src/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_avx_256_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_double 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_avx_256_double.h"
  34 #include "kernelutil_x86_avx_256_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
  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 AVX, 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              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  63     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  64     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  65     real             rcutoff_scalar;
  66     real             *shiftvec,*fshift,*x,*f;
  67     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  68     real             scratch[4*DIM];
  69     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  70     real *           vdwioffsetptr0;
  71     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  73     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  75     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  76     real             *charge;
  77     int              nvdwtype;
  78     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  79     int              *vdwtype;
  80     real             *vdwparam;
  81     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  82     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  83     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  84     real             rswitch_scalar,d_scalar;
  85     __m256d          dummy_mask,cutoff_mask;
  86     __m128           tmpmask0,tmpmask1;
  87     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  88     __m256d          one     = _mm256_set1_pd(1.0);
  89     __m256d          two     = _mm256_set1_pd(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            = _mm256_set1_pd(fr->epsfac);
 102     charge           = mdatoms->chargeA;
 103     krf              = _mm256_set1_pd(fr->ic->k_rf);
 104     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 105     crf              = _mm256_set1_pd(fr->ic->c_rf);
 106     nvdwtype         = fr->ntype;
 107     vdwparam         = fr->nbfp;
 108     vdwtype          = mdatoms->typeA;
 109
 110     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 111     rcutoff_scalar   = fr->rcoulomb;
 112     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 113     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 114
 115     rswitch_scalar   = fr->rvdw_switch;
 116     rswitch          = _mm256_set1_pd(rswitch_scalar);
 117     /* Setup switch parameters */
 118     d_scalar         = rcutoff_scalar-rswitch_scalar;
 119     d                = _mm256_set1_pd(d_scalar);
 120     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 121     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 122     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 123     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 124     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 125     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 126
 127     /* Avoid stupid compiler warnings */
 128     jnrA = jnrB = jnrC = jnrD = 0;
 129     j_coord_offsetA = 0;
 130     j_coord_offsetB = 0;
 131     j_coord_offsetC = 0;
 132     j_coord_offsetD = 0;
 133
 134     outeriter        = 0;
 135     inneriter        = 0;
 136
 137     for(iidx=0;iidx<4*DIM;iidx++)
 138     {
 139         scratch[iidx] = 0.0;
 140     }
 141
 142     /* Start outer loop over neighborlists */
 143     for(iidx=0; iidx<nri; iidx++)
 144     {
 145         /* Load shift vector for this list */
 146         i_shift_offset   = DIM*shiftidx[iidx];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 158
 159         fix0             = _mm256_setzero_pd();
 160         fiy0             = _mm256_setzero_pd();
 161         fiz0             = _mm256_setzero_pd();
 162
 163         /* Load parameters for i particles */
 164         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 165         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 166
 167         /* Reset potential sums */
 168         velecsum         = _mm256_setzero_pd();
 169         vvdwsum          = _mm256_setzero_pd();
 170
 171         /* Start inner kernel loop */
 172         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 173         {
 174
 175             /* Get j neighbor index, and coordinate index */
 176             jnrA             = jjnr[jidx];
 177             jnrB             = jjnr[jidx+1];
 178             jnrC             = jjnr[jidx+2];
 179             jnrD             = jjnr[jidx+3];
 180             j_coord_offsetA  = DIM*jnrA;
 181             j_coord_offsetB  = DIM*jnrB;
 182             j_coord_offsetC  = DIM*jnrC;
 183             j_coord_offsetD  = DIM*jnrD;
 184
 185             /* load j atom coordinates */
 186             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 187                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 188                                                  &jx0,&jy0,&jz0);
 189
 190             /* Calculate displacement vector */
 191             dx00             = _mm256_sub_pd(ix0,jx0);
 192             dy00             = _mm256_sub_pd(iy0,jy0);
 193             dz00             = _mm256_sub_pd(iz0,jz0);
 194
 195             /* Calculate squared distance and things based on it */
 196             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 197
 198             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 199
 200             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 201
 202             /* Load parameters for j particles */
 203             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 204                                                                  charge+jnrC+0,charge+jnrD+0);
 205             vdwjidx0A        = 2*vdwtype[jnrA+0];
 206             vdwjidx0B        = 2*vdwtype[jnrB+0];
 207             vdwjidx0C        = 2*vdwtype[jnrC+0];
 208             vdwjidx0D        = 2*vdwtype[jnrD+0];
 209
 210             /**************************
 211              * CALCULATE INTERACTIONS *
 212              **************************/
 213
 214             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 215             {
 216
 217             r00              = _mm256_mul_pd(rsq00,rinv00);
 218
 219             /* Compute parameters for interactions between i and j atoms */
 220             qq00             = _mm256_mul_pd(iq0,jq0);
 221             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 222                                             vdwioffsetptr0+vdwjidx0B,
 223                                             vdwioffsetptr0+vdwjidx0C,
 224                                             vdwioffsetptr0+vdwjidx0D,
 225                                             &c6_00,&c12_00);
 226
 227             /* REACTION-FIELD ELECTROSTATICS */
 228             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 229             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 230
 231             /* LENNARD-JONES DISPERSION/REPULSION */
 232
 233             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 234             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 235             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 236             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 237             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 238
 239             d                = _mm256_sub_pd(r00,rswitch);
 240             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 241             d2               = _mm256_mul_pd(d,d);
 242             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 243
 244             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 245
 246             /* Evaluate switch function */
 247             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 248             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 249             vvdw             = _mm256_mul_pd(vvdw,sw);
 250             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 251
 252             /* Update potential sum for this i atom from the interaction with this j atom. */
 253             velec            = _mm256_and_pd(velec,cutoff_mask);
 254             velecsum         = _mm256_add_pd(velecsum,velec);
 255             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 256             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 257
 258             fscal            = _mm256_add_pd(felec,fvdw);
 259
 260             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 261
 262             /* Calculate temporary vectorial force */
 263             tx               = _mm256_mul_pd(fscal,dx00);
 264             ty               = _mm256_mul_pd(fscal,dy00);
 265             tz               = _mm256_mul_pd(fscal,dz00);
 266
 267             /* Update vectorial force */
 268             fix0             = _mm256_add_pd(fix0,tx);
 269             fiy0             = _mm256_add_pd(fiy0,ty);
 270             fiz0             = _mm256_add_pd(fiz0,tz);
 271
 272             fjptrA             = f+j_coord_offsetA;
 273             fjptrB             = f+j_coord_offsetB;
 274             fjptrC             = f+j_coord_offsetC;
 275             fjptrD             = f+j_coord_offsetD;
 276             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 277
 278             }
 279
 280             /* Inner loop uses 70 flops */
 281         }
 282
 283         if(jidx<j_index_end)
 284         {
 285
 286             /* Get j neighbor index, and coordinate index */
 287             jnrlistA         = jjnr[jidx];
 288             jnrlistB         = jjnr[jidx+1];
 289             jnrlistC         = jjnr[jidx+2];
 290             jnrlistD         = jjnr[jidx+3];
 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_pd(mask,val) to clear dummy entries.
 294              */
 295             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 296
 297             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 298             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 299             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 300
 301             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 302             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 303             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 304             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 305             j_coord_offsetA  = DIM*jnrA;
 306             j_coord_offsetB  = DIM*jnrB;
 307             j_coord_offsetC  = DIM*jnrC;
 308             j_coord_offsetD  = DIM*jnrD;
 309
 310             /* load j atom coordinates */
 311             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 312                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 313                                                  &jx0,&jy0,&jz0);
 314
 315             /* Calculate displacement vector */
 316             dx00             = _mm256_sub_pd(ix0,jx0);
 317             dy00             = _mm256_sub_pd(iy0,jy0);
 318             dz00             = _mm256_sub_pd(iz0,jz0);
 319
 320             /* Calculate squared distance and things based on it */
 321             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 322
 323             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 324
 325             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 326
 327             /* Load parameters for j particles */
 328             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 329                                                                  charge+jnrC+0,charge+jnrD+0);
 330             vdwjidx0A        = 2*vdwtype[jnrA+0];
 331             vdwjidx0B        = 2*vdwtype[jnrB+0];
 332             vdwjidx0C        = 2*vdwtype[jnrC+0];
 333             vdwjidx0D        = 2*vdwtype[jnrD+0];
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 340             {
 341
 342             r00              = _mm256_mul_pd(rsq00,rinv00);
 343             r00              = _mm256_andnot_pd(dummy_mask,r00);
 344
 345             /* Compute parameters for interactions between i and j atoms */
 346             qq00             = _mm256_mul_pd(iq0,jq0);
 347             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 348                                             vdwioffsetptr0+vdwjidx0B,
 349                                             vdwioffsetptr0+vdwjidx0C,
 350                                             vdwioffsetptr0+vdwjidx0D,
 351                                             &c6_00,&c12_00);
 352
 353             /* REACTION-FIELD ELECTROSTATICS */
 354             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 355             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 356
 357             /* LENNARD-JONES DISPERSION/REPULSION */
 358
 359             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 360             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 361             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 362             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 363             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 364
 365             d                = _mm256_sub_pd(r00,rswitch);
 366             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 367             d2               = _mm256_mul_pd(d,d);
 368             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 369
 370             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 371
 372             /* Evaluate switch function */
 373             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 374             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 375             vvdw             = _mm256_mul_pd(vvdw,sw);
 376             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 377
 378             /* Update potential sum for this i atom from the interaction with this j atom. */
 379             velec            = _mm256_and_pd(velec,cutoff_mask);
 380             velec            = _mm256_andnot_pd(dummy_mask,velec);
 381             velecsum         = _mm256_add_pd(velecsum,velec);
 382             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 383             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 384             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 385
 386             fscal            = _mm256_add_pd(felec,fvdw);
 387
 388             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 389
 390             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 391
 392             /* Calculate temporary vectorial force */
 393             tx               = _mm256_mul_pd(fscal,dx00);
 394             ty               = _mm256_mul_pd(fscal,dy00);
 395             tz               = _mm256_mul_pd(fscal,dz00);
 396
 397             /* Update vectorial force */
 398             fix0             = _mm256_add_pd(fix0,tx);
 399             fiy0             = _mm256_add_pd(fiy0,ty);
 400             fiz0             = _mm256_add_pd(fiz0,tz);
 401
 402             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 403             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 404             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 405             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 406             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 407
 408             }
 409
 410             /* Inner loop uses 71 flops */
 411         }
 412
 413         /* End of innermost loop */
 414
 415         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 416                                                  f+i_coord_offset,fshift+i_shift_offset);
 417
 418         ggid                        = gid[iidx];
 419         /* Update potential energies */
 420         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 421         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 422
 423         /* Increment number of inner iterations */
 424         inneriter                  += j_index_end - j_index_start;
 425
 426         /* Outer loop uses 9 flops */
 427     }
 428
 429     /* Increment number of outer iterations */
 430     outeriter        += nri;
 431
 432     /* Update outer/inner flops */
 433
 434     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
 435 }
 436 /*
 437  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 438  * Electrostatics interaction: ReactionField
 439  * VdW interaction:            LennardJones
 440  * Geometry:                   Particle-Particle
 441  * Calculate force/pot:        Force
 442  */
 443 void
 444 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 445                     (t_nblist * gmx_restrict                nlist,
 446                      rvec * gmx_restrict                    xx,
 447                      rvec * gmx_restrict                    ff,
 448                      t_forcerec * gmx_restrict              fr,
 449                      t_mdatoms * gmx_restrict               mdatoms,
 450                      nb_kernel_data_t * gmx_restrict        kernel_data,
 451                      t_nrnb * gmx_restrict                  nrnb)
 452 {
 453     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 454      * just 0 for non-waters.
 455      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 456      * jnr indices corresponding to data put in the four positions in the SIMD register.
 457      */
 458     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 459     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 460     int              jnrA,jnrB,jnrC,jnrD;
 461     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 462     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 463     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 464     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 465     real             rcutoff_scalar;
 466     real             *shiftvec,*fshift,*x,*f;
 467     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 468     real             scratch[4*DIM];
 469     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 470     real *           vdwioffsetptr0;
 471     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 472     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 473     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 474     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 475     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 476     real             *charge;
 477     int              nvdwtype;
 478     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 479     int              *vdwtype;
 480     real             *vdwparam;
 481     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 482     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 483     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 484     real             rswitch_scalar,d_scalar;
 485     __m256d          dummy_mask,cutoff_mask;
 486     __m128           tmpmask0,tmpmask1;
 487     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 488     __m256d          one     = _mm256_set1_pd(1.0);
 489     __m256d          two     = _mm256_set1_pd(2.0);
 490     x                = xx[0];
 491     f                = ff[0];
 492
 493     nri              = nlist->nri;
 494     iinr             = nlist->iinr;
 495     jindex           = nlist->jindex;
 496     jjnr             = nlist->jjnr;
 497     shiftidx         = nlist->shift;
 498     gid              = nlist->gid;
 499     shiftvec         = fr->shift_vec[0];
 500     fshift           = fr->fshift[0];
 501     facel            = _mm256_set1_pd(fr->epsfac);
 502     charge           = mdatoms->chargeA;
 503     krf              = _mm256_set1_pd(fr->ic->k_rf);
 504     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 505     crf              = _mm256_set1_pd(fr->ic->c_rf);
 506     nvdwtype         = fr->ntype;
 507     vdwparam         = fr->nbfp;
 508     vdwtype          = mdatoms->typeA;
 509
 510     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 511     rcutoff_scalar   = fr->rcoulomb;
 512     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 513     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 514
 515     rswitch_scalar   = fr->rvdw_switch;
 516     rswitch          = _mm256_set1_pd(rswitch_scalar);
 517     /* Setup switch parameters */
 518     d_scalar         = rcutoff_scalar-rswitch_scalar;
 519     d                = _mm256_set1_pd(d_scalar);
 520     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 521     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 522     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 523     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 524     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 525     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 526
 527     /* Avoid stupid compiler warnings */
 528     jnrA = jnrB = jnrC = jnrD = 0;
 529     j_coord_offsetA = 0;
 530     j_coord_offsetB = 0;
 531     j_coord_offsetC = 0;
 532     j_coord_offsetD = 0;
 533
 534     outeriter        = 0;
 535     inneriter        = 0;
 536
 537     for(iidx=0;iidx<4*DIM;iidx++)
 538     {
 539         scratch[iidx] = 0.0;
 540     }
 541
 542     /* Start outer loop over neighborlists */
 543     for(iidx=0; iidx<nri; iidx++)
 544     {
 545         /* Load shift vector for this list */
 546         i_shift_offset   = DIM*shiftidx[iidx];
 547
 548         /* Load limits for loop over neighbors */
 549         j_index_start    = jindex[iidx];
 550         j_index_end      = jindex[iidx+1];
 551
 552         /* Get outer coordinate index */
 553         inr              = iinr[iidx];
 554         i_coord_offset   = DIM*inr;
 555
 556         /* Load i particle coords and add shift vector */
 557         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 558
 559         fix0             = _mm256_setzero_pd();
 560         fiy0             = _mm256_setzero_pd();
 561         fiz0             = _mm256_setzero_pd();
 562
 563         /* Load parameters for i particles */
 564         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 565         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 566
 567         /* Start inner kernel loop */
 568         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 569         {
 570
 571             /* Get j neighbor index, and coordinate index */
 572             jnrA             = jjnr[jidx];
 573             jnrB             = jjnr[jidx+1];
 574             jnrC             = jjnr[jidx+2];
 575             jnrD             = jjnr[jidx+3];
 576             j_coord_offsetA  = DIM*jnrA;
 577             j_coord_offsetB  = DIM*jnrB;
 578             j_coord_offsetC  = DIM*jnrC;
 579             j_coord_offsetD  = DIM*jnrD;
 580
 581             /* load j atom coordinates */
 582             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 583                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 584                                                  &jx0,&jy0,&jz0);
 585
 586             /* Calculate displacement vector */
 587             dx00             = _mm256_sub_pd(ix0,jx0);
 588             dy00             = _mm256_sub_pd(iy0,jy0);
 589             dz00             = _mm256_sub_pd(iz0,jz0);
 590
 591             /* Calculate squared distance and things based on it */
 592             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 593
 594             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 595
 596             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 597
 598             /* Load parameters for j particles */
 599             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 600                                                                  charge+jnrC+0,charge+jnrD+0);
 601             vdwjidx0A        = 2*vdwtype[jnrA+0];
 602             vdwjidx0B        = 2*vdwtype[jnrB+0];
 603             vdwjidx0C        = 2*vdwtype[jnrC+0];
 604             vdwjidx0D        = 2*vdwtype[jnrD+0];
 605
 606             /**************************
 607              * CALCULATE INTERACTIONS *
 608              **************************/
 609
 610             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 611             {
 612
 613             r00              = _mm256_mul_pd(rsq00,rinv00);
 614
 615             /* Compute parameters for interactions between i and j atoms */
 616             qq00             = _mm256_mul_pd(iq0,jq0);
 617             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 618                                             vdwioffsetptr0+vdwjidx0B,
 619                                             vdwioffsetptr0+vdwjidx0C,
 620                                             vdwioffsetptr0+vdwjidx0D,
 621                                             &c6_00,&c12_00);
 622
 623             /* REACTION-FIELD ELECTROSTATICS */
 624             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 625
 626             /* LENNARD-JONES DISPERSION/REPULSION */
 627
 628             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 629             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 630             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 631             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 632             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 633
 634             d                = _mm256_sub_pd(r00,rswitch);
 635             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 636             d2               = _mm256_mul_pd(d,d);
 637             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 638
 639             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 640
 641             /* Evaluate switch function */
 642             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 643             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 644             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 645
 646             fscal            = _mm256_add_pd(felec,fvdw);
 647
 648             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 649
 650             /* Calculate temporary vectorial force */
 651             tx               = _mm256_mul_pd(fscal,dx00);
 652             ty               = _mm256_mul_pd(fscal,dy00);
 653             tz               = _mm256_mul_pd(fscal,dz00);
 654
 655             /* Update vectorial force */
 656             fix0             = _mm256_add_pd(fix0,tx);
 657             fiy0             = _mm256_add_pd(fiy0,ty);
 658             fiz0             = _mm256_add_pd(fiz0,tz);
 659
 660             fjptrA             = f+j_coord_offsetA;
 661             fjptrB             = f+j_coord_offsetB;
 662             fjptrC             = f+j_coord_offsetC;
 663             fjptrD             = f+j_coord_offsetD;
 664             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 665
 666             }
 667
 668             /* Inner loop uses 61 flops */
 669         }
 670
 671         if(jidx<j_index_end)
 672         {
 673
 674             /* Get j neighbor index, and coordinate index */
 675             jnrlistA         = jjnr[jidx];
 676             jnrlistB         = jjnr[jidx+1];
 677             jnrlistC         = jjnr[jidx+2];
 678             jnrlistD         = jjnr[jidx+3];
 679             /* Sign of each element will be negative for non-real atoms.
 680              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 681              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 682              */
 683             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 684
 685             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 686             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 687             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 688
 689             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 690             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 691             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 692             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 693             j_coord_offsetA  = DIM*jnrA;
 694             j_coord_offsetB  = DIM*jnrB;
 695             j_coord_offsetC  = DIM*jnrC;
 696             j_coord_offsetD  = DIM*jnrD;
 697
 698             /* load j atom coordinates */
 699             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 700                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 701                                                  &jx0,&jy0,&jz0);
 702
 703             /* Calculate displacement vector */
 704             dx00             = _mm256_sub_pd(ix0,jx0);
 705             dy00             = _mm256_sub_pd(iy0,jy0);
 706             dz00             = _mm256_sub_pd(iz0,jz0);
 707
 708             /* Calculate squared distance and things based on it */
 709             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 710
 711             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 712
 713             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 714
 715             /* Load parameters for j particles */
 716             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 717                                                                  charge+jnrC+0,charge+jnrD+0);
 718             vdwjidx0A        = 2*vdwtype[jnrA+0];
 719             vdwjidx0B        = 2*vdwtype[jnrB+0];
 720             vdwjidx0C        = 2*vdwtype[jnrC+0];
 721             vdwjidx0D        = 2*vdwtype[jnrD+0];
 722
 723             /**************************
 724              * CALCULATE INTERACTIONS *
 725              **************************/
 726
 727             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 728             {
 729
 730             r00              = _mm256_mul_pd(rsq00,rinv00);
 731             r00              = _mm256_andnot_pd(dummy_mask,r00);
 732
 733             /* Compute parameters for interactions between i and j atoms */
 734             qq00             = _mm256_mul_pd(iq0,jq0);
 735             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 736                                             vdwioffsetptr0+vdwjidx0B,
 737                                             vdwioffsetptr0+vdwjidx0C,
 738                                             vdwioffsetptr0+vdwjidx0D,
 739                                             &c6_00,&c12_00);
 740
 741             /* REACTION-FIELD ELECTROSTATICS */
 742             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 743
 744             /* LENNARD-JONES DISPERSION/REPULSION */
 745
 746             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 747             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 748             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 749             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 750             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 751
 752             d                = _mm256_sub_pd(r00,rswitch);
 753             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 754             d2               = _mm256_mul_pd(d,d);
 755             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 756
 757             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 758
 759             /* Evaluate switch function */
 760             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 761             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 762             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 763
 764             fscal            = _mm256_add_pd(felec,fvdw);
 765
 766             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 767
 768             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 769
 770             /* Calculate temporary vectorial force */
 771             tx               = _mm256_mul_pd(fscal,dx00);
 772             ty               = _mm256_mul_pd(fscal,dy00);
 773             tz               = _mm256_mul_pd(fscal,dz00);
 774
 775             /* Update vectorial force */
 776             fix0             = _mm256_add_pd(fix0,tx);
 777             fiy0             = _mm256_add_pd(fiy0,ty);
 778             fiz0             = _mm256_add_pd(fiz0,tz);
 779
 780             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 781             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 782             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 783             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 784             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 785
 786             }
 787
 788             /* Inner loop uses 62 flops */
 789         }
 790
 791         /* End of innermost loop */
 792
 793         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 794                                                  f+i_coord_offset,fshift+i_shift_offset);
 795
 796         /* Increment number of inner iterations */
 797         inneriter                  += j_index_end - j_index_start;
 798
 799         /* Outer loop uses 7 flops */
 800     }
 801
 802     /* Increment number of outer iterations */
 803     outeriter        += nri;
 804
 805     /* Update outer/inner flops */
 806
 807     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);
 808 }