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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  90     real             *charge;
  91     int              nvdwtype;
  92     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  93     int              *vdwtype;
  94     real             *vdwparam;
  95     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  96     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  97     __m128i          vfitab;
  98     __m128i          ifour       = _mm_set1_epi32(4);
  99     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 100     real             *vftab;
 101     __m256d          dummy_mask,cutoff_mask;
 102     __m128           tmpmask0,tmpmask1;
 103     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 104     __m256d          one     = _mm256_set1_pd(1.0);
 105     __m256d          two     = _mm256_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm256_set1_pd(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = _mm256_set1_pd(fr->ic->k_rf);
 120     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 121     crf              = _mm256_set1_pd(fr->ic->c_rf);
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     vftab            = kernel_data->table_vdw->data;
 127     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = jnrC = jnrD = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133     j_coord_offsetC = 0;
 134     j_coord_offsetD = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     for(iidx=0;iidx<4*DIM;iidx++)
 140     {
 141         scratch[iidx] = 0.0;
 142     }
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 160
 161         fix0             = _mm256_setzero_pd();
 162         fiy0             = _mm256_setzero_pd();
 163         fiz0             = _mm256_setzero_pd();
 164
 165         /* Load parameters for i particles */
 166         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 167         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 168
 169         /* Reset potential sums */
 170         velecsum         = _mm256_setzero_pd();
 171         vvdwsum          = _mm256_setzero_pd();
 172
 173         /* Start inner kernel loop */
 174         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 175         {
 176
 177             /* Get j neighbor index, and coordinate index */
 178             jnrA             = jjnr[jidx];
 179             jnrB             = jjnr[jidx+1];
 180             jnrC             = jjnr[jidx+2];
 181             jnrD             = jjnr[jidx+3];
 182             j_coord_offsetA  = DIM*jnrA;
 183             j_coord_offsetB  = DIM*jnrB;
 184             j_coord_offsetC  = DIM*jnrC;
 185             j_coord_offsetD  = DIM*jnrD;
 186
 187             /* load j atom coordinates */
 188             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 190                                                  &jx0,&jy0,&jz0);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm256_sub_pd(ix0,jx0);
 194             dy00             = _mm256_sub_pd(iy0,jy0);
 195             dz00             = _mm256_sub_pd(iz0,jz0);
 196
 197             /* Calculate squared distance and things based on it */
 198             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 199
 200             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 201
 202             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 203
 204             /* Load parameters for j particles */
 205             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 206                                                                  charge+jnrC+0,charge+jnrD+0);
 207             vdwjidx0A        = 2*vdwtype[jnrA+0];
 208             vdwjidx0B        = 2*vdwtype[jnrB+0];
 209             vdwjidx0C        = 2*vdwtype[jnrC+0];
 210             vdwjidx0D        = 2*vdwtype[jnrD+0];
 211
 212             /**************************
 213              * CALCULATE INTERACTIONS *
 214              **************************/
 215
 216             r00              = _mm256_mul_pd(rsq00,rinv00);
 217
 218             /* Compute parameters for interactions between i and j atoms */
 219             qq00             = _mm256_mul_pd(iq0,jq0);
 220             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 221                                             vdwioffsetptr0+vdwjidx0B,
 222                                             vdwioffsetptr0+vdwjidx0C,
 223                                             vdwioffsetptr0+vdwjidx0D,
 224                                             &c6_00,&c12_00);
 225
 226             /* Calculate table index by multiplying r with table scale and truncate to integer */
 227             rt               = _mm256_mul_pd(r00,vftabscale);
 228             vfitab           = _mm256_cvttpd_epi32(rt);
 229             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 230             vfitab           = _mm_slli_epi32(vfitab,3);
 231
 232             /* REACTION-FIELD ELECTROSTATICS */
 233             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 234             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 235
 236             /* CUBIC SPLINE TABLE DISPERSION */
 237             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 238             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 239             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 240             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 241             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 242             Heps             = _mm256_mul_pd(vfeps,H);
 243             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 244             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 245             vvdw6            = _mm256_mul_pd(c6_00,VV);
 246             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 247             fvdw6            = _mm256_mul_pd(c6_00,FF);
 248
 249             /* CUBIC SPLINE TABLE REPULSION */
 250             vfitab           = _mm_add_epi32(vfitab,ifour);
 251             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 252             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 253             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 254             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 255             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 256             Heps             = _mm256_mul_pd(vfeps,H);
 257             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 258             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 259             vvdw12           = _mm256_mul_pd(c12_00,VV);
 260             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 261             fvdw12           = _mm256_mul_pd(c12_00,FF);
 262             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 263             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 264
 265             /* Update potential sum for this i atom from the interaction with this j atom. */
 266             velecsum         = _mm256_add_pd(velecsum,velec);
 267             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 268
 269             fscal            = _mm256_add_pd(felec,fvdw);
 270
 271             /* Calculate temporary vectorial force */
 272             tx               = _mm256_mul_pd(fscal,dx00);
 273             ty               = _mm256_mul_pd(fscal,dy00);
 274             tz               = _mm256_mul_pd(fscal,dz00);
 275
 276             /* Update vectorial force */
 277             fix0             = _mm256_add_pd(fix0,tx);
 278             fiy0             = _mm256_add_pd(fiy0,ty);
 279             fiz0             = _mm256_add_pd(fiz0,tz);
 280
 281             fjptrA             = f+j_coord_offsetA;
 282             fjptrB             = f+j_coord_offsetB;
 283             fjptrC             = f+j_coord_offsetC;
 284             fjptrD             = f+j_coord_offsetD;
 285             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 286
 287             /* Inner loop uses 67 flops */
 288         }
 289
 290         if(jidx<j_index_end)
 291         {
 292
 293             /* Get j neighbor index, and coordinate index */
 294             jnrlistA         = jjnr[jidx];
 295             jnrlistB         = jjnr[jidx+1];
 296             jnrlistC         = jjnr[jidx+2];
 297             jnrlistD         = jjnr[jidx+3];
 298             /* Sign of each element will be negative for non-real atoms.
 299              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 300              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 301              */
 302             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 303
 304             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 305             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 306             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 307
 308             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 309             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 310             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 311             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 312             j_coord_offsetA  = DIM*jnrA;
 313             j_coord_offsetB  = DIM*jnrB;
 314             j_coord_offsetC  = DIM*jnrC;
 315             j_coord_offsetD  = DIM*jnrD;
 316
 317             /* load j atom coordinates */
 318             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 319                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 320                                                  &jx0,&jy0,&jz0);
 321
 322             /* Calculate displacement vector */
 323             dx00             = _mm256_sub_pd(ix0,jx0);
 324             dy00             = _mm256_sub_pd(iy0,jy0);
 325             dz00             = _mm256_sub_pd(iz0,jz0);
 326
 327             /* Calculate squared distance and things based on it */
 328             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 329
 330             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 331
 332             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 333
 334             /* Load parameters for j particles */
 335             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 336                                                                  charge+jnrC+0,charge+jnrD+0);
 337             vdwjidx0A        = 2*vdwtype[jnrA+0];
 338             vdwjidx0B        = 2*vdwtype[jnrB+0];
 339             vdwjidx0C        = 2*vdwtype[jnrC+0];
 340             vdwjidx0D        = 2*vdwtype[jnrD+0];
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             r00              = _mm256_mul_pd(rsq00,rinv00);
 347             r00              = _mm256_andnot_pd(dummy_mask,r00);
 348
 349             /* Compute parameters for interactions between i and j atoms */
 350             qq00             = _mm256_mul_pd(iq0,jq0);
 351             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 352                                             vdwioffsetptr0+vdwjidx0B,
 353                                             vdwioffsetptr0+vdwjidx0C,
 354                                             vdwioffsetptr0+vdwjidx0D,
 355                                             &c6_00,&c12_00);
 356
 357             /* Calculate table index by multiplying r with table scale and truncate to integer */
 358             rt               = _mm256_mul_pd(r00,vftabscale);
 359             vfitab           = _mm256_cvttpd_epi32(rt);
 360             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 361             vfitab           = _mm_slli_epi32(vfitab,3);
 362
 363             /* REACTION-FIELD ELECTROSTATICS */
 364             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 365             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 366
 367             /* CUBIC SPLINE TABLE DISPERSION */
 368             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 369             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 370             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 371             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 372             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 373             Heps             = _mm256_mul_pd(vfeps,H);
 374             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 375             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 376             vvdw6            = _mm256_mul_pd(c6_00,VV);
 377             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 378             fvdw6            = _mm256_mul_pd(c6_00,FF);
 379
 380             /* CUBIC SPLINE TABLE REPULSION */
 381             vfitab           = _mm_add_epi32(vfitab,ifour);
 382             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 383             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 384             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 385             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 386             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 387             Heps             = _mm256_mul_pd(vfeps,H);
 388             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 389             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 390             vvdw12           = _mm256_mul_pd(c12_00,VV);
 391             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 392             fvdw12           = _mm256_mul_pd(c12_00,FF);
 393             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 394             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velec            = _mm256_andnot_pd(dummy_mask,velec);
 398             velecsum         = _mm256_add_pd(velecsum,velec);
 399             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 400             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 401
 402             fscal            = _mm256_add_pd(felec,fvdw);
 403
 404             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 405
 406             /* Calculate temporary vectorial force */
 407             tx               = _mm256_mul_pd(fscal,dx00);
 408             ty               = _mm256_mul_pd(fscal,dy00);
 409             tz               = _mm256_mul_pd(fscal,dz00);
 410
 411             /* Update vectorial force */
 412             fix0             = _mm256_add_pd(fix0,tx);
 413             fiy0             = _mm256_add_pd(fiy0,ty);
 414             fiz0             = _mm256_add_pd(fiz0,tz);
 415
 416             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 417             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 418             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 419             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 420             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 421
 422             /* Inner loop uses 68 flops */
 423         }
 424
 425         /* End of innermost loop */
 426
 427         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 428                                                  f+i_coord_offset,fshift+i_shift_offset);
 429
 430         ggid                        = gid[iidx];
 431         /* Update potential energies */
 432         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 433         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 434
 435         /* Increment number of inner iterations */
 436         inneriter                  += j_index_end - j_index_start;
 437
 438         /* Outer loop uses 9 flops */
 439     }
 440
 441     /* Increment number of outer iterations */
 442     outeriter        += nri;
 443
 444     /* Update outer/inner flops */
 445
 446     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*68);
 447 }
 448 /*
 449  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_256_double
 450  * Electrostatics interaction: ReactionField
 451  * VdW interaction:            CubicSplineTable
 452  * Geometry:                   Particle-Particle
 453  * Calculate force/pot:        Force
 454  */
 455 void
 456 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_256_double
 457                     (t_nblist                    * gmx_restrict       nlist,
 458                      rvec                        * gmx_restrict          xx,
 459                      rvec                        * gmx_restrict          ff,
 460                      t_forcerec                  * gmx_restrict          fr,
 461                      t_mdatoms                   * gmx_restrict     mdatoms,
 462                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 463                      t_nrnb                      * gmx_restrict        nrnb)
 464 {
 465     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 466      * just 0 for non-waters.
 467      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 468      * jnr indices corresponding to data put in the four positions in the SIMD register.
 469      */
 470     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 471     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 472     int              jnrA,jnrB,jnrC,jnrD;
 473     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 474     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 475     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 476     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 477     real             rcutoff_scalar;
 478     real             *shiftvec,*fshift,*x,*f;
 479     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 480     real             scratch[4*DIM];
 481     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 482     real *           vdwioffsetptr0;
 483     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 484     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 485     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 486     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 487     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 488     real             *charge;
 489     int              nvdwtype;
 490     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 491     int              *vdwtype;
 492     real             *vdwparam;
 493     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 494     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 495     __m128i          vfitab;
 496     __m128i          ifour       = _mm_set1_epi32(4);
 497     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 498     real             *vftab;
 499     __m256d          dummy_mask,cutoff_mask;
 500     __m128           tmpmask0,tmpmask1;
 501     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 502     __m256d          one     = _mm256_set1_pd(1.0);
 503     __m256d          two     = _mm256_set1_pd(2.0);
 504     x                = xx[0];
 505     f                = ff[0];
 506
 507     nri              = nlist->nri;
 508     iinr             = nlist->iinr;
 509     jindex           = nlist->jindex;
 510     jjnr             = nlist->jjnr;
 511     shiftidx         = nlist->shift;
 512     gid              = nlist->gid;
 513     shiftvec         = fr->shift_vec[0];
 514     fshift           = fr->fshift[0];
 515     facel            = _mm256_set1_pd(fr->epsfac);
 516     charge           = mdatoms->chargeA;
 517     krf              = _mm256_set1_pd(fr->ic->k_rf);
 518     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 519     crf              = _mm256_set1_pd(fr->ic->c_rf);
 520     nvdwtype         = fr->ntype;
 521     vdwparam         = fr->nbfp;
 522     vdwtype          = mdatoms->typeA;
 523
 524     vftab            = kernel_data->table_vdw->data;
 525     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 526
 527     /* Avoid stupid compiler warnings */
 528     jnrA = jnrB = jnrC = jnrD = 0;
 529     j_coord_offsetA = 0;
 530     j_coord_offsetB = 0;
 531     j_coord_offsetC = 0;
 532     j_coord_offsetD = 0;
 533
 534     outeriter        = 0;
 535     inneriter        = 0;
 536
 537     for(iidx=0;iidx<4*DIM;iidx++)
 538     {
 539         scratch[iidx] = 0.0;
 540     }
 541
 542     /* Start outer loop over neighborlists */
 543     for(iidx=0; iidx<nri; iidx++)
 544     {
 545         /* Load shift vector for this list */
 546         i_shift_offset   = DIM*shiftidx[iidx];
 547
 548         /* Load limits for loop over neighbors */
 549         j_index_start    = jindex[iidx];
 550         j_index_end      = jindex[iidx+1];
 551
 552         /* Get outer coordinate index */
 553         inr              = iinr[iidx];
 554         i_coord_offset   = DIM*inr;
 555
 556         /* Load i particle coords and add shift vector */
 557         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 558
 559         fix0             = _mm256_setzero_pd();
 560         fiy0             = _mm256_setzero_pd();
 561         fiz0             = _mm256_setzero_pd();
 562
 563         /* Load parameters for i particles */
 564         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 565         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 566
 567         /* Start inner kernel loop */
 568         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 569         {
 570
 571             /* Get j neighbor index, and coordinate index */
 572             jnrA             = jjnr[jidx];
 573             jnrB             = jjnr[jidx+1];
 574             jnrC             = jjnr[jidx+2];
 575             jnrD             = jjnr[jidx+3];
 576             j_coord_offsetA  = DIM*jnrA;
 577             j_coord_offsetB  = DIM*jnrB;
 578             j_coord_offsetC  = DIM*jnrC;
 579             j_coord_offsetD  = DIM*jnrD;
 580
 581             /* load j atom coordinates */
 582             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 583                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 584                                                  &jx0,&jy0,&jz0);
 585
 586             /* Calculate displacement vector */
 587             dx00             = _mm256_sub_pd(ix0,jx0);
 588             dy00             = _mm256_sub_pd(iy0,jy0);
 589             dz00             = _mm256_sub_pd(iz0,jz0);
 590
 591             /* Calculate squared distance and things based on it */
 592             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 593
 594             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 595
 596             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 597
 598             /* Load parameters for j particles */
 599             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 600                                                                  charge+jnrC+0,charge+jnrD+0);
 601             vdwjidx0A        = 2*vdwtype[jnrA+0];
 602             vdwjidx0B        = 2*vdwtype[jnrB+0];
 603             vdwjidx0C        = 2*vdwtype[jnrC+0];
 604             vdwjidx0D        = 2*vdwtype[jnrD+0];
 605
 606             /**************************
 607              * CALCULATE INTERACTIONS *
 608              **************************/
 609
 610             r00              = _mm256_mul_pd(rsq00,rinv00);
 611
 612             /* Compute parameters for interactions between i and j atoms */
 613             qq00             = _mm256_mul_pd(iq0,jq0);
 614             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 615                                             vdwioffsetptr0+vdwjidx0B,
 616                                             vdwioffsetptr0+vdwjidx0C,
 617                                             vdwioffsetptr0+vdwjidx0D,
 618                                             &c6_00,&c12_00);
 619
 620             /* Calculate table index by multiplying r with table scale and truncate to integer */
 621             rt               = _mm256_mul_pd(r00,vftabscale);
 622             vfitab           = _mm256_cvttpd_epi32(rt);
 623             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 624             vfitab           = _mm_slli_epi32(vfitab,3);
 625
 626             /* REACTION-FIELD ELECTROSTATICS */
 627             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 628
 629             /* CUBIC SPLINE TABLE DISPERSION */
 630             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 631             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 632             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 633             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 634             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 635             Heps             = _mm256_mul_pd(vfeps,H);
 636             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 637             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 638             fvdw6            = _mm256_mul_pd(c6_00,FF);
 639
 640             /* CUBIC SPLINE TABLE REPULSION */
 641             vfitab           = _mm_add_epi32(vfitab,ifour);
 642             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 643             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 644             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 645             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 646             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 647             Heps             = _mm256_mul_pd(vfeps,H);
 648             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 649             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 650             fvdw12           = _mm256_mul_pd(c12_00,FF);
 651             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 652
 653             fscal            = _mm256_add_pd(felec,fvdw);
 654
 655             /* Calculate temporary vectorial force */
 656             tx               = _mm256_mul_pd(fscal,dx00);
 657             ty               = _mm256_mul_pd(fscal,dy00);
 658             tz               = _mm256_mul_pd(fscal,dz00);
 659
 660             /* Update vectorial force */
 661             fix0             = _mm256_add_pd(fix0,tx);
 662             fiy0             = _mm256_add_pd(fiy0,ty);
 663             fiz0             = _mm256_add_pd(fiz0,tz);
 664
 665             fjptrA             = f+j_coord_offsetA;
 666             fjptrB             = f+j_coord_offsetB;
 667             fjptrC             = f+j_coord_offsetC;
 668             fjptrD             = f+j_coord_offsetD;
 669             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 670
 671             /* Inner loop uses 54 flops */
 672         }
 673
 674         if(jidx<j_index_end)
 675         {
 676
 677             /* Get j neighbor index, and coordinate index */
 678             jnrlistA         = jjnr[jidx];
 679             jnrlistB         = jjnr[jidx+1];
 680             jnrlistC         = jjnr[jidx+2];
 681             jnrlistD         = jjnr[jidx+3];
 682             /* Sign of each element will be negative for non-real atoms.
 683              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 684              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 685              */
 686             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 687
 688             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 689             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 690             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 691
 692             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 693             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 694             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 695             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 696             j_coord_offsetA  = DIM*jnrA;
 697             j_coord_offsetB  = DIM*jnrB;
 698             j_coord_offsetC  = DIM*jnrC;
 699             j_coord_offsetD  = DIM*jnrD;
 700
 701             /* load j atom coordinates */
 702             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 703                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 704                                                  &jx0,&jy0,&jz0);
 705
 706             /* Calculate displacement vector */
 707             dx00             = _mm256_sub_pd(ix0,jx0);
 708             dy00             = _mm256_sub_pd(iy0,jy0);
 709             dz00             = _mm256_sub_pd(iz0,jz0);
 710
 711             /* Calculate squared distance and things based on it */
 712             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 713
 714             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 715
 716             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 717
 718             /* Load parameters for j particles */
 719             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 720                                                                  charge+jnrC+0,charge+jnrD+0);
 721             vdwjidx0A        = 2*vdwtype[jnrA+0];
 722             vdwjidx0B        = 2*vdwtype[jnrB+0];
 723             vdwjidx0C        = 2*vdwtype[jnrC+0];
 724             vdwjidx0D        = 2*vdwtype[jnrD+0];
 725
 726             /**************************
 727              * CALCULATE INTERACTIONS *
 728              **************************/
 729
 730             r00              = _mm256_mul_pd(rsq00,rinv00);
 731             r00              = _mm256_andnot_pd(dummy_mask,r00);
 732
 733             /* Compute parameters for interactions between i and j atoms */
 734             qq00             = _mm256_mul_pd(iq0,jq0);
 735             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 736                                             vdwioffsetptr0+vdwjidx0B,
 737                                             vdwioffsetptr0+vdwjidx0C,
 738                                             vdwioffsetptr0+vdwjidx0D,
 739                                             &c6_00,&c12_00);
 740
 741             /* Calculate table index by multiplying r with table scale and truncate to integer */
 742             rt               = _mm256_mul_pd(r00,vftabscale);
 743             vfitab           = _mm256_cvttpd_epi32(rt);
 744             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 745             vfitab           = _mm_slli_epi32(vfitab,3);
 746
 747             /* REACTION-FIELD ELECTROSTATICS */
 748             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 749
 750             /* CUBIC SPLINE TABLE DISPERSION */
 751             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 752             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 753             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 754             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 755             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 756             Heps             = _mm256_mul_pd(vfeps,H);
 757             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 758             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 759             fvdw6            = _mm256_mul_pd(c6_00,FF);
 760
 761             /* CUBIC SPLINE TABLE REPULSION */
 762             vfitab           = _mm_add_epi32(vfitab,ifour);
 763             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 764             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 765             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 766             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 767             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 768             Heps             = _mm256_mul_pd(vfeps,H);
 769             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 770             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 771             fvdw12           = _mm256_mul_pd(c12_00,FF);
 772             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 773
 774             fscal            = _mm256_add_pd(felec,fvdw);
 775
 776             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 777
 778             /* Calculate temporary vectorial force */
 779             tx               = _mm256_mul_pd(fscal,dx00);
 780             ty               = _mm256_mul_pd(fscal,dy00);
 781             tz               = _mm256_mul_pd(fscal,dz00);
 782
 783             /* Update vectorial force */
 784             fix0             = _mm256_add_pd(fix0,tx);
 785             fiy0             = _mm256_add_pd(fiy0,ty);
 786             fiz0             = _mm256_add_pd(fiz0,tz);
 787
 788             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 789             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 790             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 791             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 792             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 793
 794             /* Inner loop uses 55 flops */
 795         }
 796
 797         /* End of innermost loop */
 798
 799         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 800                                                  f+i_coord_offset,fshift+i_shift_offset);
 801
 802         /* Increment number of inner iterations */
 803         inneriter                  += j_index_end - j_index_start;
 804
 805         /* Outer loop uses 7 flops */
 806     }
 807
 808     /* Increment number of outer iterations */
 809     outeriter        += nri;
 810
 811     /* Update outer/inner flops */
 812
 813     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 814 }