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) 2012,2013,2014, 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_single 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_single.h"
  48 #include "kernelutil_x86_avx_256_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_single
  52  * Electrostatics interaction: CubicSplineTable
  53  * VdW interaction:            None
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_single
  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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
  76     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  77     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  84     real             scratch[4*DIM];
  85     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  89     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     __m256i          vfitab;
  94     __m128i          vfitab_lo,vfitab_hi;
  95     __m128i          ifour       = _mm_set1_epi32(4);
  96     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  97     real             *vftab;
  98     __m256           dummy_mask,cutoff_mask;
  99     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 100     __m256           one     = _mm256_set1_ps(1.0);
 101     __m256           two     = _mm256_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm256_set1_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115
 116     vftab            = kernel_data->table_elec->data;
 117     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 118
 119     /* Avoid stupid compiler warnings */
 120     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 121     j_coord_offsetA = 0;
 122     j_coord_offsetB = 0;
 123     j_coord_offsetC = 0;
 124     j_coord_offsetD = 0;
 125     j_coord_offsetE = 0;
 126     j_coord_offsetF = 0;
 127     j_coord_offsetG = 0;
 128     j_coord_offsetH = 0;
 129
 130     outeriter        = 0;
 131     inneriter        = 0;
 132
 133     for(iidx=0;iidx<4*DIM;iidx++)
 134     {
 135         scratch[iidx] = 0.0;
 136     }
 137
 138     /* Start outer loop over neighborlists */
 139     for(iidx=0; iidx<nri; iidx++)
 140     {
 141         /* Load shift vector for this list */
 142         i_shift_offset   = DIM*shiftidx[iidx];
 143
 144         /* Load limits for loop over neighbors */
 145         j_index_start    = jindex[iidx];
 146         j_index_end      = jindex[iidx+1];
 147
 148         /* Get outer coordinate index */
 149         inr              = iinr[iidx];
 150         i_coord_offset   = DIM*inr;
 151
 152         /* Load i particle coords and add shift vector */
 153         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 154
 155         fix0             = _mm256_setzero_ps();
 156         fiy0             = _mm256_setzero_ps();
 157         fiz0             = _mm256_setzero_ps();
 158
 159         /* Load parameters for i particles */
 160         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm256_setzero_ps();
 164
 165         /* Start inner kernel loop */
 166         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 167         {
 168
 169             /* Get j neighbor index, and coordinate index */
 170             jnrA             = jjnr[jidx];
 171             jnrB             = jjnr[jidx+1];
 172             jnrC             = jjnr[jidx+2];
 173             jnrD             = jjnr[jidx+3];
 174             jnrE             = jjnr[jidx+4];
 175             jnrF             = jjnr[jidx+5];
 176             jnrG             = jjnr[jidx+6];
 177             jnrH             = jjnr[jidx+7];
 178             j_coord_offsetA  = DIM*jnrA;
 179             j_coord_offsetB  = DIM*jnrB;
 180             j_coord_offsetC  = DIM*jnrC;
 181             j_coord_offsetD  = DIM*jnrD;
 182             j_coord_offsetE  = DIM*jnrE;
 183             j_coord_offsetF  = DIM*jnrF;
 184             j_coord_offsetG  = DIM*jnrG;
 185             j_coord_offsetH  = DIM*jnrH;
 186
 187             /* load j atom coordinates */
 188             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 190                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 191                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 192                                                  &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm256_sub_ps(ix0,jx0);
 196             dy00             = _mm256_sub_ps(iy0,jy0);
 197             dz00             = _mm256_sub_ps(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 201
 202             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 203
 204             /* Load parameters for j particles */
 205             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 206                                                                  charge+jnrC+0,charge+jnrD+0,
 207                                                                  charge+jnrE+0,charge+jnrF+0,
 208                                                                  charge+jnrG+0,charge+jnrH+0);
 209
 210             /**************************
 211              * CALCULATE INTERACTIONS *
 212              **************************/
 213
 214             r00              = _mm256_mul_ps(rsq00,rinv00);
 215
 216             /* Compute parameters for interactions between i and j atoms */
 217             qq00             = _mm256_mul_ps(iq0,jq0);
 218
 219             /* Calculate table index by multiplying r with table scale and truncate to integer */
 220             rt               = _mm256_mul_ps(r00,vftabscale);
 221             vfitab           = _mm256_cvttps_epi32(rt);
 222             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 223             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 224             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 225             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 226             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 227             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 228
 229             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 230             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 231                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 232             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 233                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 234             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 235                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 236             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 237                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 238             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 239             Heps             = _mm256_mul_ps(vfeps,H);
 240             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 241             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 242             velec            = _mm256_mul_ps(qq00,VV);
 243             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 244             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 245
 246             /* Update potential sum for this i atom from the interaction with this j atom. */
 247             velecsum         = _mm256_add_ps(velecsum,velec);
 248
 249             fscal            = felec;
 250
 251             /* Calculate temporary vectorial force */
 252             tx               = _mm256_mul_ps(fscal,dx00);
 253             ty               = _mm256_mul_ps(fscal,dy00);
 254             tz               = _mm256_mul_ps(fscal,dz00);
 255
 256             /* Update vectorial force */
 257             fix0             = _mm256_add_ps(fix0,tx);
 258             fiy0             = _mm256_add_ps(fiy0,ty);
 259             fiz0             = _mm256_add_ps(fiz0,tz);
 260
 261             fjptrA             = f+j_coord_offsetA;
 262             fjptrB             = f+j_coord_offsetB;
 263             fjptrC             = f+j_coord_offsetC;
 264             fjptrD             = f+j_coord_offsetD;
 265             fjptrE             = f+j_coord_offsetE;
 266             fjptrF             = f+j_coord_offsetF;
 267             fjptrG             = f+j_coord_offsetG;
 268             fjptrH             = f+j_coord_offsetH;
 269             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 270
 271             /* Inner loop uses 43 flops */
 272         }
 273
 274         if(jidx<j_index_end)
 275         {
 276
 277             /* Get j neighbor index, and coordinate index */
 278             jnrlistA         = jjnr[jidx];
 279             jnrlistB         = jjnr[jidx+1];
 280             jnrlistC         = jjnr[jidx+2];
 281             jnrlistD         = jjnr[jidx+3];
 282             jnrlistE         = jjnr[jidx+4];
 283             jnrlistF         = jjnr[jidx+5];
 284             jnrlistG         = jjnr[jidx+6];
 285             jnrlistH         = jjnr[jidx+7];
 286             /* Sign of each element will be negative for non-real atoms.
 287              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 288              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 289              */
 290             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 291                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 292
 293             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 294             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 295             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 296             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 297             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 298             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 299             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 300             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 301             j_coord_offsetA  = DIM*jnrA;
 302             j_coord_offsetB  = DIM*jnrB;
 303             j_coord_offsetC  = DIM*jnrC;
 304             j_coord_offsetD  = DIM*jnrD;
 305             j_coord_offsetE  = DIM*jnrE;
 306             j_coord_offsetF  = DIM*jnrF;
 307             j_coord_offsetG  = DIM*jnrG;
 308             j_coord_offsetH  = DIM*jnrH;
 309
 310             /* load j atom coordinates */
 311             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 312                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 313                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 314                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 315                                                  &jx0,&jy0,&jz0);
 316
 317             /* Calculate displacement vector */
 318             dx00             = _mm256_sub_ps(ix0,jx0);
 319             dy00             = _mm256_sub_ps(iy0,jy0);
 320             dz00             = _mm256_sub_ps(iz0,jz0);
 321
 322             /* Calculate squared distance and things based on it */
 323             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 324
 325             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 326
 327             /* Load parameters for j particles */
 328             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 329                                                                  charge+jnrC+0,charge+jnrD+0,
 330                                                                  charge+jnrE+0,charge+jnrF+0,
 331                                                                  charge+jnrG+0,charge+jnrH+0);
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             r00              = _mm256_mul_ps(rsq00,rinv00);
 338             r00              = _mm256_andnot_ps(dummy_mask,r00);
 339
 340             /* Compute parameters for interactions between i and j atoms */
 341             qq00             = _mm256_mul_ps(iq0,jq0);
 342
 343             /* Calculate table index by multiplying r with table scale and truncate to integer */
 344             rt               = _mm256_mul_ps(r00,vftabscale);
 345             vfitab           = _mm256_cvttps_epi32(rt);
 346             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 347             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 348             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 349             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 350             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 351             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 352
 353             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 354             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 355                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 356             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 357                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 358             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 359                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 360             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 361                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 362             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 363             Heps             = _mm256_mul_ps(vfeps,H);
 364             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 365             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 366             velec            = _mm256_mul_ps(qq00,VV);
 367             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 368             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velec            = _mm256_andnot_ps(dummy_mask,velec);
 372             velecsum         = _mm256_add_ps(velecsum,velec);
 373
 374             fscal            = felec;
 375
 376             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 377
 378             /* Calculate temporary vectorial force */
 379             tx               = _mm256_mul_ps(fscal,dx00);
 380             ty               = _mm256_mul_ps(fscal,dy00);
 381             tz               = _mm256_mul_ps(fscal,dz00);
 382
 383             /* Update vectorial force */
 384             fix0             = _mm256_add_ps(fix0,tx);
 385             fiy0             = _mm256_add_ps(fiy0,ty);
 386             fiz0             = _mm256_add_ps(fiz0,tz);
 387
 388             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 389             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 390             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 391             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 392             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 393             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 394             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 395             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 396             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 397
 398             /* Inner loop uses 44 flops */
 399         }
 400
 401         /* End of innermost loop */
 402
 403         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 404                                                  f+i_coord_offset,fshift+i_shift_offset);
 405
 406         ggid                        = gid[iidx];
 407         /* Update potential energies */
 408         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 409
 410         /* Increment number of inner iterations */
 411         inneriter                  += j_index_end - j_index_start;
 412
 413         /* Outer loop uses 8 flops */
 414     }
 415
 416     /* Increment number of outer iterations */
 417     outeriter        += nri;
 418
 419     /* Update outer/inner flops */
 420
 421     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*44);
 422 }
 423 /*
 424  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 425  * Electrostatics interaction: CubicSplineTable
 426  * VdW interaction:            None
 427  * Geometry:                   Particle-Particle
 428  * Calculate force/pot:        Force
 429  */
 430 void
 431 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 432                     (t_nblist                    * gmx_restrict       nlist,
 433                      rvec                        * gmx_restrict          xx,
 434                      rvec                        * gmx_restrict          ff,
 435                      t_forcerec                  * gmx_restrict          fr,
 436                      t_mdatoms                   * gmx_restrict     mdatoms,
 437                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 438                      t_nrnb                      * gmx_restrict        nrnb)
 439 {
 440     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 441      * just 0 for non-waters.
 442      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 443      * jnr indices corresponding to data put in the four positions in the SIMD register.
 444      */
 445     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 446     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 447     int              jnrA,jnrB,jnrC,jnrD;
 448     int              jnrE,jnrF,jnrG,jnrH;
 449     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 450     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 451     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 452     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 453     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 454     real             rcutoff_scalar;
 455     real             *shiftvec,*fshift,*x,*f;
 456     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 457     real             scratch[4*DIM];
 458     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 459     real *           vdwioffsetptr0;
 460     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 461     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 462     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 463     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 464     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 465     real             *charge;
 466     __m256i          vfitab;
 467     __m128i          vfitab_lo,vfitab_hi;
 468     __m128i          ifour       = _mm_set1_epi32(4);
 469     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 470     real             *vftab;
 471     __m256           dummy_mask,cutoff_mask;
 472     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 473     __m256           one     = _mm256_set1_ps(1.0);
 474     __m256           two     = _mm256_set1_ps(2.0);
 475     x                = xx[0];
 476     f                = ff[0];
 477
 478     nri              = nlist->nri;
 479     iinr             = nlist->iinr;
 480     jindex           = nlist->jindex;
 481     jjnr             = nlist->jjnr;
 482     shiftidx         = nlist->shift;
 483     gid              = nlist->gid;
 484     shiftvec         = fr->shift_vec[0];
 485     fshift           = fr->fshift[0];
 486     facel            = _mm256_set1_ps(fr->epsfac);
 487     charge           = mdatoms->chargeA;
 488
 489     vftab            = kernel_data->table_elec->data;
 490     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 491
 492     /* Avoid stupid compiler warnings */
 493     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 494     j_coord_offsetA = 0;
 495     j_coord_offsetB = 0;
 496     j_coord_offsetC = 0;
 497     j_coord_offsetD = 0;
 498     j_coord_offsetE = 0;
 499     j_coord_offsetF = 0;
 500     j_coord_offsetG = 0;
 501     j_coord_offsetH = 0;
 502
 503     outeriter        = 0;
 504     inneriter        = 0;
 505
 506     for(iidx=0;iidx<4*DIM;iidx++)
 507     {
 508         scratch[iidx] = 0.0;
 509     }
 510
 511     /* Start outer loop over neighborlists */
 512     for(iidx=0; iidx<nri; iidx++)
 513     {
 514         /* Load shift vector for this list */
 515         i_shift_offset   = DIM*shiftidx[iidx];
 516
 517         /* Load limits for loop over neighbors */
 518         j_index_start    = jindex[iidx];
 519         j_index_end      = jindex[iidx+1];
 520
 521         /* Get outer coordinate index */
 522         inr              = iinr[iidx];
 523         i_coord_offset   = DIM*inr;
 524
 525         /* Load i particle coords and add shift vector */
 526         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 527
 528         fix0             = _mm256_setzero_ps();
 529         fiy0             = _mm256_setzero_ps();
 530         fiz0             = _mm256_setzero_ps();
 531
 532         /* Load parameters for i particles */
 533         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 534
 535         /* Start inner kernel loop */
 536         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 537         {
 538
 539             /* Get j neighbor index, and coordinate index */
 540             jnrA             = jjnr[jidx];
 541             jnrB             = jjnr[jidx+1];
 542             jnrC             = jjnr[jidx+2];
 543             jnrD             = jjnr[jidx+3];
 544             jnrE             = jjnr[jidx+4];
 545             jnrF             = jjnr[jidx+5];
 546             jnrG             = jjnr[jidx+6];
 547             jnrH             = jjnr[jidx+7];
 548             j_coord_offsetA  = DIM*jnrA;
 549             j_coord_offsetB  = DIM*jnrB;
 550             j_coord_offsetC  = DIM*jnrC;
 551             j_coord_offsetD  = DIM*jnrD;
 552             j_coord_offsetE  = DIM*jnrE;
 553             j_coord_offsetF  = DIM*jnrF;
 554             j_coord_offsetG  = DIM*jnrG;
 555             j_coord_offsetH  = DIM*jnrH;
 556
 557             /* load j atom coordinates */
 558             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 559                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 560                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 561                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 562                                                  &jx0,&jy0,&jz0);
 563
 564             /* Calculate displacement vector */
 565             dx00             = _mm256_sub_ps(ix0,jx0);
 566             dy00             = _mm256_sub_ps(iy0,jy0);
 567             dz00             = _mm256_sub_ps(iz0,jz0);
 568
 569             /* Calculate squared distance and things based on it */
 570             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 571
 572             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 573
 574             /* Load parameters for j particles */
 575             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 576                                                                  charge+jnrC+0,charge+jnrD+0,
 577                                                                  charge+jnrE+0,charge+jnrF+0,
 578                                                                  charge+jnrG+0,charge+jnrH+0);
 579
 580             /**************************
 581              * CALCULATE INTERACTIONS *
 582              **************************/
 583
 584             r00              = _mm256_mul_ps(rsq00,rinv00);
 585
 586             /* Compute parameters for interactions between i and j atoms */
 587             qq00             = _mm256_mul_ps(iq0,jq0);
 588
 589             /* Calculate table index by multiplying r with table scale and truncate to integer */
 590             rt               = _mm256_mul_ps(r00,vftabscale);
 591             vfitab           = _mm256_cvttps_epi32(rt);
 592             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 593             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 594             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 595             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 596             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 597             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 598
 599             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 600             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 601                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 602             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 603                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 604             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 605                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 606             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 607                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 608             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 609             Heps             = _mm256_mul_ps(vfeps,H);
 610             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 611             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 612             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 613
 614             fscal            = felec;
 615
 616             /* Calculate temporary vectorial force */
 617             tx               = _mm256_mul_ps(fscal,dx00);
 618             ty               = _mm256_mul_ps(fscal,dy00);
 619             tz               = _mm256_mul_ps(fscal,dz00);
 620
 621             /* Update vectorial force */
 622             fix0             = _mm256_add_ps(fix0,tx);
 623             fiy0             = _mm256_add_ps(fiy0,ty);
 624             fiz0             = _mm256_add_ps(fiz0,tz);
 625
 626             fjptrA             = f+j_coord_offsetA;
 627             fjptrB             = f+j_coord_offsetB;
 628             fjptrC             = f+j_coord_offsetC;
 629             fjptrD             = f+j_coord_offsetD;
 630             fjptrE             = f+j_coord_offsetE;
 631             fjptrF             = f+j_coord_offsetF;
 632             fjptrG             = f+j_coord_offsetG;
 633             fjptrH             = f+j_coord_offsetH;
 634             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 635
 636             /* Inner loop uses 39 flops */
 637         }
 638
 639         if(jidx<j_index_end)
 640         {
 641
 642             /* Get j neighbor index, and coordinate index */
 643             jnrlistA         = jjnr[jidx];
 644             jnrlistB         = jjnr[jidx+1];
 645             jnrlistC         = jjnr[jidx+2];
 646             jnrlistD         = jjnr[jidx+3];
 647             jnrlistE         = jjnr[jidx+4];
 648             jnrlistF         = jjnr[jidx+5];
 649             jnrlistG         = jjnr[jidx+6];
 650             jnrlistH         = jjnr[jidx+7];
 651             /* Sign of each element will be negative for non-real atoms.
 652              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 653              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 654              */
 655             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 656                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 657
 658             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 659             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 660             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 661             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 662             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 663             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 664             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 665             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 666             j_coord_offsetA  = DIM*jnrA;
 667             j_coord_offsetB  = DIM*jnrB;
 668             j_coord_offsetC  = DIM*jnrC;
 669             j_coord_offsetD  = DIM*jnrD;
 670             j_coord_offsetE  = DIM*jnrE;
 671             j_coord_offsetF  = DIM*jnrF;
 672             j_coord_offsetG  = DIM*jnrG;
 673             j_coord_offsetH  = DIM*jnrH;
 674
 675             /* load j atom coordinates */
 676             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 677                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 678                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 679                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 680                                                  &jx0,&jy0,&jz0);
 681
 682             /* Calculate displacement vector */
 683             dx00             = _mm256_sub_ps(ix0,jx0);
 684             dy00             = _mm256_sub_ps(iy0,jy0);
 685             dz00             = _mm256_sub_ps(iz0,jz0);
 686
 687             /* Calculate squared distance and things based on it */
 688             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 689
 690             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 691
 692             /* Load parameters for j particles */
 693             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 694                                                                  charge+jnrC+0,charge+jnrD+0,
 695                                                                  charge+jnrE+0,charge+jnrF+0,
 696                                                                  charge+jnrG+0,charge+jnrH+0);
 697
 698             /**************************
 699              * CALCULATE INTERACTIONS *
 700              **************************/
 701
 702             r00              = _mm256_mul_ps(rsq00,rinv00);
 703             r00              = _mm256_andnot_ps(dummy_mask,r00);
 704
 705             /* Compute parameters for interactions between i and j atoms */
 706             qq00             = _mm256_mul_ps(iq0,jq0);
 707
 708             /* Calculate table index by multiplying r with table scale and truncate to integer */
 709             rt               = _mm256_mul_ps(r00,vftabscale);
 710             vfitab           = _mm256_cvttps_epi32(rt);
 711             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 712             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 713             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 714             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 715             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 716             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 717
 718             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 719             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 720                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 721             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 722                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 723             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 724                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 725             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 726                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 727             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 728             Heps             = _mm256_mul_ps(vfeps,H);
 729             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 730             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 731             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 736
 737             /* Calculate temporary vectorial force */
 738             tx               = _mm256_mul_ps(fscal,dx00);
 739             ty               = _mm256_mul_ps(fscal,dy00);
 740             tz               = _mm256_mul_ps(fscal,dz00);
 741
 742             /* Update vectorial force */
 743             fix0             = _mm256_add_ps(fix0,tx);
 744             fiy0             = _mm256_add_ps(fiy0,ty);
 745             fiz0             = _mm256_add_ps(fiz0,tz);
 746
 747             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 748             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 749             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 750             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 751             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 752             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 753             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 754             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 755             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 756
 757             /* Inner loop uses 40 flops */
 758         }
 759
 760         /* End of innermost loop */
 761
 762         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 763                                                  f+i_coord_offset,fshift+i_shift_offset);
 764
 765         /* Increment number of inner iterations */
 766         inneriter                  += j_index_end - j_index_start;
 767
 768         /* Outer loop uses 7 flops */
 769     }
 770
 771     /* Increment number of outer iterations */
 772     outeriter        += nri;
 773
 774     /* Update outer/inner flops */
 775
 776     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 777 }