src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_128_fma_double
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            None
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwjidx0A,vdwjidx0B;
  83     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  84     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  85     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  86     real             *charge;
  87     __m128i          ewitab;
  88     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  89     real             *ewtab;
  90     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  91     real             rswitch_scalar,d_scalar;
  92     __m128d          dummy_mask,cutoff_mask;
  93     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  94     __m128d          one     = _mm_set1_pd(1.0);
  95     __m128d          two     = _mm_set1_pd(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_pd(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109
 110     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 111     ewtab            = fr->ic->tabq_coul_FDV0;
 112     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 113     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 114
 115     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 116     rcutoff_scalar   = fr->rcoulomb;
 117     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 118     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 119
 120     rswitch_scalar   = fr->rcoulomb_switch;
 121     rswitch          = _mm_set1_pd(rswitch_scalar);
 122     /* Setup switch parameters */
 123     d_scalar         = rcutoff_scalar-rswitch_scalar;
 124     d                = _mm_set1_pd(d_scalar);
 125     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 126     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 127     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 128     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 129     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 130     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 131
 132     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136
 137     outeriter        = 0;
 138     inneriter        = 0;
 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_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 156
 157         fix0             = _mm_setzero_pd();
 158         fiy0             = _mm_setzero_pd();
 159         fiz0             = _mm_setzero_pd();
 160
 161         /* Load parameters for i particles */
 162         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_pd();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             j_coord_offsetA  = DIM*jnrA;
 175             j_coord_offsetB  = DIM*jnrB;
 176
 177             /* load j atom coordinates */
 178             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 179                                               &jx0,&jy0,&jz0);
 180
 181             /* Calculate displacement vector */
 182             dx00             = _mm_sub_pd(ix0,jx0);
 183             dy00             = _mm_sub_pd(iy0,jy0);
 184             dz00             = _mm_sub_pd(iz0,jz0);
 185
 186             /* Calculate squared distance and things based on it */
 187             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 188
 189             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 190
 191             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 192
 193             /* Load parameters for j particles */
 194             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 195
 196             /**************************
 197              * CALCULATE INTERACTIONS *
 198              **************************/
 199
 200             if (gmx_mm_any_lt(rsq00,rcutoff2))
 201             {
 202
 203             r00              = _mm_mul_pd(rsq00,rinv00);
 204
 205             /* Compute parameters for interactions between i and j atoms */
 206             qq00             = _mm_mul_pd(iq0,jq0);
 207
 208             /* EWALD ELECTROSTATICS */
 209
 210             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 211             ewrt             = _mm_mul_pd(r00,ewtabscale);
 212             ewitab           = _mm_cvttpd_epi32(ewrt);
 213 #ifdef __XOP__
 214             eweps            = _mm_frcz_pd(ewrt);
 215 #else
 216             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 217 #endif
 218             twoeweps         = _mm_add_pd(eweps,eweps);
 219             ewitab           = _mm_slli_epi32(ewitab,2);
 220             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 221             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 222             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 223             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 224             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 225             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 226             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 227             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 228             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 229             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 230
 231             d                = _mm_sub_pd(r00,rswitch);
 232             d                = _mm_max_pd(d,_mm_setzero_pd());
 233             d2               = _mm_mul_pd(d,d);
 234             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 235
 236             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 237
 238             /* Evaluate switch function */
 239             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 240             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 241             velec            = _mm_mul_pd(velec,sw);
 242             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 243
 244             /* Update potential sum for this i atom from the interaction with this j atom. */
 245             velec            = _mm_and_pd(velec,cutoff_mask);
 246             velecsum         = _mm_add_pd(velecsum,velec);
 247
 248             fscal            = felec;
 249
 250             fscal            = _mm_and_pd(fscal,cutoff_mask);
 251
 252             /* Update vectorial force */
 253             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 254             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 255             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 256
 257             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 258                                                    _mm_mul_pd(dx00,fscal),
 259                                                    _mm_mul_pd(dy00,fscal),
 260                                                    _mm_mul_pd(dz00,fscal));
 261
 262             }
 263
 264             /* Inner loop uses 68 flops */
 265         }
 266
 267         if(jidx<j_index_end)
 268         {
 269
 270             jnrA             = jjnr[jidx];
 271             j_coord_offsetA  = DIM*jnrA;
 272
 273             /* load j atom coordinates */
 274             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 275                                               &jx0,&jy0,&jz0);
 276
 277             /* Calculate displacement vector */
 278             dx00             = _mm_sub_pd(ix0,jx0);
 279             dy00             = _mm_sub_pd(iy0,jy0);
 280             dz00             = _mm_sub_pd(iz0,jz0);
 281
 282             /* Calculate squared distance and things based on it */
 283             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 284
 285             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 286
 287             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 288
 289             /* Load parameters for j particles */
 290             jq0              = _mm_load_sd(charge+jnrA+0);
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             if (gmx_mm_any_lt(rsq00,rcutoff2))
 297             {
 298
 299             r00              = _mm_mul_pd(rsq00,rinv00);
 300
 301             /* Compute parameters for interactions between i and j atoms */
 302             qq00             = _mm_mul_pd(iq0,jq0);
 303
 304             /* EWALD ELECTROSTATICS */
 305
 306             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 307             ewrt             = _mm_mul_pd(r00,ewtabscale);
 308             ewitab           = _mm_cvttpd_epi32(ewrt);
 309 #ifdef __XOP__
 310             eweps            = _mm_frcz_pd(ewrt);
 311 #else
 312             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 313 #endif
 314             twoeweps         = _mm_add_pd(eweps,eweps);
 315             ewitab           = _mm_slli_epi32(ewitab,2);
 316             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 317             ewtabD           = _mm_setzero_pd();
 318             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 319             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 320             ewtabFn          = _mm_setzero_pd();
 321             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 322             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 323             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 324             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 325             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 326
 327             d                = _mm_sub_pd(r00,rswitch);
 328             d                = _mm_max_pd(d,_mm_setzero_pd());
 329             d2               = _mm_mul_pd(d,d);
 330             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 331
 332             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 333
 334             /* Evaluate switch function */
 335             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 336             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 337             velec            = _mm_mul_pd(velec,sw);
 338             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 339
 340             /* Update potential sum for this i atom from the interaction with this j atom. */
 341             velec            = _mm_and_pd(velec,cutoff_mask);
 342             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 343             velecsum         = _mm_add_pd(velecsum,velec);
 344
 345             fscal            = felec;
 346
 347             fscal            = _mm_and_pd(fscal,cutoff_mask);
 348
 349             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 350
 351             /* Update vectorial force */
 352             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 353             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 354             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 355
 356             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 357                                                    _mm_mul_pd(dx00,fscal),
 358                                                    _mm_mul_pd(dy00,fscal),
 359                                                    _mm_mul_pd(dz00,fscal));
 360
 361             }
 362
 363             /* Inner loop uses 68 flops */
 364         }
 365
 366         /* End of innermost loop */
 367
 368         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 369                                               f+i_coord_offset,fshift+i_shift_offset);
 370
 371         ggid                        = gid[iidx];
 372         /* Update potential energies */
 373         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 374
 375         /* Increment number of inner iterations */
 376         inneriter                  += j_index_end - j_index_start;
 377
 378         /* Outer loop uses 8 flops */
 379     }
 380
 381     /* Increment number of outer iterations */
 382     outeriter        += nri;
 383
 384     /* Update outer/inner flops */
 385
 386     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*68);
 387 }
 388 /*
 389  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_double
 390  * Electrostatics interaction: Ewald
 391  * VdW interaction:            None
 392  * Geometry:                   Particle-Particle
 393  * Calculate force/pot:        Force
 394  */
 395 void
 396 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_double
 397                     (t_nblist                    * gmx_restrict       nlist,
 398                      rvec                        * gmx_restrict          xx,
 399                      rvec                        * gmx_restrict          ff,
 400                      t_forcerec                  * gmx_restrict          fr,
 401                      t_mdatoms                   * gmx_restrict     mdatoms,
 402                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 403                      t_nrnb                      * gmx_restrict        nrnb)
 404 {
 405     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 406      * just 0 for non-waters.
 407      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 408      * jnr indices corresponding to data put in the four positions in the SIMD register.
 409      */
 410     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 411     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 412     int              jnrA,jnrB;
 413     int              j_coord_offsetA,j_coord_offsetB;
 414     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 415     real             rcutoff_scalar;
 416     real             *shiftvec,*fshift,*x,*f;
 417     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 418     int              vdwioffset0;
 419     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 420     int              vdwjidx0A,vdwjidx0B;
 421     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 422     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 423     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 424     real             *charge;
 425     __m128i          ewitab;
 426     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 427     real             *ewtab;
 428     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 429     real             rswitch_scalar,d_scalar;
 430     __m128d          dummy_mask,cutoff_mask;
 431     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 432     __m128d          one     = _mm_set1_pd(1.0);
 433     __m128d          two     = _mm_set1_pd(2.0);
 434     x                = xx[0];
 435     f                = ff[0];
 436
 437     nri              = nlist->nri;
 438     iinr             = nlist->iinr;
 439     jindex           = nlist->jindex;
 440     jjnr             = nlist->jjnr;
 441     shiftidx         = nlist->shift;
 442     gid              = nlist->gid;
 443     shiftvec         = fr->shift_vec[0];
 444     fshift           = fr->fshift[0];
 445     facel            = _mm_set1_pd(fr->epsfac);
 446     charge           = mdatoms->chargeA;
 447
 448     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 449     ewtab            = fr->ic->tabq_coul_FDV0;
 450     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 451     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 452
 453     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 454     rcutoff_scalar   = fr->rcoulomb;
 455     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 456     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 457
 458     rswitch_scalar   = fr->rcoulomb_switch;
 459     rswitch          = _mm_set1_pd(rswitch_scalar);
 460     /* Setup switch parameters */
 461     d_scalar         = rcutoff_scalar-rswitch_scalar;
 462     d                = _mm_set1_pd(d_scalar);
 463     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 464     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 465     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 466     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 467     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 468     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 469
 470     /* Avoid stupid compiler warnings */
 471     jnrA = jnrB = 0;
 472     j_coord_offsetA = 0;
 473     j_coord_offsetB = 0;
 474
 475     outeriter        = 0;
 476     inneriter        = 0;
 477
 478     /* Start outer loop over neighborlists */
 479     for(iidx=0; iidx<nri; iidx++)
 480     {
 481         /* Load shift vector for this list */
 482         i_shift_offset   = DIM*shiftidx[iidx];
 483
 484         /* Load limits for loop over neighbors */
 485         j_index_start    = jindex[iidx];
 486         j_index_end      = jindex[iidx+1];
 487
 488         /* Get outer coordinate index */
 489         inr              = iinr[iidx];
 490         i_coord_offset   = DIM*inr;
 491
 492         /* Load i particle coords and add shift vector */
 493         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 494
 495         fix0             = _mm_setzero_pd();
 496         fiy0             = _mm_setzero_pd();
 497         fiz0             = _mm_setzero_pd();
 498
 499         /* Load parameters for i particles */
 500         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 501
 502         /* Start inner kernel loop */
 503         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 504         {
 505
 506             /* Get j neighbor index, and coordinate index */
 507             jnrA             = jjnr[jidx];
 508             jnrB             = jjnr[jidx+1];
 509             j_coord_offsetA  = DIM*jnrA;
 510             j_coord_offsetB  = DIM*jnrB;
 511
 512             /* load j atom coordinates */
 513             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 514                                               &jx0,&jy0,&jz0);
 515
 516             /* Calculate displacement vector */
 517             dx00             = _mm_sub_pd(ix0,jx0);
 518             dy00             = _mm_sub_pd(iy0,jy0);
 519             dz00             = _mm_sub_pd(iz0,jz0);
 520
 521             /* Calculate squared distance and things based on it */
 522             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 523
 524             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 525
 526             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 527
 528             /* Load parameters for j particles */
 529             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             if (gmx_mm_any_lt(rsq00,rcutoff2))
 536             {
 537
 538             r00              = _mm_mul_pd(rsq00,rinv00);
 539
 540             /* Compute parameters for interactions between i and j atoms */
 541             qq00             = _mm_mul_pd(iq0,jq0);
 542
 543             /* EWALD ELECTROSTATICS */
 544
 545             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 546             ewrt             = _mm_mul_pd(r00,ewtabscale);
 547             ewitab           = _mm_cvttpd_epi32(ewrt);
 548 #ifdef __XOP__
 549             eweps            = _mm_frcz_pd(ewrt);
 550 #else
 551             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 552 #endif
 553             twoeweps         = _mm_add_pd(eweps,eweps);
 554             ewitab           = _mm_slli_epi32(ewitab,2);
 555             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 556             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 557             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 558             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 559             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 560             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 561             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 562             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 563             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 564             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 565
 566             d                = _mm_sub_pd(r00,rswitch);
 567             d                = _mm_max_pd(d,_mm_setzero_pd());
 568             d2               = _mm_mul_pd(d,d);
 569             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 570
 571             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 572
 573             /* Evaluate switch function */
 574             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 575             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 576             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 577
 578             fscal            = felec;
 579
 580             fscal            = _mm_and_pd(fscal,cutoff_mask);
 581
 582             /* Update vectorial force */
 583             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 584             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 585             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 586
 587             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 588                                                    _mm_mul_pd(dx00,fscal),
 589                                                    _mm_mul_pd(dy00,fscal),
 590                                                    _mm_mul_pd(dz00,fscal));
 591
 592             }
 593
 594             /* Inner loop uses 65 flops */
 595         }
 596
 597         if(jidx<j_index_end)
 598         {
 599
 600             jnrA             = jjnr[jidx];
 601             j_coord_offsetA  = DIM*jnrA;
 602
 603             /* load j atom coordinates */
 604             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 605                                               &jx0,&jy0,&jz0);
 606
 607             /* Calculate displacement vector */
 608             dx00             = _mm_sub_pd(ix0,jx0);
 609             dy00             = _mm_sub_pd(iy0,jy0);
 610             dz00             = _mm_sub_pd(iz0,jz0);
 611
 612             /* Calculate squared distance and things based on it */
 613             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 614
 615             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 616
 617             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 618
 619             /* Load parameters for j particles */
 620             jq0              = _mm_load_sd(charge+jnrA+0);
 621
 622             /**************************
 623              * CALCULATE INTERACTIONS *
 624              **************************/
 625
 626             if (gmx_mm_any_lt(rsq00,rcutoff2))
 627             {
 628
 629             r00              = _mm_mul_pd(rsq00,rinv00);
 630
 631             /* Compute parameters for interactions between i and j atoms */
 632             qq00             = _mm_mul_pd(iq0,jq0);
 633
 634             /* EWALD ELECTROSTATICS */
 635
 636             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 637             ewrt             = _mm_mul_pd(r00,ewtabscale);
 638             ewitab           = _mm_cvttpd_epi32(ewrt);
 639 #ifdef __XOP__
 640             eweps            = _mm_frcz_pd(ewrt);
 641 #else
 642             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 643 #endif
 644             twoeweps         = _mm_add_pd(eweps,eweps);
 645             ewitab           = _mm_slli_epi32(ewitab,2);
 646             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 647             ewtabD           = _mm_setzero_pd();
 648             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 649             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 650             ewtabFn          = _mm_setzero_pd();
 651             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 652             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 653             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 654             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 655             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 656
 657             d                = _mm_sub_pd(r00,rswitch);
 658             d                = _mm_max_pd(d,_mm_setzero_pd());
 659             d2               = _mm_mul_pd(d,d);
 660             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 661
 662             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 663
 664             /* Evaluate switch function */
 665             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 666             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 667             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 668
 669             fscal            = felec;
 670
 671             fscal            = _mm_and_pd(fscal,cutoff_mask);
 672
 673             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 674
 675             /* Update vectorial force */
 676             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 677             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 678             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 679
 680             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 681                                                    _mm_mul_pd(dx00,fscal),
 682                                                    _mm_mul_pd(dy00,fscal),
 683                                                    _mm_mul_pd(dz00,fscal));
 684
 685             }
 686
 687             /* Inner loop uses 65 flops */
 688         }
 689
 690         /* End of innermost loop */
 691
 692         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 693                                               f+i_coord_offset,fshift+i_shift_offset);
 694
 695         /* Increment number of inner iterations */
 696         inneriter                  += j_index_end - j_index_start;
 697
 698         /* Outer loop uses 7 flops */
 699     }
 700
 701     /* Increment number of outer iterations */
 702     outeriter        += nri;
 703
 704     /* Update outer/inner flops */
 705
 706     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*65);
 707 }