src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_VdwCSTab_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_ElecCoul_VdwCSTab_GeomP1P1_VF_sse2_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwCSTab_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     int              nvdwtype;
  77     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  78     int              *vdwtype;
  79     real             *vdwparam;
  80     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  81     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  82     __m128i          vfitab;
  83     __m128i          ifour       = _mm_set1_epi32(4);
  84     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  85     real             *vftab;
  86     __m128           dummy_mask,cutoff_mask;
  87     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  88     __m128           one     = _mm_set1_ps(1.0);
  89     __m128           two     = _mm_set1_ps(2.0);
  90     x                = xx[0];
  91     f                = ff[0];
  92
  93     nri              = nlist->nri;
  94     iinr             = nlist->iinr;
  95     jindex           = nlist->jindex;
  96     jjnr             = nlist->jjnr;
  97     shiftidx         = nlist->shift;
  98     gid              = nlist->gid;
  99     shiftvec         = fr->shift_vec[0];
 100     fshift           = fr->fshift[0];
 101     facel            = _mm_set1_ps(fr->epsfac);
 102     charge           = mdatoms->chargeA;
 103     nvdwtype         = fr->ntype;
 104     vdwparam         = fr->nbfp;
 105     vdwtype          = mdatoms->typeA;
 106
 107     vftab            = kernel_data->table_vdw->data;
 108     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 109
 110     /* Avoid stupid compiler warnings */
 111     jnrA = jnrB = jnrC = jnrD = 0;
 112     j_coord_offsetA = 0;
 113     j_coord_offsetB = 0;
 114     j_coord_offsetC = 0;
 115     j_coord_offsetD = 0;
 116
 117     outeriter        = 0;
 118     inneriter        = 0;
 119
 120     for(iidx=0;iidx<4*DIM;iidx++)
 121     {
 122         scratch[iidx] = 0.0;
 123     }
 124
 125     /* Start outer loop over neighborlists */
 126     for(iidx=0; iidx<nri; iidx++)
 127     {
 128         /* Load shift vector for this list */
 129         i_shift_offset   = DIM*shiftidx[iidx];
 130
 131         /* Load limits for loop over neighbors */
 132         j_index_start    = jindex[iidx];
 133         j_index_end      = jindex[iidx+1];
 134
 135         /* Get outer coordinate index */
 136         inr              = iinr[iidx];
 137         i_coord_offset   = DIM*inr;
 138
 139         /* Load i particle coords and add shift vector */
 140         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 141
 142         fix0             = _mm_setzero_ps();
 143         fiy0             = _mm_setzero_ps();
 144         fiz0             = _mm_setzero_ps();
 145
 146         /* Load parameters for i particles */
 147         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 148         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 149
 150         /* Reset potential sums */
 151         velecsum         = _mm_setzero_ps();
 152         vvdwsum          = _mm_setzero_ps();
 153
 154         /* Start inner kernel loop */
 155         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 156         {
 157
 158             /* Get j neighbor index, and coordinate index */
 159             jnrA             = jjnr[jidx];
 160             jnrB             = jjnr[jidx+1];
 161             jnrC             = jjnr[jidx+2];
 162             jnrD             = jjnr[jidx+3];
 163             j_coord_offsetA  = DIM*jnrA;
 164             j_coord_offsetB  = DIM*jnrB;
 165             j_coord_offsetC  = DIM*jnrC;
 166             j_coord_offsetD  = DIM*jnrD;
 167
 168             /* load j atom coordinates */
 169             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 170                                               x+j_coord_offsetC,x+j_coord_offsetD,
 171                                               &jx0,&jy0,&jz0);
 172
 173             /* Calculate displacement vector */
 174             dx00             = _mm_sub_ps(ix0,jx0);
 175             dy00             = _mm_sub_ps(iy0,jy0);
 176             dz00             = _mm_sub_ps(iz0,jz0);
 177
 178             /* Calculate squared distance and things based on it */
 179             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 180
 181             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 182
 183             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 184
 185             /* Load parameters for j particles */
 186             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 187                                                               charge+jnrC+0,charge+jnrD+0);
 188             vdwjidx0A        = 2*vdwtype[jnrA+0];
 189             vdwjidx0B        = 2*vdwtype[jnrB+0];
 190             vdwjidx0C        = 2*vdwtype[jnrC+0];
 191             vdwjidx0D        = 2*vdwtype[jnrD+0];
 192
 193             /**************************
 194              * CALCULATE INTERACTIONS *
 195              **************************/
 196
 197             r00              = _mm_mul_ps(rsq00,rinv00);
 198
 199             /* Compute parameters for interactions between i and j atoms */
 200             qq00             = _mm_mul_ps(iq0,jq0);
 201             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 202                                          vdwparam+vdwioffset0+vdwjidx0B,
 203                                          vdwparam+vdwioffset0+vdwjidx0C,
 204                                          vdwparam+vdwioffset0+vdwjidx0D,
 205                                          &c6_00,&c12_00);
 206
 207             /* Calculate table index by multiplying r with table scale and truncate to integer */
 208             rt               = _mm_mul_ps(r00,vftabscale);
 209             vfitab           = _mm_cvttps_epi32(rt);
 210             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 211             vfitab           = _mm_slli_epi32(vfitab,3);
 212
 213             /* COULOMB ELECTROSTATICS */
 214             velec            = _mm_mul_ps(qq00,rinv00);
 215             felec            = _mm_mul_ps(velec,rinvsq00);
 216
 217             /* CUBIC SPLINE TABLE DISPERSION */
 218             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 219             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 220             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 221             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 222             _MM_TRANSPOSE4_PS(Y,F,G,H);
 223             Heps             = _mm_mul_ps(vfeps,H);
 224             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 225             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 226             vvdw6            = _mm_mul_ps(c6_00,VV);
 227             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 228             fvdw6            = _mm_mul_ps(c6_00,FF);
 229
 230             /* CUBIC SPLINE TABLE REPULSION */
 231             vfitab           = _mm_add_epi32(vfitab,ifour);
 232             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 233             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 234             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 235             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 236             _MM_TRANSPOSE4_PS(Y,F,G,H);
 237             Heps             = _mm_mul_ps(vfeps,H);
 238             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 239             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 240             vvdw12           = _mm_mul_ps(c12_00,VV);
 241             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 242             fvdw12           = _mm_mul_ps(c12_00,FF);
 243             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 244             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 245
 246             /* Update potential sum for this i atom from the interaction with this j atom. */
 247             velecsum         = _mm_add_ps(velecsum,velec);
 248             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 249
 250             fscal            = _mm_add_ps(felec,fvdw);
 251
 252             /* Calculate temporary vectorial force */
 253             tx               = _mm_mul_ps(fscal,dx00);
 254             ty               = _mm_mul_ps(fscal,dy00);
 255             tz               = _mm_mul_ps(fscal,dz00);
 256
 257             /* Update vectorial force */
 258             fix0             = _mm_add_ps(fix0,tx);
 259             fiy0             = _mm_add_ps(fiy0,ty);
 260             fiz0             = _mm_add_ps(fiz0,tz);
 261
 262             fjptrA             = f+j_coord_offsetA;
 263             fjptrB             = f+j_coord_offsetB;
 264             fjptrC             = f+j_coord_offsetC;
 265             fjptrD             = f+j_coord_offsetD;
 266             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 267
 268             /* Inner loop uses 63 flops */
 269         }
 270
 271         if(jidx<j_index_end)
 272         {
 273
 274             /* Get j neighbor index, and coordinate index */
 275             jnrlistA         = jjnr[jidx];
 276             jnrlistB         = jjnr[jidx+1];
 277             jnrlistC         = jjnr[jidx+2];
 278             jnrlistD         = jjnr[jidx+3];
 279             /* Sign of each element will be negative for non-real atoms.
 280              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 281              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 282              */
 283             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 284             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 285             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 286             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 287             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 288             j_coord_offsetA  = DIM*jnrA;
 289             j_coord_offsetB  = DIM*jnrB;
 290             j_coord_offsetC  = DIM*jnrC;
 291             j_coord_offsetD  = DIM*jnrD;
 292
 293             /* load j atom coordinates */
 294             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 295                                               x+j_coord_offsetC,x+j_coord_offsetD,
 296                                               &jx0,&jy0,&jz0);
 297
 298             /* Calculate displacement vector */
 299             dx00             = _mm_sub_ps(ix0,jx0);
 300             dy00             = _mm_sub_ps(iy0,jy0);
 301             dz00             = _mm_sub_ps(iz0,jz0);
 302
 303             /* Calculate squared distance and things based on it */
 304             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 305
 306             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 307
 308             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 309
 310             /* Load parameters for j particles */
 311             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 312                                                               charge+jnrC+0,charge+jnrD+0);
 313             vdwjidx0A        = 2*vdwtype[jnrA+0];
 314             vdwjidx0B        = 2*vdwtype[jnrB+0];
 315             vdwjidx0C        = 2*vdwtype[jnrC+0];
 316             vdwjidx0D        = 2*vdwtype[jnrD+0];
 317
 318             /**************************
 319              * CALCULATE INTERACTIONS *
 320              **************************/
 321
 322             r00              = _mm_mul_ps(rsq00,rinv00);
 323             r00              = _mm_andnot_ps(dummy_mask,r00);
 324
 325             /* Compute parameters for interactions between i and j atoms */
 326             qq00             = _mm_mul_ps(iq0,jq0);
 327             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 328                                          vdwparam+vdwioffset0+vdwjidx0B,
 329                                          vdwparam+vdwioffset0+vdwjidx0C,
 330                                          vdwparam+vdwioffset0+vdwjidx0D,
 331                                          &c6_00,&c12_00);
 332
 333             /* Calculate table index by multiplying r with table scale and truncate to integer */
 334             rt               = _mm_mul_ps(r00,vftabscale);
 335             vfitab           = _mm_cvttps_epi32(rt);
 336             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 337             vfitab           = _mm_slli_epi32(vfitab,3);
 338
 339             /* COULOMB ELECTROSTATICS */
 340             velec            = _mm_mul_ps(qq00,rinv00);
 341             felec            = _mm_mul_ps(velec,rinvsq00);
 342
 343             /* CUBIC SPLINE TABLE DISPERSION */
 344             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 345             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 346             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 347             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 348             _MM_TRANSPOSE4_PS(Y,F,G,H);
 349             Heps             = _mm_mul_ps(vfeps,H);
 350             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 351             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 352             vvdw6            = _mm_mul_ps(c6_00,VV);
 353             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 354             fvdw6            = _mm_mul_ps(c6_00,FF);
 355
 356             /* CUBIC SPLINE TABLE REPULSION */
 357             vfitab           = _mm_add_epi32(vfitab,ifour);
 358             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 359             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 360             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 361             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 362             _MM_TRANSPOSE4_PS(Y,F,G,H);
 363             Heps             = _mm_mul_ps(vfeps,H);
 364             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 365             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 366             vvdw12           = _mm_mul_ps(c12_00,VV);
 367             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 368             fvdw12           = _mm_mul_ps(c12_00,FF);
 369             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 370             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velec            = _mm_andnot_ps(dummy_mask,velec);
 374             velecsum         = _mm_add_ps(velecsum,velec);
 375             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 376             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 377
 378             fscal            = _mm_add_ps(felec,fvdw);
 379
 380             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 381
 382             /* Calculate temporary vectorial force */
 383             tx               = _mm_mul_ps(fscal,dx00);
 384             ty               = _mm_mul_ps(fscal,dy00);
 385             tz               = _mm_mul_ps(fscal,dz00);
 386
 387             /* Update vectorial force */
 388             fix0             = _mm_add_ps(fix0,tx);
 389             fiy0             = _mm_add_ps(fiy0,ty);
 390             fiz0             = _mm_add_ps(fiz0,tz);
 391
 392             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 393             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 394             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 395             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 396             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 397
 398             /* Inner loop uses 64 flops */
 399         }
 400
 401         /* End of innermost loop */
 402
 403         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 404                                               f+i_coord_offset,fshift+i_shift_offset);
 405
 406         ggid                        = gid[iidx];
 407         /* Update potential energies */
 408         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 409         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 410
 411         /* Increment number of inner iterations */
 412         inneriter                  += j_index_end - j_index_start;
 413
 414         /* Outer loop uses 9 flops */
 415     }
 416
 417     /* Increment number of outer iterations */
 418     outeriter        += nri;
 419
 420     /* Update outer/inner flops */
 421
 422     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*64);
 423 }
 424 /*
 425  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_sse2_single
 426  * Electrostatics interaction: Coulomb
 427  * VdW interaction:            CubicSplineTable
 428  * Geometry:                   Particle-Particle
 429  * Calculate force/pot:        Force
 430  */
 431 void
 432 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_sse2_single
 433                     (t_nblist * gmx_restrict                nlist,
 434                      rvec * gmx_restrict                    xx,
 435                      rvec * gmx_restrict                    ff,
 436                      t_forcerec * gmx_restrict              fr,
 437                      t_mdatoms * gmx_restrict               mdatoms,
 438                      nb_kernel_data_t * gmx_restrict        kernel_data,
 439                      t_nrnb * gmx_restrict                  nrnb)
 440 {
 441     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 442      * just 0 for non-waters.
 443      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 444      * jnr indices corresponding to data put in the four positions in the SIMD register.
 445      */
 446     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 447     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 448     int              jnrA,jnrB,jnrC,jnrD;
 449     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 450     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 451     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 452     real             rcutoff_scalar;
 453     real             *shiftvec,*fshift,*x,*f;
 454     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 455     real             scratch[4*DIM];
 456     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 457     int              vdwioffset0;
 458     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 459     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 460     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 461     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 462     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 463     real             *charge;
 464     int              nvdwtype;
 465     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 466     int              *vdwtype;
 467     real             *vdwparam;
 468     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 469     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 470     __m128i          vfitab;
 471     __m128i          ifour       = _mm_set1_epi32(4);
 472     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 473     real             *vftab;
 474     __m128           dummy_mask,cutoff_mask;
 475     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 476     __m128           one     = _mm_set1_ps(1.0);
 477     __m128           two     = _mm_set1_ps(2.0);
 478     x                = xx[0];
 479     f                = ff[0];
 480
 481     nri              = nlist->nri;
 482     iinr             = nlist->iinr;
 483     jindex           = nlist->jindex;
 484     jjnr             = nlist->jjnr;
 485     shiftidx         = nlist->shift;
 486     gid              = nlist->gid;
 487     shiftvec         = fr->shift_vec[0];
 488     fshift           = fr->fshift[0];
 489     facel            = _mm_set1_ps(fr->epsfac);
 490     charge           = mdatoms->chargeA;
 491     nvdwtype         = fr->ntype;
 492     vdwparam         = fr->nbfp;
 493     vdwtype          = mdatoms->typeA;
 494
 495     vftab            = kernel_data->table_vdw->data;
 496     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 497
 498     /* Avoid stupid compiler warnings */
 499     jnrA = jnrB = jnrC = jnrD = 0;
 500     j_coord_offsetA = 0;
 501     j_coord_offsetB = 0;
 502     j_coord_offsetC = 0;
 503     j_coord_offsetD = 0;
 504
 505     outeriter        = 0;
 506     inneriter        = 0;
 507
 508     for(iidx=0;iidx<4*DIM;iidx++)
 509     {
 510         scratch[iidx] = 0.0;
 511     }
 512
 513     /* Start outer loop over neighborlists */
 514     for(iidx=0; iidx<nri; iidx++)
 515     {
 516         /* Load shift vector for this list */
 517         i_shift_offset   = DIM*shiftidx[iidx];
 518
 519         /* Load limits for loop over neighbors */
 520         j_index_start    = jindex[iidx];
 521         j_index_end      = jindex[iidx+1];
 522
 523         /* Get outer coordinate index */
 524         inr              = iinr[iidx];
 525         i_coord_offset   = DIM*inr;
 526
 527         /* Load i particle coords and add shift vector */
 528         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 529
 530         fix0             = _mm_setzero_ps();
 531         fiy0             = _mm_setzero_ps();
 532         fiz0             = _mm_setzero_ps();
 533
 534         /* Load parameters for i particles */
 535         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 536         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 537
 538         /* Start inner kernel loop */
 539         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 540         {
 541
 542             /* Get j neighbor index, and coordinate index */
 543             jnrA             = jjnr[jidx];
 544             jnrB             = jjnr[jidx+1];
 545             jnrC             = jjnr[jidx+2];
 546             jnrD             = jjnr[jidx+3];
 547             j_coord_offsetA  = DIM*jnrA;
 548             j_coord_offsetB  = DIM*jnrB;
 549             j_coord_offsetC  = DIM*jnrC;
 550             j_coord_offsetD  = DIM*jnrD;
 551
 552             /* load j atom coordinates */
 553             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 554                                               x+j_coord_offsetC,x+j_coord_offsetD,
 555                                               &jx0,&jy0,&jz0);
 556
 557             /* Calculate displacement vector */
 558             dx00             = _mm_sub_ps(ix0,jx0);
 559             dy00             = _mm_sub_ps(iy0,jy0);
 560             dz00             = _mm_sub_ps(iz0,jz0);
 561
 562             /* Calculate squared distance and things based on it */
 563             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 564
 565             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 566
 567             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 568
 569             /* Load parameters for j particles */
 570             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 571                                                               charge+jnrC+0,charge+jnrD+0);
 572             vdwjidx0A        = 2*vdwtype[jnrA+0];
 573             vdwjidx0B        = 2*vdwtype[jnrB+0];
 574             vdwjidx0C        = 2*vdwtype[jnrC+0];
 575             vdwjidx0D        = 2*vdwtype[jnrD+0];
 576
 577             /**************************
 578              * CALCULATE INTERACTIONS *
 579              **************************/
 580
 581             r00              = _mm_mul_ps(rsq00,rinv00);
 582
 583             /* Compute parameters for interactions between i and j atoms */
 584             qq00             = _mm_mul_ps(iq0,jq0);
 585             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 586                                          vdwparam+vdwioffset0+vdwjidx0B,
 587                                          vdwparam+vdwioffset0+vdwjidx0C,
 588                                          vdwparam+vdwioffset0+vdwjidx0D,
 589                                          &c6_00,&c12_00);
 590
 591             /* Calculate table index by multiplying r with table scale and truncate to integer */
 592             rt               = _mm_mul_ps(r00,vftabscale);
 593             vfitab           = _mm_cvttps_epi32(rt);
 594             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 595             vfitab           = _mm_slli_epi32(vfitab,3);
 596
 597             /* COULOMB ELECTROSTATICS */
 598             velec            = _mm_mul_ps(qq00,rinv00);
 599             felec            = _mm_mul_ps(velec,rinvsq00);
 600
 601             /* CUBIC SPLINE TABLE DISPERSION */
 602             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 603             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 604             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 605             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 606             _MM_TRANSPOSE4_PS(Y,F,G,H);
 607             Heps             = _mm_mul_ps(vfeps,H);
 608             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 609             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 610             fvdw6            = _mm_mul_ps(c6_00,FF);
 611
 612             /* CUBIC SPLINE TABLE REPULSION */
 613             vfitab           = _mm_add_epi32(vfitab,ifour);
 614             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 615             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 616             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 617             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 618             _MM_TRANSPOSE4_PS(Y,F,G,H);
 619             Heps             = _mm_mul_ps(vfeps,H);
 620             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 621             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 622             fvdw12           = _mm_mul_ps(c12_00,FF);
 623             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 624
 625             fscal            = _mm_add_ps(felec,fvdw);
 626
 627             /* Calculate temporary vectorial force */
 628             tx               = _mm_mul_ps(fscal,dx00);
 629             ty               = _mm_mul_ps(fscal,dy00);
 630             tz               = _mm_mul_ps(fscal,dz00);
 631
 632             /* Update vectorial force */
 633             fix0             = _mm_add_ps(fix0,tx);
 634             fiy0             = _mm_add_ps(fiy0,ty);
 635             fiz0             = _mm_add_ps(fiz0,tz);
 636
 637             fjptrA             = f+j_coord_offsetA;
 638             fjptrB             = f+j_coord_offsetB;
 639             fjptrC             = f+j_coord_offsetC;
 640             fjptrD             = f+j_coord_offsetD;
 641             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 642
 643             /* Inner loop uses 54 flops */
 644         }
 645
 646         if(jidx<j_index_end)
 647         {
 648
 649             /* Get j neighbor index, and coordinate index */
 650             jnrlistA         = jjnr[jidx];
 651             jnrlistB         = jjnr[jidx+1];
 652             jnrlistC         = jjnr[jidx+2];
 653             jnrlistD         = jjnr[jidx+3];
 654             /* Sign of each element will be negative for non-real atoms.
 655              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 656              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 657              */
 658             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 659             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 660             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 661             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 662             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 663             j_coord_offsetA  = DIM*jnrA;
 664             j_coord_offsetB  = DIM*jnrB;
 665             j_coord_offsetC  = DIM*jnrC;
 666             j_coord_offsetD  = DIM*jnrD;
 667
 668             /* load j atom coordinates */
 669             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 670                                               x+j_coord_offsetC,x+j_coord_offsetD,
 671                                               &jx0,&jy0,&jz0);
 672
 673             /* Calculate displacement vector */
 674             dx00             = _mm_sub_ps(ix0,jx0);
 675             dy00             = _mm_sub_ps(iy0,jy0);
 676             dz00             = _mm_sub_ps(iz0,jz0);
 677
 678             /* Calculate squared distance and things based on it */
 679             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 680
 681             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 682
 683             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 684
 685             /* Load parameters for j particles */
 686             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 687                                                               charge+jnrC+0,charge+jnrD+0);
 688             vdwjidx0A        = 2*vdwtype[jnrA+0];
 689             vdwjidx0B        = 2*vdwtype[jnrB+0];
 690             vdwjidx0C        = 2*vdwtype[jnrC+0];
 691             vdwjidx0D        = 2*vdwtype[jnrD+0];
 692
 693             /**************************
 694              * CALCULATE INTERACTIONS *
 695              **************************/
 696
 697             r00              = _mm_mul_ps(rsq00,rinv00);
 698             r00              = _mm_andnot_ps(dummy_mask,r00);
 699
 700             /* Compute parameters for interactions between i and j atoms */
 701             qq00             = _mm_mul_ps(iq0,jq0);
 702             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 703                                          vdwparam+vdwioffset0+vdwjidx0B,
 704                                          vdwparam+vdwioffset0+vdwjidx0C,
 705                                          vdwparam+vdwioffset0+vdwjidx0D,
 706                                          &c6_00,&c12_00);
 707
 708             /* Calculate table index by multiplying r with table scale and truncate to integer */
 709             rt               = _mm_mul_ps(r00,vftabscale);
 710             vfitab           = _mm_cvttps_epi32(rt);
 711             vfeps            = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
 712             vfitab           = _mm_slli_epi32(vfitab,3);
 713
 714             /* COULOMB ELECTROSTATICS */
 715             velec            = _mm_mul_ps(qq00,rinv00);
 716             felec            = _mm_mul_ps(velec,rinvsq00);
 717
 718             /* CUBIC SPLINE TABLE DISPERSION */
 719             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 720             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 721             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 722             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 723             _MM_TRANSPOSE4_PS(Y,F,G,H);
 724             Heps             = _mm_mul_ps(vfeps,H);
 725             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 726             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 727             fvdw6            = _mm_mul_ps(c6_00,FF);
 728
 729             /* CUBIC SPLINE TABLE REPULSION */
 730             vfitab           = _mm_add_epi32(vfitab,ifour);
 731             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 732             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 733             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 734             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 735             _MM_TRANSPOSE4_PS(Y,F,G,H);
 736             Heps             = _mm_mul_ps(vfeps,H);
 737             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 738             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 739             fvdw12           = _mm_mul_ps(c12_00,FF);
 740             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 741
 742             fscal            = _mm_add_ps(felec,fvdw);
 743
 744             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 745
 746             /* Calculate temporary vectorial force */
 747             tx               = _mm_mul_ps(fscal,dx00);
 748             ty               = _mm_mul_ps(fscal,dy00);
 749             tz               = _mm_mul_ps(fscal,dz00);
 750
 751             /* Update vectorial force */
 752             fix0             = _mm_add_ps(fix0,tx);
 753             fiy0             = _mm_add_ps(fiy0,ty);
 754             fiz0             = _mm_add_ps(fiz0,tz);
 755
 756             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 757             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 758             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 759             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 760             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 761
 762             /* Inner loop uses 55 flops */
 763         }
 764
 765         /* End of innermost loop */
 766
 767         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 768                                               f+i_coord_offset,fshift+i_shift_offset);
 769
 770         /* Increment number of inner iterations */
 771         inneriter                  += j_index_end - j_index_start;
 772
 773         /* Outer loop uses 7 flops */
 774     }
 775
 776     /* Increment number of outer iterations */
 777     outeriter        += nri;
 778
 779     /* Update outer/inner flops */
 780
 781     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 782 }