src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_256_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  79     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  84     real             scratch[4*DIM];
  85     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  98     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  99     __m128i          vfitab;
 100     __m128i          ifour       = _mm_set1_epi32(4);
 101     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 102     real             *vftab;
 103     __m256d          dummy_mask,cutoff_mask;
 104     __m128           tmpmask0,tmpmask1;
 105     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 106     __m256d          one     = _mm256_set1_pd(1.0);
 107     __m256d          two     = _mm256_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm256_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     vftab            = kernel_data->table_vdw->data;
 126     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 127
 128     /* Avoid stupid compiler warnings */
 129     jnrA = jnrB = jnrC = jnrD = 0;
 130     j_coord_offsetA = 0;
 131     j_coord_offsetB = 0;
 132     j_coord_offsetC = 0;
 133     j_coord_offsetD = 0;
 134
 135     outeriter        = 0;
 136     inneriter        = 0;
 137
 138     for(iidx=0;iidx<4*DIM;iidx++)
 139     {
 140         scratch[iidx] = 0.0;
 141     }
 142
 143     /* Start outer loop over neighborlists */
 144     for(iidx=0; iidx<nri; iidx++)
 145     {
 146         /* Load shift vector for this list */
 147         i_shift_offset   = DIM*shiftidx[iidx];
 148
 149         /* Load limits for loop over neighbors */
 150         j_index_start    = jindex[iidx];
 151         j_index_end      = jindex[iidx+1];
 152
 153         /* Get outer coordinate index */
 154         inr              = iinr[iidx];
 155         i_coord_offset   = DIM*inr;
 156
 157         /* Load i particle coords and add shift vector */
 158         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 159
 160         fix0             = _mm256_setzero_pd();
 161         fiy0             = _mm256_setzero_pd();
 162         fiz0             = _mm256_setzero_pd();
 163
 164         /* Load parameters for i particles */
 165         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 166         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 167
 168         /* Reset potential sums */
 169         velecsum         = _mm256_setzero_pd();
 170         vvdwsum          = _mm256_setzero_pd();
 171
 172         /* Start inner kernel loop */
 173         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 174         {
 175
 176             /* Get j neighbor index, and coordinate index */
 177             jnrA             = jjnr[jidx];
 178             jnrB             = jjnr[jidx+1];
 179             jnrC             = jjnr[jidx+2];
 180             jnrD             = jjnr[jidx+3];
 181             j_coord_offsetA  = DIM*jnrA;
 182             j_coord_offsetB  = DIM*jnrB;
 183             j_coord_offsetC  = DIM*jnrC;
 184             j_coord_offsetD  = DIM*jnrD;
 185
 186             /* load j atom coordinates */
 187             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 188                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 189                                                  &jx0,&jy0,&jz0);
 190
 191             /* Calculate displacement vector */
 192             dx00             = _mm256_sub_pd(ix0,jx0);
 193             dy00             = _mm256_sub_pd(iy0,jy0);
 194             dz00             = _mm256_sub_pd(iz0,jz0);
 195
 196             /* Calculate squared distance and things based on it */
 197             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 198
 199             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 200
 201             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 202
 203             /* Load parameters for j particles */
 204             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 205                                                                  charge+jnrC+0,charge+jnrD+0);
 206             vdwjidx0A        = 2*vdwtype[jnrA+0];
 207             vdwjidx0B        = 2*vdwtype[jnrB+0];
 208             vdwjidx0C        = 2*vdwtype[jnrC+0];
 209             vdwjidx0D        = 2*vdwtype[jnrD+0];
 210
 211             /**************************
 212              * CALCULATE INTERACTIONS *
 213              **************************/
 214
 215             r00              = _mm256_mul_pd(rsq00,rinv00);
 216
 217             /* Compute parameters for interactions between i and j atoms */
 218             qq00             = _mm256_mul_pd(iq0,jq0);
 219             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 220                                             vdwioffsetptr0+vdwjidx0B,
 221                                             vdwioffsetptr0+vdwjidx0C,
 222                                             vdwioffsetptr0+vdwjidx0D,
 223                                             &c6_00,&c12_00);
 224
 225             /* Calculate table index by multiplying r with table scale and truncate to integer */
 226             rt               = _mm256_mul_pd(r00,vftabscale);
 227             vfitab           = _mm256_cvttpd_epi32(rt);
 228             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 229             vfitab           = _mm_slli_epi32(vfitab,3);
 230
 231             /* COULOMB ELECTROSTATICS */
 232             velec            = _mm256_mul_pd(qq00,rinv00);
 233             felec            = _mm256_mul_pd(velec,rinvsq00);
 234
 235             /* CUBIC SPLINE TABLE DISPERSION */
 236             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 237             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 238             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 239             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 240             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 241             Heps             = _mm256_mul_pd(vfeps,H);
 242             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 243             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 244             vvdw6            = _mm256_mul_pd(c6_00,VV);
 245             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 246             fvdw6            = _mm256_mul_pd(c6_00,FF);
 247
 248             /* CUBIC SPLINE TABLE REPULSION */
 249             vfitab           = _mm_add_epi32(vfitab,ifour);
 250             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 251             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 252             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 253             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 254             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 255             Heps             = _mm256_mul_pd(vfeps,H);
 256             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 257             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 258             vvdw12           = _mm256_mul_pd(c12_00,VV);
 259             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 260             fvdw12           = _mm256_mul_pd(c12_00,FF);
 261             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 262             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 263
 264             /* Update potential sum for this i atom from the interaction with this j atom. */
 265             velecsum         = _mm256_add_pd(velecsum,velec);
 266             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 267
 268             fscal            = _mm256_add_pd(felec,fvdw);
 269
 270             /* Calculate temporary vectorial force */
 271             tx               = _mm256_mul_pd(fscal,dx00);
 272             ty               = _mm256_mul_pd(fscal,dy00);
 273             tz               = _mm256_mul_pd(fscal,dz00);
 274
 275             /* Update vectorial force */
 276             fix0             = _mm256_add_pd(fix0,tx);
 277             fiy0             = _mm256_add_pd(fiy0,ty);
 278             fiz0             = _mm256_add_pd(fiz0,tz);
 279
 280             fjptrA             = f+j_coord_offsetA;
 281             fjptrB             = f+j_coord_offsetB;
 282             fjptrC             = f+j_coord_offsetC;
 283             fjptrD             = f+j_coord_offsetD;
 284             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 285
 286             /* Inner loop uses 62 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             /* Sign of each element will be negative for non-real atoms.
 298              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 299              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 300              */
 301             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 302
 303             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 304             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 305             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 306
 307             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 308             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 309             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 310             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 311             j_coord_offsetA  = DIM*jnrA;
 312             j_coord_offsetB  = DIM*jnrB;
 313             j_coord_offsetC  = DIM*jnrC;
 314             j_coord_offsetD  = DIM*jnrD;
 315
 316             /* load j atom coordinates */
 317             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 318                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 319                                                  &jx0,&jy0,&jz0);
 320
 321             /* Calculate displacement vector */
 322             dx00             = _mm256_sub_pd(ix0,jx0);
 323             dy00             = _mm256_sub_pd(iy0,jy0);
 324             dz00             = _mm256_sub_pd(iz0,jz0);
 325
 326             /* Calculate squared distance and things based on it */
 327             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 328
 329             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 330
 331             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 332
 333             /* Load parameters for j particles */
 334             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 335                                                                  charge+jnrC+0,charge+jnrD+0);
 336             vdwjidx0A        = 2*vdwtype[jnrA+0];
 337             vdwjidx0B        = 2*vdwtype[jnrB+0];
 338             vdwjidx0C        = 2*vdwtype[jnrC+0];
 339             vdwjidx0D        = 2*vdwtype[jnrD+0];
 340
 341             /**************************
 342              * CALCULATE INTERACTIONS *
 343              **************************/
 344
 345             r00              = _mm256_mul_pd(rsq00,rinv00);
 346             r00              = _mm256_andnot_pd(dummy_mask,r00);
 347
 348             /* Compute parameters for interactions between i and j atoms */
 349             qq00             = _mm256_mul_pd(iq0,jq0);
 350             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 351                                             vdwioffsetptr0+vdwjidx0B,
 352                                             vdwioffsetptr0+vdwjidx0C,
 353                                             vdwioffsetptr0+vdwjidx0D,
 354                                             &c6_00,&c12_00);
 355
 356             /* Calculate table index by multiplying r with table scale and truncate to integer */
 357             rt               = _mm256_mul_pd(r00,vftabscale);
 358             vfitab           = _mm256_cvttpd_epi32(rt);
 359             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 360             vfitab           = _mm_slli_epi32(vfitab,3);
 361
 362             /* COULOMB ELECTROSTATICS */
 363             velec            = _mm256_mul_pd(qq00,rinv00);
 364             felec            = _mm256_mul_pd(velec,rinvsq00);
 365
 366             /* CUBIC SPLINE TABLE DISPERSION */
 367             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 368             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 369             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 370             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 371             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 372             Heps             = _mm256_mul_pd(vfeps,H);
 373             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 374             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 375             vvdw6            = _mm256_mul_pd(c6_00,VV);
 376             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 377             fvdw6            = _mm256_mul_pd(c6_00,FF);
 378
 379             /* CUBIC SPLINE TABLE REPULSION */
 380             vfitab           = _mm_add_epi32(vfitab,ifour);
 381             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 382             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 383             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 384             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 385             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 386             Heps             = _mm256_mul_pd(vfeps,H);
 387             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 388             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 389             vvdw12           = _mm256_mul_pd(c12_00,VV);
 390             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 391             fvdw12           = _mm256_mul_pd(c12_00,FF);
 392             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 393             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velec            = _mm256_andnot_pd(dummy_mask,velec);
 397             velecsum         = _mm256_add_pd(velecsum,velec);
 398             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 399             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 400
 401             fscal            = _mm256_add_pd(felec,fvdw);
 402
 403             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 404
 405             /* Calculate temporary vectorial force */
 406             tx               = _mm256_mul_pd(fscal,dx00);
 407             ty               = _mm256_mul_pd(fscal,dy00);
 408             tz               = _mm256_mul_pd(fscal,dz00);
 409
 410             /* Update vectorial force */
 411             fix0             = _mm256_add_pd(fix0,tx);
 412             fiy0             = _mm256_add_pd(fiy0,ty);
 413             fiz0             = _mm256_add_pd(fiz0,tz);
 414
 415             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 416             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 417             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 418             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 419             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 420
 421             /* Inner loop uses 63 flops */
 422         }
 423
 424         /* End of innermost loop */
 425
 426         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 427                                                  f+i_coord_offset,fshift+i_shift_offset);
 428
 429         ggid                        = gid[iidx];
 430         /* Update potential energies */
 431         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 432         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 433
 434         /* Increment number of inner iterations */
 435         inneriter                  += j_index_end - j_index_start;
 436
 437         /* Outer loop uses 9 flops */
 438     }
 439
 440     /* Increment number of outer iterations */
 441     outeriter        += nri;
 442
 443     /* Update outer/inner flops */
 444
 445     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*63);
 446 }
 447 /*
 448  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_256_double
 449  * Electrostatics interaction: Coulomb
 450  * VdW interaction:            CubicSplineTable
 451  * Geometry:                   Particle-Particle
 452  * Calculate force/pot:        Force
 453  */
 454 void
 455 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_256_double
 456                     (t_nblist                    * gmx_restrict       nlist,
 457                      rvec                        * gmx_restrict          xx,
 458                      rvec                        * gmx_restrict          ff,
 459                      t_forcerec                  * gmx_restrict          fr,
 460                      t_mdatoms                   * gmx_restrict     mdatoms,
 461                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 462                      t_nrnb                      * gmx_restrict        nrnb)
 463 {
 464     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 465      * just 0 for non-waters.
 466      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 467      * jnr indices corresponding to data put in the four positions in the SIMD register.
 468      */
 469     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 470     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 471     int              jnrA,jnrB,jnrC,jnrD;
 472     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 473     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 474     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 475     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 476     real             rcutoff_scalar;
 477     real             *shiftvec,*fshift,*x,*f;
 478     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 479     real             scratch[4*DIM];
 480     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 481     real *           vdwioffsetptr0;
 482     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 483     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 484     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 485     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 486     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 487     real             *charge;
 488     int              nvdwtype;
 489     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 490     int              *vdwtype;
 491     real             *vdwparam;
 492     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 493     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 494     __m128i          vfitab;
 495     __m128i          ifour       = _mm_set1_epi32(4);
 496     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 497     real             *vftab;
 498     __m256d          dummy_mask,cutoff_mask;
 499     __m128           tmpmask0,tmpmask1;
 500     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 501     __m256d          one     = _mm256_set1_pd(1.0);
 502     __m256d          two     = _mm256_set1_pd(2.0);
 503     x                = xx[0];
 504     f                = ff[0];
 505
 506     nri              = nlist->nri;
 507     iinr             = nlist->iinr;
 508     jindex           = nlist->jindex;
 509     jjnr             = nlist->jjnr;
 510     shiftidx         = nlist->shift;
 511     gid              = nlist->gid;
 512     shiftvec         = fr->shift_vec[0];
 513     fshift           = fr->fshift[0];
 514     facel            = _mm256_set1_pd(fr->epsfac);
 515     charge           = mdatoms->chargeA;
 516     nvdwtype         = fr->ntype;
 517     vdwparam         = fr->nbfp;
 518     vdwtype          = mdatoms->typeA;
 519
 520     vftab            = kernel_data->table_vdw->data;
 521     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 522
 523     /* Avoid stupid compiler warnings */
 524     jnrA = jnrB = jnrC = jnrD = 0;
 525     j_coord_offsetA = 0;
 526     j_coord_offsetB = 0;
 527     j_coord_offsetC = 0;
 528     j_coord_offsetD = 0;
 529
 530     outeriter        = 0;
 531     inneriter        = 0;
 532
 533     for(iidx=0;iidx<4*DIM;iidx++)
 534     {
 535         scratch[iidx] = 0.0;
 536     }
 537
 538     /* Start outer loop over neighborlists */
 539     for(iidx=0; iidx<nri; iidx++)
 540     {
 541         /* Load shift vector for this list */
 542         i_shift_offset   = DIM*shiftidx[iidx];
 543
 544         /* Load limits for loop over neighbors */
 545         j_index_start    = jindex[iidx];
 546         j_index_end      = jindex[iidx+1];
 547
 548         /* Get outer coordinate index */
 549         inr              = iinr[iidx];
 550         i_coord_offset   = DIM*inr;
 551
 552         /* Load i particle coords and add shift vector */
 553         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 554
 555         fix0             = _mm256_setzero_pd();
 556         fiy0             = _mm256_setzero_pd();
 557         fiz0             = _mm256_setzero_pd();
 558
 559         /* Load parameters for i particles */
 560         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 561         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 562
 563         /* Start inner kernel loop */
 564         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 565         {
 566
 567             /* Get j neighbor index, and coordinate index */
 568             jnrA             = jjnr[jidx];
 569             jnrB             = jjnr[jidx+1];
 570             jnrC             = jjnr[jidx+2];
 571             jnrD             = jjnr[jidx+3];
 572             j_coord_offsetA  = DIM*jnrA;
 573             j_coord_offsetB  = DIM*jnrB;
 574             j_coord_offsetC  = DIM*jnrC;
 575             j_coord_offsetD  = DIM*jnrD;
 576
 577             /* load j atom coordinates */
 578             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 579                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 580                                                  &jx0,&jy0,&jz0);
 581
 582             /* Calculate displacement vector */
 583             dx00             = _mm256_sub_pd(ix0,jx0);
 584             dy00             = _mm256_sub_pd(iy0,jy0);
 585             dz00             = _mm256_sub_pd(iz0,jz0);
 586
 587             /* Calculate squared distance and things based on it */
 588             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 589
 590             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 591
 592             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 593
 594             /* Load parameters for j particles */
 595             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 596                                                                  charge+jnrC+0,charge+jnrD+0);
 597             vdwjidx0A        = 2*vdwtype[jnrA+0];
 598             vdwjidx0B        = 2*vdwtype[jnrB+0];
 599             vdwjidx0C        = 2*vdwtype[jnrC+0];
 600             vdwjidx0D        = 2*vdwtype[jnrD+0];
 601
 602             /**************************
 603              * CALCULATE INTERACTIONS *
 604              **************************/
 605
 606             r00              = _mm256_mul_pd(rsq00,rinv00);
 607
 608             /* Compute parameters for interactions between i and j atoms */
 609             qq00             = _mm256_mul_pd(iq0,jq0);
 610             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 611                                             vdwioffsetptr0+vdwjidx0B,
 612                                             vdwioffsetptr0+vdwjidx0C,
 613                                             vdwioffsetptr0+vdwjidx0D,
 614                                             &c6_00,&c12_00);
 615
 616             /* Calculate table index by multiplying r with table scale and truncate to integer */
 617             rt               = _mm256_mul_pd(r00,vftabscale);
 618             vfitab           = _mm256_cvttpd_epi32(rt);
 619             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 620             vfitab           = _mm_slli_epi32(vfitab,3);
 621
 622             /* COULOMB ELECTROSTATICS */
 623             velec            = _mm256_mul_pd(qq00,rinv00);
 624             felec            = _mm256_mul_pd(velec,rinvsq00);
 625
 626             /* CUBIC SPLINE TABLE DISPERSION */
 627             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 628             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 629             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 630             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 631             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 632             Heps             = _mm256_mul_pd(vfeps,H);
 633             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 634             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 635             fvdw6            = _mm256_mul_pd(c6_00,FF);
 636
 637             /* CUBIC SPLINE TABLE REPULSION */
 638             vfitab           = _mm_add_epi32(vfitab,ifour);
 639             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 640             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 641             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 642             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 643             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 644             Heps             = _mm256_mul_pd(vfeps,H);
 645             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 646             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 647             fvdw12           = _mm256_mul_pd(c12_00,FF);
 648             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 649
 650             fscal            = _mm256_add_pd(felec,fvdw);
 651
 652             /* Calculate temporary vectorial force */
 653             tx               = _mm256_mul_pd(fscal,dx00);
 654             ty               = _mm256_mul_pd(fscal,dy00);
 655             tz               = _mm256_mul_pd(fscal,dz00);
 656
 657             /* Update vectorial force */
 658             fix0             = _mm256_add_pd(fix0,tx);
 659             fiy0             = _mm256_add_pd(fiy0,ty);
 660             fiz0             = _mm256_add_pd(fiz0,tz);
 661
 662             fjptrA             = f+j_coord_offsetA;
 663             fjptrB             = f+j_coord_offsetB;
 664             fjptrC             = f+j_coord_offsetC;
 665             fjptrD             = f+j_coord_offsetD;
 666             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 667
 668             /* Inner loop uses 53 flops */
 669         }
 670
 671         if(jidx<j_index_end)
 672         {
 673
 674             /* Get j neighbor index, and coordinate index */
 675             jnrlistA         = jjnr[jidx];
 676             jnrlistB         = jjnr[jidx+1];
 677             jnrlistC         = jjnr[jidx+2];
 678             jnrlistD         = jjnr[jidx+3];
 679             /* Sign of each element will be negative for non-real atoms.
 680              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 681              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 682              */
 683             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 684
 685             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 686             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 687             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 688
 689             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 690             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 691             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 692             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 693             j_coord_offsetA  = DIM*jnrA;
 694             j_coord_offsetB  = DIM*jnrB;
 695             j_coord_offsetC  = DIM*jnrC;
 696             j_coord_offsetD  = DIM*jnrD;
 697
 698             /* load j atom coordinates */
 699             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 700                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 701                                                  &jx0,&jy0,&jz0);
 702
 703             /* Calculate displacement vector */
 704             dx00             = _mm256_sub_pd(ix0,jx0);
 705             dy00             = _mm256_sub_pd(iy0,jy0);
 706             dz00             = _mm256_sub_pd(iz0,jz0);
 707
 708             /* Calculate squared distance and things based on it */
 709             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 710
 711             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 712
 713             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 714
 715             /* Load parameters for j particles */
 716             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 717                                                                  charge+jnrC+0,charge+jnrD+0);
 718             vdwjidx0A        = 2*vdwtype[jnrA+0];
 719             vdwjidx0B        = 2*vdwtype[jnrB+0];
 720             vdwjidx0C        = 2*vdwtype[jnrC+0];
 721             vdwjidx0D        = 2*vdwtype[jnrD+0];
 722
 723             /**************************
 724              * CALCULATE INTERACTIONS *
 725              **************************/
 726
 727             r00              = _mm256_mul_pd(rsq00,rinv00);
 728             r00              = _mm256_andnot_pd(dummy_mask,r00);
 729
 730             /* Compute parameters for interactions between i and j atoms */
 731             qq00             = _mm256_mul_pd(iq0,jq0);
 732             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 733                                             vdwioffsetptr0+vdwjidx0B,
 734                                             vdwioffsetptr0+vdwjidx0C,
 735                                             vdwioffsetptr0+vdwjidx0D,
 736                                             &c6_00,&c12_00);
 737
 738             /* Calculate table index by multiplying r with table scale and truncate to integer */
 739             rt               = _mm256_mul_pd(r00,vftabscale);
 740             vfitab           = _mm256_cvttpd_epi32(rt);
 741             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 742             vfitab           = _mm_slli_epi32(vfitab,3);
 743
 744             /* COULOMB ELECTROSTATICS */
 745             velec            = _mm256_mul_pd(qq00,rinv00);
 746             felec            = _mm256_mul_pd(velec,rinvsq00);
 747
 748             /* CUBIC SPLINE TABLE DISPERSION */
 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             fvdw6            = _mm256_mul_pd(c6_00,FF);
 758
 759             /* CUBIC SPLINE TABLE REPULSION */
 760             vfitab           = _mm_add_epi32(vfitab,ifour);
 761             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 762             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 763             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 764             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 765             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 766             Heps             = _mm256_mul_pd(vfeps,H);
 767             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 768             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 769             fvdw12           = _mm256_mul_pd(c12_00,FF);
 770             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 771
 772             fscal            = _mm256_add_pd(felec,fvdw);
 773
 774             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 775
 776             /* Calculate temporary vectorial force */
 777             tx               = _mm256_mul_pd(fscal,dx00);
 778             ty               = _mm256_mul_pd(fscal,dy00);
 779             tz               = _mm256_mul_pd(fscal,dz00);
 780
 781             /* Update vectorial force */
 782             fix0             = _mm256_add_pd(fix0,tx);
 783             fiy0             = _mm256_add_pd(fiy0,ty);
 784             fiz0             = _mm256_add_pd(fiz0,tz);
 785
 786             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 787             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 788             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 789             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 790             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 791
 792             /* Inner loop uses 54 flops */
 793         }
 794
 795         /* End of innermost loop */
 796
 797         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 798                                                  f+i_coord_offset,fshift+i_shift_offset);
 799
 800         /* Increment number of inner iterations */
 801         inneriter                  += j_index_end - j_index_start;
 802
 803         /* Outer loop uses 7 flops */
 804     }
 805
 806     /* Increment number of outer iterations */
 807     outeriter        += nri;
 808
 809     /* Update outer/inner flops */
 810
 811     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*54);
 812 }