src/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_avx_256_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_256_single.h"
  34 #include "kernelutil_x86_avx_256_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_256_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_256_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrE,jnrF,jnrG,jnrH;
  62     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  63     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  64     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  65     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  66     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  67     real             rcutoff_scalar;
  68     real             *shiftvec,*fshift,*x,*f;
  69     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  70     real             scratch[4*DIM];
  71     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  72     real *           vdwioffsetptr0;
  73     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  75     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  77     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     __m256i          ewitab;
  80     __m128i          ewitab_lo,ewitab_hi;
  81     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  82     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  83     real             *ewtab;
  84     __m256           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  85     real             rswitch_scalar,d_scalar;
  86     __m256           dummy_mask,cutoff_mask;
  87     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  88     __m256           one     = _mm256_set1_ps(1.0);
  89     __m256           two     = _mm256_set1_ps(2.0);
  90     x                = xx[0];
  91     f                = ff[0];
  92
  93     nri              = nlist->nri;
  94     iinr             = nlist->iinr;
  95     jindex           = nlist->jindex;
  96     jjnr             = nlist->jjnr;
  97     shiftidx         = nlist->shift;
  98     gid              = nlist->gid;
  99     shiftvec         = fr->shift_vec[0];
 100     fshift           = fr->fshift[0];
 101     facel            = _mm256_set1_ps(fr->epsfac);
 102     charge           = mdatoms->chargeA;
 103
 104     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 105     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff);
 106     beta2            = _mm256_mul_ps(beta,beta);
 107     beta3            = _mm256_mul_ps(beta,beta2);
 108
 109     ewtab            = fr->ic->tabq_coul_FDV0;
 110     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 111     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 112
 113     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 114     rcutoff_scalar   = fr->rcoulomb;
 115     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 116     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 117
 118     rswitch_scalar   = fr->rcoulomb_switch;
 119     rswitch          = _mm256_set1_ps(rswitch_scalar);
 120     /* Setup switch parameters */
 121     d_scalar         = rcutoff_scalar-rswitch_scalar;
 122     d                = _mm256_set1_ps(d_scalar);
 123     swV3             = _mm256_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 124     swV4             = _mm256_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 125     swV5             = _mm256_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 126     swF2             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 127     swF3             = _mm256_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 128     swF4             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 129
 130     /* Avoid stupid compiler warnings */
 131     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 132     j_coord_offsetA = 0;
 133     j_coord_offsetB = 0;
 134     j_coord_offsetC = 0;
 135     j_coord_offsetD = 0;
 136     j_coord_offsetE = 0;
 137     j_coord_offsetF = 0;
 138     j_coord_offsetG = 0;
 139     j_coord_offsetH = 0;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     for(iidx=0;iidx<4*DIM;iidx++)
 145     {
 146         scratch[iidx] = 0.0;
 147     }
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 165
 166         fix0             = _mm256_setzero_ps();
 167         fiy0             = _mm256_setzero_ps();
 168         fiz0             = _mm256_setzero_ps();
 169
 170         /* Load parameters for i particles */
 171         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 172
 173         /* Reset potential sums */
 174         velecsum         = _mm256_setzero_ps();
 175
 176         /* Start inner kernel loop */
 177         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 178         {
 179
 180             /* Get j neighbor index, and coordinate index */
 181             jnrA             = jjnr[jidx];
 182             jnrB             = jjnr[jidx+1];
 183             jnrC             = jjnr[jidx+2];
 184             jnrD             = jjnr[jidx+3];
 185             jnrE             = jjnr[jidx+4];
 186             jnrF             = jjnr[jidx+5];
 187             jnrG             = jjnr[jidx+6];
 188             jnrH             = jjnr[jidx+7];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191             j_coord_offsetC  = DIM*jnrC;
 192             j_coord_offsetD  = DIM*jnrD;
 193             j_coord_offsetE  = DIM*jnrE;
 194             j_coord_offsetF  = DIM*jnrF;
 195             j_coord_offsetG  = DIM*jnrG;
 196             j_coord_offsetH  = DIM*jnrH;
 197
 198             /* load j atom coordinates */
 199             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 200                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 201                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 202                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 203                                                  &jx0,&jy0,&jz0);
 204
 205             /* Calculate displacement vector */
 206             dx00             = _mm256_sub_ps(ix0,jx0);
 207             dy00             = _mm256_sub_ps(iy0,jy0);
 208             dz00             = _mm256_sub_ps(iz0,jz0);
 209
 210             /* Calculate squared distance and things based on it */
 211             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 212
 213             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 214
 215             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 216
 217             /* Load parameters for j particles */
 218             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 219                                                                  charge+jnrC+0,charge+jnrD+0,
 220                                                                  charge+jnrE+0,charge+jnrF+0,
 221                                                                  charge+jnrG+0,charge+jnrH+0);
 222
 223             /**************************
 224              * CALCULATE INTERACTIONS *
 225              **************************/
 226
 227             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 228             {
 229
 230             r00              = _mm256_mul_ps(rsq00,rinv00);
 231
 232             /* Compute parameters for interactions between i and j atoms */
 233             qq00             = _mm256_mul_ps(iq0,jq0);
 234
 235             /* EWALD ELECTROSTATICS */
 236
 237             /* Analytical PME correction */
 238             zeta2            = _mm256_mul_ps(beta2,rsq00);
 239             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 240             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 241             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 242             felec            = _mm256_mul_ps(qq00,felec);
 243             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 244             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 245             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 246             velec            = _mm256_mul_ps(qq00,velec);
 247
 248             d                = _mm256_sub_ps(r00,rswitch);
 249             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 250             d2               = _mm256_mul_ps(d,d);
 251             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 252
 253             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 254
 255             /* Evaluate switch function */
 256             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 257             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 258             velec            = _mm256_mul_ps(velec,sw);
 259             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 260
 261             /* Update potential sum for this i atom from the interaction with this j atom. */
 262             velec            = _mm256_and_ps(velec,cutoff_mask);
 263             velecsum         = _mm256_add_ps(velecsum,velec);
 264
 265             fscal            = felec;
 266
 267             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 268
 269             /* Calculate temporary vectorial force */
 270             tx               = _mm256_mul_ps(fscal,dx00);
 271             ty               = _mm256_mul_ps(fscal,dy00);
 272             tz               = _mm256_mul_ps(fscal,dz00);
 273
 274             /* Update vectorial force */
 275             fix0             = _mm256_add_ps(fix0,tx);
 276             fiy0             = _mm256_add_ps(fiy0,ty);
 277             fiz0             = _mm256_add_ps(fiz0,tz);
 278
 279             fjptrA             = f+j_coord_offsetA;
 280             fjptrB             = f+j_coord_offsetB;
 281             fjptrC             = f+j_coord_offsetC;
 282             fjptrD             = f+j_coord_offsetD;
 283             fjptrE             = f+j_coord_offsetE;
 284             fjptrF             = f+j_coord_offsetF;
 285             fjptrG             = f+j_coord_offsetG;
 286             fjptrH             = f+j_coord_offsetH;
 287             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 288
 289             }
 290
 291             /* Inner loop uses 108 flops */
 292         }
 293
 294         if(jidx<j_index_end)
 295         {
 296
 297             /* Get j neighbor index, and coordinate index */
 298             jnrlistA         = jjnr[jidx];
 299             jnrlistB         = jjnr[jidx+1];
 300             jnrlistC         = jjnr[jidx+2];
 301             jnrlistD         = jjnr[jidx+3];
 302             jnrlistE         = jjnr[jidx+4];
 303             jnrlistF         = jjnr[jidx+5];
 304             jnrlistG         = jjnr[jidx+6];
 305             jnrlistH         = jjnr[jidx+7];
 306             /* Sign of each element will be negative for non-real atoms.
 307              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 308              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 309              */
 310             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 311                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 312
 313             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 314             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 315             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 316             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 317             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 318             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 319             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 320             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 321             j_coord_offsetA  = DIM*jnrA;
 322             j_coord_offsetB  = DIM*jnrB;
 323             j_coord_offsetC  = DIM*jnrC;
 324             j_coord_offsetD  = DIM*jnrD;
 325             j_coord_offsetE  = DIM*jnrE;
 326             j_coord_offsetF  = DIM*jnrF;
 327             j_coord_offsetG  = DIM*jnrG;
 328             j_coord_offsetH  = DIM*jnrH;
 329
 330             /* load j atom coordinates */
 331             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 332                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 333                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 334                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 335                                                  &jx0,&jy0,&jz0);
 336
 337             /* Calculate displacement vector */
 338             dx00             = _mm256_sub_ps(ix0,jx0);
 339             dy00             = _mm256_sub_ps(iy0,jy0);
 340             dz00             = _mm256_sub_ps(iz0,jz0);
 341
 342             /* Calculate squared distance and things based on it */
 343             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 344
 345             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 346
 347             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 348
 349             /* Load parameters for j particles */
 350             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 351                                                                  charge+jnrC+0,charge+jnrD+0,
 352                                                                  charge+jnrE+0,charge+jnrF+0,
 353                                                                  charge+jnrG+0,charge+jnrH+0);
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 360             {
 361
 362             r00              = _mm256_mul_ps(rsq00,rinv00);
 363             r00              = _mm256_andnot_ps(dummy_mask,r00);
 364
 365             /* Compute parameters for interactions between i and j atoms */
 366             qq00             = _mm256_mul_ps(iq0,jq0);
 367
 368             /* EWALD ELECTROSTATICS */
 369
 370             /* Analytical PME correction */
 371             zeta2            = _mm256_mul_ps(beta2,rsq00);
 372             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 373             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 374             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 375             felec            = _mm256_mul_ps(qq00,felec);
 376             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 377             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 378             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 379             velec            = _mm256_mul_ps(qq00,velec);
 380
 381             d                = _mm256_sub_ps(r00,rswitch);
 382             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 383             d2               = _mm256_mul_ps(d,d);
 384             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 385
 386             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 387
 388             /* Evaluate switch function */
 389             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 390             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 391             velec            = _mm256_mul_ps(velec,sw);
 392             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 393
 394             /* Update potential sum for this i atom from the interaction with this j atom. */
 395             velec            = _mm256_and_ps(velec,cutoff_mask);
 396             velec            = _mm256_andnot_ps(dummy_mask,velec);
 397             velecsum         = _mm256_add_ps(velecsum,velec);
 398
 399             fscal            = felec;
 400
 401             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 402
 403             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 404
 405             /* Calculate temporary vectorial force */
 406             tx               = _mm256_mul_ps(fscal,dx00);
 407             ty               = _mm256_mul_ps(fscal,dy00);
 408             tz               = _mm256_mul_ps(fscal,dz00);
 409
 410             /* Update vectorial force */
 411             fix0             = _mm256_add_ps(fix0,tx);
 412             fiy0             = _mm256_add_ps(fiy0,ty);
 413             fiz0             = _mm256_add_ps(fiz0,tz);
 414
 415             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 416             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 417             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 418             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 419             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 420             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 421             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 422             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 423             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 424
 425             }
 426
 427             /* Inner loop uses 109 flops */
 428         }
 429
 430         /* End of innermost loop */
 431
 432         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 433                                                  f+i_coord_offset,fshift+i_shift_offset);
 434
 435         ggid                        = gid[iidx];
 436         /* Update potential energies */
 437         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 438
 439         /* Increment number of inner iterations */
 440         inneriter                  += j_index_end - j_index_start;
 441
 442         /* Outer loop uses 8 flops */
 443     }
 444
 445     /* Increment number of outer iterations */
 446     outeriter        += nri;
 447
 448     /* Update outer/inner flops */
 449
 450     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*109);
 451 }
 452 /*
 453  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_single
 454  * Electrostatics interaction: Ewald
 455  * VdW interaction:            None
 456  * Geometry:                   Particle-Particle
 457  * Calculate force/pot:        Force
 458  */
 459 void
 460 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_single
 461                     (t_nblist * gmx_restrict                nlist,
 462                      rvec * gmx_restrict                    xx,
 463                      rvec * gmx_restrict                    ff,
 464                      t_forcerec * gmx_restrict              fr,
 465                      t_mdatoms * gmx_restrict               mdatoms,
 466                      nb_kernel_data_t * gmx_restrict        kernel_data,
 467                      t_nrnb * gmx_restrict                  nrnb)
 468 {
 469     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 470      * just 0 for non-waters.
 471      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 472      * jnr indices corresponding to data put in the four positions in the SIMD register.
 473      */
 474     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 475     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 476     int              jnrA,jnrB,jnrC,jnrD;
 477     int              jnrE,jnrF,jnrG,jnrH;
 478     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 479     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 480     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 481     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 482     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 483     real             rcutoff_scalar;
 484     real             *shiftvec,*fshift,*x,*f;
 485     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 486     real             scratch[4*DIM];
 487     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 488     real *           vdwioffsetptr0;
 489     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 490     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 491     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 492     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 493     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 494     real             *charge;
 495     __m256i          ewitab;
 496     __m128i          ewitab_lo,ewitab_hi;
 497     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 498     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 499     real             *ewtab;
 500     __m256           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 501     real             rswitch_scalar,d_scalar;
 502     __m256           dummy_mask,cutoff_mask;
 503     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 504     __m256           one     = _mm256_set1_ps(1.0);
 505     __m256           two     = _mm256_set1_ps(2.0);
 506     x                = xx[0];
 507     f                = ff[0];
 508
 509     nri              = nlist->nri;
 510     iinr             = nlist->iinr;
 511     jindex           = nlist->jindex;
 512     jjnr             = nlist->jjnr;
 513     shiftidx         = nlist->shift;
 514     gid              = nlist->gid;
 515     shiftvec         = fr->shift_vec[0];
 516     fshift           = fr->fshift[0];
 517     facel            = _mm256_set1_ps(fr->epsfac);
 518     charge           = mdatoms->chargeA;
 519
 520     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 521     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff);
 522     beta2            = _mm256_mul_ps(beta,beta);
 523     beta3            = _mm256_mul_ps(beta,beta2);
 524
 525     ewtab            = fr->ic->tabq_coul_FDV0;
 526     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 527     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 528
 529     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 530     rcutoff_scalar   = fr->rcoulomb;
 531     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 532     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 533
 534     rswitch_scalar   = fr->rcoulomb_switch;
 535     rswitch          = _mm256_set1_ps(rswitch_scalar);
 536     /* Setup switch parameters */
 537     d_scalar         = rcutoff_scalar-rswitch_scalar;
 538     d                = _mm256_set1_ps(d_scalar);
 539     swV3             = _mm256_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 540     swV4             = _mm256_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 541     swV5             = _mm256_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 542     swF2             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 543     swF3             = _mm256_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 544     swF4             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 545
 546     /* Avoid stupid compiler warnings */
 547     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 548     j_coord_offsetA = 0;
 549     j_coord_offsetB = 0;
 550     j_coord_offsetC = 0;
 551     j_coord_offsetD = 0;
 552     j_coord_offsetE = 0;
 553     j_coord_offsetF = 0;
 554     j_coord_offsetG = 0;
 555     j_coord_offsetH = 0;
 556
 557     outeriter        = 0;
 558     inneriter        = 0;
 559
 560     for(iidx=0;iidx<4*DIM;iidx++)
 561     {
 562         scratch[iidx] = 0.0;
 563     }
 564
 565     /* Start outer loop over neighborlists */
 566     for(iidx=0; iidx<nri; iidx++)
 567     {
 568         /* Load shift vector for this list */
 569         i_shift_offset   = DIM*shiftidx[iidx];
 570
 571         /* Load limits for loop over neighbors */
 572         j_index_start    = jindex[iidx];
 573         j_index_end      = jindex[iidx+1];
 574
 575         /* Get outer coordinate index */
 576         inr              = iinr[iidx];
 577         i_coord_offset   = DIM*inr;
 578
 579         /* Load i particle coords and add shift vector */
 580         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 581
 582         fix0             = _mm256_setzero_ps();
 583         fiy0             = _mm256_setzero_ps();
 584         fiz0             = _mm256_setzero_ps();
 585
 586         /* Load parameters for i particles */
 587         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 588
 589         /* Start inner kernel loop */
 590         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 591         {
 592
 593             /* Get j neighbor index, and coordinate index */
 594             jnrA             = jjnr[jidx];
 595             jnrB             = jjnr[jidx+1];
 596             jnrC             = jjnr[jidx+2];
 597             jnrD             = jjnr[jidx+3];
 598             jnrE             = jjnr[jidx+4];
 599             jnrF             = jjnr[jidx+5];
 600             jnrG             = jjnr[jidx+6];
 601             jnrH             = jjnr[jidx+7];
 602             j_coord_offsetA  = DIM*jnrA;
 603             j_coord_offsetB  = DIM*jnrB;
 604             j_coord_offsetC  = DIM*jnrC;
 605             j_coord_offsetD  = DIM*jnrD;
 606             j_coord_offsetE  = DIM*jnrE;
 607             j_coord_offsetF  = DIM*jnrF;
 608             j_coord_offsetG  = DIM*jnrG;
 609             j_coord_offsetH  = DIM*jnrH;
 610
 611             /* load j atom coordinates */
 612             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 613                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 614                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 615                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 616                                                  &jx0,&jy0,&jz0);
 617
 618             /* Calculate displacement vector */
 619             dx00             = _mm256_sub_ps(ix0,jx0);
 620             dy00             = _mm256_sub_ps(iy0,jy0);
 621             dz00             = _mm256_sub_ps(iz0,jz0);
 622
 623             /* Calculate squared distance and things based on it */
 624             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 625
 626             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 627
 628             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 629
 630             /* Load parameters for j particles */
 631             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 632                                                                  charge+jnrC+0,charge+jnrD+0,
 633                                                                  charge+jnrE+0,charge+jnrF+0,
 634                                                                  charge+jnrG+0,charge+jnrH+0);
 635
 636             /**************************
 637              * CALCULATE INTERACTIONS *
 638              **************************/
 639
 640             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 641             {
 642
 643             r00              = _mm256_mul_ps(rsq00,rinv00);
 644
 645             /* Compute parameters for interactions between i and j atoms */
 646             qq00             = _mm256_mul_ps(iq0,jq0);
 647
 648             /* EWALD ELECTROSTATICS */
 649
 650             /* Analytical PME correction */
 651             zeta2            = _mm256_mul_ps(beta2,rsq00);
 652             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 653             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 654             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 655             felec            = _mm256_mul_ps(qq00,felec);
 656             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 657             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 658             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 659             velec            = _mm256_mul_ps(qq00,velec);
 660
 661             d                = _mm256_sub_ps(r00,rswitch);
 662             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 663             d2               = _mm256_mul_ps(d,d);
 664             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 665
 666             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 667
 668             /* Evaluate switch function */
 669             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 670             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 671             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 672
 673             fscal            = felec;
 674
 675             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 676
 677             /* Calculate temporary vectorial force */
 678             tx               = _mm256_mul_ps(fscal,dx00);
 679             ty               = _mm256_mul_ps(fscal,dy00);
 680             tz               = _mm256_mul_ps(fscal,dz00);
 681
 682             /* Update vectorial force */
 683             fix0             = _mm256_add_ps(fix0,tx);
 684             fiy0             = _mm256_add_ps(fiy0,ty);
 685             fiz0             = _mm256_add_ps(fiz0,tz);
 686
 687             fjptrA             = f+j_coord_offsetA;
 688             fjptrB             = f+j_coord_offsetB;
 689             fjptrC             = f+j_coord_offsetC;
 690             fjptrD             = f+j_coord_offsetD;
 691             fjptrE             = f+j_coord_offsetE;
 692             fjptrF             = f+j_coord_offsetF;
 693             fjptrG             = f+j_coord_offsetG;
 694             fjptrH             = f+j_coord_offsetH;
 695             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 696
 697             }
 698
 699             /* Inner loop uses 105 flops */
 700         }
 701
 702         if(jidx<j_index_end)
 703         {
 704
 705             /* Get j neighbor index, and coordinate index */
 706             jnrlistA         = jjnr[jidx];
 707             jnrlistB         = jjnr[jidx+1];
 708             jnrlistC         = jjnr[jidx+2];
 709             jnrlistD         = jjnr[jidx+3];
 710             jnrlistE         = jjnr[jidx+4];
 711             jnrlistF         = jjnr[jidx+5];
 712             jnrlistG         = jjnr[jidx+6];
 713             jnrlistH         = jjnr[jidx+7];
 714             /* Sign of each element will be negative for non-real atoms.
 715              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 716              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 717              */
 718             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 719                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 720
 721             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 722             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 723             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 724             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 725             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 726             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 727             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 728             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 729             j_coord_offsetA  = DIM*jnrA;
 730             j_coord_offsetB  = DIM*jnrB;
 731             j_coord_offsetC  = DIM*jnrC;
 732             j_coord_offsetD  = DIM*jnrD;
 733             j_coord_offsetE  = DIM*jnrE;
 734             j_coord_offsetF  = DIM*jnrF;
 735             j_coord_offsetG  = DIM*jnrG;
 736             j_coord_offsetH  = DIM*jnrH;
 737
 738             /* load j atom coordinates */
 739             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 740                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 741                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 742                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 743                                                  &jx0,&jy0,&jz0);
 744
 745             /* Calculate displacement vector */
 746             dx00             = _mm256_sub_ps(ix0,jx0);
 747             dy00             = _mm256_sub_ps(iy0,jy0);
 748             dz00             = _mm256_sub_ps(iz0,jz0);
 749
 750             /* Calculate squared distance and things based on it */
 751             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 752
 753             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 754
 755             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 756
 757             /* Load parameters for j particles */
 758             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 759                                                                  charge+jnrC+0,charge+jnrD+0,
 760                                                                  charge+jnrE+0,charge+jnrF+0,
 761                                                                  charge+jnrG+0,charge+jnrH+0);
 762
 763             /**************************
 764              * CALCULATE INTERACTIONS *
 765              **************************/
 766
 767             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 768             {
 769
 770             r00              = _mm256_mul_ps(rsq00,rinv00);
 771             r00              = _mm256_andnot_ps(dummy_mask,r00);
 772
 773             /* Compute parameters for interactions between i and j atoms */
 774             qq00             = _mm256_mul_ps(iq0,jq0);
 775
 776             /* EWALD ELECTROSTATICS */
 777
 778             /* Analytical PME correction */
 779             zeta2            = _mm256_mul_ps(beta2,rsq00);
 780             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 781             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 782             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 783             felec            = _mm256_mul_ps(qq00,felec);
 784             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 785             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 786             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 787             velec            = _mm256_mul_ps(qq00,velec);
 788
 789             d                = _mm256_sub_ps(r00,rswitch);
 790             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 791             d2               = _mm256_mul_ps(d,d);
 792             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 793
 794             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 795
 796             /* Evaluate switch function */
 797             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 798             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 799             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 800
 801             fscal            = felec;
 802
 803             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 804
 805             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 806
 807             /* Calculate temporary vectorial force */
 808             tx               = _mm256_mul_ps(fscal,dx00);
 809             ty               = _mm256_mul_ps(fscal,dy00);
 810             tz               = _mm256_mul_ps(fscal,dz00);
 811
 812             /* Update vectorial force */
 813             fix0             = _mm256_add_ps(fix0,tx);
 814             fiy0             = _mm256_add_ps(fiy0,ty);
 815             fiz0             = _mm256_add_ps(fiz0,tz);
 816
 817             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 818             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 819             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 820             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 821             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 822             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 823             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 824             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 825             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 826
 827             }
 828
 829             /* Inner loop uses 106 flops */
 830         }
 831
 832         /* End of innermost loop */
 833
 834         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 835                                                  f+i_coord_offset,fshift+i_shift_offset);
 836
 837         /* Increment number of inner iterations */
 838         inneriter                  += j_index_end - j_index_start;
 839
 840         /* Outer loop uses 7 flops */
 841     }
 842
 843     /* Increment number of outer iterations */
 844     outeriter        += nri;
 845
 846     /* Update outer/inner flops */
 847
 848     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*106);
 849 }