src/gromacs/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              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             shX,shY,shZ,rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  69     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  70     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  71     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  72     real             *charge;
  73     __m128i          gbitab;
  74     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,dvdatmp;
  75     __m128           minushalf = _mm_set1_ps(-0.5);
  76     real             *invsqrta,*dvda,*gbtab;
  77     __m128i          vfitab;
  78     __m128i          ifour       = _mm_set1_epi32(4);
  79     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  80     real             *vftab;
  81     __m128           dummy_mask,cutoff_mask;
  82     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  83     __m128           one     = _mm_set1_ps(1.0);
  84     __m128           two     = _mm_set1_ps(2.0);
  85     x                = xx[0];
  86     f                = ff[0];
  87
  88     nri              = nlist->nri;
  89     iinr             = nlist->iinr;
  90     jindex           = nlist->jindex;
  91     jjnr             = nlist->jjnr;
  92     shiftidx         = nlist->shift;
  93     gid              = nlist->gid;
  94     shiftvec         = fr->shift_vec[0];
  95     fshift           = fr->fshift[0];
  96     facel            = _mm_set1_ps(fr->epsfac);
  97     charge           = mdatoms->chargeA;
  98
  99     invsqrta         = fr->invsqrta;
 100     dvda             = fr->dvda;
 101     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 102     gbtab            = fr->gbtab.data;
 103     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 104
 105     /* Avoid stupid compiler warnings */
 106     jnrA = jnrB = jnrC = jnrD = 0;
 107     j_coord_offsetA = 0;
 108     j_coord_offsetB = 0;
 109     j_coord_offsetC = 0;
 110     j_coord_offsetD = 0;
 111
 112     outeriter        = 0;
 113     inneriter        = 0;
 114
 115     /* Start outer loop over neighborlists */
 116     for(iidx=0; iidx<nri; iidx++)
 117     {
 118         /* Load shift vector for this list */
 119         i_shift_offset   = DIM*shiftidx[iidx];
 120         shX              = shiftvec[i_shift_offset+XX];
 121         shY              = shiftvec[i_shift_offset+YY];
 122         shZ              = shiftvec[i_shift_offset+ZZ];
 123
 124         /* Load limits for loop over neighbors */
 125         j_index_start    = jindex[iidx];
 126         j_index_end      = jindex[iidx+1];
 127
 128         /* Get outer coordinate index */
 129         inr              = iinr[iidx];
 130         i_coord_offset   = DIM*inr;
 131
 132         /* Load i particle coords and add shift vector */
 133         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 134         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 135         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 136
 137         fix0             = _mm_setzero_ps();
 138         fiy0             = _mm_setzero_ps();
 139         fiz0             = _mm_setzero_ps();
 140
 141         /* Load parameters for i particles */
 142         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 143         isai0            = _mm_load1_ps(invsqrta+inr+0);
 144
 145         /* Reset potential sums */
 146         velecsum         = _mm_setzero_ps();
 147         vgbsum           = _mm_setzero_ps();
 148         dvdasum          = _mm_setzero_ps();
 149
 150         /* Start inner kernel loop */
 151         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 152         {
 153
 154             /* Get j neighbor index, and coordinate index */
 155             jnrA             = jjnr[jidx];
 156             jnrB             = jjnr[jidx+1];
 157             jnrC             = jjnr[jidx+2];
 158             jnrD             = jjnr[jidx+3];
 159
 160             j_coord_offsetA  = DIM*jnrA;
 161             j_coord_offsetB  = DIM*jnrB;
 162             j_coord_offsetC  = DIM*jnrC;
 163             j_coord_offsetD  = DIM*jnrD;
 164
 165             /* load j atom coordinates */
 166             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 167                                               x+j_coord_offsetC,x+j_coord_offsetD,
 168                                               &jx0,&jy0,&jz0);
 169
 170             /* Calculate displacement vector */
 171             dx00             = _mm_sub_ps(ix0,jx0);
 172             dy00             = _mm_sub_ps(iy0,jy0);
 173             dz00             = _mm_sub_ps(iz0,jz0);
 174
 175             /* Calculate squared distance and things based on it */
 176             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 177
 178             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 179
 180             /* Load parameters for j particles */
 181             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 182                                                               charge+jnrC+0,charge+jnrD+0);
 183             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 184                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 185
 186             /**************************
 187              * CALCULATE INTERACTIONS *
 188              **************************/
 189
 190             r00              = _mm_mul_ps(rsq00,rinv00);
 191
 192             /* Compute parameters for interactions between i and j atoms */
 193             qq00             = _mm_mul_ps(iq0,jq0);
 194
 195             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 196             isaprod          = _mm_mul_ps(isai0,isaj0);
 197             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 198             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 199             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 200
 201             /* Calculate generalized born table index - this is a separate table from the normal one,
 202              * but we use the same procedure by multiplying r with scale and truncating to integer.
 203              */
 204             rt               = _mm_mul_ps(r00,gbscale);
 205             gbitab           = _mm_cvttps_epi32(rt);
 206             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 207             gbitab           = _mm_slli_epi32(gbitab,2);
 208
 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             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 224                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 225             velec            = _mm_mul_ps(qq00,rinv00);
 226             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 227
 228             /* Update potential sum for this i atom from the interaction with this j atom. */
 229             velecsum         = _mm_add_ps(velecsum,velec);
 230             vgbsum           = _mm_add_ps(vgbsum,vgb);
 231
 232             fscal            = felec;
 233
 234             /* Calculate temporary vectorial force */
 235             tx               = _mm_mul_ps(fscal,dx00);
 236             ty               = _mm_mul_ps(fscal,dy00);
 237             tz               = _mm_mul_ps(fscal,dz00);
 238
 239             /* Update vectorial force */
 240             fix0             = _mm_add_ps(fix0,tx);
 241             fiy0             = _mm_add_ps(fiy0,ty);
 242             fiz0             = _mm_add_ps(fiz0,tz);
 243
 244             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 245                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 246                                                    tx,ty,tz);
 247
 248             /* Inner loop uses 58 flops */
 249         }
 250
 251         if(jidx<j_index_end)
 252         {
 253
 254             /* Get j neighbor index, and coordinate index */
 255             jnrA             = jjnr[jidx];
 256             jnrB             = jjnr[jidx+1];
 257             jnrC             = jjnr[jidx+2];
 258             jnrD             = jjnr[jidx+3];
 259
 260             /* Sign of each element will be negative for non-real atoms.
 261              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 262              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 263              */
 264             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 265             jnrA       = (jnrA>=0) ? jnrA : 0;
 266             jnrB       = (jnrB>=0) ? jnrB : 0;
 267             jnrC       = (jnrC>=0) ? jnrC : 0;
 268             jnrD       = (jnrD>=0) ? jnrD : 0;
 269
 270             j_coord_offsetA  = DIM*jnrA;
 271             j_coord_offsetB  = DIM*jnrB;
 272             j_coord_offsetC  = DIM*jnrC;
 273             j_coord_offsetD  = DIM*jnrD;
 274
 275             /* load j atom coordinates */
 276             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 277                                               x+j_coord_offsetC,x+j_coord_offsetD,
 278                                               &jx0,&jy0,&jz0);
 279
 280             /* Calculate displacement vector */
 281             dx00             = _mm_sub_ps(ix0,jx0);
 282             dy00             = _mm_sub_ps(iy0,jy0);
 283             dz00             = _mm_sub_ps(iz0,jz0);
 284
 285             /* Calculate squared distance and things based on it */
 286             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 287
 288             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 289
 290             /* Load parameters for j particles */
 291             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 292                                                               charge+jnrC+0,charge+jnrD+0);
 293             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 294                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 295
 296             /**************************
 297              * CALCULATE INTERACTIONS *
 298              **************************/
 299
 300             r00              = _mm_mul_ps(rsq00,rinv00);
 301             r00              = _mm_andnot_ps(dummy_mask,r00);
 302
 303             /* Compute parameters for interactions between i and j atoms */
 304             qq00             = _mm_mul_ps(iq0,jq0);
 305
 306             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 307             isaprod          = _mm_mul_ps(isai0,isaj0);
 308             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 309             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 310             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 311
 312             /* Calculate generalized born table index - this is a separate table from the normal one,
 313              * but we use the same procedure by multiplying r with scale and truncating to integer.
 314              */
 315             rt               = _mm_mul_ps(r00,gbscale);
 316             gbitab           = _mm_cvttps_epi32(rt);
 317             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 318             gbitab           = _mm_slli_epi32(gbitab,2);
 319
 320             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 321             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 322             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 323             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 324             _MM_TRANSPOSE4_PS(Y,F,G,H);
 325             Heps             = _mm_mul_ps(gbeps,H);
 326             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 327             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 328             vgb              = _mm_mul_ps(gbqqfactor,VV);
 329
 330             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 331             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 332             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 333             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 334             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 335                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 336             velec            = _mm_mul_ps(qq00,rinv00);
 337             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 338
 339             /* Update potential sum for this i atom from the interaction with this j atom. */
 340             velec            = _mm_andnot_ps(dummy_mask,velec);
 341             velecsum         = _mm_add_ps(velecsum,velec);
 342             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 343             vgbsum           = _mm_add_ps(vgbsum,vgb);
 344
 345             fscal            = felec;
 346
 347             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 348
 349             /* Calculate temporary vectorial force */
 350             tx               = _mm_mul_ps(fscal,dx00);
 351             ty               = _mm_mul_ps(fscal,dy00);
 352             tz               = _mm_mul_ps(fscal,dz00);
 353
 354             /* Update vectorial force */
 355             fix0             = _mm_add_ps(fix0,tx);
 356             fiy0             = _mm_add_ps(fiy0,ty);
 357             fiz0             = _mm_add_ps(fiz0,tz);
 358
 359             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 360                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 361                                                    tx,ty,tz);
 362
 363             /* Inner loop uses 59 flops */
 364         }
 365
 366         /* End of innermost loop */
 367
 368         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 369                                               f+i_coord_offset,fshift+i_shift_offset);
 370
 371         ggid                        = gid[iidx];
 372         /* Update potential energies */
 373         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 374         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 375         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 376         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 377
 378         /* Increment number of inner iterations */
 379         inneriter                  += j_index_end - j_index_start;
 380
 381         /* Outer loop uses 12 flops */
 382     }
 383
 384     /* Increment number of outer iterations */
 385     outeriter        += nri;
 386
 387     /* Update outer/inner flops */
 388
 389     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*12 + inneriter*59);
 390 }
 391 /*
 392  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse2_single
 393  * Electrostatics interaction: GeneralizedBorn
 394  * VdW interaction:            None
 395  * Geometry:                   Particle-Particle
 396  * Calculate force/pot:        Force
 397  */
 398 void
 399 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_sse2_single
 400                     (t_nblist * gmx_restrict                nlist,
 401                      rvec * gmx_restrict                    xx,
 402                      rvec * gmx_restrict                    ff,
 403                      t_forcerec * gmx_restrict              fr,
 404                      t_mdatoms * gmx_restrict               mdatoms,
 405                      nb_kernel_data_t * gmx_restrict        kernel_data,
 406                      t_nrnb * gmx_restrict                  nrnb)
 407 {
 408     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 409      * just 0 for non-waters.
 410      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 411      * jnr indices corresponding to data put in the four positions in the SIMD register.
 412      */
 413     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 414     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 415     int              jnrA,jnrB,jnrC,jnrD;
 416     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 417     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 418     real             shX,shY,shZ,rcutoff_scalar;
 419     real             *shiftvec,*fshift,*x,*f;
 420     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 421     int              vdwioffset0;
 422     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 423     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 424     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 425     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 426     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 427     real             *charge;
 428     __m128i          gbitab;
 429     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,dvdatmp;
 430     __m128           minushalf = _mm_set1_ps(-0.5);
 431     real             *invsqrta,*dvda,*gbtab;
 432     __m128i          vfitab;
 433     __m128i          ifour       = _mm_set1_epi32(4);
 434     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 435     real             *vftab;
 436     __m128           dummy_mask,cutoff_mask;
 437     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 438     __m128           one     = _mm_set1_ps(1.0);
 439     __m128           two     = _mm_set1_ps(2.0);
 440     x                = xx[0];
 441     f                = ff[0];
 442
 443     nri              = nlist->nri;
 444     iinr             = nlist->iinr;
 445     jindex           = nlist->jindex;
 446     jjnr             = nlist->jjnr;
 447     shiftidx         = nlist->shift;
 448     gid              = nlist->gid;
 449     shiftvec         = fr->shift_vec[0];
 450     fshift           = fr->fshift[0];
 451     facel            = _mm_set1_ps(fr->epsfac);
 452     charge           = mdatoms->chargeA;
 453
 454     invsqrta         = fr->invsqrta;
 455     dvda             = fr->dvda;
 456     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 457     gbtab            = fr->gbtab.data;
 458     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 459
 460     /* Avoid stupid compiler warnings */
 461     jnrA = jnrB = jnrC = jnrD = 0;
 462     j_coord_offsetA = 0;
 463     j_coord_offsetB = 0;
 464     j_coord_offsetC = 0;
 465     j_coord_offsetD = 0;
 466
 467     outeriter        = 0;
 468     inneriter        = 0;
 469
 470     /* Start outer loop over neighborlists */
 471     for(iidx=0; iidx<nri; iidx++)
 472     {
 473         /* Load shift vector for this list */
 474         i_shift_offset   = DIM*shiftidx[iidx];
 475         shX              = shiftvec[i_shift_offset+XX];
 476         shY              = shiftvec[i_shift_offset+YY];
 477         shZ              = shiftvec[i_shift_offset+ZZ];
 478
 479         /* Load limits for loop over neighbors */
 480         j_index_start    = jindex[iidx];
 481         j_index_end      = jindex[iidx+1];
 482
 483         /* Get outer coordinate index */
 484         inr              = iinr[iidx];
 485         i_coord_offset   = DIM*inr;
 486
 487         /* Load i particle coords and add shift vector */
 488         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 489         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 490         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 491
 492         fix0             = _mm_setzero_ps();
 493         fiy0             = _mm_setzero_ps();
 494         fiz0             = _mm_setzero_ps();
 495
 496         /* Load parameters for i particles */
 497         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 498         isai0            = _mm_load1_ps(invsqrta+inr+0);
 499
 500         dvdasum          = _mm_setzero_ps();
 501
 502         /* Start inner kernel loop */
 503         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 504         {
 505
 506             /* Get j neighbor index, and coordinate index */
 507             jnrA             = jjnr[jidx];
 508             jnrB             = jjnr[jidx+1];
 509             jnrC             = jjnr[jidx+2];
 510             jnrD             = jjnr[jidx+3];
 511
 512             j_coord_offsetA  = DIM*jnrA;
 513             j_coord_offsetB  = DIM*jnrB;
 514             j_coord_offsetC  = DIM*jnrC;
 515             j_coord_offsetD  = DIM*jnrD;
 516
 517             /* load j atom coordinates */
 518             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 519                                               x+j_coord_offsetC,x+j_coord_offsetD,
 520                                               &jx0,&jy0,&jz0);
 521
 522             /* Calculate displacement vector */
 523             dx00             = _mm_sub_ps(ix0,jx0);
 524             dy00             = _mm_sub_ps(iy0,jy0);
 525             dz00             = _mm_sub_ps(iz0,jz0);
 526
 527             /* Calculate squared distance and things based on it */
 528             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 529
 530             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 531
 532             /* Load parameters for j particles */
 533             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 534                                                               charge+jnrC+0,charge+jnrD+0);
 535             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 536                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 537
 538             /**************************
 539              * CALCULATE INTERACTIONS *
 540              **************************/
 541
 542             r00              = _mm_mul_ps(rsq00,rinv00);
 543
 544             /* Compute parameters for interactions between i and j atoms */
 545             qq00             = _mm_mul_ps(iq0,jq0);
 546
 547             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 548             isaprod          = _mm_mul_ps(isai0,isaj0);
 549             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 550             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 551             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 552
 553             /* Calculate generalized born table index - this is a separate table from the normal one,
 554              * but we use the same procedure by multiplying r with scale and truncating to integer.
 555              */
 556             rt               = _mm_mul_ps(r00,gbscale);
 557             gbitab           = _mm_cvttps_epi32(rt);
 558             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 559             gbitab           = _mm_slli_epi32(gbitab,2);
 560
 561             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 562             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 563             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 564             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 565             _MM_TRANSPOSE4_PS(Y,F,G,H);
 566             Heps             = _mm_mul_ps(gbeps,H);
 567             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 568             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 569             vgb              = _mm_mul_ps(gbqqfactor,VV);
 570
 571             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 572             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 573             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 574             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 575             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 576                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 577             velec            = _mm_mul_ps(qq00,rinv00);
 578             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 579
 580             fscal            = felec;
 581
 582             /* Calculate temporary vectorial force */
 583             tx               = _mm_mul_ps(fscal,dx00);
 584             ty               = _mm_mul_ps(fscal,dy00);
 585             tz               = _mm_mul_ps(fscal,dz00);
 586
 587             /* Update vectorial force */
 588             fix0             = _mm_add_ps(fix0,tx);
 589             fiy0             = _mm_add_ps(fiy0,ty);
 590             fiz0             = _mm_add_ps(fiz0,tz);
 591
 592             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 593                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 594                                                    tx,ty,tz);
 595
 596             /* Inner loop uses 56 flops */
 597         }
 598
 599         if(jidx<j_index_end)
 600         {
 601
 602             /* Get j neighbor index, and coordinate index */
 603             jnrA             = jjnr[jidx];
 604             jnrB             = jjnr[jidx+1];
 605             jnrC             = jjnr[jidx+2];
 606             jnrD             = jjnr[jidx+3];
 607
 608             /* Sign of each element will be negative for non-real atoms.
 609              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 610              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 611              */
 612             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 613             jnrA       = (jnrA>=0) ? jnrA : 0;
 614             jnrB       = (jnrB>=0) ? jnrB : 0;
 615             jnrC       = (jnrC>=0) ? jnrC : 0;
 616             jnrD       = (jnrD>=0) ? jnrD : 0;
 617
 618             j_coord_offsetA  = DIM*jnrA;
 619             j_coord_offsetB  = DIM*jnrB;
 620             j_coord_offsetC  = DIM*jnrC;
 621             j_coord_offsetD  = DIM*jnrD;
 622
 623             /* load j atom coordinates */
 624             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 625                                               x+j_coord_offsetC,x+j_coord_offsetD,
 626                                               &jx0,&jy0,&jz0);
 627
 628             /* Calculate displacement vector */
 629             dx00             = _mm_sub_ps(ix0,jx0);
 630             dy00             = _mm_sub_ps(iy0,jy0);
 631             dz00             = _mm_sub_ps(iz0,jz0);
 632
 633             /* Calculate squared distance and things based on it */
 634             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 635
 636             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 637
 638             /* Load parameters for j particles */
 639             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 640                                                               charge+jnrC+0,charge+jnrD+0);
 641             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 642                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 643
 644             /**************************
 645              * CALCULATE INTERACTIONS *
 646              **************************/
 647
 648             r00              = _mm_mul_ps(rsq00,rinv00);
 649             r00              = _mm_andnot_ps(dummy_mask,r00);
 650
 651             /* Compute parameters for interactions between i and j atoms */
 652             qq00             = _mm_mul_ps(iq0,jq0);
 653
 654             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 655             isaprod          = _mm_mul_ps(isai0,isaj0);
 656             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 657             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 658             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 659
 660             /* Calculate generalized born table index - this is a separate table from the normal one,
 661              * but we use the same procedure by multiplying r with scale and truncating to integer.
 662              */
 663             rt               = _mm_mul_ps(r00,gbscale);
 664             gbitab           = _mm_cvttps_epi32(rt);
 665             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 666             gbitab           = _mm_slli_epi32(gbitab,2);
 667
 668             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 669             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 670             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 671             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 672             _MM_TRANSPOSE4_PS(Y,F,G,H);
 673             Heps             = _mm_mul_ps(gbeps,H);
 674             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 675             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 676             vgb              = _mm_mul_ps(gbqqfactor,VV);
 677
 678             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 679             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 680             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 681             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 682             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 683                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 684             velec            = _mm_mul_ps(qq00,rinv00);
 685             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 686
 687             fscal            = felec;
 688
 689             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 690
 691             /* Calculate temporary vectorial force */
 692             tx               = _mm_mul_ps(fscal,dx00);
 693             ty               = _mm_mul_ps(fscal,dy00);
 694             tz               = _mm_mul_ps(fscal,dz00);
 695
 696             /* Update vectorial force */
 697             fix0             = _mm_add_ps(fix0,tx);
 698             fiy0             = _mm_add_ps(fiy0,ty);
 699             fiz0             = _mm_add_ps(fiz0,tz);
 700
 701             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 702                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 703                                                    tx,ty,tz);
 704
 705             /* Inner loop uses 57 flops */
 706         }
 707
 708         /* End of innermost loop */
 709
 710         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 711                                               f+i_coord_offset,fshift+i_shift_offset);
 712
 713         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 714         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 715
 716         /* Increment number of inner iterations */
 717         inneriter                  += j_index_end - j_index_start;
 718
 719         /* Outer loop uses 10 flops */
 720     }
 721
 722     /* Increment number of outer iterations */
 723     outeriter        += nri;
 724
 725     /* Update outer/inner flops */
 726
 727     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*10 + inneriter*57);
 728 }