src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecCSTab_VdwNone_GeomP1P1_avx_256_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_single 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/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_single
  51  * Electrostatics interaction: CubicSplineTable
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrE,jnrF,jnrG,jnrH;
  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              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  83     real             scratch[4*DIM];
  84     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     real *           vdwioffsetptr0;
  86     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  88     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m256i          vfitab;
  93     __m128i          vfitab_lo,vfitab_hi;
  94     __m128i          ifour       = _mm_set1_epi32(4);
  95     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  96     real             *vftab;
  97     __m256           dummy_mask,cutoff_mask;
  98     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  99     __m256           one     = _mm256_set1_ps(1.0);
 100     __m256           two     = _mm256_set1_ps(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = _mm256_set1_ps(fr->ic->epsfac);
 113     charge           = mdatoms->chargeA;
 114
 115     vftab            = kernel_data->table_elec->data;
 116     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 117
 118     /* Avoid stupid compiler warnings */
 119     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 120     j_coord_offsetA = 0;
 121     j_coord_offsetB = 0;
 122     j_coord_offsetC = 0;
 123     j_coord_offsetD = 0;
 124     j_coord_offsetE = 0;
 125     j_coord_offsetF = 0;
 126     j_coord_offsetG = 0;
 127     j_coord_offsetH = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm256_setzero_ps();
 155         fiy0             = _mm256_setzero_ps();
 156         fiz0             = _mm256_setzero_ps();
 157
 158         /* Load parameters for i particles */
 159         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 160
 161         /* Reset potential sums */
 162         velecsum         = _mm256_setzero_ps();
 163
 164         /* Start inner kernel loop */
 165         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 166         {
 167
 168             /* Get j neighbor index, and coordinate index */
 169             jnrA             = jjnr[jidx];
 170             jnrB             = jjnr[jidx+1];
 171             jnrC             = jjnr[jidx+2];
 172             jnrD             = jjnr[jidx+3];
 173             jnrE             = jjnr[jidx+4];
 174             jnrF             = jjnr[jidx+5];
 175             jnrG             = jjnr[jidx+6];
 176             jnrH             = jjnr[jidx+7];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179             j_coord_offsetC  = DIM*jnrC;
 180             j_coord_offsetD  = DIM*jnrD;
 181             j_coord_offsetE  = DIM*jnrE;
 182             j_coord_offsetF  = DIM*jnrF;
 183             j_coord_offsetG  = DIM*jnrG;
 184             j_coord_offsetH  = DIM*jnrH;
 185
 186             /* load j atom coordinates */
 187             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 188                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 189                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 190                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 191                                                  &jx0,&jy0,&jz0);
 192
 193             /* Calculate displacement vector */
 194             dx00             = _mm256_sub_ps(ix0,jx0);
 195             dy00             = _mm256_sub_ps(iy0,jy0);
 196             dz00             = _mm256_sub_ps(iz0,jz0);
 197
 198             /* Calculate squared distance and things based on it */
 199             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 200
 201             rinv00           = avx256_invsqrt_f(rsq00);
 202
 203             /* Load parameters for j particles */
 204             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 205                                                                  charge+jnrC+0,charge+jnrD+0,
 206                                                                  charge+jnrE+0,charge+jnrF+0,
 207                                                                  charge+jnrG+0,charge+jnrH+0);
 208
 209             /**************************
 210              * CALCULATE INTERACTIONS *
 211              **************************/
 212
 213             r00              = _mm256_mul_ps(rsq00,rinv00);
 214
 215             /* Compute parameters for interactions between i and j atoms */
 216             qq00             = _mm256_mul_ps(iq0,jq0);
 217
 218             /* Calculate table index by multiplying r with table scale and truncate to integer */
 219             rt               = _mm256_mul_ps(r00,vftabscale);
 220             vfitab           = _mm256_cvttps_epi32(rt);
 221             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 222             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 223             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 224             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 225             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 226             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 227
 228             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 229             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 230                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 231             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 232                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 233             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 234                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 235             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 236                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 237             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 238             Heps             = _mm256_mul_ps(vfeps,H);
 239             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 240             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 241             velec            = _mm256_mul_ps(qq00,VV);
 242             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 243             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 244
 245             /* Update potential sum for this i atom from the interaction with this j atom. */
 246             velecsum         = _mm256_add_ps(velecsum,velec);
 247
 248             fscal            = felec;
 249
 250             /* Calculate temporary vectorial force */
 251             tx               = _mm256_mul_ps(fscal,dx00);
 252             ty               = _mm256_mul_ps(fscal,dy00);
 253             tz               = _mm256_mul_ps(fscal,dz00);
 254
 255             /* Update vectorial force */
 256             fix0             = _mm256_add_ps(fix0,tx);
 257             fiy0             = _mm256_add_ps(fiy0,ty);
 258             fiz0             = _mm256_add_ps(fiz0,tz);
 259
 260             fjptrA             = f+j_coord_offsetA;
 261             fjptrB             = f+j_coord_offsetB;
 262             fjptrC             = f+j_coord_offsetC;
 263             fjptrD             = f+j_coord_offsetD;
 264             fjptrE             = f+j_coord_offsetE;
 265             fjptrF             = f+j_coord_offsetF;
 266             fjptrG             = f+j_coord_offsetG;
 267             fjptrH             = f+j_coord_offsetH;
 268             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 269
 270             /* Inner loop uses 43 flops */
 271         }
 272
 273         if(jidx<j_index_end)
 274         {
 275
 276             /* Get j neighbor index, and coordinate index */
 277             jnrlistA         = jjnr[jidx];
 278             jnrlistB         = jjnr[jidx+1];
 279             jnrlistC         = jjnr[jidx+2];
 280             jnrlistD         = jjnr[jidx+3];
 281             jnrlistE         = jjnr[jidx+4];
 282             jnrlistF         = jjnr[jidx+5];
 283             jnrlistG         = jjnr[jidx+6];
 284             jnrlistH         = jjnr[jidx+7];
 285             /* Sign of each element will be negative for non-real atoms.
 286              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 287              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 288              */
 289             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 290                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 291
 292             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 293             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 294             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 295             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 296             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 297             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 298             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 299             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 300             j_coord_offsetA  = DIM*jnrA;
 301             j_coord_offsetB  = DIM*jnrB;
 302             j_coord_offsetC  = DIM*jnrC;
 303             j_coord_offsetD  = DIM*jnrD;
 304             j_coord_offsetE  = DIM*jnrE;
 305             j_coord_offsetF  = DIM*jnrF;
 306             j_coord_offsetG  = DIM*jnrG;
 307             j_coord_offsetH  = DIM*jnrH;
 308
 309             /* load j atom coordinates */
 310             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 311                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 312                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 313                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 314                                                  &jx0,&jy0,&jz0);
 315
 316             /* Calculate displacement vector */
 317             dx00             = _mm256_sub_ps(ix0,jx0);
 318             dy00             = _mm256_sub_ps(iy0,jy0);
 319             dz00             = _mm256_sub_ps(iz0,jz0);
 320
 321             /* Calculate squared distance and things based on it */
 322             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 323
 324             rinv00           = avx256_invsqrt_f(rsq00);
 325
 326             /* Load parameters for j particles */
 327             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 328                                                                  charge+jnrC+0,charge+jnrD+0,
 329                                                                  charge+jnrE+0,charge+jnrF+0,
 330                                                                  charge+jnrG+0,charge+jnrH+0);
 331
 332             /**************************
 333              * CALCULATE INTERACTIONS *
 334              **************************/
 335
 336             r00              = _mm256_mul_ps(rsq00,rinv00);
 337             r00              = _mm256_andnot_ps(dummy_mask,r00);
 338
 339             /* Compute parameters for interactions between i and j atoms */
 340             qq00             = _mm256_mul_ps(iq0,jq0);
 341
 342             /* Calculate table index by multiplying r with table scale and truncate to integer */
 343             rt               = _mm256_mul_ps(r00,vftabscale);
 344             vfitab           = _mm256_cvttps_epi32(rt);
 345             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 346             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 347             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 348             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 349             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 350             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 351
 352             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 353             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 354                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 355             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 356                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 357             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 358                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 359             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 360                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 361             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 362             Heps             = _mm256_mul_ps(vfeps,H);
 363             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 364             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 365             velec            = _mm256_mul_ps(qq00,VV);
 366             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 367             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm256_andnot_ps(dummy_mask,velec);
 371             velecsum         = _mm256_add_ps(velecsum,velec);
 372
 373             fscal            = felec;
 374
 375             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 376
 377             /* Calculate temporary vectorial force */
 378             tx               = _mm256_mul_ps(fscal,dx00);
 379             ty               = _mm256_mul_ps(fscal,dy00);
 380             tz               = _mm256_mul_ps(fscal,dz00);
 381
 382             /* Update vectorial force */
 383             fix0             = _mm256_add_ps(fix0,tx);
 384             fiy0             = _mm256_add_ps(fiy0,ty);
 385             fiz0             = _mm256_add_ps(fiz0,tz);
 386
 387             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 388             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 389             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 390             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 391             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 392             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 393             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 394             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 395             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 396
 397             /* Inner loop uses 44 flops */
 398         }
 399
 400         /* End of innermost loop */
 401
 402         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 403                                                  f+i_coord_offset,fshift+i_shift_offset);
 404
 405         ggid                        = gid[iidx];
 406         /* Update potential energies */
 407         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 408
 409         /* Increment number of inner iterations */
 410         inneriter                  += j_index_end - j_index_start;
 411
 412         /* Outer loop uses 8 flops */
 413     }
 414
 415     /* Increment number of outer iterations */
 416     outeriter        += nri;
 417
 418     /* Update outer/inner flops */
 419
 420     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*44);
 421 }
 422 /*
 423  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 424  * Electrostatics interaction: CubicSplineTable
 425  * VdW interaction:            None
 426  * Geometry:                   Particle-Particle
 427  * Calculate force/pot:        Force
 428  */
 429 void
 430 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 431                     (t_nblist                    * gmx_restrict       nlist,
 432                      rvec                        * gmx_restrict          xx,
 433                      rvec                        * gmx_restrict          ff,
 434                      struct t_forcerec           * gmx_restrict          fr,
 435                      t_mdatoms                   * gmx_restrict     mdatoms,
 436                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 437                      t_nrnb                      * gmx_restrict        nrnb)
 438 {
 439     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 440      * just 0 for non-waters.
 441      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 442      * jnr indices corresponding to data put in the four positions in the SIMD register.
 443      */
 444     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 445     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 446     int              jnrA,jnrB,jnrC,jnrD;
 447     int              jnrE,jnrF,jnrG,jnrH;
 448     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 449     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 450     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 451     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 452     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 453     real             rcutoff_scalar;
 454     real             *shiftvec,*fshift,*x,*f;
 455     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 456     real             scratch[4*DIM];
 457     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 458     real *           vdwioffsetptr0;
 459     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 460     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 461     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 462     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 463     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 464     real             *charge;
 465     __m256i          vfitab;
 466     __m128i          vfitab_lo,vfitab_hi;
 467     __m128i          ifour       = _mm_set1_epi32(4);
 468     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 469     real             *vftab;
 470     __m256           dummy_mask,cutoff_mask;
 471     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 472     __m256           one     = _mm256_set1_ps(1.0);
 473     __m256           two     = _mm256_set1_ps(2.0);
 474     x                = xx[0];
 475     f                = ff[0];
 476
 477     nri              = nlist->nri;
 478     iinr             = nlist->iinr;
 479     jindex           = nlist->jindex;
 480     jjnr             = nlist->jjnr;
 481     shiftidx         = nlist->shift;
 482     gid              = nlist->gid;
 483     shiftvec         = fr->shift_vec[0];
 484     fshift           = fr->fshift[0];
 485     facel            = _mm256_set1_ps(fr->ic->epsfac);
 486     charge           = mdatoms->chargeA;
 487
 488     vftab            = kernel_data->table_elec->data;
 489     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 490
 491     /* Avoid stupid compiler warnings */
 492     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 493     j_coord_offsetA = 0;
 494     j_coord_offsetB = 0;
 495     j_coord_offsetC = 0;
 496     j_coord_offsetD = 0;
 497     j_coord_offsetE = 0;
 498     j_coord_offsetF = 0;
 499     j_coord_offsetG = 0;
 500     j_coord_offsetH = 0;
 501
 502     outeriter        = 0;
 503     inneriter        = 0;
 504
 505     for(iidx=0;iidx<4*DIM;iidx++)
 506     {
 507         scratch[iidx] = 0.0;
 508     }
 509
 510     /* Start outer loop over neighborlists */
 511     for(iidx=0; iidx<nri; iidx++)
 512     {
 513         /* Load shift vector for this list */
 514         i_shift_offset   = DIM*shiftidx[iidx];
 515
 516         /* Load limits for loop over neighbors */
 517         j_index_start    = jindex[iidx];
 518         j_index_end      = jindex[iidx+1];
 519
 520         /* Get outer coordinate index */
 521         inr              = iinr[iidx];
 522         i_coord_offset   = DIM*inr;
 523
 524         /* Load i particle coords and add shift vector */
 525         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 526
 527         fix0             = _mm256_setzero_ps();
 528         fiy0             = _mm256_setzero_ps();
 529         fiz0             = _mm256_setzero_ps();
 530
 531         /* Load parameters for i particles */
 532         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 533
 534         /* Start inner kernel loop */
 535         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 536         {
 537
 538             /* Get j neighbor index, and coordinate index */
 539             jnrA             = jjnr[jidx];
 540             jnrB             = jjnr[jidx+1];
 541             jnrC             = jjnr[jidx+2];
 542             jnrD             = jjnr[jidx+3];
 543             jnrE             = jjnr[jidx+4];
 544             jnrF             = jjnr[jidx+5];
 545             jnrG             = jjnr[jidx+6];
 546             jnrH             = jjnr[jidx+7];
 547             j_coord_offsetA  = DIM*jnrA;
 548             j_coord_offsetB  = DIM*jnrB;
 549             j_coord_offsetC  = DIM*jnrC;
 550             j_coord_offsetD  = DIM*jnrD;
 551             j_coord_offsetE  = DIM*jnrE;
 552             j_coord_offsetF  = DIM*jnrF;
 553             j_coord_offsetG  = DIM*jnrG;
 554             j_coord_offsetH  = DIM*jnrH;
 555
 556             /* load j atom coordinates */
 557             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 558                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 559                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 560                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 561                                                  &jx0,&jy0,&jz0);
 562
 563             /* Calculate displacement vector */
 564             dx00             = _mm256_sub_ps(ix0,jx0);
 565             dy00             = _mm256_sub_ps(iy0,jy0);
 566             dz00             = _mm256_sub_ps(iz0,jz0);
 567
 568             /* Calculate squared distance and things based on it */
 569             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 570
 571             rinv00           = avx256_invsqrt_f(rsq00);
 572
 573             /* Load parameters for j particles */
 574             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 575                                                                  charge+jnrC+0,charge+jnrD+0,
 576                                                                  charge+jnrE+0,charge+jnrF+0,
 577                                                                  charge+jnrG+0,charge+jnrH+0);
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             r00              = _mm256_mul_ps(rsq00,rinv00);
 584
 585             /* Compute parameters for interactions between i and j atoms */
 586             qq00             = _mm256_mul_ps(iq0,jq0);
 587
 588             /* Calculate table index by multiplying r with table scale and truncate to integer */
 589             rt               = _mm256_mul_ps(r00,vftabscale);
 590             vfitab           = _mm256_cvttps_epi32(rt);
 591             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 592             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 593             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 594             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 595             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 596             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 597
 598             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 599             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 600                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 601             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 602                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 603             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 604                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 605             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 606                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 607             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 608             Heps             = _mm256_mul_ps(vfeps,H);
 609             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 610             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 611             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 612
 613             fscal            = felec;
 614
 615             /* Calculate temporary vectorial force */
 616             tx               = _mm256_mul_ps(fscal,dx00);
 617             ty               = _mm256_mul_ps(fscal,dy00);
 618             tz               = _mm256_mul_ps(fscal,dz00);
 619
 620             /* Update vectorial force */
 621             fix0             = _mm256_add_ps(fix0,tx);
 622             fiy0             = _mm256_add_ps(fiy0,ty);
 623             fiz0             = _mm256_add_ps(fiz0,tz);
 624
 625             fjptrA             = f+j_coord_offsetA;
 626             fjptrB             = f+j_coord_offsetB;
 627             fjptrC             = f+j_coord_offsetC;
 628             fjptrD             = f+j_coord_offsetD;
 629             fjptrE             = f+j_coord_offsetE;
 630             fjptrF             = f+j_coord_offsetF;
 631             fjptrG             = f+j_coord_offsetG;
 632             fjptrH             = f+j_coord_offsetH;
 633             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 634
 635             /* Inner loop uses 39 flops */
 636         }
 637
 638         if(jidx<j_index_end)
 639         {
 640
 641             /* Get j neighbor index, and coordinate index */
 642             jnrlistA         = jjnr[jidx];
 643             jnrlistB         = jjnr[jidx+1];
 644             jnrlistC         = jjnr[jidx+2];
 645             jnrlistD         = jjnr[jidx+3];
 646             jnrlistE         = jjnr[jidx+4];
 647             jnrlistF         = jjnr[jidx+5];
 648             jnrlistG         = jjnr[jidx+6];
 649             jnrlistH         = jjnr[jidx+7];
 650             /* Sign of each element will be negative for non-real atoms.
 651              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 652              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 653              */
 654             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 655                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 656
 657             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 658             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 659             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 660             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 661             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 662             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 663             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 664             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 665             j_coord_offsetA  = DIM*jnrA;
 666             j_coord_offsetB  = DIM*jnrB;
 667             j_coord_offsetC  = DIM*jnrC;
 668             j_coord_offsetD  = DIM*jnrD;
 669             j_coord_offsetE  = DIM*jnrE;
 670             j_coord_offsetF  = DIM*jnrF;
 671             j_coord_offsetG  = DIM*jnrG;
 672             j_coord_offsetH  = DIM*jnrH;
 673
 674             /* load j atom coordinates */
 675             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 676                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 677                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 678                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 679                                                  &jx0,&jy0,&jz0);
 680
 681             /* Calculate displacement vector */
 682             dx00             = _mm256_sub_ps(ix0,jx0);
 683             dy00             = _mm256_sub_ps(iy0,jy0);
 684             dz00             = _mm256_sub_ps(iz0,jz0);
 685
 686             /* Calculate squared distance and things based on it */
 687             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 688
 689             rinv00           = avx256_invsqrt_f(rsq00);
 690
 691             /* Load parameters for j particles */
 692             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 693                                                                  charge+jnrC+0,charge+jnrD+0,
 694                                                                  charge+jnrE+0,charge+jnrF+0,
 695                                                                  charge+jnrG+0,charge+jnrH+0);
 696
 697             /**************************
 698              * CALCULATE INTERACTIONS *
 699              **************************/
 700
 701             r00              = _mm256_mul_ps(rsq00,rinv00);
 702             r00              = _mm256_andnot_ps(dummy_mask,r00);
 703
 704             /* Compute parameters for interactions between i and j atoms */
 705             qq00             = _mm256_mul_ps(iq0,jq0);
 706
 707             /* Calculate table index by multiplying r with table scale and truncate to integer */
 708             rt               = _mm256_mul_ps(r00,vftabscale);
 709             vfitab           = _mm256_cvttps_epi32(rt);
 710             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 711             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 712             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 713             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 714             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 715             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 716
 717             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 718             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 719                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 720             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 721                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 722             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 723                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 724             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 725                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 726             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 727             Heps             = _mm256_mul_ps(vfeps,H);
 728             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 729             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 730             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 731
 732             fscal            = felec;
 733
 734             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 735
 736             /* Calculate temporary vectorial force */
 737             tx               = _mm256_mul_ps(fscal,dx00);
 738             ty               = _mm256_mul_ps(fscal,dy00);
 739             tz               = _mm256_mul_ps(fscal,dz00);
 740
 741             /* Update vectorial force */
 742             fix0             = _mm256_add_ps(fix0,tx);
 743             fiy0             = _mm256_add_ps(fiy0,ty);
 744             fiz0             = _mm256_add_ps(fiz0,tz);
 745
 746             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 747             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 748             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 749             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 750             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 751             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 752             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 753             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 754             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 755
 756             /* Inner loop uses 40 flops */
 757         }
 758
 759         /* End of innermost loop */
 760
 761         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 762                                                  f+i_coord_offset,fshift+i_shift_offset);
 763
 764         /* Increment number of inner iterations */
 765         inneriter                  += j_index_end - j_index_start;
 766
 767         /* Outer loop uses 7 flops */
 768     }
 769
 770     /* Increment number of outer iterations */
 771     outeriter        += nri;
 772
 773     /* Update outer/inner flops */
 774
 775     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 776 }