src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecGB_VdwNone_GeomP1P1_sse4_1_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 sse4_1_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_sse4_1_single.h"
  48 #include "kernelutil_x86_sse4_1_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_sse4_1_single
  52  * Electrostatics interaction: GeneralizedBorn
  53  * VdW interaction:            None
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_sse4_1_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          gbitab;
  91     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  92     __m128           minushalf = _mm_set1_ps(-0.5);
  93     real             *invsqrta,*dvda,*gbtab;
  94     __m128i          vfitab;
  95     __m128i          ifour       = _mm_set1_epi32(4);
  96     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  97     real             *vftab;
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm_set1_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115
 116     invsqrta         = fr->invsqrta;
 117     dvda             = fr->dvda;
 118     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 119     gbtab            = fr->gbtab.data;
 120     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm_setzero_ps();
 155         fiy0             = _mm_setzero_ps();
 156         fiz0             = _mm_setzero_ps();
 157
 158         /* Load parameters for i particles */
 159         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 160         isai0            = _mm_load1_ps(invsqrta+inr+0);
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm_setzero_ps();
 164         vgbsum           = _mm_setzero_ps();
 165         dvdasum          = _mm_setzero_ps();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178             j_coord_offsetC  = DIM*jnrC;
 179             j_coord_offsetD  = DIM*jnrD;
 180
 181             /* load j atom coordinates */
 182             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 183                                               x+j_coord_offsetC,x+j_coord_offsetD,
 184                                               &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm_sub_ps(ix0,jx0);
 188             dy00             = _mm_sub_ps(iy0,jy0);
 189             dz00             = _mm_sub_ps(iz0,jz0);
 190
 191             /* Calculate squared distance and things based on it */
 192             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 193
 194             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 195
 196             /* Load parameters for j particles */
 197             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 198                                                               charge+jnrC+0,charge+jnrD+0);
 199             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 200                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 201
 202             /**************************
 203              * CALCULATE INTERACTIONS *
 204              **************************/
 205
 206             r00              = _mm_mul_ps(rsq00,rinv00);
 207
 208             /* Compute parameters for interactions between i and j atoms */
 209             qq00             = _mm_mul_ps(iq0,jq0);
 210
 211             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 212             isaprod          = _mm_mul_ps(isai0,isaj0);
 213             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 214             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 215
 216             /* Calculate generalized born table index - this is a separate table from the normal one,
 217              * but we use the same procedure by multiplying r with scale and truncating to integer.
 218              */
 219             rt               = _mm_mul_ps(r00,gbscale);
 220             gbitab           = _mm_cvttps_epi32(rt);
 221             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 222             gbitab           = _mm_slli_epi32(gbitab,2);
 223             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 224             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 225             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 226             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 227             _MM_TRANSPOSE4_PS(Y,F,G,H);
 228             Heps             = _mm_mul_ps(gbeps,H);
 229             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 230             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 231             vgb              = _mm_mul_ps(gbqqfactor,VV);
 232
 233             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 234             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 235             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 236             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 237             fjptrA           = dvda+jnrA;
 238             fjptrB           = dvda+jnrB;
 239             fjptrC           = dvda+jnrC;
 240             fjptrD           = dvda+jnrD;
 241             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 242             velec            = _mm_mul_ps(qq00,rinv00);
 243             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 244
 245             /* Update potential sum for this i atom from the interaction with this j atom. */
 246             velecsum         = _mm_add_ps(velecsum,velec);
 247             vgbsum           = _mm_add_ps(vgbsum,vgb);
 248
 249             fscal            = felec;
 250
 251             /* Calculate temporary vectorial force */
 252             tx               = _mm_mul_ps(fscal,dx00);
 253             ty               = _mm_mul_ps(fscal,dy00);
 254             tz               = _mm_mul_ps(fscal,dz00);
 255
 256             /* Update vectorial force */
 257             fix0             = _mm_add_ps(fix0,tx);
 258             fiy0             = _mm_add_ps(fiy0,ty);
 259             fiz0             = _mm_add_ps(fiz0,tz);
 260
 261             fjptrA             = f+j_coord_offsetA;
 262             fjptrB             = f+j_coord_offsetB;
 263             fjptrC             = f+j_coord_offsetC;
 264             fjptrD             = f+j_coord_offsetD;
 265             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 266
 267             /* Inner loop uses 58 flops */
 268         }
 269
 270         if(jidx<j_index_end)
 271         {
 272
 273             /* Get j neighbor index, and coordinate index */
 274             jnrlistA         = jjnr[jidx];
 275             jnrlistB         = jjnr[jidx+1];
 276             jnrlistC         = jjnr[jidx+2];
 277             jnrlistD         = jjnr[jidx+3];
 278             /* Sign of each element will be negative for non-real atoms.
 279              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 280              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 281              */
 282             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 283             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 284             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 285             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 286             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 287             j_coord_offsetA  = DIM*jnrA;
 288             j_coord_offsetB  = DIM*jnrB;
 289             j_coord_offsetC  = DIM*jnrC;
 290             j_coord_offsetD  = DIM*jnrD;
 291
 292             /* load j atom coordinates */
 293             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 294                                               x+j_coord_offsetC,x+j_coord_offsetD,
 295                                               &jx0,&jy0,&jz0);
 296
 297             /* Calculate displacement vector */
 298             dx00             = _mm_sub_ps(ix0,jx0);
 299             dy00             = _mm_sub_ps(iy0,jy0);
 300             dz00             = _mm_sub_ps(iz0,jz0);
 301
 302             /* Calculate squared distance and things based on it */
 303             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 304
 305             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 306
 307             /* Load parameters for j particles */
 308             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 309                                                               charge+jnrC+0,charge+jnrD+0);
 310             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 311                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             r00              = _mm_mul_ps(rsq00,rinv00);
 318             r00              = _mm_andnot_ps(dummy_mask,r00);
 319
 320             /* Compute parameters for interactions between i and j atoms */
 321             qq00             = _mm_mul_ps(iq0,jq0);
 322
 323             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 324             isaprod          = _mm_mul_ps(isai0,isaj0);
 325             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 326             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 327
 328             /* Calculate generalized born table index - this is a separate table from the normal one,
 329              * but we use the same procedure by multiplying r with scale and truncating to integer.
 330              */
 331             rt               = _mm_mul_ps(r00,gbscale);
 332             gbitab           = _mm_cvttps_epi32(rt);
 333             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 334             gbitab           = _mm_slli_epi32(gbitab,2);
 335             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 336             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 337             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 338             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 339             _MM_TRANSPOSE4_PS(Y,F,G,H);
 340             Heps             = _mm_mul_ps(gbeps,H);
 341             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 342             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 343             vgb              = _mm_mul_ps(gbqqfactor,VV);
 344
 345             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 346             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 347             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 348             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 349             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 350             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 351             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 352             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 353             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 354             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 355             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 356             velec            = _mm_mul_ps(qq00,rinv00);
 357             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velec            = _mm_andnot_ps(dummy_mask,velec);
 361             velecsum         = _mm_add_ps(velecsum,velec);
 362             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 363             vgbsum           = _mm_add_ps(vgbsum,vgb);
 364
 365             fscal            = felec;
 366
 367             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 368
 369             /* Calculate temporary vectorial force */
 370             tx               = _mm_mul_ps(fscal,dx00);
 371             ty               = _mm_mul_ps(fscal,dy00);
 372             tz               = _mm_mul_ps(fscal,dz00);
 373
 374             /* Update vectorial force */
 375             fix0             = _mm_add_ps(fix0,tx);
 376             fiy0             = _mm_add_ps(fiy0,ty);
 377             fiz0             = _mm_add_ps(fiz0,tz);
 378
 379             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 380             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 381             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 382             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 383             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 384
 385             /* Inner loop uses 59 flops */
 386         }
 387
 388         /* End of innermost loop */
 389
 390         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 391                                               f+i_coord_offset,fshift+i_shift_offset);
 392
 393         ggid                        = gid[iidx];
 394         /* Update potential energies */
 395         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 396         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 397         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 398         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 399
 400         /* Increment number of inner iterations */
 401         inneriter                  += j_index_end - j_index_start;
 402
 403         /* Outer loop uses 9 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*9 + inneriter*59);
 412 }
 413 /*
 414  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse4_1_single
 415  * Electrostatics interaction: GeneralizedBorn
 416  * VdW interaction:            None
 417  * Geometry:                   Particle-Particle
 418  * Calculate force/pot:        Force
 419  */
 420 void
 421 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse4_1_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 SSE, 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           tx,ty,tz,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          gbitab;
 454     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 455     __m128           minushalf = _mm_set1_ps(-0.5);
 456     real             *invsqrta,*dvda,*gbtab;
 457     __m128i          vfitab;
 458     __m128i          ifour       = _mm_set1_epi32(4);
 459     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 460     real             *vftab;
 461     __m128           dummy_mask,cutoff_mask;
 462     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 463     __m128           one     = _mm_set1_ps(1.0);
 464     __m128           two     = _mm_set1_ps(2.0);
 465     x                = xx[0];
 466     f                = ff[0];
 467
 468     nri              = nlist->nri;
 469     iinr             = nlist->iinr;
 470     jindex           = nlist->jindex;
 471     jjnr             = nlist->jjnr;
 472     shiftidx         = nlist->shift;
 473     gid              = nlist->gid;
 474     shiftvec         = fr->shift_vec[0];
 475     fshift           = fr->fshift[0];
 476     facel            = _mm_set1_ps(fr->epsfac);
 477     charge           = mdatoms->chargeA;
 478
 479     invsqrta         = fr->invsqrta;
 480     dvda             = fr->dvda;
 481     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 482     gbtab            = fr->gbtab.data;
 483     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 484
 485     /* Avoid stupid compiler warnings */
 486     jnrA = jnrB = jnrC = jnrD = 0;
 487     j_coord_offsetA = 0;
 488     j_coord_offsetB = 0;
 489     j_coord_offsetC = 0;
 490     j_coord_offsetD = 0;
 491
 492     outeriter        = 0;
 493     inneriter        = 0;
 494
 495     for(iidx=0;iidx<4*DIM;iidx++)
 496     {
 497         scratch[iidx] = 0.0;
 498     }
 499
 500     /* Start outer loop over neighborlists */
 501     for(iidx=0; iidx<nri; iidx++)
 502     {
 503         /* Load shift vector for this list */
 504         i_shift_offset   = DIM*shiftidx[iidx];
 505
 506         /* Load limits for loop over neighbors */
 507         j_index_start    = jindex[iidx];
 508         j_index_end      = jindex[iidx+1];
 509
 510         /* Get outer coordinate index */
 511         inr              = iinr[iidx];
 512         i_coord_offset   = DIM*inr;
 513
 514         /* Load i particle coords and add shift vector */
 515         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 516
 517         fix0             = _mm_setzero_ps();
 518         fiy0             = _mm_setzero_ps();
 519         fiz0             = _mm_setzero_ps();
 520
 521         /* Load parameters for i particles */
 522         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 523         isai0            = _mm_load1_ps(invsqrta+inr+0);
 524
 525         dvdasum          = _mm_setzero_ps();
 526
 527         /* Start inner kernel loop */
 528         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 529         {
 530
 531             /* Get j neighbor index, and coordinate index */
 532             jnrA             = jjnr[jidx];
 533             jnrB             = jjnr[jidx+1];
 534             jnrC             = jjnr[jidx+2];
 535             jnrD             = jjnr[jidx+3];
 536             j_coord_offsetA  = DIM*jnrA;
 537             j_coord_offsetB  = DIM*jnrB;
 538             j_coord_offsetC  = DIM*jnrC;
 539             j_coord_offsetD  = DIM*jnrD;
 540
 541             /* load j atom coordinates */
 542             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 543                                               x+j_coord_offsetC,x+j_coord_offsetD,
 544                                               &jx0,&jy0,&jz0);
 545
 546             /* Calculate displacement vector */
 547             dx00             = _mm_sub_ps(ix0,jx0);
 548             dy00             = _mm_sub_ps(iy0,jy0);
 549             dz00             = _mm_sub_ps(iz0,jz0);
 550
 551             /* Calculate squared distance and things based on it */
 552             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 553
 554             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 555
 556             /* Load parameters for j particles */
 557             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 558                                                               charge+jnrC+0,charge+jnrD+0);
 559             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 560                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 561
 562             /**************************
 563              * CALCULATE INTERACTIONS *
 564              **************************/
 565
 566             r00              = _mm_mul_ps(rsq00,rinv00);
 567
 568             /* Compute parameters for interactions between i and j atoms */
 569             qq00             = _mm_mul_ps(iq0,jq0);
 570
 571             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 572             isaprod          = _mm_mul_ps(isai0,isaj0);
 573             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 574             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 575
 576             /* Calculate generalized born table index - this is a separate table from the normal one,
 577              * but we use the same procedure by multiplying r with scale and truncating to integer.
 578              */
 579             rt               = _mm_mul_ps(r00,gbscale);
 580             gbitab           = _mm_cvttps_epi32(rt);
 581             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 582             gbitab           = _mm_slli_epi32(gbitab,2);
 583             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 584             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 585             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 586             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 587             _MM_TRANSPOSE4_PS(Y,F,G,H);
 588             Heps             = _mm_mul_ps(gbeps,H);
 589             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 590             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 591             vgb              = _mm_mul_ps(gbqqfactor,VV);
 592
 593             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 594             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 595             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 596             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 597             fjptrA           = dvda+jnrA;
 598             fjptrB           = dvda+jnrB;
 599             fjptrC           = dvda+jnrC;
 600             fjptrD           = dvda+jnrD;
 601             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 602             velec            = _mm_mul_ps(qq00,rinv00);
 603             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 604
 605             fscal            = felec;
 606
 607             /* Calculate temporary vectorial force */
 608             tx               = _mm_mul_ps(fscal,dx00);
 609             ty               = _mm_mul_ps(fscal,dy00);
 610             tz               = _mm_mul_ps(fscal,dz00);
 611
 612             /* Update vectorial force */
 613             fix0             = _mm_add_ps(fix0,tx);
 614             fiy0             = _mm_add_ps(fiy0,ty);
 615             fiz0             = _mm_add_ps(fiz0,tz);
 616
 617             fjptrA             = f+j_coord_offsetA;
 618             fjptrB             = f+j_coord_offsetB;
 619             fjptrC             = f+j_coord_offsetC;
 620             fjptrD             = f+j_coord_offsetD;
 621             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 622
 623             /* Inner loop uses 56 flops */
 624         }
 625
 626         if(jidx<j_index_end)
 627         {
 628
 629             /* Get j neighbor index, and coordinate index */
 630             jnrlistA         = jjnr[jidx];
 631             jnrlistB         = jjnr[jidx+1];
 632             jnrlistC         = jjnr[jidx+2];
 633             jnrlistD         = jjnr[jidx+3];
 634             /* Sign of each element will be negative for non-real atoms.
 635              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 636              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 637              */
 638             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 639             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 640             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 641             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 642             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 643             j_coord_offsetA  = DIM*jnrA;
 644             j_coord_offsetB  = DIM*jnrB;
 645             j_coord_offsetC  = DIM*jnrC;
 646             j_coord_offsetD  = DIM*jnrD;
 647
 648             /* load j atom coordinates */
 649             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 650                                               x+j_coord_offsetC,x+j_coord_offsetD,
 651                                               &jx0,&jy0,&jz0);
 652
 653             /* Calculate displacement vector */
 654             dx00             = _mm_sub_ps(ix0,jx0);
 655             dy00             = _mm_sub_ps(iy0,jy0);
 656             dz00             = _mm_sub_ps(iz0,jz0);
 657
 658             /* Calculate squared distance and things based on it */
 659             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 660
 661             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 662
 663             /* Load parameters for j particles */
 664             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 665                                                               charge+jnrC+0,charge+jnrD+0);
 666             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 667                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 668
 669             /**************************
 670              * CALCULATE INTERACTIONS *
 671              **************************/
 672
 673             r00              = _mm_mul_ps(rsq00,rinv00);
 674             r00              = _mm_andnot_ps(dummy_mask,r00);
 675
 676             /* Compute parameters for interactions between i and j atoms */
 677             qq00             = _mm_mul_ps(iq0,jq0);
 678
 679             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 680             isaprod          = _mm_mul_ps(isai0,isaj0);
 681             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 682             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 683
 684             /* Calculate generalized born table index - this is a separate table from the normal one,
 685              * but we use the same procedure by multiplying r with scale and truncating to integer.
 686              */
 687             rt               = _mm_mul_ps(r00,gbscale);
 688             gbitab           = _mm_cvttps_epi32(rt);
 689             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 690             gbitab           = _mm_slli_epi32(gbitab,2);
 691             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 692             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 693             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 694             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 695             _MM_TRANSPOSE4_PS(Y,F,G,H);
 696             Heps             = _mm_mul_ps(gbeps,H);
 697             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 698             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 699             vgb              = _mm_mul_ps(gbqqfactor,VV);
 700
 701             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 702             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 703             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 704             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 705             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 706             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 707             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 708             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 709             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 710             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 711             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 712             velec            = _mm_mul_ps(qq00,rinv00);
 713             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 714
 715             fscal            = felec;
 716
 717             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 718
 719             /* Calculate temporary vectorial force */
 720             tx               = _mm_mul_ps(fscal,dx00);
 721             ty               = _mm_mul_ps(fscal,dy00);
 722             tz               = _mm_mul_ps(fscal,dz00);
 723
 724             /* Update vectorial force */
 725             fix0             = _mm_add_ps(fix0,tx);
 726             fiy0             = _mm_add_ps(fiy0,ty);
 727             fiz0             = _mm_add_ps(fiz0,tz);
 728
 729             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 730             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 731             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 732             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 733             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 734
 735             /* Inner loop uses 57 flops */
 736         }
 737
 738         /* End of innermost loop */
 739
 740         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 741                                               f+i_coord_offset,fshift+i_shift_offset);
 742
 743         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 744         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 745
 746         /* Increment number of inner iterations */
 747         inneriter                  += j_index_end - j_index_start;
 748
 749         /* Outer loop uses 7 flops */
 750     }
 751
 752     /* Increment number of outer iterations */
 753     outeriter        += nri;
 754
 755     /* Update outer/inner flops */
 756
 757     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*57);
 758 }