src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCoul_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_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_VF_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     nvdwtype         = fr->ntype;
 106     vdwparam         = fr->nbfp;
 107     vdwtype          = mdatoms->typeA;
 108
 109     vftab            = kernel_data->table_vdw->data;
 110     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 111
 112     /* Avoid stupid compiler warnings */
 113     jnrA = jnrB = jnrC = jnrD = 0;
 114     j_coord_offsetA = 0;
 115     j_coord_offsetB = 0;
 116     j_coord_offsetC = 0;
 117     j_coord_offsetD = 0;
 118
 119     outeriter        = 0;
 120     inneriter        = 0;
 121
 122     for(iidx=0;iidx<4*DIM;iidx++)
 123     {
 124         scratch[iidx] = 0.0;
 125     }
 126
 127     /* Start outer loop over neighborlists */
 128     for(iidx=0; iidx<nri; iidx++)
 129     {
 130         /* Load shift vector for this list */
 131         i_shift_offset   = DIM*shiftidx[iidx];
 132
 133         /* Load limits for loop over neighbors */
 134         j_index_start    = jindex[iidx];
 135         j_index_end      = jindex[iidx+1];
 136
 137         /* Get outer coordinate index */
 138         inr              = iinr[iidx];
 139         i_coord_offset   = DIM*inr;
 140
 141         /* Load i particle coords and add shift vector */
 142         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 143
 144         fix0             = _mm256_setzero_pd();
 145         fiy0             = _mm256_setzero_pd();
 146         fiz0             = _mm256_setzero_pd();
 147
 148         /* Load parameters for i particles */
 149         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 150         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 151
 152         /* Reset potential sums */
 153         velecsum         = _mm256_setzero_pd();
 154         vvdwsum          = _mm256_setzero_pd();
 155
 156         /* Start inner kernel loop */
 157         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 158         {
 159
 160             /* Get j neighbor index, and coordinate index */
 161             jnrA             = jjnr[jidx];
 162             jnrB             = jjnr[jidx+1];
 163             jnrC             = jjnr[jidx+2];
 164             jnrD             = jjnr[jidx+3];
 165             j_coord_offsetA  = DIM*jnrA;
 166             j_coord_offsetB  = DIM*jnrB;
 167             j_coord_offsetC  = DIM*jnrC;
 168             j_coord_offsetD  = DIM*jnrD;
 169
 170             /* load j atom coordinates */
 171             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 172                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 173                                                  &jx0,&jy0,&jz0);
 174
 175             /* Calculate displacement vector */
 176             dx00             = _mm256_sub_pd(ix0,jx0);
 177             dy00             = _mm256_sub_pd(iy0,jy0);
 178             dz00             = _mm256_sub_pd(iz0,jz0);
 179
 180             /* Calculate squared distance and things based on it */
 181             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 182
 183             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 184
 185             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 186
 187             /* Load parameters for j particles */
 188             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 189                                                                  charge+jnrC+0,charge+jnrD+0);
 190             vdwjidx0A        = 2*vdwtype[jnrA+0];
 191             vdwjidx0B        = 2*vdwtype[jnrB+0];
 192             vdwjidx0C        = 2*vdwtype[jnrC+0];
 193             vdwjidx0D        = 2*vdwtype[jnrD+0];
 194
 195             /**************************
 196              * CALCULATE INTERACTIONS *
 197              **************************/
 198
 199             r00              = _mm256_mul_pd(rsq00,rinv00);
 200
 201             /* Compute parameters for interactions between i and j atoms */
 202             qq00             = _mm256_mul_pd(iq0,jq0);
 203             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 204                                             vdwioffsetptr0+vdwjidx0B,
 205                                             vdwioffsetptr0+vdwjidx0C,
 206                                             vdwioffsetptr0+vdwjidx0D,
 207                                             &c6_00,&c12_00);
 208
 209             /* Calculate table index by multiplying r with table scale and truncate to integer */
 210             rt               = _mm256_mul_pd(r00,vftabscale);
 211             vfitab           = _mm256_cvttpd_epi32(rt);
 212             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 213             vfitab           = _mm_slli_epi32(vfitab,3);
 214
 215             /* COULOMB ELECTROSTATICS */
 216             velec            = _mm256_mul_pd(qq00,rinv00);
 217             felec            = _mm256_mul_pd(velec,rinvsq00);
 218
 219             /* CUBIC SPLINE TABLE DISPERSION */
 220             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 221             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 222             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 223             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 224             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 225             Heps             = _mm256_mul_pd(vfeps,H);
 226             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 227             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 228             vvdw6            = _mm256_mul_pd(c6_00,VV);
 229             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 230             fvdw6            = _mm256_mul_pd(c6_00,FF);
 231
 232             /* CUBIC SPLINE TABLE REPULSION */
 233             vfitab           = _mm_add_epi32(vfitab,ifour);
 234             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 235             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 236             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 237             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 238             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 239             Heps             = _mm256_mul_pd(vfeps,H);
 240             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 241             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 242             vvdw12           = _mm256_mul_pd(c12_00,VV);
 243             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 244             fvdw12           = _mm256_mul_pd(c12_00,FF);
 245             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 246             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 247
 248             /* Update potential sum for this i atom from the interaction with this j atom. */
 249             velecsum         = _mm256_add_pd(velecsum,velec);
 250             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 251
 252             fscal            = _mm256_add_pd(felec,fvdw);
 253
 254             /* Calculate temporary vectorial force */
 255             tx               = _mm256_mul_pd(fscal,dx00);
 256             ty               = _mm256_mul_pd(fscal,dy00);
 257             tz               = _mm256_mul_pd(fscal,dz00);
 258
 259             /* Update vectorial force */
 260             fix0             = _mm256_add_pd(fix0,tx);
 261             fiy0             = _mm256_add_pd(fiy0,ty);
 262             fiz0             = _mm256_add_pd(fiz0,tz);
 263
 264             fjptrA             = f+j_coord_offsetA;
 265             fjptrB             = f+j_coord_offsetB;
 266             fjptrC             = f+j_coord_offsetC;
 267             fjptrD             = f+j_coord_offsetD;
 268             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 269
 270             /* Inner loop uses 62 flops */
 271         }
 272
 273         if(jidx<j_index_end)
 274         {
 275
 276             /* Get j neighbor index, and coordinate index */
 277             jnrlistA         = jjnr[jidx];
 278             jnrlistB         = jjnr[jidx+1];
 279             jnrlistC         = jjnr[jidx+2];
 280             jnrlistD         = jjnr[jidx+3];
 281             /* Sign of each element will be negative for non-real atoms.
 282              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 283              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 284              */
 285             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 286
 287             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 288             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 289             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 290
 291             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 292             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 293             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 294             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 295             j_coord_offsetA  = DIM*jnrA;
 296             j_coord_offsetB  = DIM*jnrB;
 297             j_coord_offsetC  = DIM*jnrC;
 298             j_coord_offsetD  = DIM*jnrD;
 299
 300             /* load j atom coordinates */
 301             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 302                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 303                                                  &jx0,&jy0,&jz0);
 304
 305             /* Calculate displacement vector */
 306             dx00             = _mm256_sub_pd(ix0,jx0);
 307             dy00             = _mm256_sub_pd(iy0,jy0);
 308             dz00             = _mm256_sub_pd(iz0,jz0);
 309
 310             /* Calculate squared distance and things based on it */
 311             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 312
 313             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 314
 315             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 316
 317             /* Load parameters for j particles */
 318             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 319                                                                  charge+jnrC+0,charge+jnrD+0);
 320             vdwjidx0A        = 2*vdwtype[jnrA+0];
 321             vdwjidx0B        = 2*vdwtype[jnrB+0];
 322             vdwjidx0C        = 2*vdwtype[jnrC+0];
 323             vdwjidx0D        = 2*vdwtype[jnrD+0];
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             r00              = _mm256_mul_pd(rsq00,rinv00);
 330             r00              = _mm256_andnot_pd(dummy_mask,r00);
 331
 332             /* Compute parameters for interactions between i and j atoms */
 333             qq00             = _mm256_mul_pd(iq0,jq0);
 334             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 335                                             vdwioffsetptr0+vdwjidx0B,
 336                                             vdwioffsetptr0+vdwjidx0C,
 337                                             vdwioffsetptr0+vdwjidx0D,
 338                                             &c6_00,&c12_00);
 339
 340             /* Calculate table index by multiplying r with table scale and truncate to integer */
 341             rt               = _mm256_mul_pd(r00,vftabscale);
 342             vfitab           = _mm256_cvttpd_epi32(rt);
 343             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 344             vfitab           = _mm_slli_epi32(vfitab,3);
 345
 346             /* COULOMB ELECTROSTATICS */
 347             velec            = _mm256_mul_pd(qq00,rinv00);
 348             felec            = _mm256_mul_pd(velec,rinvsq00);
 349
 350             /* CUBIC SPLINE TABLE DISPERSION */
 351             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 352             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 353             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 354             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 355             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 356             Heps             = _mm256_mul_pd(vfeps,H);
 357             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 358             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 359             vvdw6            = _mm256_mul_pd(c6_00,VV);
 360             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 361             fvdw6            = _mm256_mul_pd(c6_00,FF);
 362
 363             /* CUBIC SPLINE TABLE REPULSION */
 364             vfitab           = _mm_add_epi32(vfitab,ifour);
 365             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 366             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 367             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 368             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 369             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 370             Heps             = _mm256_mul_pd(vfeps,H);
 371             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 372             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 373             vvdw12           = _mm256_mul_pd(c12_00,VV);
 374             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 375             fvdw12           = _mm256_mul_pd(c12_00,FF);
 376             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 377             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm256_andnot_pd(dummy_mask,velec);
 381             velecsum         = _mm256_add_pd(velecsum,velec);
 382             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 383             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 384
 385             fscal            = _mm256_add_pd(felec,fvdw);
 386
 387             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 388
 389             /* Calculate temporary vectorial force */
 390             tx               = _mm256_mul_pd(fscal,dx00);
 391             ty               = _mm256_mul_pd(fscal,dy00);
 392             tz               = _mm256_mul_pd(fscal,dz00);
 393
 394             /* Update vectorial force */
 395             fix0             = _mm256_add_pd(fix0,tx);
 396             fiy0             = _mm256_add_pd(fiy0,ty);
 397             fiz0             = _mm256_add_pd(fiz0,tz);
 398
 399             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 400             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 401             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 402             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 403             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 404
 405             /* Inner loop uses 63 flops */
 406         }
 407
 408         /* End of innermost loop */
 409
 410         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 411                                                  f+i_coord_offset,fshift+i_shift_offset);
 412
 413         ggid                        = gid[iidx];
 414         /* Update potential energies */
 415         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 416         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 417
 418         /* Increment number of inner iterations */
 419         inneriter                  += j_index_end - j_index_start;
 420
 421         /* Outer loop uses 9 flops */
 422     }
 423
 424     /* Increment number of outer iterations */
 425     outeriter        += nri;
 426
 427     /* Update outer/inner flops */
 428
 429     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*63);
 430 }
 431 /*
 432  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_256_double
 433  * Electrostatics interaction: Coulomb
 434  * VdW interaction:            CubicSplineTable
 435  * Geometry:                   Particle-Particle
 436  * Calculate force/pot:        Force
 437  */
 438 void
 439 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_256_double
 440                     (t_nblist * gmx_restrict                nlist,
 441                      rvec * gmx_restrict                    xx,
 442                      rvec * gmx_restrict                    ff,
 443                      t_forcerec * gmx_restrict              fr,
 444                      t_mdatoms * gmx_restrict               mdatoms,
 445                      nb_kernel_data_t * gmx_restrict        kernel_data,
 446                      t_nrnb * gmx_restrict                  nrnb)
 447 {
 448     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 449      * just 0 for non-waters.
 450      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 451      * jnr indices corresponding to data put in the four positions in the SIMD register.
 452      */
 453     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 454     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 455     int              jnrA,jnrB,jnrC,jnrD;
 456     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 457     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 458     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 459     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 460     real             rcutoff_scalar;
 461     real             *shiftvec,*fshift,*x,*f;
 462     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 463     real             scratch[4*DIM];
 464     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 465     real *           vdwioffsetptr0;
 466     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 467     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 468     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 469     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 470     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 471     real             *charge;
 472     int              nvdwtype;
 473     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 474     int              *vdwtype;
 475     real             *vdwparam;
 476     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 477     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 478     __m128i          vfitab;
 479     __m128i          ifour       = _mm_set1_epi32(4);
 480     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 481     real             *vftab;
 482     __m256d          dummy_mask,cutoff_mask;
 483     __m128           tmpmask0,tmpmask1;
 484     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 485     __m256d          one     = _mm256_set1_pd(1.0);
 486     __m256d          two     = _mm256_set1_pd(2.0);
 487     x                = xx[0];
 488     f                = ff[0];
 489
 490     nri              = nlist->nri;
 491     iinr             = nlist->iinr;
 492     jindex           = nlist->jindex;
 493     jjnr             = nlist->jjnr;
 494     shiftidx         = nlist->shift;
 495     gid              = nlist->gid;
 496     shiftvec         = fr->shift_vec[0];
 497     fshift           = fr->fshift[0];
 498     facel            = _mm256_set1_pd(fr->epsfac);
 499     charge           = mdatoms->chargeA;
 500     nvdwtype         = fr->ntype;
 501     vdwparam         = fr->nbfp;
 502     vdwtype          = mdatoms->typeA;
 503
 504     vftab            = kernel_data->table_vdw->data;
 505     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 506
 507     /* Avoid stupid compiler warnings */
 508     jnrA = jnrB = jnrC = jnrD = 0;
 509     j_coord_offsetA = 0;
 510     j_coord_offsetB = 0;
 511     j_coord_offsetC = 0;
 512     j_coord_offsetD = 0;
 513
 514     outeriter        = 0;
 515     inneriter        = 0;
 516
 517     for(iidx=0;iidx<4*DIM;iidx++)
 518     {
 519         scratch[iidx] = 0.0;
 520     }
 521
 522     /* Start outer loop over neighborlists */
 523     for(iidx=0; iidx<nri; iidx++)
 524     {
 525         /* Load shift vector for this list */
 526         i_shift_offset   = DIM*shiftidx[iidx];
 527
 528         /* Load limits for loop over neighbors */
 529         j_index_start    = jindex[iidx];
 530         j_index_end      = jindex[iidx+1];
 531
 532         /* Get outer coordinate index */
 533         inr              = iinr[iidx];
 534         i_coord_offset   = DIM*inr;
 535
 536         /* Load i particle coords and add shift vector */
 537         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 538
 539         fix0             = _mm256_setzero_pd();
 540         fiy0             = _mm256_setzero_pd();
 541         fiz0             = _mm256_setzero_pd();
 542
 543         /* Load parameters for i particles */
 544         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 545         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 546
 547         /* Start inner kernel loop */
 548         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 549         {
 550
 551             /* Get j neighbor index, and coordinate index */
 552             jnrA             = jjnr[jidx];
 553             jnrB             = jjnr[jidx+1];
 554             jnrC             = jjnr[jidx+2];
 555             jnrD             = jjnr[jidx+3];
 556             j_coord_offsetA  = DIM*jnrA;
 557             j_coord_offsetB  = DIM*jnrB;
 558             j_coord_offsetC  = DIM*jnrC;
 559             j_coord_offsetD  = DIM*jnrD;
 560
 561             /* load j atom coordinates */
 562             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 563                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 564                                                  &jx0,&jy0,&jz0);
 565
 566             /* Calculate displacement vector */
 567             dx00             = _mm256_sub_pd(ix0,jx0);
 568             dy00             = _mm256_sub_pd(iy0,jy0);
 569             dz00             = _mm256_sub_pd(iz0,jz0);
 570
 571             /* Calculate squared distance and things based on it */
 572             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 573
 574             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 575
 576             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 577
 578             /* Load parameters for j particles */
 579             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 580                                                                  charge+jnrC+0,charge+jnrD+0);
 581             vdwjidx0A        = 2*vdwtype[jnrA+0];
 582             vdwjidx0B        = 2*vdwtype[jnrB+0];
 583             vdwjidx0C        = 2*vdwtype[jnrC+0];
 584             vdwjidx0D        = 2*vdwtype[jnrD+0];
 585
 586             /**************************
 587              * CALCULATE INTERACTIONS *
 588              **************************/
 589
 590             r00              = _mm256_mul_pd(rsq00,rinv00);
 591
 592             /* Compute parameters for interactions between i and j atoms */
 593             qq00             = _mm256_mul_pd(iq0,jq0);
 594             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 595                                             vdwioffsetptr0+vdwjidx0B,
 596                                             vdwioffsetptr0+vdwjidx0C,
 597                                             vdwioffsetptr0+vdwjidx0D,
 598                                             &c6_00,&c12_00);
 599
 600             /* Calculate table index by multiplying r with table scale and truncate to integer */
 601             rt               = _mm256_mul_pd(r00,vftabscale);
 602             vfitab           = _mm256_cvttpd_epi32(rt);
 603             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 604             vfitab           = _mm_slli_epi32(vfitab,3);
 605
 606             /* COULOMB ELECTROSTATICS */
 607             velec            = _mm256_mul_pd(qq00,rinv00);
 608             felec            = _mm256_mul_pd(velec,rinvsq00);
 609
 610             /* CUBIC SPLINE TABLE DISPERSION */
 611             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 612             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 613             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 614             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 615             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 616             Heps             = _mm256_mul_pd(vfeps,H);
 617             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 618             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 619             fvdw6            = _mm256_mul_pd(c6_00,FF);
 620
 621             /* CUBIC SPLINE TABLE REPULSION */
 622             vfitab           = _mm_add_epi32(vfitab,ifour);
 623             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 624             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 625             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 626             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 627             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 628             Heps             = _mm256_mul_pd(vfeps,H);
 629             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 630             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 631             fvdw12           = _mm256_mul_pd(c12_00,FF);
 632             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 633
 634             fscal            = _mm256_add_pd(felec,fvdw);
 635
 636             /* Calculate temporary vectorial force */
 637             tx               = _mm256_mul_pd(fscal,dx00);
 638             ty               = _mm256_mul_pd(fscal,dy00);
 639             tz               = _mm256_mul_pd(fscal,dz00);
 640
 641             /* Update vectorial force */
 642             fix0             = _mm256_add_pd(fix0,tx);
 643             fiy0             = _mm256_add_pd(fiy0,ty);
 644             fiz0             = _mm256_add_pd(fiz0,tz);
 645
 646             fjptrA             = f+j_coord_offsetA;
 647             fjptrB             = f+j_coord_offsetB;
 648             fjptrC             = f+j_coord_offsetC;
 649             fjptrD             = f+j_coord_offsetD;
 650             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 651
 652             /* Inner loop uses 53 flops */
 653         }
 654
 655         if(jidx<j_index_end)
 656         {
 657
 658             /* Get j neighbor index, and coordinate index */
 659             jnrlistA         = jjnr[jidx];
 660             jnrlistB         = jjnr[jidx+1];
 661             jnrlistC         = jjnr[jidx+2];
 662             jnrlistD         = jjnr[jidx+3];
 663             /* Sign of each element will be negative for non-real atoms.
 664              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 665              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 666              */
 667             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 668
 669             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 670             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 671             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 672
 673             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 674             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 675             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 676             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 677             j_coord_offsetA  = DIM*jnrA;
 678             j_coord_offsetB  = DIM*jnrB;
 679             j_coord_offsetC  = DIM*jnrC;
 680             j_coord_offsetD  = DIM*jnrD;
 681
 682             /* load j atom coordinates */
 683             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 684                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 685                                                  &jx0,&jy0,&jz0);
 686
 687             /* Calculate displacement vector */
 688             dx00             = _mm256_sub_pd(ix0,jx0);
 689             dy00             = _mm256_sub_pd(iy0,jy0);
 690             dz00             = _mm256_sub_pd(iz0,jz0);
 691
 692             /* Calculate squared distance and things based on it */
 693             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 694
 695             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 696
 697             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 698
 699             /* Load parameters for j particles */
 700             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 701                                                                  charge+jnrC+0,charge+jnrD+0);
 702             vdwjidx0A        = 2*vdwtype[jnrA+0];
 703             vdwjidx0B        = 2*vdwtype[jnrB+0];
 704             vdwjidx0C        = 2*vdwtype[jnrC+0];
 705             vdwjidx0D        = 2*vdwtype[jnrD+0];
 706
 707             /**************************
 708              * CALCULATE INTERACTIONS *
 709              **************************/
 710
 711             r00              = _mm256_mul_pd(rsq00,rinv00);
 712             r00              = _mm256_andnot_pd(dummy_mask,r00);
 713
 714             /* Compute parameters for interactions between i and j atoms */
 715             qq00             = _mm256_mul_pd(iq0,jq0);
 716             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 717                                             vdwioffsetptr0+vdwjidx0B,
 718                                             vdwioffsetptr0+vdwjidx0C,
 719                                             vdwioffsetptr0+vdwjidx0D,
 720                                             &c6_00,&c12_00);
 721
 722             /* Calculate table index by multiplying r with table scale and truncate to integer */
 723             rt               = _mm256_mul_pd(r00,vftabscale);
 724             vfitab           = _mm256_cvttpd_epi32(rt);
 725             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 726             vfitab           = _mm_slli_epi32(vfitab,3);
 727
 728             /* COULOMB ELECTROSTATICS */
 729             velec            = _mm256_mul_pd(qq00,rinv00);
 730             felec            = _mm256_mul_pd(velec,rinvsq00);
 731
 732             /* CUBIC SPLINE TABLE DISPERSION */
 733             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 734             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 735             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 736             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 737             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 738             Heps             = _mm256_mul_pd(vfeps,H);
 739             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 740             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 741             fvdw6            = _mm256_mul_pd(c6_00,FF);
 742
 743             /* CUBIC SPLINE TABLE REPULSION */
 744             vfitab           = _mm_add_epi32(vfitab,ifour);
 745             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 746             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 747             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 748             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 749             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 750             Heps             = _mm256_mul_pd(vfeps,H);
 751             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 752             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 753             fvdw12           = _mm256_mul_pd(c12_00,FF);
 754             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 755
 756             fscal            = _mm256_add_pd(felec,fvdw);
 757
 758             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 759
 760             /* Calculate temporary vectorial force */
 761             tx               = _mm256_mul_pd(fscal,dx00);
 762             ty               = _mm256_mul_pd(fscal,dy00);
 763             tz               = _mm256_mul_pd(fscal,dz00);
 764
 765             /* Update vectorial force */
 766             fix0             = _mm256_add_pd(fix0,tx);
 767             fiy0             = _mm256_add_pd(fiy0,ty);
 768             fiz0             = _mm256_add_pd(fiz0,tz);
 769
 770             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 771             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 772             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 773             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 774             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 775
 776             /* Inner loop uses 54 flops */
 777         }
 778
 779         /* End of innermost loop */
 780
 781         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 782                                                  f+i_coord_offset,fshift+i_shift_offset);
 783
 784         /* Increment number of inner iterations */
 785         inneriter                  += j_index_end - j_index_start;
 786
 787         /* Outer loop uses 7 flops */
 788     }
 789
 790     /* Increment number of outer iterations */
 791     outeriter        += nri;
 792
 793     /* Update outer/inner flops */
 794
 795     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*54);
 796 }