src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_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, 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_ElecRFCut_VdwNone_GeomW4W4_VF_avx_128_fma_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            None
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_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              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  96     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  97     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 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     krf              = _mm_set1_ps(fr->ic->k_rf);
 126     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 127     crf              = _mm_set1_ps(fr->ic->c_rf);
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 132     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 133     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 134
 135     jq1              = _mm_set1_ps(charge[inr+1]);
 136     jq2              = _mm_set1_ps(charge[inr+2]);
 137     jq3              = _mm_set1_ps(charge[inr+3]);
 138     qq11             = _mm_mul_ps(iq1,jq1);
 139     qq12             = _mm_mul_ps(iq1,jq2);
 140     qq13             = _mm_mul_ps(iq1,jq3);
 141     qq21             = _mm_mul_ps(iq2,jq1);
 142     qq22             = _mm_mul_ps(iq2,jq2);
 143     qq23             = _mm_mul_ps(iq2,jq3);
 144     qq31             = _mm_mul_ps(iq3,jq1);
 145     qq32             = _mm_mul_ps(iq3,jq2);
 146     qq33             = _mm_mul_ps(iq3,jq3);
 147
 148     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 149     rcutoff_scalar   = fr->rcoulomb;
 150     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 151     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 152
 153     /* Avoid stupid compiler warnings */
 154     jnrA = jnrB = jnrC = jnrD = 0;
 155     j_coord_offsetA = 0;
 156     j_coord_offsetB = 0;
 157     j_coord_offsetC = 0;
 158     j_coord_offsetD = 0;
 159
 160     outeriter        = 0;
 161     inneriter        = 0;
 162
 163     for(iidx=0;iidx<4*DIM;iidx++)
 164     {
 165         scratch[iidx] = 0.0;
 166     }
 167
 168     /* Start outer loop over neighborlists */
 169     for(iidx=0; iidx<nri; iidx++)
 170     {
 171         /* Load shift vector for this list */
 172         i_shift_offset   = DIM*shiftidx[iidx];
 173
 174         /* Load limits for loop over neighbors */
 175         j_index_start    = jindex[iidx];
 176         j_index_end      = jindex[iidx+1];
 177
 178         /* Get outer coordinate index */
 179         inr              = iinr[iidx];
 180         i_coord_offset   = DIM*inr;
 181
 182         /* Load i particle coords and add shift vector */
 183         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 184                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 185
 186         fix1             = _mm_setzero_ps();
 187         fiy1             = _mm_setzero_ps();
 188         fiz1             = _mm_setzero_ps();
 189         fix2             = _mm_setzero_ps();
 190         fiy2             = _mm_setzero_ps();
 191         fiz2             = _mm_setzero_ps();
 192         fix3             = _mm_setzero_ps();
 193         fiy3             = _mm_setzero_ps();
 194         fiz3             = _mm_setzero_ps();
 195
 196         /* Reset potential sums */
 197         velecsum         = _mm_setzero_ps();
 198
 199         /* Start inner kernel loop */
 200         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 201         {
 202
 203             /* Get j neighbor index, and coordinate index */
 204             jnrA             = jjnr[jidx];
 205             jnrB             = jjnr[jidx+1];
 206             jnrC             = jjnr[jidx+2];
 207             jnrD             = jjnr[jidx+3];
 208             j_coord_offsetA  = DIM*jnrA;
 209             j_coord_offsetB  = DIM*jnrB;
 210             j_coord_offsetC  = DIM*jnrC;
 211             j_coord_offsetD  = DIM*jnrD;
 212
 213             /* load j atom coordinates */
 214             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 215                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 216                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 217
 218             /* Calculate displacement vector */
 219             dx11             = _mm_sub_ps(ix1,jx1);
 220             dy11             = _mm_sub_ps(iy1,jy1);
 221             dz11             = _mm_sub_ps(iz1,jz1);
 222             dx12             = _mm_sub_ps(ix1,jx2);
 223             dy12             = _mm_sub_ps(iy1,jy2);
 224             dz12             = _mm_sub_ps(iz1,jz2);
 225             dx13             = _mm_sub_ps(ix1,jx3);
 226             dy13             = _mm_sub_ps(iy1,jy3);
 227             dz13             = _mm_sub_ps(iz1,jz3);
 228             dx21             = _mm_sub_ps(ix2,jx1);
 229             dy21             = _mm_sub_ps(iy2,jy1);
 230             dz21             = _mm_sub_ps(iz2,jz1);
 231             dx22             = _mm_sub_ps(ix2,jx2);
 232             dy22             = _mm_sub_ps(iy2,jy2);
 233             dz22             = _mm_sub_ps(iz2,jz2);
 234             dx23             = _mm_sub_ps(ix2,jx3);
 235             dy23             = _mm_sub_ps(iy2,jy3);
 236             dz23             = _mm_sub_ps(iz2,jz3);
 237             dx31             = _mm_sub_ps(ix3,jx1);
 238             dy31             = _mm_sub_ps(iy3,jy1);
 239             dz31             = _mm_sub_ps(iz3,jz1);
 240             dx32             = _mm_sub_ps(ix3,jx2);
 241             dy32             = _mm_sub_ps(iy3,jy2);
 242             dz32             = _mm_sub_ps(iz3,jz2);
 243             dx33             = _mm_sub_ps(ix3,jx3);
 244             dy33             = _mm_sub_ps(iy3,jy3);
 245             dz33             = _mm_sub_ps(iz3,jz3);
 246
 247             /* Calculate squared distance and things based on it */
 248             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 249             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 250             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 251             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 252             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 253             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 254             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 255             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 256             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 257
 258             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 259             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 260             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 261             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 262             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 263             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 264             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 265             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 266             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 267
 268             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 269             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 270             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 271             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 272             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 273             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 274             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 275             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 276             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 277
 278             fjx1             = _mm_setzero_ps();
 279             fjy1             = _mm_setzero_ps();
 280             fjz1             = _mm_setzero_ps();
 281             fjx2             = _mm_setzero_ps();
 282             fjy2             = _mm_setzero_ps();
 283             fjz2             = _mm_setzero_ps();
 284             fjx3             = _mm_setzero_ps();
 285             fjy3             = _mm_setzero_ps();
 286             fjz3             = _mm_setzero_ps();
 287
 288             /**************************
 289              * CALCULATE INTERACTIONS *
 290              **************************/
 291
 292             if (gmx_mm_any_lt(rsq11,rcutoff2))
 293             {
 294
 295             /* REACTION-FIELD ELECTROSTATICS */
 296             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 297             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 298
 299             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 300
 301             /* Update potential sum for this i atom from the interaction with this j atom. */
 302             velec            = _mm_and_ps(velec,cutoff_mask);
 303             velecsum         = _mm_add_ps(velecsum,velec);
 304
 305             fscal            = felec;
 306
 307             fscal            = _mm_and_ps(fscal,cutoff_mask);
 308
 309              /* Update vectorial force */
 310             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 311             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 312             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 313
 314             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 315             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 316             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 317
 318             }
 319
 320             /**************************
 321              * CALCULATE INTERACTIONS *
 322              **************************/
 323
 324             if (gmx_mm_any_lt(rsq12,rcutoff2))
 325             {
 326
 327             /* REACTION-FIELD ELECTROSTATICS */
 328             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 329             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 330
 331             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             velec            = _mm_and_ps(velec,cutoff_mask);
 335             velecsum         = _mm_add_ps(velecsum,velec);
 336
 337             fscal            = felec;
 338
 339             fscal            = _mm_and_ps(fscal,cutoff_mask);
 340
 341              /* Update vectorial force */
 342             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 343             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 344             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 345
 346             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 347             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 348             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 349
 350             }
 351
 352             /**************************
 353              * CALCULATE INTERACTIONS *
 354              **************************/
 355
 356             if (gmx_mm_any_lt(rsq13,rcutoff2))
 357             {
 358
 359             /* REACTION-FIELD ELECTROSTATICS */
 360             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 361             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 362
 363             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             velec            = _mm_and_ps(velec,cutoff_mask);
 367             velecsum         = _mm_add_ps(velecsum,velec);
 368
 369             fscal            = felec;
 370
 371             fscal            = _mm_and_ps(fscal,cutoff_mask);
 372
 373              /* Update vectorial force */
 374             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 375             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 376             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 377
 378             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 379             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 380             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 381
 382             }
 383
 384             /**************************
 385              * CALCULATE INTERACTIONS *
 386              **************************/
 387
 388             if (gmx_mm_any_lt(rsq21,rcutoff2))
 389             {
 390
 391             /* REACTION-FIELD ELECTROSTATICS */
 392             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 393             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 394
 395             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velec            = _mm_and_ps(velec,cutoff_mask);
 399             velecsum         = _mm_add_ps(velecsum,velec);
 400
 401             fscal            = felec;
 402
 403             fscal            = _mm_and_ps(fscal,cutoff_mask);
 404
 405              /* Update vectorial force */
 406             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 407             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 408             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 409
 410             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 411             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 412             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 413
 414             }
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             if (gmx_mm_any_lt(rsq22,rcutoff2))
 421             {
 422
 423             /* REACTION-FIELD ELECTROSTATICS */
 424             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 425             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 426
 427             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 428
 429             /* Update potential sum for this i atom from the interaction with this j atom. */
 430             velec            = _mm_and_ps(velec,cutoff_mask);
 431             velecsum         = _mm_add_ps(velecsum,velec);
 432
 433             fscal            = felec;
 434
 435             fscal            = _mm_and_ps(fscal,cutoff_mask);
 436
 437              /* Update vectorial force */
 438             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 439             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 440             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 441
 442             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 443             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 444             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 445
 446             }
 447
 448             /**************************
 449              * CALCULATE INTERACTIONS *
 450              **************************/
 451
 452             if (gmx_mm_any_lt(rsq23,rcutoff2))
 453             {
 454
 455             /* REACTION-FIELD ELECTROSTATICS */
 456             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 457             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 458
 459             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             velec            = _mm_and_ps(velec,cutoff_mask);
 463             velecsum         = _mm_add_ps(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467             fscal            = _mm_and_ps(fscal,cutoff_mask);
 468
 469              /* Update vectorial force */
 470             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 471             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 472             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 473
 474             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 475             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 476             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 477
 478             }
 479
 480             /**************************
 481              * CALCULATE INTERACTIONS *
 482              **************************/
 483
 484             if (gmx_mm_any_lt(rsq31,rcutoff2))
 485             {
 486
 487             /* REACTION-FIELD ELECTROSTATICS */
 488             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 489             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 490
 491             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 492
 493             /* Update potential sum for this i atom from the interaction with this j atom. */
 494             velec            = _mm_and_ps(velec,cutoff_mask);
 495             velecsum         = _mm_add_ps(velecsum,velec);
 496
 497             fscal            = felec;
 498
 499             fscal            = _mm_and_ps(fscal,cutoff_mask);
 500
 501              /* Update vectorial force */
 502             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 503             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 504             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 505
 506             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 507             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 508             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 509
 510             }
 511
 512             /**************************
 513              * CALCULATE INTERACTIONS *
 514              **************************/
 515
 516             if (gmx_mm_any_lt(rsq32,rcutoff2))
 517             {
 518
 519             /* REACTION-FIELD ELECTROSTATICS */
 520             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 521             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 522
 523             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 524
 525             /* Update potential sum for this i atom from the interaction with this j atom. */
 526             velec            = _mm_and_ps(velec,cutoff_mask);
 527             velecsum         = _mm_add_ps(velecsum,velec);
 528
 529             fscal            = felec;
 530
 531             fscal            = _mm_and_ps(fscal,cutoff_mask);
 532
 533              /* Update vectorial force */
 534             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 535             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 536             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 537
 538             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 539             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 540             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 541
 542             }
 543
 544             /**************************
 545              * CALCULATE INTERACTIONS *
 546              **************************/
 547
 548             if (gmx_mm_any_lt(rsq33,rcutoff2))
 549             {
 550
 551             /* REACTION-FIELD ELECTROSTATICS */
 552             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 553             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 554
 555             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 556
 557             /* Update potential sum for this i atom from the interaction with this j atom. */
 558             velec            = _mm_and_ps(velec,cutoff_mask);
 559             velecsum         = _mm_add_ps(velecsum,velec);
 560
 561             fscal            = felec;
 562
 563             fscal            = _mm_and_ps(fscal,cutoff_mask);
 564
 565              /* Update vectorial force */
 566             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 567             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 568             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 569
 570             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 571             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 572             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 573
 574             }
 575
 576             fjptrA             = f+j_coord_offsetA;
 577             fjptrB             = f+j_coord_offsetB;
 578             fjptrC             = f+j_coord_offsetC;
 579             fjptrD             = f+j_coord_offsetD;
 580
 581             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 582                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 583
 584             /* Inner loop uses 351 flops */
 585         }
 586
 587         if(jidx<j_index_end)
 588         {
 589
 590             /* Get j neighbor index, and coordinate index */
 591             jnrlistA         = jjnr[jidx];
 592             jnrlistB         = jjnr[jidx+1];
 593             jnrlistC         = jjnr[jidx+2];
 594             jnrlistD         = jjnr[jidx+3];
 595             /* Sign of each element will be negative for non-real atoms.
 596              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 597              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 598              */
 599             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 600             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 601             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 602             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 603             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 604             j_coord_offsetA  = DIM*jnrA;
 605             j_coord_offsetB  = DIM*jnrB;
 606             j_coord_offsetC  = DIM*jnrC;
 607             j_coord_offsetD  = DIM*jnrD;
 608
 609             /* load j atom coordinates */
 610             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 611                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 612                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 613
 614             /* Calculate displacement vector */
 615             dx11             = _mm_sub_ps(ix1,jx1);
 616             dy11             = _mm_sub_ps(iy1,jy1);
 617             dz11             = _mm_sub_ps(iz1,jz1);
 618             dx12             = _mm_sub_ps(ix1,jx2);
 619             dy12             = _mm_sub_ps(iy1,jy2);
 620             dz12             = _mm_sub_ps(iz1,jz2);
 621             dx13             = _mm_sub_ps(ix1,jx3);
 622             dy13             = _mm_sub_ps(iy1,jy3);
 623             dz13             = _mm_sub_ps(iz1,jz3);
 624             dx21             = _mm_sub_ps(ix2,jx1);
 625             dy21             = _mm_sub_ps(iy2,jy1);
 626             dz21             = _mm_sub_ps(iz2,jz1);
 627             dx22             = _mm_sub_ps(ix2,jx2);
 628             dy22             = _mm_sub_ps(iy2,jy2);
 629             dz22             = _mm_sub_ps(iz2,jz2);
 630             dx23             = _mm_sub_ps(ix2,jx3);
 631             dy23             = _mm_sub_ps(iy2,jy3);
 632             dz23             = _mm_sub_ps(iz2,jz3);
 633             dx31             = _mm_sub_ps(ix3,jx1);
 634             dy31             = _mm_sub_ps(iy3,jy1);
 635             dz31             = _mm_sub_ps(iz3,jz1);
 636             dx32             = _mm_sub_ps(ix3,jx2);
 637             dy32             = _mm_sub_ps(iy3,jy2);
 638             dz32             = _mm_sub_ps(iz3,jz2);
 639             dx33             = _mm_sub_ps(ix3,jx3);
 640             dy33             = _mm_sub_ps(iy3,jy3);
 641             dz33             = _mm_sub_ps(iz3,jz3);
 642
 643             /* Calculate squared distance and things based on it */
 644             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 645             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 646             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 647             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 648             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 649             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 650             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 651             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 652             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 653
 654             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 655             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 656             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 657             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 658             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 659             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 660             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 661             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 662             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 663
 664             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 665             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 666             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 667             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 668             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 669             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 670             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 671             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 672             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 673
 674             fjx1             = _mm_setzero_ps();
 675             fjy1             = _mm_setzero_ps();
 676             fjz1             = _mm_setzero_ps();
 677             fjx2             = _mm_setzero_ps();
 678             fjy2             = _mm_setzero_ps();
 679             fjz2             = _mm_setzero_ps();
 680             fjx3             = _mm_setzero_ps();
 681             fjy3             = _mm_setzero_ps();
 682             fjz3             = _mm_setzero_ps();
 683
 684             /**************************
 685              * CALCULATE INTERACTIONS *
 686              **************************/
 687
 688             if (gmx_mm_any_lt(rsq11,rcutoff2))
 689             {
 690
 691             /* REACTION-FIELD ELECTROSTATICS */
 692             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 693             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 694
 695             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 696
 697             /* Update potential sum for this i atom from the interaction with this j atom. */
 698             velec            = _mm_and_ps(velec,cutoff_mask);
 699             velec            = _mm_andnot_ps(dummy_mask,velec);
 700             velecsum         = _mm_add_ps(velecsum,velec);
 701
 702             fscal            = felec;
 703
 704             fscal            = _mm_and_ps(fscal,cutoff_mask);
 705
 706             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 707
 708              /* Update vectorial force */
 709             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 710             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 711             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 712
 713             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 714             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 715             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 716
 717             }
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             if (gmx_mm_any_lt(rsq12,rcutoff2))
 724             {
 725
 726             /* REACTION-FIELD ELECTROSTATICS */
 727             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 728             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 729
 730             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 731
 732             /* Update potential sum for this i atom from the interaction with this j atom. */
 733             velec            = _mm_and_ps(velec,cutoff_mask);
 734             velec            = _mm_andnot_ps(dummy_mask,velec);
 735             velecsum         = _mm_add_ps(velecsum,velec);
 736
 737             fscal            = felec;
 738
 739             fscal            = _mm_and_ps(fscal,cutoff_mask);
 740
 741             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 742
 743              /* Update vectorial force */
 744             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 745             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 746             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 747
 748             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 749             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 750             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 751
 752             }
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             if (gmx_mm_any_lt(rsq13,rcutoff2))
 759             {
 760
 761             /* REACTION-FIELD ELECTROSTATICS */
 762             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 763             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 764
 765             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 766
 767             /* Update potential sum for this i atom from the interaction with this j atom. */
 768             velec            = _mm_and_ps(velec,cutoff_mask);
 769             velec            = _mm_andnot_ps(dummy_mask,velec);
 770             velecsum         = _mm_add_ps(velecsum,velec);
 771
 772             fscal            = felec;
 773
 774             fscal            = _mm_and_ps(fscal,cutoff_mask);
 775
 776             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 777
 778              /* Update vectorial force */
 779             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 780             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 781             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 782
 783             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 784             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 785             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 786
 787             }
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             if (gmx_mm_any_lt(rsq21,rcutoff2))
 794             {
 795
 796             /* REACTION-FIELD ELECTROSTATICS */
 797             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 798             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 799
 800             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_and_ps(velec,cutoff_mask);
 804             velec            = _mm_andnot_ps(dummy_mask,velec);
 805             velecsum         = _mm_add_ps(velecsum,velec);
 806
 807             fscal            = felec;
 808
 809             fscal            = _mm_and_ps(fscal,cutoff_mask);
 810
 811             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 812
 813              /* Update vectorial force */
 814             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 815             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 816             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 817
 818             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 819             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 820             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 821
 822             }
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (gmx_mm_any_lt(rsq22,rcutoff2))
 829             {
 830
 831             /* REACTION-FIELD ELECTROSTATICS */
 832             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 833             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 834
 835             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_and_ps(velec,cutoff_mask);
 839             velec            = _mm_andnot_ps(dummy_mask,velec);
 840             velecsum         = _mm_add_ps(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_and_ps(fscal,cutoff_mask);
 845
 846             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 847
 848              /* Update vectorial force */
 849             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 850             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 851             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 852
 853             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 854             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 855             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 856
 857             }
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             if (gmx_mm_any_lt(rsq23,rcutoff2))
 864             {
 865
 866             /* REACTION-FIELD ELECTROSTATICS */
 867             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 868             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 869
 870             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 871
 872             /* Update potential sum for this i atom from the interaction with this j atom. */
 873             velec            = _mm_and_ps(velec,cutoff_mask);
 874             velec            = _mm_andnot_ps(dummy_mask,velec);
 875             velecsum         = _mm_add_ps(velecsum,velec);
 876
 877             fscal            = felec;
 878
 879             fscal            = _mm_and_ps(fscal,cutoff_mask);
 880
 881             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 882
 883              /* Update vectorial force */
 884             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 885             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 886             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 887
 888             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 889             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 890             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 891
 892             }
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             if (gmx_mm_any_lt(rsq31,rcutoff2))
 899             {
 900
 901             /* REACTION-FIELD ELECTROSTATICS */
 902             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 903             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 904
 905             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 906
 907             /* Update potential sum for this i atom from the interaction with this j atom. */
 908             velec            = _mm_and_ps(velec,cutoff_mask);
 909             velec            = _mm_andnot_ps(dummy_mask,velec);
 910             velecsum         = _mm_add_ps(velecsum,velec);
 911
 912             fscal            = felec;
 913
 914             fscal            = _mm_and_ps(fscal,cutoff_mask);
 915
 916             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 917
 918              /* Update vectorial force */
 919             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 920             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 921             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 922
 923             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 924             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 925             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 926
 927             }
 928
 929             /**************************
 930              * CALCULATE INTERACTIONS *
 931              **************************/
 932
 933             if (gmx_mm_any_lt(rsq32,rcutoff2))
 934             {
 935
 936             /* REACTION-FIELD ELECTROSTATICS */
 937             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 938             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 939
 940             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 941
 942             /* Update potential sum for this i atom from the interaction with this j atom. */
 943             velec            = _mm_and_ps(velec,cutoff_mask);
 944             velec            = _mm_andnot_ps(dummy_mask,velec);
 945             velecsum         = _mm_add_ps(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_and_ps(fscal,cutoff_mask);
 950
 951             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 952
 953              /* Update vectorial force */
 954             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 955             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 956             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 957
 958             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 959             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 960             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 961
 962             }
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             if (gmx_mm_any_lt(rsq33,rcutoff2))
 969             {
 970
 971             /* REACTION-FIELD ELECTROSTATICS */
 972             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 973             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 974
 975             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 976
 977             /* Update potential sum for this i atom from the interaction with this j atom. */
 978             velec            = _mm_and_ps(velec,cutoff_mask);
 979             velec            = _mm_andnot_ps(dummy_mask,velec);
 980             velecsum         = _mm_add_ps(velecsum,velec);
 981
 982             fscal            = felec;
 983
 984             fscal            = _mm_and_ps(fscal,cutoff_mask);
 985
 986             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 987
 988              /* Update vectorial force */
 989             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 990             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 991             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 992
 993             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 994             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 995             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 996
 997             }
 998
 999             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1000             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1001             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1002             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1003
1004             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1005                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1006
1007             /* Inner loop uses 351 flops */
1008         }
1009
1010         /* End of innermost loop */
1011
1012         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1013                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1014
1015         ggid                        = gid[iidx];
1016         /* Update potential energies */
1017         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1018
1019         /* Increment number of inner iterations */
1020         inneriter                  += j_index_end - j_index_start;
1021
1022         /* Outer loop uses 19 flops */
1023     }
1024
1025     /* Increment number of outer iterations */
1026     outeriter        += nri;
1027
1028     /* Update outer/inner flops */
1029
1030     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*351);
1031 }
1032 /*
1033  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_single
1034  * Electrostatics interaction: ReactionField
1035  * VdW interaction:            None
1036  * Geometry:                   Water4-Water4
1037  * Calculate force/pot:        Force
1038  */
1039 void
1040 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_single
1041                     (t_nblist                    * gmx_restrict       nlist,
1042                      rvec                        * gmx_restrict          xx,
1043                      rvec                        * gmx_restrict          ff,
1044                      t_forcerec                  * gmx_restrict          fr,
1045                      t_mdatoms                   * gmx_restrict     mdatoms,
1046                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1047                      t_nrnb                      * gmx_restrict        nrnb)
1048 {
1049     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1050      * just 0 for non-waters.
1051      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1052      * jnr indices corresponding to data put in the four positions in the SIMD register.
1053      */
1054     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1055     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1056     int              jnrA,jnrB,jnrC,jnrD;
1057     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1058     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1059     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1060     real             rcutoff_scalar;
1061     real             *shiftvec,*fshift,*x,*f;
1062     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1063     real             scratch[4*DIM];
1064     __m128           fscal,rcutoff,rcutoff2,jidxall;
1065     int              vdwioffset1;
1066     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1067     int              vdwioffset2;
1068     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1069     int              vdwioffset3;
1070     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1071     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1072     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1073     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1074     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1075     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1076     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1077     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1078     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1079     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1080     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1081     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1082     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1083     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1084     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1085     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1086     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1087     real             *charge;
1088     __m128           dummy_mask,cutoff_mask;
1089     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1090     __m128           one     = _mm_set1_ps(1.0);
1091     __m128           two     = _mm_set1_ps(2.0);
1092     x                = xx[0];
1093     f                = ff[0];
1094
1095     nri              = nlist->nri;
1096     iinr             = nlist->iinr;
1097     jindex           = nlist->jindex;
1098     jjnr             = nlist->jjnr;
1099     shiftidx         = nlist->shift;
1100     gid              = nlist->gid;
1101     shiftvec         = fr->shift_vec[0];
1102     fshift           = fr->fshift[0];
1103     facel            = _mm_set1_ps(fr->epsfac);
1104     charge           = mdatoms->chargeA;
1105     krf              = _mm_set1_ps(fr->ic->k_rf);
1106     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1107     crf              = _mm_set1_ps(fr->ic->c_rf);
1108
1109     /* Setup water-specific parameters */
1110     inr              = nlist->iinr[0];
1111     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1112     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1113     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1114
1115     jq1              = _mm_set1_ps(charge[inr+1]);
1116     jq2              = _mm_set1_ps(charge[inr+2]);
1117     jq3              = _mm_set1_ps(charge[inr+3]);
1118     qq11             = _mm_mul_ps(iq1,jq1);
1119     qq12             = _mm_mul_ps(iq1,jq2);
1120     qq13             = _mm_mul_ps(iq1,jq3);
1121     qq21             = _mm_mul_ps(iq2,jq1);
1122     qq22             = _mm_mul_ps(iq2,jq2);
1123     qq23             = _mm_mul_ps(iq2,jq3);
1124     qq31             = _mm_mul_ps(iq3,jq1);
1125     qq32             = _mm_mul_ps(iq3,jq2);
1126     qq33             = _mm_mul_ps(iq3,jq3);
1127
1128     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1129     rcutoff_scalar   = fr->rcoulomb;
1130     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1131     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1132
1133     /* Avoid stupid compiler warnings */
1134     jnrA = jnrB = jnrC = jnrD = 0;
1135     j_coord_offsetA = 0;
1136     j_coord_offsetB = 0;
1137     j_coord_offsetC = 0;
1138     j_coord_offsetD = 0;
1139
1140     outeriter        = 0;
1141     inneriter        = 0;
1142
1143     for(iidx=0;iidx<4*DIM;iidx++)
1144     {
1145         scratch[iidx] = 0.0;
1146     }
1147
1148     /* Start outer loop over neighborlists */
1149     for(iidx=0; iidx<nri; iidx++)
1150     {
1151         /* Load shift vector for this list */
1152         i_shift_offset   = DIM*shiftidx[iidx];
1153
1154         /* Load limits for loop over neighbors */
1155         j_index_start    = jindex[iidx];
1156         j_index_end      = jindex[iidx+1];
1157
1158         /* Get outer coordinate index */
1159         inr              = iinr[iidx];
1160         i_coord_offset   = DIM*inr;
1161
1162         /* Load i particle coords and add shift vector */
1163         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1164                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1165
1166         fix1             = _mm_setzero_ps();
1167         fiy1             = _mm_setzero_ps();
1168         fiz1             = _mm_setzero_ps();
1169         fix2             = _mm_setzero_ps();
1170         fiy2             = _mm_setzero_ps();
1171         fiz2             = _mm_setzero_ps();
1172         fix3             = _mm_setzero_ps();
1173         fiy3             = _mm_setzero_ps();
1174         fiz3             = _mm_setzero_ps();
1175
1176         /* Start inner kernel loop */
1177         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1178         {
1179
1180             /* Get j neighbor index, and coordinate index */
1181             jnrA             = jjnr[jidx];
1182             jnrB             = jjnr[jidx+1];
1183             jnrC             = jjnr[jidx+2];
1184             jnrD             = jjnr[jidx+3];
1185             j_coord_offsetA  = DIM*jnrA;
1186             j_coord_offsetB  = DIM*jnrB;
1187             j_coord_offsetC  = DIM*jnrC;
1188             j_coord_offsetD  = DIM*jnrD;
1189
1190             /* load j atom coordinates */
1191             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1192                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1193                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1194
1195             /* Calculate displacement vector */
1196             dx11             = _mm_sub_ps(ix1,jx1);
1197             dy11             = _mm_sub_ps(iy1,jy1);
1198             dz11             = _mm_sub_ps(iz1,jz1);
1199             dx12             = _mm_sub_ps(ix1,jx2);
1200             dy12             = _mm_sub_ps(iy1,jy2);
1201             dz12             = _mm_sub_ps(iz1,jz2);
1202             dx13             = _mm_sub_ps(ix1,jx3);
1203             dy13             = _mm_sub_ps(iy1,jy3);
1204             dz13             = _mm_sub_ps(iz1,jz3);
1205             dx21             = _mm_sub_ps(ix2,jx1);
1206             dy21             = _mm_sub_ps(iy2,jy1);
1207             dz21             = _mm_sub_ps(iz2,jz1);
1208             dx22             = _mm_sub_ps(ix2,jx2);
1209             dy22             = _mm_sub_ps(iy2,jy2);
1210             dz22             = _mm_sub_ps(iz2,jz2);
1211             dx23             = _mm_sub_ps(ix2,jx3);
1212             dy23             = _mm_sub_ps(iy2,jy3);
1213             dz23             = _mm_sub_ps(iz2,jz3);
1214             dx31             = _mm_sub_ps(ix3,jx1);
1215             dy31             = _mm_sub_ps(iy3,jy1);
1216             dz31             = _mm_sub_ps(iz3,jz1);
1217             dx32             = _mm_sub_ps(ix3,jx2);
1218             dy32             = _mm_sub_ps(iy3,jy2);
1219             dz32             = _mm_sub_ps(iz3,jz2);
1220             dx33             = _mm_sub_ps(ix3,jx3);
1221             dy33             = _mm_sub_ps(iy3,jy3);
1222             dz33             = _mm_sub_ps(iz3,jz3);
1223
1224             /* Calculate squared distance and things based on it */
1225             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1226             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1227             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1228             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1229             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1230             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1231             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1232             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1233             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1234
1235             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1236             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1237             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1238             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1239             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1240             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1241             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1242             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1243             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1244
1245             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1246             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1247             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1248             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1249             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1250             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1251             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1252             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1253             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1254
1255             fjx1             = _mm_setzero_ps();
1256             fjy1             = _mm_setzero_ps();
1257             fjz1             = _mm_setzero_ps();
1258             fjx2             = _mm_setzero_ps();
1259             fjy2             = _mm_setzero_ps();
1260             fjz2             = _mm_setzero_ps();
1261             fjx3             = _mm_setzero_ps();
1262             fjy3             = _mm_setzero_ps();
1263             fjz3             = _mm_setzero_ps();
1264
1265             /**************************
1266              * CALCULATE INTERACTIONS *
1267              **************************/
1268
1269             if (gmx_mm_any_lt(rsq11,rcutoff2))
1270             {
1271
1272             /* REACTION-FIELD ELECTROSTATICS */
1273             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1274
1275             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1276
1277             fscal            = felec;
1278
1279             fscal            = _mm_and_ps(fscal,cutoff_mask);
1280
1281              /* Update vectorial force */
1282             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1283             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1284             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1285
1286             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1287             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1288             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1289
1290             }
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             if (gmx_mm_any_lt(rsq12,rcutoff2))
1297             {
1298
1299             /* REACTION-FIELD ELECTROSTATICS */
1300             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1301
1302             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1303
1304             fscal            = felec;
1305
1306             fscal            = _mm_and_ps(fscal,cutoff_mask);
1307
1308              /* Update vectorial force */
1309             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1310             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1311             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1312
1313             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1314             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1315             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1316
1317             }
1318
1319             /**************************
1320              * CALCULATE INTERACTIONS *
1321              **************************/
1322
1323             if (gmx_mm_any_lt(rsq13,rcutoff2))
1324             {
1325
1326             /* REACTION-FIELD ELECTROSTATICS */
1327             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1328
1329             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1330
1331             fscal            = felec;
1332
1333             fscal            = _mm_and_ps(fscal,cutoff_mask);
1334
1335              /* Update vectorial force */
1336             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1337             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1338             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1339
1340             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1341             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1342             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1343
1344             }
1345
1346             /**************************
1347              * CALCULATE INTERACTIONS *
1348              **************************/
1349
1350             if (gmx_mm_any_lt(rsq21,rcutoff2))
1351             {
1352
1353             /* REACTION-FIELD ELECTROSTATICS */
1354             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1355
1356             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1357
1358             fscal            = felec;
1359
1360             fscal            = _mm_and_ps(fscal,cutoff_mask);
1361
1362              /* Update vectorial force */
1363             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1364             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1365             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1366
1367             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1368             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1369             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1370
1371             }
1372
1373             /**************************
1374              * CALCULATE INTERACTIONS *
1375              **************************/
1376
1377             if (gmx_mm_any_lt(rsq22,rcutoff2))
1378             {
1379
1380             /* REACTION-FIELD ELECTROSTATICS */
1381             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1382
1383             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1384
1385             fscal            = felec;
1386
1387             fscal            = _mm_and_ps(fscal,cutoff_mask);
1388
1389              /* Update vectorial force */
1390             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1391             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1392             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1393
1394             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1395             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1396             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1397
1398             }
1399
1400             /**************************
1401              * CALCULATE INTERACTIONS *
1402              **************************/
1403
1404             if (gmx_mm_any_lt(rsq23,rcutoff2))
1405             {
1406
1407             /* REACTION-FIELD ELECTROSTATICS */
1408             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1409
1410             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1411
1412             fscal            = felec;
1413
1414             fscal            = _mm_and_ps(fscal,cutoff_mask);
1415
1416              /* Update vectorial force */
1417             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1418             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1419             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1420
1421             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1422             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1423             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1424
1425             }
1426
1427             /**************************
1428              * CALCULATE INTERACTIONS *
1429              **************************/
1430
1431             if (gmx_mm_any_lt(rsq31,rcutoff2))
1432             {
1433
1434             /* REACTION-FIELD ELECTROSTATICS */
1435             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1436
1437             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1438
1439             fscal            = felec;
1440
1441             fscal            = _mm_and_ps(fscal,cutoff_mask);
1442
1443              /* Update vectorial force */
1444             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1445             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1446             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1447
1448             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1449             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1450             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1451
1452             }
1453
1454             /**************************
1455              * CALCULATE INTERACTIONS *
1456              **************************/
1457
1458             if (gmx_mm_any_lt(rsq32,rcutoff2))
1459             {
1460
1461             /* REACTION-FIELD ELECTROSTATICS */
1462             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1463
1464             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1465
1466             fscal            = felec;
1467
1468             fscal            = _mm_and_ps(fscal,cutoff_mask);
1469
1470              /* Update vectorial force */
1471             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1472             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1473             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1474
1475             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1476             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1477             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1478
1479             }
1480
1481             /**************************
1482              * CALCULATE INTERACTIONS *
1483              **************************/
1484
1485             if (gmx_mm_any_lt(rsq33,rcutoff2))
1486             {
1487
1488             /* REACTION-FIELD ELECTROSTATICS */
1489             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1490
1491             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1492
1493             fscal            = felec;
1494
1495             fscal            = _mm_and_ps(fscal,cutoff_mask);
1496
1497              /* Update vectorial force */
1498             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1499             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1500             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1501
1502             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1503             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1504             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1505
1506             }
1507
1508             fjptrA             = f+j_coord_offsetA;
1509             fjptrB             = f+j_coord_offsetB;
1510             fjptrC             = f+j_coord_offsetC;
1511             fjptrD             = f+j_coord_offsetD;
1512
1513             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1514                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1515
1516             /* Inner loop uses 297 flops */
1517         }
1518
1519         if(jidx<j_index_end)
1520         {
1521
1522             /* Get j neighbor index, and coordinate index */
1523             jnrlistA         = jjnr[jidx];
1524             jnrlistB         = jjnr[jidx+1];
1525             jnrlistC         = jjnr[jidx+2];
1526             jnrlistD         = jjnr[jidx+3];
1527             /* Sign of each element will be negative for non-real atoms.
1528              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1529              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1530              */
1531             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1532             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1533             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1534             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1535             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1536             j_coord_offsetA  = DIM*jnrA;
1537             j_coord_offsetB  = DIM*jnrB;
1538             j_coord_offsetC  = DIM*jnrC;
1539             j_coord_offsetD  = DIM*jnrD;
1540
1541             /* load j atom coordinates */
1542             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1543                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1544                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1545
1546             /* Calculate displacement vector */
1547             dx11             = _mm_sub_ps(ix1,jx1);
1548             dy11             = _mm_sub_ps(iy1,jy1);
1549             dz11             = _mm_sub_ps(iz1,jz1);
1550             dx12             = _mm_sub_ps(ix1,jx2);
1551             dy12             = _mm_sub_ps(iy1,jy2);
1552             dz12             = _mm_sub_ps(iz1,jz2);
1553             dx13             = _mm_sub_ps(ix1,jx3);
1554             dy13             = _mm_sub_ps(iy1,jy3);
1555             dz13             = _mm_sub_ps(iz1,jz3);
1556             dx21             = _mm_sub_ps(ix2,jx1);
1557             dy21             = _mm_sub_ps(iy2,jy1);
1558             dz21             = _mm_sub_ps(iz2,jz1);
1559             dx22             = _mm_sub_ps(ix2,jx2);
1560             dy22             = _mm_sub_ps(iy2,jy2);
1561             dz22             = _mm_sub_ps(iz2,jz2);
1562             dx23             = _mm_sub_ps(ix2,jx3);
1563             dy23             = _mm_sub_ps(iy2,jy3);
1564             dz23             = _mm_sub_ps(iz2,jz3);
1565             dx31             = _mm_sub_ps(ix3,jx1);
1566             dy31             = _mm_sub_ps(iy3,jy1);
1567             dz31             = _mm_sub_ps(iz3,jz1);
1568             dx32             = _mm_sub_ps(ix3,jx2);
1569             dy32             = _mm_sub_ps(iy3,jy2);
1570             dz32             = _mm_sub_ps(iz3,jz2);
1571             dx33             = _mm_sub_ps(ix3,jx3);
1572             dy33             = _mm_sub_ps(iy3,jy3);
1573             dz33             = _mm_sub_ps(iz3,jz3);
1574
1575             /* Calculate squared distance and things based on it */
1576             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1577             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1578             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1579             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1580             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1581             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1582             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1583             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1584             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1585
1586             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1587             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1588             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1589             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1590             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1591             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1592             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1593             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1594             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1595
1596             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1597             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1598             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1599             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1600             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1601             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1602             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1603             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1604             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1605
1606             fjx1             = _mm_setzero_ps();
1607             fjy1             = _mm_setzero_ps();
1608             fjz1             = _mm_setzero_ps();
1609             fjx2             = _mm_setzero_ps();
1610             fjy2             = _mm_setzero_ps();
1611             fjz2             = _mm_setzero_ps();
1612             fjx3             = _mm_setzero_ps();
1613             fjy3             = _mm_setzero_ps();
1614             fjz3             = _mm_setzero_ps();
1615
1616             /**************************
1617              * CALCULATE INTERACTIONS *
1618              **************************/
1619
1620             if (gmx_mm_any_lt(rsq11,rcutoff2))
1621             {
1622
1623             /* REACTION-FIELD ELECTROSTATICS */
1624             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1625
1626             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1627
1628             fscal            = felec;
1629
1630             fscal            = _mm_and_ps(fscal,cutoff_mask);
1631
1632             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1633
1634              /* Update vectorial force */
1635             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1636             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1637             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1638
1639             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1640             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1641             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1642
1643             }
1644
1645             /**************************
1646              * CALCULATE INTERACTIONS *
1647              **************************/
1648
1649             if (gmx_mm_any_lt(rsq12,rcutoff2))
1650             {
1651
1652             /* REACTION-FIELD ELECTROSTATICS */
1653             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1654
1655             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1656
1657             fscal            = felec;
1658
1659             fscal            = _mm_and_ps(fscal,cutoff_mask);
1660
1661             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1662
1663              /* Update vectorial force */
1664             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1665             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1666             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1667
1668             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1669             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1670             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1671
1672             }
1673
1674             /**************************
1675              * CALCULATE INTERACTIONS *
1676              **************************/
1677
1678             if (gmx_mm_any_lt(rsq13,rcutoff2))
1679             {
1680
1681             /* REACTION-FIELD ELECTROSTATICS */
1682             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1683
1684             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1685
1686             fscal            = felec;
1687
1688             fscal            = _mm_and_ps(fscal,cutoff_mask);
1689
1690             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1691
1692              /* Update vectorial force */
1693             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1694             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1695             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1696
1697             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1698             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1699             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1700
1701             }
1702
1703             /**************************
1704              * CALCULATE INTERACTIONS *
1705              **************************/
1706
1707             if (gmx_mm_any_lt(rsq21,rcutoff2))
1708             {
1709
1710             /* REACTION-FIELD ELECTROSTATICS */
1711             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1712
1713             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1714
1715             fscal            = felec;
1716
1717             fscal            = _mm_and_ps(fscal,cutoff_mask);
1718
1719             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1720
1721              /* Update vectorial force */
1722             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1723             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1724             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1725
1726             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1727             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1728             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1729
1730             }
1731
1732             /**************************
1733              * CALCULATE INTERACTIONS *
1734              **************************/
1735
1736             if (gmx_mm_any_lt(rsq22,rcutoff2))
1737             {
1738
1739             /* REACTION-FIELD ELECTROSTATICS */
1740             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1741
1742             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1743
1744             fscal            = felec;
1745
1746             fscal            = _mm_and_ps(fscal,cutoff_mask);
1747
1748             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1749
1750              /* Update vectorial force */
1751             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1752             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1753             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1754
1755             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1756             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1757             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1758
1759             }
1760
1761             /**************************
1762              * CALCULATE INTERACTIONS *
1763              **************************/
1764
1765             if (gmx_mm_any_lt(rsq23,rcutoff2))
1766             {
1767
1768             /* REACTION-FIELD ELECTROSTATICS */
1769             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1770
1771             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1772
1773             fscal            = felec;
1774
1775             fscal            = _mm_and_ps(fscal,cutoff_mask);
1776
1777             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1778
1779              /* Update vectorial force */
1780             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1781             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1782             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1783
1784             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1785             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1786             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1787
1788             }
1789
1790             /**************************
1791              * CALCULATE INTERACTIONS *
1792              **************************/
1793
1794             if (gmx_mm_any_lt(rsq31,rcutoff2))
1795             {
1796
1797             /* REACTION-FIELD ELECTROSTATICS */
1798             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1799
1800             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1801
1802             fscal            = felec;
1803
1804             fscal            = _mm_and_ps(fscal,cutoff_mask);
1805
1806             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1807
1808              /* Update vectorial force */
1809             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1810             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1811             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1812
1813             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1814             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1815             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1816
1817             }
1818
1819             /**************************
1820              * CALCULATE INTERACTIONS *
1821              **************************/
1822
1823             if (gmx_mm_any_lt(rsq32,rcutoff2))
1824             {
1825
1826             /* REACTION-FIELD ELECTROSTATICS */
1827             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1828
1829             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1830
1831             fscal            = felec;
1832
1833             fscal            = _mm_and_ps(fscal,cutoff_mask);
1834
1835             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1836
1837              /* Update vectorial force */
1838             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1839             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1840             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1841
1842             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1843             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1844             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1845
1846             }
1847
1848             /**************************
1849              * CALCULATE INTERACTIONS *
1850              **************************/
1851
1852             if (gmx_mm_any_lt(rsq33,rcutoff2))
1853             {
1854
1855             /* REACTION-FIELD ELECTROSTATICS */
1856             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1857
1858             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1859
1860             fscal            = felec;
1861
1862             fscal            = _mm_and_ps(fscal,cutoff_mask);
1863
1864             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1865
1866              /* Update vectorial force */
1867             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1868             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1869             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1870
1871             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1872             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1873             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1874
1875             }
1876
1877             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1878             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1879             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1880             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1881
1882             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1883                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1884
1885             /* Inner loop uses 297 flops */
1886         }
1887
1888         /* End of innermost loop */
1889
1890         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1891                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1892
1893         /* Increment number of inner iterations */
1894         inneriter                  += j_index_end - j_index_start;
1895
1896         /* Outer loop uses 18 flops */
1897     }
1898
1899     /* Increment number of outer iterations */
1900     outeriter        += nri;
1901
1902     /* Update outer/inner flops */
1903
1904     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*297);
1905 }