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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_single
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            None
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     __m128           dummy_mask,cutoff_mask;
 109     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 110     __m128           one     = _mm_set1_ps(1.0);
 111     __m128           two     = _mm_set1_ps(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm_set1_ps(fr->epsfac);
 124     charge           = mdatoms->chargeA;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131
 132     jq0              = _mm_set1_ps(charge[inr+0]);
 133     jq1              = _mm_set1_ps(charge[inr+1]);
 134     jq2              = _mm_set1_ps(charge[inr+2]);
 135     qq00             = _mm_mul_ps(iq0,jq0);
 136     qq01             = _mm_mul_ps(iq0,jq1);
 137     qq02             = _mm_mul_ps(iq0,jq2);
 138     qq10             = _mm_mul_ps(iq1,jq0);
 139     qq11             = _mm_mul_ps(iq1,jq1);
 140     qq12             = _mm_mul_ps(iq1,jq2);
 141     qq20             = _mm_mul_ps(iq2,jq0);
 142     qq21             = _mm_mul_ps(iq2,jq1);
 143     qq22             = _mm_mul_ps(iq2,jq2);
 144
 145     /* Avoid stupid compiler warnings */
 146     jnrA = jnrB = jnrC = jnrD = 0;
 147     j_coord_offsetA = 0;
 148     j_coord_offsetB = 0;
 149     j_coord_offsetC = 0;
 150     j_coord_offsetD = 0;
 151
 152     outeriter        = 0;
 153     inneriter        = 0;
 154
 155     for(iidx=0;iidx<4*DIM;iidx++)
 156     {
 157         scratch[iidx] = 0.0;
 158     }
 159
 160     /* Start outer loop over neighborlists */
 161     for(iidx=0; iidx<nri; iidx++)
 162     {
 163         /* Load shift vector for this list */
 164         i_shift_offset   = DIM*shiftidx[iidx];
 165
 166         /* Load limits for loop over neighbors */
 167         j_index_start    = jindex[iidx];
 168         j_index_end      = jindex[iidx+1];
 169
 170         /* Get outer coordinate index */
 171         inr              = iinr[iidx];
 172         i_coord_offset   = DIM*inr;
 173
 174         /* Load i particle coords and add shift vector */
 175         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 176                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 177
 178         fix0             = _mm_setzero_ps();
 179         fiy0             = _mm_setzero_ps();
 180         fiz0             = _mm_setzero_ps();
 181         fix1             = _mm_setzero_ps();
 182         fiy1             = _mm_setzero_ps();
 183         fiz1             = _mm_setzero_ps();
 184         fix2             = _mm_setzero_ps();
 185         fiy2             = _mm_setzero_ps();
 186         fiz2             = _mm_setzero_ps();
 187
 188         /* Reset potential sums */
 189         velecsum         = _mm_setzero_ps();
 190
 191         /* Start inner kernel loop */
 192         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 193         {
 194
 195             /* Get j neighbor index, and coordinate index */
 196             jnrA             = jjnr[jidx];
 197             jnrB             = jjnr[jidx+1];
 198             jnrC             = jjnr[jidx+2];
 199             jnrD             = jjnr[jidx+3];
 200             j_coord_offsetA  = DIM*jnrA;
 201             j_coord_offsetB  = DIM*jnrB;
 202             j_coord_offsetC  = DIM*jnrC;
 203             j_coord_offsetD  = DIM*jnrD;
 204
 205             /* load j atom coordinates */
 206             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 207                                               x+j_coord_offsetC,x+j_coord_offsetD,
 208                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 209
 210             /* Calculate displacement vector */
 211             dx00             = _mm_sub_ps(ix0,jx0);
 212             dy00             = _mm_sub_ps(iy0,jy0);
 213             dz00             = _mm_sub_ps(iz0,jz0);
 214             dx01             = _mm_sub_ps(ix0,jx1);
 215             dy01             = _mm_sub_ps(iy0,jy1);
 216             dz01             = _mm_sub_ps(iz0,jz1);
 217             dx02             = _mm_sub_ps(ix0,jx2);
 218             dy02             = _mm_sub_ps(iy0,jy2);
 219             dz02             = _mm_sub_ps(iz0,jz2);
 220             dx10             = _mm_sub_ps(ix1,jx0);
 221             dy10             = _mm_sub_ps(iy1,jy0);
 222             dz10             = _mm_sub_ps(iz1,jz0);
 223             dx11             = _mm_sub_ps(ix1,jx1);
 224             dy11             = _mm_sub_ps(iy1,jy1);
 225             dz11             = _mm_sub_ps(iz1,jz1);
 226             dx12             = _mm_sub_ps(ix1,jx2);
 227             dy12             = _mm_sub_ps(iy1,jy2);
 228             dz12             = _mm_sub_ps(iz1,jz2);
 229             dx20             = _mm_sub_ps(ix2,jx0);
 230             dy20             = _mm_sub_ps(iy2,jy0);
 231             dz20             = _mm_sub_ps(iz2,jz0);
 232             dx21             = _mm_sub_ps(ix2,jx1);
 233             dy21             = _mm_sub_ps(iy2,jy1);
 234             dz21             = _mm_sub_ps(iz2,jz1);
 235             dx22             = _mm_sub_ps(ix2,jx2);
 236             dy22             = _mm_sub_ps(iy2,jy2);
 237             dz22             = _mm_sub_ps(iz2,jz2);
 238
 239             /* Calculate squared distance and things based on it */
 240             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 241             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 242             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 243             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 244             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 245             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 246             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 247             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 248             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 249
 250             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 251             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 252             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 253             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 254             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 255             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 256             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 257             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 258             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 259
 260             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 261             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 262             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 263             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 264             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 265             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 266             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 267             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 268             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 269
 270             fjx0             = _mm_setzero_ps();
 271             fjy0             = _mm_setzero_ps();
 272             fjz0             = _mm_setzero_ps();
 273             fjx1             = _mm_setzero_ps();
 274             fjy1             = _mm_setzero_ps();
 275             fjz1             = _mm_setzero_ps();
 276             fjx2             = _mm_setzero_ps();
 277             fjy2             = _mm_setzero_ps();
 278             fjz2             = _mm_setzero_ps();
 279
 280             /**************************
 281              * CALCULATE INTERACTIONS *
 282              **************************/
 283
 284             /* COULOMB ELECTROSTATICS */
 285             velec            = _mm_mul_ps(qq00,rinv00);
 286             felec            = _mm_mul_ps(velec,rinvsq00);
 287
 288             /* Update potential sum for this i atom from the interaction with this j atom. */
 289             velecsum         = _mm_add_ps(velecsum,velec);
 290
 291             fscal            = felec;
 292
 293              /* Update vectorial force */
 294             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 295             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 296             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 297
 298             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 299             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 300             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             /* COULOMB ELECTROSTATICS */
 307             velec            = _mm_mul_ps(qq01,rinv01);
 308             felec            = _mm_mul_ps(velec,rinvsq01);
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_ps(velecsum,velec);
 312
 313             fscal            = felec;
 314
 315              /* Update vectorial force */
 316             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 317             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 318             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 319
 320             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 321             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 322             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             /* COULOMB ELECTROSTATICS */
 329             velec            = _mm_mul_ps(qq02,rinv02);
 330             felec            = _mm_mul_ps(velec,rinvsq02);
 331
 332             /* Update potential sum for this i atom from the interaction with this j atom. */
 333             velecsum         = _mm_add_ps(velecsum,velec);
 334
 335             fscal            = felec;
 336
 337              /* Update vectorial force */
 338             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 339             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 340             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 341
 342             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 343             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 344             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 345
 346             /**************************
 347              * CALCULATE INTERACTIONS *
 348              **************************/
 349
 350             /* COULOMB ELECTROSTATICS */
 351             velec            = _mm_mul_ps(qq10,rinv10);
 352             felec            = _mm_mul_ps(velec,rinvsq10);
 353
 354             /* Update potential sum for this i atom from the interaction with this j atom. */
 355             velecsum         = _mm_add_ps(velecsum,velec);
 356
 357             fscal            = felec;
 358
 359              /* Update vectorial force */
 360             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 361             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 362             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 363
 364             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 365             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 366             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             /* COULOMB ELECTROSTATICS */
 373             velec            = _mm_mul_ps(qq11,rinv11);
 374             felec            = _mm_mul_ps(velec,rinvsq11);
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velecsum         = _mm_add_ps(velecsum,velec);
 378
 379             fscal            = felec;
 380
 381              /* Update vectorial force */
 382             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 383             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 384             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 385
 386             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 387             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 388             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 389
 390             /**************************
 391              * CALCULATE INTERACTIONS *
 392              **************************/
 393
 394             /* COULOMB ELECTROSTATICS */
 395             velec            = _mm_mul_ps(qq12,rinv12);
 396             felec            = _mm_mul_ps(velec,rinvsq12);
 397
 398             /* Update potential sum for this i atom from the interaction with this j atom. */
 399             velecsum         = _mm_add_ps(velecsum,velec);
 400
 401             fscal            = felec;
 402
 403              /* Update vectorial force */
 404             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 405             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 406             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 407
 408             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 409             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 410             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 411
 412             /**************************
 413              * CALCULATE INTERACTIONS *
 414              **************************/
 415
 416             /* COULOMB ELECTROSTATICS */
 417             velec            = _mm_mul_ps(qq20,rinv20);
 418             felec            = _mm_mul_ps(velec,rinvsq20);
 419
 420             /* Update potential sum for this i atom from the interaction with this j atom. */
 421             velecsum         = _mm_add_ps(velecsum,velec);
 422
 423             fscal            = felec;
 424
 425              /* Update vectorial force */
 426             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 427             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 428             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 429
 430             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 431             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 432             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 433
 434             /**************************
 435              * CALCULATE INTERACTIONS *
 436              **************************/
 437
 438             /* COULOMB ELECTROSTATICS */
 439             velec            = _mm_mul_ps(qq21,rinv21);
 440             felec            = _mm_mul_ps(velec,rinvsq21);
 441
 442             /* Update potential sum for this i atom from the interaction with this j atom. */
 443             velecsum         = _mm_add_ps(velecsum,velec);
 444
 445             fscal            = felec;
 446
 447              /* Update vectorial force */
 448             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 449             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 450             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 451
 452             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 453             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 454             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             /* COULOMB ELECTROSTATICS */
 461             velec            = _mm_mul_ps(qq22,rinv22);
 462             felec            = _mm_mul_ps(velec,rinvsq22);
 463
 464             /* Update potential sum for this i atom from the interaction with this j atom. */
 465             velecsum         = _mm_add_ps(velecsum,velec);
 466
 467             fscal            = felec;
 468
 469              /* Update vectorial force */
 470             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 471             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 472             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 473
 474             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 475             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 476             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 477
 478             fjptrA             = f+j_coord_offsetA;
 479             fjptrB             = f+j_coord_offsetB;
 480             fjptrC             = f+j_coord_offsetC;
 481             fjptrD             = f+j_coord_offsetD;
 482
 483             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 484                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 485
 486             /* Inner loop uses 279 flops */
 487         }
 488
 489         if(jidx<j_index_end)
 490         {
 491
 492             /* Get j neighbor index, and coordinate index */
 493             jnrlistA         = jjnr[jidx];
 494             jnrlistB         = jjnr[jidx+1];
 495             jnrlistC         = jjnr[jidx+2];
 496             jnrlistD         = jjnr[jidx+3];
 497             /* Sign of each element will be negative for non-real atoms.
 498              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 499              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 500              */
 501             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 502             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 503             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 504             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 505             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 506             j_coord_offsetA  = DIM*jnrA;
 507             j_coord_offsetB  = DIM*jnrB;
 508             j_coord_offsetC  = DIM*jnrC;
 509             j_coord_offsetD  = DIM*jnrD;
 510
 511             /* load j atom coordinates */
 512             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 513                                               x+j_coord_offsetC,x+j_coord_offsetD,
 514                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 515
 516             /* Calculate displacement vector */
 517             dx00             = _mm_sub_ps(ix0,jx0);
 518             dy00             = _mm_sub_ps(iy0,jy0);
 519             dz00             = _mm_sub_ps(iz0,jz0);
 520             dx01             = _mm_sub_ps(ix0,jx1);
 521             dy01             = _mm_sub_ps(iy0,jy1);
 522             dz01             = _mm_sub_ps(iz0,jz1);
 523             dx02             = _mm_sub_ps(ix0,jx2);
 524             dy02             = _mm_sub_ps(iy0,jy2);
 525             dz02             = _mm_sub_ps(iz0,jz2);
 526             dx10             = _mm_sub_ps(ix1,jx0);
 527             dy10             = _mm_sub_ps(iy1,jy0);
 528             dz10             = _mm_sub_ps(iz1,jz0);
 529             dx11             = _mm_sub_ps(ix1,jx1);
 530             dy11             = _mm_sub_ps(iy1,jy1);
 531             dz11             = _mm_sub_ps(iz1,jz1);
 532             dx12             = _mm_sub_ps(ix1,jx2);
 533             dy12             = _mm_sub_ps(iy1,jy2);
 534             dz12             = _mm_sub_ps(iz1,jz2);
 535             dx20             = _mm_sub_ps(ix2,jx0);
 536             dy20             = _mm_sub_ps(iy2,jy0);
 537             dz20             = _mm_sub_ps(iz2,jz0);
 538             dx21             = _mm_sub_ps(ix2,jx1);
 539             dy21             = _mm_sub_ps(iy2,jy1);
 540             dz21             = _mm_sub_ps(iz2,jz1);
 541             dx22             = _mm_sub_ps(ix2,jx2);
 542             dy22             = _mm_sub_ps(iy2,jy2);
 543             dz22             = _mm_sub_ps(iz2,jz2);
 544
 545             /* Calculate squared distance and things based on it */
 546             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 547             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 548             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 549             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 550             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 551             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 552             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 553             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 554             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 555
 556             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 557             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 558             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 559             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 560             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 561             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 562             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 563             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 564             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 565
 566             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 567             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 568             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 569             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 570             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 571             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 572             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 573             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 574             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 575
 576             fjx0             = _mm_setzero_ps();
 577             fjy0             = _mm_setzero_ps();
 578             fjz0             = _mm_setzero_ps();
 579             fjx1             = _mm_setzero_ps();
 580             fjy1             = _mm_setzero_ps();
 581             fjz1             = _mm_setzero_ps();
 582             fjx2             = _mm_setzero_ps();
 583             fjy2             = _mm_setzero_ps();
 584             fjz2             = _mm_setzero_ps();
 585
 586             /**************************
 587              * CALCULATE INTERACTIONS *
 588              **************************/
 589
 590             /* COULOMB ELECTROSTATICS */
 591             velec            = _mm_mul_ps(qq00,rinv00);
 592             felec            = _mm_mul_ps(velec,rinvsq00);
 593
 594             /* Update potential sum for this i atom from the interaction with this j atom. */
 595             velec            = _mm_andnot_ps(dummy_mask,velec);
 596             velecsum         = _mm_add_ps(velecsum,velec);
 597
 598             fscal            = felec;
 599
 600             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 601
 602              /* Update vectorial force */
 603             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 604             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 605             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 606
 607             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 608             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 609             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 610
 611             /**************************
 612              * CALCULATE INTERACTIONS *
 613              **************************/
 614
 615             /* COULOMB ELECTROSTATICS */
 616             velec            = _mm_mul_ps(qq01,rinv01);
 617             felec            = _mm_mul_ps(velec,rinvsq01);
 618
 619             /* Update potential sum for this i atom from the interaction with this j atom. */
 620             velec            = _mm_andnot_ps(dummy_mask,velec);
 621             velecsum         = _mm_add_ps(velecsum,velec);
 622
 623             fscal            = felec;
 624
 625             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 626
 627              /* Update vectorial force */
 628             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 629             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 630             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 631
 632             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 633             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 634             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 635
 636             /**************************
 637              * CALCULATE INTERACTIONS *
 638              **************************/
 639
 640             /* COULOMB ELECTROSTATICS */
 641             velec            = _mm_mul_ps(qq02,rinv02);
 642             felec            = _mm_mul_ps(velec,rinvsq02);
 643
 644             /* Update potential sum for this i atom from the interaction with this j atom. */
 645             velec            = _mm_andnot_ps(dummy_mask,velec);
 646             velecsum         = _mm_add_ps(velecsum,velec);
 647
 648             fscal            = felec;
 649
 650             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 651
 652              /* Update vectorial force */
 653             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 654             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 655             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 656
 657             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 658             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 659             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 660
 661             /**************************
 662              * CALCULATE INTERACTIONS *
 663              **************************/
 664
 665             /* COULOMB ELECTROSTATICS */
 666             velec            = _mm_mul_ps(qq10,rinv10);
 667             felec            = _mm_mul_ps(velec,rinvsq10);
 668
 669             /* Update potential sum for this i atom from the interaction with this j atom. */
 670             velec            = _mm_andnot_ps(dummy_mask,velec);
 671             velecsum         = _mm_add_ps(velecsum,velec);
 672
 673             fscal            = felec;
 674
 675             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 676
 677              /* Update vectorial force */
 678             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 679             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 680             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 681
 682             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 683             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 684             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 685
 686             /**************************
 687              * CALCULATE INTERACTIONS *
 688              **************************/
 689
 690             /* COULOMB ELECTROSTATICS */
 691             velec            = _mm_mul_ps(qq11,rinv11);
 692             felec            = _mm_mul_ps(velec,rinvsq11);
 693
 694             /* Update potential sum for this i atom from the interaction with this j atom. */
 695             velec            = _mm_andnot_ps(dummy_mask,velec);
 696             velecsum         = _mm_add_ps(velecsum,velec);
 697
 698             fscal            = felec;
 699
 700             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 701
 702              /* Update vectorial force */
 703             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 704             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 705             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 706
 707             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 708             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 709             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 710
 711             /**************************
 712              * CALCULATE INTERACTIONS *
 713              **************************/
 714
 715             /* COULOMB ELECTROSTATICS */
 716             velec            = _mm_mul_ps(qq12,rinv12);
 717             felec            = _mm_mul_ps(velec,rinvsq12);
 718
 719             /* Update potential sum for this i atom from the interaction with this j atom. */
 720             velec            = _mm_andnot_ps(dummy_mask,velec);
 721             velecsum         = _mm_add_ps(velecsum,velec);
 722
 723             fscal            = felec;
 724
 725             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 726
 727              /* Update vectorial force */
 728             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 729             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 730             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 731
 732             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 733             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 734             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 735
 736             /**************************
 737              * CALCULATE INTERACTIONS *
 738              **************************/
 739
 740             /* COULOMB ELECTROSTATICS */
 741             velec            = _mm_mul_ps(qq20,rinv20);
 742             felec            = _mm_mul_ps(velec,rinvsq20);
 743
 744             /* Update potential sum for this i atom from the interaction with this j atom. */
 745             velec            = _mm_andnot_ps(dummy_mask,velec);
 746             velecsum         = _mm_add_ps(velecsum,velec);
 747
 748             fscal            = felec;
 749
 750             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 751
 752              /* Update vectorial force */
 753             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 754             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 755             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 756
 757             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 758             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 759             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 760
 761             /**************************
 762              * CALCULATE INTERACTIONS *
 763              **************************/
 764
 765             /* COULOMB ELECTROSTATICS */
 766             velec            = _mm_mul_ps(qq21,rinv21);
 767             felec            = _mm_mul_ps(velec,rinvsq21);
 768
 769             /* Update potential sum for this i atom from the interaction with this j atom. */
 770             velec            = _mm_andnot_ps(dummy_mask,velec);
 771             velecsum         = _mm_add_ps(velecsum,velec);
 772
 773             fscal            = felec;
 774
 775             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 776
 777              /* Update vectorial force */
 778             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 779             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 780             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 781
 782             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 783             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 784             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 785
 786             /**************************
 787              * CALCULATE INTERACTIONS *
 788              **************************/
 789
 790             /* COULOMB ELECTROSTATICS */
 791             velec            = _mm_mul_ps(qq22,rinv22);
 792             felec            = _mm_mul_ps(velec,rinvsq22);
 793
 794             /* Update potential sum for this i atom from the interaction with this j atom. */
 795             velec            = _mm_andnot_ps(dummy_mask,velec);
 796             velecsum         = _mm_add_ps(velecsum,velec);
 797
 798             fscal            = felec;
 799
 800             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 801
 802              /* Update vectorial force */
 803             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 804             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 805             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 806
 807             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 808             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 809             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 810
 811             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 812             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 813             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 814             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 815
 816             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 817                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 818
 819             /* Inner loop uses 279 flops */
 820         }
 821
 822         /* End of innermost loop */
 823
 824         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 825                                               f+i_coord_offset,fshift+i_shift_offset);
 826
 827         ggid                        = gid[iidx];
 828         /* Update potential energies */
 829         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 830
 831         /* Increment number of inner iterations */
 832         inneriter                  += j_index_end - j_index_start;
 833
 834         /* Outer loop uses 19 flops */
 835     }
 836
 837     /* Increment number of outer iterations */
 838     outeriter        += nri;
 839
 840     /* Update outer/inner flops */
 841
 842     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*279);
 843 }
 844 /*
 845  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_single
 846  * Electrostatics interaction: Coulomb
 847  * VdW interaction:            None
 848  * Geometry:                   Water3-Water3
 849  * Calculate force/pot:        Force
 850  */
 851 void
 852 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_single
 853                     (t_nblist                    * gmx_restrict       nlist,
 854                      rvec                        * gmx_restrict          xx,
 855                      rvec                        * gmx_restrict          ff,
 856                      t_forcerec                  * gmx_restrict          fr,
 857                      t_mdatoms                   * gmx_restrict     mdatoms,
 858                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 859                      t_nrnb                      * gmx_restrict        nrnb)
 860 {
 861     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 862      * just 0 for non-waters.
 863      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 864      * jnr indices corresponding to data put in the four positions in the SIMD register.
 865      */
 866     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 867     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 868     int              jnrA,jnrB,jnrC,jnrD;
 869     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 870     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 871     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 872     real             rcutoff_scalar;
 873     real             *shiftvec,*fshift,*x,*f;
 874     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 875     real             scratch[4*DIM];
 876     __m128           fscal,rcutoff,rcutoff2,jidxall;
 877     int              vdwioffset0;
 878     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 879     int              vdwioffset1;
 880     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 881     int              vdwioffset2;
 882     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 883     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 884     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 885     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 886     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 887     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 888     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 889     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 890     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 891     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 892     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 893     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 894     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 895     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 896     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 897     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 898     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 899     real             *charge;
 900     __m128           dummy_mask,cutoff_mask;
 901     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 902     __m128           one     = _mm_set1_ps(1.0);
 903     __m128           two     = _mm_set1_ps(2.0);
 904     x                = xx[0];
 905     f                = ff[0];
 906
 907     nri              = nlist->nri;
 908     iinr             = nlist->iinr;
 909     jindex           = nlist->jindex;
 910     jjnr             = nlist->jjnr;
 911     shiftidx         = nlist->shift;
 912     gid              = nlist->gid;
 913     shiftvec         = fr->shift_vec[0];
 914     fshift           = fr->fshift[0];
 915     facel            = _mm_set1_ps(fr->epsfac);
 916     charge           = mdatoms->chargeA;
 917
 918     /* Setup water-specific parameters */
 919     inr              = nlist->iinr[0];
 920     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 921     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 922     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 923
 924     jq0              = _mm_set1_ps(charge[inr+0]);
 925     jq1              = _mm_set1_ps(charge[inr+1]);
 926     jq2              = _mm_set1_ps(charge[inr+2]);
 927     qq00             = _mm_mul_ps(iq0,jq0);
 928     qq01             = _mm_mul_ps(iq0,jq1);
 929     qq02             = _mm_mul_ps(iq0,jq2);
 930     qq10             = _mm_mul_ps(iq1,jq0);
 931     qq11             = _mm_mul_ps(iq1,jq1);
 932     qq12             = _mm_mul_ps(iq1,jq2);
 933     qq20             = _mm_mul_ps(iq2,jq0);
 934     qq21             = _mm_mul_ps(iq2,jq1);
 935     qq22             = _mm_mul_ps(iq2,jq2);
 936
 937     /* Avoid stupid compiler warnings */
 938     jnrA = jnrB = jnrC = jnrD = 0;
 939     j_coord_offsetA = 0;
 940     j_coord_offsetB = 0;
 941     j_coord_offsetC = 0;
 942     j_coord_offsetD = 0;
 943
 944     outeriter        = 0;
 945     inneriter        = 0;
 946
 947     for(iidx=0;iidx<4*DIM;iidx++)
 948     {
 949         scratch[iidx] = 0.0;
 950     }
 951
 952     /* Start outer loop over neighborlists */
 953     for(iidx=0; iidx<nri; iidx++)
 954     {
 955         /* Load shift vector for this list */
 956         i_shift_offset   = DIM*shiftidx[iidx];
 957
 958         /* Load limits for loop over neighbors */
 959         j_index_start    = jindex[iidx];
 960         j_index_end      = jindex[iidx+1];
 961
 962         /* Get outer coordinate index */
 963         inr              = iinr[iidx];
 964         i_coord_offset   = DIM*inr;
 965
 966         /* Load i particle coords and add shift vector */
 967         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 968                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 969
 970         fix0             = _mm_setzero_ps();
 971         fiy0             = _mm_setzero_ps();
 972         fiz0             = _mm_setzero_ps();
 973         fix1             = _mm_setzero_ps();
 974         fiy1             = _mm_setzero_ps();
 975         fiz1             = _mm_setzero_ps();
 976         fix2             = _mm_setzero_ps();
 977         fiy2             = _mm_setzero_ps();
 978         fiz2             = _mm_setzero_ps();
 979
 980         /* Start inner kernel loop */
 981         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 982         {
 983
 984             /* Get j neighbor index, and coordinate index */
 985             jnrA             = jjnr[jidx];
 986             jnrB             = jjnr[jidx+1];
 987             jnrC             = jjnr[jidx+2];
 988             jnrD             = jjnr[jidx+3];
 989             j_coord_offsetA  = DIM*jnrA;
 990             j_coord_offsetB  = DIM*jnrB;
 991             j_coord_offsetC  = DIM*jnrC;
 992             j_coord_offsetD  = DIM*jnrD;
 993
 994             /* load j atom coordinates */
 995             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 996                                               x+j_coord_offsetC,x+j_coord_offsetD,
 997                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 998
 999             /* Calculate displacement vector */
1000             dx00             = _mm_sub_ps(ix0,jx0);
1001             dy00             = _mm_sub_ps(iy0,jy0);
1002             dz00             = _mm_sub_ps(iz0,jz0);
1003             dx01             = _mm_sub_ps(ix0,jx1);
1004             dy01             = _mm_sub_ps(iy0,jy1);
1005             dz01             = _mm_sub_ps(iz0,jz1);
1006             dx02             = _mm_sub_ps(ix0,jx2);
1007             dy02             = _mm_sub_ps(iy0,jy2);
1008             dz02             = _mm_sub_ps(iz0,jz2);
1009             dx10             = _mm_sub_ps(ix1,jx0);
1010             dy10             = _mm_sub_ps(iy1,jy0);
1011             dz10             = _mm_sub_ps(iz1,jz0);
1012             dx11             = _mm_sub_ps(ix1,jx1);
1013             dy11             = _mm_sub_ps(iy1,jy1);
1014             dz11             = _mm_sub_ps(iz1,jz1);
1015             dx12             = _mm_sub_ps(ix1,jx2);
1016             dy12             = _mm_sub_ps(iy1,jy2);
1017             dz12             = _mm_sub_ps(iz1,jz2);
1018             dx20             = _mm_sub_ps(ix2,jx0);
1019             dy20             = _mm_sub_ps(iy2,jy0);
1020             dz20             = _mm_sub_ps(iz2,jz0);
1021             dx21             = _mm_sub_ps(ix2,jx1);
1022             dy21             = _mm_sub_ps(iy2,jy1);
1023             dz21             = _mm_sub_ps(iz2,jz1);
1024             dx22             = _mm_sub_ps(ix2,jx2);
1025             dy22             = _mm_sub_ps(iy2,jy2);
1026             dz22             = _mm_sub_ps(iz2,jz2);
1027
1028             /* Calculate squared distance and things based on it */
1029             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1030             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1031             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1032             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1033             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1034             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1035             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1036             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1037             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1038
1039             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1040             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1041             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1042             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1043             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1044             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1045             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1046             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1047             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1048
1049             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1050             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1051             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1052             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1053             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1054             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1055             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1056             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1057             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1058
1059             fjx0             = _mm_setzero_ps();
1060             fjy0             = _mm_setzero_ps();
1061             fjz0             = _mm_setzero_ps();
1062             fjx1             = _mm_setzero_ps();
1063             fjy1             = _mm_setzero_ps();
1064             fjz1             = _mm_setzero_ps();
1065             fjx2             = _mm_setzero_ps();
1066             fjy2             = _mm_setzero_ps();
1067             fjz2             = _mm_setzero_ps();
1068
1069             /**************************
1070              * CALCULATE INTERACTIONS *
1071              **************************/
1072
1073             /* COULOMB ELECTROSTATICS */
1074             velec            = _mm_mul_ps(qq00,rinv00);
1075             felec            = _mm_mul_ps(velec,rinvsq00);
1076
1077             fscal            = felec;
1078
1079              /* Update vectorial force */
1080             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1081             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1082             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1083
1084             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1085             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1086             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1087
1088             /**************************
1089              * CALCULATE INTERACTIONS *
1090              **************************/
1091
1092             /* COULOMB ELECTROSTATICS */
1093             velec            = _mm_mul_ps(qq01,rinv01);
1094             felec            = _mm_mul_ps(velec,rinvsq01);
1095
1096             fscal            = felec;
1097
1098              /* Update vectorial force */
1099             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1100             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1101             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1102
1103             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1104             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1105             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1106
1107             /**************************
1108              * CALCULATE INTERACTIONS *
1109              **************************/
1110
1111             /* COULOMB ELECTROSTATICS */
1112             velec            = _mm_mul_ps(qq02,rinv02);
1113             felec            = _mm_mul_ps(velec,rinvsq02);
1114
1115             fscal            = felec;
1116
1117              /* Update vectorial force */
1118             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1119             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1120             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1121
1122             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1123             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1124             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1125
1126             /**************************
1127              * CALCULATE INTERACTIONS *
1128              **************************/
1129
1130             /* COULOMB ELECTROSTATICS */
1131             velec            = _mm_mul_ps(qq10,rinv10);
1132             felec            = _mm_mul_ps(velec,rinvsq10);
1133
1134             fscal            = felec;
1135
1136              /* Update vectorial force */
1137             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1138             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1139             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1140
1141             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1142             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1143             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1144
1145             /**************************
1146              * CALCULATE INTERACTIONS *
1147              **************************/
1148
1149             /* COULOMB ELECTROSTATICS */
1150             velec            = _mm_mul_ps(qq11,rinv11);
1151             felec            = _mm_mul_ps(velec,rinvsq11);
1152
1153             fscal            = felec;
1154
1155              /* Update vectorial force */
1156             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1157             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1158             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1159
1160             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1161             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1162             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1163
1164             /**************************
1165              * CALCULATE INTERACTIONS *
1166              **************************/
1167
1168             /* COULOMB ELECTROSTATICS */
1169             velec            = _mm_mul_ps(qq12,rinv12);
1170             felec            = _mm_mul_ps(velec,rinvsq12);
1171
1172             fscal            = felec;
1173
1174              /* Update vectorial force */
1175             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1176             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1177             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1178
1179             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1180             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1181             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1182
1183             /**************************
1184              * CALCULATE INTERACTIONS *
1185              **************************/
1186
1187             /* COULOMB ELECTROSTATICS */
1188             velec            = _mm_mul_ps(qq20,rinv20);
1189             felec            = _mm_mul_ps(velec,rinvsq20);
1190
1191             fscal            = felec;
1192
1193              /* Update vectorial force */
1194             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1195             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1196             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1197
1198             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1199             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1200             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1201
1202             /**************************
1203              * CALCULATE INTERACTIONS *
1204              **************************/
1205
1206             /* COULOMB ELECTROSTATICS */
1207             velec            = _mm_mul_ps(qq21,rinv21);
1208             felec            = _mm_mul_ps(velec,rinvsq21);
1209
1210             fscal            = felec;
1211
1212              /* Update vectorial force */
1213             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1214             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1215             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1216
1217             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1218             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1219             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1220
1221             /**************************
1222              * CALCULATE INTERACTIONS *
1223              **************************/
1224
1225             /* COULOMB ELECTROSTATICS */
1226             velec            = _mm_mul_ps(qq22,rinv22);
1227             felec            = _mm_mul_ps(velec,rinvsq22);
1228
1229             fscal            = felec;
1230
1231              /* Update vectorial force */
1232             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1233             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1234             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1235
1236             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1237             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1238             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1239
1240             fjptrA             = f+j_coord_offsetA;
1241             fjptrB             = f+j_coord_offsetB;
1242             fjptrC             = f+j_coord_offsetC;
1243             fjptrD             = f+j_coord_offsetD;
1244
1245             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1246                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1247
1248             /* Inner loop uses 270 flops */
1249         }
1250
1251         if(jidx<j_index_end)
1252         {
1253
1254             /* Get j neighbor index, and coordinate index */
1255             jnrlistA         = jjnr[jidx];
1256             jnrlistB         = jjnr[jidx+1];
1257             jnrlistC         = jjnr[jidx+2];
1258             jnrlistD         = jjnr[jidx+3];
1259             /* Sign of each element will be negative for non-real atoms.
1260              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1261              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1262              */
1263             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1264             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1265             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1266             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1267             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1268             j_coord_offsetA  = DIM*jnrA;
1269             j_coord_offsetB  = DIM*jnrB;
1270             j_coord_offsetC  = DIM*jnrC;
1271             j_coord_offsetD  = DIM*jnrD;
1272
1273             /* load j atom coordinates */
1274             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1275                                               x+j_coord_offsetC,x+j_coord_offsetD,
1276                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1277
1278             /* Calculate displacement vector */
1279             dx00             = _mm_sub_ps(ix0,jx0);
1280             dy00             = _mm_sub_ps(iy0,jy0);
1281             dz00             = _mm_sub_ps(iz0,jz0);
1282             dx01             = _mm_sub_ps(ix0,jx1);
1283             dy01             = _mm_sub_ps(iy0,jy1);
1284             dz01             = _mm_sub_ps(iz0,jz1);
1285             dx02             = _mm_sub_ps(ix0,jx2);
1286             dy02             = _mm_sub_ps(iy0,jy2);
1287             dz02             = _mm_sub_ps(iz0,jz2);
1288             dx10             = _mm_sub_ps(ix1,jx0);
1289             dy10             = _mm_sub_ps(iy1,jy0);
1290             dz10             = _mm_sub_ps(iz1,jz0);
1291             dx11             = _mm_sub_ps(ix1,jx1);
1292             dy11             = _mm_sub_ps(iy1,jy1);
1293             dz11             = _mm_sub_ps(iz1,jz1);
1294             dx12             = _mm_sub_ps(ix1,jx2);
1295             dy12             = _mm_sub_ps(iy1,jy2);
1296             dz12             = _mm_sub_ps(iz1,jz2);
1297             dx20             = _mm_sub_ps(ix2,jx0);
1298             dy20             = _mm_sub_ps(iy2,jy0);
1299             dz20             = _mm_sub_ps(iz2,jz0);
1300             dx21             = _mm_sub_ps(ix2,jx1);
1301             dy21             = _mm_sub_ps(iy2,jy1);
1302             dz21             = _mm_sub_ps(iz2,jz1);
1303             dx22             = _mm_sub_ps(ix2,jx2);
1304             dy22             = _mm_sub_ps(iy2,jy2);
1305             dz22             = _mm_sub_ps(iz2,jz2);
1306
1307             /* Calculate squared distance and things based on it */
1308             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1309             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1310             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1311             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1312             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1313             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1314             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1315             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1316             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1317
1318             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1319             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1320             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1321             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1322             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1323             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1324             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1325             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1326             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1327
1328             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1329             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1330             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1331             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1332             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1333             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1334             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1335             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1336             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1337
1338             fjx0             = _mm_setzero_ps();
1339             fjy0             = _mm_setzero_ps();
1340             fjz0             = _mm_setzero_ps();
1341             fjx1             = _mm_setzero_ps();
1342             fjy1             = _mm_setzero_ps();
1343             fjz1             = _mm_setzero_ps();
1344             fjx2             = _mm_setzero_ps();
1345             fjy2             = _mm_setzero_ps();
1346             fjz2             = _mm_setzero_ps();
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* COULOMB ELECTROSTATICS */
1353             velec            = _mm_mul_ps(qq00,rinv00);
1354             felec            = _mm_mul_ps(velec,rinvsq00);
1355
1356             fscal            = felec;
1357
1358             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1359
1360              /* Update vectorial force */
1361             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1362             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1363             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1364
1365             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1366             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1367             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1368
1369             /**************************
1370              * CALCULATE INTERACTIONS *
1371              **************************/
1372
1373             /* COULOMB ELECTROSTATICS */
1374             velec            = _mm_mul_ps(qq01,rinv01);
1375             felec            = _mm_mul_ps(velec,rinvsq01);
1376
1377             fscal            = felec;
1378
1379             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1380
1381              /* Update vectorial force */
1382             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1383             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1384             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1385
1386             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1387             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1388             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1389
1390             /**************************
1391              * CALCULATE INTERACTIONS *
1392              **************************/
1393
1394             /* COULOMB ELECTROSTATICS */
1395             velec            = _mm_mul_ps(qq02,rinv02);
1396             felec            = _mm_mul_ps(velec,rinvsq02);
1397
1398             fscal            = felec;
1399
1400             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1401
1402              /* Update vectorial force */
1403             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1404             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1405             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1406
1407             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1408             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1409             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1410
1411             /**************************
1412              * CALCULATE INTERACTIONS *
1413              **************************/
1414
1415             /* COULOMB ELECTROSTATICS */
1416             velec            = _mm_mul_ps(qq10,rinv10);
1417             felec            = _mm_mul_ps(velec,rinvsq10);
1418
1419             fscal            = felec;
1420
1421             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1422
1423              /* Update vectorial force */
1424             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1425             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1426             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1427
1428             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1429             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1430             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1431
1432             /**************************
1433              * CALCULATE INTERACTIONS *
1434              **************************/
1435
1436             /* COULOMB ELECTROSTATICS */
1437             velec            = _mm_mul_ps(qq11,rinv11);
1438             felec            = _mm_mul_ps(velec,rinvsq11);
1439
1440             fscal            = felec;
1441
1442             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1443
1444              /* Update vectorial force */
1445             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1446             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1447             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1448
1449             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1450             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1451             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1452
1453             /**************************
1454              * CALCULATE INTERACTIONS *
1455              **************************/
1456
1457             /* COULOMB ELECTROSTATICS */
1458             velec            = _mm_mul_ps(qq12,rinv12);
1459             felec            = _mm_mul_ps(velec,rinvsq12);
1460
1461             fscal            = felec;
1462
1463             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1464
1465              /* Update vectorial force */
1466             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1467             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1468             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1469
1470             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1471             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1472             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1473
1474             /**************************
1475              * CALCULATE INTERACTIONS *
1476              **************************/
1477
1478             /* COULOMB ELECTROSTATICS */
1479             velec            = _mm_mul_ps(qq20,rinv20);
1480             felec            = _mm_mul_ps(velec,rinvsq20);
1481
1482             fscal            = felec;
1483
1484             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1485
1486              /* Update vectorial force */
1487             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1488             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1489             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1490
1491             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1492             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1493             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1494
1495             /**************************
1496              * CALCULATE INTERACTIONS *
1497              **************************/
1498
1499             /* COULOMB ELECTROSTATICS */
1500             velec            = _mm_mul_ps(qq21,rinv21);
1501             felec            = _mm_mul_ps(velec,rinvsq21);
1502
1503             fscal            = felec;
1504
1505             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1506
1507              /* Update vectorial force */
1508             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1509             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1510             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1511
1512             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1513             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1514             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1515
1516             /**************************
1517              * CALCULATE INTERACTIONS *
1518              **************************/
1519
1520             /* COULOMB ELECTROSTATICS */
1521             velec            = _mm_mul_ps(qq22,rinv22);
1522             felec            = _mm_mul_ps(velec,rinvsq22);
1523
1524             fscal            = felec;
1525
1526             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1527
1528              /* Update vectorial force */
1529             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1530             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1531             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1532
1533             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1534             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1535             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1536
1537             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1538             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1539             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1540             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1541
1542             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1543                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1544
1545             /* Inner loop uses 270 flops */
1546         }
1547
1548         /* End of innermost loop */
1549
1550         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1551                                               f+i_coord_offset,fshift+i_shift_offset);
1552
1553         /* Increment number of inner iterations */
1554         inneriter                  += j_index_end - j_index_start;
1555
1556         /* Outer loop uses 18 flops */
1557     }
1558
1559     /* Increment number of outer iterations */
1560     outeriter        += nri;
1561
1562     /* Update outer/inner flops */
1563
1564     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
1565 }