src/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_avx_256_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_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_avx_256_single.h"
  34 #include "kernelutil_x86_avx_256_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_256_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_256_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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
  62     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  63     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  64     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  65     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  66     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  67     real             rcutoff_scalar;
  68     real             *shiftvec,*fshift,*x,*f;
  69     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  70     real             scratch[4*DIM];
  71     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  72     real *           vdwioffsetptr0;
  73     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  75     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  77     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     int              nvdwtype;
  80     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  81     int              *vdwtype;
  82     real             *vdwparam;
  83     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  84     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  85     __m256           dummy_mask,cutoff_mask;
  86     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  87     __m256           one     = _mm256_set1_ps(1.0);
  88     __m256           two     = _mm256_set1_ps(2.0);
  89     x                = xx[0];
  90     f                = ff[0];
  91
  92     nri              = nlist->nri;
  93     iinr             = nlist->iinr;
  94     jindex           = nlist->jindex;
  95     jjnr             = nlist->jjnr;
  96     shiftidx         = nlist->shift;
  97     gid              = nlist->gid;
  98     shiftvec         = fr->shift_vec[0];
  99     fshift           = fr->fshift[0];
 100     facel            = _mm256_set1_ps(fr->epsfac);
 101     charge           = mdatoms->chargeA;
 102     krf              = _mm256_set1_ps(fr->ic->k_rf);
 103     krf2             = _mm256_set1_ps(fr->ic->k_rf*2.0);
 104     crf              = _mm256_set1_ps(fr->ic->c_rf);
 105     nvdwtype         = fr->ntype;
 106     vdwparam         = fr->nbfp;
 107     vdwtype          = mdatoms->typeA;
 108
 109     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 110     rcutoff_scalar   = fr->rcoulomb;
 111     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 112     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 113
 114     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 115     rvdw             = _mm256_set1_ps(fr->rvdw);
 116
 117     /* Avoid stupid compiler warnings */
 118     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 119     j_coord_offsetA = 0;
 120     j_coord_offsetB = 0;
 121     j_coord_offsetC = 0;
 122     j_coord_offsetD = 0;
 123     j_coord_offsetE = 0;
 124     j_coord_offsetF = 0;
 125     j_coord_offsetG = 0;
 126     j_coord_offsetH = 0;
 127
 128     outeriter        = 0;
 129     inneriter        = 0;
 130
 131     for(iidx=0;iidx<4*DIM;iidx++)
 132     {
 133         scratch[iidx] = 0.0;
 134     }
 135
 136     /* Start outer loop over neighborlists */
 137     for(iidx=0; iidx<nri; iidx++)
 138     {
 139         /* Load shift vector for this list */
 140         i_shift_offset   = DIM*shiftidx[iidx];
 141
 142         /* Load limits for loop over neighbors */
 143         j_index_start    = jindex[iidx];
 144         j_index_end      = jindex[iidx+1];
 145
 146         /* Get outer coordinate index */
 147         inr              = iinr[iidx];
 148         i_coord_offset   = DIM*inr;
 149
 150         /* Load i particle coords and add shift vector */
 151         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 152
 153         fix0             = _mm256_setzero_ps();
 154         fiy0             = _mm256_setzero_ps();
 155         fiz0             = _mm256_setzero_ps();
 156
 157         /* Load parameters for i particles */
 158         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 159         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 160
 161         /* Reset potential sums */
 162         velecsum         = _mm256_setzero_ps();
 163         vvdwsum          = _mm256_setzero_ps();
 164
 165         /* Start inner kernel loop */
 166         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 167         {
 168
 169             /* Get j neighbor index, and coordinate index */
 170             jnrA             = jjnr[jidx];
 171             jnrB             = jjnr[jidx+1];
 172             jnrC             = jjnr[jidx+2];
 173             jnrD             = jjnr[jidx+3];
 174             jnrE             = jjnr[jidx+4];
 175             jnrF             = jjnr[jidx+5];
 176             jnrG             = jjnr[jidx+6];
 177             jnrH             = jjnr[jidx+7];
 178             j_coord_offsetA  = DIM*jnrA;
 179             j_coord_offsetB  = DIM*jnrB;
 180             j_coord_offsetC  = DIM*jnrC;
 181             j_coord_offsetD  = DIM*jnrD;
 182             j_coord_offsetE  = DIM*jnrE;
 183             j_coord_offsetF  = DIM*jnrF;
 184             j_coord_offsetG  = DIM*jnrG;
 185             j_coord_offsetH  = DIM*jnrH;
 186
 187             /* load j atom coordinates */
 188             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 190                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 191                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 192                                                  &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm256_sub_ps(ix0,jx0);
 196             dy00             = _mm256_sub_ps(iy0,jy0);
 197             dz00             = _mm256_sub_ps(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 201
 202             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 203
 204             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 208                                                                  charge+jnrC+0,charge+jnrD+0,
 209                                                                  charge+jnrE+0,charge+jnrF+0,
 210                                                                  charge+jnrG+0,charge+jnrH+0);
 211             vdwjidx0A        = 2*vdwtype[jnrA+0];
 212             vdwjidx0B        = 2*vdwtype[jnrB+0];
 213             vdwjidx0C        = 2*vdwtype[jnrC+0];
 214             vdwjidx0D        = 2*vdwtype[jnrD+0];
 215             vdwjidx0E        = 2*vdwtype[jnrE+0];
 216             vdwjidx0F        = 2*vdwtype[jnrF+0];
 217             vdwjidx0G        = 2*vdwtype[jnrG+0];
 218             vdwjidx0H        = 2*vdwtype[jnrH+0];
 219
 220             /**************************
 221              * CALCULATE INTERACTIONS *
 222              **************************/
 223
 224             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 225             {
 226
 227             /* Compute parameters for interactions between i and j atoms */
 228             qq00             = _mm256_mul_ps(iq0,jq0);
 229             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 230                                             vdwioffsetptr0+vdwjidx0B,
 231                                             vdwioffsetptr0+vdwjidx0C,
 232                                             vdwioffsetptr0+vdwjidx0D,
 233                                             vdwioffsetptr0+vdwjidx0E,
 234                                             vdwioffsetptr0+vdwjidx0F,
 235                                             vdwioffsetptr0+vdwjidx0G,
 236                                             vdwioffsetptr0+vdwjidx0H,
 237                                             &c6_00,&c12_00);
 238
 239             /* REACTION-FIELD ELECTROSTATICS */
 240             velec            = _mm256_mul_ps(qq00,_mm256_sub_ps(_mm256_add_ps(rinv00,_mm256_mul_ps(krf,rsq00)),crf));
 241             felec            = _mm256_mul_ps(qq00,_mm256_sub_ps(_mm256_mul_ps(rinv00,rinvsq00),krf2));
 242
 243             /* LENNARD-JONES DISPERSION/REPULSION */
 244
 245             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 246             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 247             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 248             vvdw             = _mm256_sub_ps(_mm256_mul_ps( _mm256_sub_ps(vvdw12 , _mm256_mul_ps(c12_00,_mm256_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 249                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 250             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 251
 252             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 253
 254             /* Update potential sum for this i atom from the interaction with this j atom. */
 255             velec            = _mm256_and_ps(velec,cutoff_mask);
 256             velecsum         = _mm256_add_ps(velecsum,velec);
 257             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 258             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 259
 260             fscal            = _mm256_add_ps(felec,fvdw);
 261
 262             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 263
 264             /* Calculate temporary vectorial force */
 265             tx               = _mm256_mul_ps(fscal,dx00);
 266             ty               = _mm256_mul_ps(fscal,dy00);
 267             tz               = _mm256_mul_ps(fscal,dz00);
 268
 269             /* Update vectorial force */
 270             fix0             = _mm256_add_ps(fix0,tx);
 271             fiy0             = _mm256_add_ps(fiy0,ty);
 272             fiz0             = _mm256_add_ps(fiz0,tz);
 273
 274             fjptrA             = f+j_coord_offsetA;
 275             fjptrB             = f+j_coord_offsetB;
 276             fjptrC             = f+j_coord_offsetC;
 277             fjptrD             = f+j_coord_offsetD;
 278             fjptrE             = f+j_coord_offsetE;
 279             fjptrF             = f+j_coord_offsetF;
 280             fjptrG             = f+j_coord_offsetG;
 281             fjptrH             = f+j_coord_offsetH;
 282             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 283
 284             }
 285
 286             /* Inner loop uses 54 flops */
 287         }
 288
 289         if(jidx<j_index_end)
 290         {
 291
 292             /* Get j neighbor index, and coordinate index */
 293             jnrlistA         = jjnr[jidx];
 294             jnrlistB         = jjnr[jidx+1];
 295             jnrlistC         = jjnr[jidx+2];
 296             jnrlistD         = jjnr[jidx+3];
 297             jnrlistE         = jjnr[jidx+4];
 298             jnrlistF         = jjnr[jidx+5];
 299             jnrlistG         = jjnr[jidx+6];
 300             jnrlistH         = jjnr[jidx+7];
 301             /* Sign of each element will be negative for non-real atoms.
 302              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 303              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 304              */
 305             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 306                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 307
 308             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 309             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 310             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 311             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 312             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 313             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 314             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 315             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 316             j_coord_offsetA  = DIM*jnrA;
 317             j_coord_offsetB  = DIM*jnrB;
 318             j_coord_offsetC  = DIM*jnrC;
 319             j_coord_offsetD  = DIM*jnrD;
 320             j_coord_offsetE  = DIM*jnrE;
 321             j_coord_offsetF  = DIM*jnrF;
 322             j_coord_offsetG  = DIM*jnrG;
 323             j_coord_offsetH  = DIM*jnrH;
 324
 325             /* load j atom coordinates */
 326             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 327                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 328                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 329                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 330                                                  &jx0,&jy0,&jz0);
 331
 332             /* Calculate displacement vector */
 333             dx00             = _mm256_sub_ps(ix0,jx0);
 334             dy00             = _mm256_sub_ps(iy0,jy0);
 335             dz00             = _mm256_sub_ps(iz0,jz0);
 336
 337             /* Calculate squared distance and things based on it */
 338             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 339
 340             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 341
 342             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 343
 344             /* Load parameters for j particles */
 345             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 346                                                                  charge+jnrC+0,charge+jnrD+0,
 347                                                                  charge+jnrE+0,charge+jnrF+0,
 348                                                                  charge+jnrG+0,charge+jnrH+0);
 349             vdwjidx0A        = 2*vdwtype[jnrA+0];
 350             vdwjidx0B        = 2*vdwtype[jnrB+0];
 351             vdwjidx0C        = 2*vdwtype[jnrC+0];
 352             vdwjidx0D        = 2*vdwtype[jnrD+0];
 353             vdwjidx0E        = 2*vdwtype[jnrE+0];
 354             vdwjidx0F        = 2*vdwtype[jnrF+0];
 355             vdwjidx0G        = 2*vdwtype[jnrG+0];
 356             vdwjidx0H        = 2*vdwtype[jnrH+0];
 357
 358             /**************************
 359              * CALCULATE INTERACTIONS *
 360              **************************/
 361
 362             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 363             {
 364
 365             /* Compute parameters for interactions between i and j atoms */
 366             qq00             = _mm256_mul_ps(iq0,jq0);
 367             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 368                                             vdwioffsetptr0+vdwjidx0B,
 369                                             vdwioffsetptr0+vdwjidx0C,
 370                                             vdwioffsetptr0+vdwjidx0D,
 371                                             vdwioffsetptr0+vdwjidx0E,
 372                                             vdwioffsetptr0+vdwjidx0F,
 373                                             vdwioffsetptr0+vdwjidx0G,
 374                                             vdwioffsetptr0+vdwjidx0H,
 375                                             &c6_00,&c12_00);
 376
 377             /* REACTION-FIELD ELECTROSTATICS */
 378             velec            = _mm256_mul_ps(qq00,_mm256_sub_ps(_mm256_add_ps(rinv00,_mm256_mul_ps(krf,rsq00)),crf));
 379             felec            = _mm256_mul_ps(qq00,_mm256_sub_ps(_mm256_mul_ps(rinv00,rinvsq00),krf2));
 380
 381             /* LENNARD-JONES DISPERSION/REPULSION */
 382
 383             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 384             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 385             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 386             vvdw             = _mm256_sub_ps(_mm256_mul_ps( _mm256_sub_ps(vvdw12 , _mm256_mul_ps(c12_00,_mm256_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 387                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 388             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 389
 390             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 391
 392             /* Update potential sum for this i atom from the interaction with this j atom. */
 393             velec            = _mm256_and_ps(velec,cutoff_mask);
 394             velec            = _mm256_andnot_ps(dummy_mask,velec);
 395             velecsum         = _mm256_add_ps(velecsum,velec);
 396             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 397             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 398             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 399
 400             fscal            = _mm256_add_ps(felec,fvdw);
 401
 402             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 403
 404             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 405
 406             /* Calculate temporary vectorial force */
 407             tx               = _mm256_mul_ps(fscal,dx00);
 408             ty               = _mm256_mul_ps(fscal,dy00);
 409             tz               = _mm256_mul_ps(fscal,dz00);
 410
 411             /* Update vectorial force */
 412             fix0             = _mm256_add_ps(fix0,tx);
 413             fiy0             = _mm256_add_ps(fiy0,ty);
 414             fiz0             = _mm256_add_ps(fiz0,tz);
 415
 416             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 417             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 418             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 419             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 420             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 421             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 422             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 423             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 424             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 425
 426             }
 427
 428             /* Inner loop uses 54 flops */
 429         }
 430
 431         /* End of innermost loop */
 432
 433         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 434                                                  f+i_coord_offset,fshift+i_shift_offset);
 435
 436         ggid                        = gid[iidx];
 437         /* Update potential energies */
 438         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 439         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 440
 441         /* Increment number of inner iterations */
 442         inneriter                  += j_index_end - j_index_start;
 443
 444         /* Outer loop uses 9 flops */
 445     }
 446
 447     /* Increment number of outer iterations */
 448     outeriter        += nri;
 449
 450     /* Update outer/inner flops */
 451
 452     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 453 }
 454 /*
 455  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_256_single
 456  * Electrostatics interaction: ReactionField
 457  * VdW interaction:            LennardJones
 458  * Geometry:                   Particle-Particle
 459  * Calculate force/pot:        Force
 460  */
 461 void
 462 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_256_single
 463                     (t_nblist * gmx_restrict                nlist,
 464                      rvec * gmx_restrict                    xx,
 465                      rvec * gmx_restrict                    ff,
 466                      t_forcerec * gmx_restrict              fr,
 467                      t_mdatoms * gmx_restrict               mdatoms,
 468                      nb_kernel_data_t * gmx_restrict        kernel_data,
 469                      t_nrnb * gmx_restrict                  nrnb)
 470 {
 471     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 472      * just 0 for non-waters.
 473      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 474      * jnr indices corresponding to data put in the four positions in the SIMD register.
 475      */
 476     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 477     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 478     int              jnrA,jnrB,jnrC,jnrD;
 479     int              jnrE,jnrF,jnrG,jnrH;
 480     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 481     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 482     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 483     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 484     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 485     real             rcutoff_scalar;
 486     real             *shiftvec,*fshift,*x,*f;
 487     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 488     real             scratch[4*DIM];
 489     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 490     real *           vdwioffsetptr0;
 491     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 492     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 493     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 494     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 495     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 496     real             *charge;
 497     int              nvdwtype;
 498     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 499     int              *vdwtype;
 500     real             *vdwparam;
 501     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 502     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 503     __m256           dummy_mask,cutoff_mask;
 504     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 505     __m256           one     = _mm256_set1_ps(1.0);
 506     __m256           two     = _mm256_set1_ps(2.0);
 507     x                = xx[0];
 508     f                = ff[0];
 509
 510     nri              = nlist->nri;
 511     iinr             = nlist->iinr;
 512     jindex           = nlist->jindex;
 513     jjnr             = nlist->jjnr;
 514     shiftidx         = nlist->shift;
 515     gid              = nlist->gid;
 516     shiftvec         = fr->shift_vec[0];
 517     fshift           = fr->fshift[0];
 518     facel            = _mm256_set1_ps(fr->epsfac);
 519     charge           = mdatoms->chargeA;
 520     krf              = _mm256_set1_ps(fr->ic->k_rf);
 521     krf2             = _mm256_set1_ps(fr->ic->k_rf*2.0);
 522     crf              = _mm256_set1_ps(fr->ic->c_rf);
 523     nvdwtype         = fr->ntype;
 524     vdwparam         = fr->nbfp;
 525     vdwtype          = mdatoms->typeA;
 526
 527     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 528     rcutoff_scalar   = fr->rcoulomb;
 529     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 530     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 531
 532     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 533     rvdw             = _mm256_set1_ps(fr->rvdw);
 534
 535     /* Avoid stupid compiler warnings */
 536     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 537     j_coord_offsetA = 0;
 538     j_coord_offsetB = 0;
 539     j_coord_offsetC = 0;
 540     j_coord_offsetD = 0;
 541     j_coord_offsetE = 0;
 542     j_coord_offsetF = 0;
 543     j_coord_offsetG = 0;
 544     j_coord_offsetH = 0;
 545
 546     outeriter        = 0;
 547     inneriter        = 0;
 548
 549     for(iidx=0;iidx<4*DIM;iidx++)
 550     {
 551         scratch[iidx] = 0.0;
 552     }
 553
 554     /* Start outer loop over neighborlists */
 555     for(iidx=0; iidx<nri; iidx++)
 556     {
 557         /* Load shift vector for this list */
 558         i_shift_offset   = DIM*shiftidx[iidx];
 559
 560         /* Load limits for loop over neighbors */
 561         j_index_start    = jindex[iidx];
 562         j_index_end      = jindex[iidx+1];
 563
 564         /* Get outer coordinate index */
 565         inr              = iinr[iidx];
 566         i_coord_offset   = DIM*inr;
 567
 568         /* Load i particle coords and add shift vector */
 569         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 570
 571         fix0             = _mm256_setzero_ps();
 572         fiy0             = _mm256_setzero_ps();
 573         fiz0             = _mm256_setzero_ps();
 574
 575         /* Load parameters for i particles */
 576         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 577         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 578
 579         /* Start inner kernel loop */
 580         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 581         {
 582
 583             /* Get j neighbor index, and coordinate index */
 584             jnrA             = jjnr[jidx];
 585             jnrB             = jjnr[jidx+1];
 586             jnrC             = jjnr[jidx+2];
 587             jnrD             = jjnr[jidx+3];
 588             jnrE             = jjnr[jidx+4];
 589             jnrF             = jjnr[jidx+5];
 590             jnrG             = jjnr[jidx+6];
 591             jnrH             = jjnr[jidx+7];
 592             j_coord_offsetA  = DIM*jnrA;
 593             j_coord_offsetB  = DIM*jnrB;
 594             j_coord_offsetC  = DIM*jnrC;
 595             j_coord_offsetD  = DIM*jnrD;
 596             j_coord_offsetE  = DIM*jnrE;
 597             j_coord_offsetF  = DIM*jnrF;
 598             j_coord_offsetG  = DIM*jnrG;
 599             j_coord_offsetH  = DIM*jnrH;
 600
 601             /* load j atom coordinates */
 602             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 603                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 604                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 605                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 606                                                  &jx0,&jy0,&jz0);
 607
 608             /* Calculate displacement vector */
 609             dx00             = _mm256_sub_ps(ix0,jx0);
 610             dy00             = _mm256_sub_ps(iy0,jy0);
 611             dz00             = _mm256_sub_ps(iz0,jz0);
 612
 613             /* Calculate squared distance and things based on it */
 614             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 615
 616             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 617
 618             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 619
 620             /* Load parameters for j particles */
 621             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 622                                                                  charge+jnrC+0,charge+jnrD+0,
 623                                                                  charge+jnrE+0,charge+jnrF+0,
 624                                                                  charge+jnrG+0,charge+jnrH+0);
 625             vdwjidx0A        = 2*vdwtype[jnrA+0];
 626             vdwjidx0B        = 2*vdwtype[jnrB+0];
 627             vdwjidx0C        = 2*vdwtype[jnrC+0];
 628             vdwjidx0D        = 2*vdwtype[jnrD+0];
 629             vdwjidx0E        = 2*vdwtype[jnrE+0];
 630             vdwjidx0F        = 2*vdwtype[jnrF+0];
 631             vdwjidx0G        = 2*vdwtype[jnrG+0];
 632             vdwjidx0H        = 2*vdwtype[jnrH+0];
 633
 634             /**************************
 635              * CALCULATE INTERACTIONS *
 636              **************************/
 637
 638             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 639             {
 640
 641             /* Compute parameters for interactions between i and j atoms */
 642             qq00             = _mm256_mul_ps(iq0,jq0);
 643             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 644                                             vdwioffsetptr0+vdwjidx0B,
 645                                             vdwioffsetptr0+vdwjidx0C,
 646                                             vdwioffsetptr0+vdwjidx0D,
 647                                             vdwioffsetptr0+vdwjidx0E,
 648                                             vdwioffsetptr0+vdwjidx0F,
 649                                             vdwioffsetptr0+vdwjidx0G,
 650                                             vdwioffsetptr0+vdwjidx0H,
 651                                             &c6_00,&c12_00);
 652
 653             /* REACTION-FIELD ELECTROSTATICS */
 654             felec            = _mm256_mul_ps(qq00,_mm256_sub_ps(_mm256_mul_ps(rinv00,rinvsq00),krf2));
 655
 656             /* LENNARD-JONES DISPERSION/REPULSION */
 657
 658             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 659             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 660
 661             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 662
 663             fscal            = _mm256_add_ps(felec,fvdw);
 664
 665             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 666
 667             /* Calculate temporary vectorial force */
 668             tx               = _mm256_mul_ps(fscal,dx00);
 669             ty               = _mm256_mul_ps(fscal,dy00);
 670             tz               = _mm256_mul_ps(fscal,dz00);
 671
 672             /* Update vectorial force */
 673             fix0             = _mm256_add_ps(fix0,tx);
 674             fiy0             = _mm256_add_ps(fiy0,ty);
 675             fiz0             = _mm256_add_ps(fiz0,tz);
 676
 677             fjptrA             = f+j_coord_offsetA;
 678             fjptrB             = f+j_coord_offsetB;
 679             fjptrC             = f+j_coord_offsetC;
 680             fjptrD             = f+j_coord_offsetD;
 681             fjptrE             = f+j_coord_offsetE;
 682             fjptrF             = f+j_coord_offsetF;
 683             fjptrG             = f+j_coord_offsetG;
 684             fjptrH             = f+j_coord_offsetH;
 685             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 686
 687             }
 688
 689             /* Inner loop uses 37 flops */
 690         }
 691
 692         if(jidx<j_index_end)
 693         {
 694
 695             /* Get j neighbor index, and coordinate index */
 696             jnrlistA         = jjnr[jidx];
 697             jnrlistB         = jjnr[jidx+1];
 698             jnrlistC         = jjnr[jidx+2];
 699             jnrlistD         = jjnr[jidx+3];
 700             jnrlistE         = jjnr[jidx+4];
 701             jnrlistF         = jjnr[jidx+5];
 702             jnrlistG         = jjnr[jidx+6];
 703             jnrlistH         = jjnr[jidx+7];
 704             /* Sign of each element will be negative for non-real atoms.
 705              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 706              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 707              */
 708             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 709                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 710
 711             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 712             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 713             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 714             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 715             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 716             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 717             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 718             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 719             j_coord_offsetA  = DIM*jnrA;
 720             j_coord_offsetB  = DIM*jnrB;
 721             j_coord_offsetC  = DIM*jnrC;
 722             j_coord_offsetD  = DIM*jnrD;
 723             j_coord_offsetE  = DIM*jnrE;
 724             j_coord_offsetF  = DIM*jnrF;
 725             j_coord_offsetG  = DIM*jnrG;
 726             j_coord_offsetH  = DIM*jnrH;
 727
 728             /* load j atom coordinates */
 729             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 730                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 731                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 732                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 733                                                  &jx0,&jy0,&jz0);
 734
 735             /* Calculate displacement vector */
 736             dx00             = _mm256_sub_ps(ix0,jx0);
 737             dy00             = _mm256_sub_ps(iy0,jy0);
 738             dz00             = _mm256_sub_ps(iz0,jz0);
 739
 740             /* Calculate squared distance and things based on it */
 741             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 742
 743             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 744
 745             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 746
 747             /* Load parameters for j particles */
 748             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 749                                                                  charge+jnrC+0,charge+jnrD+0,
 750                                                                  charge+jnrE+0,charge+jnrF+0,
 751                                                                  charge+jnrG+0,charge+jnrH+0);
 752             vdwjidx0A        = 2*vdwtype[jnrA+0];
 753             vdwjidx0B        = 2*vdwtype[jnrB+0];
 754             vdwjidx0C        = 2*vdwtype[jnrC+0];
 755             vdwjidx0D        = 2*vdwtype[jnrD+0];
 756             vdwjidx0E        = 2*vdwtype[jnrE+0];
 757             vdwjidx0F        = 2*vdwtype[jnrF+0];
 758             vdwjidx0G        = 2*vdwtype[jnrG+0];
 759             vdwjidx0H        = 2*vdwtype[jnrH+0];
 760
 761             /**************************
 762              * CALCULATE INTERACTIONS *
 763              **************************/
 764
 765             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 766             {
 767
 768             /* Compute parameters for interactions between i and j atoms */
 769             qq00             = _mm256_mul_ps(iq0,jq0);
 770             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 771                                             vdwioffsetptr0+vdwjidx0B,
 772                                             vdwioffsetptr0+vdwjidx0C,
 773                                             vdwioffsetptr0+vdwjidx0D,
 774                                             vdwioffsetptr0+vdwjidx0E,
 775                                             vdwioffsetptr0+vdwjidx0F,
 776                                             vdwioffsetptr0+vdwjidx0G,
 777                                             vdwioffsetptr0+vdwjidx0H,
 778                                             &c6_00,&c12_00);
 779
 780             /* REACTION-FIELD ELECTROSTATICS */
 781             felec            = _mm256_mul_ps(qq00,_mm256_sub_ps(_mm256_mul_ps(rinv00,rinvsq00),krf2));
 782
 783             /* LENNARD-JONES DISPERSION/REPULSION */
 784
 785             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 786             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 787
 788             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 789
 790             fscal            = _mm256_add_ps(felec,fvdw);
 791
 792             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 793
 794             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 795
 796             /* Calculate temporary vectorial force */
 797             tx               = _mm256_mul_ps(fscal,dx00);
 798             ty               = _mm256_mul_ps(fscal,dy00);
 799             tz               = _mm256_mul_ps(fscal,dz00);
 800
 801             /* Update vectorial force */
 802             fix0             = _mm256_add_ps(fix0,tx);
 803             fiy0             = _mm256_add_ps(fiy0,ty);
 804             fiz0             = _mm256_add_ps(fiz0,tz);
 805
 806             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 807             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 808             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 809             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 810             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 811             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 812             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 813             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 814             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 815
 816             }
 817
 818             /* Inner loop uses 37 flops */
 819         }
 820
 821         /* End of innermost loop */
 822
 823         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 824                                                  f+i_coord_offset,fshift+i_shift_offset);
 825
 826         /* Increment number of inner iterations */
 827         inneriter                  += j_index_end - j_index_start;
 828
 829         /* Outer loop uses 7 flops */
 830     }
 831
 832     /* Increment number of outer iterations */
 833     outeriter        += nri;
 834
 835     /* Update outer/inner flops */
 836
 837     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*37);
 838 }