src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_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_ElecEwSw_VdwNone_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_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              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     __m128i          ewitab;
  77     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  78     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  79     real             *ewtab;
  80     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  81     real             rswitch_scalar,d_scalar;
  82     __m128           dummy_mask,cutoff_mask;
  83     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  84     __m128           one     = _mm_set1_ps(1.0);
  85     __m128           two     = _mm_set1_ps(2.0);
  86     x                = xx[0];
  87     f                = ff[0];
  88
  89     nri              = nlist->nri;
  90     iinr             = nlist->iinr;
  91     jindex           = nlist->jindex;
  92     jjnr             = nlist->jjnr;
  93     shiftidx         = nlist->shift;
  94     gid              = nlist->gid;
  95     shiftvec         = fr->shift_vec[0];
  96     fshift           = fr->fshift[0];
  97     facel            = _mm_set1_ps(fr->epsfac);
  98     charge           = mdatoms->chargeA;
  99
 100     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 101     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 102     beta2            = _mm_mul_ps(beta,beta);
 103     beta3            = _mm_mul_ps(beta,beta2);
 104     ewtab            = fr->ic->tabq_coul_FDV0;
 105     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 106     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 107
 108     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 109     rcutoff_scalar   = fr->rcoulomb;
 110     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 111     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 112
 113     rswitch_scalar   = fr->rcoulomb_switch;
 114     rswitch          = _mm_set1_ps(rswitch_scalar);
 115     /* Setup switch parameters */
 116     d_scalar         = rcutoff_scalar-rswitch_scalar;
 117     d                = _mm_set1_ps(d_scalar);
 118     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 119     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 120     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 121     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 122     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 123     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 124
 125     /* Avoid stupid compiler warnings */
 126     jnrA = jnrB = jnrC = jnrD = 0;
 127     j_coord_offsetA = 0;
 128     j_coord_offsetB = 0;
 129     j_coord_offsetC = 0;
 130     j_coord_offsetD = 0;
 131
 132     outeriter        = 0;
 133     inneriter        = 0;
 134
 135     for(iidx=0;iidx<4*DIM;iidx++)
 136     {
 137         scratch[iidx] = 0.0;
 138     }
 139
 140     /* Start outer loop over neighborlists */
 141     for(iidx=0; iidx<nri; iidx++)
 142     {
 143         /* Load shift vector for this list */
 144         i_shift_offset   = DIM*shiftidx[iidx];
 145
 146         /* Load limits for loop over neighbors */
 147         j_index_start    = jindex[iidx];
 148         j_index_end      = jindex[iidx+1];
 149
 150         /* Get outer coordinate index */
 151         inr              = iinr[iidx];
 152         i_coord_offset   = DIM*inr;
 153
 154         /* Load i particle coords and add shift vector */
 155         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 156
 157         fix0             = _mm_setzero_ps();
 158         fiy0             = _mm_setzero_ps();
 159         fiz0             = _mm_setzero_ps();
 160
 161         /* Load parameters for i particles */
 162         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_ps();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178             j_coord_offsetC  = DIM*jnrC;
 179             j_coord_offsetD  = DIM*jnrD;
 180
 181             /* load j atom coordinates */
 182             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 183                                               x+j_coord_offsetC,x+j_coord_offsetD,
 184                                               &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm_sub_ps(ix0,jx0);
 188             dy00             = _mm_sub_ps(iy0,jy0);
 189             dz00             = _mm_sub_ps(iz0,jz0);
 190
 191             /* Calculate squared distance and things based on it */
 192             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 193
 194             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 195
 196             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 197
 198             /* Load parameters for j particles */
 199             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 200                                                               charge+jnrC+0,charge+jnrD+0);
 201
 202             /**************************
 203              * CALCULATE INTERACTIONS *
 204              **************************/
 205
 206             if (gmx_mm_any_lt(rsq00,rcutoff2))
 207             {
 208
 209             r00              = _mm_mul_ps(rsq00,rinv00);
 210
 211             /* Compute parameters for interactions between i and j atoms */
 212             qq00             = _mm_mul_ps(iq0,jq0);
 213
 214             /* EWALD ELECTROSTATICS */
 215
 216             /* Analytical PME correction */
 217             zeta2            = _mm_mul_ps(beta2,rsq00);
 218             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 219             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 220             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 221             felec            = _mm_mul_ps(qq00,felec);
 222             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 223             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 224             velec            = _mm_mul_ps(qq00,velec);
 225
 226             d                = _mm_sub_ps(r00,rswitch);
 227             d                = _mm_max_ps(d,_mm_setzero_ps());
 228             d2               = _mm_mul_ps(d,d);
 229             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 230
 231             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 232
 233             /* Evaluate switch function */
 234             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 235             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 236             velec            = _mm_mul_ps(velec,sw);
 237             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 238
 239             /* Update potential sum for this i atom from the interaction with this j atom. */
 240             velec            = _mm_and_ps(velec,cutoff_mask);
 241             velecsum         = _mm_add_ps(velecsum,velec);
 242
 243             fscal            = felec;
 244
 245             fscal            = _mm_and_ps(fscal,cutoff_mask);
 246
 247              /* Update vectorial force */
 248             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 249             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 250             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 251
 252             fjptrA             = f+j_coord_offsetA;
 253             fjptrB             = f+j_coord_offsetB;
 254             fjptrC             = f+j_coord_offsetC;
 255             fjptrD             = f+j_coord_offsetD;
 256             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 257                                                    _mm_mul_ps(dx00,fscal),
 258                                                    _mm_mul_ps(dy00,fscal),
 259                                                    _mm_mul_ps(dz00,fscal));
 260
 261             }
 262
 263             /* Inner loop uses 53 flops */
 264         }
 265
 266         if(jidx<j_index_end)
 267         {
 268
 269             /* Get j neighbor index, and coordinate index */
 270             jnrlistA         = jjnr[jidx];
 271             jnrlistB         = jjnr[jidx+1];
 272             jnrlistC         = jjnr[jidx+2];
 273             jnrlistD         = jjnr[jidx+3];
 274             /* Sign of each element will be negative for non-real atoms.
 275              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 276              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 277              */
 278             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 279             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 280             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 281             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 282             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 283             j_coord_offsetA  = DIM*jnrA;
 284             j_coord_offsetB  = DIM*jnrB;
 285             j_coord_offsetC  = DIM*jnrC;
 286             j_coord_offsetD  = DIM*jnrD;
 287
 288             /* load j atom coordinates */
 289             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 290                                               x+j_coord_offsetC,x+j_coord_offsetD,
 291                                               &jx0,&jy0,&jz0);
 292
 293             /* Calculate displacement vector */
 294             dx00             = _mm_sub_ps(ix0,jx0);
 295             dy00             = _mm_sub_ps(iy0,jy0);
 296             dz00             = _mm_sub_ps(iz0,jz0);
 297
 298             /* Calculate squared distance and things based on it */
 299             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 300
 301             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 302
 303             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 304
 305             /* Load parameters for j particles */
 306             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 307                                                               charge+jnrC+0,charge+jnrD+0);
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             if (gmx_mm_any_lt(rsq00,rcutoff2))
 314             {
 315
 316             r00              = _mm_mul_ps(rsq00,rinv00);
 317             r00              = _mm_andnot_ps(dummy_mask,r00);
 318
 319             /* Compute parameters for interactions between i and j atoms */
 320             qq00             = _mm_mul_ps(iq0,jq0);
 321
 322             /* EWALD ELECTROSTATICS */
 323
 324             /* Analytical PME correction */
 325             zeta2            = _mm_mul_ps(beta2,rsq00);
 326             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 327             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 328             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 329             felec            = _mm_mul_ps(qq00,felec);
 330             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 331             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 332             velec            = _mm_mul_ps(qq00,velec);
 333
 334             d                = _mm_sub_ps(r00,rswitch);
 335             d                = _mm_max_ps(d,_mm_setzero_ps());
 336             d2               = _mm_mul_ps(d,d);
 337             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 338
 339             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 340
 341             /* Evaluate switch function */
 342             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 343             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 344             velec            = _mm_mul_ps(velec,sw);
 345             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velec            = _mm_and_ps(velec,cutoff_mask);
 349             velec            = _mm_andnot_ps(dummy_mask,velec);
 350             velecsum         = _mm_add_ps(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             fscal            = _mm_and_ps(fscal,cutoff_mask);
 355
 356             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 357
 358              /* Update vectorial force */
 359             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 360             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 361             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 362
 363             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 364             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 365             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 366             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 367             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 368                                                    _mm_mul_ps(dx00,fscal),
 369                                                    _mm_mul_ps(dy00,fscal),
 370                                                    _mm_mul_ps(dz00,fscal));
 371
 372             }
 373
 374             /* Inner loop uses 54 flops */
 375         }
 376
 377         /* End of innermost loop */
 378
 379         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 380                                               f+i_coord_offset,fshift+i_shift_offset);
 381
 382         ggid                        = gid[iidx];
 383         /* Update potential energies */
 384         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 385
 386         /* Increment number of inner iterations */
 387         inneriter                  += j_index_end - j_index_start;
 388
 389         /* Outer loop uses 8 flops */
 390     }
 391
 392     /* Increment number of outer iterations */
 393     outeriter        += nri;
 394
 395     /* Update outer/inner flops */
 396
 397     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*54);
 398 }
 399 /*
 400  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 401  * Electrostatics interaction: Ewald
 402  * VdW interaction:            None
 403  * Geometry:                   Particle-Particle
 404  * Calculate force/pot:        Force
 405  */
 406 void
 407 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 408                     (t_nblist * gmx_restrict                nlist,
 409                      rvec * gmx_restrict                    xx,
 410                      rvec * gmx_restrict                    ff,
 411                      t_forcerec * gmx_restrict              fr,
 412                      t_mdatoms * gmx_restrict               mdatoms,
 413                      nb_kernel_data_t * gmx_restrict        kernel_data,
 414                      t_nrnb * gmx_restrict                  nrnb)
 415 {
 416     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 417      * just 0 for non-waters.
 418      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 419      * jnr indices corresponding to data put in the four positions in the SIMD register.
 420      */
 421     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 422     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 423     int              jnrA,jnrB,jnrC,jnrD;
 424     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 425     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 426     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 427     real             rcutoff_scalar;
 428     real             *shiftvec,*fshift,*x,*f;
 429     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 430     real             scratch[4*DIM];
 431     __m128           fscal,rcutoff,rcutoff2,jidxall;
 432     int              vdwioffset0;
 433     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 434     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 435     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 436     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 437     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 438     real             *charge;
 439     __m128i          ewitab;
 440     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 441     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 442     real             *ewtab;
 443     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 444     real             rswitch_scalar,d_scalar;
 445     __m128           dummy_mask,cutoff_mask;
 446     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 447     __m128           one     = _mm_set1_ps(1.0);
 448     __m128           two     = _mm_set1_ps(2.0);
 449     x                = xx[0];
 450     f                = ff[0];
 451
 452     nri              = nlist->nri;
 453     iinr             = nlist->iinr;
 454     jindex           = nlist->jindex;
 455     jjnr             = nlist->jjnr;
 456     shiftidx         = nlist->shift;
 457     gid              = nlist->gid;
 458     shiftvec         = fr->shift_vec[0];
 459     fshift           = fr->fshift[0];
 460     facel            = _mm_set1_ps(fr->epsfac);
 461     charge           = mdatoms->chargeA;
 462
 463     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 464     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 465     beta2            = _mm_mul_ps(beta,beta);
 466     beta3            = _mm_mul_ps(beta,beta2);
 467     ewtab            = fr->ic->tabq_coul_FDV0;
 468     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 469     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 470
 471     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 472     rcutoff_scalar   = fr->rcoulomb;
 473     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 474     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 475
 476     rswitch_scalar   = fr->rcoulomb_switch;
 477     rswitch          = _mm_set1_ps(rswitch_scalar);
 478     /* Setup switch parameters */
 479     d_scalar         = rcutoff_scalar-rswitch_scalar;
 480     d                = _mm_set1_ps(d_scalar);
 481     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 482     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 483     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 484     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 485     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 486     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 487
 488     /* Avoid stupid compiler warnings */
 489     jnrA = jnrB = jnrC = jnrD = 0;
 490     j_coord_offsetA = 0;
 491     j_coord_offsetB = 0;
 492     j_coord_offsetC = 0;
 493     j_coord_offsetD = 0;
 494
 495     outeriter        = 0;
 496     inneriter        = 0;
 497
 498     for(iidx=0;iidx<4*DIM;iidx++)
 499     {
 500         scratch[iidx] = 0.0;
 501     }
 502
 503     /* Start outer loop over neighborlists */
 504     for(iidx=0; iidx<nri; iidx++)
 505     {
 506         /* Load shift vector for this list */
 507         i_shift_offset   = DIM*shiftidx[iidx];
 508
 509         /* Load limits for loop over neighbors */
 510         j_index_start    = jindex[iidx];
 511         j_index_end      = jindex[iidx+1];
 512
 513         /* Get outer coordinate index */
 514         inr              = iinr[iidx];
 515         i_coord_offset   = DIM*inr;
 516
 517         /* Load i particle coords and add shift vector */
 518         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 519
 520         fix0             = _mm_setzero_ps();
 521         fiy0             = _mm_setzero_ps();
 522         fiz0             = _mm_setzero_ps();
 523
 524         /* Load parameters for i particles */
 525         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 526
 527         /* Start inner kernel loop */
 528         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 529         {
 530
 531             /* Get j neighbor index, and coordinate index */
 532             jnrA             = jjnr[jidx];
 533             jnrB             = jjnr[jidx+1];
 534             jnrC             = jjnr[jidx+2];
 535             jnrD             = jjnr[jidx+3];
 536             j_coord_offsetA  = DIM*jnrA;
 537             j_coord_offsetB  = DIM*jnrB;
 538             j_coord_offsetC  = DIM*jnrC;
 539             j_coord_offsetD  = DIM*jnrD;
 540
 541             /* load j atom coordinates */
 542             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 543                                               x+j_coord_offsetC,x+j_coord_offsetD,
 544                                               &jx0,&jy0,&jz0);
 545
 546             /* Calculate displacement vector */
 547             dx00             = _mm_sub_ps(ix0,jx0);
 548             dy00             = _mm_sub_ps(iy0,jy0);
 549             dz00             = _mm_sub_ps(iz0,jz0);
 550
 551             /* Calculate squared distance and things based on it */
 552             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 553
 554             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 555
 556             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 557
 558             /* Load parameters for j particles */
 559             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 560                                                               charge+jnrC+0,charge+jnrD+0);
 561
 562             /**************************
 563              * CALCULATE INTERACTIONS *
 564              **************************/
 565
 566             if (gmx_mm_any_lt(rsq00,rcutoff2))
 567             {
 568
 569             r00              = _mm_mul_ps(rsq00,rinv00);
 570
 571             /* Compute parameters for interactions between i and j atoms */
 572             qq00             = _mm_mul_ps(iq0,jq0);
 573
 574             /* EWALD ELECTROSTATICS */
 575
 576             /* Analytical PME correction */
 577             zeta2            = _mm_mul_ps(beta2,rsq00);
 578             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 579             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 580             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 581             felec            = _mm_mul_ps(qq00,felec);
 582             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 583             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 584             velec            = _mm_mul_ps(qq00,velec);
 585
 586             d                = _mm_sub_ps(r00,rswitch);
 587             d                = _mm_max_ps(d,_mm_setzero_ps());
 588             d2               = _mm_mul_ps(d,d);
 589             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 590
 591             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 592
 593             /* Evaluate switch function */
 594             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 595             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 596             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 597
 598             fscal            = felec;
 599
 600             fscal            = _mm_and_ps(fscal,cutoff_mask);
 601
 602              /* Update vectorial force */
 603             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 604             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 605             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 606
 607             fjptrA             = f+j_coord_offsetA;
 608             fjptrB             = f+j_coord_offsetB;
 609             fjptrC             = f+j_coord_offsetC;
 610             fjptrD             = f+j_coord_offsetD;
 611             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 612                                                    _mm_mul_ps(dx00,fscal),
 613                                                    _mm_mul_ps(dy00,fscal),
 614                                                    _mm_mul_ps(dz00,fscal));
 615
 616             }
 617
 618             /* Inner loop uses 50 flops */
 619         }
 620
 621         if(jidx<j_index_end)
 622         {
 623
 624             /* Get j neighbor index, and coordinate index */
 625             jnrlistA         = jjnr[jidx];
 626             jnrlistB         = jjnr[jidx+1];
 627             jnrlistC         = jjnr[jidx+2];
 628             jnrlistD         = jjnr[jidx+3];
 629             /* Sign of each element will be negative for non-real atoms.
 630              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 631              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 632              */
 633             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 634             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 635             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 636             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 637             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 638             j_coord_offsetA  = DIM*jnrA;
 639             j_coord_offsetB  = DIM*jnrB;
 640             j_coord_offsetC  = DIM*jnrC;
 641             j_coord_offsetD  = DIM*jnrD;
 642
 643             /* load j atom coordinates */
 644             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 645                                               x+j_coord_offsetC,x+j_coord_offsetD,
 646                                               &jx0,&jy0,&jz0);
 647
 648             /* Calculate displacement vector */
 649             dx00             = _mm_sub_ps(ix0,jx0);
 650             dy00             = _mm_sub_ps(iy0,jy0);
 651             dz00             = _mm_sub_ps(iz0,jz0);
 652
 653             /* Calculate squared distance and things based on it */
 654             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 655
 656             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 657
 658             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 659
 660             /* Load parameters for j particles */
 661             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 662                                                               charge+jnrC+0,charge+jnrD+0);
 663
 664             /**************************
 665              * CALCULATE INTERACTIONS *
 666              **************************/
 667
 668             if (gmx_mm_any_lt(rsq00,rcutoff2))
 669             {
 670
 671             r00              = _mm_mul_ps(rsq00,rinv00);
 672             r00              = _mm_andnot_ps(dummy_mask,r00);
 673
 674             /* Compute parameters for interactions between i and j atoms */
 675             qq00             = _mm_mul_ps(iq0,jq0);
 676
 677             /* EWALD ELECTROSTATICS */
 678
 679             /* Analytical PME correction */
 680             zeta2            = _mm_mul_ps(beta2,rsq00);
 681             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 682             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 683             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 684             felec            = _mm_mul_ps(qq00,felec);
 685             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 686             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 687             velec            = _mm_mul_ps(qq00,velec);
 688
 689             d                = _mm_sub_ps(r00,rswitch);
 690             d                = _mm_max_ps(d,_mm_setzero_ps());
 691             d2               = _mm_mul_ps(d,d);
 692             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 693
 694             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 695
 696             /* Evaluate switch function */
 697             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 698             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 699             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 700
 701             fscal            = felec;
 702
 703             fscal            = _mm_and_ps(fscal,cutoff_mask);
 704
 705             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 706
 707              /* Update vectorial force */
 708             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 709             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 710             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 711
 712             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 713             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 714             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 715             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 716             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 717                                                    _mm_mul_ps(dx00,fscal),
 718                                                    _mm_mul_ps(dy00,fscal),
 719                                                    _mm_mul_ps(dz00,fscal));
 720
 721             }
 722
 723             /* Inner loop uses 51 flops */
 724         }
 725
 726         /* End of innermost loop */
 727
 728         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 729                                               f+i_coord_offset,fshift+i_shift_offset);
 730
 731         /* Increment number of inner iterations */
 732         inneriter                  += j_index_end - j_index_start;
 733
 734         /* Outer loop uses 7 flops */
 735     }
 736
 737     /* Increment number of outer iterations */
 738     outeriter        += nri;
 739
 740     /* Update outer/inner flops */
 741
 742     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*51);
 743 }