
1686              /* Update vectorial force */
1687             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1688             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1689             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1690
1691             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1692             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1693             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1694
1695             /**************************
1696              * CALCULATE INTERACTIONS *
1697              **************************/
1698
1699             r10              = _mm_mul_ps(rsq10,rinv10);
1700             r10              = _mm_andnot_ps(dummy_mask,r10);
1701
1702             /* EWALD ELECTROSTATICS */
1703
1704             /* Analytical PME correction */
1705             zeta2            = _mm_mul_ps(beta2,rsq10);
1706             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1707             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1708             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1709             felec            = _mm_mul_ps(qq10,felec);
1710
1711             fscal            = felec;
1712
1713             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1714
1715              /* Update vectorial force */
1716             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1717             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1718             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1719
1720             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1721             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1722             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1723
1724             /**************************
1725              * CALCULATE INTERACTIONS *
1726              **************************/
1727
1728             r11              = _mm_mul_ps(rsq11,rinv11);
1729             r11              = _mm_andnot_ps(dummy_mask,r11);
1730
1731             /* EWALD ELECTROSTATICS */
1732
1733             /* Analytical PME correction */
1734             zeta2            = _mm_mul_ps(beta2,rsq11);
1735             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1736             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1737             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1738             felec            = _mm_mul_ps(qq11,felec);
1739
1740             fscal            = felec;
1741
1742             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1743
1744              /* Update vectorial force */
1745             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1746             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1747             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1748
1749             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1750             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1751             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1752
1753             /**************************
1754              * CALCULATE INTERACTIONS *
1755              **************************/
1756
1757             r12              = _mm_mul_ps(rsq12,rinv12);
1758             r12              = _mm_andnot_ps(dummy_mask,r12);
1759
1760             /* EWALD ELECTROSTATICS */
1761
1762             /* Analytical PME correction */
1763             zeta2            = _mm_mul_ps(beta2,rsq12);
1764             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1765             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1766             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1767             felec            = _mm_mul_ps(qq12,felec);
1768
1769             fscal            = felec;
1770
1771             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1772
1773              /* Update vectorial force */
1774             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1775             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1776             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1777
1778             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1779             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1780             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1781
1782             /**************************
1783              * CALCULATE INTERACTIONS *
1784              **************************/
1785
1786             r20              = _mm_mul_ps(rsq20,rinv20);
1787             r20              = _mm_andnot_ps(dummy_mask,r20);
1788
1789             /* EWALD ELECTROSTATICS */
1790
1791             /* Analytical PME correction */
1792             zeta2            = _mm_mul_ps(beta2,rsq20);
1793             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1794             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1795             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1796             felec            = _mm_mul_ps(qq20,felec);
1797
1798             fscal            = felec;
1799
1800             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1801
1802              /* Update vectorial force */
1803             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1804             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1805             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1806
1807             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1808             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1809             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1810
1811             /**************************
1812              * CALCULATE INTERACTIONS *
1813              **************************/
1814
1815             r21              = _mm_mul_ps(rsq21,rinv21);
1816             r21              = _mm_andnot_ps(dummy_mask,r21);
1817
1818             /* EWALD ELECTROSTATICS */
1819
1820             /* Analytical PME correction */
1821             zeta2            = _mm_mul_ps(beta2,rsq21);
1822             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1823             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1824             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1825             felec            = _mm_mul_ps(qq21,felec);
1826
1827             fscal            = felec;
1828
1829             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1830
1831              /* Update vectorial force */
1832             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1833             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1834             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1835
1836             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1837             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1838             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1839
1840             /**************************
1841              * CALCULATE INTERACTIONS *
1842              **************************/
1843
1844             r22              = _mm_mul_ps(rsq22,rinv22);
1845             r22              = _mm_andnot_ps(dummy_mask,r22);
1846
1847             /* EWALD ELECTROSTATICS */
1848
1849             /* Analytical PME correction */
1850             zeta2            = _mm_mul_ps(beta2,rsq22);
1851             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1852             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1853             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1854             felec            = _mm_mul_ps(qq22,felec);
1855
1856             fscal            = felec;
1857
1858             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1859
1860              /* Update vectorial force */
1861             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1862             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1863             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1864
1865             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1866             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1867             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1868
1869             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1870             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1871             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1872             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1873
1874             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1875                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1876
1877             /* Inner loop uses 261 flops */
1878         }
1879
1880         /* End of innermost loop */
1881
1882         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1883                                               f+i_coord_offset,fshift+i_shift_offset);
1884
1885         /* Increment number of inner iterations */
1886         inneriter                  += j_index_end - j_index_start;
1887
1888         /* Outer loop uses 18 flops */
1889     }
1890
1891     /* Increment number of outer iterations */
1892     outeriter        += nri;
1893
1894     /* Update outer/inner flops */
1895
1896     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*261);
1897 }