src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRF_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_ElecRF_VdwCSTab_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_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     krf              = _mm256_set1_pd(fr->ic->k_rf);
 122     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 123     crf              = _mm256_set1_pd(fr->ic->c_rf);
 124     nvdwtype         = fr->ntype;
 125     vdwparam         = fr->nbfp;
 126     vdwtype          = mdatoms->typeA;
 127
 128     vftab            = kernel_data->table_vdw->data;
 129     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 130
 131     /* Avoid stupid compiler warnings */
 132     jnrA = jnrB = jnrC = jnrD = 0;
 133     j_coord_offsetA = 0;
 134     j_coord_offsetB = 0;
 135     j_coord_offsetC = 0;
 136     j_coord_offsetD = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     for(iidx=0;iidx<4*DIM;iidx++)
 142     {
 143         scratch[iidx] = 0.0;
 144     }
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 162
 163         fix0             = _mm256_setzero_pd();
 164         fiy0             = _mm256_setzero_pd();
 165         fiz0             = _mm256_setzero_pd();
 166
 167         /* Load parameters for i particles */
 168         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 169         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 170
 171         /* Reset potential sums */
 172         velecsum         = _mm256_setzero_pd();
 173         vvdwsum          = _mm256_setzero_pd();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188
 189             /* load j atom coordinates */
 190             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 192                                                  &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm256_sub_pd(ix0,jx0);
 196             dy00             = _mm256_sub_pd(iy0,jy0);
 197             dz00             = _mm256_sub_pd(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 201
 202             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 203
 204             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 208                                                                  charge+jnrC+0,charge+jnrD+0);
 209             vdwjidx0A        = 2*vdwtype[jnrA+0];
 210             vdwjidx0B        = 2*vdwtype[jnrB+0];
 211             vdwjidx0C        = 2*vdwtype[jnrC+0];
 212             vdwjidx0D        = 2*vdwtype[jnrD+0];
 213
 214             /**************************
 215              * CALCULATE INTERACTIONS *
 216              **************************/
 217
 218             r00              = _mm256_mul_pd(rsq00,rinv00);
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             qq00             = _mm256_mul_pd(iq0,jq0);
 222             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 223                                             vdwioffsetptr0+vdwjidx0B,
 224                                             vdwioffsetptr0+vdwjidx0C,
 225                                             vdwioffsetptr0+vdwjidx0D,
 226                                             &c6_00,&c12_00);
 227
 228             /* Calculate table index by multiplying r with table scale and truncate to integer */
 229             rt               = _mm256_mul_pd(r00,vftabscale);
 230             vfitab           = _mm256_cvttpd_epi32(rt);
 231             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 232             vfitab           = _mm_slli_epi32(vfitab,3);
 233
 234             /* REACTION-FIELD ELECTROSTATICS */
 235             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 236             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 237
 238             /* CUBIC SPLINE TABLE DISPERSION */
 239             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 240             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 241             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 242             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 243             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 244             Heps             = _mm256_mul_pd(vfeps,H);
 245             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 246             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 247             vvdw6            = _mm256_mul_pd(c6_00,VV);
 248             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 249             fvdw6            = _mm256_mul_pd(c6_00,FF);
 250
 251             /* CUBIC SPLINE TABLE REPULSION */
 252             vfitab           = _mm_add_epi32(vfitab,ifour);
 253             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 254             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 255             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 256             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 257             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 258             Heps             = _mm256_mul_pd(vfeps,H);
 259             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 260             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 261             vvdw12           = _mm256_mul_pd(c12_00,VV);
 262             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 263             fvdw12           = _mm256_mul_pd(c12_00,FF);
 264             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 265             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 266
 267             /* Update potential sum for this i atom from the interaction with this j atom. */
 268             velecsum         = _mm256_add_pd(velecsum,velec);
 269             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 270
 271             fscal            = _mm256_add_pd(felec,fvdw);
 272
 273             /* Calculate temporary vectorial force */
 274             tx               = _mm256_mul_pd(fscal,dx00);
 275             ty               = _mm256_mul_pd(fscal,dy00);
 276             tz               = _mm256_mul_pd(fscal,dz00);
 277
 278             /* Update vectorial force */
 279             fix0             = _mm256_add_pd(fix0,tx);
 280             fiy0             = _mm256_add_pd(fiy0,ty);
 281             fiz0             = _mm256_add_pd(fiz0,tz);
 282
 283             fjptrA             = f+j_coord_offsetA;
 284             fjptrB             = f+j_coord_offsetB;
 285             fjptrC             = f+j_coord_offsetC;
 286             fjptrD             = f+j_coord_offsetD;
 287             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 288
 289             /* Inner loop uses 67 flops */
 290         }
 291
 292         if(jidx<j_index_end)
 293         {
 294
 295             /* Get j neighbor index, and coordinate index */
 296             jnrlistA         = jjnr[jidx];
 297             jnrlistB         = jjnr[jidx+1];
 298             jnrlistC         = jjnr[jidx+2];
 299             jnrlistD         = jjnr[jidx+3];
 300             /* Sign of each element will be negative for non-real atoms.
 301              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 302              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 303              */
 304             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 305
 306             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 307             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 308             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 309
 310             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 311             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 312             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 313             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 314             j_coord_offsetA  = DIM*jnrA;
 315             j_coord_offsetB  = DIM*jnrB;
 316             j_coord_offsetC  = DIM*jnrC;
 317             j_coord_offsetD  = DIM*jnrD;
 318
 319             /* load j atom coordinates */
 320             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 321                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 322                                                  &jx0,&jy0,&jz0);
 323
 324             /* Calculate displacement vector */
 325             dx00             = _mm256_sub_pd(ix0,jx0);
 326             dy00             = _mm256_sub_pd(iy0,jy0);
 327             dz00             = _mm256_sub_pd(iz0,jz0);
 328
 329             /* Calculate squared distance and things based on it */
 330             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 331
 332             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 333
 334             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 335
 336             /* Load parameters for j particles */
 337             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 338                                                                  charge+jnrC+0,charge+jnrD+0);
 339             vdwjidx0A        = 2*vdwtype[jnrA+0];
 340             vdwjidx0B        = 2*vdwtype[jnrB+0];
 341             vdwjidx0C        = 2*vdwtype[jnrC+0];
 342             vdwjidx0D        = 2*vdwtype[jnrD+0];
 343
 344             /**************************
 345              * CALCULATE INTERACTIONS *
 346              **************************/
 347
 348             r00              = _mm256_mul_pd(rsq00,rinv00);
 349             r00              = _mm256_andnot_pd(dummy_mask,r00);
 350
 351             /* Compute parameters for interactions between i and j atoms */
 352             qq00             = _mm256_mul_pd(iq0,jq0);
 353             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 354                                             vdwioffsetptr0+vdwjidx0B,
 355                                             vdwioffsetptr0+vdwjidx0C,
 356                                             vdwioffsetptr0+vdwjidx0D,
 357                                             &c6_00,&c12_00);
 358
 359             /* Calculate table index by multiplying r with table scale and truncate to integer */
 360             rt               = _mm256_mul_pd(r00,vftabscale);
 361             vfitab           = _mm256_cvttpd_epi32(rt);
 362             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 363             vfitab           = _mm_slli_epi32(vfitab,3);
 364
 365             /* REACTION-FIELD ELECTROSTATICS */
 366             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 367             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 368
 369             /* CUBIC SPLINE TABLE DISPERSION */
 370             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 371             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 372             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 373             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 374             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 375             Heps             = _mm256_mul_pd(vfeps,H);
 376             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 377             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 378             vvdw6            = _mm256_mul_pd(c6_00,VV);
 379             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 380             fvdw6            = _mm256_mul_pd(c6_00,FF);
 381
 382             /* CUBIC SPLINE TABLE REPULSION */
 383             vfitab           = _mm_add_epi32(vfitab,ifour);
 384             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 385             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 386             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 387             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 388             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 389             Heps             = _mm256_mul_pd(vfeps,H);
 390             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 391             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 392             vvdw12           = _mm256_mul_pd(c12_00,VV);
 393             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 394             fvdw12           = _mm256_mul_pd(c12_00,FF);
 395             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 396             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 397
 398             /* Update potential sum for this i atom from the interaction with this j atom. */
 399             velec            = _mm256_andnot_pd(dummy_mask,velec);
 400             velecsum         = _mm256_add_pd(velecsum,velec);
 401             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 402             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 403
 404             fscal            = _mm256_add_pd(felec,fvdw);
 405
 406             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 407
 408             /* Calculate temporary vectorial force */
 409             tx               = _mm256_mul_pd(fscal,dx00);
 410             ty               = _mm256_mul_pd(fscal,dy00);
 411             tz               = _mm256_mul_pd(fscal,dz00);
 412
 413             /* Update vectorial force */
 414             fix0             = _mm256_add_pd(fix0,tx);
 415             fiy0             = _mm256_add_pd(fiy0,ty);
 416             fiz0             = _mm256_add_pd(fiz0,tz);
 417
 418             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 419             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 420             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 421             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 422             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 423
 424             /* Inner loop uses 68 flops */
 425         }
 426
 427         /* End of innermost loop */
 428
 429         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 430                                                  f+i_coord_offset,fshift+i_shift_offset);
 431
 432         ggid                        = gid[iidx];
 433         /* Update potential energies */
 434         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 435         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 436
 437         /* Increment number of inner iterations */
 438         inneriter                  += j_index_end - j_index_start;
 439
 440         /* Outer loop uses 9 flops */
 441     }
 442
 443     /* Increment number of outer iterations */
 444     outeriter        += nri;
 445
 446     /* Update outer/inner flops */
 447
 448     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*68);
 449 }
 450 /*
 451  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_256_double
 452  * Electrostatics interaction: ReactionField
 453  * VdW interaction:            CubicSplineTable
 454  * Geometry:                   Particle-Particle
 455  * Calculate force/pot:        Force
 456  */
 457 void
 458 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_256_double
 459                     (t_nblist                    * gmx_restrict       nlist,
 460                      rvec                        * gmx_restrict          xx,
 461                      rvec                        * gmx_restrict          ff,
 462                      t_forcerec                  * gmx_restrict          fr,
 463                      t_mdatoms                   * gmx_restrict     mdatoms,
 464                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 465                      t_nrnb                      * gmx_restrict        nrnb)
 466 {
 467     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 468      * just 0 for non-waters.
 469      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 470      * jnr indices corresponding to data put in the four positions in the SIMD register.
 471      */
 472     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 473     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 474     int              jnrA,jnrB,jnrC,jnrD;
 475     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 476     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 477     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 478     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 479     real             rcutoff_scalar;
 480     real             *shiftvec,*fshift,*x,*f;
 481     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 482     real             scratch[4*DIM];
 483     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 484     real *           vdwioffsetptr0;
 485     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 486     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 487     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 488     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 489     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 490     real             *charge;
 491     int              nvdwtype;
 492     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 493     int              *vdwtype;
 494     real             *vdwparam;
 495     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 496     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 497     __m128i          vfitab;
 498     __m128i          ifour       = _mm_set1_epi32(4);
 499     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 500     real             *vftab;
 501     __m256d          dummy_mask,cutoff_mask;
 502     __m128           tmpmask0,tmpmask1;
 503     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 504     __m256d          one     = _mm256_set1_pd(1.0);
 505     __m256d          two     = _mm256_set1_pd(2.0);
 506     x                = xx[0];
 507     f                = ff[0];
 508
 509     nri              = nlist->nri;
 510     iinr             = nlist->iinr;
 511     jindex           = nlist->jindex;
 512     jjnr             = nlist->jjnr;
 513     shiftidx         = nlist->shift;
 514     gid              = nlist->gid;
 515     shiftvec         = fr->shift_vec[0];
 516     fshift           = fr->fshift[0];
 517     facel            = _mm256_set1_pd(fr->epsfac);
 518     charge           = mdatoms->chargeA;
 519     krf              = _mm256_set1_pd(fr->ic->k_rf);
 520     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 521     crf              = _mm256_set1_pd(fr->ic->c_rf);
 522     nvdwtype         = fr->ntype;
 523     vdwparam         = fr->nbfp;
 524     vdwtype          = mdatoms->typeA;
 525
 526     vftab            = kernel_data->table_vdw->data;
 527     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 528
 529     /* Avoid stupid compiler warnings */
 530     jnrA = jnrB = jnrC = jnrD = 0;
 531     j_coord_offsetA = 0;
 532     j_coord_offsetB = 0;
 533     j_coord_offsetC = 0;
 534     j_coord_offsetD = 0;
 535
 536     outeriter        = 0;
 537     inneriter        = 0;
 538
 539     for(iidx=0;iidx<4*DIM;iidx++)
 540     {
 541         scratch[iidx] = 0.0;
 542     }
 543
 544     /* Start outer loop over neighborlists */
 545     for(iidx=0; iidx<nri; iidx++)
 546     {
 547         /* Load shift vector for this list */
 548         i_shift_offset   = DIM*shiftidx[iidx];
 549
 550         /* Load limits for loop over neighbors */
 551         j_index_start    = jindex[iidx];
 552         j_index_end      = jindex[iidx+1];
 553
 554         /* Get outer coordinate index */
 555         inr              = iinr[iidx];
 556         i_coord_offset   = DIM*inr;
 557
 558         /* Load i particle coords and add shift vector */
 559         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 560
 561         fix0             = _mm256_setzero_pd();
 562         fiy0             = _mm256_setzero_pd();
 563         fiz0             = _mm256_setzero_pd();
 564
 565         /* Load parameters for i particles */
 566         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 567         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 568
 569         /* Start inner kernel loop */
 570         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 571         {
 572
 573             /* Get j neighbor index, and coordinate index */
 574             jnrA             = jjnr[jidx];
 575             jnrB             = jjnr[jidx+1];
 576             jnrC             = jjnr[jidx+2];
 577             jnrD             = jjnr[jidx+3];
 578             j_coord_offsetA  = DIM*jnrA;
 579             j_coord_offsetB  = DIM*jnrB;
 580             j_coord_offsetC  = DIM*jnrC;
 581             j_coord_offsetD  = DIM*jnrD;
 582
 583             /* load j atom coordinates */
 584             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 585                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 586                                                  &jx0,&jy0,&jz0);
 587
 588             /* Calculate displacement vector */
 589             dx00             = _mm256_sub_pd(ix0,jx0);
 590             dy00             = _mm256_sub_pd(iy0,jy0);
 591             dz00             = _mm256_sub_pd(iz0,jz0);
 592
 593             /* Calculate squared distance and things based on it */
 594             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 595
 596             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 597
 598             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 599
 600             /* Load parameters for j particles */
 601             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 602                                                                  charge+jnrC+0,charge+jnrD+0);
 603             vdwjidx0A        = 2*vdwtype[jnrA+0];
 604             vdwjidx0B        = 2*vdwtype[jnrB+0];
 605             vdwjidx0C        = 2*vdwtype[jnrC+0];
 606             vdwjidx0D        = 2*vdwtype[jnrD+0];
 607
 608             /**************************
 609              * CALCULATE INTERACTIONS *
 610              **************************/
 611
 612             r00              = _mm256_mul_pd(rsq00,rinv00);
 613
 614             /* Compute parameters for interactions between i and j atoms */
 615             qq00             = _mm256_mul_pd(iq0,jq0);
 616             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 617                                             vdwioffsetptr0+vdwjidx0B,
 618                                             vdwioffsetptr0+vdwjidx0C,
 619                                             vdwioffsetptr0+vdwjidx0D,
 620                                             &c6_00,&c12_00);
 621
 622             /* Calculate table index by multiplying r with table scale and truncate to integer */
 623             rt               = _mm256_mul_pd(r00,vftabscale);
 624             vfitab           = _mm256_cvttpd_epi32(rt);
 625             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 626             vfitab           = _mm_slli_epi32(vfitab,3);
 627
 628             /* REACTION-FIELD ELECTROSTATICS */
 629             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 630
 631             /* CUBIC SPLINE TABLE DISPERSION */
 632             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 633             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 634             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 635             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 636             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 637             Heps             = _mm256_mul_pd(vfeps,H);
 638             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 639             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 640             fvdw6            = _mm256_mul_pd(c6_00,FF);
 641
 642             /* CUBIC SPLINE TABLE REPULSION */
 643             vfitab           = _mm_add_epi32(vfitab,ifour);
 644             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 645             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 646             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 647             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 648             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 649             Heps             = _mm256_mul_pd(vfeps,H);
 650             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 651             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 652             fvdw12           = _mm256_mul_pd(c12_00,FF);
 653             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 654
 655             fscal            = _mm256_add_pd(felec,fvdw);
 656
 657             /* Calculate temporary vectorial force */
 658             tx               = _mm256_mul_pd(fscal,dx00);
 659             ty               = _mm256_mul_pd(fscal,dy00);
 660             tz               = _mm256_mul_pd(fscal,dz00);
 661
 662             /* Update vectorial force */
 663             fix0             = _mm256_add_pd(fix0,tx);
 664             fiy0             = _mm256_add_pd(fiy0,ty);
 665             fiz0             = _mm256_add_pd(fiz0,tz);
 666
 667             fjptrA             = f+j_coord_offsetA;
 668             fjptrB             = f+j_coord_offsetB;
 669             fjptrC             = f+j_coord_offsetC;
 670             fjptrD             = f+j_coord_offsetD;
 671             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 672
 673             /* Inner loop uses 54 flops */
 674         }
 675
 676         if(jidx<j_index_end)
 677         {
 678
 679             /* Get j neighbor index, and coordinate index */
 680             jnrlistA         = jjnr[jidx];
 681             jnrlistB         = jjnr[jidx+1];
 682             jnrlistC         = jjnr[jidx+2];
 683             jnrlistD         = jjnr[jidx+3];
 684             /* Sign of each element will be negative for non-real atoms.
 685              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 686              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 687              */
 688             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 689
 690             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 691             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 692             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 693
 694             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 695             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 696             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 697             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 698             j_coord_offsetA  = DIM*jnrA;
 699             j_coord_offsetB  = DIM*jnrB;
 700             j_coord_offsetC  = DIM*jnrC;
 701             j_coord_offsetD  = DIM*jnrD;
 702
 703             /* load j atom coordinates */
 704             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 705                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 706                                                  &jx0,&jy0,&jz0);
 707
 708             /* Calculate displacement vector */
 709             dx00             = _mm256_sub_pd(ix0,jx0);
 710             dy00             = _mm256_sub_pd(iy0,jy0);
 711             dz00             = _mm256_sub_pd(iz0,jz0);
 712
 713             /* Calculate squared distance and things based on it */
 714             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 715
 716             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 717
 718             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 719
 720             /* Load parameters for j particles */
 721             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 722                                                                  charge+jnrC+0,charge+jnrD+0);
 723             vdwjidx0A        = 2*vdwtype[jnrA+0];
 724             vdwjidx0B        = 2*vdwtype[jnrB+0];
 725             vdwjidx0C        = 2*vdwtype[jnrC+0];
 726             vdwjidx0D        = 2*vdwtype[jnrD+0];
 727
 728             /**************************
 729              * CALCULATE INTERACTIONS *
 730              **************************/
 731
 732             r00              = _mm256_mul_pd(rsq00,rinv00);
 733             r00              = _mm256_andnot_pd(dummy_mask,r00);
 734
 735             /* Compute parameters for interactions between i and j atoms */
 736             qq00             = _mm256_mul_pd(iq0,jq0);
 737             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 738                                             vdwioffsetptr0+vdwjidx0B,
 739                                             vdwioffsetptr0+vdwjidx0C,
 740                                             vdwioffsetptr0+vdwjidx0D,
 741                                             &c6_00,&c12_00);
 742
 743             /* Calculate table index by multiplying r with table scale and truncate to integer */
 744             rt               = _mm256_mul_pd(r00,vftabscale);
 745             vfitab           = _mm256_cvttpd_epi32(rt);
 746             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 747             vfitab           = _mm_slli_epi32(vfitab,3);
 748
 749             /* REACTION-FIELD ELECTROSTATICS */
 750             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 751
 752             /* CUBIC SPLINE TABLE DISPERSION */
 753             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 754             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 755             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 756             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 757             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 758             Heps             = _mm256_mul_pd(vfeps,H);
 759             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 760             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 761             fvdw6            = _mm256_mul_pd(c6_00,FF);
 762
 763             /* CUBIC SPLINE TABLE REPULSION */
 764             vfitab           = _mm_add_epi32(vfitab,ifour);
 765             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 766             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 767             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 768             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 769             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 770             Heps             = _mm256_mul_pd(vfeps,H);
 771             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 772             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 773             fvdw12           = _mm256_mul_pd(c12_00,FF);
 774             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 775
 776             fscal            = _mm256_add_pd(felec,fvdw);
 777
 778             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 779
 780             /* Calculate temporary vectorial force */
 781             tx               = _mm256_mul_pd(fscal,dx00);
 782             ty               = _mm256_mul_pd(fscal,dy00);
 783             tz               = _mm256_mul_pd(fscal,dz00);
 784
 785             /* Update vectorial force */
 786             fix0             = _mm256_add_pd(fix0,tx);
 787             fiy0             = _mm256_add_pd(fiy0,ty);
 788             fiz0             = _mm256_add_pd(fiz0,tz);
 789
 790             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 791             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 792             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 793             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 794             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 795
 796             /* Inner loop uses 55 flops */
 797         }
 798
 799         /* End of innermost loop */
 800
 801         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 802                                                  f+i_coord_offset,fshift+i_shift_offset);
 803
 804         /* Increment number of inner iterations */
 805         inneriter                  += j_index_end - j_index_start;
 806
 807         /* Outer loop uses 7 flops */
 808     }
 809
 810     /* Increment number of outer iterations */
 811     outeriter        += nri;
 812
 813     /* Update outer/inner flops */
 814
 815     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 816 }