src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRF_VdwLJ_GeomW3W3_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  91     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  93     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     int              nvdwtype;
 106     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 107     int              *vdwtype;
 108     real             *vdwparam;
 109     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 110     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 111     __m128           dummy_mask,cutoff_mask;
 112     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 113     __m128           one     = _mm_set1_ps(1.0);
 114     __m128           two     = _mm_set1_ps(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm_set1_ps(fr->ic->epsfac);
 127     charge           = mdatoms->chargeA;
 128     krf              = _mm_set1_ps(fr->ic->k_rf);
 129     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 130     crf              = _mm_set1_ps(fr->ic->c_rf);
 131     nvdwtype         = fr->ntype;
 132     vdwparam         = fr->nbfp;
 133     vdwtype          = mdatoms->typeA;
 134
 135     /* Setup water-specific parameters */
 136     inr              = nlist->iinr[0];
 137     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 138     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 139     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 140     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 141
 142     jq0              = _mm_set1_ps(charge[inr+0]);
 143     jq1              = _mm_set1_ps(charge[inr+1]);
 144     jq2              = _mm_set1_ps(charge[inr+2]);
 145     vdwjidx0A        = 2*vdwtype[inr+0];
 146     qq00             = _mm_mul_ps(iq0,jq0);
 147     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 148     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 149     qq01             = _mm_mul_ps(iq0,jq1);
 150     qq02             = _mm_mul_ps(iq0,jq2);
 151     qq10             = _mm_mul_ps(iq1,jq0);
 152     qq11             = _mm_mul_ps(iq1,jq1);
 153     qq12             = _mm_mul_ps(iq1,jq2);
 154     qq20             = _mm_mul_ps(iq2,jq0);
 155     qq21             = _mm_mul_ps(iq2,jq1);
 156     qq22             = _mm_mul_ps(iq2,jq2);
 157
 158     /* Avoid stupid compiler warnings */
 159     jnrA = jnrB = jnrC = jnrD = 0;
 160     j_coord_offsetA = 0;
 161     j_coord_offsetB = 0;
 162     j_coord_offsetC = 0;
 163     j_coord_offsetD = 0;
 164
 165     outeriter        = 0;
 166     inneriter        = 0;
 167
 168     for(iidx=0;iidx<4*DIM;iidx++)
 169     {
 170         scratch[iidx] = 0.0;
 171     }
 172
 173     /* Start outer loop over neighborlists */
 174     for(iidx=0; iidx<nri; iidx++)
 175     {
 176         /* Load shift vector for this list */
 177         i_shift_offset   = DIM*shiftidx[iidx];
 178
 179         /* Load limits for loop over neighbors */
 180         j_index_start    = jindex[iidx];
 181         j_index_end      = jindex[iidx+1];
 182
 183         /* Get outer coordinate index */
 184         inr              = iinr[iidx];
 185         i_coord_offset   = DIM*inr;
 186
 187         /* Load i particle coords and add shift vector */
 188         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 189                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 190
 191         fix0             = _mm_setzero_ps();
 192         fiy0             = _mm_setzero_ps();
 193         fiz0             = _mm_setzero_ps();
 194         fix1             = _mm_setzero_ps();
 195         fiy1             = _mm_setzero_ps();
 196         fiz1             = _mm_setzero_ps();
 197         fix2             = _mm_setzero_ps();
 198         fiy2             = _mm_setzero_ps();
 199         fiz2             = _mm_setzero_ps();
 200
 201         /* Reset potential sums */
 202         velecsum         = _mm_setzero_ps();
 203         vvdwsum          = _mm_setzero_ps();
 204
 205         /* Start inner kernel loop */
 206         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 207         {
 208
 209             /* Get j neighbor index, and coordinate index */
 210             jnrA             = jjnr[jidx];
 211             jnrB             = jjnr[jidx+1];
 212             jnrC             = jjnr[jidx+2];
 213             jnrD             = jjnr[jidx+3];
 214             j_coord_offsetA  = DIM*jnrA;
 215             j_coord_offsetB  = DIM*jnrB;
 216             j_coord_offsetC  = DIM*jnrC;
 217             j_coord_offsetD  = DIM*jnrD;
 218
 219             /* load j atom coordinates */
 220             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 221                                               x+j_coord_offsetC,x+j_coord_offsetD,
 222                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 223
 224             /* Calculate displacement vector */
 225             dx00             = _mm_sub_ps(ix0,jx0);
 226             dy00             = _mm_sub_ps(iy0,jy0);
 227             dz00             = _mm_sub_ps(iz0,jz0);
 228             dx01             = _mm_sub_ps(ix0,jx1);
 229             dy01             = _mm_sub_ps(iy0,jy1);
 230             dz01             = _mm_sub_ps(iz0,jz1);
 231             dx02             = _mm_sub_ps(ix0,jx2);
 232             dy02             = _mm_sub_ps(iy0,jy2);
 233             dz02             = _mm_sub_ps(iz0,jz2);
 234             dx10             = _mm_sub_ps(ix1,jx0);
 235             dy10             = _mm_sub_ps(iy1,jy0);
 236             dz10             = _mm_sub_ps(iz1,jz0);
 237             dx11             = _mm_sub_ps(ix1,jx1);
 238             dy11             = _mm_sub_ps(iy1,jy1);
 239             dz11             = _mm_sub_ps(iz1,jz1);
 240             dx12             = _mm_sub_ps(ix1,jx2);
 241             dy12             = _mm_sub_ps(iy1,jy2);
 242             dz12             = _mm_sub_ps(iz1,jz2);
 243             dx20             = _mm_sub_ps(ix2,jx0);
 244             dy20             = _mm_sub_ps(iy2,jy0);
 245             dz20             = _mm_sub_ps(iz2,jz0);
 246             dx21             = _mm_sub_ps(ix2,jx1);
 247             dy21             = _mm_sub_ps(iy2,jy1);
 248             dz21             = _mm_sub_ps(iz2,jz1);
 249             dx22             = _mm_sub_ps(ix2,jx2);
 250             dy22             = _mm_sub_ps(iy2,jy2);
 251             dz22             = _mm_sub_ps(iz2,jz2);
 252
 253             /* Calculate squared distance and things based on it */
 254             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 255             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 256             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 257             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 258             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 259             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 260             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 261             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 262             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 263
 264             rinv00           = avx128fma_invsqrt_f(rsq00);
 265             rinv01           = avx128fma_invsqrt_f(rsq01);
 266             rinv02           = avx128fma_invsqrt_f(rsq02);
 267             rinv10           = avx128fma_invsqrt_f(rsq10);
 268             rinv11           = avx128fma_invsqrt_f(rsq11);
 269             rinv12           = avx128fma_invsqrt_f(rsq12);
 270             rinv20           = avx128fma_invsqrt_f(rsq20);
 271             rinv21           = avx128fma_invsqrt_f(rsq21);
 272             rinv22           = avx128fma_invsqrt_f(rsq22);
 273
 274             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 275             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 276             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 277             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 278             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 279             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 280             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 281             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 282             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 283
 284             fjx0             = _mm_setzero_ps();
 285             fjy0             = _mm_setzero_ps();
 286             fjz0             = _mm_setzero_ps();
 287             fjx1             = _mm_setzero_ps();
 288             fjy1             = _mm_setzero_ps();
 289             fjz1             = _mm_setzero_ps();
 290             fjx2             = _mm_setzero_ps();
 291             fjy2             = _mm_setzero_ps();
 292             fjz2             = _mm_setzero_ps();
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             /* REACTION-FIELD ELECTROSTATICS */
 299             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 300             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 301
 302             /* LENNARD-JONES DISPERSION/REPULSION */
 303
 304             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 305             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 306             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 307             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 308             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_ps(velecsum,velec);
 312             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 313
 314             fscal            = _mm_add_ps(felec,fvdw);
 315
 316              /* Update vectorial force */
 317             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 318             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 319             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 320
 321             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 322             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 323             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             /* REACTION-FIELD ELECTROSTATICS */
 330             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 331             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             velecsum         = _mm_add_ps(velecsum,velec);
 335
 336             fscal            = felec;
 337
 338              /* Update vectorial force */
 339             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 340             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 341             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 342
 343             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 344             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 345             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 346
 347             /**************************
 348              * CALCULATE INTERACTIONS *
 349              **************************/
 350
 351             /* REACTION-FIELD ELECTROSTATICS */
 352             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 353             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velecsum         = _mm_add_ps(velecsum,velec);
 357
 358             fscal            = felec;
 359
 360              /* Update vectorial force */
 361             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 362             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 363             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 364
 365             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 366             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 367             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 375             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 376
 377             /* Update potential sum for this i atom from the interaction with this j atom. */
 378             velecsum         = _mm_add_ps(velecsum,velec);
 379
 380             fscal            = felec;
 381
 382              /* Update vectorial force */
 383             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 384             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 385             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 386
 387             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 388             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 389             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             /* REACTION-FIELD ELECTROSTATICS */
 396             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 397             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velecsum         = _mm_add_ps(velecsum,velec);
 401
 402             fscal            = felec;
 403
 404              /* Update vectorial force */
 405             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 406             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 407             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 408
 409             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 410             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 411             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 412
 413             /**************************
 414              * CALCULATE INTERACTIONS *
 415              **************************/
 416
 417             /* REACTION-FIELD ELECTROSTATICS */
 418             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 419             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 420
 421             /* Update potential sum for this i atom from the interaction with this j atom. */
 422             velecsum         = _mm_add_ps(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426              /* Update vectorial force */
 427             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 428             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 429             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 430
 431             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 432             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 433             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             /* REACTION-FIELD ELECTROSTATICS */
 440             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 441             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velecsum         = _mm_add_ps(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448              /* Update vectorial force */
 449             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 450             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 451             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 452
 453             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 454             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 455             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             /* REACTION-FIELD ELECTROSTATICS */
 462             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 463             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 464
 465             /* Update potential sum for this i atom from the interaction with this j atom. */
 466             velecsum         = _mm_add_ps(velecsum,velec);
 467
 468             fscal            = felec;
 469
 470              /* Update vectorial force */
 471             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 472             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 473             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 474
 475             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 476             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 477             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             /* REACTION-FIELD ELECTROSTATICS */
 484             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 485             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 486
 487             /* Update potential sum for this i atom from the interaction with this j atom. */
 488             velecsum         = _mm_add_ps(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492              /* Update vectorial force */
 493             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 494             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 495             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 496
 497             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 498             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 499             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 500
 501             fjptrA             = f+j_coord_offsetA;
 502             fjptrB             = f+j_coord_offsetB;
 503             fjptrC             = f+j_coord_offsetC;
 504             fjptrD             = f+j_coord_offsetD;
 505
 506             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 507                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 508
 509             /* Inner loop uses 327 flops */
 510         }
 511
 512         if(jidx<j_index_end)
 513         {
 514
 515             /* Get j neighbor index, and coordinate index */
 516             jnrlistA         = jjnr[jidx];
 517             jnrlistB         = jjnr[jidx+1];
 518             jnrlistC         = jjnr[jidx+2];
 519             jnrlistD         = jjnr[jidx+3];
 520             /* Sign of each element will be negative for non-real atoms.
 521              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 522              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 523              */
 524             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 525             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 526             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 527             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 528             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 529             j_coord_offsetA  = DIM*jnrA;
 530             j_coord_offsetB  = DIM*jnrB;
 531             j_coord_offsetC  = DIM*jnrC;
 532             j_coord_offsetD  = DIM*jnrD;
 533
 534             /* load j atom coordinates */
 535             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 536                                               x+j_coord_offsetC,x+j_coord_offsetD,
 537                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 538
 539             /* Calculate displacement vector */
 540             dx00             = _mm_sub_ps(ix0,jx0);
 541             dy00             = _mm_sub_ps(iy0,jy0);
 542             dz00             = _mm_sub_ps(iz0,jz0);
 543             dx01             = _mm_sub_ps(ix0,jx1);
 544             dy01             = _mm_sub_ps(iy0,jy1);
 545             dz01             = _mm_sub_ps(iz0,jz1);
 546             dx02             = _mm_sub_ps(ix0,jx2);
 547             dy02             = _mm_sub_ps(iy0,jy2);
 548             dz02             = _mm_sub_ps(iz0,jz2);
 549             dx10             = _mm_sub_ps(ix1,jx0);
 550             dy10             = _mm_sub_ps(iy1,jy0);
 551             dz10             = _mm_sub_ps(iz1,jz0);
 552             dx11             = _mm_sub_ps(ix1,jx1);
 553             dy11             = _mm_sub_ps(iy1,jy1);
 554             dz11             = _mm_sub_ps(iz1,jz1);
 555             dx12             = _mm_sub_ps(ix1,jx2);
 556             dy12             = _mm_sub_ps(iy1,jy2);
 557             dz12             = _mm_sub_ps(iz1,jz2);
 558             dx20             = _mm_sub_ps(ix2,jx0);
 559             dy20             = _mm_sub_ps(iy2,jy0);
 560             dz20             = _mm_sub_ps(iz2,jz0);
 561             dx21             = _mm_sub_ps(ix2,jx1);
 562             dy21             = _mm_sub_ps(iy2,jy1);
 563             dz21             = _mm_sub_ps(iz2,jz1);
 564             dx22             = _mm_sub_ps(ix2,jx2);
 565             dy22             = _mm_sub_ps(iy2,jy2);
 566             dz22             = _mm_sub_ps(iz2,jz2);
 567
 568             /* Calculate squared distance and things based on it */
 569             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 570             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 571             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 572             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 573             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 574             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 575             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 576             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 577             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 578
 579             rinv00           = avx128fma_invsqrt_f(rsq00);
 580             rinv01           = avx128fma_invsqrt_f(rsq01);
 581             rinv02           = avx128fma_invsqrt_f(rsq02);
 582             rinv10           = avx128fma_invsqrt_f(rsq10);
 583             rinv11           = avx128fma_invsqrt_f(rsq11);
 584             rinv12           = avx128fma_invsqrt_f(rsq12);
 585             rinv20           = avx128fma_invsqrt_f(rsq20);
 586             rinv21           = avx128fma_invsqrt_f(rsq21);
 587             rinv22           = avx128fma_invsqrt_f(rsq22);
 588
 589             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 590             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 591             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 592             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 593             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 594             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 595             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 596             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 597             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 598
 599             fjx0             = _mm_setzero_ps();
 600             fjy0             = _mm_setzero_ps();
 601             fjz0             = _mm_setzero_ps();
 602             fjx1             = _mm_setzero_ps();
 603             fjy1             = _mm_setzero_ps();
 604             fjz1             = _mm_setzero_ps();
 605             fjx2             = _mm_setzero_ps();
 606             fjy2             = _mm_setzero_ps();
 607             fjz2             = _mm_setzero_ps();
 608
 609             /**************************
 610              * CALCULATE INTERACTIONS *
 611              **************************/
 612
 613             /* REACTION-FIELD ELECTROSTATICS */
 614             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 615             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 616
 617             /* LENNARD-JONES DISPERSION/REPULSION */
 618
 619             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 620             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 621             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 622             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 623             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 624
 625             /* Update potential sum for this i atom from the interaction with this j atom. */
 626             velec            = _mm_andnot_ps(dummy_mask,velec);
 627             velecsum         = _mm_add_ps(velecsum,velec);
 628             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 629             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 630
 631             fscal            = _mm_add_ps(felec,fvdw);
 632
 633             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 634
 635              /* Update vectorial force */
 636             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 637             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 638             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 639
 640             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 641             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 642             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 643
 644             /**************************
 645              * CALCULATE INTERACTIONS *
 646              **************************/
 647
 648             /* REACTION-FIELD ELECTROSTATICS */
 649             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 650             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 651
 652             /* Update potential sum for this i atom from the interaction with this j atom. */
 653             velec            = _mm_andnot_ps(dummy_mask,velec);
 654             velecsum         = _mm_add_ps(velecsum,velec);
 655
 656             fscal            = felec;
 657
 658             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 659
 660              /* Update vectorial force */
 661             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 662             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 663             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 664
 665             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 666             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 667             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 668
 669             /**************************
 670              * CALCULATE INTERACTIONS *
 671              **************************/
 672
 673             /* REACTION-FIELD ELECTROSTATICS */
 674             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 675             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 676
 677             /* Update potential sum for this i atom from the interaction with this j atom. */
 678             velec            = _mm_andnot_ps(dummy_mask,velec);
 679             velecsum         = _mm_add_ps(velecsum,velec);
 680
 681             fscal            = felec;
 682
 683             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 684
 685              /* Update vectorial force */
 686             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 687             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 688             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 689
 690             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 691             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 692             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 693
 694             /**************************
 695              * CALCULATE INTERACTIONS *
 696              **************************/
 697
 698             /* REACTION-FIELD ELECTROSTATICS */
 699             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 700             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 701
 702             /* Update potential sum for this i atom from the interaction with this j atom. */
 703             velec            = _mm_andnot_ps(dummy_mask,velec);
 704             velecsum         = _mm_add_ps(velecsum,velec);
 705
 706             fscal            = felec;
 707
 708             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 709
 710              /* Update vectorial force */
 711             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 712             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 713             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 714
 715             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 716             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 717             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 725             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 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             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 737             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 738             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 739
 740             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 741             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 742             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 743
 744             /**************************
 745              * CALCULATE INTERACTIONS *
 746              **************************/
 747
 748             /* REACTION-FIELD ELECTROSTATICS */
 749             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 750             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 751
 752             /* Update potential sum for this i atom from the interaction with this j atom. */
 753             velec            = _mm_andnot_ps(dummy_mask,velec);
 754             velecsum         = _mm_add_ps(velecsum,velec);
 755
 756             fscal            = felec;
 757
 758             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 759
 760              /* Update vectorial force */
 761             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 762             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 763             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 764
 765             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 766             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 767             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* REACTION-FIELD ELECTROSTATICS */
 774             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 775             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_andnot_ps(dummy_mask,velec);
 779             velecsum         = _mm_add_ps(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 784
 785              /* Update vectorial force */
 786             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 787             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 788             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 789
 790             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 791             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 792             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 800             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_andnot_ps(dummy_mask,velec);
 804             velecsum         = _mm_add_ps(velecsum,velec);
 805
 806             fscal            = felec;
 807
 808             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 809
 810              /* Update vectorial force */
 811             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 812             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 813             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 814
 815             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 816             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 817             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             /* REACTION-FIELD ELECTROSTATICS */
 824             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 825             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 826
 827             /* Update potential sum for this i atom from the interaction with this j atom. */
 828             velec            = _mm_andnot_ps(dummy_mask,velec);
 829             velecsum         = _mm_add_ps(velecsum,velec);
 830
 831             fscal            = felec;
 832
 833             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 834
 835              /* Update vectorial force */
 836             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 837             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 838             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 839
 840             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 841             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 842             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 843
 844             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 845             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 846             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 847             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 848
 849             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 850                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 851
 852             /* Inner loop uses 327 flops */
 853         }
 854
 855         /* End of innermost loop */
 856
 857         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 858                                               f+i_coord_offset,fshift+i_shift_offset);
 859
 860         ggid                        = gid[iidx];
 861         /* Update potential energies */
 862         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 863         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 864
 865         /* Increment number of inner iterations */
 866         inneriter                  += j_index_end - j_index_start;
 867
 868         /* Outer loop uses 20 flops */
 869     }
 870
 871     /* Increment number of outer iterations */
 872     outeriter        += nri;
 873
 874     /* Update outer/inner flops */
 875
 876     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*327);
 877 }
 878 /*
 879  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_single
 880  * Electrostatics interaction: ReactionField
 881  * VdW interaction:            LennardJones
 882  * Geometry:                   Water3-Water3
 883  * Calculate force/pot:        Force
 884  */
 885 void
 886 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_single
 887                     (t_nblist                    * gmx_restrict       nlist,
 888                      rvec                        * gmx_restrict          xx,
 889                      rvec                        * gmx_restrict          ff,
 890                      struct t_forcerec           * gmx_restrict          fr,
 891                      t_mdatoms                   * gmx_restrict     mdatoms,
 892                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 893                      t_nrnb                      * gmx_restrict        nrnb)
 894 {
 895     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 896      * just 0 for non-waters.
 897      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 898      * jnr indices corresponding to data put in the four positions in the SIMD register.
 899      */
 900     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 901     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 902     int              jnrA,jnrB,jnrC,jnrD;
 903     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 904     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 905     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 906     real             rcutoff_scalar;
 907     real             *shiftvec,*fshift,*x,*f;
 908     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 909     real             scratch[4*DIM];
 910     __m128           fscal,rcutoff,rcutoff2,jidxall;
 911     int              vdwioffset0;
 912     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 913     int              vdwioffset1;
 914     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 915     int              vdwioffset2;
 916     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 917     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 918     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 919     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 920     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 921     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 922     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 923     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 924     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 925     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 926     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 927     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 928     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 929     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 930     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 931     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 932     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 933     real             *charge;
 934     int              nvdwtype;
 935     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 936     int              *vdwtype;
 937     real             *vdwparam;
 938     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 939     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 940     __m128           dummy_mask,cutoff_mask;
 941     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 942     __m128           one     = _mm_set1_ps(1.0);
 943     __m128           two     = _mm_set1_ps(2.0);
 944     x                = xx[0];
 945     f                = ff[0];
 946
 947     nri              = nlist->nri;
 948     iinr             = nlist->iinr;
 949     jindex           = nlist->jindex;
 950     jjnr             = nlist->jjnr;
 951     shiftidx         = nlist->shift;
 952     gid              = nlist->gid;
 953     shiftvec         = fr->shift_vec[0];
 954     fshift           = fr->fshift[0];
 955     facel            = _mm_set1_ps(fr->ic->epsfac);
 956     charge           = mdatoms->chargeA;
 957     krf              = _mm_set1_ps(fr->ic->k_rf);
 958     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 959     crf              = _mm_set1_ps(fr->ic->c_rf);
 960     nvdwtype         = fr->ntype;
 961     vdwparam         = fr->nbfp;
 962     vdwtype          = mdatoms->typeA;
 963
 964     /* Setup water-specific parameters */
 965     inr              = nlist->iinr[0];
 966     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 967     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 968     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 969     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 970
 971     jq0              = _mm_set1_ps(charge[inr+0]);
 972     jq1              = _mm_set1_ps(charge[inr+1]);
 973     jq2              = _mm_set1_ps(charge[inr+2]);
 974     vdwjidx0A        = 2*vdwtype[inr+0];
 975     qq00             = _mm_mul_ps(iq0,jq0);
 976     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 977     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 978     qq01             = _mm_mul_ps(iq0,jq1);
 979     qq02             = _mm_mul_ps(iq0,jq2);
 980     qq10             = _mm_mul_ps(iq1,jq0);
 981     qq11             = _mm_mul_ps(iq1,jq1);
 982     qq12             = _mm_mul_ps(iq1,jq2);
 983     qq20             = _mm_mul_ps(iq2,jq0);
 984     qq21             = _mm_mul_ps(iq2,jq1);
 985     qq22             = _mm_mul_ps(iq2,jq2);
 986
 987     /* Avoid stupid compiler warnings */
 988     jnrA = jnrB = jnrC = jnrD = 0;
 989     j_coord_offsetA = 0;
 990     j_coord_offsetB = 0;
 991     j_coord_offsetC = 0;
 992     j_coord_offsetD = 0;
 993
 994     outeriter        = 0;
 995     inneriter        = 0;
 996
 997     for(iidx=0;iidx<4*DIM;iidx++)
 998     {
 999         scratch[iidx] = 0.0;
1000     }
1001
1002     /* Start outer loop over neighborlists */
1003     for(iidx=0; iidx<nri; iidx++)
1004     {
1005         /* Load shift vector for this list */
1006         i_shift_offset   = DIM*shiftidx[iidx];
1007
1008         /* Load limits for loop over neighbors */
1009         j_index_start    = jindex[iidx];
1010         j_index_end      = jindex[iidx+1];
1011
1012         /* Get outer coordinate index */
1013         inr              = iinr[iidx];
1014         i_coord_offset   = DIM*inr;
1015
1016         /* Load i particle coords and add shift vector */
1017         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1018                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1019
1020         fix0             = _mm_setzero_ps();
1021         fiy0             = _mm_setzero_ps();
1022         fiz0             = _mm_setzero_ps();
1023         fix1             = _mm_setzero_ps();
1024         fiy1             = _mm_setzero_ps();
1025         fiz1             = _mm_setzero_ps();
1026         fix2             = _mm_setzero_ps();
1027         fiy2             = _mm_setzero_ps();
1028         fiz2             = _mm_setzero_ps();
1029
1030         /* Start inner kernel loop */
1031         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1032         {
1033
1034             /* Get j neighbor index, and coordinate index */
1035             jnrA             = jjnr[jidx];
1036             jnrB             = jjnr[jidx+1];
1037             jnrC             = jjnr[jidx+2];
1038             jnrD             = jjnr[jidx+3];
1039             j_coord_offsetA  = DIM*jnrA;
1040             j_coord_offsetB  = DIM*jnrB;
1041             j_coord_offsetC  = DIM*jnrC;
1042             j_coord_offsetD  = DIM*jnrD;
1043
1044             /* load j atom coordinates */
1045             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1046                                               x+j_coord_offsetC,x+j_coord_offsetD,
1047                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1048
1049             /* Calculate displacement vector */
1050             dx00             = _mm_sub_ps(ix0,jx0);
1051             dy00             = _mm_sub_ps(iy0,jy0);
1052             dz00             = _mm_sub_ps(iz0,jz0);
1053             dx01             = _mm_sub_ps(ix0,jx1);
1054             dy01             = _mm_sub_ps(iy0,jy1);
1055             dz01             = _mm_sub_ps(iz0,jz1);
1056             dx02             = _mm_sub_ps(ix0,jx2);
1057             dy02             = _mm_sub_ps(iy0,jy2);
1058             dz02             = _mm_sub_ps(iz0,jz2);
1059             dx10             = _mm_sub_ps(ix1,jx0);
1060             dy10             = _mm_sub_ps(iy1,jy0);
1061             dz10             = _mm_sub_ps(iz1,jz0);
1062             dx11             = _mm_sub_ps(ix1,jx1);
1063             dy11             = _mm_sub_ps(iy1,jy1);
1064             dz11             = _mm_sub_ps(iz1,jz1);
1065             dx12             = _mm_sub_ps(ix1,jx2);
1066             dy12             = _mm_sub_ps(iy1,jy2);
1067             dz12             = _mm_sub_ps(iz1,jz2);
1068             dx20             = _mm_sub_ps(ix2,jx0);
1069             dy20             = _mm_sub_ps(iy2,jy0);
1070             dz20             = _mm_sub_ps(iz2,jz0);
1071             dx21             = _mm_sub_ps(ix2,jx1);
1072             dy21             = _mm_sub_ps(iy2,jy1);
1073             dz21             = _mm_sub_ps(iz2,jz1);
1074             dx22             = _mm_sub_ps(ix2,jx2);
1075             dy22             = _mm_sub_ps(iy2,jy2);
1076             dz22             = _mm_sub_ps(iz2,jz2);
1077
1078             /* Calculate squared distance and things based on it */
1079             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1080             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1081             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1082             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1083             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1084             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1085             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1086             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1087             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1088
1089             rinv00           = avx128fma_invsqrt_f(rsq00);
1090             rinv01           = avx128fma_invsqrt_f(rsq01);
1091             rinv02           = avx128fma_invsqrt_f(rsq02);
1092             rinv10           = avx128fma_invsqrt_f(rsq10);
1093             rinv11           = avx128fma_invsqrt_f(rsq11);
1094             rinv12           = avx128fma_invsqrt_f(rsq12);
1095             rinv20           = avx128fma_invsqrt_f(rsq20);
1096             rinv21           = avx128fma_invsqrt_f(rsq21);
1097             rinv22           = avx128fma_invsqrt_f(rsq22);
1098
1099             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1100             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1101             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1102             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1103             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1104             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1105             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1106             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1107             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1108
1109             fjx0             = _mm_setzero_ps();
1110             fjy0             = _mm_setzero_ps();
1111             fjz0             = _mm_setzero_ps();
1112             fjx1             = _mm_setzero_ps();
1113             fjy1             = _mm_setzero_ps();
1114             fjz1             = _mm_setzero_ps();
1115             fjx2             = _mm_setzero_ps();
1116             fjy2             = _mm_setzero_ps();
1117             fjz2             = _mm_setzero_ps();
1118
1119             /**************************
1120              * CALCULATE INTERACTIONS *
1121              **************************/
1122
1123             /* REACTION-FIELD ELECTROSTATICS */
1124             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1125
1126             /* LENNARD-JONES DISPERSION/REPULSION */
1127
1128             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1129             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1130
1131             fscal            = _mm_add_ps(felec,fvdw);
1132
1133              /* Update vectorial force */
1134             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1135             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1136             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1137
1138             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1139             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1140             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1141
1142             /**************************
1143              * CALCULATE INTERACTIONS *
1144              **************************/
1145
1146             /* REACTION-FIELD ELECTROSTATICS */
1147             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1148
1149             fscal            = felec;
1150
1151              /* Update vectorial force */
1152             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1153             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1154             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1155
1156             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1157             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1158             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1159
1160             /**************************
1161              * CALCULATE INTERACTIONS *
1162              **************************/
1163
1164             /* REACTION-FIELD ELECTROSTATICS */
1165             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1166
1167             fscal            = felec;
1168
1169              /* Update vectorial force */
1170             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1171             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1172             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1173
1174             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1175             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1176             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1177
1178             /**************************
1179              * CALCULATE INTERACTIONS *
1180              **************************/
1181
1182             /* REACTION-FIELD ELECTROSTATICS */
1183             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1184
1185             fscal            = felec;
1186
1187              /* Update vectorial force */
1188             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1189             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1190             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1191
1192             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1193             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1194             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1195
1196             /**************************
1197              * CALCULATE INTERACTIONS *
1198              **************************/
1199
1200             /* REACTION-FIELD ELECTROSTATICS */
1201             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1202
1203             fscal            = felec;
1204
1205              /* Update vectorial force */
1206             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1207             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1208             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1209
1210             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1211             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1212             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1213
1214             /**************************
1215              * CALCULATE INTERACTIONS *
1216              **************************/
1217
1218             /* REACTION-FIELD ELECTROSTATICS */
1219             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1220
1221             fscal            = felec;
1222
1223              /* Update vectorial force */
1224             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1225             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1226             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1227
1228             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1229             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1230             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1231
1232             /**************************
1233              * CALCULATE INTERACTIONS *
1234              **************************/
1235
1236             /* REACTION-FIELD ELECTROSTATICS */
1237             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1238
1239             fscal            = felec;
1240
1241              /* Update vectorial force */
1242             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1243             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1244             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1245
1246             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1247             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1248             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1249
1250             /**************************
1251              * CALCULATE INTERACTIONS *
1252              **************************/
1253
1254             /* REACTION-FIELD ELECTROSTATICS */
1255             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1256
1257             fscal            = felec;
1258
1259              /* Update vectorial force */
1260             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1261             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1262             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1263
1264             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1265             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1266             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1267
1268             /**************************
1269              * CALCULATE INTERACTIONS *
1270              **************************/
1271
1272             /* REACTION-FIELD ELECTROSTATICS */
1273             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1274
1275             fscal            = felec;
1276
1277              /* Update vectorial force */
1278             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1279             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1280             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1281
1282             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1283             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1284             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1285
1286             fjptrA             = f+j_coord_offsetA;
1287             fjptrB             = f+j_coord_offsetB;
1288             fjptrC             = f+j_coord_offsetC;
1289             fjptrD             = f+j_coord_offsetD;
1290
1291             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1292                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1293
1294             /* Inner loop uses 277 flops */
1295         }
1296
1297         if(jidx<j_index_end)
1298         {
1299
1300             /* Get j neighbor index, and coordinate index */
1301             jnrlistA         = jjnr[jidx];
1302             jnrlistB         = jjnr[jidx+1];
1303             jnrlistC         = jjnr[jidx+2];
1304             jnrlistD         = jjnr[jidx+3];
1305             /* Sign of each element will be negative for non-real atoms.
1306              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1307              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1308              */
1309             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1310             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1311             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1312             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1313             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1314             j_coord_offsetA  = DIM*jnrA;
1315             j_coord_offsetB  = DIM*jnrB;
1316             j_coord_offsetC  = DIM*jnrC;
1317             j_coord_offsetD  = DIM*jnrD;
1318
1319             /* load j atom coordinates */
1320             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1321                                               x+j_coord_offsetC,x+j_coord_offsetD,
1322                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1323
1324             /* Calculate displacement vector */
1325             dx00             = _mm_sub_ps(ix0,jx0);
1326             dy00             = _mm_sub_ps(iy0,jy0);
1327             dz00             = _mm_sub_ps(iz0,jz0);
1328             dx01             = _mm_sub_ps(ix0,jx1);
1329             dy01             = _mm_sub_ps(iy0,jy1);
1330             dz01             = _mm_sub_ps(iz0,jz1);
1331             dx02             = _mm_sub_ps(ix0,jx2);
1332             dy02             = _mm_sub_ps(iy0,jy2);
1333             dz02             = _mm_sub_ps(iz0,jz2);
1334             dx10             = _mm_sub_ps(ix1,jx0);
1335             dy10             = _mm_sub_ps(iy1,jy0);
1336             dz10             = _mm_sub_ps(iz1,jz0);
1337             dx11             = _mm_sub_ps(ix1,jx1);
1338             dy11             = _mm_sub_ps(iy1,jy1);
1339             dz11             = _mm_sub_ps(iz1,jz1);
1340             dx12             = _mm_sub_ps(ix1,jx2);
1341             dy12             = _mm_sub_ps(iy1,jy2);
1342             dz12             = _mm_sub_ps(iz1,jz2);
1343             dx20             = _mm_sub_ps(ix2,jx0);
1344             dy20             = _mm_sub_ps(iy2,jy0);
1345             dz20             = _mm_sub_ps(iz2,jz0);
1346             dx21             = _mm_sub_ps(ix2,jx1);
1347             dy21             = _mm_sub_ps(iy2,jy1);
1348             dz21             = _mm_sub_ps(iz2,jz1);
1349             dx22             = _mm_sub_ps(ix2,jx2);
1350             dy22             = _mm_sub_ps(iy2,jy2);
1351             dz22             = _mm_sub_ps(iz2,jz2);
1352
1353             /* Calculate squared distance and things based on it */
1354             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1355             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1356             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1357             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1358             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1359             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1360             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1361             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1362             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1363
1364             rinv00           = avx128fma_invsqrt_f(rsq00);
1365             rinv01           = avx128fma_invsqrt_f(rsq01);
1366             rinv02           = avx128fma_invsqrt_f(rsq02);
1367             rinv10           = avx128fma_invsqrt_f(rsq10);
1368             rinv11           = avx128fma_invsqrt_f(rsq11);
1369             rinv12           = avx128fma_invsqrt_f(rsq12);
1370             rinv20           = avx128fma_invsqrt_f(rsq20);
1371             rinv21           = avx128fma_invsqrt_f(rsq21);
1372             rinv22           = avx128fma_invsqrt_f(rsq22);
1373
1374             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1375             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1376             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1377             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1378             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1379             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1380             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1381             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1382             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1383
1384             fjx0             = _mm_setzero_ps();
1385             fjy0             = _mm_setzero_ps();
1386             fjz0             = _mm_setzero_ps();
1387             fjx1             = _mm_setzero_ps();
1388             fjy1             = _mm_setzero_ps();
1389             fjz1             = _mm_setzero_ps();
1390             fjx2             = _mm_setzero_ps();
1391             fjy2             = _mm_setzero_ps();
1392             fjz2             = _mm_setzero_ps();
1393
1394             /**************************
1395              * CALCULATE INTERACTIONS *
1396              **************************/
1397
1398             /* REACTION-FIELD ELECTROSTATICS */
1399             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1400
1401             /* LENNARD-JONES DISPERSION/REPULSION */
1402
1403             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1404             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1405
1406             fscal            = _mm_add_ps(felec,fvdw);
1407
1408             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1409
1410              /* Update vectorial force */
1411             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1412             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1413             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1414
1415             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1416             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1417             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1418
1419             /**************************
1420              * CALCULATE INTERACTIONS *
1421              **************************/
1422
1423             /* REACTION-FIELD ELECTROSTATICS */
1424             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1425
1426             fscal            = felec;
1427
1428             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1429
1430              /* Update vectorial force */
1431             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1432             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1433             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1434
1435             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1436             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1437             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1438
1439             /**************************
1440              * CALCULATE INTERACTIONS *
1441              **************************/
1442
1443             /* REACTION-FIELD ELECTROSTATICS */
1444             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1445
1446             fscal            = felec;
1447
1448             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1449
1450              /* Update vectorial force */
1451             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1452             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1453             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1454
1455             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1456             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1457             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1458
1459             /**************************
1460              * CALCULATE INTERACTIONS *
1461              **************************/
1462
1463             /* REACTION-FIELD ELECTROSTATICS */
1464             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1465
1466             fscal            = felec;
1467
1468             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1469
1470              /* Update vectorial force */
1471             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1472             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1473             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1474
1475             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1476             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1477             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1478
1479             /**************************
1480              * CALCULATE INTERACTIONS *
1481              **************************/
1482
1483             /* REACTION-FIELD ELECTROSTATICS */
1484             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1485
1486             fscal            = felec;
1487
1488             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1489
1490              /* Update vectorial force */
1491             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1492             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1493             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1494
1495             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1496             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1497             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1498
1499             /**************************
1500              * CALCULATE INTERACTIONS *
1501              **************************/
1502
1503             /* REACTION-FIELD ELECTROSTATICS */
1504             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1505
1506             fscal            = felec;
1507
1508             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1509
1510              /* Update vectorial force */
1511             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1512             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1513             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1514
1515             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1516             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1517             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1518
1519             /**************************
1520              * CALCULATE INTERACTIONS *
1521              **************************/
1522
1523             /* REACTION-FIELD ELECTROSTATICS */
1524             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1525
1526             fscal            = felec;
1527
1528             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1529
1530              /* Update vectorial force */
1531             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1532             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1533             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1534
1535             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1536             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1537             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1538
1539             /**************************
1540              * CALCULATE INTERACTIONS *
1541              **************************/
1542
1543             /* REACTION-FIELD ELECTROSTATICS */
1544             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1549
1550              /* Update vectorial force */
1551             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1552             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1553             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1554
1555             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1556             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1557             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1558
1559             /**************************
1560              * CALCULATE INTERACTIONS *
1561              **************************/
1562
1563             /* REACTION-FIELD ELECTROSTATICS */
1564             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1565
1566             fscal            = felec;
1567
1568             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1569
1570              /* Update vectorial force */
1571             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1572             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1573             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1574
1575             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1576             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1577             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1578
1579             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1580             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1581             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1582             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1583
1584             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1585                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1586
1587             /* Inner loop uses 277 flops */
1588         }
1589
1590         /* End of innermost loop */
1591
1592         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1593                                               f+i_coord_offset,fshift+i_shift_offset);
1594
1595         /* Increment number of inner iterations */
1596         inneriter                  += j_index_end - j_index_start;
1597
1598         /* Outer loop uses 18 flops */
1599     }
1600
1601     /* Increment number of outer iterations */
1602     outeriter        += nri;
1603
1604     /* Update outer/inner flops */
1605
1606     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1607 }