src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_VdwNone_GeomW4W4_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_avx_128_fma_single
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            None
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset1;
  84     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwioffset3;
  88     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  89     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  90     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  92     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  94     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  95     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  98     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 101     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 102     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 103     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     __m128           dummy_mask,cutoff_mask;
 107     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 108     __m128           one     = _mm_set1_ps(1.0);
 109     __m128           two     = _mm_set1_ps(2.0);
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = _mm_set1_ps(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 127     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 128     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 129
 130     jq1              = _mm_set1_ps(charge[inr+1]);
 131     jq2              = _mm_set1_ps(charge[inr+2]);
 132     jq3              = _mm_set1_ps(charge[inr+3]);
 133     qq11             = _mm_mul_ps(iq1,jq1);
 134     qq12             = _mm_mul_ps(iq1,jq2);
 135     qq13             = _mm_mul_ps(iq1,jq3);
 136     qq21             = _mm_mul_ps(iq2,jq1);
 137     qq22             = _mm_mul_ps(iq2,jq2);
 138     qq23             = _mm_mul_ps(iq2,jq3);
 139     qq31             = _mm_mul_ps(iq3,jq1);
 140     qq32             = _mm_mul_ps(iq3,jq2);
 141     qq33             = _mm_mul_ps(iq3,jq3);
 142
 143     /* Avoid stupid compiler warnings */
 144     jnrA = jnrB = jnrC = jnrD = 0;
 145     j_coord_offsetA = 0;
 146     j_coord_offsetB = 0;
 147     j_coord_offsetC = 0;
 148     j_coord_offsetD = 0;
 149
 150     outeriter        = 0;
 151     inneriter        = 0;
 152
 153     for(iidx=0;iidx<4*DIM;iidx++)
 154     {
 155         scratch[iidx] = 0.0;
 156     }
 157
 158     /* Start outer loop over neighborlists */
 159     for(iidx=0; iidx<nri; iidx++)
 160     {
 161         /* Load shift vector for this list */
 162         i_shift_offset   = DIM*shiftidx[iidx];
 163
 164         /* Load limits for loop over neighbors */
 165         j_index_start    = jindex[iidx];
 166         j_index_end      = jindex[iidx+1];
 167
 168         /* Get outer coordinate index */
 169         inr              = iinr[iidx];
 170         i_coord_offset   = DIM*inr;
 171
 172         /* Load i particle coords and add shift vector */
 173         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 174                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 175
 176         fix1             = _mm_setzero_ps();
 177         fiy1             = _mm_setzero_ps();
 178         fiz1             = _mm_setzero_ps();
 179         fix2             = _mm_setzero_ps();
 180         fiy2             = _mm_setzero_ps();
 181         fiz2             = _mm_setzero_ps();
 182         fix3             = _mm_setzero_ps();
 183         fiy3             = _mm_setzero_ps();
 184         fiz3             = _mm_setzero_ps();
 185
 186         /* Reset potential sums */
 187         velecsum         = _mm_setzero_ps();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             jnrC             = jjnr[jidx+2];
 197             jnrD             = jjnr[jidx+3];
 198             j_coord_offsetA  = DIM*jnrA;
 199             j_coord_offsetB  = DIM*jnrB;
 200             j_coord_offsetC  = DIM*jnrC;
 201             j_coord_offsetD  = DIM*jnrD;
 202
 203             /* load j atom coordinates */
 204             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 205                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 206                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 207
 208             /* Calculate displacement vector */
 209             dx11             = _mm_sub_ps(ix1,jx1);
 210             dy11             = _mm_sub_ps(iy1,jy1);
 211             dz11             = _mm_sub_ps(iz1,jz1);
 212             dx12             = _mm_sub_ps(ix1,jx2);
 213             dy12             = _mm_sub_ps(iy1,jy2);
 214             dz12             = _mm_sub_ps(iz1,jz2);
 215             dx13             = _mm_sub_ps(ix1,jx3);
 216             dy13             = _mm_sub_ps(iy1,jy3);
 217             dz13             = _mm_sub_ps(iz1,jz3);
 218             dx21             = _mm_sub_ps(ix2,jx1);
 219             dy21             = _mm_sub_ps(iy2,jy1);
 220             dz21             = _mm_sub_ps(iz2,jz1);
 221             dx22             = _mm_sub_ps(ix2,jx2);
 222             dy22             = _mm_sub_ps(iy2,jy2);
 223             dz22             = _mm_sub_ps(iz2,jz2);
 224             dx23             = _mm_sub_ps(ix2,jx3);
 225             dy23             = _mm_sub_ps(iy2,jy3);
 226             dz23             = _mm_sub_ps(iz2,jz3);
 227             dx31             = _mm_sub_ps(ix3,jx1);
 228             dy31             = _mm_sub_ps(iy3,jy1);
 229             dz31             = _mm_sub_ps(iz3,jz1);
 230             dx32             = _mm_sub_ps(ix3,jx2);
 231             dy32             = _mm_sub_ps(iy3,jy2);
 232             dz32             = _mm_sub_ps(iz3,jz2);
 233             dx33             = _mm_sub_ps(ix3,jx3);
 234             dy33             = _mm_sub_ps(iy3,jy3);
 235             dz33             = _mm_sub_ps(iz3,jz3);
 236
 237             /* Calculate squared distance and things based on it */
 238             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 239             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 240             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 241             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 242             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 243             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 244             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 245             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 246             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 247
 248             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 249             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 250             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 251             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 252             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 253             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 254             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 255             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 256             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 257
 258             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 259             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 260             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 261             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 262             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 263             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 264             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 265             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 266             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 267
 268             fjx1             = _mm_setzero_ps();
 269             fjy1             = _mm_setzero_ps();
 270             fjz1             = _mm_setzero_ps();
 271             fjx2             = _mm_setzero_ps();
 272             fjy2             = _mm_setzero_ps();
 273             fjz2             = _mm_setzero_ps();
 274             fjx3             = _mm_setzero_ps();
 275             fjy3             = _mm_setzero_ps();
 276             fjz3             = _mm_setzero_ps();
 277
 278             /**************************
 279              * CALCULATE INTERACTIONS *
 280              **************************/
 281
 282             /* COULOMB ELECTROSTATICS */
 283             velec            = _mm_mul_ps(qq11,rinv11);
 284             felec            = _mm_mul_ps(velec,rinvsq11);
 285
 286             /* Update potential sum for this i atom from the interaction with this j atom. */
 287             velecsum         = _mm_add_ps(velecsum,velec);
 288
 289             fscal            = felec;
 290
 291              /* Update vectorial force */
 292             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 293             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 294             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 295
 296             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 297             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 298             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 299
 300             /**************************
 301              * CALCULATE INTERACTIONS *
 302              **************************/
 303
 304             /* COULOMB ELECTROSTATICS */
 305             velec            = _mm_mul_ps(qq12,rinv12);
 306             felec            = _mm_mul_ps(velec,rinvsq12);
 307
 308             /* Update potential sum for this i atom from the interaction with this j atom. */
 309             velecsum         = _mm_add_ps(velecsum,velec);
 310
 311             fscal            = felec;
 312
 313              /* Update vectorial force */
 314             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 315             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 316             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 317
 318             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 319             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 320             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             /* COULOMB ELECTROSTATICS */
 327             velec            = _mm_mul_ps(qq13,rinv13);
 328             felec            = _mm_mul_ps(velec,rinvsq13);
 329
 330             /* Update potential sum for this i atom from the interaction with this j atom. */
 331             velecsum         = _mm_add_ps(velecsum,velec);
 332
 333             fscal            = felec;
 334
 335              /* Update vectorial force */
 336             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 337             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 338             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 339
 340             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 341             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 342             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 343
 344             /**************************
 345              * CALCULATE INTERACTIONS *
 346              **************************/
 347
 348             /* COULOMB ELECTROSTATICS */
 349             velec            = _mm_mul_ps(qq21,rinv21);
 350             felec            = _mm_mul_ps(velec,rinvsq21);
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             velecsum         = _mm_add_ps(velecsum,velec);
 354
 355             fscal            = felec;
 356
 357              /* Update vectorial force */
 358             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 359             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 360             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 361
 362             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 363             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 364             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             /* COULOMB ELECTROSTATICS */
 371             velec            = _mm_mul_ps(qq22,rinv22);
 372             felec            = _mm_mul_ps(velec,rinvsq22);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _mm_add_ps(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379              /* Update vectorial force */
 380             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 381             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 382             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 383
 384             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 385             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 386             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             /* COULOMB ELECTROSTATICS */
 393             velec            = _mm_mul_ps(qq23,rinv23);
 394             felec            = _mm_mul_ps(velec,rinvsq23);
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velecsum         = _mm_add_ps(velecsum,velec);
 398
 399             fscal            = felec;
 400
 401              /* Update vectorial force */
 402             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 403             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 404             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 405
 406             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 407             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 408             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             /* COULOMB ELECTROSTATICS */
 415             velec            = _mm_mul_ps(qq31,rinv31);
 416             felec            = _mm_mul_ps(velec,rinvsq31);
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velecsum         = _mm_add_ps(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423              /* Update vectorial force */
 424             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 425             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 426             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 427
 428             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 429             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 430             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             /* COULOMB ELECTROSTATICS */
 437             velec            = _mm_mul_ps(qq32,rinv32);
 438             felec            = _mm_mul_ps(velec,rinvsq32);
 439
 440             /* Update potential sum for this i atom from the interaction with this j atom. */
 441             velecsum         = _mm_add_ps(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445              /* Update vectorial force */
 446             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 447             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 448             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 449
 450             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 451             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 452             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 453
 454             /**************************
 455              * CALCULATE INTERACTIONS *
 456              **************************/
 457
 458             /* COULOMB ELECTROSTATICS */
 459             velec            = _mm_mul_ps(qq33,rinv33);
 460             felec            = _mm_mul_ps(velec,rinvsq33);
 461
 462             /* Update potential sum for this i atom from the interaction with this j atom. */
 463             velecsum         = _mm_add_ps(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467              /* Update vectorial force */
 468             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 469             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 470             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 471
 472             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 473             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 474             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 475
 476             fjptrA             = f+j_coord_offsetA;
 477             fjptrB             = f+j_coord_offsetB;
 478             fjptrC             = f+j_coord_offsetC;
 479             fjptrD             = f+j_coord_offsetD;
 480
 481             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 482                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 483
 484             /* Inner loop uses 279 flops */
 485         }
 486
 487         if(jidx<j_index_end)
 488         {
 489
 490             /* Get j neighbor index, and coordinate index */
 491             jnrlistA         = jjnr[jidx];
 492             jnrlistB         = jjnr[jidx+1];
 493             jnrlistC         = jjnr[jidx+2];
 494             jnrlistD         = jjnr[jidx+3];
 495             /* Sign of each element will be negative for non-real atoms.
 496              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 497              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 498              */
 499             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 500             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 501             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 502             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 503             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 504             j_coord_offsetA  = DIM*jnrA;
 505             j_coord_offsetB  = DIM*jnrB;
 506             j_coord_offsetC  = DIM*jnrC;
 507             j_coord_offsetD  = DIM*jnrD;
 508
 509             /* load j atom coordinates */
 510             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 511                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 512                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 513
 514             /* Calculate displacement vector */
 515             dx11             = _mm_sub_ps(ix1,jx1);
 516             dy11             = _mm_sub_ps(iy1,jy1);
 517             dz11             = _mm_sub_ps(iz1,jz1);
 518             dx12             = _mm_sub_ps(ix1,jx2);
 519             dy12             = _mm_sub_ps(iy1,jy2);
 520             dz12             = _mm_sub_ps(iz1,jz2);
 521             dx13             = _mm_sub_ps(ix1,jx3);
 522             dy13             = _mm_sub_ps(iy1,jy3);
 523             dz13             = _mm_sub_ps(iz1,jz3);
 524             dx21             = _mm_sub_ps(ix2,jx1);
 525             dy21             = _mm_sub_ps(iy2,jy1);
 526             dz21             = _mm_sub_ps(iz2,jz1);
 527             dx22             = _mm_sub_ps(ix2,jx2);
 528             dy22             = _mm_sub_ps(iy2,jy2);
 529             dz22             = _mm_sub_ps(iz2,jz2);
 530             dx23             = _mm_sub_ps(ix2,jx3);
 531             dy23             = _mm_sub_ps(iy2,jy3);
 532             dz23             = _mm_sub_ps(iz2,jz3);
 533             dx31             = _mm_sub_ps(ix3,jx1);
 534             dy31             = _mm_sub_ps(iy3,jy1);
 535             dz31             = _mm_sub_ps(iz3,jz1);
 536             dx32             = _mm_sub_ps(ix3,jx2);
 537             dy32             = _mm_sub_ps(iy3,jy2);
 538             dz32             = _mm_sub_ps(iz3,jz2);
 539             dx33             = _mm_sub_ps(ix3,jx3);
 540             dy33             = _mm_sub_ps(iy3,jy3);
 541             dz33             = _mm_sub_ps(iz3,jz3);
 542
 543             /* Calculate squared distance and things based on it */
 544             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 545             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 546             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 547             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 548             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 549             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 550             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 551             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 552             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 553
 554             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 555             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 556             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 557             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 558             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 559             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 560             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 561             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 562             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 563
 564             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 565             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 566             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 567             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 568             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 569             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 570             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 571             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 572             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 573
 574             fjx1             = _mm_setzero_ps();
 575             fjy1             = _mm_setzero_ps();
 576             fjz1             = _mm_setzero_ps();
 577             fjx2             = _mm_setzero_ps();
 578             fjy2             = _mm_setzero_ps();
 579             fjz2             = _mm_setzero_ps();
 580             fjx3             = _mm_setzero_ps();
 581             fjy3             = _mm_setzero_ps();
 582             fjz3             = _mm_setzero_ps();
 583
 584             /**************************
 585              * CALCULATE INTERACTIONS *
 586              **************************/
 587
 588             /* COULOMB ELECTROSTATICS */
 589             velec            = _mm_mul_ps(qq11,rinv11);
 590             felec            = _mm_mul_ps(velec,rinvsq11);
 591
 592             /* Update potential sum for this i atom from the interaction with this j atom. */
 593             velec            = _mm_andnot_ps(dummy_mask,velec);
 594             velecsum         = _mm_add_ps(velecsum,velec);
 595
 596             fscal            = felec;
 597
 598             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 599
 600              /* Update vectorial force */
 601             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 602             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 603             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 604
 605             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 606             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 607             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 608
 609             /**************************
 610              * CALCULATE INTERACTIONS *
 611              **************************/
 612
 613             /* COULOMB ELECTROSTATICS */
 614             velec            = _mm_mul_ps(qq12,rinv12);
 615             felec            = _mm_mul_ps(velec,rinvsq12);
 616
 617             /* Update potential sum for this i atom from the interaction with this j atom. */
 618             velec            = _mm_andnot_ps(dummy_mask,velec);
 619             velecsum         = _mm_add_ps(velecsum,velec);
 620
 621             fscal            = felec;
 622
 623             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 624
 625              /* Update vectorial force */
 626             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 627             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 628             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 629
 630             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 631             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 632             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 633
 634             /**************************
 635              * CALCULATE INTERACTIONS *
 636              **************************/
 637
 638             /* COULOMB ELECTROSTATICS */
 639             velec            = _mm_mul_ps(qq13,rinv13);
 640             felec            = _mm_mul_ps(velec,rinvsq13);
 641
 642             /* Update potential sum for this i atom from the interaction with this j atom. */
 643             velec            = _mm_andnot_ps(dummy_mask,velec);
 644             velecsum         = _mm_add_ps(velecsum,velec);
 645
 646             fscal            = felec;
 647
 648             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 649
 650              /* Update vectorial force */
 651             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 652             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 653             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 654
 655             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 656             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 657             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 658
 659             /**************************
 660              * CALCULATE INTERACTIONS *
 661              **************************/
 662
 663             /* COULOMB ELECTROSTATICS */
 664             velec            = _mm_mul_ps(qq21,rinv21);
 665             felec            = _mm_mul_ps(velec,rinvsq21);
 666
 667             /* Update potential sum for this i atom from the interaction with this j atom. */
 668             velec            = _mm_andnot_ps(dummy_mask,velec);
 669             velecsum         = _mm_add_ps(velecsum,velec);
 670
 671             fscal            = felec;
 672
 673             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 674
 675              /* Update vectorial force */
 676             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 677             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 678             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 679
 680             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 681             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 682             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 683
 684             /**************************
 685              * CALCULATE INTERACTIONS *
 686              **************************/
 687
 688             /* COULOMB ELECTROSTATICS */
 689             velec            = _mm_mul_ps(qq22,rinv22);
 690             felec            = _mm_mul_ps(velec,rinvsq22);
 691
 692             /* Update potential sum for this i atom from the interaction with this j atom. */
 693             velec            = _mm_andnot_ps(dummy_mask,velec);
 694             velecsum         = _mm_add_ps(velecsum,velec);
 695
 696             fscal            = felec;
 697
 698             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 699
 700              /* Update vectorial force */
 701             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 702             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 703             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 704
 705             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 706             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 707             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 708
 709             /**************************
 710              * CALCULATE INTERACTIONS *
 711              **************************/
 712
 713             /* COULOMB ELECTROSTATICS */
 714             velec            = _mm_mul_ps(qq23,rinv23);
 715             felec            = _mm_mul_ps(velec,rinvsq23);
 716
 717             /* Update potential sum for this i atom from the interaction with this j atom. */
 718             velec            = _mm_andnot_ps(dummy_mask,velec);
 719             velecsum         = _mm_add_ps(velecsum,velec);
 720
 721             fscal            = felec;
 722
 723             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 724
 725              /* Update vectorial force */
 726             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 727             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 728             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 729
 730             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 731             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 732             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 733
 734             /**************************
 735              * CALCULATE INTERACTIONS *
 736              **************************/
 737
 738             /* COULOMB ELECTROSTATICS */
 739             velec            = _mm_mul_ps(qq31,rinv31);
 740             felec            = _mm_mul_ps(velec,rinvsq31);
 741
 742             /* Update potential sum for this i atom from the interaction with this j atom. */
 743             velec            = _mm_andnot_ps(dummy_mask,velec);
 744             velecsum         = _mm_add_ps(velecsum,velec);
 745
 746             fscal            = felec;
 747
 748             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 749
 750              /* Update vectorial force */
 751             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 752             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 753             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 754
 755             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 756             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 757             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 758
 759             /**************************
 760              * CALCULATE INTERACTIONS *
 761              **************************/
 762
 763             /* COULOMB ELECTROSTATICS */
 764             velec            = _mm_mul_ps(qq32,rinv32);
 765             felec            = _mm_mul_ps(velec,rinvsq32);
 766
 767             /* Update potential sum for this i atom from the interaction with this j atom. */
 768             velec            = _mm_andnot_ps(dummy_mask,velec);
 769             velecsum         = _mm_add_ps(velecsum,velec);
 770
 771             fscal            = felec;
 772
 773             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 774
 775              /* Update vectorial force */
 776             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 777             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 778             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 779
 780             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 781             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 782             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             /* COULOMB ELECTROSTATICS */
 789             velec            = _mm_mul_ps(qq33,rinv33);
 790             felec            = _mm_mul_ps(velec,rinvsq33);
 791
 792             /* Update potential sum for this i atom from the interaction with this j atom. */
 793             velec            = _mm_andnot_ps(dummy_mask,velec);
 794             velecsum         = _mm_add_ps(velecsum,velec);
 795
 796             fscal            = felec;
 797
 798             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 799
 800              /* Update vectorial force */
 801             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 802             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 803             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 804
 805             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 806             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 807             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 808
 809             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 810             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 811             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 812             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 813
 814             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 815                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 816
 817             /* Inner loop uses 279 flops */
 818         }
 819
 820         /* End of innermost loop */
 821
 822         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 823                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 824
 825         ggid                        = gid[iidx];
 826         /* Update potential energies */
 827         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 828
 829         /* Increment number of inner iterations */
 830         inneriter                  += j_index_end - j_index_start;
 831
 832         /* Outer loop uses 19 flops */
 833     }
 834
 835     /* Increment number of outer iterations */
 836     outeriter        += nri;
 837
 838     /* Update outer/inner flops */
 839
 840     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*279);
 841 }
 842 /*
 843  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_single
 844  * Electrostatics interaction: Coulomb
 845  * VdW interaction:            None
 846  * Geometry:                   Water4-Water4
 847  * Calculate force/pot:        Force
 848  */
 849 void
 850 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_single
 851                     (t_nblist                    * gmx_restrict       nlist,
 852                      rvec                        * gmx_restrict          xx,
 853                      rvec                        * gmx_restrict          ff,
 854                      t_forcerec                  * gmx_restrict          fr,
 855                      t_mdatoms                   * gmx_restrict     mdatoms,
 856                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 857                      t_nrnb                      * gmx_restrict        nrnb)
 858 {
 859     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 860      * just 0 for non-waters.
 861      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 862      * jnr indices corresponding to data put in the four positions in the SIMD register.
 863      */
 864     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 865     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 866     int              jnrA,jnrB,jnrC,jnrD;
 867     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 868     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 869     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 870     real             rcutoff_scalar;
 871     real             *shiftvec,*fshift,*x,*f;
 872     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 873     real             scratch[4*DIM];
 874     __m128           fscal,rcutoff,rcutoff2,jidxall;
 875     int              vdwioffset1;
 876     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 877     int              vdwioffset2;
 878     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 879     int              vdwioffset3;
 880     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 881     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 882     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 883     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 884     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 885     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 886     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 887     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 888     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 889     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 890     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 891     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 892     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 893     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 894     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 895     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 896     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 897     real             *charge;
 898     __m128           dummy_mask,cutoff_mask;
 899     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 900     __m128           one     = _mm_set1_ps(1.0);
 901     __m128           two     = _mm_set1_ps(2.0);
 902     x                = xx[0];
 903     f                = ff[0];
 904
 905     nri              = nlist->nri;
 906     iinr             = nlist->iinr;
 907     jindex           = nlist->jindex;
 908     jjnr             = nlist->jjnr;
 909     shiftidx         = nlist->shift;
 910     gid              = nlist->gid;
 911     shiftvec         = fr->shift_vec[0];
 912     fshift           = fr->fshift[0];
 913     facel            = _mm_set1_ps(fr->epsfac);
 914     charge           = mdatoms->chargeA;
 915
 916     /* Setup water-specific parameters */
 917     inr              = nlist->iinr[0];
 918     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 919     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 920     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 921
 922     jq1              = _mm_set1_ps(charge[inr+1]);
 923     jq2              = _mm_set1_ps(charge[inr+2]);
 924     jq3              = _mm_set1_ps(charge[inr+3]);
 925     qq11             = _mm_mul_ps(iq1,jq1);
 926     qq12             = _mm_mul_ps(iq1,jq2);
 927     qq13             = _mm_mul_ps(iq1,jq3);
 928     qq21             = _mm_mul_ps(iq2,jq1);
 929     qq22             = _mm_mul_ps(iq2,jq2);
 930     qq23             = _mm_mul_ps(iq2,jq3);
 931     qq31             = _mm_mul_ps(iq3,jq1);
 932     qq32             = _mm_mul_ps(iq3,jq2);
 933     qq33             = _mm_mul_ps(iq3,jq3);
 934
 935     /* Avoid stupid compiler warnings */
 936     jnrA = jnrB = jnrC = jnrD = 0;
 937     j_coord_offsetA = 0;
 938     j_coord_offsetB = 0;
 939     j_coord_offsetC = 0;
 940     j_coord_offsetD = 0;
 941
 942     outeriter        = 0;
 943     inneriter        = 0;
 944
 945     for(iidx=0;iidx<4*DIM;iidx++)
 946     {
 947         scratch[iidx] = 0.0;
 948     }
 949
 950     /* Start outer loop over neighborlists */
 951     for(iidx=0; iidx<nri; iidx++)
 952     {
 953         /* Load shift vector for this list */
 954         i_shift_offset   = DIM*shiftidx[iidx];
 955
 956         /* Load limits for loop over neighbors */
 957         j_index_start    = jindex[iidx];
 958         j_index_end      = jindex[iidx+1];
 959
 960         /* Get outer coordinate index */
 961         inr              = iinr[iidx];
 962         i_coord_offset   = DIM*inr;
 963
 964         /* Load i particle coords and add shift vector */
 965         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 966                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 967
 968         fix1             = _mm_setzero_ps();
 969         fiy1             = _mm_setzero_ps();
 970         fiz1             = _mm_setzero_ps();
 971         fix2             = _mm_setzero_ps();
 972         fiy2             = _mm_setzero_ps();
 973         fiz2             = _mm_setzero_ps();
 974         fix3             = _mm_setzero_ps();
 975         fiy3             = _mm_setzero_ps();
 976         fiz3             = _mm_setzero_ps();
 977
 978         /* Start inner kernel loop */
 979         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 980         {
 981
 982             /* Get j neighbor index, and coordinate index */
 983             jnrA             = jjnr[jidx];
 984             jnrB             = jjnr[jidx+1];
 985             jnrC             = jjnr[jidx+2];
 986             jnrD             = jjnr[jidx+3];
 987             j_coord_offsetA  = DIM*jnrA;
 988             j_coord_offsetB  = DIM*jnrB;
 989             j_coord_offsetC  = DIM*jnrC;
 990             j_coord_offsetD  = DIM*jnrD;
 991
 992             /* load j atom coordinates */
 993             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 994                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 995                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 996
 997             /* Calculate displacement vector */
 998             dx11             = _mm_sub_ps(ix1,jx1);
 999             dy11             = _mm_sub_ps(iy1,jy1);
1000             dz11             = _mm_sub_ps(iz1,jz1);
1001             dx12             = _mm_sub_ps(ix1,jx2);
1002             dy12             = _mm_sub_ps(iy1,jy2);
1003             dz12             = _mm_sub_ps(iz1,jz2);
1004             dx13             = _mm_sub_ps(ix1,jx3);
1005             dy13             = _mm_sub_ps(iy1,jy3);
1006             dz13             = _mm_sub_ps(iz1,jz3);
1007             dx21             = _mm_sub_ps(ix2,jx1);
1008             dy21             = _mm_sub_ps(iy2,jy1);
1009             dz21             = _mm_sub_ps(iz2,jz1);
1010             dx22             = _mm_sub_ps(ix2,jx2);
1011             dy22             = _mm_sub_ps(iy2,jy2);
1012             dz22             = _mm_sub_ps(iz2,jz2);
1013             dx23             = _mm_sub_ps(ix2,jx3);
1014             dy23             = _mm_sub_ps(iy2,jy3);
1015             dz23             = _mm_sub_ps(iz2,jz3);
1016             dx31             = _mm_sub_ps(ix3,jx1);
1017             dy31             = _mm_sub_ps(iy3,jy1);
1018             dz31             = _mm_sub_ps(iz3,jz1);
1019             dx32             = _mm_sub_ps(ix3,jx2);
1020             dy32             = _mm_sub_ps(iy3,jy2);
1021             dz32             = _mm_sub_ps(iz3,jz2);
1022             dx33             = _mm_sub_ps(ix3,jx3);
1023             dy33             = _mm_sub_ps(iy3,jy3);
1024             dz33             = _mm_sub_ps(iz3,jz3);
1025
1026             /* Calculate squared distance and things based on it */
1027             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1028             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1029             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1030             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1031             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1032             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1033             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1034             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1035             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1036
1037             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1038             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1039             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1040             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1041             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1042             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1043             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1044             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1045             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1046
1047             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1048             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1049             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1050             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1051             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1052             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1053             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1054             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1055             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1056
1057             fjx1             = _mm_setzero_ps();
1058             fjy1             = _mm_setzero_ps();
1059             fjz1             = _mm_setzero_ps();
1060             fjx2             = _mm_setzero_ps();
1061             fjy2             = _mm_setzero_ps();
1062             fjz2             = _mm_setzero_ps();
1063             fjx3             = _mm_setzero_ps();
1064             fjy3             = _mm_setzero_ps();
1065             fjz3             = _mm_setzero_ps();
1066
1067             /**************************
1068              * CALCULATE INTERACTIONS *
1069              **************************/
1070
1071             /* COULOMB ELECTROSTATICS */
1072             velec            = _mm_mul_ps(qq11,rinv11);
1073             felec            = _mm_mul_ps(velec,rinvsq11);
1074
1075             fscal            = felec;
1076
1077              /* Update vectorial force */
1078             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1079             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1080             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1081
1082             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1083             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1084             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1085
1086             /**************************
1087              * CALCULATE INTERACTIONS *
1088              **************************/
1089
1090             /* COULOMB ELECTROSTATICS */
1091             velec            = _mm_mul_ps(qq12,rinv12);
1092             felec            = _mm_mul_ps(velec,rinvsq12);
1093
1094             fscal            = felec;
1095
1096              /* Update vectorial force */
1097             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1098             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1099             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1100
1101             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1102             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1103             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1104
1105             /**************************
1106              * CALCULATE INTERACTIONS *
1107              **************************/
1108
1109             /* COULOMB ELECTROSTATICS */
1110             velec            = _mm_mul_ps(qq13,rinv13);
1111             felec            = _mm_mul_ps(velec,rinvsq13);
1112
1113             fscal            = felec;
1114
1115              /* Update vectorial force */
1116             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1117             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1118             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1119
1120             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1121             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1122             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1123
1124             /**************************
1125              * CALCULATE INTERACTIONS *
1126              **************************/
1127
1128             /* COULOMB ELECTROSTATICS */
1129             velec            = _mm_mul_ps(qq21,rinv21);
1130             felec            = _mm_mul_ps(velec,rinvsq21);
1131
1132             fscal            = felec;
1133
1134              /* Update vectorial force */
1135             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1136             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1137             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1138
1139             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1140             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1141             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             /* COULOMB ELECTROSTATICS */
1148             velec            = _mm_mul_ps(qq22,rinv22);
1149             felec            = _mm_mul_ps(velec,rinvsq22);
1150
1151             fscal            = felec;
1152
1153              /* Update vectorial force */
1154             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1155             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1156             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1157
1158             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1159             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1160             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1161
1162             /**************************
1163              * CALCULATE INTERACTIONS *
1164              **************************/
1165
1166             /* COULOMB ELECTROSTATICS */
1167             velec            = _mm_mul_ps(qq23,rinv23);
1168             felec            = _mm_mul_ps(velec,rinvsq23);
1169
1170             fscal            = felec;
1171
1172              /* Update vectorial force */
1173             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1174             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1175             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1176
1177             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1178             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1179             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1180
1181             /**************************
1182              * CALCULATE INTERACTIONS *
1183              **************************/
1184
1185             /* COULOMB ELECTROSTATICS */
1186             velec            = _mm_mul_ps(qq31,rinv31);
1187             felec            = _mm_mul_ps(velec,rinvsq31);
1188
1189             fscal            = felec;
1190
1191              /* Update vectorial force */
1192             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1193             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1194             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1195
1196             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1197             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1198             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1199
1200             /**************************
1201              * CALCULATE INTERACTIONS *
1202              **************************/
1203
1204             /* COULOMB ELECTROSTATICS */
1205             velec            = _mm_mul_ps(qq32,rinv32);
1206             felec            = _mm_mul_ps(velec,rinvsq32);
1207
1208             fscal            = felec;
1209
1210              /* Update vectorial force */
1211             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1212             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1213             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1214
1215             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1216             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1217             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1218
1219             /**************************
1220              * CALCULATE INTERACTIONS *
1221              **************************/
1222
1223             /* COULOMB ELECTROSTATICS */
1224             velec            = _mm_mul_ps(qq33,rinv33);
1225             felec            = _mm_mul_ps(velec,rinvsq33);
1226
1227             fscal            = felec;
1228
1229              /* Update vectorial force */
1230             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1231             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1232             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1233
1234             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1235             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1236             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1237
1238             fjptrA             = f+j_coord_offsetA;
1239             fjptrB             = f+j_coord_offsetB;
1240             fjptrC             = f+j_coord_offsetC;
1241             fjptrD             = f+j_coord_offsetD;
1242
1243             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1244                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1245
1246             /* Inner loop uses 270 flops */
1247         }
1248
1249         if(jidx<j_index_end)
1250         {
1251
1252             /* Get j neighbor index, and coordinate index */
1253             jnrlistA         = jjnr[jidx];
1254             jnrlistB         = jjnr[jidx+1];
1255             jnrlistC         = jjnr[jidx+2];
1256             jnrlistD         = jjnr[jidx+3];
1257             /* Sign of each element will be negative for non-real atoms.
1258              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1259              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1260              */
1261             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1262             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1263             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1264             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1265             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1266             j_coord_offsetA  = DIM*jnrA;
1267             j_coord_offsetB  = DIM*jnrB;
1268             j_coord_offsetC  = DIM*jnrC;
1269             j_coord_offsetD  = DIM*jnrD;
1270
1271             /* load j atom coordinates */
1272             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1273                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1274                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1275
1276             /* Calculate displacement vector */
1277             dx11             = _mm_sub_ps(ix1,jx1);
1278             dy11             = _mm_sub_ps(iy1,jy1);
1279             dz11             = _mm_sub_ps(iz1,jz1);
1280             dx12             = _mm_sub_ps(ix1,jx2);
1281             dy12             = _mm_sub_ps(iy1,jy2);
1282             dz12             = _mm_sub_ps(iz1,jz2);
1283             dx13             = _mm_sub_ps(ix1,jx3);
1284             dy13             = _mm_sub_ps(iy1,jy3);
1285             dz13             = _mm_sub_ps(iz1,jz3);
1286             dx21             = _mm_sub_ps(ix2,jx1);
1287             dy21             = _mm_sub_ps(iy2,jy1);
1288             dz21             = _mm_sub_ps(iz2,jz1);
1289             dx22             = _mm_sub_ps(ix2,jx2);
1290             dy22             = _mm_sub_ps(iy2,jy2);
1291             dz22             = _mm_sub_ps(iz2,jz2);
1292             dx23             = _mm_sub_ps(ix2,jx3);
1293             dy23             = _mm_sub_ps(iy2,jy3);
1294             dz23             = _mm_sub_ps(iz2,jz3);
1295             dx31             = _mm_sub_ps(ix3,jx1);
1296             dy31             = _mm_sub_ps(iy3,jy1);
1297             dz31             = _mm_sub_ps(iz3,jz1);
1298             dx32             = _mm_sub_ps(ix3,jx2);
1299             dy32             = _mm_sub_ps(iy3,jy2);
1300             dz32             = _mm_sub_ps(iz3,jz2);
1301             dx33             = _mm_sub_ps(ix3,jx3);
1302             dy33             = _mm_sub_ps(iy3,jy3);
1303             dz33             = _mm_sub_ps(iz3,jz3);
1304
1305             /* Calculate squared distance and things based on it */
1306             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1307             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1308             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1309             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1310             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1311             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1312             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1313             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1314             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1315
1316             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1317             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1318             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1319             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1320             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1321             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1322             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1323             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1324             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1325
1326             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1327             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1328             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1329             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1330             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1331             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1332             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1333             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1334             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1335
1336             fjx1             = _mm_setzero_ps();
1337             fjy1             = _mm_setzero_ps();
1338             fjz1             = _mm_setzero_ps();
1339             fjx2             = _mm_setzero_ps();
1340             fjy2             = _mm_setzero_ps();
1341             fjz2             = _mm_setzero_ps();
1342             fjx3             = _mm_setzero_ps();
1343             fjy3             = _mm_setzero_ps();
1344             fjz3             = _mm_setzero_ps();
1345
1346             /**************************
1347              * CALCULATE INTERACTIONS *
1348              **************************/
1349
1350             /* COULOMB ELECTROSTATICS */
1351             velec            = _mm_mul_ps(qq11,rinv11);
1352             felec            = _mm_mul_ps(velec,rinvsq11);
1353
1354             fscal            = felec;
1355
1356             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1357
1358              /* Update vectorial force */
1359             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1360             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1361             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1362
1363             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1364             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1365             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1366
1367             /**************************
1368              * CALCULATE INTERACTIONS *
1369              **************************/
1370
1371             /* COULOMB ELECTROSTATICS */
1372             velec            = _mm_mul_ps(qq12,rinv12);
1373             felec            = _mm_mul_ps(velec,rinvsq12);
1374
1375             fscal            = felec;
1376
1377             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1378
1379              /* Update vectorial force */
1380             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1381             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1382             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1383
1384             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1385             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1386             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1387
1388             /**************************
1389              * CALCULATE INTERACTIONS *
1390              **************************/
1391
1392             /* COULOMB ELECTROSTATICS */
1393             velec            = _mm_mul_ps(qq13,rinv13);
1394             felec            = _mm_mul_ps(velec,rinvsq13);
1395
1396             fscal            = felec;
1397
1398             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1399
1400              /* Update vectorial force */
1401             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1402             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1403             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1404
1405             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1406             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1407             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1408
1409             /**************************
1410              * CALCULATE INTERACTIONS *
1411              **************************/
1412
1413             /* COULOMB ELECTROSTATICS */
1414             velec            = _mm_mul_ps(qq21,rinv21);
1415             felec            = _mm_mul_ps(velec,rinvsq21);
1416
1417             fscal            = felec;
1418
1419             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1420
1421              /* Update vectorial force */
1422             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1423             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1424             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1425
1426             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1427             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1428             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1429
1430             /**************************
1431              * CALCULATE INTERACTIONS *
1432              **************************/
1433
1434             /* COULOMB ELECTROSTATICS */
1435             velec            = _mm_mul_ps(qq22,rinv22);
1436             felec            = _mm_mul_ps(velec,rinvsq22);
1437
1438             fscal            = felec;
1439
1440             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1441
1442              /* Update vectorial force */
1443             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1444             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1445             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1446
1447             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1448             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1449             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1450
1451             /**************************
1452              * CALCULATE INTERACTIONS *
1453              **************************/
1454
1455             /* COULOMB ELECTROSTATICS */
1456             velec            = _mm_mul_ps(qq23,rinv23);
1457             felec            = _mm_mul_ps(velec,rinvsq23);
1458
1459             fscal            = felec;
1460
1461             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1462
1463              /* Update vectorial force */
1464             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1465             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1466             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1467
1468             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1469             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1470             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1471
1472             /**************************
1473              * CALCULATE INTERACTIONS *
1474              **************************/
1475
1476             /* COULOMB ELECTROSTATICS */
1477             velec            = _mm_mul_ps(qq31,rinv31);
1478             felec            = _mm_mul_ps(velec,rinvsq31);
1479
1480             fscal            = felec;
1481
1482             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1483
1484              /* Update vectorial force */
1485             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1486             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1487             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1488
1489             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1490             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1491             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1492
1493             /**************************
1494              * CALCULATE INTERACTIONS *
1495              **************************/
1496
1497             /* COULOMB ELECTROSTATICS */
1498             velec            = _mm_mul_ps(qq32,rinv32);
1499             felec            = _mm_mul_ps(velec,rinvsq32);
1500
1501             fscal            = felec;
1502
1503             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1504
1505              /* Update vectorial force */
1506             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1507             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1508             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1509
1510             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1511             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1512             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1513
1514             /**************************
1515              * CALCULATE INTERACTIONS *
1516              **************************/
1517
1518             /* COULOMB ELECTROSTATICS */
1519             velec            = _mm_mul_ps(qq33,rinv33);
1520             felec            = _mm_mul_ps(velec,rinvsq33);
1521
1522             fscal            = felec;
1523
1524             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1525
1526              /* Update vectorial force */
1527             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1528             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1529             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1530
1531             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1532             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1533             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1534
1535             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1536             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1537             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1538             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1539
1540             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1541                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1542
1543             /* Inner loop uses 270 flops */
1544         }
1545
1546         /* End of innermost loop */
1547
1548         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1549                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1550
1551         /* Increment number of inner iterations */
1552         inneriter                  += j_index_end - j_index_start;
1553
1554         /* Outer loop uses 18 flops */
1555     }
1556
1557     /* Increment number of outer iterations */
1558     outeriter        += nri;
1559
1560     /* Update outer/inner flops */
1561
1562     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1563 }