src/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRF_VdwCSTab_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_ElecRF_VdwCSTab_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwCSTab_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     __m128i          vfitab;
  84     __m128i          ifour       = _mm_set1_epi32(4);
  85     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  86     real             *vftab;
  87     __m256d          dummy_mask,cutoff_mask;
  88     __m128           tmpmask0,tmpmask1;
  89     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  90     __m256d          one     = _mm256_set1_pd(1.0);
  91     __m256d          two     = _mm256_set1_pd(2.0);
  92     x                = xx[0];
  93     f                = ff[0];
  94
  95     nri              = nlist->nri;
  96     iinr             = nlist->iinr;
  97     jindex           = nlist->jindex;
  98     jjnr             = nlist->jjnr;
  99     shiftidx         = nlist->shift;
 100     gid              = nlist->gid;
 101     shiftvec         = fr->shift_vec[0];
 102     fshift           = fr->fshift[0];
 103     facel            = _mm256_set1_pd(fr->epsfac);
 104     charge           = mdatoms->chargeA;
 105     krf              = _mm256_set1_pd(fr->ic->k_rf);
 106     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 107     crf              = _mm256_set1_pd(fr->ic->c_rf);
 108     nvdwtype         = fr->ntype;
 109     vdwparam         = fr->nbfp;
 110     vdwtype          = mdatoms->typeA;
 111
 112     vftab            = kernel_data->table_vdw->data;
 113     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 114
 115     /* Avoid stupid compiler warnings */
 116     jnrA = jnrB = jnrC = jnrD = 0;
 117     j_coord_offsetA = 0;
 118     j_coord_offsetB = 0;
 119     j_coord_offsetC = 0;
 120     j_coord_offsetD = 0;
 121
 122     outeriter        = 0;
 123     inneriter        = 0;
 124
 125     for(iidx=0;iidx<4*DIM;iidx++)
 126     {
 127         scratch[iidx] = 0.0;
 128     }
 129
 130     /* Start outer loop over neighborlists */
 131     for(iidx=0; iidx<nri; iidx++)
 132     {
 133         /* Load shift vector for this list */
 134         i_shift_offset   = DIM*shiftidx[iidx];
 135
 136         /* Load limits for loop over neighbors */
 137         j_index_start    = jindex[iidx];
 138         j_index_end      = jindex[iidx+1];
 139
 140         /* Get outer coordinate index */
 141         inr              = iinr[iidx];
 142         i_coord_offset   = DIM*inr;
 143
 144         /* Load i particle coords and add shift vector */
 145         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 146
 147         fix0             = _mm256_setzero_pd();
 148         fiy0             = _mm256_setzero_pd();
 149         fiz0             = _mm256_setzero_pd();
 150
 151         /* Load parameters for i particles */
 152         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 153         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 154
 155         /* Reset potential sums */
 156         velecsum         = _mm256_setzero_pd();
 157         vvdwsum          = _mm256_setzero_pd();
 158
 159         /* Start inner kernel loop */
 160         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 161         {
 162
 163             /* Get j neighbor index, and coordinate index */
 164             jnrA             = jjnr[jidx];
 165             jnrB             = jjnr[jidx+1];
 166             jnrC             = jjnr[jidx+2];
 167             jnrD             = jjnr[jidx+3];
 168             j_coord_offsetA  = DIM*jnrA;
 169             j_coord_offsetB  = DIM*jnrB;
 170             j_coord_offsetC  = DIM*jnrC;
 171             j_coord_offsetD  = DIM*jnrD;
 172
 173             /* load j atom coordinates */
 174             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 175                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 176                                                  &jx0,&jy0,&jz0);
 177
 178             /* Calculate displacement vector */
 179             dx00             = _mm256_sub_pd(ix0,jx0);
 180             dy00             = _mm256_sub_pd(iy0,jy0);
 181             dz00             = _mm256_sub_pd(iz0,jz0);
 182
 183             /* Calculate squared distance and things based on it */
 184             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 185
 186             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 187
 188             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 189
 190             /* Load parameters for j particles */
 191             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 192                                                                  charge+jnrC+0,charge+jnrD+0);
 193             vdwjidx0A        = 2*vdwtype[jnrA+0];
 194             vdwjidx0B        = 2*vdwtype[jnrB+0];
 195             vdwjidx0C        = 2*vdwtype[jnrC+0];
 196             vdwjidx0D        = 2*vdwtype[jnrD+0];
 197
 198             /**************************
 199              * CALCULATE INTERACTIONS *
 200              **************************/
 201
 202             r00              = _mm256_mul_pd(rsq00,rinv00);
 203
 204             /* Compute parameters for interactions between i and j atoms */
 205             qq00             = _mm256_mul_pd(iq0,jq0);
 206             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 207                                             vdwioffsetptr0+vdwjidx0B,
 208                                             vdwioffsetptr0+vdwjidx0C,
 209                                             vdwioffsetptr0+vdwjidx0D,
 210                                             &c6_00,&c12_00);
 211
 212             /* Calculate table index by multiplying r with table scale and truncate to integer */
 213             rt               = _mm256_mul_pd(r00,vftabscale);
 214             vfitab           = _mm256_cvttpd_epi32(rt);
 215             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 216             vfitab           = _mm_slli_epi32(vfitab,3);
 217
 218             /* REACTION-FIELD ELECTROSTATICS */
 219             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 220             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 221
 222             /* CUBIC SPLINE TABLE DISPERSION */
 223             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 224             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 225             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 226             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 227             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 228             Heps             = _mm256_mul_pd(vfeps,H);
 229             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 230             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 231             vvdw6            = _mm256_mul_pd(c6_00,VV);
 232             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 233             fvdw6            = _mm256_mul_pd(c6_00,FF);
 234
 235             /* CUBIC SPLINE TABLE REPULSION */
 236             vfitab           = _mm_add_epi32(vfitab,ifour);
 237             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 238             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 239             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 240             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 241             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 242             Heps             = _mm256_mul_pd(vfeps,H);
 243             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 244             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 245             vvdw12           = _mm256_mul_pd(c12_00,VV);
 246             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 247             fvdw12           = _mm256_mul_pd(c12_00,FF);
 248             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 249             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 250
 251             /* Update potential sum for this i atom from the interaction with this j atom. */
 252             velecsum         = _mm256_add_pd(velecsum,velec);
 253             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 254
 255             fscal            = _mm256_add_pd(felec,fvdw);
 256
 257             /* Calculate temporary vectorial force */
 258             tx               = _mm256_mul_pd(fscal,dx00);
 259             ty               = _mm256_mul_pd(fscal,dy00);
 260             tz               = _mm256_mul_pd(fscal,dz00);
 261
 262             /* Update vectorial force */
 263             fix0             = _mm256_add_pd(fix0,tx);
 264             fiy0             = _mm256_add_pd(fiy0,ty);
 265             fiz0             = _mm256_add_pd(fiz0,tz);
 266
 267             fjptrA             = f+j_coord_offsetA;
 268             fjptrB             = f+j_coord_offsetB;
 269             fjptrC             = f+j_coord_offsetC;
 270             fjptrD             = f+j_coord_offsetD;
 271             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 272
 273             /* Inner loop uses 67 flops */
 274         }
 275
 276         if(jidx<j_index_end)
 277         {
 278
 279             /* Get j neighbor index, and coordinate index */
 280             jnrlistA         = jjnr[jidx];
 281             jnrlistB         = jjnr[jidx+1];
 282             jnrlistC         = jjnr[jidx+2];
 283             jnrlistD         = jjnr[jidx+3];
 284             /* Sign of each element will be negative for non-real atoms.
 285              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 286              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 287              */
 288             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 289
 290             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 291             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 292             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 293
 294             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 295             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 296             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 297             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 298             j_coord_offsetA  = DIM*jnrA;
 299             j_coord_offsetB  = DIM*jnrB;
 300             j_coord_offsetC  = DIM*jnrC;
 301             j_coord_offsetD  = DIM*jnrD;
 302
 303             /* load j atom coordinates */
 304             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 305                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 306                                                  &jx0,&jy0,&jz0);
 307
 308             /* Calculate displacement vector */
 309             dx00             = _mm256_sub_pd(ix0,jx0);
 310             dy00             = _mm256_sub_pd(iy0,jy0);
 311             dz00             = _mm256_sub_pd(iz0,jz0);
 312
 313             /* Calculate squared distance and things based on it */
 314             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 315
 316             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 317
 318             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 319
 320             /* Load parameters for j particles */
 321             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 322                                                                  charge+jnrC+0,charge+jnrD+0);
 323             vdwjidx0A        = 2*vdwtype[jnrA+0];
 324             vdwjidx0B        = 2*vdwtype[jnrB+0];
 325             vdwjidx0C        = 2*vdwtype[jnrC+0];
 326             vdwjidx0D        = 2*vdwtype[jnrD+0];
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             r00              = _mm256_mul_pd(rsq00,rinv00);
 333             r00              = _mm256_andnot_pd(dummy_mask,r00);
 334
 335             /* Compute parameters for interactions between i and j atoms */
 336             qq00             = _mm256_mul_pd(iq0,jq0);
 337             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 338                                             vdwioffsetptr0+vdwjidx0B,
 339                                             vdwioffsetptr0+vdwjidx0C,
 340                                             vdwioffsetptr0+vdwjidx0D,
 341                                             &c6_00,&c12_00);
 342
 343             /* Calculate table index by multiplying r with table scale and truncate to integer */
 344             rt               = _mm256_mul_pd(r00,vftabscale);
 345             vfitab           = _mm256_cvttpd_epi32(rt);
 346             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 347             vfitab           = _mm_slli_epi32(vfitab,3);
 348
 349             /* REACTION-FIELD ELECTROSTATICS */
 350             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 351             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 352
 353             /* CUBIC SPLINE TABLE DISPERSION */
 354             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 355             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 356             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 357             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 358             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 359             Heps             = _mm256_mul_pd(vfeps,H);
 360             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 361             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 362             vvdw6            = _mm256_mul_pd(c6_00,VV);
 363             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 364             fvdw6            = _mm256_mul_pd(c6_00,FF);
 365
 366             /* CUBIC SPLINE TABLE REPULSION */
 367             vfitab           = _mm_add_epi32(vfitab,ifour);
 368             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 369             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 370             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 371             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 372             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 373             Heps             = _mm256_mul_pd(vfeps,H);
 374             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 375             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 376             vvdw12           = _mm256_mul_pd(c12_00,VV);
 377             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 378             fvdw12           = _mm256_mul_pd(c12_00,FF);
 379             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 380             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velec            = _mm256_andnot_pd(dummy_mask,velec);
 384             velecsum         = _mm256_add_pd(velecsum,velec);
 385             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 386             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 387
 388             fscal            = _mm256_add_pd(felec,fvdw);
 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             /* Inner loop uses 68 flops */
 409         }
 410
 411         /* End of innermost loop */
 412
 413         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 414                                                  f+i_coord_offset,fshift+i_shift_offset);
 415
 416         ggid                        = gid[iidx];
 417         /* Update potential energies */
 418         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 419         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 420
 421         /* Increment number of inner iterations */
 422         inneriter                  += j_index_end - j_index_start;
 423
 424         /* Outer loop uses 9 flops */
 425     }
 426
 427     /* Increment number of outer iterations */
 428     outeriter        += nri;
 429
 430     /* Update outer/inner flops */
 431
 432     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*68);
 433 }
 434 /*
 435  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_256_double
 436  * Electrostatics interaction: ReactionField
 437  * VdW interaction:            CubicSplineTable
 438  * Geometry:                   Particle-Particle
 439  * Calculate force/pot:        Force
 440  */
 441 void
 442 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_256_double
 443                     (t_nblist * gmx_restrict                nlist,
 444                      rvec * gmx_restrict                    xx,
 445                      rvec * gmx_restrict                    ff,
 446                      t_forcerec * gmx_restrict              fr,
 447                      t_mdatoms * gmx_restrict               mdatoms,
 448                      nb_kernel_data_t * gmx_restrict        kernel_data,
 449                      t_nrnb * gmx_restrict                  nrnb)
 450 {
 451     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 452      * just 0 for non-waters.
 453      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 454      * jnr indices corresponding to data put in the four positions in the SIMD register.
 455      */
 456     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 457     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 458     int              jnrA,jnrB,jnrC,jnrD;
 459     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 460     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 461     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 462     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 463     real             rcutoff_scalar;
 464     real             *shiftvec,*fshift,*x,*f;
 465     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 466     real             scratch[4*DIM];
 467     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 468     real *           vdwioffsetptr0;
 469     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 470     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 471     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 472     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 473     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 474     real             *charge;
 475     int              nvdwtype;
 476     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 477     int              *vdwtype;
 478     real             *vdwparam;
 479     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 480     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 481     __m128i          vfitab;
 482     __m128i          ifour       = _mm_set1_epi32(4);
 483     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 484     real             *vftab;
 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     vftab            = kernel_data->table_vdw->data;
 511     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 512
 513     /* Avoid stupid compiler warnings */
 514     jnrA = jnrB = jnrC = jnrD = 0;
 515     j_coord_offsetA = 0;
 516     j_coord_offsetB = 0;
 517     j_coord_offsetC = 0;
 518     j_coord_offsetD = 0;
 519
 520     outeriter        = 0;
 521     inneriter        = 0;
 522
 523     for(iidx=0;iidx<4*DIM;iidx++)
 524     {
 525         scratch[iidx] = 0.0;
 526     }
 527
 528     /* Start outer loop over neighborlists */
 529     for(iidx=0; iidx<nri; iidx++)
 530     {
 531         /* Load shift vector for this list */
 532         i_shift_offset   = DIM*shiftidx[iidx];
 533
 534         /* Load limits for loop over neighbors */
 535         j_index_start    = jindex[iidx];
 536         j_index_end      = jindex[iidx+1];
 537
 538         /* Get outer coordinate index */
 539         inr              = iinr[iidx];
 540         i_coord_offset   = DIM*inr;
 541
 542         /* Load i particle coords and add shift vector */
 543         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 544
 545         fix0             = _mm256_setzero_pd();
 546         fiy0             = _mm256_setzero_pd();
 547         fiz0             = _mm256_setzero_pd();
 548
 549         /* Load parameters for i particles */
 550         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 551         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 552
 553         /* Start inner kernel loop */
 554         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 555         {
 556
 557             /* Get j neighbor index, and coordinate index */
 558             jnrA             = jjnr[jidx];
 559             jnrB             = jjnr[jidx+1];
 560             jnrC             = jjnr[jidx+2];
 561             jnrD             = jjnr[jidx+3];
 562             j_coord_offsetA  = DIM*jnrA;
 563             j_coord_offsetB  = DIM*jnrB;
 564             j_coord_offsetC  = DIM*jnrC;
 565             j_coord_offsetD  = DIM*jnrD;
 566
 567             /* load j atom coordinates */
 568             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 569                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 570                                                  &jx0,&jy0,&jz0);
 571
 572             /* Calculate displacement vector */
 573             dx00             = _mm256_sub_pd(ix0,jx0);
 574             dy00             = _mm256_sub_pd(iy0,jy0);
 575             dz00             = _mm256_sub_pd(iz0,jz0);
 576
 577             /* Calculate squared distance and things based on it */
 578             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 579
 580             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 581
 582             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 583
 584             /* Load parameters for j particles */
 585             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 586                                                                  charge+jnrC+0,charge+jnrD+0);
 587             vdwjidx0A        = 2*vdwtype[jnrA+0];
 588             vdwjidx0B        = 2*vdwtype[jnrB+0];
 589             vdwjidx0C        = 2*vdwtype[jnrC+0];
 590             vdwjidx0D        = 2*vdwtype[jnrD+0];
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             r00              = _mm256_mul_pd(rsq00,rinv00);
 597
 598             /* Compute parameters for interactions between i and j atoms */
 599             qq00             = _mm256_mul_pd(iq0,jq0);
 600             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 601                                             vdwioffsetptr0+vdwjidx0B,
 602                                             vdwioffsetptr0+vdwjidx0C,
 603                                             vdwioffsetptr0+vdwjidx0D,
 604                                             &c6_00,&c12_00);
 605
 606             /* Calculate table index by multiplying r with table scale and truncate to integer */
 607             rt               = _mm256_mul_pd(r00,vftabscale);
 608             vfitab           = _mm256_cvttpd_epi32(rt);
 609             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 610             vfitab           = _mm_slli_epi32(vfitab,3);
 611
 612             /* REACTION-FIELD ELECTROSTATICS */
 613             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 614
 615             /* CUBIC SPLINE TABLE DISPERSION */
 616             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 617             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 618             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 619             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 620             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 621             Heps             = _mm256_mul_pd(vfeps,H);
 622             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 623             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 624             fvdw6            = _mm256_mul_pd(c6_00,FF);
 625
 626             /* CUBIC SPLINE TABLE REPULSION */
 627             vfitab           = _mm_add_epi32(vfitab,ifour);
 628             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 629             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 630             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 631             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 632             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 633             Heps             = _mm256_mul_pd(vfeps,H);
 634             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 635             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 636             fvdw12           = _mm256_mul_pd(c12_00,FF);
 637             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 638
 639             fscal            = _mm256_add_pd(felec,fvdw);
 640
 641             /* Calculate temporary vectorial force */
 642             tx               = _mm256_mul_pd(fscal,dx00);
 643             ty               = _mm256_mul_pd(fscal,dy00);
 644             tz               = _mm256_mul_pd(fscal,dz00);
 645
 646             /* Update vectorial force */
 647             fix0             = _mm256_add_pd(fix0,tx);
 648             fiy0             = _mm256_add_pd(fiy0,ty);
 649             fiz0             = _mm256_add_pd(fiz0,tz);
 650
 651             fjptrA             = f+j_coord_offsetA;
 652             fjptrB             = f+j_coord_offsetB;
 653             fjptrC             = f+j_coord_offsetC;
 654             fjptrD             = f+j_coord_offsetD;
 655             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 656
 657             /* Inner loop uses 54 flops */
 658         }
 659
 660         if(jidx<j_index_end)
 661         {
 662
 663             /* Get j neighbor index, and coordinate index */
 664             jnrlistA         = jjnr[jidx];
 665             jnrlistB         = jjnr[jidx+1];
 666             jnrlistC         = jjnr[jidx+2];
 667             jnrlistD         = jjnr[jidx+3];
 668             /* Sign of each element will be negative for non-real atoms.
 669              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 670              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 671              */
 672             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 673
 674             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 675             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 676             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 677
 678             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 679             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 680             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 681             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 682             j_coord_offsetA  = DIM*jnrA;
 683             j_coord_offsetB  = DIM*jnrB;
 684             j_coord_offsetC  = DIM*jnrC;
 685             j_coord_offsetD  = DIM*jnrD;
 686
 687             /* load j atom coordinates */
 688             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 689                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 690                                                  &jx0,&jy0,&jz0);
 691
 692             /* Calculate displacement vector */
 693             dx00             = _mm256_sub_pd(ix0,jx0);
 694             dy00             = _mm256_sub_pd(iy0,jy0);
 695             dz00             = _mm256_sub_pd(iz0,jz0);
 696
 697             /* Calculate squared distance and things based on it */
 698             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 699
 700             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 701
 702             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 703
 704             /* Load parameters for j particles */
 705             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 706                                                                  charge+jnrC+0,charge+jnrD+0);
 707             vdwjidx0A        = 2*vdwtype[jnrA+0];
 708             vdwjidx0B        = 2*vdwtype[jnrB+0];
 709             vdwjidx0C        = 2*vdwtype[jnrC+0];
 710             vdwjidx0D        = 2*vdwtype[jnrD+0];
 711
 712             /**************************
 713              * CALCULATE INTERACTIONS *
 714              **************************/
 715
 716             r00              = _mm256_mul_pd(rsq00,rinv00);
 717             r00              = _mm256_andnot_pd(dummy_mask,r00);
 718
 719             /* Compute parameters for interactions between i and j atoms */
 720             qq00             = _mm256_mul_pd(iq0,jq0);
 721             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 722                                             vdwioffsetptr0+vdwjidx0B,
 723                                             vdwioffsetptr0+vdwjidx0C,
 724                                             vdwioffsetptr0+vdwjidx0D,
 725                                             &c6_00,&c12_00);
 726
 727             /* Calculate table index by multiplying r with table scale and truncate to integer */
 728             rt               = _mm256_mul_pd(r00,vftabscale);
 729             vfitab           = _mm256_cvttpd_epi32(rt);
 730             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 731             vfitab           = _mm_slli_epi32(vfitab,3);
 732
 733             /* REACTION-FIELD ELECTROSTATICS */
 734             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 735
 736             /* CUBIC SPLINE TABLE DISPERSION */
 737             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 738             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 739             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 740             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 741             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 742             Heps             = _mm256_mul_pd(vfeps,H);
 743             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 744             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 745             fvdw6            = _mm256_mul_pd(c6_00,FF);
 746
 747             /* CUBIC SPLINE TABLE REPULSION */
 748             vfitab           = _mm_add_epi32(vfitab,ifour);
 749             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 750             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 751             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 752             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 753             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 754             Heps             = _mm256_mul_pd(vfeps,H);
 755             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 756             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 757             fvdw12           = _mm256_mul_pd(c12_00,FF);
 758             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 759
 760             fscal            = _mm256_add_pd(felec,fvdw);
 761
 762             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 763
 764             /* Calculate temporary vectorial force */
 765             tx               = _mm256_mul_pd(fscal,dx00);
 766             ty               = _mm256_mul_pd(fscal,dy00);
 767             tz               = _mm256_mul_pd(fscal,dz00);
 768
 769             /* Update vectorial force */
 770             fix0             = _mm256_add_pd(fix0,tx);
 771             fiy0             = _mm256_add_pd(fiy0,ty);
 772             fiz0             = _mm256_add_pd(fiz0,tz);
 773
 774             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 775             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 776             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 777             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 778             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 779
 780             /* Inner loop uses 55 flops */
 781         }
 782
 783         /* End of innermost loop */
 784
 785         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 786                                                  f+i_coord_offset,fshift+i_shift_offset);
 787
 788         /* Increment number of inner iterations */
 789         inneriter                  += j_index_end - j_index_start;
 790
 791         /* Outer loop uses 7 flops */
 792     }
 793
 794     /* Increment number of outer iterations */
 795     outeriter        += nri;
 796
 797     /* Update outer/inner flops */
 798
 799     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 800 }