src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecGB_VdwLJ_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_VdwLJ_GeomP1P1_VF_sse2_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwLJ_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     int              nvdwtype;
  78     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  79     int              *vdwtype;
  80     real             *vdwparam;
  81     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  82     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  83     __m128i          vfitab;
  84     __m128i          ifour       = _mm_set1_epi32(4);
  85     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  86     real             *vftab;
  87     __m128           dummy_mask,cutoff_mask;
  88     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  89     __m128           one     = _mm_set1_ps(1.0);
  90     __m128           two     = _mm_set1_ps(2.0);
  91     x                = xx[0];
  92     f                = ff[0];
  93
  94     nri              = nlist->nri;
  95     iinr             = nlist->iinr;
  96     jindex           = nlist->jindex;
  97     jjnr             = nlist->jjnr;
  98     shiftidx         = nlist->shift;
  99     gid              = nlist->gid;
 100     shiftvec         = fr->shift_vec[0];
 101     fshift           = fr->fshift[0];
 102     facel            = _mm_set1_ps(fr->epsfac);
 103     charge           = mdatoms->chargeA;
 104     nvdwtype         = fr->ntype;
 105     vdwparam         = fr->nbfp;
 106     vdwtype          = mdatoms->typeA;
 107
 108     invsqrta         = fr->invsqrta;
 109     dvda             = fr->dvda;
 110     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 111     gbtab            = fr->gbtab.data;
 112     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 113
 114     /* Avoid stupid compiler warnings */
 115     jnrA = jnrB = jnrC = jnrD = 0;
 116     j_coord_offsetA = 0;
 117     j_coord_offsetB = 0;
 118     j_coord_offsetC = 0;
 119     j_coord_offsetD = 0;
 120
 121     outeriter        = 0;
 122     inneriter        = 0;
 123
 124     /* Start outer loop over neighborlists */
 125     for(iidx=0; iidx<nri; iidx++)
 126     {
 127         /* Load shift vector for this list */
 128         i_shift_offset   = DIM*shiftidx[iidx];
 129         shX              = shiftvec[i_shift_offset+XX];
 130         shY              = shiftvec[i_shift_offset+YY];
 131         shZ              = shiftvec[i_shift_offset+ZZ];
 132
 133         /* Load limits for loop over neighbors */
 134         j_index_start    = jindex[iidx];
 135         j_index_end      = jindex[iidx+1];
 136
 137         /* Get outer coordinate index */
 138         inr              = iinr[iidx];
 139         i_coord_offset   = DIM*inr;
 140
 141         /* Load i particle coords and add shift vector */
 142         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 143         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 144         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 145
 146         fix0             = _mm_setzero_ps();
 147         fiy0             = _mm_setzero_ps();
 148         fiz0             = _mm_setzero_ps();
 149
 150         /* Load parameters for i particles */
 151         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 152         isai0            = _mm_load1_ps(invsqrta+inr+0);
 153         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 154
 155         /* Reset potential sums */
 156         velecsum         = _mm_setzero_ps();
 157         vgbsum           = _mm_setzero_ps();
 158         vvdwsum          = _mm_setzero_ps();
 159         dvdasum          = _mm_setzero_ps();
 160
 161         /* Start inner kernel loop */
 162         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 163         {
 164
 165             /* Get j neighbor index, and coordinate index */
 166             jnrA             = jjnr[jidx];
 167             jnrB             = jjnr[jidx+1];
 168             jnrC             = jjnr[jidx+2];
 169             jnrD             = jjnr[jidx+3];
 170
 171             j_coord_offsetA  = DIM*jnrA;
 172             j_coord_offsetB  = DIM*jnrB;
 173             j_coord_offsetC  = DIM*jnrC;
 174             j_coord_offsetD  = DIM*jnrD;
 175
 176             /* load j atom coordinates */
 177             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 178                                               x+j_coord_offsetC,x+j_coord_offsetD,
 179                                               &jx0,&jy0,&jz0);
 180
 181             /* Calculate displacement vector */
 182             dx00             = _mm_sub_ps(ix0,jx0);
 183             dy00             = _mm_sub_ps(iy0,jy0);
 184             dz00             = _mm_sub_ps(iz0,jz0);
 185
 186             /* Calculate squared distance and things based on it */
 187             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 188
 189             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 190
 191             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 192
 193             /* Load parameters for j particles */
 194             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 195                                                               charge+jnrC+0,charge+jnrD+0);
 196             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 197                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 198             vdwjidx0A        = 2*vdwtype[jnrA+0];
 199             vdwjidx0B        = 2*vdwtype[jnrB+0];
 200             vdwjidx0C        = 2*vdwtype[jnrC+0];
 201             vdwjidx0D        = 2*vdwtype[jnrD+0];
 202
 203             /**************************
 204              * CALCULATE INTERACTIONS *
 205              **************************/
 206
 207             r00              = _mm_mul_ps(rsq00,rinv00);
 208
 209             /* Compute parameters for interactions between i and j atoms */
 210             qq00             = _mm_mul_ps(iq0,jq0);
 211             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 212                                          vdwparam+vdwioffset0+vdwjidx0B,
 213                                          vdwparam+vdwioffset0+vdwjidx0C,
 214                                          vdwparam+vdwioffset0+vdwjidx0D,
 215                                          &c6_00,&c12_00);
 216
 217             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 218             isaprod          = _mm_mul_ps(isai0,isaj0);
 219             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 220             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 221             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 222
 223             /* Calculate generalized born table index - this is a separate table from the normal one,
 224              * but we use the same procedure by multiplying r with scale and truncating to integer.
 225              */
 226             rt               = _mm_mul_ps(r00,gbscale);
 227             gbitab           = _mm_cvttps_epi32(rt);
 228             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 229             gbitab           = _mm_slli_epi32(gbitab,2);
 230
 231             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 232             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 233             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 234             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 235             _MM_TRANSPOSE4_PS(Y,F,G,H);
 236             Heps             = _mm_mul_ps(gbeps,H);
 237             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 238             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 239             vgb              = _mm_mul_ps(gbqqfactor,VV);
 240
 241             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 242             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 243             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 244             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 245             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 246                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 247             velec            = _mm_mul_ps(qq00,rinv00);
 248             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 249
 250             /* LENNARD-JONES DISPERSION/REPULSION */
 251
 252             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 253             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 254             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 255             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 256             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 257
 258             /* Update potential sum for this i atom from the interaction with this j atom. */
 259             velecsum         = _mm_add_ps(velecsum,velec);
 260             vgbsum           = _mm_add_ps(vgbsum,vgb);
 261             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 262
 263             fscal            = _mm_add_ps(felec,fvdw);
 264
 265             /* Calculate temporary vectorial force */
 266             tx               = _mm_mul_ps(fscal,dx00);
 267             ty               = _mm_mul_ps(fscal,dy00);
 268             tz               = _mm_mul_ps(fscal,dz00);
 269
 270             /* Update vectorial force */
 271             fix0             = _mm_add_ps(fix0,tx);
 272             fiy0             = _mm_add_ps(fiy0,ty);
 273             fiz0             = _mm_add_ps(fiz0,tz);
 274
 275             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 276                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 277                                                    tx,ty,tz);
 278
 279             /* Inner loop uses 71 flops */
 280         }
 281
 282         if(jidx<j_index_end)
 283         {
 284
 285             /* Get j neighbor index, and coordinate index */
 286             jnrA             = jjnr[jidx];
 287             jnrB             = jjnr[jidx+1];
 288             jnrC             = jjnr[jidx+2];
 289             jnrD             = jjnr[jidx+3];
 290
 291             /* Sign of each element will be negative for non-real atoms.
 292              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 293              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 294              */
 295             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 296             jnrA       = (jnrA>=0) ? jnrA : 0;
 297             jnrB       = (jnrB>=0) ? jnrB : 0;
 298             jnrC       = (jnrC>=0) ? jnrC : 0;
 299             jnrD       = (jnrD>=0) ? jnrD : 0;
 300
 301             j_coord_offsetA  = DIM*jnrA;
 302             j_coord_offsetB  = DIM*jnrB;
 303             j_coord_offsetC  = DIM*jnrC;
 304             j_coord_offsetD  = DIM*jnrD;
 305
 306             /* load j atom coordinates */
 307             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 308                                               x+j_coord_offsetC,x+j_coord_offsetD,
 309                                               &jx0,&jy0,&jz0);
 310
 311             /* Calculate displacement vector */
 312             dx00             = _mm_sub_ps(ix0,jx0);
 313             dy00             = _mm_sub_ps(iy0,jy0);
 314             dz00             = _mm_sub_ps(iz0,jz0);
 315
 316             /* Calculate squared distance and things based on it */
 317             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 318
 319             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 320
 321             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 322
 323             /* Load parameters for j particles */
 324             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 325                                                               charge+jnrC+0,charge+jnrD+0);
 326             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 327                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 328             vdwjidx0A        = 2*vdwtype[jnrA+0];
 329             vdwjidx0B        = 2*vdwtype[jnrB+0];
 330             vdwjidx0C        = 2*vdwtype[jnrC+0];
 331             vdwjidx0D        = 2*vdwtype[jnrD+0];
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             r00              = _mm_mul_ps(rsq00,rinv00);
 338             r00              = _mm_andnot_ps(dummy_mask,r00);
 339
 340             /* Compute parameters for interactions between i and j atoms */
 341             qq00             = _mm_mul_ps(iq0,jq0);
 342             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 343                                          vdwparam+vdwioffset0+vdwjidx0B,
 344                                          vdwparam+vdwioffset0+vdwjidx0C,
 345                                          vdwparam+vdwioffset0+vdwjidx0D,
 346                                          &c6_00,&c12_00);
 347
 348             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 349             isaprod          = _mm_mul_ps(isai0,isaj0);
 350             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 351             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 352             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 353
 354             /* Calculate generalized born table index - this is a separate table from the normal one,
 355              * but we use the same procedure by multiplying r with scale and truncating to integer.
 356              */
 357             rt               = _mm_mul_ps(r00,gbscale);
 358             gbitab           = _mm_cvttps_epi32(rt);
 359             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 360             gbitab           = _mm_slli_epi32(gbitab,2);
 361
 362             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 363             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 364             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 365             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 366             _MM_TRANSPOSE4_PS(Y,F,G,H);
 367             Heps             = _mm_mul_ps(gbeps,H);
 368             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 369             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 370             vgb              = _mm_mul_ps(gbqqfactor,VV);
 371
 372             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 373             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 374             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 375             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 376             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 377                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 378             velec            = _mm_mul_ps(qq00,rinv00);
 379             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 380
 381             /* LENNARD-JONES DISPERSION/REPULSION */
 382
 383             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 384             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 385             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 386             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 387             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 388
 389             /* Update potential sum for this i atom from the interaction with this j atom. */
 390             velec            = _mm_andnot_ps(dummy_mask,velec);
 391             velecsum         = _mm_add_ps(velecsum,velec);
 392             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 393             vgbsum           = _mm_add_ps(vgbsum,vgb);
 394             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 395             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 396
 397             fscal            = _mm_add_ps(felec,fvdw);
 398
 399             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 400
 401             /* Calculate temporary vectorial force */
 402             tx               = _mm_mul_ps(fscal,dx00);
 403             ty               = _mm_mul_ps(fscal,dy00);
 404             tz               = _mm_mul_ps(fscal,dz00);
 405
 406             /* Update vectorial force */
 407             fix0             = _mm_add_ps(fix0,tx);
 408             fiy0             = _mm_add_ps(fiy0,ty);
 409             fiz0             = _mm_add_ps(fiz0,tz);
 410
 411             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 412                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 413                                                    tx,ty,tz);
 414
 415             /* Inner loop uses 72 flops */
 416         }
 417
 418         /* End of innermost loop */
 419
 420         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 421                                               f+i_coord_offset,fshift+i_shift_offset);
 422
 423         ggid                        = gid[iidx];
 424         /* Update potential energies */
 425         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 426         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 427         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 428         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 429         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 430
 431         /* Increment number of inner iterations */
 432         inneriter                  += j_index_end - j_index_start;
 433
 434         /* Outer loop uses 13 flops */
 435     }
 436
 437     /* Increment number of outer iterations */
 438     outeriter        += nri;
 439
 440     /* Update outer/inner flops */
 441
 442     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*13 + inneriter*72);
 443 }
 444 /*
 445  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_sse2_single
 446  * Electrostatics interaction: GeneralizedBorn
 447  * VdW interaction:            LennardJones
 448  * Geometry:                   Particle-Particle
 449  * Calculate force/pot:        Force
 450  */
 451 void
 452 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_sse2_single
 453                     (t_nblist * gmx_restrict                nlist,
 454                      rvec * gmx_restrict                    xx,
 455                      rvec * gmx_restrict                    ff,
 456                      t_forcerec * gmx_restrict              fr,
 457                      t_mdatoms * gmx_restrict               mdatoms,
 458                      nb_kernel_data_t * gmx_restrict        kernel_data,
 459                      t_nrnb * gmx_restrict                  nrnb)
 460 {
 461     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 462      * just 0 for non-waters.
 463      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 464      * jnr indices corresponding to data put in the four positions in the SIMD register.
 465      */
 466     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 467     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 468     int              jnrA,jnrB,jnrC,jnrD;
 469     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 470     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 471     real             shX,shY,shZ,rcutoff_scalar;
 472     real             *shiftvec,*fshift,*x,*f;
 473     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 474     int              vdwioffset0;
 475     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 476     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 477     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 478     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 479     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 480     real             *charge;
 481     __m128i          gbitab;
 482     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,dvdatmp;
 483     __m128           minushalf = _mm_set1_ps(-0.5);
 484     real             *invsqrta,*dvda,*gbtab;
 485     int              nvdwtype;
 486     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 487     int              *vdwtype;
 488     real             *vdwparam;
 489     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 490     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 491     __m128i          vfitab;
 492     __m128i          ifour       = _mm_set1_epi32(4);
 493     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 494     real             *vftab;
 495     __m128           dummy_mask,cutoff_mask;
 496     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 497     __m128           one     = _mm_set1_ps(1.0);
 498     __m128           two     = _mm_set1_ps(2.0);
 499     x                = xx[0];
 500     f                = ff[0];
 501
 502     nri              = nlist->nri;
 503     iinr             = nlist->iinr;
 504     jindex           = nlist->jindex;
 505     jjnr             = nlist->jjnr;
 506     shiftidx         = nlist->shift;
 507     gid              = nlist->gid;
 508     shiftvec         = fr->shift_vec[0];
 509     fshift           = fr->fshift[0];
 510     facel            = _mm_set1_ps(fr->epsfac);
 511     charge           = mdatoms->chargeA;
 512     nvdwtype         = fr->ntype;
 513     vdwparam         = fr->nbfp;
 514     vdwtype          = mdatoms->typeA;
 515
 516     invsqrta         = fr->invsqrta;
 517     dvda             = fr->dvda;
 518     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 519     gbtab            = fr->gbtab.data;
 520     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 521
 522     /* Avoid stupid compiler warnings */
 523     jnrA = jnrB = jnrC = jnrD = 0;
 524     j_coord_offsetA = 0;
 525     j_coord_offsetB = 0;
 526     j_coord_offsetC = 0;
 527     j_coord_offsetD = 0;
 528
 529     outeriter        = 0;
 530     inneriter        = 0;
 531
 532     /* Start outer loop over neighborlists */
 533     for(iidx=0; iidx<nri; iidx++)
 534     {
 535         /* Load shift vector for this list */
 536         i_shift_offset   = DIM*shiftidx[iidx];
 537         shX              = shiftvec[i_shift_offset+XX];
 538         shY              = shiftvec[i_shift_offset+YY];
 539         shZ              = shiftvec[i_shift_offset+ZZ];
 540
 541         /* Load limits for loop over neighbors */
 542         j_index_start    = jindex[iidx];
 543         j_index_end      = jindex[iidx+1];
 544
 545         /* Get outer coordinate index */
 546         inr              = iinr[iidx];
 547         i_coord_offset   = DIM*inr;
 548
 549         /* Load i particle coords and add shift vector */
 550         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 551         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 552         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 553
 554         fix0             = _mm_setzero_ps();
 555         fiy0             = _mm_setzero_ps();
 556         fiz0             = _mm_setzero_ps();
 557
 558         /* Load parameters for i particles */
 559         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 560         isai0            = _mm_load1_ps(invsqrta+inr+0);
 561         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 562
 563         dvdasum          = _mm_setzero_ps();
 564
 565         /* Start inner kernel loop */
 566         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 567         {
 568
 569             /* Get j neighbor index, and coordinate index */
 570             jnrA             = jjnr[jidx];
 571             jnrB             = jjnr[jidx+1];
 572             jnrC             = jjnr[jidx+2];
 573             jnrD             = jjnr[jidx+3];
 574
 575             j_coord_offsetA  = DIM*jnrA;
 576             j_coord_offsetB  = DIM*jnrB;
 577             j_coord_offsetC  = DIM*jnrC;
 578             j_coord_offsetD  = DIM*jnrD;
 579
 580             /* load j atom coordinates */
 581             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 582                                               x+j_coord_offsetC,x+j_coord_offsetD,
 583                                               &jx0,&jy0,&jz0);
 584
 585             /* Calculate displacement vector */
 586             dx00             = _mm_sub_ps(ix0,jx0);
 587             dy00             = _mm_sub_ps(iy0,jy0);
 588             dz00             = _mm_sub_ps(iz0,jz0);
 589
 590             /* Calculate squared distance and things based on it */
 591             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 592
 593             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 594
 595             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 596
 597             /* Load parameters for j particles */
 598             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 599                                                               charge+jnrC+0,charge+jnrD+0);
 600             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 601                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 602             vdwjidx0A        = 2*vdwtype[jnrA+0];
 603             vdwjidx0B        = 2*vdwtype[jnrB+0];
 604             vdwjidx0C        = 2*vdwtype[jnrC+0];
 605             vdwjidx0D        = 2*vdwtype[jnrD+0];
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             r00              = _mm_mul_ps(rsq00,rinv00);
 612
 613             /* Compute parameters for interactions between i and j atoms */
 614             qq00             = _mm_mul_ps(iq0,jq0);
 615             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 616                                          vdwparam+vdwioffset0+vdwjidx0B,
 617                                          vdwparam+vdwioffset0+vdwjidx0C,
 618                                          vdwparam+vdwioffset0+vdwjidx0D,
 619                                          &c6_00,&c12_00);
 620
 621             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 622             isaprod          = _mm_mul_ps(isai0,isaj0);
 623             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 624             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 625             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 626
 627             /* Calculate generalized born table index - this is a separate table from the normal one,
 628              * but we use the same procedure by multiplying r with scale and truncating to integer.
 629              */
 630             rt               = _mm_mul_ps(r00,gbscale);
 631             gbitab           = _mm_cvttps_epi32(rt);
 632             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 633             gbitab           = _mm_slli_epi32(gbitab,2);
 634
 635             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 636             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 637             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 638             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 639             _MM_TRANSPOSE4_PS(Y,F,G,H);
 640             Heps             = _mm_mul_ps(gbeps,H);
 641             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 642             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 643             vgb              = _mm_mul_ps(gbqqfactor,VV);
 644
 645             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 646             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 647             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 648             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 649             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 650                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 651             velec            = _mm_mul_ps(qq00,rinv00);
 652             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 653
 654             /* LENNARD-JONES DISPERSION/REPULSION */
 655
 656             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 657             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 658
 659             fscal            = _mm_add_ps(felec,fvdw);
 660
 661             /* Calculate temporary vectorial force */
 662             tx               = _mm_mul_ps(fscal,dx00);
 663             ty               = _mm_mul_ps(fscal,dy00);
 664             tz               = _mm_mul_ps(fscal,dz00);
 665
 666             /* Update vectorial force */
 667             fix0             = _mm_add_ps(fix0,tx);
 668             fiy0             = _mm_add_ps(fiy0,ty);
 669             fiz0             = _mm_add_ps(fiz0,tz);
 670
 671             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 672                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 673                                                    tx,ty,tz);
 674
 675             /* Inner loop uses 64 flops */
 676         }
 677
 678         if(jidx<j_index_end)
 679         {
 680
 681             /* Get j neighbor index, and coordinate index */
 682             jnrA             = jjnr[jidx];
 683             jnrB             = jjnr[jidx+1];
 684             jnrC             = jjnr[jidx+2];
 685             jnrD             = jjnr[jidx+3];
 686
 687             /* Sign of each element will be negative for non-real atoms.
 688              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 689              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 690              */
 691             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 692             jnrA       = (jnrA>=0) ? jnrA : 0;
 693             jnrB       = (jnrB>=0) ? jnrB : 0;
 694             jnrC       = (jnrC>=0) ? jnrC : 0;
 695             jnrD       = (jnrD>=0) ? jnrD : 0;
 696
 697             j_coord_offsetA  = DIM*jnrA;
 698             j_coord_offsetB  = DIM*jnrB;
 699             j_coord_offsetC  = DIM*jnrC;
 700             j_coord_offsetD  = DIM*jnrD;
 701
 702             /* load j atom coordinates */
 703             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 704                                               x+j_coord_offsetC,x+j_coord_offsetD,
 705                                               &jx0,&jy0,&jz0);
 706
 707             /* Calculate displacement vector */
 708             dx00             = _mm_sub_ps(ix0,jx0);
 709             dy00             = _mm_sub_ps(iy0,jy0);
 710             dz00             = _mm_sub_ps(iz0,jz0);
 711
 712             /* Calculate squared distance and things based on it */
 713             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 714
 715             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 716
 717             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 718
 719             /* Load parameters for j particles */
 720             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 721                                                               charge+jnrC+0,charge+jnrD+0);
 722             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 723                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 724             vdwjidx0A        = 2*vdwtype[jnrA+0];
 725             vdwjidx0B        = 2*vdwtype[jnrB+0];
 726             vdwjidx0C        = 2*vdwtype[jnrC+0];
 727             vdwjidx0D        = 2*vdwtype[jnrD+0];
 728
 729             /**************************
 730              * CALCULATE INTERACTIONS *
 731              **************************/
 732
 733             r00              = _mm_mul_ps(rsq00,rinv00);
 734             r00              = _mm_andnot_ps(dummy_mask,r00);
 735
 736             /* Compute parameters for interactions between i and j atoms */
 737             qq00             = _mm_mul_ps(iq0,jq0);
 738             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 739                                          vdwparam+vdwioffset0+vdwjidx0B,
 740                                          vdwparam+vdwioffset0+vdwjidx0C,
 741                                          vdwparam+vdwioffset0+vdwjidx0D,
 742                                          &c6_00,&c12_00);
 743
 744             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 745             isaprod          = _mm_mul_ps(isai0,isaj0);
 746             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 747             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 748             dvdaj            = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+0,dvda+jnrB+0,dvda+jnrC+0,dvda+jnrD+0);
 749
 750             /* Calculate generalized born table index - this is a separate table from the normal one,
 751              * but we use the same procedure by multiplying r with scale and truncating to integer.
 752              */
 753             rt               = _mm_mul_ps(r00,gbscale);
 754             gbitab           = _mm_cvttps_epi32(rt);
 755             gbeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
 756             gbitab           = _mm_slli_epi32(gbitab,2);
 757
 758             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 759             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 760             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 761             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 762             _MM_TRANSPOSE4_PS(Y,F,G,H);
 763             Heps             = _mm_mul_ps(gbeps,H);
 764             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 765             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 766             vgb              = _mm_mul_ps(gbqqfactor,VV);
 767
 768             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 769             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 770             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 771             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 772             gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
 773                                           _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 774             velec            = _mm_mul_ps(qq00,rinv00);
 775             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 776
 777             /* LENNARD-JONES DISPERSION/REPULSION */
 778
 779             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 780             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 781
 782             fscal            = _mm_add_ps(felec,fvdw);
 783
 784             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 785
 786             /* Calculate temporary vectorial force */
 787             tx               = _mm_mul_ps(fscal,dx00);
 788             ty               = _mm_mul_ps(fscal,dy00);
 789             tz               = _mm_mul_ps(fscal,dz00);
 790
 791             /* Update vectorial force */
 792             fix0             = _mm_add_ps(fix0,tx);
 793             fiy0             = _mm_add_ps(fiy0,ty);
 794             fiz0             = _mm_add_ps(fiz0,tz);
 795
 796             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 797                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 798                                                    tx,ty,tz);
 799
 800             /* Inner loop uses 65 flops */
 801         }
 802
 803         /* End of innermost loop */
 804
 805         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 806                                               f+i_coord_offset,fshift+i_shift_offset);
 807
 808         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 809         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 810
 811         /* Increment number of inner iterations */
 812         inneriter                  += j_index_end - j_index_start;
 813
 814         /* Outer loop uses 10 flops */
 815     }
 816
 817     /* Increment number of outer iterations */
 818     outeriter        += nri;
 819
 820     /* Update outer/inner flops */
 821
 822     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*10 + inneriter*65);
 823 }