src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecEwSh_VdwLJSh_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_ElecEwSh_VdwLJSh_GeomP1P1_VF_sse2_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSh_VdwLJSh_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     int              nvdwtype;
  74     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  75     int              *vdwtype;
  76     real             *vdwparam;
  77     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  78     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  79     __m128i          ewitab;
  80     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  81     real             *ewtab;
  82     __m128           dummy_mask,cutoff_mask;
  83     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  84     __m128           one     = _mm_set1_ps(1.0);
  85     __m128           two     = _mm_set1_ps(2.0);
  86     x                = xx[0];
  87     f                = ff[0];
  88
  89     nri              = nlist->nri;
  90     iinr             = nlist->iinr;
  91     jindex           = nlist->jindex;
  92     jjnr             = nlist->jjnr;
  93     shiftidx         = nlist->shift;
  94     gid              = nlist->gid;
  95     shiftvec         = fr->shift_vec[0];
  96     fshift           = fr->fshift[0];
  97     facel            = _mm_set1_ps(fr->epsfac);
  98     charge           = mdatoms->chargeA;
  99     nvdwtype         = fr->ntype;
 100     vdwparam         = fr->nbfp;
 101     vdwtype          = mdatoms->typeA;
 102
 103     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 104     ewtab            = fr->ic->tabq_coul_FDV0;
 105     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 106     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 107
 108     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 109     rcutoff_scalar   = fr->rcoulomb;
 110     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 111     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 112
 113     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 114     rvdw             = _mm_set1_ps(fr->rvdw);
 115
 116     /* Avoid stupid compiler warnings */
 117     jnrA = jnrB = jnrC = jnrD = 0;
 118     j_coord_offsetA = 0;
 119     j_coord_offsetB = 0;
 120     j_coord_offsetC = 0;
 121     j_coord_offsetD = 0;
 122
 123     outeriter        = 0;
 124     inneriter        = 0;
 125
 126     /* Start outer loop over neighborlists */
 127     for(iidx=0; iidx<nri; iidx++)
 128     {
 129         /* Load shift vector for this list */
 130         i_shift_offset   = DIM*shiftidx[iidx];
 131         shX              = shiftvec[i_shift_offset+XX];
 132         shY              = shiftvec[i_shift_offset+YY];
 133         shZ              = shiftvec[i_shift_offset+ZZ];
 134
 135         /* Load limits for loop over neighbors */
 136         j_index_start    = jindex[iidx];
 137         j_index_end      = jindex[iidx+1];
 138
 139         /* Get outer coordinate index */
 140         inr              = iinr[iidx];
 141         i_coord_offset   = DIM*inr;
 142
 143         /* Load i particle coords and add shift vector */
 144         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 145         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 146         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 147
 148         fix0             = _mm_setzero_ps();
 149         fiy0             = _mm_setzero_ps();
 150         fiz0             = _mm_setzero_ps();
 151
 152         /* Load parameters for i particles */
 153         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 154         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 155
 156         /* Reset potential sums */
 157         velecsum         = _mm_setzero_ps();
 158         vvdwsum          = _mm_setzero_ps();
 159
 160         /* Start inner kernel loop */
 161         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 162         {
 163
 164             /* Get j neighbor index, and coordinate index */
 165             jnrA             = jjnr[jidx];
 166             jnrB             = jjnr[jidx+1];
 167             jnrC             = jjnr[jidx+2];
 168             jnrD             = jjnr[jidx+3];
 169
 170             j_coord_offsetA  = DIM*jnrA;
 171             j_coord_offsetB  = DIM*jnrB;
 172             j_coord_offsetC  = DIM*jnrC;
 173             j_coord_offsetD  = DIM*jnrD;
 174
 175             /* load j atom coordinates */
 176             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 177                                               x+j_coord_offsetC,x+j_coord_offsetD,
 178                                               &jx0,&jy0,&jz0);
 179
 180             /* Calculate displacement vector */
 181             dx00             = _mm_sub_ps(ix0,jx0);
 182             dy00             = _mm_sub_ps(iy0,jy0);
 183             dz00             = _mm_sub_ps(iz0,jz0);
 184
 185             /* Calculate squared distance and things based on it */
 186             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 187
 188             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 189
 190             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 191
 192             /* Load parameters for j particles */
 193             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 194                                                               charge+jnrC+0,charge+jnrD+0);
 195             vdwjidx0A        = 2*vdwtype[jnrA+0];
 196             vdwjidx0B        = 2*vdwtype[jnrB+0];
 197             vdwjidx0C        = 2*vdwtype[jnrC+0];
 198             vdwjidx0D        = 2*vdwtype[jnrD+0];
 199
 200             /**************************
 201              * CALCULATE INTERACTIONS *
 202              **************************/
 203
 204             if (gmx_mm_any_lt(rsq00,rcutoff2))
 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             /* EWALD ELECTROSTATICS */
 218
 219             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 220             ewrt             = _mm_mul_ps(r00,ewtabscale);
 221             ewitab           = _mm_cvttps_epi32(ewrt);
 222             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 223             ewitab           = _mm_slli_epi32(ewitab,2);
 224             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 225             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 226             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 227             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 228             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 229             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 230             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 231             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_sub_ps(rinv00,sh_ewald),velec));
 232             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 233
 234             /* LENNARD-JONES DISPERSION/REPULSION */
 235
 236             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 237             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 238             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 239             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 240                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 241             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 242
 243             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 244
 245             /* Update potential sum for this i atom from the interaction with this j atom. */
 246             velec            = _mm_and_ps(velec,cutoff_mask);
 247             velecsum         = _mm_add_ps(velecsum,velec);
 248             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 249             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 250
 251             fscal            = _mm_add_ps(felec,fvdw);
 252
 253             fscal            = _mm_and_ps(fscal,cutoff_mask);
 254
 255             /* Calculate temporary vectorial force */
 256             tx               = _mm_mul_ps(fscal,dx00);
 257             ty               = _mm_mul_ps(fscal,dy00);
 258             tz               = _mm_mul_ps(fscal,dz00);
 259
 260             /* Update vectorial force */
 261             fix0             = _mm_add_ps(fix0,tx);
 262             fiy0             = _mm_add_ps(fiy0,ty);
 263             fiz0             = _mm_add_ps(fiz0,tz);
 264
 265             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 266                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 267                                                    tx,ty,tz);
 268
 269             }
 270
 271             /* Inner loop uses 64 flops */
 272         }
 273
 274         if(jidx<j_index_end)
 275         {
 276
 277             /* Get j neighbor index, and coordinate index */
 278             jnrA             = jjnr[jidx];
 279             jnrB             = jjnr[jidx+1];
 280             jnrC             = jjnr[jidx+2];
 281             jnrD             = jjnr[jidx+3];
 282
 283             /* Sign of each element will be negative for non-real atoms.
 284              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 285              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 286              */
 287             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 288             jnrA       = (jnrA>=0) ? jnrA : 0;
 289             jnrB       = (jnrB>=0) ? jnrB : 0;
 290             jnrC       = (jnrC>=0) ? jnrC : 0;
 291             jnrD       = (jnrD>=0) ? jnrD : 0;
 292
 293             j_coord_offsetA  = DIM*jnrA;
 294             j_coord_offsetB  = DIM*jnrB;
 295             j_coord_offsetC  = DIM*jnrC;
 296             j_coord_offsetD  = DIM*jnrD;
 297
 298             /* load j atom coordinates */
 299             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 300                                               x+j_coord_offsetC,x+j_coord_offsetD,
 301                                               &jx0,&jy0,&jz0);
 302
 303             /* Calculate displacement vector */
 304             dx00             = _mm_sub_ps(ix0,jx0);
 305             dy00             = _mm_sub_ps(iy0,jy0);
 306             dz00             = _mm_sub_ps(iz0,jz0);
 307
 308             /* Calculate squared distance and things based on it */
 309             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 310
 311             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 312
 313             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 314
 315             /* Load parameters for j particles */
 316             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 317                                                               charge+jnrC+0,charge+jnrD+0);
 318             vdwjidx0A        = 2*vdwtype[jnrA+0];
 319             vdwjidx0B        = 2*vdwtype[jnrB+0];
 320             vdwjidx0C        = 2*vdwtype[jnrC+0];
 321             vdwjidx0D        = 2*vdwtype[jnrD+0];
 322
 323             /**************************
 324              * CALCULATE INTERACTIONS *
 325              **************************/
 326
 327             if (gmx_mm_any_lt(rsq00,rcutoff2))
 328             {
 329
 330             r00              = _mm_mul_ps(rsq00,rinv00);
 331             r00              = _mm_andnot_ps(dummy_mask,r00);
 332
 333             /* Compute parameters for interactions between i and j atoms */
 334             qq00             = _mm_mul_ps(iq0,jq0);
 335             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 336                                          vdwparam+vdwioffset0+vdwjidx0B,
 337                                          vdwparam+vdwioffset0+vdwjidx0C,
 338                                          vdwparam+vdwioffset0+vdwjidx0D,
 339                                          &c6_00,&c12_00);
 340
 341             /* EWALD ELECTROSTATICS */
 342
 343             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 344             ewrt             = _mm_mul_ps(r00,ewtabscale);
 345             ewitab           = _mm_cvttps_epi32(ewrt);
 346             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 347             ewitab           = _mm_slli_epi32(ewitab,2);
 348             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 349             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 350             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 351             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 352             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 353             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 354             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 355             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_sub_ps(rinv00,sh_ewald),velec));
 356             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 357
 358             /* LENNARD-JONES DISPERSION/REPULSION */
 359
 360             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 361             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 362             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 363             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 364                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 365             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 366
 367             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm_and_ps(velec,cutoff_mask);
 371             velec            = _mm_andnot_ps(dummy_mask,velec);
 372             velecsum         = _mm_add_ps(velecsum,velec);
 373             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 374             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 375             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 376
 377             fscal            = _mm_add_ps(felec,fvdw);
 378
 379             fscal            = _mm_and_ps(fscal,cutoff_mask);
 380
 381             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 382
 383             /* Calculate temporary vectorial force */
 384             tx               = _mm_mul_ps(fscal,dx00);
 385             ty               = _mm_mul_ps(fscal,dy00);
 386             tz               = _mm_mul_ps(fscal,dz00);
 387
 388             /* Update vectorial force */
 389             fix0             = _mm_add_ps(fix0,tx);
 390             fiy0             = _mm_add_ps(fiy0,ty);
 391             fiz0             = _mm_add_ps(fiz0,tz);
 392
 393             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 394                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 395                                                    tx,ty,tz);
 396
 397             }
 398
 399             /* Inner loop uses 65 flops */
 400         }
 401
 402         /* End of innermost loop */
 403
 404         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 405                                               f+i_coord_offset,fshift+i_shift_offset);
 406
 407         ggid                        = gid[iidx];
 408         /* Update potential energies */
 409         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 410         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 411
 412         /* Increment number of inner iterations */
 413         inneriter                  += j_index_end - j_index_start;
 414
 415         /* Outer loop uses 12 flops */
 416     }
 417
 418     /* Increment number of outer iterations */
 419     outeriter        += nri;
 420
 421     /* Update outer/inner flops */
 422
 423     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*12 + inneriter*65);
 424 }
 425 /*
 426  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_sse2_single
 427  * Electrostatics interaction: Ewald
 428  * VdW interaction:            LennardJones
 429  * Geometry:                   Particle-Particle
 430  * Calculate force/pot:        Force
 431  */
 432 void
 433 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_sse2_single
 434                     (t_nblist * gmx_restrict                nlist,
 435                      rvec * gmx_restrict                    xx,
 436                      rvec * gmx_restrict                    ff,
 437                      t_forcerec * gmx_restrict              fr,
 438                      t_mdatoms * gmx_restrict               mdatoms,
 439                      nb_kernel_data_t * gmx_restrict        kernel_data,
 440                      t_nrnb * gmx_restrict                  nrnb)
 441 {
 442     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 443      * just 0 for non-waters.
 444      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 445      * jnr indices corresponding to data put in the four positions in the SIMD register.
 446      */
 447     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 448     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 449     int              jnrA,jnrB,jnrC,jnrD;
 450     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 451     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 452     real             shX,shY,shZ,rcutoff_scalar;
 453     real             *shiftvec,*fshift,*x,*f;
 454     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 455     int              vdwioffset0;
 456     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 457     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 458     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 459     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 460     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 461     real             *charge;
 462     int              nvdwtype;
 463     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 464     int              *vdwtype;
 465     real             *vdwparam;
 466     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 467     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 468     __m128i          ewitab;
 469     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 470     real             *ewtab;
 471     __m128           dummy_mask,cutoff_mask;
 472     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 473     __m128           one     = _mm_set1_ps(1.0);
 474     __m128           two     = _mm_set1_ps(2.0);
 475     x                = xx[0];
 476     f                = ff[0];
 477
 478     nri              = nlist->nri;
 479     iinr             = nlist->iinr;
 480     jindex           = nlist->jindex;
 481     jjnr             = nlist->jjnr;
 482     shiftidx         = nlist->shift;
 483     gid              = nlist->gid;
 484     shiftvec         = fr->shift_vec[0];
 485     fshift           = fr->fshift[0];
 486     facel            = _mm_set1_ps(fr->epsfac);
 487     charge           = mdatoms->chargeA;
 488     nvdwtype         = fr->ntype;
 489     vdwparam         = fr->nbfp;
 490     vdwtype          = mdatoms->typeA;
 491
 492     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 493     ewtab            = fr->ic->tabq_coul_F;
 494     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 495     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 496
 497     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 498     rcutoff_scalar   = fr->rcoulomb;
 499     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 500     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 501
 502     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 503     rvdw             = _mm_set1_ps(fr->rvdw);
 504
 505     /* Avoid stupid compiler warnings */
 506     jnrA = jnrB = jnrC = jnrD = 0;
 507     j_coord_offsetA = 0;
 508     j_coord_offsetB = 0;
 509     j_coord_offsetC = 0;
 510     j_coord_offsetD = 0;
 511
 512     outeriter        = 0;
 513     inneriter        = 0;
 514
 515     /* Start outer loop over neighborlists */
 516     for(iidx=0; iidx<nri; iidx++)
 517     {
 518         /* Load shift vector for this list */
 519         i_shift_offset   = DIM*shiftidx[iidx];
 520         shX              = shiftvec[i_shift_offset+XX];
 521         shY              = shiftvec[i_shift_offset+YY];
 522         shZ              = shiftvec[i_shift_offset+ZZ];
 523
 524         /* Load limits for loop over neighbors */
 525         j_index_start    = jindex[iidx];
 526         j_index_end      = jindex[iidx+1];
 527
 528         /* Get outer coordinate index */
 529         inr              = iinr[iidx];
 530         i_coord_offset   = DIM*inr;
 531
 532         /* Load i particle coords and add shift vector */
 533         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 534         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 535         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 536
 537         fix0             = _mm_setzero_ps();
 538         fiy0             = _mm_setzero_ps();
 539         fiz0             = _mm_setzero_ps();
 540
 541         /* Load parameters for i particles */
 542         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 543         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 544
 545         /* Start inner kernel loop */
 546         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 547         {
 548
 549             /* Get j neighbor index, and coordinate index */
 550             jnrA             = jjnr[jidx];
 551             jnrB             = jjnr[jidx+1];
 552             jnrC             = jjnr[jidx+2];
 553             jnrD             = jjnr[jidx+3];
 554
 555             j_coord_offsetA  = DIM*jnrA;
 556             j_coord_offsetB  = DIM*jnrB;
 557             j_coord_offsetC  = DIM*jnrC;
 558             j_coord_offsetD  = DIM*jnrD;
 559
 560             /* load j atom coordinates */
 561             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 562                                               x+j_coord_offsetC,x+j_coord_offsetD,
 563                                               &jx0,&jy0,&jz0);
 564
 565             /* Calculate displacement vector */
 566             dx00             = _mm_sub_ps(ix0,jx0);
 567             dy00             = _mm_sub_ps(iy0,jy0);
 568             dz00             = _mm_sub_ps(iz0,jz0);
 569
 570             /* Calculate squared distance and things based on it */
 571             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 572
 573             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 574
 575             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 576
 577             /* Load parameters for j particles */
 578             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 579                                                               charge+jnrC+0,charge+jnrD+0);
 580             vdwjidx0A        = 2*vdwtype[jnrA+0];
 581             vdwjidx0B        = 2*vdwtype[jnrB+0];
 582             vdwjidx0C        = 2*vdwtype[jnrC+0];
 583             vdwjidx0D        = 2*vdwtype[jnrD+0];
 584
 585             /**************************
 586              * CALCULATE INTERACTIONS *
 587              **************************/
 588
 589             if (gmx_mm_any_lt(rsq00,rcutoff2))
 590             {
 591
 592             r00              = _mm_mul_ps(rsq00,rinv00);
 593
 594             /* Compute parameters for interactions between i and j atoms */
 595             qq00             = _mm_mul_ps(iq0,jq0);
 596             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 597                                          vdwparam+vdwioffset0+vdwjidx0B,
 598                                          vdwparam+vdwioffset0+vdwjidx0C,
 599                                          vdwparam+vdwioffset0+vdwjidx0D,
 600                                          &c6_00,&c12_00);
 601
 602             /* EWALD ELECTROSTATICS */
 603
 604             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 605             ewrt             = _mm_mul_ps(r00,ewtabscale);
 606             ewitab           = _mm_cvttps_epi32(ewrt);
 607             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 608             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
 609                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
 610                                          &ewtabF,&ewtabFn);
 611             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
 612             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 613
 614             /* LENNARD-JONES DISPERSION/REPULSION */
 615
 616             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 617             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 618
 619             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 620
 621             fscal            = _mm_add_ps(felec,fvdw);
 622
 623             fscal            = _mm_and_ps(fscal,cutoff_mask);
 624
 625             /* Calculate temporary vectorial force */
 626             tx               = _mm_mul_ps(fscal,dx00);
 627             ty               = _mm_mul_ps(fscal,dy00);
 628             tz               = _mm_mul_ps(fscal,dz00);
 629
 630             /* Update vectorial force */
 631             fix0             = _mm_add_ps(fix0,tx);
 632             fiy0             = _mm_add_ps(fiy0,ty);
 633             fiz0             = _mm_add_ps(fiz0,tz);
 634
 635             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 636                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 637                                                    tx,ty,tz);
 638
 639             }
 640
 641             /* Inner loop uses 46 flops */
 642         }
 643
 644         if(jidx<j_index_end)
 645         {
 646
 647             /* Get j neighbor index, and coordinate index */
 648             jnrA             = jjnr[jidx];
 649             jnrB             = jjnr[jidx+1];
 650             jnrC             = jjnr[jidx+2];
 651             jnrD             = jjnr[jidx+3];
 652
 653             /* Sign of each element will be negative for non-real atoms.
 654              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 655              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 656              */
 657             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 658             jnrA       = (jnrA>=0) ? jnrA : 0;
 659             jnrB       = (jnrB>=0) ? jnrB : 0;
 660             jnrC       = (jnrC>=0) ? jnrC : 0;
 661             jnrD       = (jnrD>=0) ? jnrD : 0;
 662
 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             if (gmx_mm_any_lt(rsq00,rcutoff2))
 698             {
 699
 700             r00              = _mm_mul_ps(rsq00,rinv00);
 701             r00              = _mm_andnot_ps(dummy_mask,r00);
 702
 703             /* Compute parameters for interactions between i and j atoms */
 704             qq00             = _mm_mul_ps(iq0,jq0);
 705             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 706                                          vdwparam+vdwioffset0+vdwjidx0B,
 707                                          vdwparam+vdwioffset0+vdwjidx0C,
 708                                          vdwparam+vdwioffset0+vdwjidx0D,
 709                                          &c6_00,&c12_00);
 710
 711             /* EWALD ELECTROSTATICS */
 712
 713             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 714             ewrt             = _mm_mul_ps(r00,ewtabscale);
 715             ewitab           = _mm_cvttps_epi32(ewrt);
 716             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 717             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
 718                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
 719                                          &ewtabF,&ewtabFn);
 720             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
 721             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 722
 723             /* LENNARD-JONES DISPERSION/REPULSION */
 724
 725             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 726             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 727
 728             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 729
 730             fscal            = _mm_add_ps(felec,fvdw);
 731
 732             fscal            = _mm_and_ps(fscal,cutoff_mask);
 733
 734             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 735
 736             /* Calculate temporary vectorial force */
 737             tx               = _mm_mul_ps(fscal,dx00);
 738             ty               = _mm_mul_ps(fscal,dy00);
 739             tz               = _mm_mul_ps(fscal,dz00);
 740
 741             /* Update vectorial force */
 742             fix0             = _mm_add_ps(fix0,tx);
 743             fiy0             = _mm_add_ps(fiy0,ty);
 744             fiz0             = _mm_add_ps(fiz0,tz);
 745
 746             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 747                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 748                                                    tx,ty,tz);
 749
 750             }
 751
 752             /* Inner loop uses 47 flops */
 753         }
 754
 755         /* End of innermost loop */
 756
 757         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 758                                               f+i_coord_offset,fshift+i_shift_offset);
 759
 760         /* Increment number of inner iterations */
 761         inneriter                  += j_index_end - j_index_start;
 762
 763         /* Outer loop uses 10 flops */
 764     }
 765
 766     /* Increment number of outer iterations */
 767     outeriter        += nri;
 768
 769     /* Update outer/inner flops */
 770
 771     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*10 + inneriter*47);
 772 }