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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_256_single.h"
  50 #include "kernelutil_x86_avx_256_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_single
  54  * Electrostatics interaction: CubicSplineTable
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_single
  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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
  78     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  79     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  80     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  81     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  82     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  83     real             rcutoff_scalar;
  84     real             *shiftvec,*fshift,*x,*f;
  85     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  86     real             scratch[4*DIM];
  87     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  88     real *           vdwioffsetptr0;
  89     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  91     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     __m256i          vfitab;
  96     __m128i          vfitab_lo,vfitab_hi;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  99     real             *vftab;
 100     __m256           dummy_mask,cutoff_mask;
 101     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 102     __m256           one     = _mm256_set1_ps(1.0);
 103     __m256           two     = _mm256_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm256_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117
 118     vftab            = kernel_data->table_elec->data;
 119     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 120
 121     /* Avoid stupid compiler warnings */
 122     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 123     j_coord_offsetA = 0;
 124     j_coord_offsetB = 0;
 125     j_coord_offsetC = 0;
 126     j_coord_offsetD = 0;
 127     j_coord_offsetE = 0;
 128     j_coord_offsetF = 0;
 129     j_coord_offsetG = 0;
 130     j_coord_offsetH = 0;
 131
 132     outeriter        = 0;
 133     inneriter        = 0;
 134
 135     for(iidx=0;iidx<4*DIM;iidx++)
 136     {
 137         scratch[iidx] = 0.0;
 138     }
 139
 140     /* Start outer loop over neighborlists */
 141     for(iidx=0; iidx<nri; iidx++)
 142     {
 143         /* Load shift vector for this list */
 144         i_shift_offset   = DIM*shiftidx[iidx];
 145
 146         /* Load limits for loop over neighbors */
 147         j_index_start    = jindex[iidx];
 148         j_index_end      = jindex[iidx+1];
 149
 150         /* Get outer coordinate index */
 151         inr              = iinr[iidx];
 152         i_coord_offset   = DIM*inr;
 153
 154         /* Load i particle coords and add shift vector */
 155         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 156
 157         fix0             = _mm256_setzero_ps();
 158         fiy0             = _mm256_setzero_ps();
 159         fiz0             = _mm256_setzero_ps();
 160
 161         /* Load parameters for i particles */
 162         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm256_setzero_ps();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             jnrE             = jjnr[jidx+4];
 177             jnrF             = jjnr[jidx+5];
 178             jnrG             = jjnr[jidx+6];
 179             jnrH             = jjnr[jidx+7];
 180             j_coord_offsetA  = DIM*jnrA;
 181             j_coord_offsetB  = DIM*jnrB;
 182             j_coord_offsetC  = DIM*jnrC;
 183             j_coord_offsetD  = DIM*jnrD;
 184             j_coord_offsetE  = DIM*jnrE;
 185             j_coord_offsetF  = DIM*jnrF;
 186             j_coord_offsetG  = DIM*jnrG;
 187             j_coord_offsetH  = DIM*jnrH;
 188
 189             /* load j atom coordinates */
 190             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 192                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 193                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 194                                                  &jx0,&jy0,&jz0);
 195
 196             /* Calculate displacement vector */
 197             dx00             = _mm256_sub_ps(ix0,jx0);
 198             dy00             = _mm256_sub_ps(iy0,jy0);
 199             dz00             = _mm256_sub_ps(iz0,jz0);
 200
 201             /* Calculate squared distance and things based on it */
 202             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 203
 204             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 205
 206             /* Load parameters for j particles */
 207             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 208                                                                  charge+jnrC+0,charge+jnrD+0,
 209                                                                  charge+jnrE+0,charge+jnrF+0,
 210                                                                  charge+jnrG+0,charge+jnrH+0);
 211
 212             /**************************
 213              * CALCULATE INTERACTIONS *
 214              **************************/
 215
 216             r00              = _mm256_mul_ps(rsq00,rinv00);
 217
 218             /* Compute parameters for interactions between i and j atoms */
 219             qq00             = _mm256_mul_ps(iq0,jq0);
 220
 221             /* Calculate table index by multiplying r with table scale and truncate to integer */
 222             rt               = _mm256_mul_ps(r00,vftabscale);
 223             vfitab           = _mm256_cvttps_epi32(rt);
 224             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 225             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 226             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 227             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 228             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 229             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 230
 231             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 232             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 233                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 234             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 235                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 236             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 237                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 238             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 239                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 240             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 241             Heps             = _mm256_mul_ps(vfeps,H);
 242             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 243             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 244             velec            = _mm256_mul_ps(qq00,VV);
 245             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 246             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 247
 248             /* Update potential sum for this i atom from the interaction with this j atom. */
 249             velecsum         = _mm256_add_ps(velecsum,velec);
 250
 251             fscal            = felec;
 252
 253             /* Calculate temporary vectorial force */
 254             tx               = _mm256_mul_ps(fscal,dx00);
 255             ty               = _mm256_mul_ps(fscal,dy00);
 256             tz               = _mm256_mul_ps(fscal,dz00);
 257
 258             /* Update vectorial force */
 259             fix0             = _mm256_add_ps(fix0,tx);
 260             fiy0             = _mm256_add_ps(fiy0,ty);
 261             fiz0             = _mm256_add_ps(fiz0,tz);
 262
 263             fjptrA             = f+j_coord_offsetA;
 264             fjptrB             = f+j_coord_offsetB;
 265             fjptrC             = f+j_coord_offsetC;
 266             fjptrD             = f+j_coord_offsetD;
 267             fjptrE             = f+j_coord_offsetE;
 268             fjptrF             = f+j_coord_offsetF;
 269             fjptrG             = f+j_coord_offsetG;
 270             fjptrH             = f+j_coord_offsetH;
 271             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 272
 273             /* Inner loop uses 43 flops */
 274         }
 275
 276         if(jidx<j_index_end)
 277         {
 278
 279             /* Get j neighbor index, and coordinate index */
 280             jnrlistA         = jjnr[jidx];
 281             jnrlistB         = jjnr[jidx+1];
 282             jnrlistC         = jjnr[jidx+2];
 283             jnrlistD         = jjnr[jidx+3];
 284             jnrlistE         = jjnr[jidx+4];
 285             jnrlistF         = jjnr[jidx+5];
 286             jnrlistG         = jjnr[jidx+6];
 287             jnrlistH         = jjnr[jidx+7];
 288             /* Sign of each element will be negative for non-real atoms.
 289              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 290              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 291              */
 292             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 293                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 294
 295             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 296             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 297             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 298             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 299             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 300             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 301             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 302             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 303             j_coord_offsetA  = DIM*jnrA;
 304             j_coord_offsetB  = DIM*jnrB;
 305             j_coord_offsetC  = DIM*jnrC;
 306             j_coord_offsetD  = DIM*jnrD;
 307             j_coord_offsetE  = DIM*jnrE;
 308             j_coord_offsetF  = DIM*jnrF;
 309             j_coord_offsetG  = DIM*jnrG;
 310             j_coord_offsetH  = DIM*jnrH;
 311
 312             /* load j atom coordinates */
 313             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 314                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 315                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 316                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 317                                                  &jx0,&jy0,&jz0);
 318
 319             /* Calculate displacement vector */
 320             dx00             = _mm256_sub_ps(ix0,jx0);
 321             dy00             = _mm256_sub_ps(iy0,jy0);
 322             dz00             = _mm256_sub_ps(iz0,jz0);
 323
 324             /* Calculate squared distance and things based on it */
 325             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 326
 327             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 328
 329             /* Load parameters for j particles */
 330             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 331                                                                  charge+jnrC+0,charge+jnrD+0,
 332                                                                  charge+jnrE+0,charge+jnrF+0,
 333                                                                  charge+jnrG+0,charge+jnrH+0);
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             r00              = _mm256_mul_ps(rsq00,rinv00);
 340             r00              = _mm256_andnot_ps(dummy_mask,r00);
 341
 342             /* Compute parameters for interactions between i and j atoms */
 343             qq00             = _mm256_mul_ps(iq0,jq0);
 344
 345             /* Calculate table index by multiplying r with table scale and truncate to integer */
 346             rt               = _mm256_mul_ps(r00,vftabscale);
 347             vfitab           = _mm256_cvttps_epi32(rt);
 348             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 349             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 350             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 351             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 352             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 353             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 354
 355             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 356             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 357                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 358             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 359                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 360             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 361                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 362             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 363                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 364             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 365             Heps             = _mm256_mul_ps(vfeps,H);
 366             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 367             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 368             velec            = _mm256_mul_ps(qq00,VV);
 369             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 370             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velec            = _mm256_andnot_ps(dummy_mask,velec);
 374             velecsum         = _mm256_add_ps(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 379
 380             /* Calculate temporary vectorial force */
 381             tx               = _mm256_mul_ps(fscal,dx00);
 382             ty               = _mm256_mul_ps(fscal,dy00);
 383             tz               = _mm256_mul_ps(fscal,dz00);
 384
 385             /* Update vectorial force */
 386             fix0             = _mm256_add_ps(fix0,tx);
 387             fiy0             = _mm256_add_ps(fiy0,ty);
 388             fiz0             = _mm256_add_ps(fiz0,tz);
 389
 390             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 391             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 392             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 393             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 394             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 395             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 396             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 397             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 398             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 399
 400             /* Inner loop uses 44 flops */
 401         }
 402
 403         /* End of innermost loop */
 404
 405         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 406                                                  f+i_coord_offset,fshift+i_shift_offset);
 407
 408         ggid                        = gid[iidx];
 409         /* Update potential energies */
 410         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 411
 412         /* Increment number of inner iterations */
 413         inneriter                  += j_index_end - j_index_start;
 414
 415         /* Outer loop uses 8 flops */
 416     }
 417
 418     /* Increment number of outer iterations */
 419     outeriter        += nri;
 420
 421     /* Update outer/inner flops */
 422
 423     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*44);
 424 }
 425 /*
 426  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 427  * Electrostatics interaction: CubicSplineTable
 428  * VdW interaction:            None
 429  * Geometry:                   Particle-Particle
 430  * Calculate force/pot:        Force
 431  */
 432 void
 433 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 434                     (t_nblist                    * gmx_restrict       nlist,
 435                      rvec                        * gmx_restrict          xx,
 436                      rvec                        * gmx_restrict          ff,
 437                      t_forcerec                  * gmx_restrict          fr,
 438                      t_mdatoms                   * gmx_restrict     mdatoms,
 439                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 440                      t_nrnb                      * gmx_restrict        nrnb)
 441 {
 442     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 443      * just 0 for non-waters.
 444      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 445      * jnr indices corresponding to data put in the four positions in the SIMD register.
 446      */
 447     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 448     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 449     int              jnrA,jnrB,jnrC,jnrD;
 450     int              jnrE,jnrF,jnrG,jnrH;
 451     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 452     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 453     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 454     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 455     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 456     real             rcutoff_scalar;
 457     real             *shiftvec,*fshift,*x,*f;
 458     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 459     real             scratch[4*DIM];
 460     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 461     real *           vdwioffsetptr0;
 462     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 463     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 464     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 465     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 466     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 467     real             *charge;
 468     __m256i          vfitab;
 469     __m128i          vfitab_lo,vfitab_hi;
 470     __m128i          ifour       = _mm_set1_epi32(4);
 471     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 472     real             *vftab;
 473     __m256           dummy_mask,cutoff_mask;
 474     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 475     __m256           one     = _mm256_set1_ps(1.0);
 476     __m256           two     = _mm256_set1_ps(2.0);
 477     x                = xx[0];
 478     f                = ff[0];
 479
 480     nri              = nlist->nri;
 481     iinr             = nlist->iinr;
 482     jindex           = nlist->jindex;
 483     jjnr             = nlist->jjnr;
 484     shiftidx         = nlist->shift;
 485     gid              = nlist->gid;
 486     shiftvec         = fr->shift_vec[0];
 487     fshift           = fr->fshift[0];
 488     facel            = _mm256_set1_ps(fr->epsfac);
 489     charge           = mdatoms->chargeA;
 490
 491     vftab            = kernel_data->table_elec->data;
 492     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 493
 494     /* Avoid stupid compiler warnings */
 495     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 496     j_coord_offsetA = 0;
 497     j_coord_offsetB = 0;
 498     j_coord_offsetC = 0;
 499     j_coord_offsetD = 0;
 500     j_coord_offsetE = 0;
 501     j_coord_offsetF = 0;
 502     j_coord_offsetG = 0;
 503     j_coord_offsetH = 0;
 504
 505     outeriter        = 0;
 506     inneriter        = 0;
 507
 508     for(iidx=0;iidx<4*DIM;iidx++)
 509     {
 510         scratch[iidx] = 0.0;
 511     }
 512
 513     /* Start outer loop over neighborlists */
 514     for(iidx=0; iidx<nri; iidx++)
 515     {
 516         /* Load shift vector for this list */
 517         i_shift_offset   = DIM*shiftidx[iidx];
 518
 519         /* Load limits for loop over neighbors */
 520         j_index_start    = jindex[iidx];
 521         j_index_end      = jindex[iidx+1];
 522
 523         /* Get outer coordinate index */
 524         inr              = iinr[iidx];
 525         i_coord_offset   = DIM*inr;
 526
 527         /* Load i particle coords and add shift vector */
 528         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 529
 530         fix0             = _mm256_setzero_ps();
 531         fiy0             = _mm256_setzero_ps();
 532         fiz0             = _mm256_setzero_ps();
 533
 534         /* Load parameters for i particles */
 535         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 536
 537         /* Start inner kernel loop */
 538         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 539         {
 540
 541             /* Get j neighbor index, and coordinate index */
 542             jnrA             = jjnr[jidx];
 543             jnrB             = jjnr[jidx+1];
 544             jnrC             = jjnr[jidx+2];
 545             jnrD             = jjnr[jidx+3];
 546             jnrE             = jjnr[jidx+4];
 547             jnrF             = jjnr[jidx+5];
 548             jnrG             = jjnr[jidx+6];
 549             jnrH             = jjnr[jidx+7];
 550             j_coord_offsetA  = DIM*jnrA;
 551             j_coord_offsetB  = DIM*jnrB;
 552             j_coord_offsetC  = DIM*jnrC;
 553             j_coord_offsetD  = DIM*jnrD;
 554             j_coord_offsetE  = DIM*jnrE;
 555             j_coord_offsetF  = DIM*jnrF;
 556             j_coord_offsetG  = DIM*jnrG;
 557             j_coord_offsetH  = DIM*jnrH;
 558
 559             /* load j atom coordinates */
 560             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 561                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 562                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 563                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 564                                                  &jx0,&jy0,&jz0);
 565
 566             /* Calculate displacement vector */
 567             dx00             = _mm256_sub_ps(ix0,jx0);
 568             dy00             = _mm256_sub_ps(iy0,jy0);
 569             dz00             = _mm256_sub_ps(iz0,jz0);
 570
 571             /* Calculate squared distance and things based on it */
 572             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 573
 574             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 575
 576             /* Load parameters for j particles */
 577             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 578                                                                  charge+jnrC+0,charge+jnrD+0,
 579                                                                  charge+jnrE+0,charge+jnrF+0,
 580                                                                  charge+jnrG+0,charge+jnrH+0);
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             r00              = _mm256_mul_ps(rsq00,rinv00);
 587
 588             /* Compute parameters for interactions between i and j atoms */
 589             qq00             = _mm256_mul_ps(iq0,jq0);
 590
 591             /* Calculate table index by multiplying r with table scale and truncate to integer */
 592             rt               = _mm256_mul_ps(r00,vftabscale);
 593             vfitab           = _mm256_cvttps_epi32(rt);
 594             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 595             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 596             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 597             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 598             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 599             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 600
 601             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 602             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 603                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 604             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 605                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 606             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 607                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 608             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 609                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 610             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 611             Heps             = _mm256_mul_ps(vfeps,H);
 612             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 613             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 614             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 615
 616             fscal            = felec;
 617
 618             /* Calculate temporary vectorial force */
 619             tx               = _mm256_mul_ps(fscal,dx00);
 620             ty               = _mm256_mul_ps(fscal,dy00);
 621             tz               = _mm256_mul_ps(fscal,dz00);
 622
 623             /* Update vectorial force */
 624             fix0             = _mm256_add_ps(fix0,tx);
 625             fiy0             = _mm256_add_ps(fiy0,ty);
 626             fiz0             = _mm256_add_ps(fiz0,tz);
 627
 628             fjptrA             = f+j_coord_offsetA;
 629             fjptrB             = f+j_coord_offsetB;
 630             fjptrC             = f+j_coord_offsetC;
 631             fjptrD             = f+j_coord_offsetD;
 632             fjptrE             = f+j_coord_offsetE;
 633             fjptrF             = f+j_coord_offsetF;
 634             fjptrG             = f+j_coord_offsetG;
 635             fjptrH             = f+j_coord_offsetH;
 636             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 637
 638             /* Inner loop uses 39 flops */
 639         }
 640
 641         if(jidx<j_index_end)
 642         {
 643
 644             /* Get j neighbor index, and coordinate index */
 645             jnrlistA         = jjnr[jidx];
 646             jnrlistB         = jjnr[jidx+1];
 647             jnrlistC         = jjnr[jidx+2];
 648             jnrlistD         = jjnr[jidx+3];
 649             jnrlistE         = jjnr[jidx+4];
 650             jnrlistF         = jjnr[jidx+5];
 651             jnrlistG         = jjnr[jidx+6];
 652             jnrlistH         = jjnr[jidx+7];
 653             /* Sign of each element will be negative for non-real atoms.
 654              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 655              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 656              */
 657             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 658                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 659
 660             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 661             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 662             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 663             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 664             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 665             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 666             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 667             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 668             j_coord_offsetA  = DIM*jnrA;
 669             j_coord_offsetB  = DIM*jnrB;
 670             j_coord_offsetC  = DIM*jnrC;
 671             j_coord_offsetD  = DIM*jnrD;
 672             j_coord_offsetE  = DIM*jnrE;
 673             j_coord_offsetF  = DIM*jnrF;
 674             j_coord_offsetG  = DIM*jnrG;
 675             j_coord_offsetH  = DIM*jnrH;
 676
 677             /* load j atom coordinates */
 678             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 679                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 680                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 681                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 682                                                  &jx0,&jy0,&jz0);
 683
 684             /* Calculate displacement vector */
 685             dx00             = _mm256_sub_ps(ix0,jx0);
 686             dy00             = _mm256_sub_ps(iy0,jy0);
 687             dz00             = _mm256_sub_ps(iz0,jz0);
 688
 689             /* Calculate squared distance and things based on it */
 690             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 691
 692             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 693
 694             /* Load parameters for j particles */
 695             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 696                                                                  charge+jnrC+0,charge+jnrD+0,
 697                                                                  charge+jnrE+0,charge+jnrF+0,
 698                                                                  charge+jnrG+0,charge+jnrH+0);
 699
 700             /**************************
 701              * CALCULATE INTERACTIONS *
 702              **************************/
 703
 704             r00              = _mm256_mul_ps(rsq00,rinv00);
 705             r00              = _mm256_andnot_ps(dummy_mask,r00);
 706
 707             /* Compute parameters for interactions between i and j atoms */
 708             qq00             = _mm256_mul_ps(iq0,jq0);
 709
 710             /* Calculate table index by multiplying r with table scale and truncate to integer */
 711             rt               = _mm256_mul_ps(r00,vftabscale);
 712             vfitab           = _mm256_cvttps_epi32(rt);
 713             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 714             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 715             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 716             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 717             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 718             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 719
 720             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 721             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 722                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 723             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 724                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 725             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 726                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 727             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 728                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 729             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 730             Heps             = _mm256_mul_ps(vfeps,H);
 731             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 732             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 733             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 734
 735             fscal            = felec;
 736
 737             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 738
 739             /* Calculate temporary vectorial force */
 740             tx               = _mm256_mul_ps(fscal,dx00);
 741             ty               = _mm256_mul_ps(fscal,dy00);
 742             tz               = _mm256_mul_ps(fscal,dz00);
 743
 744             /* Update vectorial force */
 745             fix0             = _mm256_add_ps(fix0,tx);
 746             fiy0             = _mm256_add_ps(fiy0,ty);
 747             fiz0             = _mm256_add_ps(fiz0,tz);
 748
 749             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 750             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 751             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 752             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 753             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 754             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 755             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 756             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 757             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 758
 759             /* Inner loop uses 40 flops */
 760         }
 761
 762         /* End of innermost loop */
 763
 764         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 765                                                  f+i_coord_offset,fshift+i_shift_offset);
 766
 767         /* Increment number of inner iterations */
 768         inneriter                  += j_index_end - j_index_start;
 769
 770         /* Outer loop uses 7 flops */
 771     }
 772
 773     /* Increment number of outer iterations */
 774     outeriter        += nri;
 775
 776     /* Update outer/inner flops */
 777
 778     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 779 }