src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwLJEw_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJEw_GeomP1P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  85     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  94     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  95     __m128           c6grid_00;
  96     real             *vdwgridparam;
  97     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  98     __m128           one_half = _mm_set1_ps(0.5);
  99     __m128           minus_one = _mm_set1_ps(-1.0);
 100     __m128i          ewitab;
 101     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 102     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 103     real             *ewtab;
 104     __m128           dummy_mask,cutoff_mask;
 105     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 106     __m128           one     = _mm_set1_ps(1.0);
 107     __m128           two     = _mm_set1_ps(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_ps(fr->ic->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124     vdwgridparam     = fr->ljpme_c6grid;
 125     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 126     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 127     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 128
 129     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 130     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 131     beta2            = _mm_mul_ps(beta,beta);
 132     beta3            = _mm_mul_ps(beta,beta2);
 133     ewtab            = fr->ic->tabq_coul_FDV0;
 134     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 135     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 136
 137     /* Avoid stupid compiler warnings */
 138     jnrA = jnrB = jnrC = jnrD = 0;
 139     j_coord_offsetA = 0;
 140     j_coord_offsetB = 0;
 141     j_coord_offsetC = 0;
 142     j_coord_offsetD = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     for(iidx=0;iidx<4*DIM;iidx++)
 148     {
 149         scratch[iidx] = 0.0;
 150     }
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157
 158         /* Load limits for loop over neighbors */
 159         j_index_start    = jindex[iidx];
 160         j_index_end      = jindex[iidx+1];
 161
 162         /* Get outer coordinate index */
 163         inr              = iinr[iidx];
 164         i_coord_offset   = DIM*inr;
 165
 166         /* Load i particle coords and add shift vector */
 167         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 168
 169         fix0             = _mm_setzero_ps();
 170         fiy0             = _mm_setzero_ps();
 171         fiz0             = _mm_setzero_ps();
 172
 173         /* Load parameters for i particles */
 174         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 175         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 176
 177         /* Reset potential sums */
 178         velecsum         = _mm_setzero_ps();
 179         vvdwsum          = _mm_setzero_ps();
 180
 181         /* Start inner kernel loop */
 182         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 183         {
 184
 185             /* Get j neighbor index, and coordinate index */
 186             jnrA             = jjnr[jidx];
 187             jnrB             = jjnr[jidx+1];
 188             jnrC             = jjnr[jidx+2];
 189             jnrD             = jjnr[jidx+3];
 190             j_coord_offsetA  = DIM*jnrA;
 191             j_coord_offsetB  = DIM*jnrB;
 192             j_coord_offsetC  = DIM*jnrC;
 193             j_coord_offsetD  = DIM*jnrD;
 194
 195             /* load j atom coordinates */
 196             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 197                                               x+j_coord_offsetC,x+j_coord_offsetD,
 198                                               &jx0,&jy0,&jz0);
 199
 200             /* Calculate displacement vector */
 201             dx00             = _mm_sub_ps(ix0,jx0);
 202             dy00             = _mm_sub_ps(iy0,jy0);
 203             dz00             = _mm_sub_ps(iz0,jz0);
 204
 205             /* Calculate squared distance and things based on it */
 206             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 207
 208             rinv00           = avx128fma_invsqrt_f(rsq00);
 209
 210             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 211
 212             /* Load parameters for j particles */
 213             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 214                                                               charge+jnrC+0,charge+jnrD+0);
 215             vdwjidx0A        = 2*vdwtype[jnrA+0];
 216             vdwjidx0B        = 2*vdwtype[jnrB+0];
 217             vdwjidx0C        = 2*vdwtype[jnrC+0];
 218             vdwjidx0D        = 2*vdwtype[jnrD+0];
 219
 220             /**************************
 221              * CALCULATE INTERACTIONS *
 222              **************************/
 223
 224             r00              = _mm_mul_ps(rsq00,rinv00);
 225
 226             /* Compute parameters for interactions between i and j atoms */
 227             qq00             = _mm_mul_ps(iq0,jq0);
 228             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 229                                          vdwparam+vdwioffset0+vdwjidx0B,
 230                                          vdwparam+vdwioffset0+vdwjidx0C,
 231                                          vdwparam+vdwioffset0+vdwjidx0D,
 232                                          &c6_00,&c12_00);
 233
 234             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 235                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 236                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 237                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 238
 239             /* EWALD ELECTROSTATICS */
 240
 241             /* Analytical PME correction */
 242             zeta2            = _mm_mul_ps(beta2,rsq00);
 243             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 244             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 245             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 246             felec            = _mm_mul_ps(qq00,felec);
 247             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 248             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 249             velec            = _mm_mul_ps(qq00,velec);
 250
 251             /* Analytical LJ-PME */
 252             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 253             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 254             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 255             exponent         = avx128fma_exp_f(ewcljrsq);
 256             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 257             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 258             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 259             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 260             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 261             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 262             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 263             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 264
 265             /* Update potential sum for this i atom from the interaction with this j atom. */
 266             velecsum         = _mm_add_ps(velecsum,velec);
 267             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 268
 269             fscal            = _mm_add_ps(felec,fvdw);
 270
 271              /* Update vectorial force */
 272             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 273             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 274             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 275
 276             fjptrA             = f+j_coord_offsetA;
 277             fjptrB             = f+j_coord_offsetB;
 278             fjptrC             = f+j_coord_offsetC;
 279             fjptrD             = f+j_coord_offsetD;
 280             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 281                                                    _mm_mul_ps(dx00,fscal),
 282                                                    _mm_mul_ps(dy00,fscal),
 283                                                    _mm_mul_ps(dz00,fscal));
 284
 285             /* Inner loop uses 53 flops */
 286         }
 287
 288         if(jidx<j_index_end)
 289         {
 290
 291             /* Get j neighbor index, and coordinate index */
 292             jnrlistA         = jjnr[jidx];
 293             jnrlistB         = jjnr[jidx+1];
 294             jnrlistC         = jjnr[jidx+2];
 295             jnrlistD         = jjnr[jidx+3];
 296             /* Sign of each element will be negative for non-real atoms.
 297              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 298              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 299              */
 300             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 301             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 302             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 303             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 304             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 305             j_coord_offsetA  = DIM*jnrA;
 306             j_coord_offsetB  = DIM*jnrB;
 307             j_coord_offsetC  = DIM*jnrC;
 308             j_coord_offsetD  = DIM*jnrD;
 309
 310             /* load j atom coordinates */
 311             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 312                                               x+j_coord_offsetC,x+j_coord_offsetD,
 313                                               &jx0,&jy0,&jz0);
 314
 315             /* Calculate displacement vector */
 316             dx00             = _mm_sub_ps(ix0,jx0);
 317             dy00             = _mm_sub_ps(iy0,jy0);
 318             dz00             = _mm_sub_ps(iz0,jz0);
 319
 320             /* Calculate squared distance and things based on it */
 321             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 322
 323             rinv00           = avx128fma_invsqrt_f(rsq00);
 324
 325             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 326
 327             /* Load parameters for j particles */
 328             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 329                                                               charge+jnrC+0,charge+jnrD+0);
 330             vdwjidx0A        = 2*vdwtype[jnrA+0];
 331             vdwjidx0B        = 2*vdwtype[jnrB+0];
 332             vdwjidx0C        = 2*vdwtype[jnrC+0];
 333             vdwjidx0D        = 2*vdwtype[jnrD+0];
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             r00              = _mm_mul_ps(rsq00,rinv00);
 340             r00              = _mm_andnot_ps(dummy_mask,r00);
 341
 342             /* Compute parameters for interactions between i and j atoms */
 343             qq00             = _mm_mul_ps(iq0,jq0);
 344             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 345                                          vdwparam+vdwioffset0+vdwjidx0B,
 346                                          vdwparam+vdwioffset0+vdwjidx0C,
 347                                          vdwparam+vdwioffset0+vdwjidx0D,
 348                                          &c6_00,&c12_00);
 349
 350             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 351                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 352                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 353                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 354
 355             /* EWALD ELECTROSTATICS */
 356
 357             /* Analytical PME correction */
 358             zeta2            = _mm_mul_ps(beta2,rsq00);
 359             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 360             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 361             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 362             felec            = _mm_mul_ps(qq00,felec);
 363             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 364             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 365             velec            = _mm_mul_ps(qq00,velec);
 366
 367             /* Analytical LJ-PME */
 368             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 369             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 370             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 371             exponent         = avx128fma_exp_f(ewcljrsq);
 372             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 373             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 374             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 375             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 376             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 377             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 378             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 379             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velec            = _mm_andnot_ps(dummy_mask,velec);
 383             velecsum         = _mm_add_ps(velecsum,velec);
 384             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 385             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 386
 387             fscal            = _mm_add_ps(felec,fvdw);
 388
 389             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 390
 391              /* Update vectorial force */
 392             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 393             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 394             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 395
 396             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 397             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 398             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 399             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 400             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 401                                                    _mm_mul_ps(dx00,fscal),
 402                                                    _mm_mul_ps(dy00,fscal),
 403                                                    _mm_mul_ps(dz00,fscal));
 404
 405             /* Inner loop uses 54 flops */
 406         }
 407
 408         /* End of innermost loop */
 409
 410         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 411                                               f+i_coord_offset,fshift+i_shift_offset);
 412
 413         ggid                        = gid[iidx];
 414         /* Update potential energies */
 415         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 416         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 417
 418         /* Increment number of inner iterations */
 419         inneriter                  += j_index_end - j_index_start;
 420
 421         /* Outer loop uses 9 flops */
 422     }
 423
 424     /* Increment number of outer iterations */
 425     outeriter        += nri;
 426
 427     /* Update outer/inner flops */
 428
 429     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 430 }
 431 /*
 432  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 433  * Electrostatics interaction: Ewald
 434  * VdW interaction:            LJEwald
 435  * Geometry:                   Particle-Particle
 436  * Calculate force/pot:        Force
 437  */
 438 void
 439 nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 440                     (t_nblist                    * gmx_restrict       nlist,
 441                      rvec                        * gmx_restrict          xx,
 442                      rvec                        * gmx_restrict          ff,
 443                      struct t_forcerec           * gmx_restrict          fr,
 444                      t_mdatoms                   * gmx_restrict     mdatoms,
 445                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 446                      t_nrnb                      * gmx_restrict        nrnb)
 447 {
 448     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 449      * just 0 for non-waters.
 450      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 451      * jnr indices corresponding to data put in the four positions in the SIMD register.
 452      */
 453     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 454     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 455     int              jnrA,jnrB,jnrC,jnrD;
 456     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 457     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 458     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 459     real             rcutoff_scalar;
 460     real             *shiftvec,*fshift,*x,*f;
 461     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 462     real             scratch[4*DIM];
 463     __m128           fscal,rcutoff,rcutoff2,jidxall;
 464     int              vdwioffset0;
 465     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 466     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 467     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 468     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 469     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 470     real             *charge;
 471     int              nvdwtype;
 472     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 473     int              *vdwtype;
 474     real             *vdwparam;
 475     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 476     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 477     __m128           c6grid_00;
 478     real             *vdwgridparam;
 479     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 480     __m128           one_half = _mm_set1_ps(0.5);
 481     __m128           minus_one = _mm_set1_ps(-1.0);
 482     __m128i          ewitab;
 483     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 484     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 485     real             *ewtab;
 486     __m128           dummy_mask,cutoff_mask;
 487     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 488     __m128           one     = _mm_set1_ps(1.0);
 489     __m128           two     = _mm_set1_ps(2.0);
 490     x                = xx[0];
 491     f                = ff[0];
 492
 493     nri              = nlist->nri;
 494     iinr             = nlist->iinr;
 495     jindex           = nlist->jindex;
 496     jjnr             = nlist->jjnr;
 497     shiftidx         = nlist->shift;
 498     gid              = nlist->gid;
 499     shiftvec         = fr->shift_vec[0];
 500     fshift           = fr->fshift[0];
 501     facel            = _mm_set1_ps(fr->ic->epsfac);
 502     charge           = mdatoms->chargeA;
 503     nvdwtype         = fr->ntype;
 504     vdwparam         = fr->nbfp;
 505     vdwtype          = mdatoms->typeA;
 506     vdwgridparam     = fr->ljpme_c6grid;
 507     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 508     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 509     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 510
 511     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 512     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 513     beta2            = _mm_mul_ps(beta,beta);
 514     beta3            = _mm_mul_ps(beta,beta2);
 515     ewtab            = fr->ic->tabq_coul_F;
 516     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 517     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 518
 519     /* Avoid stupid compiler warnings */
 520     jnrA = jnrB = jnrC = jnrD = 0;
 521     j_coord_offsetA = 0;
 522     j_coord_offsetB = 0;
 523     j_coord_offsetC = 0;
 524     j_coord_offsetD = 0;
 525
 526     outeriter        = 0;
 527     inneriter        = 0;
 528
 529     for(iidx=0;iidx<4*DIM;iidx++)
 530     {
 531         scratch[iidx] = 0.0;
 532     }
 533
 534     /* Start outer loop over neighborlists */
 535     for(iidx=0; iidx<nri; iidx++)
 536     {
 537         /* Load shift vector for this list */
 538         i_shift_offset   = DIM*shiftidx[iidx];
 539
 540         /* Load limits for loop over neighbors */
 541         j_index_start    = jindex[iidx];
 542         j_index_end      = jindex[iidx+1];
 543
 544         /* Get outer coordinate index */
 545         inr              = iinr[iidx];
 546         i_coord_offset   = DIM*inr;
 547
 548         /* Load i particle coords and add shift vector */
 549         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 550
 551         fix0             = _mm_setzero_ps();
 552         fiy0             = _mm_setzero_ps();
 553         fiz0             = _mm_setzero_ps();
 554
 555         /* Load parameters for i particles */
 556         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 557         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 558
 559         /* Start inner kernel loop */
 560         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 561         {
 562
 563             /* Get j neighbor index, and coordinate index */
 564             jnrA             = jjnr[jidx];
 565             jnrB             = jjnr[jidx+1];
 566             jnrC             = jjnr[jidx+2];
 567             jnrD             = jjnr[jidx+3];
 568             j_coord_offsetA  = DIM*jnrA;
 569             j_coord_offsetB  = DIM*jnrB;
 570             j_coord_offsetC  = DIM*jnrC;
 571             j_coord_offsetD  = DIM*jnrD;
 572
 573             /* load j atom coordinates */
 574             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 575                                               x+j_coord_offsetC,x+j_coord_offsetD,
 576                                               &jx0,&jy0,&jz0);
 577
 578             /* Calculate displacement vector */
 579             dx00             = _mm_sub_ps(ix0,jx0);
 580             dy00             = _mm_sub_ps(iy0,jy0);
 581             dz00             = _mm_sub_ps(iz0,jz0);
 582
 583             /* Calculate squared distance and things based on it */
 584             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 585
 586             rinv00           = avx128fma_invsqrt_f(rsq00);
 587
 588             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 589
 590             /* Load parameters for j particles */
 591             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 592                                                               charge+jnrC+0,charge+jnrD+0);
 593             vdwjidx0A        = 2*vdwtype[jnrA+0];
 594             vdwjidx0B        = 2*vdwtype[jnrB+0];
 595             vdwjidx0C        = 2*vdwtype[jnrC+0];
 596             vdwjidx0D        = 2*vdwtype[jnrD+0];
 597
 598             /**************************
 599              * CALCULATE INTERACTIONS *
 600              **************************/
 601
 602             r00              = _mm_mul_ps(rsq00,rinv00);
 603
 604             /* Compute parameters for interactions between i and j atoms */
 605             qq00             = _mm_mul_ps(iq0,jq0);
 606             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 607                                          vdwparam+vdwioffset0+vdwjidx0B,
 608                                          vdwparam+vdwioffset0+vdwjidx0C,
 609                                          vdwparam+vdwioffset0+vdwjidx0D,
 610                                          &c6_00,&c12_00);
 611
 612             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 613                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 614                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 615                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 616
 617             /* EWALD ELECTROSTATICS */
 618
 619             /* Analytical PME correction */
 620             zeta2            = _mm_mul_ps(beta2,rsq00);
 621             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 622             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 623             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 624             felec            = _mm_mul_ps(qq00,felec);
 625
 626             /* Analytical LJ-PME */
 627             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 628             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 629             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 630             exponent         = avx128fma_exp_f(ewcljrsq);
 631             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 632             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 633             /* f6A = 6 * C6grid * (1 - poly) */
 634             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 635             /* f6B = C6grid * exponent * beta^6 */
 636             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 637             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 638             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 639
 640             fscal            = _mm_add_ps(felec,fvdw);
 641
 642              /* Update vectorial force */
 643             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 644             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 645             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 646
 647             fjptrA             = f+j_coord_offsetA;
 648             fjptrB             = f+j_coord_offsetB;
 649             fjptrC             = f+j_coord_offsetC;
 650             fjptrD             = f+j_coord_offsetD;
 651             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 652                                                    _mm_mul_ps(dx00,fscal),
 653                                                    _mm_mul_ps(dy00,fscal),
 654                                                    _mm_mul_ps(dz00,fscal));
 655
 656             /* Inner loop uses 49 flops */
 657         }
 658
 659         if(jidx<j_index_end)
 660         {
 661
 662             /* Get j neighbor index, and coordinate index */
 663             jnrlistA         = jjnr[jidx];
 664             jnrlistB         = jjnr[jidx+1];
 665             jnrlistC         = jjnr[jidx+2];
 666             jnrlistD         = jjnr[jidx+3];
 667             /* Sign of each element will be negative for non-real atoms.
 668              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 669              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 670              */
 671             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 672             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 673             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 674             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 675             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 676             j_coord_offsetA  = DIM*jnrA;
 677             j_coord_offsetB  = DIM*jnrB;
 678             j_coord_offsetC  = DIM*jnrC;
 679             j_coord_offsetD  = DIM*jnrD;
 680
 681             /* load j atom coordinates */
 682             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 683                                               x+j_coord_offsetC,x+j_coord_offsetD,
 684                                               &jx0,&jy0,&jz0);
 685
 686             /* Calculate displacement vector */
 687             dx00             = _mm_sub_ps(ix0,jx0);
 688             dy00             = _mm_sub_ps(iy0,jy0);
 689             dz00             = _mm_sub_ps(iz0,jz0);
 690
 691             /* Calculate squared distance and things based on it */
 692             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 693
 694             rinv00           = avx128fma_invsqrt_f(rsq00);
 695
 696             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 697
 698             /* Load parameters for j particles */
 699             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 700                                                               charge+jnrC+0,charge+jnrD+0);
 701             vdwjidx0A        = 2*vdwtype[jnrA+0];
 702             vdwjidx0B        = 2*vdwtype[jnrB+0];
 703             vdwjidx0C        = 2*vdwtype[jnrC+0];
 704             vdwjidx0D        = 2*vdwtype[jnrD+0];
 705
 706             /**************************
 707              * CALCULATE INTERACTIONS *
 708              **************************/
 709
 710             r00              = _mm_mul_ps(rsq00,rinv00);
 711             r00              = _mm_andnot_ps(dummy_mask,r00);
 712
 713             /* Compute parameters for interactions between i and j atoms */
 714             qq00             = _mm_mul_ps(iq0,jq0);
 715             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 716                                          vdwparam+vdwioffset0+vdwjidx0B,
 717                                          vdwparam+vdwioffset0+vdwjidx0C,
 718                                          vdwparam+vdwioffset0+vdwjidx0D,
 719                                          &c6_00,&c12_00);
 720
 721             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 722                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 723                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 724                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 725
 726             /* EWALD ELECTROSTATICS */
 727
 728             /* Analytical PME correction */
 729             zeta2            = _mm_mul_ps(beta2,rsq00);
 730             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 731             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 732             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 733             felec            = _mm_mul_ps(qq00,felec);
 734
 735             /* Analytical LJ-PME */
 736             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 737             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 738             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 739             exponent         = avx128fma_exp_f(ewcljrsq);
 740             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 741             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 742             /* f6A = 6 * C6grid * (1 - poly) */
 743             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 744             /* f6B = C6grid * exponent * beta^6 */
 745             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 746             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 747             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 748
 749             fscal            = _mm_add_ps(felec,fvdw);
 750
 751             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 752
 753              /* Update vectorial force */
 754             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 755             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 756             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 757
 758             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 759             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 760             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 761             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 762             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 763                                                    _mm_mul_ps(dx00,fscal),
 764                                                    _mm_mul_ps(dy00,fscal),
 765                                                    _mm_mul_ps(dz00,fscal));
 766
 767             /* Inner loop uses 50 flops */
 768         }
 769
 770         /* End of innermost loop */
 771
 772         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 773                                               f+i_coord_offset,fshift+i_shift_offset);
 774
 775         /* Increment number of inner iterations */
 776         inneriter                  += j_index_end - j_index_start;
 777
 778         /* Outer loop uses 7 flops */
 779     }
 780
 781     /* Increment number of outer iterations */
 782     outeriter        += nri;
 783
 784     /* Update outer/inner flops */
 785
 786     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*50);
 787 }