src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecGB_VdwNone_GeomP1P1_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_sse4_1_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_sse4_1_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     __m128i          gbitab;
  77     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  78     __m128           minushalf = _mm_set1_ps(-0.5);
  79     real             *invsqrta,*dvda,*gbtab;
  80     __m128i          vfitab;
  81     __m128i          ifour       = _mm_set1_epi32(4);
  82     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  83     real             *vftab;
  84     __m128           dummy_mask,cutoff_mask;
  85     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  86     __m128           one     = _mm_set1_ps(1.0);
  87     __m128           two     = _mm_set1_ps(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     facel            = _mm_set1_ps(fr->epsfac);
 100     charge           = mdatoms->chargeA;
 101
 102     invsqrta         = fr->invsqrta;
 103     dvda             = fr->dvda;
 104     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 105     gbtab            = fr->gbtab.data;
 106     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 107
 108     /* Avoid stupid compiler warnings */
 109     jnrA = jnrB = jnrC = jnrD = 0;
 110     j_coord_offsetA = 0;
 111     j_coord_offsetB = 0;
 112     j_coord_offsetC = 0;
 113     j_coord_offsetD = 0;
 114
 115     outeriter        = 0;
 116     inneriter        = 0;
 117
 118     for(iidx=0;iidx<4*DIM;iidx++)
 119     {
 120         scratch[iidx] = 0.0;
 121     }
 122
 123     /* Start outer loop over neighborlists */
 124     for(iidx=0; iidx<nri; iidx++)
 125     {
 126         /* Load shift vector for this list */
 127         i_shift_offset   = DIM*shiftidx[iidx];
 128
 129         /* Load limits for loop over neighbors */
 130         j_index_start    = jindex[iidx];
 131         j_index_end      = jindex[iidx+1];
 132
 133         /* Get outer coordinate index */
 134         inr              = iinr[iidx];
 135         i_coord_offset   = DIM*inr;
 136
 137         /* Load i particle coords and add shift vector */
 138         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 139
 140         fix0             = _mm_setzero_ps();
 141         fiy0             = _mm_setzero_ps();
 142         fiz0             = _mm_setzero_ps();
 143
 144         /* Load parameters for i particles */
 145         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 146         isai0            = _mm_load1_ps(invsqrta+inr+0);
 147
 148         /* Reset potential sums */
 149         velecsum         = _mm_setzero_ps();
 150         vgbsum           = _mm_setzero_ps();
 151         dvdasum          = _mm_setzero_ps();
 152
 153         /* Start inner kernel loop */
 154         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 155         {
 156
 157             /* Get j neighbor index, and coordinate index */
 158             jnrA             = jjnr[jidx];
 159             jnrB             = jjnr[jidx+1];
 160             jnrC             = jjnr[jidx+2];
 161             jnrD             = jjnr[jidx+3];
 162             j_coord_offsetA  = DIM*jnrA;
 163             j_coord_offsetB  = DIM*jnrB;
 164             j_coord_offsetC  = DIM*jnrC;
 165             j_coord_offsetD  = DIM*jnrD;
 166
 167             /* load j atom coordinates */
 168             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 169                                               x+j_coord_offsetC,x+j_coord_offsetD,
 170                                               &jx0,&jy0,&jz0);
 171
 172             /* Calculate displacement vector */
 173             dx00             = _mm_sub_ps(ix0,jx0);
 174             dy00             = _mm_sub_ps(iy0,jy0);
 175             dz00             = _mm_sub_ps(iz0,jz0);
 176
 177             /* Calculate squared distance and things based on it */
 178             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 179
 180             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 181
 182             /* Load parameters for j particles */
 183             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 184                                                               charge+jnrC+0,charge+jnrD+0);
 185             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 186                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 187
 188             /**************************
 189              * CALCULATE INTERACTIONS *
 190              **************************/
 191
 192             r00              = _mm_mul_ps(rsq00,rinv00);
 193
 194             /* Compute parameters for interactions between i and j atoms */
 195             qq00             = _mm_mul_ps(iq0,jq0);
 196
 197             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 198             isaprod          = _mm_mul_ps(isai0,isaj0);
 199             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 200             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 201
 202             /* Calculate generalized born table index - this is a separate table from the normal one,
 203              * but we use the same procedure by multiplying r with scale and truncating to integer.
 204              */
 205             rt               = _mm_mul_ps(r00,gbscale);
 206             gbitab           = _mm_cvttps_epi32(rt);
 207             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 208             gbitab           = _mm_slli_epi32(gbitab,2);
 209             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 210             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 211             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 212             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 213             _MM_TRANSPOSE4_PS(Y,F,G,H);
 214             Heps             = _mm_mul_ps(gbeps,H);
 215             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 216             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 217             vgb              = _mm_mul_ps(gbqqfactor,VV);
 218
 219             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 220             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 221             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 222             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 223             fjptrA           = dvda+jnrA;
 224             fjptrB           = dvda+jnrB;
 225             fjptrC           = dvda+jnrC;
 226             fjptrD           = dvda+jnrD;
 227             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 228             velec            = _mm_mul_ps(qq00,rinv00);
 229             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 230
 231             /* Update potential sum for this i atom from the interaction with this j atom. */
 232             velecsum         = _mm_add_ps(velecsum,velec);
 233             vgbsum           = _mm_add_ps(vgbsum,vgb);
 234
 235             fscal            = felec;
 236
 237             /* Calculate temporary vectorial force */
 238             tx               = _mm_mul_ps(fscal,dx00);
 239             ty               = _mm_mul_ps(fscal,dy00);
 240             tz               = _mm_mul_ps(fscal,dz00);
 241
 242             /* Update vectorial force */
 243             fix0             = _mm_add_ps(fix0,tx);
 244             fiy0             = _mm_add_ps(fiy0,ty);
 245             fiz0             = _mm_add_ps(fiz0,tz);
 246
 247             fjptrA             = f+j_coord_offsetA;
 248             fjptrB             = f+j_coord_offsetB;
 249             fjptrC             = f+j_coord_offsetC;
 250             fjptrD             = f+j_coord_offsetD;
 251             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 252
 253             /* Inner loop uses 58 flops */
 254         }
 255
 256         if(jidx<j_index_end)
 257         {
 258
 259             /* Get j neighbor index, and coordinate index */
 260             jnrlistA         = jjnr[jidx];
 261             jnrlistB         = jjnr[jidx+1];
 262             jnrlistC         = jjnr[jidx+2];
 263             jnrlistD         = jjnr[jidx+3];
 264             /* Sign of each element will be negative for non-real atoms.
 265              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 266              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 267              */
 268             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 269             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 270             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 271             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 272             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 273             j_coord_offsetA  = DIM*jnrA;
 274             j_coord_offsetB  = DIM*jnrB;
 275             j_coord_offsetC  = DIM*jnrC;
 276             j_coord_offsetD  = DIM*jnrD;
 277
 278             /* load j atom coordinates */
 279             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 280                                               x+j_coord_offsetC,x+j_coord_offsetD,
 281                                               &jx0,&jy0,&jz0);
 282
 283             /* Calculate displacement vector */
 284             dx00             = _mm_sub_ps(ix0,jx0);
 285             dy00             = _mm_sub_ps(iy0,jy0);
 286             dz00             = _mm_sub_ps(iz0,jz0);
 287
 288             /* Calculate squared distance and things based on it */
 289             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 290
 291             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 292
 293             /* Load parameters for j particles */
 294             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 295                                                               charge+jnrC+0,charge+jnrD+0);
 296             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 297                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 298
 299             /**************************
 300              * CALCULATE INTERACTIONS *
 301              **************************/
 302
 303             r00              = _mm_mul_ps(rsq00,rinv00);
 304             r00              = _mm_andnot_ps(dummy_mask,r00);
 305
 306             /* Compute parameters for interactions between i and j atoms */
 307             qq00             = _mm_mul_ps(iq0,jq0);
 308
 309             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 310             isaprod          = _mm_mul_ps(isai0,isaj0);
 311             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 312             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 313
 314             /* Calculate generalized born table index - this is a separate table from the normal one,
 315              * but we use the same procedure by multiplying r with scale and truncating to integer.
 316              */
 317             rt               = _mm_mul_ps(r00,gbscale);
 318             gbitab           = _mm_cvttps_epi32(rt);
 319             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 320             gbitab           = _mm_slli_epi32(gbitab,2);
 321             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 322             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 323             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 324             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 325             _MM_TRANSPOSE4_PS(Y,F,G,H);
 326             Heps             = _mm_mul_ps(gbeps,H);
 327             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 328             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 329             vgb              = _mm_mul_ps(gbqqfactor,VV);
 330
 331             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 332             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 333             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 334             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 335             /* 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. */
 336             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 337             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 338             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 339             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 340             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 341             velec            = _mm_mul_ps(qq00,rinv00);
 342             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _mm_andnot_ps(dummy_mask,velec);
 346             velecsum         = _mm_add_ps(velecsum,velec);
 347             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 348             vgbsum           = _mm_add_ps(vgbsum,vgb);
 349
 350             fscal            = felec;
 351
 352             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 353
 354             /* Calculate temporary vectorial force */
 355             tx               = _mm_mul_ps(fscal,dx00);
 356             ty               = _mm_mul_ps(fscal,dy00);
 357             tz               = _mm_mul_ps(fscal,dz00);
 358
 359             /* Update vectorial force */
 360             fix0             = _mm_add_ps(fix0,tx);
 361             fiy0             = _mm_add_ps(fiy0,ty);
 362             fiz0             = _mm_add_ps(fiz0,tz);
 363
 364             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 365             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 366             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 367             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 368             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 369
 370             /* Inner loop uses 59 flops */
 371         }
 372
 373         /* End of innermost loop */
 374
 375         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 376                                               f+i_coord_offset,fshift+i_shift_offset);
 377
 378         ggid                        = gid[iidx];
 379         /* Update potential energies */
 380         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 381         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 382         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 383         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 384
 385         /* Increment number of inner iterations */
 386         inneriter                  += j_index_end - j_index_start;
 387
 388         /* Outer loop uses 9 flops */
 389     }
 390
 391     /* Increment number of outer iterations */
 392     outeriter        += nri;
 393
 394     /* Update outer/inner flops */
 395
 396     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*59);
 397 }
 398 /*
 399  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse4_1_single
 400  * Electrostatics interaction: GeneralizedBorn
 401  * VdW interaction:            None
 402  * Geometry:                   Particle-Particle
 403  * Calculate force/pot:        Force
 404  */
 405 void
 406 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse4_1_single
 407                     (t_nblist * gmx_restrict                nlist,
 408                      rvec * gmx_restrict                    xx,
 409                      rvec * gmx_restrict                    ff,
 410                      t_forcerec * gmx_restrict              fr,
 411                      t_mdatoms * gmx_restrict               mdatoms,
 412                      nb_kernel_data_t * gmx_restrict        kernel_data,
 413                      t_nrnb * gmx_restrict                  nrnb)
 414 {
 415     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 416      * just 0 for non-waters.
 417      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 418      * jnr indices corresponding to data put in the four positions in the SIMD register.
 419      */
 420     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 421     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 422     int              jnrA,jnrB,jnrC,jnrD;
 423     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 424     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 425     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 426     real             rcutoff_scalar;
 427     real             *shiftvec,*fshift,*x,*f;
 428     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 429     real             scratch[4*DIM];
 430     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 431     int              vdwioffset0;
 432     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 433     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 434     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 435     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 436     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 437     real             *charge;
 438     __m128i          gbitab;
 439     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 440     __m128           minushalf = _mm_set1_ps(-0.5);
 441     real             *invsqrta,*dvda,*gbtab;
 442     __m128i          vfitab;
 443     __m128i          ifour       = _mm_set1_epi32(4);
 444     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 445     real             *vftab;
 446     __m128           dummy_mask,cutoff_mask;
 447     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 448     __m128           one     = _mm_set1_ps(1.0);
 449     __m128           two     = _mm_set1_ps(2.0);
 450     x                = xx[0];
 451     f                = ff[0];
 452
 453     nri              = nlist->nri;
 454     iinr             = nlist->iinr;
 455     jindex           = nlist->jindex;
 456     jjnr             = nlist->jjnr;
 457     shiftidx         = nlist->shift;
 458     gid              = nlist->gid;
 459     shiftvec         = fr->shift_vec[0];
 460     fshift           = fr->fshift[0];
 461     facel            = _mm_set1_ps(fr->epsfac);
 462     charge           = mdatoms->chargeA;
 463
 464     invsqrta         = fr->invsqrta;
 465     dvda             = fr->dvda;
 466     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 467     gbtab            = fr->gbtab.data;
 468     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 469
 470     /* Avoid stupid compiler warnings */
 471     jnrA = jnrB = jnrC = jnrD = 0;
 472     j_coord_offsetA = 0;
 473     j_coord_offsetB = 0;
 474     j_coord_offsetC = 0;
 475     j_coord_offsetD = 0;
 476
 477     outeriter        = 0;
 478     inneriter        = 0;
 479
 480     for(iidx=0;iidx<4*DIM;iidx++)
 481     {
 482         scratch[iidx] = 0.0;
 483     }
 484
 485     /* Start outer loop over neighborlists */
 486     for(iidx=0; iidx<nri; iidx++)
 487     {
 488         /* Load shift vector for this list */
 489         i_shift_offset   = DIM*shiftidx[iidx];
 490
 491         /* Load limits for loop over neighbors */
 492         j_index_start    = jindex[iidx];
 493         j_index_end      = jindex[iidx+1];
 494
 495         /* Get outer coordinate index */
 496         inr              = iinr[iidx];
 497         i_coord_offset   = DIM*inr;
 498
 499         /* Load i particle coords and add shift vector */
 500         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 501
 502         fix0             = _mm_setzero_ps();
 503         fiy0             = _mm_setzero_ps();
 504         fiz0             = _mm_setzero_ps();
 505
 506         /* Load parameters for i particles */
 507         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 508         isai0            = _mm_load1_ps(invsqrta+inr+0);
 509
 510         dvdasum          = _mm_setzero_ps();
 511
 512         /* Start inner kernel loop */
 513         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 514         {
 515
 516             /* Get j neighbor index, and coordinate index */
 517             jnrA             = jjnr[jidx];
 518             jnrB             = jjnr[jidx+1];
 519             jnrC             = jjnr[jidx+2];
 520             jnrD             = jjnr[jidx+3];
 521             j_coord_offsetA  = DIM*jnrA;
 522             j_coord_offsetB  = DIM*jnrB;
 523             j_coord_offsetC  = DIM*jnrC;
 524             j_coord_offsetD  = DIM*jnrD;
 525
 526             /* load j atom coordinates */
 527             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 528                                               x+j_coord_offsetC,x+j_coord_offsetD,
 529                                               &jx0,&jy0,&jz0);
 530
 531             /* Calculate displacement vector */
 532             dx00             = _mm_sub_ps(ix0,jx0);
 533             dy00             = _mm_sub_ps(iy0,jy0);
 534             dz00             = _mm_sub_ps(iz0,jz0);
 535
 536             /* Calculate squared distance and things based on it */
 537             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 538
 539             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 540
 541             /* Load parameters for j particles */
 542             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 543                                                               charge+jnrC+0,charge+jnrD+0);
 544             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 545                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 546
 547             /**************************
 548              * CALCULATE INTERACTIONS *
 549              **************************/
 550
 551             r00              = _mm_mul_ps(rsq00,rinv00);
 552
 553             /* Compute parameters for interactions between i and j atoms */
 554             qq00             = _mm_mul_ps(iq0,jq0);
 555
 556             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 557             isaprod          = _mm_mul_ps(isai0,isaj0);
 558             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 559             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 560
 561             /* Calculate generalized born table index - this is a separate table from the normal one,
 562              * but we use the same procedure by multiplying r with scale and truncating to integer.
 563              */
 564             rt               = _mm_mul_ps(r00,gbscale);
 565             gbitab           = _mm_cvttps_epi32(rt);
 566             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 567             gbitab           = _mm_slli_epi32(gbitab,2);
 568             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 569             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 570             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 571             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 572             _MM_TRANSPOSE4_PS(Y,F,G,H);
 573             Heps             = _mm_mul_ps(gbeps,H);
 574             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 575             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 576             vgb              = _mm_mul_ps(gbqqfactor,VV);
 577
 578             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 579             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 580             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 581             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 582             fjptrA           = dvda+jnrA;
 583             fjptrB           = dvda+jnrB;
 584             fjptrC           = dvda+jnrC;
 585             fjptrD           = dvda+jnrD;
 586             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 587             velec            = _mm_mul_ps(qq00,rinv00);
 588             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 589
 590             fscal            = felec;
 591
 592             /* Calculate temporary vectorial force */
 593             tx               = _mm_mul_ps(fscal,dx00);
 594             ty               = _mm_mul_ps(fscal,dy00);
 595             tz               = _mm_mul_ps(fscal,dz00);
 596
 597             /* Update vectorial force */
 598             fix0             = _mm_add_ps(fix0,tx);
 599             fiy0             = _mm_add_ps(fiy0,ty);
 600             fiz0             = _mm_add_ps(fiz0,tz);
 601
 602             fjptrA             = f+j_coord_offsetA;
 603             fjptrB             = f+j_coord_offsetB;
 604             fjptrC             = f+j_coord_offsetC;
 605             fjptrD             = f+j_coord_offsetD;
 606             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 607
 608             /* Inner loop uses 56 flops */
 609         }
 610
 611         if(jidx<j_index_end)
 612         {
 613
 614             /* Get j neighbor index, and coordinate index */
 615             jnrlistA         = jjnr[jidx];
 616             jnrlistB         = jjnr[jidx+1];
 617             jnrlistC         = jjnr[jidx+2];
 618             jnrlistD         = jjnr[jidx+3];
 619             /* Sign of each element will be negative for non-real atoms.
 620              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 621              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 622              */
 623             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 624             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 625             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 626             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 627             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 628             j_coord_offsetA  = DIM*jnrA;
 629             j_coord_offsetB  = DIM*jnrB;
 630             j_coord_offsetC  = DIM*jnrC;
 631             j_coord_offsetD  = DIM*jnrD;
 632
 633             /* load j atom coordinates */
 634             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 635                                               x+j_coord_offsetC,x+j_coord_offsetD,
 636                                               &jx0,&jy0,&jz0);
 637
 638             /* Calculate displacement vector */
 639             dx00             = _mm_sub_ps(ix0,jx0);
 640             dy00             = _mm_sub_ps(iy0,jy0);
 641             dz00             = _mm_sub_ps(iz0,jz0);
 642
 643             /* Calculate squared distance and things based on it */
 644             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 645
 646             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 647
 648             /* Load parameters for j particles */
 649             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 650                                                               charge+jnrC+0,charge+jnrD+0);
 651             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 652                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 653
 654             /**************************
 655              * CALCULATE INTERACTIONS *
 656              **************************/
 657
 658             r00              = _mm_mul_ps(rsq00,rinv00);
 659             r00              = _mm_andnot_ps(dummy_mask,r00);
 660
 661             /* Compute parameters for interactions between i and j atoms */
 662             qq00             = _mm_mul_ps(iq0,jq0);
 663
 664             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 665             isaprod          = _mm_mul_ps(isai0,isaj0);
 666             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 667             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 668
 669             /* Calculate generalized born table index - this is a separate table from the normal one,
 670              * but we use the same procedure by multiplying r with scale and truncating to integer.
 671              */
 672             rt               = _mm_mul_ps(r00,gbscale);
 673             gbitab           = _mm_cvttps_epi32(rt);
 674             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 675             gbitab           = _mm_slli_epi32(gbitab,2);
 676             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 677             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 678             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 679             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 680             _MM_TRANSPOSE4_PS(Y,F,G,H);
 681             Heps             = _mm_mul_ps(gbeps,H);
 682             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 683             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 684             vgb              = _mm_mul_ps(gbqqfactor,VV);
 685
 686             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 687             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 688             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 689             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 690             /* 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. */
 691             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 692             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 693             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 694             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 695             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 696             velec            = _mm_mul_ps(qq00,rinv00);
 697             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 698
 699             fscal            = felec;
 700
 701             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 702
 703             /* Calculate temporary vectorial force */
 704             tx               = _mm_mul_ps(fscal,dx00);
 705             ty               = _mm_mul_ps(fscal,dy00);
 706             tz               = _mm_mul_ps(fscal,dz00);
 707
 708             /* Update vectorial force */
 709             fix0             = _mm_add_ps(fix0,tx);
 710             fiy0             = _mm_add_ps(fiy0,ty);
 711             fiz0             = _mm_add_ps(fiz0,tz);
 712
 713             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 714             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 715             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 716             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 717             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 718
 719             /* Inner loop uses 57 flops */
 720         }
 721
 722         /* End of innermost loop */
 723
 724         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 725                                               f+i_coord_offset,fshift+i_shift_offset);
 726
 727         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 728         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 729
 730         /* Increment number of inner iterations */
 731         inneriter                  += j_index_end - j_index_start;
 732
 733         /* Outer loop uses 7 flops */
 734     }
 735
 736     /* Increment number of outer iterations */
 737     outeriter        += nri;
 738
 739     /* Update outer/inner flops */
 740
 741     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*57);
 742 }