src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "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_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_128_fma_single
  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_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     __m128i          ewitab;
  91     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  92     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  93     real             *ewtab;
  94     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  95     real             rswitch_scalar,d_scalar;
  96     __m128           dummy_mask,cutoff_mask;
  97     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  98     __m128           one     = _mm_set1_ps(1.0);
  99     __m128           two     = _mm_set1_ps(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = _mm_set1_ps(fr->epsfac);
 112     charge           = mdatoms->chargeA;
 113
 114     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 115     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 116     beta2            = _mm_mul_ps(beta,beta);
 117     beta3            = _mm_mul_ps(beta,beta2);
 118     ewtab            = fr->ic->tabq_coul_FDV0;
 119     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 120     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 121
 122     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 123     rcutoff_scalar   = fr->rcoulomb;
 124     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 125     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 126
 127     rswitch_scalar   = fr->rcoulomb_switch;
 128     rswitch          = _mm_set1_ps(rswitch_scalar);
 129     /* Setup switch parameters */
 130     d_scalar         = rcutoff_scalar-rswitch_scalar;
 131     d                = _mm_set1_ps(d_scalar);
 132     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 133     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 134     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 135     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 136     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 137     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = jnrC = jnrD = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143     j_coord_offsetC = 0;
 144     j_coord_offsetD = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     for(iidx=0;iidx<4*DIM;iidx++)
 150     {
 151         scratch[iidx] = 0.0;
 152     }
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 170
 171         fix0             = _mm_setzero_ps();
 172         fiy0             = _mm_setzero_ps();
 173         fiz0             = _mm_setzero_ps();
 174
 175         /* Load parameters for i particles */
 176         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 177
 178         /* Reset potential sums */
 179         velecsum         = _mm_setzero_ps();
 180
 181         /* Start inner kernel loop */
 182         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 183         {
 184
 185             /* Get j neighbor index, and coordinate index */
 186             jnrA             = jjnr[jidx];
 187             jnrB             = jjnr[jidx+1];
 188             jnrC             = jjnr[jidx+2];
 189             jnrD             = jjnr[jidx+3];
 190             j_coord_offsetA  = DIM*jnrA;
 191             j_coord_offsetB  = DIM*jnrB;
 192             j_coord_offsetC  = DIM*jnrC;
 193             j_coord_offsetD  = DIM*jnrD;
 194
 195             /* load j atom coordinates */
 196             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 197                                               x+j_coord_offsetC,x+j_coord_offsetD,
 198                                               &jx0,&jy0,&jz0);
 199
 200             /* Calculate displacement vector */
 201             dx00             = _mm_sub_ps(ix0,jx0);
 202             dy00             = _mm_sub_ps(iy0,jy0);
 203             dz00             = _mm_sub_ps(iz0,jz0);
 204
 205             /* Calculate squared distance and things based on it */
 206             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 207
 208             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 209
 210             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 211
 212             /* Load parameters for j particles */
 213             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 214                                                               charge+jnrC+0,charge+jnrD+0);
 215
 216             /**************************
 217              * CALCULATE INTERACTIONS *
 218              **************************/
 219
 220             if (gmx_mm_any_lt(rsq00,rcutoff2))
 221             {
 222
 223             r00              = _mm_mul_ps(rsq00,rinv00);
 224
 225             /* Compute parameters for interactions between i and j atoms */
 226             qq00             = _mm_mul_ps(iq0,jq0);
 227
 228             /* EWALD ELECTROSTATICS */
 229
 230             /* Analytical PME correction */
 231             zeta2            = _mm_mul_ps(beta2,rsq00);
 232             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 233             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 234             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 235             felec            = _mm_mul_ps(qq00,felec);
 236             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 237             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 238             velec            = _mm_mul_ps(qq00,velec);
 239
 240             d                = _mm_sub_ps(r00,rswitch);
 241             d                = _mm_max_ps(d,_mm_setzero_ps());
 242             d2               = _mm_mul_ps(d,d);
 243             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 244
 245             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 246
 247             /* Evaluate switch function */
 248             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 249             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 250             velec            = _mm_mul_ps(velec,sw);
 251             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 252
 253             /* Update potential sum for this i atom from the interaction with this j atom. */
 254             velec            = _mm_and_ps(velec,cutoff_mask);
 255             velecsum         = _mm_add_ps(velecsum,velec);
 256
 257             fscal            = felec;
 258
 259             fscal            = _mm_and_ps(fscal,cutoff_mask);
 260
 261              /* Update vectorial force */
 262             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 263             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 264             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 265
 266             fjptrA             = f+j_coord_offsetA;
 267             fjptrB             = f+j_coord_offsetB;
 268             fjptrC             = f+j_coord_offsetC;
 269             fjptrD             = f+j_coord_offsetD;
 270             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 271                                                    _mm_mul_ps(dx00,fscal),
 272                                                    _mm_mul_ps(dy00,fscal),
 273                                                    _mm_mul_ps(dz00,fscal));
 274
 275             }
 276
 277             /* Inner loop uses 53 flops */
 278         }
 279
 280         if(jidx<j_index_end)
 281         {
 282
 283             /* Get j neighbor index, and coordinate index */
 284             jnrlistA         = jjnr[jidx];
 285             jnrlistB         = jjnr[jidx+1];
 286             jnrlistC         = jjnr[jidx+2];
 287             jnrlistD         = jjnr[jidx+3];
 288             /* Sign of each element will be negative for non-real atoms.
 289              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 290              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 291              */
 292             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 293             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 294             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 295             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 296             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 297             j_coord_offsetA  = DIM*jnrA;
 298             j_coord_offsetB  = DIM*jnrB;
 299             j_coord_offsetC  = DIM*jnrC;
 300             j_coord_offsetD  = DIM*jnrD;
 301
 302             /* load j atom coordinates */
 303             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 304                                               x+j_coord_offsetC,x+j_coord_offsetD,
 305                                               &jx0,&jy0,&jz0);
 306
 307             /* Calculate displacement vector */
 308             dx00             = _mm_sub_ps(ix0,jx0);
 309             dy00             = _mm_sub_ps(iy0,jy0);
 310             dz00             = _mm_sub_ps(iz0,jz0);
 311
 312             /* Calculate squared distance and things based on it */
 313             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 314
 315             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 316
 317             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 318
 319             /* Load parameters for j particles */
 320             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 321                                                               charge+jnrC+0,charge+jnrD+0);
 322
 323             /**************************
 324              * CALCULATE INTERACTIONS *
 325              **************************/
 326
 327             if (gmx_mm_any_lt(rsq00,rcutoff2))
 328             {
 329
 330             r00              = _mm_mul_ps(rsq00,rinv00);
 331             r00              = _mm_andnot_ps(dummy_mask,r00);
 332
 333             /* Compute parameters for interactions between i and j atoms */
 334             qq00             = _mm_mul_ps(iq0,jq0);
 335
 336             /* EWALD ELECTROSTATICS */
 337
 338             /* Analytical PME correction */
 339             zeta2            = _mm_mul_ps(beta2,rsq00);
 340             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 341             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 342             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 343             felec            = _mm_mul_ps(qq00,felec);
 344             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 345             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 346             velec            = _mm_mul_ps(qq00,velec);
 347
 348             d                = _mm_sub_ps(r00,rswitch);
 349             d                = _mm_max_ps(d,_mm_setzero_ps());
 350             d2               = _mm_mul_ps(d,d);
 351             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 352
 353             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 354
 355             /* Evaluate switch function */
 356             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 357             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 358             velec            = _mm_mul_ps(velec,sw);
 359             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 360
 361             /* Update potential sum for this i atom from the interaction with this j atom. */
 362             velec            = _mm_and_ps(velec,cutoff_mask);
 363             velec            = _mm_andnot_ps(dummy_mask,velec);
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368             fscal            = _mm_and_ps(fscal,cutoff_mask);
 369
 370             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 371
 372              /* Update vectorial force */
 373             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 374             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 375             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 376
 377             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 378             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 379             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 380             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 381             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 382                                                    _mm_mul_ps(dx00,fscal),
 383                                                    _mm_mul_ps(dy00,fscal),
 384                                                    _mm_mul_ps(dz00,fscal));
 385
 386             }
 387
 388             /* Inner loop uses 54 flops */
 389         }
 390
 391         /* End of innermost loop */
 392
 393         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 394                                               f+i_coord_offset,fshift+i_shift_offset);
 395
 396         ggid                        = gid[iidx];
 397         /* Update potential energies */
 398         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 399
 400         /* Increment number of inner iterations */
 401         inneriter                  += j_index_end - j_index_start;
 402
 403         /* Outer loop uses 8 flops */
 404     }
 405
 406     /* Increment number of outer iterations */
 407     outeriter        += nri;
 408
 409     /* Update outer/inner flops */
 410
 411     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*54);
 412 }
 413 /*
 414  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 415  * Electrostatics interaction: Ewald
 416  * VdW interaction:            None
 417  * Geometry:                   Particle-Particle
 418  * Calculate force/pot:        Force
 419  */
 420 void
 421 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 422                     (t_nblist                    * gmx_restrict       nlist,
 423                      rvec                        * gmx_restrict          xx,
 424                      rvec                        * gmx_restrict          ff,
 425                      t_forcerec                  * gmx_restrict          fr,
 426                      t_mdatoms                   * gmx_restrict     mdatoms,
 427                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 428                      t_nrnb                      * gmx_restrict        nrnb)
 429 {
 430     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 431      * just 0 for non-waters.
 432      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 433      * jnr indices corresponding to data put in the four positions in the SIMD register.
 434      */
 435     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 436     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 437     int              jnrA,jnrB,jnrC,jnrD;
 438     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 439     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 440     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 441     real             rcutoff_scalar;
 442     real             *shiftvec,*fshift,*x,*f;
 443     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 444     real             scratch[4*DIM];
 445     __m128           fscal,rcutoff,rcutoff2,jidxall;
 446     int              vdwioffset0;
 447     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 448     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 449     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 450     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 451     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 452     real             *charge;
 453     __m128i          ewitab;
 454     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 455     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 456     real             *ewtab;
 457     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 458     real             rswitch_scalar,d_scalar;
 459     __m128           dummy_mask,cutoff_mask;
 460     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 461     __m128           one     = _mm_set1_ps(1.0);
 462     __m128           two     = _mm_set1_ps(2.0);
 463     x                = xx[0];
 464     f                = ff[0];
 465
 466     nri              = nlist->nri;
 467     iinr             = nlist->iinr;
 468     jindex           = nlist->jindex;
 469     jjnr             = nlist->jjnr;
 470     shiftidx         = nlist->shift;
 471     gid              = nlist->gid;
 472     shiftvec         = fr->shift_vec[0];
 473     fshift           = fr->fshift[0];
 474     facel            = _mm_set1_ps(fr->epsfac);
 475     charge           = mdatoms->chargeA;
 476
 477     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 478     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 479     beta2            = _mm_mul_ps(beta,beta);
 480     beta3            = _mm_mul_ps(beta,beta2);
 481     ewtab            = fr->ic->tabq_coul_FDV0;
 482     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 483     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 484
 485     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 486     rcutoff_scalar   = fr->rcoulomb;
 487     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 488     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 489
 490     rswitch_scalar   = fr->rcoulomb_switch;
 491     rswitch          = _mm_set1_ps(rswitch_scalar);
 492     /* Setup switch parameters */
 493     d_scalar         = rcutoff_scalar-rswitch_scalar;
 494     d                = _mm_set1_ps(d_scalar);
 495     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 496     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 497     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 498     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 499     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 500     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 501
 502     /* Avoid stupid compiler warnings */
 503     jnrA = jnrB = jnrC = jnrD = 0;
 504     j_coord_offsetA = 0;
 505     j_coord_offsetB = 0;
 506     j_coord_offsetC = 0;
 507     j_coord_offsetD = 0;
 508
 509     outeriter        = 0;
 510     inneriter        = 0;
 511
 512     for(iidx=0;iidx<4*DIM;iidx++)
 513     {
 514         scratch[iidx] = 0.0;
 515     }
 516
 517     /* Start outer loop over neighborlists */
 518     for(iidx=0; iidx<nri; iidx++)
 519     {
 520         /* Load shift vector for this list */
 521         i_shift_offset   = DIM*shiftidx[iidx];
 522
 523         /* Load limits for loop over neighbors */
 524         j_index_start    = jindex[iidx];
 525         j_index_end      = jindex[iidx+1];
 526
 527         /* Get outer coordinate index */
 528         inr              = iinr[iidx];
 529         i_coord_offset   = DIM*inr;
 530
 531         /* Load i particle coords and add shift vector */
 532         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 533
 534         fix0             = _mm_setzero_ps();
 535         fiy0             = _mm_setzero_ps();
 536         fiz0             = _mm_setzero_ps();
 537
 538         /* Load parameters for i particles */
 539         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 540
 541         /* Start inner kernel loop */
 542         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 543         {
 544
 545             /* Get j neighbor index, and coordinate index */
 546             jnrA             = jjnr[jidx];
 547             jnrB             = jjnr[jidx+1];
 548             jnrC             = jjnr[jidx+2];
 549             jnrD             = jjnr[jidx+3];
 550             j_coord_offsetA  = DIM*jnrA;
 551             j_coord_offsetB  = DIM*jnrB;
 552             j_coord_offsetC  = DIM*jnrC;
 553             j_coord_offsetD  = DIM*jnrD;
 554
 555             /* load j atom coordinates */
 556             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 557                                               x+j_coord_offsetC,x+j_coord_offsetD,
 558                                               &jx0,&jy0,&jz0);
 559
 560             /* Calculate displacement vector */
 561             dx00             = _mm_sub_ps(ix0,jx0);
 562             dy00             = _mm_sub_ps(iy0,jy0);
 563             dz00             = _mm_sub_ps(iz0,jz0);
 564
 565             /* Calculate squared distance and things based on it */
 566             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 567
 568             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 569
 570             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 571
 572             /* Load parameters for j particles */
 573             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 574                                                               charge+jnrC+0,charge+jnrD+0);
 575
 576             /**************************
 577              * CALCULATE INTERACTIONS *
 578              **************************/
 579
 580             if (gmx_mm_any_lt(rsq00,rcutoff2))
 581             {
 582
 583             r00              = _mm_mul_ps(rsq00,rinv00);
 584
 585             /* Compute parameters for interactions between i and j atoms */
 586             qq00             = _mm_mul_ps(iq0,jq0);
 587
 588             /* EWALD ELECTROSTATICS */
 589
 590             /* Analytical PME correction */
 591             zeta2            = _mm_mul_ps(beta2,rsq00);
 592             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 593             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 594             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 595             felec            = _mm_mul_ps(qq00,felec);
 596             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 597             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 598             velec            = _mm_mul_ps(qq00,velec);
 599
 600             d                = _mm_sub_ps(r00,rswitch);
 601             d                = _mm_max_ps(d,_mm_setzero_ps());
 602             d2               = _mm_mul_ps(d,d);
 603             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 604
 605             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 606
 607             /* Evaluate switch function */
 608             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 609             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 610             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 611
 612             fscal            = felec;
 613
 614             fscal            = _mm_and_ps(fscal,cutoff_mask);
 615
 616              /* Update vectorial force */
 617             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 618             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 619             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 620
 621             fjptrA             = f+j_coord_offsetA;
 622             fjptrB             = f+j_coord_offsetB;
 623             fjptrC             = f+j_coord_offsetC;
 624             fjptrD             = f+j_coord_offsetD;
 625             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 626                                                    _mm_mul_ps(dx00,fscal),
 627                                                    _mm_mul_ps(dy00,fscal),
 628                                                    _mm_mul_ps(dz00,fscal));
 629
 630             }
 631
 632             /* Inner loop uses 50 flops */
 633         }
 634
 635         if(jidx<j_index_end)
 636         {
 637
 638             /* Get j neighbor index, and coordinate index */
 639             jnrlistA         = jjnr[jidx];
 640             jnrlistB         = jjnr[jidx+1];
 641             jnrlistC         = jjnr[jidx+2];
 642             jnrlistD         = jjnr[jidx+3];
 643             /* Sign of each element will be negative for non-real atoms.
 644              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 645              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 646              */
 647             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 648             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 649             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 650             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 651             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 652             j_coord_offsetA  = DIM*jnrA;
 653             j_coord_offsetB  = DIM*jnrB;
 654             j_coord_offsetC  = DIM*jnrC;
 655             j_coord_offsetD  = DIM*jnrD;
 656
 657             /* load j atom coordinates */
 658             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 659                                               x+j_coord_offsetC,x+j_coord_offsetD,
 660                                               &jx0,&jy0,&jz0);
 661
 662             /* Calculate displacement vector */
 663             dx00             = _mm_sub_ps(ix0,jx0);
 664             dy00             = _mm_sub_ps(iy0,jy0);
 665             dz00             = _mm_sub_ps(iz0,jz0);
 666
 667             /* Calculate squared distance and things based on it */
 668             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 669
 670             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 671
 672             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 673
 674             /* Load parameters for j particles */
 675             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 676                                                               charge+jnrC+0,charge+jnrD+0);
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             if (gmx_mm_any_lt(rsq00,rcutoff2))
 683             {
 684
 685             r00              = _mm_mul_ps(rsq00,rinv00);
 686             r00              = _mm_andnot_ps(dummy_mask,r00);
 687
 688             /* Compute parameters for interactions between i and j atoms */
 689             qq00             = _mm_mul_ps(iq0,jq0);
 690
 691             /* EWALD ELECTROSTATICS */
 692
 693             /* Analytical PME correction */
 694             zeta2            = _mm_mul_ps(beta2,rsq00);
 695             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 696             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 697             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 698             felec            = _mm_mul_ps(qq00,felec);
 699             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 700             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 701             velec            = _mm_mul_ps(qq00,velec);
 702
 703             d                = _mm_sub_ps(r00,rswitch);
 704             d                = _mm_max_ps(d,_mm_setzero_ps());
 705             d2               = _mm_mul_ps(d,d);
 706             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 707
 708             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 709
 710             /* Evaluate switch function */
 711             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 712             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 713             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 714
 715             fscal            = felec;
 716
 717             fscal            = _mm_and_ps(fscal,cutoff_mask);
 718
 719             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 720
 721              /* Update vectorial force */
 722             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 723             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 724             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 725
 726             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 727             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 728             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 729             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 730             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 731                                                    _mm_mul_ps(dx00,fscal),
 732                                                    _mm_mul_ps(dy00,fscal),
 733                                                    _mm_mul_ps(dz00,fscal));
 734
 735             }
 736
 737             /* Inner loop uses 51 flops */
 738         }
 739
 740         /* End of innermost loop */
 741
 742         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 743                                               f+i_coord_offset,fshift+i_shift_offset);
 744
 745         /* Increment number of inner iterations */
 746         inneriter                  += j_index_end - j_index_start;
 747
 748         /* Outer loop uses 7 flops */
 749     }
 750
 751     /* Increment number of outer iterations */
 752     outeriter        += nri;
 753
 754     /* Update outer/inner flops */
 755
 756     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*51);
 757 }