src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_VdwNone_GeomW4W4_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_avx_128_fma_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset1;
  70     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  71     int              vdwioffset2;
  72     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  73     int              vdwioffset3;
  74     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  75     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  76     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  77     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  78     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  79     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  80     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  81     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  82     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  83     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  84     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  85     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  86     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  87     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  88     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  89     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128           dummy_mask,cutoff_mask;
  93     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  94     __m128           one     = _mm_set1_ps(1.0);
  95     __m128           two     = _mm_set1_ps(2.0);
  96     x                = xx[0];
  97     f                = ff[0];
  98
  99     nri              = nlist->nri;
 100     iinr             = nlist->iinr;
 101     jindex           = nlist->jindex;
 102     jjnr             = nlist->jjnr;
 103     shiftidx         = nlist->shift;
 104     gid              = nlist->gid;
 105     shiftvec         = fr->shift_vec[0];
 106     fshift           = fr->fshift[0];
 107     facel            = _mm_set1_ps(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109
 110     /* Setup water-specific parameters */
 111     inr              = nlist->iinr[0];
 112     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 113     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 114     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 115
 116     jq1              = _mm_set1_ps(charge[inr+1]);
 117     jq2              = _mm_set1_ps(charge[inr+2]);
 118     jq3              = _mm_set1_ps(charge[inr+3]);
 119     qq11             = _mm_mul_ps(iq1,jq1);
 120     qq12             = _mm_mul_ps(iq1,jq2);
 121     qq13             = _mm_mul_ps(iq1,jq3);
 122     qq21             = _mm_mul_ps(iq2,jq1);
 123     qq22             = _mm_mul_ps(iq2,jq2);
 124     qq23             = _mm_mul_ps(iq2,jq3);
 125     qq31             = _mm_mul_ps(iq3,jq1);
 126     qq32             = _mm_mul_ps(iq3,jq2);
 127     qq33             = _mm_mul_ps(iq3,jq3);
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = jnrC = jnrD = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133     j_coord_offsetC = 0;
 134     j_coord_offsetD = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     for(iidx=0;iidx<4*DIM;iidx++)
 140     {
 141         scratch[iidx] = 0.0;
 142     }
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 160                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 161
 162         fix1             = _mm_setzero_ps();
 163         fiy1             = _mm_setzero_ps();
 164         fiz1             = _mm_setzero_ps();
 165         fix2             = _mm_setzero_ps();
 166         fiy2             = _mm_setzero_ps();
 167         fiz2             = _mm_setzero_ps();
 168         fix3             = _mm_setzero_ps();
 169         fiy3             = _mm_setzero_ps();
 170         fiz3             = _mm_setzero_ps();
 171
 172         /* Reset potential sums */
 173         velecsum         = _mm_setzero_ps();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188
 189             /* load j atom coordinates */
 190             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 191                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 192                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 193
 194             /* Calculate displacement vector */
 195             dx11             = _mm_sub_ps(ix1,jx1);
 196             dy11             = _mm_sub_ps(iy1,jy1);
 197             dz11             = _mm_sub_ps(iz1,jz1);
 198             dx12             = _mm_sub_ps(ix1,jx2);
 199             dy12             = _mm_sub_ps(iy1,jy2);
 200             dz12             = _mm_sub_ps(iz1,jz2);
 201             dx13             = _mm_sub_ps(ix1,jx3);
 202             dy13             = _mm_sub_ps(iy1,jy3);
 203             dz13             = _mm_sub_ps(iz1,jz3);
 204             dx21             = _mm_sub_ps(ix2,jx1);
 205             dy21             = _mm_sub_ps(iy2,jy1);
 206             dz21             = _mm_sub_ps(iz2,jz1);
 207             dx22             = _mm_sub_ps(ix2,jx2);
 208             dy22             = _mm_sub_ps(iy2,jy2);
 209             dz22             = _mm_sub_ps(iz2,jz2);
 210             dx23             = _mm_sub_ps(ix2,jx3);
 211             dy23             = _mm_sub_ps(iy2,jy3);
 212             dz23             = _mm_sub_ps(iz2,jz3);
 213             dx31             = _mm_sub_ps(ix3,jx1);
 214             dy31             = _mm_sub_ps(iy3,jy1);
 215             dz31             = _mm_sub_ps(iz3,jz1);
 216             dx32             = _mm_sub_ps(ix3,jx2);
 217             dy32             = _mm_sub_ps(iy3,jy2);
 218             dz32             = _mm_sub_ps(iz3,jz2);
 219             dx33             = _mm_sub_ps(ix3,jx3);
 220             dy33             = _mm_sub_ps(iy3,jy3);
 221             dz33             = _mm_sub_ps(iz3,jz3);
 222
 223             /* Calculate squared distance and things based on it */
 224             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 225             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 226             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 227             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 228             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 229             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 230             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 231             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 232             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 233
 234             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 235             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 236             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 237             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 238             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 239             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 240             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 241             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 242             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 243
 244             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 245             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 246             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 247             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 248             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 249             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 250             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 251             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 252             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 253
 254             fjx1             = _mm_setzero_ps();
 255             fjy1             = _mm_setzero_ps();
 256             fjz1             = _mm_setzero_ps();
 257             fjx2             = _mm_setzero_ps();
 258             fjy2             = _mm_setzero_ps();
 259             fjz2             = _mm_setzero_ps();
 260             fjx3             = _mm_setzero_ps();
 261             fjy3             = _mm_setzero_ps();
 262             fjz3             = _mm_setzero_ps();
 263
 264             /**************************
 265              * CALCULATE INTERACTIONS *
 266              **************************/
 267
 268             /* COULOMB ELECTROSTATICS */
 269             velec            = _mm_mul_ps(qq11,rinv11);
 270             felec            = _mm_mul_ps(velec,rinvsq11);
 271
 272             /* Update potential sum for this i atom from the interaction with this j atom. */
 273             velecsum         = _mm_add_ps(velecsum,velec);
 274
 275             fscal            = felec;
 276
 277              /* Update vectorial force */
 278             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 279             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 280             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 281
 282             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 283             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 284             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 285
 286             /**************************
 287              * CALCULATE INTERACTIONS *
 288              **************************/
 289
 290             /* COULOMB ELECTROSTATICS */
 291             velec            = _mm_mul_ps(qq12,rinv12);
 292             felec            = _mm_mul_ps(velec,rinvsq12);
 293
 294             /* Update potential sum for this i atom from the interaction with this j atom. */
 295             velecsum         = _mm_add_ps(velecsum,velec);
 296
 297             fscal            = felec;
 298
 299              /* Update vectorial force */
 300             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 301             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 302             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 303
 304             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 305             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 306             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             /* COULOMB ELECTROSTATICS */
 313             velec            = _mm_mul_ps(qq13,rinv13);
 314             felec            = _mm_mul_ps(velec,rinvsq13);
 315
 316             /* Update potential sum for this i atom from the interaction with this j atom. */
 317             velecsum         = _mm_add_ps(velecsum,velec);
 318
 319             fscal            = felec;
 320
 321              /* Update vectorial force */
 322             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 323             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 324             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 325
 326             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 327             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 328             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             /* COULOMB ELECTROSTATICS */
 335             velec            = _mm_mul_ps(qq21,rinv21);
 336             felec            = _mm_mul_ps(velec,rinvsq21);
 337
 338             /* Update potential sum for this i atom from the interaction with this j atom. */
 339             velecsum         = _mm_add_ps(velecsum,velec);
 340
 341             fscal            = felec;
 342
 343              /* Update vectorial force */
 344             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 345             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 346             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 347
 348             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 349             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 350             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 351
 352             /**************************
 353              * CALCULATE INTERACTIONS *
 354              **************************/
 355
 356             /* COULOMB ELECTROSTATICS */
 357             velec            = _mm_mul_ps(qq22,rinv22);
 358             felec            = _mm_mul_ps(velec,rinvsq22);
 359
 360             /* Update potential sum for this i atom from the interaction with this j atom. */
 361             velecsum         = _mm_add_ps(velecsum,velec);
 362
 363             fscal            = felec;
 364
 365              /* Update vectorial force */
 366             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 367             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 368             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 369
 370             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 371             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 372             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 373
 374             /**************************
 375              * CALCULATE INTERACTIONS *
 376              **************************/
 377
 378             /* COULOMB ELECTROSTATICS */
 379             velec            = _mm_mul_ps(qq23,rinv23);
 380             felec            = _mm_mul_ps(velec,rinvsq23);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velecsum         = _mm_add_ps(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387              /* Update vectorial force */
 388             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 389             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 390             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 391
 392             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 393             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 394             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 395
 396             /**************************
 397              * CALCULATE INTERACTIONS *
 398              **************************/
 399
 400             /* COULOMB ELECTROSTATICS */
 401             velec            = _mm_mul_ps(qq31,rinv31);
 402             felec            = _mm_mul_ps(velec,rinvsq31);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velecsum         = _mm_add_ps(velecsum,velec);
 406
 407             fscal            = felec;
 408
 409              /* Update vectorial force */
 410             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 411             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 412             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 413
 414             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 415             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 416             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 417
 418             /**************************
 419              * CALCULATE INTERACTIONS *
 420              **************************/
 421
 422             /* COULOMB ELECTROSTATICS */
 423             velec            = _mm_mul_ps(qq32,rinv32);
 424             felec            = _mm_mul_ps(velec,rinvsq32);
 425
 426             /* Update potential sum for this i atom from the interaction with this j atom. */
 427             velecsum         = _mm_add_ps(velecsum,velec);
 428
 429             fscal            = felec;
 430
 431              /* Update vectorial force */
 432             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 433             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 434             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 435
 436             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 437             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 438             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 439
 440             /**************************
 441              * CALCULATE INTERACTIONS *
 442              **************************/
 443
 444             /* COULOMB ELECTROSTATICS */
 445             velec            = _mm_mul_ps(qq33,rinv33);
 446             felec            = _mm_mul_ps(velec,rinvsq33);
 447
 448             /* Update potential sum for this i atom from the interaction with this j atom. */
 449             velecsum         = _mm_add_ps(velecsum,velec);
 450
 451             fscal            = felec;
 452
 453              /* Update vectorial force */
 454             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 455             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 456             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 457
 458             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 459             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 460             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 461
 462             fjptrA             = f+j_coord_offsetA;
 463             fjptrB             = f+j_coord_offsetB;
 464             fjptrC             = f+j_coord_offsetC;
 465             fjptrD             = f+j_coord_offsetD;
 466
 467             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 468                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 469
 470             /* Inner loop uses 279 flops */
 471         }
 472
 473         if(jidx<j_index_end)
 474         {
 475
 476             /* Get j neighbor index, and coordinate index */
 477             jnrlistA         = jjnr[jidx];
 478             jnrlistB         = jjnr[jidx+1];
 479             jnrlistC         = jjnr[jidx+2];
 480             jnrlistD         = jjnr[jidx+3];
 481             /* Sign of each element will be negative for non-real atoms.
 482              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 483              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 484              */
 485             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 486             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 487             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 488             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 489             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 490             j_coord_offsetA  = DIM*jnrA;
 491             j_coord_offsetB  = DIM*jnrB;
 492             j_coord_offsetC  = DIM*jnrC;
 493             j_coord_offsetD  = DIM*jnrD;
 494
 495             /* load j atom coordinates */
 496             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 497                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 498                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 499
 500             /* Calculate displacement vector */
 501             dx11             = _mm_sub_ps(ix1,jx1);
 502             dy11             = _mm_sub_ps(iy1,jy1);
 503             dz11             = _mm_sub_ps(iz1,jz1);
 504             dx12             = _mm_sub_ps(ix1,jx2);
 505             dy12             = _mm_sub_ps(iy1,jy2);
 506             dz12             = _mm_sub_ps(iz1,jz2);
 507             dx13             = _mm_sub_ps(ix1,jx3);
 508             dy13             = _mm_sub_ps(iy1,jy3);
 509             dz13             = _mm_sub_ps(iz1,jz3);
 510             dx21             = _mm_sub_ps(ix2,jx1);
 511             dy21             = _mm_sub_ps(iy2,jy1);
 512             dz21             = _mm_sub_ps(iz2,jz1);
 513             dx22             = _mm_sub_ps(ix2,jx2);
 514             dy22             = _mm_sub_ps(iy2,jy2);
 515             dz22             = _mm_sub_ps(iz2,jz2);
 516             dx23             = _mm_sub_ps(ix2,jx3);
 517             dy23             = _mm_sub_ps(iy2,jy3);
 518             dz23             = _mm_sub_ps(iz2,jz3);
 519             dx31             = _mm_sub_ps(ix3,jx1);
 520             dy31             = _mm_sub_ps(iy3,jy1);
 521             dz31             = _mm_sub_ps(iz3,jz1);
 522             dx32             = _mm_sub_ps(ix3,jx2);
 523             dy32             = _mm_sub_ps(iy3,jy2);
 524             dz32             = _mm_sub_ps(iz3,jz2);
 525             dx33             = _mm_sub_ps(ix3,jx3);
 526             dy33             = _mm_sub_ps(iy3,jy3);
 527             dz33             = _mm_sub_ps(iz3,jz3);
 528
 529             /* Calculate squared distance and things based on it */
 530             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 531             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 532             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 533             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 534             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 535             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 536             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 537             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 538             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 539
 540             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 541             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 542             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 543             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 544             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 545             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 546             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 547             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 548             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 549
 550             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 551             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 552             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 553             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 554             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 555             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 556             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 557             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 558             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 559
 560             fjx1             = _mm_setzero_ps();
 561             fjy1             = _mm_setzero_ps();
 562             fjz1             = _mm_setzero_ps();
 563             fjx2             = _mm_setzero_ps();
 564             fjy2             = _mm_setzero_ps();
 565             fjz2             = _mm_setzero_ps();
 566             fjx3             = _mm_setzero_ps();
 567             fjy3             = _mm_setzero_ps();
 568             fjz3             = _mm_setzero_ps();
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             /* COULOMB ELECTROSTATICS */
 575             velec            = _mm_mul_ps(qq11,rinv11);
 576             felec            = _mm_mul_ps(velec,rinvsq11);
 577
 578             /* Update potential sum for this i atom from the interaction with this j atom. */
 579             velec            = _mm_andnot_ps(dummy_mask,velec);
 580             velecsum         = _mm_add_ps(velecsum,velec);
 581
 582             fscal            = felec;
 583
 584             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 585
 586              /* Update vectorial force */
 587             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 588             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 589             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 590
 591             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 592             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 593             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             /* COULOMB ELECTROSTATICS */
 600             velec            = _mm_mul_ps(qq12,rinv12);
 601             felec            = _mm_mul_ps(velec,rinvsq12);
 602
 603             /* Update potential sum for this i atom from the interaction with this j atom. */
 604             velec            = _mm_andnot_ps(dummy_mask,velec);
 605             velecsum         = _mm_add_ps(velecsum,velec);
 606
 607             fscal            = felec;
 608
 609             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 610
 611              /* Update vectorial force */
 612             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 613             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 614             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 615
 616             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 617             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 618             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 619
 620             /**************************
 621              * CALCULATE INTERACTIONS *
 622              **************************/
 623
 624             /* COULOMB ELECTROSTATICS */
 625             velec            = _mm_mul_ps(qq13,rinv13);
 626             felec            = _mm_mul_ps(velec,rinvsq13);
 627
 628             /* Update potential sum for this i atom from the interaction with this j atom. */
 629             velec            = _mm_andnot_ps(dummy_mask,velec);
 630             velecsum         = _mm_add_ps(velecsum,velec);
 631
 632             fscal            = felec;
 633
 634             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 635
 636              /* Update vectorial force */
 637             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 638             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 639             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 640
 641             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 642             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 643             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 644
 645             /**************************
 646              * CALCULATE INTERACTIONS *
 647              **************************/
 648
 649             /* COULOMB ELECTROSTATICS */
 650             velec            = _mm_mul_ps(qq21,rinv21);
 651             felec            = _mm_mul_ps(velec,rinvsq21);
 652
 653             /* Update potential sum for this i atom from the interaction with this j atom. */
 654             velec            = _mm_andnot_ps(dummy_mask,velec);
 655             velecsum         = _mm_add_ps(velecsum,velec);
 656
 657             fscal            = felec;
 658
 659             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 660
 661              /* Update vectorial force */
 662             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 663             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 664             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 665
 666             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 667             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 668             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 669
 670             /**************************
 671              * CALCULATE INTERACTIONS *
 672              **************************/
 673
 674             /* COULOMB ELECTROSTATICS */
 675             velec            = _mm_mul_ps(qq22,rinv22);
 676             felec            = _mm_mul_ps(velec,rinvsq22);
 677
 678             /* Update potential sum for this i atom from the interaction with this j atom. */
 679             velec            = _mm_andnot_ps(dummy_mask,velec);
 680             velecsum         = _mm_add_ps(velecsum,velec);
 681
 682             fscal            = felec;
 683
 684             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 685
 686              /* Update vectorial force */
 687             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 688             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 689             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 690
 691             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 692             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 693             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 694
 695             /**************************
 696              * CALCULATE INTERACTIONS *
 697              **************************/
 698
 699             /* COULOMB ELECTROSTATICS */
 700             velec            = _mm_mul_ps(qq23,rinv23);
 701             felec            = _mm_mul_ps(velec,rinvsq23);
 702
 703             /* Update potential sum for this i atom from the interaction with this j atom. */
 704             velec            = _mm_andnot_ps(dummy_mask,velec);
 705             velecsum         = _mm_add_ps(velecsum,velec);
 706
 707             fscal            = felec;
 708
 709             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 710
 711              /* Update vectorial force */
 712             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 713             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 714             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 715
 716             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 717             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 718             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 719
 720             /**************************
 721              * CALCULATE INTERACTIONS *
 722              **************************/
 723
 724             /* COULOMB ELECTROSTATICS */
 725             velec            = _mm_mul_ps(qq31,rinv31);
 726             felec            = _mm_mul_ps(velec,rinvsq31);
 727
 728             /* Update potential sum for this i atom from the interaction with this j atom. */
 729             velec            = _mm_andnot_ps(dummy_mask,velec);
 730             velecsum         = _mm_add_ps(velecsum,velec);
 731
 732             fscal            = felec;
 733
 734             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 735
 736              /* Update vectorial force */
 737             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 738             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 739             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 740
 741             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 742             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 743             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 744
 745             /**************************
 746              * CALCULATE INTERACTIONS *
 747              **************************/
 748
 749             /* COULOMB ELECTROSTATICS */
 750             velec            = _mm_mul_ps(qq32,rinv32);
 751             felec            = _mm_mul_ps(velec,rinvsq32);
 752
 753             /* Update potential sum for this i atom from the interaction with this j atom. */
 754             velec            = _mm_andnot_ps(dummy_mask,velec);
 755             velecsum         = _mm_add_ps(velecsum,velec);
 756
 757             fscal            = felec;
 758
 759             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 760
 761              /* Update vectorial force */
 762             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 763             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 764             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 765
 766             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 767             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 768             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 769
 770             /**************************
 771              * CALCULATE INTERACTIONS *
 772              **************************/
 773
 774             /* COULOMB ELECTROSTATICS */
 775             velec            = _mm_mul_ps(qq33,rinv33);
 776             felec            = _mm_mul_ps(velec,rinvsq33);
 777
 778             /* Update potential sum for this i atom from the interaction with this j atom. */
 779             velec            = _mm_andnot_ps(dummy_mask,velec);
 780             velecsum         = _mm_add_ps(velecsum,velec);
 781
 782             fscal            = felec;
 783
 784             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 785
 786              /* Update vectorial force */
 787             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 788             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 789             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 790
 791             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 792             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 793             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 794
 795             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 796             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 797             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 798             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 799
 800             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 801                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 802
 803             /* Inner loop uses 279 flops */
 804         }
 805
 806         /* End of innermost loop */
 807
 808         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 809                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 810
 811         ggid                        = gid[iidx];
 812         /* Update potential energies */
 813         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 814
 815         /* Increment number of inner iterations */
 816         inneriter                  += j_index_end - j_index_start;
 817
 818         /* Outer loop uses 19 flops */
 819     }
 820
 821     /* Increment number of outer iterations */
 822     outeriter        += nri;
 823
 824     /* Update outer/inner flops */
 825
 826     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*279);
 827 }
 828 /*
 829  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_single
 830  * Electrostatics interaction: Coulomb
 831  * VdW interaction:            None
 832  * Geometry:                   Water4-Water4
 833  * Calculate force/pot:        Force
 834  */
 835 void
 836 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_avx_128_fma_single
 837                     (t_nblist * gmx_restrict                nlist,
 838                      rvec * gmx_restrict                    xx,
 839                      rvec * gmx_restrict                    ff,
 840                      t_forcerec * gmx_restrict              fr,
 841                      t_mdatoms * gmx_restrict               mdatoms,
 842                      nb_kernel_data_t * gmx_restrict        kernel_data,
 843                      t_nrnb * gmx_restrict                  nrnb)
 844 {
 845     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 846      * just 0 for non-waters.
 847      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 848      * jnr indices corresponding to data put in the four positions in the SIMD register.
 849      */
 850     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 851     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 852     int              jnrA,jnrB,jnrC,jnrD;
 853     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 854     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 855     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 856     real             rcutoff_scalar;
 857     real             *shiftvec,*fshift,*x,*f;
 858     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 859     real             scratch[4*DIM];
 860     __m128           fscal,rcutoff,rcutoff2,jidxall;
 861     int              vdwioffset1;
 862     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 863     int              vdwioffset2;
 864     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 865     int              vdwioffset3;
 866     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 867     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 868     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 869     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 870     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 871     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 872     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 873     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 874     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 875     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 876     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 877     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 878     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 879     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 880     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 881     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 882     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 883     real             *charge;
 884     __m128           dummy_mask,cutoff_mask;
 885     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 886     __m128           one     = _mm_set1_ps(1.0);
 887     __m128           two     = _mm_set1_ps(2.0);
 888     x                = xx[0];
 889     f                = ff[0];
 890
 891     nri              = nlist->nri;
 892     iinr             = nlist->iinr;
 893     jindex           = nlist->jindex;
 894     jjnr             = nlist->jjnr;
 895     shiftidx         = nlist->shift;
 896     gid              = nlist->gid;
 897     shiftvec         = fr->shift_vec[0];
 898     fshift           = fr->fshift[0];
 899     facel            = _mm_set1_ps(fr->epsfac);
 900     charge           = mdatoms->chargeA;
 901
 902     /* Setup water-specific parameters */
 903     inr              = nlist->iinr[0];
 904     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 905     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 906     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 907
 908     jq1              = _mm_set1_ps(charge[inr+1]);
 909     jq2              = _mm_set1_ps(charge[inr+2]);
 910     jq3              = _mm_set1_ps(charge[inr+3]);
 911     qq11             = _mm_mul_ps(iq1,jq1);
 912     qq12             = _mm_mul_ps(iq1,jq2);
 913     qq13             = _mm_mul_ps(iq1,jq3);
 914     qq21             = _mm_mul_ps(iq2,jq1);
 915     qq22             = _mm_mul_ps(iq2,jq2);
 916     qq23             = _mm_mul_ps(iq2,jq3);
 917     qq31             = _mm_mul_ps(iq3,jq1);
 918     qq32             = _mm_mul_ps(iq3,jq2);
 919     qq33             = _mm_mul_ps(iq3,jq3);
 920
 921     /* Avoid stupid compiler warnings */
 922     jnrA = jnrB = jnrC = jnrD = 0;
 923     j_coord_offsetA = 0;
 924     j_coord_offsetB = 0;
 925     j_coord_offsetC = 0;
 926     j_coord_offsetD = 0;
 927
 928     outeriter        = 0;
 929     inneriter        = 0;
 930
 931     for(iidx=0;iidx<4*DIM;iidx++)
 932     {
 933         scratch[iidx] = 0.0;
 934     }
 935
 936     /* Start outer loop over neighborlists */
 937     for(iidx=0; iidx<nri; iidx++)
 938     {
 939         /* Load shift vector for this list */
 940         i_shift_offset   = DIM*shiftidx[iidx];
 941
 942         /* Load limits for loop over neighbors */
 943         j_index_start    = jindex[iidx];
 944         j_index_end      = jindex[iidx+1];
 945
 946         /* Get outer coordinate index */
 947         inr              = iinr[iidx];
 948         i_coord_offset   = DIM*inr;
 949
 950         /* Load i particle coords and add shift vector */
 951         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 952                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 953
 954         fix1             = _mm_setzero_ps();
 955         fiy1             = _mm_setzero_ps();
 956         fiz1             = _mm_setzero_ps();
 957         fix2             = _mm_setzero_ps();
 958         fiy2             = _mm_setzero_ps();
 959         fiz2             = _mm_setzero_ps();
 960         fix3             = _mm_setzero_ps();
 961         fiy3             = _mm_setzero_ps();
 962         fiz3             = _mm_setzero_ps();
 963
 964         /* Start inner kernel loop */
 965         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 966         {
 967
 968             /* Get j neighbor index, and coordinate index */
 969             jnrA             = jjnr[jidx];
 970             jnrB             = jjnr[jidx+1];
 971             jnrC             = jjnr[jidx+2];
 972             jnrD             = jjnr[jidx+3];
 973             j_coord_offsetA  = DIM*jnrA;
 974             j_coord_offsetB  = DIM*jnrB;
 975             j_coord_offsetC  = DIM*jnrC;
 976             j_coord_offsetD  = DIM*jnrD;
 977
 978             /* load j atom coordinates */
 979             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 980                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 981                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 982
 983             /* Calculate displacement vector */
 984             dx11             = _mm_sub_ps(ix1,jx1);
 985             dy11             = _mm_sub_ps(iy1,jy1);
 986             dz11             = _mm_sub_ps(iz1,jz1);
 987             dx12             = _mm_sub_ps(ix1,jx2);
 988             dy12             = _mm_sub_ps(iy1,jy2);
 989             dz12             = _mm_sub_ps(iz1,jz2);
 990             dx13             = _mm_sub_ps(ix1,jx3);
 991             dy13             = _mm_sub_ps(iy1,jy3);
 992             dz13             = _mm_sub_ps(iz1,jz3);
 993             dx21             = _mm_sub_ps(ix2,jx1);
 994             dy21             = _mm_sub_ps(iy2,jy1);
 995             dz21             = _mm_sub_ps(iz2,jz1);
 996             dx22             = _mm_sub_ps(ix2,jx2);
 997             dy22             = _mm_sub_ps(iy2,jy2);
 998             dz22             = _mm_sub_ps(iz2,jz2);
 999             dx23             = _mm_sub_ps(ix2,jx3);
1000             dy23             = _mm_sub_ps(iy2,jy3);
1001             dz23             = _mm_sub_ps(iz2,jz3);
1002             dx31             = _mm_sub_ps(ix3,jx1);
1003             dy31             = _mm_sub_ps(iy3,jy1);
1004             dz31             = _mm_sub_ps(iz3,jz1);
1005             dx32             = _mm_sub_ps(ix3,jx2);
1006             dy32             = _mm_sub_ps(iy3,jy2);
1007             dz32             = _mm_sub_ps(iz3,jz2);
1008             dx33             = _mm_sub_ps(ix3,jx3);
1009             dy33             = _mm_sub_ps(iy3,jy3);
1010             dz33             = _mm_sub_ps(iz3,jz3);
1011
1012             /* Calculate squared distance and things based on it */
1013             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1014             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1015             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1016             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1017             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1018             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1019             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1020             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1021             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1022
1023             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1024             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1025             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1026             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1027             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1028             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1029             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1030             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1031             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1032
1033             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1034             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1035             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1036             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1037             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1038             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1039             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1040             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1041             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1042
1043             fjx1             = _mm_setzero_ps();
1044             fjy1             = _mm_setzero_ps();
1045             fjz1             = _mm_setzero_ps();
1046             fjx2             = _mm_setzero_ps();
1047             fjy2             = _mm_setzero_ps();
1048             fjz2             = _mm_setzero_ps();
1049             fjx3             = _mm_setzero_ps();
1050             fjy3             = _mm_setzero_ps();
1051             fjz3             = _mm_setzero_ps();
1052
1053             /**************************
1054              * CALCULATE INTERACTIONS *
1055              **************************/
1056
1057             /* COULOMB ELECTROSTATICS */
1058             velec            = _mm_mul_ps(qq11,rinv11);
1059             felec            = _mm_mul_ps(velec,rinvsq11);
1060
1061             fscal            = felec;
1062
1063              /* Update vectorial force */
1064             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1065             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1066             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1067
1068             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1069             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1070             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1071
1072             /**************************
1073              * CALCULATE INTERACTIONS *
1074              **************************/
1075
1076             /* COULOMB ELECTROSTATICS */
1077             velec            = _mm_mul_ps(qq12,rinv12);
1078             felec            = _mm_mul_ps(velec,rinvsq12);
1079
1080             fscal            = felec;
1081
1082              /* Update vectorial force */
1083             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1084             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1085             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1086
1087             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1088             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1089             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1090
1091             /**************************
1092              * CALCULATE INTERACTIONS *
1093              **************************/
1094
1095             /* COULOMB ELECTROSTATICS */
1096             velec            = _mm_mul_ps(qq13,rinv13);
1097             felec            = _mm_mul_ps(velec,rinvsq13);
1098
1099             fscal            = felec;
1100
1101              /* Update vectorial force */
1102             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1103             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1104             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1105
1106             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1107             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1108             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1109
1110             /**************************
1111              * CALCULATE INTERACTIONS *
1112              **************************/
1113
1114             /* COULOMB ELECTROSTATICS */
1115             velec            = _mm_mul_ps(qq21,rinv21);
1116             felec            = _mm_mul_ps(velec,rinvsq21);
1117
1118             fscal            = felec;
1119
1120              /* Update vectorial force */
1121             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1122             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1123             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1124
1125             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1126             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1127             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1128
1129             /**************************
1130              * CALCULATE INTERACTIONS *
1131              **************************/
1132
1133             /* COULOMB ELECTROSTATICS */
1134             velec            = _mm_mul_ps(qq22,rinv22);
1135             felec            = _mm_mul_ps(velec,rinvsq22);
1136
1137             fscal            = felec;
1138
1139              /* Update vectorial force */
1140             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1141             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1142             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1143
1144             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1145             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1146             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1147
1148             /**************************
1149              * CALCULATE INTERACTIONS *
1150              **************************/
1151
1152             /* COULOMB ELECTROSTATICS */
1153             velec            = _mm_mul_ps(qq23,rinv23);
1154             felec            = _mm_mul_ps(velec,rinvsq23);
1155
1156             fscal            = felec;
1157
1158              /* Update vectorial force */
1159             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1160             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1161             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1162
1163             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1164             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1165             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1166
1167             /**************************
1168              * CALCULATE INTERACTIONS *
1169              **************************/
1170
1171             /* COULOMB ELECTROSTATICS */
1172             velec            = _mm_mul_ps(qq31,rinv31);
1173             felec            = _mm_mul_ps(velec,rinvsq31);
1174
1175             fscal            = felec;
1176
1177              /* Update vectorial force */
1178             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1179             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1180             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1181
1182             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1183             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1184             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1185
1186             /**************************
1187              * CALCULATE INTERACTIONS *
1188              **************************/
1189
1190             /* COULOMB ELECTROSTATICS */
1191             velec            = _mm_mul_ps(qq32,rinv32);
1192             felec            = _mm_mul_ps(velec,rinvsq32);
1193
1194             fscal            = felec;
1195
1196              /* Update vectorial force */
1197             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1198             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1199             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1200
1201             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1202             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1203             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1204
1205             /**************************
1206              * CALCULATE INTERACTIONS *
1207              **************************/
1208
1209             /* COULOMB ELECTROSTATICS */
1210             velec            = _mm_mul_ps(qq33,rinv33);
1211             felec            = _mm_mul_ps(velec,rinvsq33);
1212
1213             fscal            = felec;
1214
1215              /* Update vectorial force */
1216             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1217             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1218             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1219
1220             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1221             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1222             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1223
1224             fjptrA             = f+j_coord_offsetA;
1225             fjptrB             = f+j_coord_offsetB;
1226             fjptrC             = f+j_coord_offsetC;
1227             fjptrD             = f+j_coord_offsetD;
1228
1229             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1230                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1231
1232             /* Inner loop uses 270 flops */
1233         }
1234
1235         if(jidx<j_index_end)
1236         {
1237
1238             /* Get j neighbor index, and coordinate index */
1239             jnrlistA         = jjnr[jidx];
1240             jnrlistB         = jjnr[jidx+1];
1241             jnrlistC         = jjnr[jidx+2];
1242             jnrlistD         = jjnr[jidx+3];
1243             /* Sign of each element will be negative for non-real atoms.
1244              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1245              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1246              */
1247             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1248             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1249             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1250             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1251             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1252             j_coord_offsetA  = DIM*jnrA;
1253             j_coord_offsetB  = DIM*jnrB;
1254             j_coord_offsetC  = DIM*jnrC;
1255             j_coord_offsetD  = DIM*jnrD;
1256
1257             /* load j atom coordinates */
1258             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1259                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1260                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1261
1262             /* Calculate displacement vector */
1263             dx11             = _mm_sub_ps(ix1,jx1);
1264             dy11             = _mm_sub_ps(iy1,jy1);
1265             dz11             = _mm_sub_ps(iz1,jz1);
1266             dx12             = _mm_sub_ps(ix1,jx2);
1267             dy12             = _mm_sub_ps(iy1,jy2);
1268             dz12             = _mm_sub_ps(iz1,jz2);
1269             dx13             = _mm_sub_ps(ix1,jx3);
1270             dy13             = _mm_sub_ps(iy1,jy3);
1271             dz13             = _mm_sub_ps(iz1,jz3);
1272             dx21             = _mm_sub_ps(ix2,jx1);
1273             dy21             = _mm_sub_ps(iy2,jy1);
1274             dz21             = _mm_sub_ps(iz2,jz1);
1275             dx22             = _mm_sub_ps(ix2,jx2);
1276             dy22             = _mm_sub_ps(iy2,jy2);
1277             dz22             = _mm_sub_ps(iz2,jz2);
1278             dx23             = _mm_sub_ps(ix2,jx3);
1279             dy23             = _mm_sub_ps(iy2,jy3);
1280             dz23             = _mm_sub_ps(iz2,jz3);
1281             dx31             = _mm_sub_ps(ix3,jx1);
1282             dy31             = _mm_sub_ps(iy3,jy1);
1283             dz31             = _mm_sub_ps(iz3,jz1);
1284             dx32             = _mm_sub_ps(ix3,jx2);
1285             dy32             = _mm_sub_ps(iy3,jy2);
1286             dz32             = _mm_sub_ps(iz3,jz2);
1287             dx33             = _mm_sub_ps(ix3,jx3);
1288             dy33             = _mm_sub_ps(iy3,jy3);
1289             dz33             = _mm_sub_ps(iz3,jz3);
1290
1291             /* Calculate squared distance and things based on it */
1292             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1293             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1294             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1295             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1296             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1297             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1298             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1299             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1300             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1301
1302             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1303             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1304             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1305             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1306             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1307             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1308             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1309             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1310             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1311
1312             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1313             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1314             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1315             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1316             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1317             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1318             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1319             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1320             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1321
1322             fjx1             = _mm_setzero_ps();
1323             fjy1             = _mm_setzero_ps();
1324             fjz1             = _mm_setzero_ps();
1325             fjx2             = _mm_setzero_ps();
1326             fjy2             = _mm_setzero_ps();
1327             fjz2             = _mm_setzero_ps();
1328             fjx3             = _mm_setzero_ps();
1329             fjy3             = _mm_setzero_ps();
1330             fjz3             = _mm_setzero_ps();
1331
1332             /**************************
1333              * CALCULATE INTERACTIONS *
1334              **************************/
1335
1336             /* COULOMB ELECTROSTATICS */
1337             velec            = _mm_mul_ps(qq11,rinv11);
1338             felec            = _mm_mul_ps(velec,rinvsq11);
1339
1340             fscal            = felec;
1341
1342             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1343
1344              /* Update vectorial force */
1345             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1346             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1347             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1348
1349             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1350             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1351             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1352
1353             /**************************
1354              * CALCULATE INTERACTIONS *
1355              **************************/
1356
1357             /* COULOMB ELECTROSTATICS */
1358             velec            = _mm_mul_ps(qq12,rinv12);
1359             felec            = _mm_mul_ps(velec,rinvsq12);
1360
1361             fscal            = felec;
1362
1363             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1364
1365              /* Update vectorial force */
1366             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1367             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1368             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1369
1370             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1371             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1372             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1373
1374             /**************************
1375              * CALCULATE INTERACTIONS *
1376              **************************/
1377
1378             /* COULOMB ELECTROSTATICS */
1379             velec            = _mm_mul_ps(qq13,rinv13);
1380             felec            = _mm_mul_ps(velec,rinvsq13);
1381
1382             fscal            = felec;
1383
1384             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1385
1386              /* Update vectorial force */
1387             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1388             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1389             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1390
1391             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1392             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1393             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1394
1395             /**************************
1396              * CALCULATE INTERACTIONS *
1397              **************************/
1398
1399             /* COULOMB ELECTROSTATICS */
1400             velec            = _mm_mul_ps(qq21,rinv21);
1401             felec            = _mm_mul_ps(velec,rinvsq21);
1402
1403             fscal            = felec;
1404
1405             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1406
1407              /* Update vectorial force */
1408             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1409             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1410             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1411
1412             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1413             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1414             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             /* COULOMB ELECTROSTATICS */
1421             velec            = _mm_mul_ps(qq22,rinv22);
1422             felec            = _mm_mul_ps(velec,rinvsq22);
1423
1424             fscal            = felec;
1425
1426             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1427
1428              /* Update vectorial force */
1429             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1430             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1431             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1432
1433             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1434             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1435             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1436
1437             /**************************
1438              * CALCULATE INTERACTIONS *
1439              **************************/
1440
1441             /* COULOMB ELECTROSTATICS */
1442             velec            = _mm_mul_ps(qq23,rinv23);
1443             felec            = _mm_mul_ps(velec,rinvsq23);
1444
1445             fscal            = felec;
1446
1447             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1448
1449              /* Update vectorial force */
1450             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1451             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1452             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1453
1454             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1455             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1456             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1457
1458             /**************************
1459              * CALCULATE INTERACTIONS *
1460              **************************/
1461
1462             /* COULOMB ELECTROSTATICS */
1463             velec            = _mm_mul_ps(qq31,rinv31);
1464             felec            = _mm_mul_ps(velec,rinvsq31);
1465
1466             fscal            = felec;
1467
1468             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1469
1470              /* Update vectorial force */
1471             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1472             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1473             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1474
1475             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1476             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1477             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1478
1479             /**************************
1480              * CALCULATE INTERACTIONS *
1481              **************************/
1482
1483             /* COULOMB ELECTROSTATICS */
1484             velec            = _mm_mul_ps(qq32,rinv32);
1485             felec            = _mm_mul_ps(velec,rinvsq32);
1486
1487             fscal            = felec;
1488
1489             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1490
1491              /* Update vectorial force */
1492             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1493             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1494             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1495
1496             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1497             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1498             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1499
1500             /**************************
1501              * CALCULATE INTERACTIONS *
1502              **************************/
1503
1504             /* COULOMB ELECTROSTATICS */
1505             velec            = _mm_mul_ps(qq33,rinv33);
1506             felec            = _mm_mul_ps(velec,rinvsq33);
1507
1508             fscal            = felec;
1509
1510             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1511
1512              /* Update vectorial force */
1513             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1514             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1515             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1516
1517             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1518             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1519             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1520
1521             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1522             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1523             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1524             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1525
1526             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1527                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1528
1529             /* Inner loop uses 270 flops */
1530         }
1531
1532         /* End of innermost loop */
1533
1534         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1535                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1536
1537         /* Increment number of inner iterations */
1538         inneriter                  += j_index_end - j_index_start;
1539
1540         /* Outer loop uses 18 flops */
1541     }
1542
1543     /* Increment number of outer iterations */
1544     outeriter        += nri;
1545
1546     /* Update outer/inner flops */
1547
1548     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1549 }