src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_sse2_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 sse2_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_sse2_single.h"
  48 #include "kernelutil_x86_sse2_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_sse2_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_sse2_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 SSE, 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           tx,ty,tz,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,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  92     real             *ewtab;
  93     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  94     real             rswitch_scalar,d_scalar;
  95     __m128           dummy_mask,cutoff_mask;
  96     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  97     __m128           one     = _mm_set1_ps(1.0);
  98     __m128           two     = _mm_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_ps(fr->epsfac);
 111     charge           = mdatoms->chargeA;
 112
 113     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 114     ewtab            = fr->ic->tabq_coul_FDV0;
 115     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 116     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 117
 118     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 119     rcutoff_scalar   = fr->rcoulomb;
 120     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 121     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 122
 123     rswitch_scalar   = fr->rcoulomb_switch;
 124     rswitch          = _mm_set1_ps(rswitch_scalar);
 125     /* Setup switch parameters */
 126     d_scalar         = rcutoff_scalar-rswitch_scalar;
 127     d                = _mm_set1_ps(d_scalar);
 128     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 129     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 130     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 131     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 132     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 133     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 134
 135     /* Avoid stupid compiler warnings */
 136     jnrA = jnrB = jnrC = jnrD = 0;
 137     j_coord_offsetA = 0;
 138     j_coord_offsetB = 0;
 139     j_coord_offsetC = 0;
 140     j_coord_offsetD = 0;
 141
 142     outeriter        = 0;
 143     inneriter        = 0;
 144
 145     for(iidx=0;iidx<4*DIM;iidx++)
 146     {
 147         scratch[iidx] = 0.0;
 148     }
 149
 150     /* Start outer loop over neighborlists */
 151     for(iidx=0; iidx<nri; iidx++)
 152     {
 153         /* Load shift vector for this list */
 154         i_shift_offset   = DIM*shiftidx[iidx];
 155
 156         /* Load limits for loop over neighbors */
 157         j_index_start    = jindex[iidx];
 158         j_index_end      = jindex[iidx+1];
 159
 160         /* Get outer coordinate index */
 161         inr              = iinr[iidx];
 162         i_coord_offset   = DIM*inr;
 163
 164         /* Load i particle coords and add shift vector */
 165         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 166
 167         fix0             = _mm_setzero_ps();
 168         fiy0             = _mm_setzero_ps();
 169         fiz0             = _mm_setzero_ps();
 170
 171         /* Load parameters for i particles */
 172         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_ps();
 176
 177         /* Start inner kernel loop */
 178         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 179         {
 180
 181             /* Get j neighbor index, and coordinate index */
 182             jnrA             = jjnr[jidx];
 183             jnrB             = jjnr[jidx+1];
 184             jnrC             = jjnr[jidx+2];
 185             jnrD             = jjnr[jidx+3];
 186             j_coord_offsetA  = DIM*jnrA;
 187             j_coord_offsetB  = DIM*jnrB;
 188             j_coord_offsetC  = DIM*jnrC;
 189             j_coord_offsetD  = DIM*jnrD;
 190
 191             /* load j atom coordinates */
 192             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 193                                               x+j_coord_offsetC,x+j_coord_offsetD,
 194                                               &jx0,&jy0,&jz0);
 195
 196             /* Calculate displacement vector */
 197             dx00             = _mm_sub_ps(ix0,jx0);
 198             dy00             = _mm_sub_ps(iy0,jy0);
 199             dz00             = _mm_sub_ps(iz0,jz0);
 200
 201             /* Calculate squared distance and things based on it */
 202             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 203
 204             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 205
 206             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 207
 208             /* Load parameters for j particles */
 209             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 210                                                               charge+jnrC+0,charge+jnrD+0);
 211
 212             /**************************
 213              * CALCULATE INTERACTIONS *
 214              **************************/
 215
 216             if (gmx_mm_any_lt(rsq00,rcutoff2))
 217             {
 218
 219             r00              = _mm_mul_ps(rsq00,rinv00);
 220
 221             /* Compute parameters for interactions between i and j atoms */
 222             qq00             = _mm_mul_ps(iq0,jq0);
 223
 224             /* EWALD ELECTROSTATICS */
 225
 226             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 227             ewrt             = _mm_mul_ps(r00,ewtabscale);
 228             ewitab           = _mm_cvttps_epi32(ewrt);
 229             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 230             ewitab           = _mm_slli_epi32(ewitab,2);
 231             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 232             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 233             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 234             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 235             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 236             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 237             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 238             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 239             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 240
 241             d                = _mm_sub_ps(r00,rswitch);
 242             d                = _mm_max_ps(d,_mm_setzero_ps());
 243             d2               = _mm_mul_ps(d,d);
 244             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 245
 246             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 247
 248             /* Evaluate switch function */
 249             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 250             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 251             velec            = _mm_mul_ps(velec,sw);
 252             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 253
 254             /* Update potential sum for this i atom from the interaction with this j atom. */
 255             velec            = _mm_and_ps(velec,cutoff_mask);
 256             velecsum         = _mm_add_ps(velecsum,velec);
 257
 258             fscal            = felec;
 259
 260             fscal            = _mm_and_ps(fscal,cutoff_mask);
 261
 262             /* Calculate temporary vectorial force */
 263             tx               = _mm_mul_ps(fscal,dx00);
 264             ty               = _mm_mul_ps(fscal,dy00);
 265             tz               = _mm_mul_ps(fscal,dz00);
 266
 267             /* Update vectorial force */
 268             fix0             = _mm_add_ps(fix0,tx);
 269             fiy0             = _mm_add_ps(fiy0,ty);
 270             fiz0             = _mm_add_ps(fiz0,tz);
 271
 272             fjptrA             = f+j_coord_offsetA;
 273             fjptrB             = f+j_coord_offsetB;
 274             fjptrC             = f+j_coord_offsetC;
 275             fjptrD             = f+j_coord_offsetD;
 276             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 277
 278             }
 279
 280             /* Inner loop uses 65 flops */
 281         }
 282
 283         if(jidx<j_index_end)
 284         {
 285
 286             /* Get j neighbor index, and coordinate index */
 287             jnrlistA         = jjnr[jidx];
 288             jnrlistB         = jjnr[jidx+1];
 289             jnrlistC         = jjnr[jidx+2];
 290             jnrlistD         = jjnr[jidx+3];
 291             /* Sign of each element will be negative for non-real atoms.
 292              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 293              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 294              */
 295             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 296             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 297             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 298             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 299             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 300             j_coord_offsetA  = DIM*jnrA;
 301             j_coord_offsetB  = DIM*jnrB;
 302             j_coord_offsetC  = DIM*jnrC;
 303             j_coord_offsetD  = DIM*jnrD;
 304
 305             /* load j atom coordinates */
 306             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 307                                               x+j_coord_offsetC,x+j_coord_offsetD,
 308                                               &jx0,&jy0,&jz0);
 309
 310             /* Calculate displacement vector */
 311             dx00             = _mm_sub_ps(ix0,jx0);
 312             dy00             = _mm_sub_ps(iy0,jy0);
 313             dz00             = _mm_sub_ps(iz0,jz0);
 314
 315             /* Calculate squared distance and things based on it */
 316             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 317
 318             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 319
 320             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 321
 322             /* Load parameters for j particles */
 323             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 324                                                               charge+jnrC+0,charge+jnrD+0);
 325
 326             /**************************
 327              * CALCULATE INTERACTIONS *
 328              **************************/
 329
 330             if (gmx_mm_any_lt(rsq00,rcutoff2))
 331             {
 332
 333             r00              = _mm_mul_ps(rsq00,rinv00);
 334             r00              = _mm_andnot_ps(dummy_mask,r00);
 335
 336             /* Compute parameters for interactions between i and j atoms */
 337             qq00             = _mm_mul_ps(iq0,jq0);
 338
 339             /* EWALD ELECTROSTATICS */
 340
 341             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 342             ewrt             = _mm_mul_ps(r00,ewtabscale);
 343             ewitab           = _mm_cvttps_epi32(ewrt);
 344             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 345             ewitab           = _mm_slli_epi32(ewitab,2);
 346             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 347             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 348             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 349             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 350             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 351             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 352             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 353             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 354             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 355
 356             d                = _mm_sub_ps(r00,rswitch);
 357             d                = _mm_max_ps(d,_mm_setzero_ps());
 358             d2               = _mm_mul_ps(d,d);
 359             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 360
 361             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 362
 363             /* Evaluate switch function */
 364             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 365             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 366             velec            = _mm_mul_ps(velec,sw);
 367             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm_and_ps(velec,cutoff_mask);
 371             velec            = _mm_andnot_ps(dummy_mask,velec);
 372             velecsum         = _mm_add_ps(velecsum,velec);
 373
 374             fscal            = felec;
 375
 376             fscal            = _mm_and_ps(fscal,cutoff_mask);
 377
 378             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 379
 380             /* Calculate temporary vectorial force */
 381             tx               = _mm_mul_ps(fscal,dx00);
 382             ty               = _mm_mul_ps(fscal,dy00);
 383             tz               = _mm_mul_ps(fscal,dz00);
 384
 385             /* Update vectorial force */
 386             fix0             = _mm_add_ps(fix0,tx);
 387             fiy0             = _mm_add_ps(fiy0,ty);
 388             fiz0             = _mm_add_ps(fiz0,tz);
 389
 390             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 391             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 392             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 393             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 394             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 395
 396             }
 397
 398             /* Inner loop uses 66 flops */
 399         }
 400
 401         /* End of innermost loop */
 402
 403         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 404                                               f+i_coord_offset,fshift+i_shift_offset);
 405
 406         ggid                        = gid[iidx];
 407         /* Update potential energies */
 408         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 409
 410         /* Increment number of inner iterations */
 411         inneriter                  += j_index_end - j_index_start;
 412
 413         /* Outer loop uses 8 flops */
 414     }
 415
 416     /* Increment number of outer iterations */
 417     outeriter        += nri;
 418
 419     /* Update outer/inner flops */
 420
 421     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*66);
 422 }
 423 /*
 424  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_sse2_single
 425  * Electrostatics interaction: Ewald
 426  * VdW interaction:            None
 427  * Geometry:                   Particle-Particle
 428  * Calculate force/pot:        Force
 429  */
 430 void
 431 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_sse2_single
 432                     (t_nblist                    * gmx_restrict       nlist,
 433                      rvec                        * gmx_restrict          xx,
 434                      rvec                        * gmx_restrict          ff,
 435                      t_forcerec                  * gmx_restrict          fr,
 436                      t_mdatoms                   * gmx_restrict     mdatoms,
 437                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 438                      t_nrnb                      * gmx_restrict        nrnb)
 439 {
 440     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 441      * just 0 for non-waters.
 442      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 443      * jnr indices corresponding to data put in the four positions in the SIMD register.
 444      */
 445     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 446     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 447     int              jnrA,jnrB,jnrC,jnrD;
 448     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 449     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 450     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 451     real             rcutoff_scalar;
 452     real             *shiftvec,*fshift,*x,*f;
 453     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 454     real             scratch[4*DIM];
 455     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 456     int              vdwioffset0;
 457     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 458     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 459     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 460     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 461     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 462     real             *charge;
 463     __m128i          ewitab;
 464     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 465     real             *ewtab;
 466     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 467     real             rswitch_scalar,d_scalar;
 468     __m128           dummy_mask,cutoff_mask;
 469     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 470     __m128           one     = _mm_set1_ps(1.0);
 471     __m128           two     = _mm_set1_ps(2.0);
 472     x                = xx[0];
 473     f                = ff[0];
 474
 475     nri              = nlist->nri;
 476     iinr             = nlist->iinr;
 477     jindex           = nlist->jindex;
 478     jjnr             = nlist->jjnr;
 479     shiftidx         = nlist->shift;
 480     gid              = nlist->gid;
 481     shiftvec         = fr->shift_vec[0];
 482     fshift           = fr->fshift[0];
 483     facel            = _mm_set1_ps(fr->epsfac);
 484     charge           = mdatoms->chargeA;
 485
 486     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 487     ewtab            = fr->ic->tabq_coul_FDV0;
 488     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 489     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 490
 491     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 492     rcutoff_scalar   = fr->rcoulomb;
 493     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 494     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 495
 496     rswitch_scalar   = fr->rcoulomb_switch;
 497     rswitch          = _mm_set1_ps(rswitch_scalar);
 498     /* Setup switch parameters */
 499     d_scalar         = rcutoff_scalar-rswitch_scalar;
 500     d                = _mm_set1_ps(d_scalar);
 501     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 502     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 503     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 504     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 505     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 506     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 507
 508     /* Avoid stupid compiler warnings */
 509     jnrA = jnrB = jnrC = jnrD = 0;
 510     j_coord_offsetA = 0;
 511     j_coord_offsetB = 0;
 512     j_coord_offsetC = 0;
 513     j_coord_offsetD = 0;
 514
 515     outeriter        = 0;
 516     inneriter        = 0;
 517
 518     for(iidx=0;iidx<4*DIM;iidx++)
 519     {
 520         scratch[iidx] = 0.0;
 521     }
 522
 523     /* Start outer loop over neighborlists */
 524     for(iidx=0; iidx<nri; iidx++)
 525     {
 526         /* Load shift vector for this list */
 527         i_shift_offset   = DIM*shiftidx[iidx];
 528
 529         /* Load limits for loop over neighbors */
 530         j_index_start    = jindex[iidx];
 531         j_index_end      = jindex[iidx+1];
 532
 533         /* Get outer coordinate index */
 534         inr              = iinr[iidx];
 535         i_coord_offset   = DIM*inr;
 536
 537         /* Load i particle coords and add shift vector */
 538         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 539
 540         fix0             = _mm_setzero_ps();
 541         fiy0             = _mm_setzero_ps();
 542         fiz0             = _mm_setzero_ps();
 543
 544         /* Load parameters for i particles */
 545         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 546
 547         /* Start inner kernel loop */
 548         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 549         {
 550
 551             /* Get j neighbor index, and coordinate index */
 552             jnrA             = jjnr[jidx];
 553             jnrB             = jjnr[jidx+1];
 554             jnrC             = jjnr[jidx+2];
 555             jnrD             = jjnr[jidx+3];
 556             j_coord_offsetA  = DIM*jnrA;
 557             j_coord_offsetB  = DIM*jnrB;
 558             j_coord_offsetC  = DIM*jnrC;
 559             j_coord_offsetD  = DIM*jnrD;
 560
 561             /* load j atom coordinates */
 562             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 563                                               x+j_coord_offsetC,x+j_coord_offsetD,
 564                                               &jx0,&jy0,&jz0);
 565
 566             /* Calculate displacement vector */
 567             dx00             = _mm_sub_ps(ix0,jx0);
 568             dy00             = _mm_sub_ps(iy0,jy0);
 569             dz00             = _mm_sub_ps(iz0,jz0);
 570
 571             /* Calculate squared distance and things based on it */
 572             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 573
 574             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 575
 576             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 577
 578             /* Load parameters for j particles */
 579             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 580                                                               charge+jnrC+0,charge+jnrD+0);
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             if (gmx_mm_any_lt(rsq00,rcutoff2))
 587             {
 588
 589             r00              = _mm_mul_ps(rsq00,rinv00);
 590
 591             /* Compute parameters for interactions between i and j atoms */
 592             qq00             = _mm_mul_ps(iq0,jq0);
 593
 594             /* EWALD ELECTROSTATICS */
 595
 596             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 597             ewrt             = _mm_mul_ps(r00,ewtabscale);
 598             ewitab           = _mm_cvttps_epi32(ewrt);
 599             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 600             ewitab           = _mm_slli_epi32(ewitab,2);
 601             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 602             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 603             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 604             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 605             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 606             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 607             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 608             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 609             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 610
 611             d                = _mm_sub_ps(r00,rswitch);
 612             d                = _mm_max_ps(d,_mm_setzero_ps());
 613             d2               = _mm_mul_ps(d,d);
 614             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 615
 616             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 617
 618             /* Evaluate switch function */
 619             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 620             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 621             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 622
 623             fscal            = felec;
 624
 625             fscal            = _mm_and_ps(fscal,cutoff_mask);
 626
 627             /* Calculate temporary vectorial force */
 628             tx               = _mm_mul_ps(fscal,dx00);
 629             ty               = _mm_mul_ps(fscal,dy00);
 630             tz               = _mm_mul_ps(fscal,dz00);
 631
 632             /* Update vectorial force */
 633             fix0             = _mm_add_ps(fix0,tx);
 634             fiy0             = _mm_add_ps(fiy0,ty);
 635             fiz0             = _mm_add_ps(fiz0,tz);
 636
 637             fjptrA             = f+j_coord_offsetA;
 638             fjptrB             = f+j_coord_offsetB;
 639             fjptrC             = f+j_coord_offsetC;
 640             fjptrD             = f+j_coord_offsetD;
 641             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 642
 643             }
 644
 645             /* Inner loop uses 62 flops */
 646         }
 647
 648         if(jidx<j_index_end)
 649         {
 650
 651             /* Get j neighbor index, and coordinate index */
 652             jnrlistA         = jjnr[jidx];
 653             jnrlistB         = jjnr[jidx+1];
 654             jnrlistC         = jjnr[jidx+2];
 655             jnrlistD         = jjnr[jidx+3];
 656             /* Sign of each element will be negative for non-real atoms.
 657              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 658              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 659              */
 660             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 661             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 662             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 663             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 664             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 665             j_coord_offsetA  = DIM*jnrA;
 666             j_coord_offsetB  = DIM*jnrB;
 667             j_coord_offsetC  = DIM*jnrC;
 668             j_coord_offsetD  = DIM*jnrD;
 669
 670             /* load j atom coordinates */
 671             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 672                                               x+j_coord_offsetC,x+j_coord_offsetD,
 673                                               &jx0,&jy0,&jz0);
 674
 675             /* Calculate displacement vector */
 676             dx00             = _mm_sub_ps(ix0,jx0);
 677             dy00             = _mm_sub_ps(iy0,jy0);
 678             dz00             = _mm_sub_ps(iz0,jz0);
 679
 680             /* Calculate squared distance and things based on it */
 681             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 682
 683             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 684
 685             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 686
 687             /* Load parameters for j particles */
 688             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 689                                                               charge+jnrC+0,charge+jnrD+0);
 690
 691             /**************************
 692              * CALCULATE INTERACTIONS *
 693              **************************/
 694
 695             if (gmx_mm_any_lt(rsq00,rcutoff2))
 696             {
 697
 698             r00              = _mm_mul_ps(rsq00,rinv00);
 699             r00              = _mm_andnot_ps(dummy_mask,r00);
 700
 701             /* Compute parameters for interactions between i and j atoms */
 702             qq00             = _mm_mul_ps(iq0,jq0);
 703
 704             /* EWALD ELECTROSTATICS */
 705
 706             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 707             ewrt             = _mm_mul_ps(r00,ewtabscale);
 708             ewitab           = _mm_cvttps_epi32(ewrt);
 709             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 710             ewitab           = _mm_slli_epi32(ewitab,2);
 711             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 712             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 713             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 714             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 715             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 716             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 717             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 718             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 719             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 720
 721             d                = _mm_sub_ps(r00,rswitch);
 722             d                = _mm_max_ps(d,_mm_setzero_ps());
 723             d2               = _mm_mul_ps(d,d);
 724             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 725
 726             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 727
 728             /* Evaluate switch function */
 729             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 730             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 731             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm_and_ps(fscal,cutoff_mask);
 736
 737             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 738
 739             /* Calculate temporary vectorial force */
 740             tx               = _mm_mul_ps(fscal,dx00);
 741             ty               = _mm_mul_ps(fscal,dy00);
 742             tz               = _mm_mul_ps(fscal,dz00);
 743
 744             /* Update vectorial force */
 745             fix0             = _mm_add_ps(fix0,tx);
 746             fiy0             = _mm_add_ps(fiy0,ty);
 747             fiz0             = _mm_add_ps(fiz0,tz);
 748
 749             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 750             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 751             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 752             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 753             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 754
 755             }
 756
 757             /* Inner loop uses 63 flops */
 758         }
 759
 760         /* End of innermost loop */
 761
 762         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 763                                               f+i_coord_offset,fshift+i_shift_offset);
 764
 765         /* Increment number of inner iterations */
 766         inneriter                  += j_index_end - j_index_start;
 767
 768         /* Outer loop uses 7 flops */
 769     }
 770
 771     /* Increment number of outer iterations */
 772     outeriter        += nri;
 773
 774     /* Update outer/inner flops */
 775
 776     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*63);
 777 }