src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecGB_VdwNone_GeomP1P1_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012, by the GROMACS development team, led by
   5  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   6  * others, as listed in the AUTHORS file in the top-level source
   7  * 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_sse2_single.h"
  50 #include "kernelutil_x86_sse2_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_sse2_single
  54  * Electrostatics interaction: GeneralizedBorn
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_sse2_single
  61                     (t_nblist * gmx_restrict                nlist,
  62                      rvec * gmx_restrict                    xx,
  63                      rvec * gmx_restrict                    ff,
  64                      t_forcerec * gmx_restrict              fr,
  65                      t_mdatoms * gmx_restrict               mdatoms,
  66                      nb_kernel_data_t * gmx_restrict        kernel_data,
  67                      t_nrnb * gmx_restrict                  nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128i          gbitab;
  93     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  94     __m128           minushalf = _mm_set1_ps(-0.5);
  95     real             *invsqrta,*dvda,*gbtab;
  96     __m128i          vfitab;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  99     real             *vftab;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117
 118     invsqrta         = fr->invsqrta;
 119     dvda             = fr->dvda;
 120     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 121     gbtab            = fr->gbtab.data;
 122     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 123
 124     /* Avoid stupid compiler warnings */
 125     jnrA = jnrB = jnrC = jnrD = 0;
 126     j_coord_offsetA = 0;
 127     j_coord_offsetB = 0;
 128     j_coord_offsetC = 0;
 129     j_coord_offsetD = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm_setzero_ps();
 157         fiy0             = _mm_setzero_ps();
 158         fiz0             = _mm_setzero_ps();
 159
 160         /* Load parameters for i particles */
 161         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 162         isai0            = _mm_load1_ps(invsqrta+inr+0);
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_ps();
 166         vgbsum           = _mm_setzero_ps();
 167         dvdasum          = _mm_setzero_ps();
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 171         {
 172
 173             /* Get j neighbor index, and coordinate index */
 174             jnrA             = jjnr[jidx];
 175             jnrB             = jjnr[jidx+1];
 176             jnrC             = jjnr[jidx+2];
 177             jnrD             = jjnr[jidx+3];
 178             j_coord_offsetA  = DIM*jnrA;
 179             j_coord_offsetB  = DIM*jnrB;
 180             j_coord_offsetC  = DIM*jnrC;
 181             j_coord_offsetD  = DIM*jnrD;
 182
 183             /* load j atom coordinates */
 184             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 185                                               x+j_coord_offsetC,x+j_coord_offsetD,
 186                                               &jx0,&jy0,&jz0);
 187
 188             /* Calculate displacement vector */
 189             dx00             = _mm_sub_ps(ix0,jx0);
 190             dy00             = _mm_sub_ps(iy0,jy0);
 191             dz00             = _mm_sub_ps(iz0,jz0);
 192
 193             /* Calculate squared distance and things based on it */
 194             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 195
 196             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 197
 198             /* Load parameters for j particles */
 199             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 200                                                               charge+jnrC+0,charge+jnrD+0);
 201             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 202                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 203
 204             /**************************
 205              * CALCULATE INTERACTIONS *
 206              **************************/
 207
 208             r00              = _mm_mul_ps(rsq00,rinv00);
 209
 210             /* Compute parameters for interactions between i and j atoms */
 211             qq00             = _mm_mul_ps(iq0,jq0);
 212
 213             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 214             isaprod          = _mm_mul_ps(isai0,isaj0);
 215             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 216             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 217
 218             /* Calculate generalized born table index - this is a separate table from the normal one,
 219              * but we use the same procedure by multiplying r with scale and truncating to integer.
 220              */
 221             rt               = _mm_mul_ps(r00,gbscale);
 222             gbitab           = _mm_cvttps_epi32(rt);
 223             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 224             gbitab           = _mm_slli_epi32(gbitab,2);
 225
 226             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 227             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 228             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 229             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 230             _MM_TRANSPOSE4_PS(Y,F,G,H);
 231             Heps             = _mm_mul_ps(gbeps,H);
 232             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 233             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 234             vgb              = _mm_mul_ps(gbqqfactor,VV);
 235
 236             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 237             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 238             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 239             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 240             fjptrA           = dvda+jnrA;
 241             fjptrB           = dvda+jnrB;
 242             fjptrC           = dvda+jnrC;
 243             fjptrD           = dvda+jnrD;
 244             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 245             velec            = _mm_mul_ps(qq00,rinv00);
 246             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 247
 248             /* Update potential sum for this i atom from the interaction with this j atom. */
 249             velecsum         = _mm_add_ps(velecsum,velec);
 250             vgbsum           = _mm_add_ps(vgbsum,vgb);
 251
 252             fscal            = felec;
 253
 254             /* Calculate temporary vectorial force */
 255             tx               = _mm_mul_ps(fscal,dx00);
 256             ty               = _mm_mul_ps(fscal,dy00);
 257             tz               = _mm_mul_ps(fscal,dz00);
 258
 259             /* Update vectorial force */
 260             fix0             = _mm_add_ps(fix0,tx);
 261             fiy0             = _mm_add_ps(fiy0,ty);
 262             fiz0             = _mm_add_ps(fiz0,tz);
 263
 264             fjptrA             = f+j_coord_offsetA;
 265             fjptrB             = f+j_coord_offsetB;
 266             fjptrC             = f+j_coord_offsetC;
 267             fjptrD             = f+j_coord_offsetD;
 268             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 269
 270             /* Inner loop uses 58 flops */
 271         }
 272
 273         if(jidx<j_index_end)
 274         {
 275
 276             /* Get j neighbor index, and coordinate index */
 277             jnrlistA         = jjnr[jidx];
 278             jnrlistB         = jjnr[jidx+1];
 279             jnrlistC         = jjnr[jidx+2];
 280             jnrlistD         = jjnr[jidx+3];
 281             /* Sign of each element will be negative for non-real atoms.
 282              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 283              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 284              */
 285             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 286             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 287             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 288             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 289             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 290             j_coord_offsetA  = DIM*jnrA;
 291             j_coord_offsetB  = DIM*jnrB;
 292             j_coord_offsetC  = DIM*jnrC;
 293             j_coord_offsetD  = DIM*jnrD;
 294
 295             /* load j atom coordinates */
 296             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 297                                               x+j_coord_offsetC,x+j_coord_offsetD,
 298                                               &jx0,&jy0,&jz0);
 299
 300             /* Calculate displacement vector */
 301             dx00             = _mm_sub_ps(ix0,jx0);
 302             dy00             = _mm_sub_ps(iy0,jy0);
 303             dz00             = _mm_sub_ps(iz0,jz0);
 304
 305             /* Calculate squared distance and things based on it */
 306             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 307
 308             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 309
 310             /* Load parameters for j particles */
 311             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 312                                                               charge+jnrC+0,charge+jnrD+0);
 313             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 314                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             r00              = _mm_mul_ps(rsq00,rinv00);
 321             r00              = _mm_andnot_ps(dummy_mask,r00);
 322
 323             /* Compute parameters for interactions between i and j atoms */
 324             qq00             = _mm_mul_ps(iq0,jq0);
 325
 326             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 327             isaprod          = _mm_mul_ps(isai0,isaj0);
 328             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 329             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 330
 331             /* Calculate generalized born table index - this is a separate table from the normal one,
 332              * but we use the same procedure by multiplying r with scale and truncating to integer.
 333              */
 334             rt               = _mm_mul_ps(r00,gbscale);
 335             gbitab           = _mm_cvttps_epi32(rt);
 336             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 337             gbitab           = _mm_slli_epi32(gbitab,2);
 338
 339             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 340             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 341             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 342             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 343             _MM_TRANSPOSE4_PS(Y,F,G,H);
 344             Heps             = _mm_mul_ps(gbeps,H);
 345             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 346             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 347             vgb              = _mm_mul_ps(gbqqfactor,VV);
 348
 349             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 350             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 351             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 352             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 353             /* 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. */
 354             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 355             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 356             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 357             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 358             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 359             velec            = _mm_mul_ps(qq00,rinv00);
 360             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velec            = _mm_andnot_ps(dummy_mask,velec);
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 366             vgbsum           = _mm_add_ps(vgbsum,vgb);
 367
 368             fscal            = felec;
 369
 370             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 371
 372             /* Calculate temporary vectorial force */
 373             tx               = _mm_mul_ps(fscal,dx00);
 374             ty               = _mm_mul_ps(fscal,dy00);
 375             tz               = _mm_mul_ps(fscal,dz00);
 376
 377             /* Update vectorial force */
 378             fix0             = _mm_add_ps(fix0,tx);
 379             fiy0             = _mm_add_ps(fiy0,ty);
 380             fiz0             = _mm_add_ps(fiz0,tz);
 381
 382             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 383             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 384             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 385             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 386             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 387
 388             /* Inner loop uses 59 flops */
 389         }
 390
 391         /* End of innermost loop */
 392
 393         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 394                                               f+i_coord_offset,fshift+i_shift_offset);
 395
 396         ggid                        = gid[iidx];
 397         /* Update potential energies */
 398         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 399         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 400         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 401         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 402
 403         /* Increment number of inner iterations */
 404         inneriter                  += j_index_end - j_index_start;
 405
 406         /* Outer loop uses 9 flops */
 407     }
 408
 409     /* Increment number of outer iterations */
 410     outeriter        += nri;
 411
 412     /* Update outer/inner flops */
 413
 414     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*59);
 415 }
 416 /*
 417  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse2_single
 418  * Electrostatics interaction: GeneralizedBorn
 419  * VdW interaction:            None
 420  * Geometry:                   Particle-Particle
 421  * Calculate force/pot:        Force
 422  */
 423 void
 424 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse2_single
 425                     (t_nblist * gmx_restrict                nlist,
 426                      rvec * gmx_restrict                    xx,
 427                      rvec * gmx_restrict                    ff,
 428                      t_forcerec * gmx_restrict              fr,
 429                      t_mdatoms * gmx_restrict               mdatoms,
 430                      nb_kernel_data_t * gmx_restrict        kernel_data,
 431                      t_nrnb * gmx_restrict                  nrnb)
 432 {
 433     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 434      * just 0 for non-waters.
 435      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 436      * jnr indices corresponding to data put in the four positions in the SIMD register.
 437      */
 438     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 439     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 440     int              jnrA,jnrB,jnrC,jnrD;
 441     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 442     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 443     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 444     real             rcutoff_scalar;
 445     real             *shiftvec,*fshift,*x,*f;
 446     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 447     real             scratch[4*DIM];
 448     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 449     int              vdwioffset0;
 450     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 451     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 452     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 453     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 454     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 455     real             *charge;
 456     __m128i          gbitab;
 457     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 458     __m128           minushalf = _mm_set1_ps(-0.5);
 459     real             *invsqrta,*dvda,*gbtab;
 460     __m128i          vfitab;
 461     __m128i          ifour       = _mm_set1_epi32(4);
 462     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 463     real             *vftab;
 464     __m128           dummy_mask,cutoff_mask;
 465     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 466     __m128           one     = _mm_set1_ps(1.0);
 467     __m128           two     = _mm_set1_ps(2.0);
 468     x                = xx[0];
 469     f                = ff[0];
 470
 471     nri              = nlist->nri;
 472     iinr             = nlist->iinr;
 473     jindex           = nlist->jindex;
 474     jjnr             = nlist->jjnr;
 475     shiftidx         = nlist->shift;
 476     gid              = nlist->gid;
 477     shiftvec         = fr->shift_vec[0];
 478     fshift           = fr->fshift[0];
 479     facel            = _mm_set1_ps(fr->epsfac);
 480     charge           = mdatoms->chargeA;
 481
 482     invsqrta         = fr->invsqrta;
 483     dvda             = fr->dvda;
 484     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 485     gbtab            = fr->gbtab.data;
 486     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 487
 488     /* Avoid stupid compiler warnings */
 489     jnrA = jnrB = jnrC = jnrD = 0;
 490     j_coord_offsetA = 0;
 491     j_coord_offsetB = 0;
 492     j_coord_offsetC = 0;
 493     j_coord_offsetD = 0;
 494
 495     outeriter        = 0;
 496     inneriter        = 0;
 497
 498     for(iidx=0;iidx<4*DIM;iidx++)
 499     {
 500         scratch[iidx] = 0.0;
 501     }
 502
 503     /* Start outer loop over neighborlists */
 504     for(iidx=0; iidx<nri; iidx++)
 505     {
 506         /* Load shift vector for this list */
 507         i_shift_offset   = DIM*shiftidx[iidx];
 508
 509         /* Load limits for loop over neighbors */
 510         j_index_start    = jindex[iidx];
 511         j_index_end      = jindex[iidx+1];
 512
 513         /* Get outer coordinate index */
 514         inr              = iinr[iidx];
 515         i_coord_offset   = DIM*inr;
 516
 517         /* Load i particle coords and add shift vector */
 518         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 519
 520         fix0             = _mm_setzero_ps();
 521         fiy0             = _mm_setzero_ps();
 522         fiz0             = _mm_setzero_ps();
 523
 524         /* Load parameters for i particles */
 525         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 526         isai0            = _mm_load1_ps(invsqrta+inr+0);
 527
 528         dvdasum          = _mm_setzero_ps();
 529
 530         /* Start inner kernel loop */
 531         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 532         {
 533
 534             /* Get j neighbor index, and coordinate index */
 535             jnrA             = jjnr[jidx];
 536             jnrB             = jjnr[jidx+1];
 537             jnrC             = jjnr[jidx+2];
 538             jnrD             = jjnr[jidx+3];
 539             j_coord_offsetA  = DIM*jnrA;
 540             j_coord_offsetB  = DIM*jnrB;
 541             j_coord_offsetC  = DIM*jnrC;
 542             j_coord_offsetD  = DIM*jnrD;
 543
 544             /* load j atom coordinates */
 545             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 546                                               x+j_coord_offsetC,x+j_coord_offsetD,
 547                                               &jx0,&jy0,&jz0);
 548
 549             /* Calculate displacement vector */
 550             dx00             = _mm_sub_ps(ix0,jx0);
 551             dy00             = _mm_sub_ps(iy0,jy0);
 552             dz00             = _mm_sub_ps(iz0,jz0);
 553
 554             /* Calculate squared distance and things based on it */
 555             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 556
 557             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 558
 559             /* Load parameters for j particles */
 560             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 561                                                               charge+jnrC+0,charge+jnrD+0);
 562             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 563                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 564
 565             /**************************
 566              * CALCULATE INTERACTIONS *
 567              **************************/
 568
 569             r00              = _mm_mul_ps(rsq00,rinv00);
 570
 571             /* Compute parameters for interactions between i and j atoms */
 572             qq00             = _mm_mul_ps(iq0,jq0);
 573
 574             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 575             isaprod          = _mm_mul_ps(isai0,isaj0);
 576             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 577             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 578
 579             /* Calculate generalized born table index - this is a separate table from the normal one,
 580              * but we use the same procedure by multiplying r with scale and truncating to integer.
 581              */
 582             rt               = _mm_mul_ps(r00,gbscale);
 583             gbitab           = _mm_cvttps_epi32(rt);
 584             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 585             gbitab           = _mm_slli_epi32(gbitab,2);
 586
 587             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 588             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 589             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 590             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 591             _MM_TRANSPOSE4_PS(Y,F,G,H);
 592             Heps             = _mm_mul_ps(gbeps,H);
 593             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 594             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 595             vgb              = _mm_mul_ps(gbqqfactor,VV);
 596
 597             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 598             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 599             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 600             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 601             fjptrA           = dvda+jnrA;
 602             fjptrB           = dvda+jnrB;
 603             fjptrC           = dvda+jnrC;
 604             fjptrD           = dvda+jnrD;
 605             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 606             velec            = _mm_mul_ps(qq00,rinv00);
 607             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 608
 609             fscal            = felec;
 610
 611             /* Calculate temporary vectorial force */
 612             tx               = _mm_mul_ps(fscal,dx00);
 613             ty               = _mm_mul_ps(fscal,dy00);
 614             tz               = _mm_mul_ps(fscal,dz00);
 615
 616             /* Update vectorial force */
 617             fix0             = _mm_add_ps(fix0,tx);
 618             fiy0             = _mm_add_ps(fiy0,ty);
 619             fiz0             = _mm_add_ps(fiz0,tz);
 620
 621             fjptrA             = f+j_coord_offsetA;
 622             fjptrB             = f+j_coord_offsetB;
 623             fjptrC             = f+j_coord_offsetC;
 624             fjptrD             = f+j_coord_offsetD;
 625             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 626
 627             /* Inner loop uses 56 flops */
 628         }
 629
 630         if(jidx<j_index_end)
 631         {
 632
 633             /* Get j neighbor index, and coordinate index */
 634             jnrlistA         = jjnr[jidx];
 635             jnrlistB         = jjnr[jidx+1];
 636             jnrlistC         = jjnr[jidx+2];
 637             jnrlistD         = jjnr[jidx+3];
 638             /* Sign of each element will be negative for non-real atoms.
 639              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 640              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 641              */
 642             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 643             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 644             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 645             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 646             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 647             j_coord_offsetA  = DIM*jnrA;
 648             j_coord_offsetB  = DIM*jnrB;
 649             j_coord_offsetC  = DIM*jnrC;
 650             j_coord_offsetD  = DIM*jnrD;
 651
 652             /* load j atom coordinates */
 653             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 654                                               x+j_coord_offsetC,x+j_coord_offsetD,
 655                                               &jx0,&jy0,&jz0);
 656
 657             /* Calculate displacement vector */
 658             dx00             = _mm_sub_ps(ix0,jx0);
 659             dy00             = _mm_sub_ps(iy0,jy0);
 660             dz00             = _mm_sub_ps(iz0,jz0);
 661
 662             /* Calculate squared distance and things based on it */
 663             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 664
 665             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 666
 667             /* Load parameters for j particles */
 668             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 669                                                               charge+jnrC+0,charge+jnrD+0);
 670             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 671                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 672
 673             /**************************
 674              * CALCULATE INTERACTIONS *
 675              **************************/
 676
 677             r00              = _mm_mul_ps(rsq00,rinv00);
 678             r00              = _mm_andnot_ps(dummy_mask,r00);
 679
 680             /* Compute parameters for interactions between i and j atoms */
 681             qq00             = _mm_mul_ps(iq0,jq0);
 682
 683             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 684             isaprod          = _mm_mul_ps(isai0,isaj0);
 685             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 686             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 687
 688             /* Calculate generalized born table index - this is a separate table from the normal one,
 689              * but we use the same procedure by multiplying r with scale and truncating to integer.
 690              */
 691             rt               = _mm_mul_ps(r00,gbscale);
 692             gbitab           = _mm_cvttps_epi32(rt);
 693             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 694             gbitab           = _mm_slli_epi32(gbitab,2);
 695
 696             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 697             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 698             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 699             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 700             _MM_TRANSPOSE4_PS(Y,F,G,H);
 701             Heps             = _mm_mul_ps(gbeps,H);
 702             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 703             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 704             vgb              = _mm_mul_ps(gbqqfactor,VV);
 705
 706             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 707             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 708             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 709             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 710             /* 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. */
 711             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 712             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 713             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 714             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 715             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 716             velec            = _mm_mul_ps(qq00,rinv00);
 717             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 718
 719             fscal            = felec;
 720
 721             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 722
 723             /* Calculate temporary vectorial force */
 724             tx               = _mm_mul_ps(fscal,dx00);
 725             ty               = _mm_mul_ps(fscal,dy00);
 726             tz               = _mm_mul_ps(fscal,dz00);
 727
 728             /* Update vectorial force */
 729             fix0             = _mm_add_ps(fix0,tx);
 730             fiy0             = _mm_add_ps(fiy0,ty);
 731             fiz0             = _mm_add_ps(fiz0,tz);
 732
 733             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 734             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 735             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 736             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 737             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 738
 739             /* Inner loop uses 57 flops */
 740         }
 741
 742         /* End of innermost loop */
 743
 744         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 745                                               f+i_coord_offset,fshift+i_shift_offset);
 746
 747         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 748         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 749
 750         /* Increment number of inner iterations */
 751         inneriter                  += j_index_end - j_index_start;
 752
 753         /* Outer loop uses 7 flops */
 754     }
 755
 756     /* Increment number of outer iterations */
 757     outeriter        += nri;
 758
 759     /* Update outer/inner flops */
 760
 761     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*57);
 762 }