src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecGB_VdwNone_GeomP1P1_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_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_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_sse2_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_sse2_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_cvtepi32_ps(gbitab));
 208             gbitab           = _mm_slli_epi32(gbitab,2);
 209
 210             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 211             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 212             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 213             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 214             _MM_TRANSPOSE4_PS(Y,F,G,H);
 215             Heps             = _mm_mul_ps(gbeps,H);
 216             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 217             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 218             vgb              = _mm_mul_ps(gbqqfactor,VV);
 219
 220             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 221             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 222             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 223             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 224             fjptrA           = dvda+jnrA;
 225             fjptrB           = dvda+jnrB;
 226             fjptrC           = dvda+jnrC;
 227             fjptrD           = dvda+jnrD;
 228             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 229             velec            = _mm_mul_ps(qq00,rinv00);
 230             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 231
 232             /* Update potential sum for this i atom from the interaction with this j atom. */
 233             velecsum         = _mm_add_ps(velecsum,velec);
 234             vgbsum           = _mm_add_ps(vgbsum,vgb);
 235
 236             fscal            = felec;
 237
 238             /* Calculate temporary vectorial force */
 239             tx               = _mm_mul_ps(fscal,dx00);
 240             ty               = _mm_mul_ps(fscal,dy00);
 241             tz               = _mm_mul_ps(fscal,dz00);
 242
 243             /* Update vectorial force */
 244             fix0             = _mm_add_ps(fix0,tx);
 245             fiy0             = _mm_add_ps(fiy0,ty);
 246             fiz0             = _mm_add_ps(fiz0,tz);
 247
 248             fjptrA             = f+j_coord_offsetA;
 249             fjptrB             = f+j_coord_offsetB;
 250             fjptrC             = f+j_coord_offsetC;
 251             fjptrD             = f+j_coord_offsetD;
 252             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 253
 254             /* Inner loop uses 58 flops */
 255         }
 256
 257         if(jidx<j_index_end)
 258         {
 259
 260             /* Get j neighbor index, and coordinate index */
 261             jnrlistA         = jjnr[jidx];
 262             jnrlistB         = jjnr[jidx+1];
 263             jnrlistC         = jjnr[jidx+2];
 264             jnrlistD         = jjnr[jidx+3];
 265             /* Sign of each element will be negative for non-real atoms.
 266              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 267              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 268              */
 269             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 270             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 271             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 272             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 273             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 274             j_coord_offsetA  = DIM*jnrA;
 275             j_coord_offsetB  = DIM*jnrB;
 276             j_coord_offsetC  = DIM*jnrC;
 277             j_coord_offsetD  = DIM*jnrD;
 278
 279             /* load j atom coordinates */
 280             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 281                                               x+j_coord_offsetC,x+j_coord_offsetD,
 282                                               &jx0,&jy0,&jz0);
 283
 284             /* Calculate displacement vector */
 285             dx00             = _mm_sub_ps(ix0,jx0);
 286             dy00             = _mm_sub_ps(iy0,jy0);
 287             dz00             = _mm_sub_ps(iz0,jz0);
 288
 289             /* Calculate squared distance and things based on it */
 290             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 291
 292             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 293
 294             /* Load parameters for j particles */
 295             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 296                                                               charge+jnrC+0,charge+jnrD+0);
 297             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 298                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 299
 300             /**************************
 301              * CALCULATE INTERACTIONS *
 302              **************************/
 303
 304             r00              = _mm_mul_ps(rsq00,rinv00);
 305             r00              = _mm_andnot_ps(dummy_mask,r00);
 306
 307             /* Compute parameters for interactions between i and j atoms */
 308             qq00             = _mm_mul_ps(iq0,jq0);
 309
 310             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 311             isaprod          = _mm_mul_ps(isai0,isaj0);
 312             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 313             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 314
 315             /* Calculate generalized born table index - this is a separate table from the normal one,
 316              * but we use the same procedure by multiplying r with scale and truncating to integer.
 317              */
 318             rt               = _mm_mul_ps(r00,gbscale);
 319             gbitab           = _mm_cvttps_epi32(rt);
 320             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 321             gbitab           = _mm_slli_epi32(gbitab,2);
 322
 323             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 324             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 325             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 326             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 327             _MM_TRANSPOSE4_PS(Y,F,G,H);
 328             Heps             = _mm_mul_ps(gbeps,H);
 329             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 330             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 331             vgb              = _mm_mul_ps(gbqqfactor,VV);
 332
 333             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 334             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 335             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 336             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 337             /* 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. */
 338             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 339             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 340             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 341             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 342             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 343             velec            = _mm_mul_ps(qq00,rinv00);
 344             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velec            = _mm_andnot_ps(dummy_mask,velec);
 348             velecsum         = _mm_add_ps(velecsum,velec);
 349             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 350             vgbsum           = _mm_add_ps(vgbsum,vgb);
 351
 352             fscal            = felec;
 353
 354             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 355
 356             /* Calculate temporary vectorial force */
 357             tx               = _mm_mul_ps(fscal,dx00);
 358             ty               = _mm_mul_ps(fscal,dy00);
 359             tz               = _mm_mul_ps(fscal,dz00);
 360
 361             /* Update vectorial force */
 362             fix0             = _mm_add_ps(fix0,tx);
 363             fiy0             = _mm_add_ps(fiy0,ty);
 364             fiz0             = _mm_add_ps(fiz0,tz);
 365
 366             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 367             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 368             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 369             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 370             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 371
 372             /* Inner loop uses 59 flops */
 373         }
 374
 375         /* End of innermost loop */
 376
 377         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 378                                               f+i_coord_offset,fshift+i_shift_offset);
 379
 380         ggid                        = gid[iidx];
 381         /* Update potential energies */
 382         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 383         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 384         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 385         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 386
 387         /* Increment number of inner iterations */
 388         inneriter                  += j_index_end - j_index_start;
 389
 390         /* Outer loop uses 9 flops */
 391     }
 392
 393     /* Increment number of outer iterations */
 394     outeriter        += nri;
 395
 396     /* Update outer/inner flops */
 397
 398     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*59);
 399 }
 400 /*
 401  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse2_single
 402  * Electrostatics interaction: GeneralizedBorn
 403  * VdW interaction:            None
 404  * Geometry:                   Particle-Particle
 405  * Calculate force/pot:        Force
 406  */
 407 void
 408 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse2_single
 409                     (t_nblist * gmx_restrict                nlist,
 410                      rvec * gmx_restrict                    xx,
 411                      rvec * gmx_restrict                    ff,
 412                      t_forcerec * gmx_restrict              fr,
 413                      t_mdatoms * gmx_restrict               mdatoms,
 414                      nb_kernel_data_t * gmx_restrict        kernel_data,
 415                      t_nrnb * gmx_restrict                  nrnb)
 416 {
 417     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 418      * just 0 for non-waters.
 419      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 420      * jnr indices corresponding to data put in the four positions in the SIMD register.
 421      */
 422     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 423     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 424     int              jnrA,jnrB,jnrC,jnrD;
 425     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 426     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 427     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 428     real             rcutoff_scalar;
 429     real             *shiftvec,*fshift,*x,*f;
 430     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 431     real             scratch[4*DIM];
 432     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 433     int              vdwioffset0;
 434     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 435     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 436     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 437     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 438     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 439     real             *charge;
 440     __m128i          gbitab;
 441     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 442     __m128           minushalf = _mm_set1_ps(-0.5);
 443     real             *invsqrta,*dvda,*gbtab;
 444     __m128i          vfitab;
 445     __m128i          ifour       = _mm_set1_epi32(4);
 446     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 447     real             *vftab;
 448     __m128           dummy_mask,cutoff_mask;
 449     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 450     __m128           one     = _mm_set1_ps(1.0);
 451     __m128           two     = _mm_set1_ps(2.0);
 452     x                = xx[0];
 453     f                = ff[0];
 454
 455     nri              = nlist->nri;
 456     iinr             = nlist->iinr;
 457     jindex           = nlist->jindex;
 458     jjnr             = nlist->jjnr;
 459     shiftidx         = nlist->shift;
 460     gid              = nlist->gid;
 461     shiftvec         = fr->shift_vec[0];
 462     fshift           = fr->fshift[0];
 463     facel            = _mm_set1_ps(fr->epsfac);
 464     charge           = mdatoms->chargeA;
 465
 466     invsqrta         = fr->invsqrta;
 467     dvda             = fr->dvda;
 468     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 469     gbtab            = fr->gbtab.data;
 470     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 471
 472     /* Avoid stupid compiler warnings */
 473     jnrA = jnrB = jnrC = jnrD = 0;
 474     j_coord_offsetA = 0;
 475     j_coord_offsetB = 0;
 476     j_coord_offsetC = 0;
 477     j_coord_offsetD = 0;
 478
 479     outeriter        = 0;
 480     inneriter        = 0;
 481
 482     for(iidx=0;iidx<4*DIM;iidx++)
 483     {
 484         scratch[iidx] = 0.0;
 485     }
 486
 487     /* Start outer loop over neighborlists */
 488     for(iidx=0; iidx<nri; iidx++)
 489     {
 490         /* Load shift vector for this list */
 491         i_shift_offset   = DIM*shiftidx[iidx];
 492
 493         /* Load limits for loop over neighbors */
 494         j_index_start    = jindex[iidx];
 495         j_index_end      = jindex[iidx+1];
 496
 497         /* Get outer coordinate index */
 498         inr              = iinr[iidx];
 499         i_coord_offset   = DIM*inr;
 500
 501         /* Load i particle coords and add shift vector */
 502         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 503
 504         fix0             = _mm_setzero_ps();
 505         fiy0             = _mm_setzero_ps();
 506         fiz0             = _mm_setzero_ps();
 507
 508         /* Load parameters for i particles */
 509         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 510         isai0            = _mm_load1_ps(invsqrta+inr+0);
 511
 512         dvdasum          = _mm_setzero_ps();
 513
 514         /* Start inner kernel loop */
 515         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 516         {
 517
 518             /* Get j neighbor index, and coordinate index */
 519             jnrA             = jjnr[jidx];
 520             jnrB             = jjnr[jidx+1];
 521             jnrC             = jjnr[jidx+2];
 522             jnrD             = jjnr[jidx+3];
 523             j_coord_offsetA  = DIM*jnrA;
 524             j_coord_offsetB  = DIM*jnrB;
 525             j_coord_offsetC  = DIM*jnrC;
 526             j_coord_offsetD  = DIM*jnrD;
 527
 528             /* load j atom coordinates */
 529             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 530                                               x+j_coord_offsetC,x+j_coord_offsetD,
 531                                               &jx0,&jy0,&jz0);
 532
 533             /* Calculate displacement vector */
 534             dx00             = _mm_sub_ps(ix0,jx0);
 535             dy00             = _mm_sub_ps(iy0,jy0);
 536             dz00             = _mm_sub_ps(iz0,jz0);
 537
 538             /* Calculate squared distance and things based on it */
 539             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 540
 541             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 542
 543             /* Load parameters for j particles */
 544             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 545                                                               charge+jnrC+0,charge+jnrD+0);
 546             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 547                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 548
 549             /**************************
 550              * CALCULATE INTERACTIONS *
 551              **************************/
 552
 553             r00              = _mm_mul_ps(rsq00,rinv00);
 554
 555             /* Compute parameters for interactions between i and j atoms */
 556             qq00             = _mm_mul_ps(iq0,jq0);
 557
 558             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 559             isaprod          = _mm_mul_ps(isai0,isaj0);
 560             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 561             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 562
 563             /* Calculate generalized born table index - this is a separate table from the normal one,
 564              * but we use the same procedure by multiplying r with scale and truncating to integer.
 565              */
 566             rt               = _mm_mul_ps(r00,gbscale);
 567             gbitab           = _mm_cvttps_epi32(rt);
 568             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 569             gbitab           = _mm_slli_epi32(gbitab,2);
 570
 571             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 572             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 573             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 574             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 575             _MM_TRANSPOSE4_PS(Y,F,G,H);
 576             Heps             = _mm_mul_ps(gbeps,H);
 577             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 578             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 579             vgb              = _mm_mul_ps(gbqqfactor,VV);
 580
 581             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 582             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 583             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 584             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 585             fjptrA           = dvda+jnrA;
 586             fjptrB           = dvda+jnrB;
 587             fjptrC           = dvda+jnrC;
 588             fjptrD           = dvda+jnrD;
 589             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 590             velec            = _mm_mul_ps(qq00,rinv00);
 591             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 592
 593             fscal            = felec;
 594
 595             /* Calculate temporary vectorial force */
 596             tx               = _mm_mul_ps(fscal,dx00);
 597             ty               = _mm_mul_ps(fscal,dy00);
 598             tz               = _mm_mul_ps(fscal,dz00);
 599
 600             /* Update vectorial force */
 601             fix0             = _mm_add_ps(fix0,tx);
 602             fiy0             = _mm_add_ps(fiy0,ty);
 603             fiz0             = _mm_add_ps(fiz0,tz);
 604
 605             fjptrA             = f+j_coord_offsetA;
 606             fjptrB             = f+j_coord_offsetB;
 607             fjptrC             = f+j_coord_offsetC;
 608             fjptrD             = f+j_coord_offsetD;
 609             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 610
 611             /* Inner loop uses 56 flops */
 612         }
 613
 614         if(jidx<j_index_end)
 615         {
 616
 617             /* Get j neighbor index, and coordinate index */
 618             jnrlistA         = jjnr[jidx];
 619             jnrlistB         = jjnr[jidx+1];
 620             jnrlistC         = jjnr[jidx+2];
 621             jnrlistD         = jjnr[jidx+3];
 622             /* Sign of each element will be negative for non-real atoms.
 623              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 624              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 625              */
 626             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 627             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 628             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 629             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 630             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 631             j_coord_offsetA  = DIM*jnrA;
 632             j_coord_offsetB  = DIM*jnrB;
 633             j_coord_offsetC  = DIM*jnrC;
 634             j_coord_offsetD  = DIM*jnrD;
 635
 636             /* load j atom coordinates */
 637             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 638                                               x+j_coord_offsetC,x+j_coord_offsetD,
 639                                               &jx0,&jy0,&jz0);
 640
 641             /* Calculate displacement vector */
 642             dx00             = _mm_sub_ps(ix0,jx0);
 643             dy00             = _mm_sub_ps(iy0,jy0);
 644             dz00             = _mm_sub_ps(iz0,jz0);
 645
 646             /* Calculate squared distance and things based on it */
 647             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 648
 649             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 650
 651             /* Load parameters for j particles */
 652             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 653                                                               charge+jnrC+0,charge+jnrD+0);
 654             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 655                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 656
 657             /**************************
 658              * CALCULATE INTERACTIONS *
 659              **************************/
 660
 661             r00              = _mm_mul_ps(rsq00,rinv00);
 662             r00              = _mm_andnot_ps(dummy_mask,r00);
 663
 664             /* Compute parameters for interactions between i and j atoms */
 665             qq00             = _mm_mul_ps(iq0,jq0);
 666
 667             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 668             isaprod          = _mm_mul_ps(isai0,isaj0);
 669             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 670             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 671
 672             /* Calculate generalized born table index - this is a separate table from the normal one,
 673              * but we use the same procedure by multiplying r with scale and truncating to integer.
 674              */
 675             rt               = _mm_mul_ps(r00,gbscale);
 676             gbitab           = _mm_cvttps_epi32(rt);
 677             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 678             gbitab           = _mm_slli_epi32(gbitab,2);
 679
 680             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 681             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 682             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 683             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 684             _MM_TRANSPOSE4_PS(Y,F,G,H);
 685             Heps             = _mm_mul_ps(gbeps,H);
 686             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 687             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 688             vgb              = _mm_mul_ps(gbqqfactor,VV);
 689
 690             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 691             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 692             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 693             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 694             /* 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. */
 695             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 696             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 697             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 698             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 699             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 700             velec            = _mm_mul_ps(qq00,rinv00);
 701             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 702
 703             fscal            = felec;
 704
 705             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 706
 707             /* Calculate temporary vectorial force */
 708             tx               = _mm_mul_ps(fscal,dx00);
 709             ty               = _mm_mul_ps(fscal,dy00);
 710             tz               = _mm_mul_ps(fscal,dz00);
 711
 712             /* Update vectorial force */
 713             fix0             = _mm_add_ps(fix0,tx);
 714             fiy0             = _mm_add_ps(fiy0,ty);
 715             fiz0             = _mm_add_ps(fiz0,tz);
 716
 717             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 718             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 719             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 720             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 721             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 722
 723             /* Inner loop uses 57 flops */
 724         }
 725
 726         /* End of innermost loop */
 727
 728         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 729                                               f+i_coord_offset,fshift+i_shift_offset);
 730
 731         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 732         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 733
 734         /* Increment number of inner iterations */
 735         inneriter                  += j_index_end - j_index_start;
 736
 737         /* Outer loop uses 7 flops */
 738     }
 739
 740     /* Increment number of outer iterations */
 741     outeriter        += nri;
 742
 743     /* Update outer/inner flops */
 744
 745     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*57);
 746 }