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, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            LJEwald
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEw_VdwLJEw_GeomP1P1_VF_avx_128_fma_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 refer to j loop unrolling done with AVX_128, e.g. for the four 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              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  95     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  96     __m128           c6grid_00;
  97     real             *vdwgridparam;
  98     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  99     __m128           one_half = _mm_set1_ps(0.5);
 100     __m128           minus_one = _mm_set1_ps(-1.0);
 101     __m128i          ewitab;
 102     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 103     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 104     real             *ewtab;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125     vdwgridparam     = fr->ljpme_c6grid;
 126     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 127     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 128     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 129
 130     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 131     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 132     beta2            = _mm_mul_ps(beta,beta);
 133     beta3            = _mm_mul_ps(beta,beta2);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 136     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 137
 138     /* Avoid stupid compiler warnings */
 139     jnrA = jnrB = jnrC = jnrD = 0;
 140     j_coord_offsetA = 0;
 141     j_coord_offsetB = 0;
 142     j_coord_offsetC = 0;
 143     j_coord_offsetD = 0;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     for(iidx=0;iidx<4*DIM;iidx++)
 149     {
 150         scratch[iidx] = 0.0;
 151     }
 152
 153     /* Start outer loop over neighborlists */
 154     for(iidx=0; iidx<nri; iidx++)
 155     {
 156         /* Load shift vector for this list */
 157         i_shift_offset   = DIM*shiftidx[iidx];
 158
 159         /* Load limits for loop over neighbors */
 160         j_index_start    = jindex[iidx];
 161         j_index_end      = jindex[iidx+1];
 162
 163         /* Get outer coordinate index */
 164         inr              = iinr[iidx];
 165         i_coord_offset   = DIM*inr;
 166
 167         /* Load i particle coords and add shift vector */
 168         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 169
 170         fix0             = _mm_setzero_ps();
 171         fiy0             = _mm_setzero_ps();
 172         fiz0             = _mm_setzero_ps();
 173
 174         /* Load parameters for i particles */
 175         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 176         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 177
 178         /* Reset potential sums */
 179         velecsum         = _mm_setzero_ps();
 180         vvdwsum          = _mm_setzero_ps();
 181
 182         /* Start inner kernel loop */
 183         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 184         {
 185
 186             /* Get j neighbor index, and coordinate index */
 187             jnrA             = jjnr[jidx];
 188             jnrB             = jjnr[jidx+1];
 189             jnrC             = jjnr[jidx+2];
 190             jnrD             = jjnr[jidx+3];
 191             j_coord_offsetA  = DIM*jnrA;
 192             j_coord_offsetB  = DIM*jnrB;
 193             j_coord_offsetC  = DIM*jnrC;
 194             j_coord_offsetD  = DIM*jnrD;
 195
 196             /* load j atom coordinates */
 197             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 198                                               x+j_coord_offsetC,x+j_coord_offsetD,
 199                                               &jx0,&jy0,&jz0);
 200
 201             /* Calculate displacement vector */
 202             dx00             = _mm_sub_ps(ix0,jx0);
 203             dy00             = _mm_sub_ps(iy0,jy0);
 204             dz00             = _mm_sub_ps(iz0,jz0);
 205
 206             /* Calculate squared distance and things based on it */
 207             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 208
 209             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 210
 211             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 212
 213             /* Load parameters for j particles */
 214             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 215                                                               charge+jnrC+0,charge+jnrD+0);
 216             vdwjidx0A        = 2*vdwtype[jnrA+0];
 217             vdwjidx0B        = 2*vdwtype[jnrB+0];
 218             vdwjidx0C        = 2*vdwtype[jnrC+0];
 219             vdwjidx0D        = 2*vdwtype[jnrD+0];
 220
 221             /**************************
 222              * CALCULATE INTERACTIONS *
 223              **************************/
 224
 225             r00              = _mm_mul_ps(rsq00,rinv00);
 226
 227             /* Compute parameters for interactions between i and j atoms */
 228             qq00             = _mm_mul_ps(iq0,jq0);
 229             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 230                                          vdwparam+vdwioffset0+vdwjidx0B,
 231                                          vdwparam+vdwioffset0+vdwjidx0C,
 232                                          vdwparam+vdwioffset0+vdwjidx0D,
 233                                          &c6_00,&c12_00);
 234
 235             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 236                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 237                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 238                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 239
 240             /* EWALD ELECTROSTATICS */
 241
 242             /* Analytical PME correction */
 243             zeta2            = _mm_mul_ps(beta2,rsq00);
 244             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 245             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 246             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 247             felec            = _mm_mul_ps(qq00,felec);
 248             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 249             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 250             velec            = _mm_mul_ps(qq00,velec);
 251
 252             /* Analytical LJ-PME */
 253             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 254             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 255             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 256             exponent         = gmx_simd_exp_r(ewcljrsq);
 257             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 258             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 259             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 260             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 261             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 262             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 263             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 264             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);
 265
 266             /* Update potential sum for this i atom from the interaction with this j atom. */
 267             velecsum         = _mm_add_ps(velecsum,velec);
 268             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 269
 270             fscal            = _mm_add_ps(felec,fvdw);
 271
 272              /* Update vectorial force */
 273             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 274             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 275             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 276
 277             fjptrA             = f+j_coord_offsetA;
 278             fjptrB             = f+j_coord_offsetB;
 279             fjptrC             = f+j_coord_offsetC;
 280             fjptrD             = f+j_coord_offsetD;
 281             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 282                                                    _mm_mul_ps(dx00,fscal),
 283                                                    _mm_mul_ps(dy00,fscal),
 284                                                    _mm_mul_ps(dz00,fscal));
 285
 286             /* Inner loop uses 53 flops */
 287         }
 288
 289         if(jidx<j_index_end)
 290         {
 291
 292             /* Get j neighbor index, and coordinate index */
 293             jnrlistA         = jjnr[jidx];
 294             jnrlistB         = jjnr[jidx+1];
 295             jnrlistC         = jjnr[jidx+2];
 296             jnrlistD         = jjnr[jidx+3];
 297             /* Sign of each element will be negative for non-real atoms.
 298              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 299              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 300              */
 301             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 302             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 303             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 304             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 305             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 306             j_coord_offsetA  = DIM*jnrA;
 307             j_coord_offsetB  = DIM*jnrB;
 308             j_coord_offsetC  = DIM*jnrC;
 309             j_coord_offsetD  = DIM*jnrD;
 310
 311             /* load j atom coordinates */
 312             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 313                                               x+j_coord_offsetC,x+j_coord_offsetD,
 314                                               &jx0,&jy0,&jz0);
 315
 316             /* Calculate displacement vector */
 317             dx00             = _mm_sub_ps(ix0,jx0);
 318             dy00             = _mm_sub_ps(iy0,jy0);
 319             dz00             = _mm_sub_ps(iz0,jz0);
 320
 321             /* Calculate squared distance and things based on it */
 322             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 323
 324             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 325
 326             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 327
 328             /* Load parameters for j particles */
 329             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 330                                                               charge+jnrC+0,charge+jnrD+0);
 331             vdwjidx0A        = 2*vdwtype[jnrA+0];
 332             vdwjidx0B        = 2*vdwtype[jnrB+0];
 333             vdwjidx0C        = 2*vdwtype[jnrC+0];
 334             vdwjidx0D        = 2*vdwtype[jnrD+0];
 335
 336             /**************************
 337              * CALCULATE INTERACTIONS *
 338              **************************/
 339
 340             r00              = _mm_mul_ps(rsq00,rinv00);
 341             r00              = _mm_andnot_ps(dummy_mask,r00);
 342
 343             /* Compute parameters for interactions between i and j atoms */
 344             qq00             = _mm_mul_ps(iq0,jq0);
 345             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 346                                          vdwparam+vdwioffset0+vdwjidx0B,
 347                                          vdwparam+vdwioffset0+vdwjidx0C,
 348                                          vdwparam+vdwioffset0+vdwjidx0D,
 349                                          &c6_00,&c12_00);
 350
 351             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 352                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 353                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 354                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 355
 356             /* EWALD ELECTROSTATICS */
 357
 358             /* Analytical PME correction */
 359             zeta2            = _mm_mul_ps(beta2,rsq00);
 360             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 361             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 362             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 363             felec            = _mm_mul_ps(qq00,felec);
 364             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 365             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 366             velec            = _mm_mul_ps(qq00,velec);
 367
 368             /* Analytical LJ-PME */
 369             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 370             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 371             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 372             exponent         = gmx_simd_exp_r(ewcljrsq);
 373             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 374             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 375             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 376             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 377             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 378             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 379             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 380             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);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velec            = _mm_andnot_ps(dummy_mask,velec);
 384             velecsum         = _mm_add_ps(velecsum,velec);
 385             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 386             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 387
 388             fscal            = _mm_add_ps(felec,fvdw);
 389
 390             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 391
 392              /* Update vectorial force */
 393             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 394             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 395             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 396
 397             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 398             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 399             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 400             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 401             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 402                                                    _mm_mul_ps(dx00,fscal),
 403                                                    _mm_mul_ps(dy00,fscal),
 404                                                    _mm_mul_ps(dz00,fscal));
 405
 406             /* Inner loop uses 54 flops */
 407         }
 408
 409         /* End of innermost loop */
 410
 411         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 412                                               f+i_coord_offset,fshift+i_shift_offset);
 413
 414         ggid                        = gid[iidx];
 415         /* Update potential energies */
 416         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 417         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 418
 419         /* Increment number of inner iterations */
 420         inneriter                  += j_index_end - j_index_start;
 421
 422         /* Outer loop uses 9 flops */
 423     }
 424
 425     /* Increment number of outer iterations */
 426     outeriter        += nri;
 427
 428     /* Update outer/inner flops */
 429
 430     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 431 }
 432 /*
 433  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 434  * Electrostatics interaction: Ewald
 435  * VdW interaction:            LJEwald
 436  * Geometry:                   Particle-Particle
 437  * Calculate force/pot:        Force
 438  */
 439 void
 440 nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 441                     (t_nblist                    * gmx_restrict       nlist,
 442                      rvec                        * gmx_restrict          xx,
 443                      rvec                        * gmx_restrict          ff,
 444                      t_forcerec                  * gmx_restrict          fr,
 445                      t_mdatoms                   * gmx_restrict     mdatoms,
 446                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 447                      t_nrnb                      * gmx_restrict        nrnb)
 448 {
 449     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 450      * just 0 for non-waters.
 451      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 452      * jnr indices corresponding to data put in the four positions in the SIMD register.
 453      */
 454     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 455     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 456     int              jnrA,jnrB,jnrC,jnrD;
 457     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 458     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 459     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 460     real             rcutoff_scalar;
 461     real             *shiftvec,*fshift,*x,*f;
 462     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 463     real             scratch[4*DIM];
 464     __m128           fscal,rcutoff,rcutoff2,jidxall;
 465     int              vdwioffset0;
 466     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 467     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 468     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 469     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 470     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 471     real             *charge;
 472     int              nvdwtype;
 473     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 474     int              *vdwtype;
 475     real             *vdwparam;
 476     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 477     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 478     __m128           c6grid_00;
 479     real             *vdwgridparam;
 480     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 481     __m128           one_half = _mm_set1_ps(0.5);
 482     __m128           minus_one = _mm_set1_ps(-1.0);
 483     __m128i          ewitab;
 484     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 485     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 486     real             *ewtab;
 487     __m128           dummy_mask,cutoff_mask;
 488     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 489     __m128           one     = _mm_set1_ps(1.0);
 490     __m128           two     = _mm_set1_ps(2.0);
 491     x                = xx[0];
 492     f                = ff[0];
 493
 494     nri              = nlist->nri;
 495     iinr             = nlist->iinr;
 496     jindex           = nlist->jindex;
 497     jjnr             = nlist->jjnr;
 498     shiftidx         = nlist->shift;
 499     gid              = nlist->gid;
 500     shiftvec         = fr->shift_vec[0];
 501     fshift           = fr->fshift[0];
 502     facel            = _mm_set1_ps(fr->epsfac);
 503     charge           = mdatoms->chargeA;
 504     nvdwtype         = fr->ntype;
 505     vdwparam         = fr->nbfp;
 506     vdwtype          = mdatoms->typeA;
 507     vdwgridparam     = fr->ljpme_c6grid;
 508     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 509     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 510     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 511
 512     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 513     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 514     beta2            = _mm_mul_ps(beta,beta);
 515     beta3            = _mm_mul_ps(beta,beta2);
 516     ewtab            = fr->ic->tabq_coul_F;
 517     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 518     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 519
 520     /* Avoid stupid compiler warnings */
 521     jnrA = jnrB = jnrC = jnrD = 0;
 522     j_coord_offsetA = 0;
 523     j_coord_offsetB = 0;
 524     j_coord_offsetC = 0;
 525     j_coord_offsetD = 0;
 526
 527     outeriter        = 0;
 528     inneriter        = 0;
 529
 530     for(iidx=0;iidx<4*DIM;iidx++)
 531     {
 532         scratch[iidx] = 0.0;
 533     }
 534
 535     /* Start outer loop over neighborlists */
 536     for(iidx=0; iidx<nri; iidx++)
 537     {
 538         /* Load shift vector for this list */
 539         i_shift_offset   = DIM*shiftidx[iidx];
 540
 541         /* Load limits for loop over neighbors */
 542         j_index_start    = jindex[iidx];
 543         j_index_end      = jindex[iidx+1];
 544
 545         /* Get outer coordinate index */
 546         inr              = iinr[iidx];
 547         i_coord_offset   = DIM*inr;
 548
 549         /* Load i particle coords and add shift vector */
 550         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 551
 552         fix0             = _mm_setzero_ps();
 553         fiy0             = _mm_setzero_ps();
 554         fiz0             = _mm_setzero_ps();
 555
 556         /* Load parameters for i particles */
 557         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 558         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 559
 560         /* Start inner kernel loop */
 561         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 562         {
 563
 564             /* Get j neighbor index, and coordinate index */
 565             jnrA             = jjnr[jidx];
 566             jnrB             = jjnr[jidx+1];
 567             jnrC             = jjnr[jidx+2];
 568             jnrD             = jjnr[jidx+3];
 569             j_coord_offsetA  = DIM*jnrA;
 570             j_coord_offsetB  = DIM*jnrB;
 571             j_coord_offsetC  = DIM*jnrC;
 572             j_coord_offsetD  = DIM*jnrD;
 573
 574             /* load j atom coordinates */
 575             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 576                                               x+j_coord_offsetC,x+j_coord_offsetD,
 577                                               &jx0,&jy0,&jz0);
 578
 579             /* Calculate displacement vector */
 580             dx00             = _mm_sub_ps(ix0,jx0);
 581             dy00             = _mm_sub_ps(iy0,jy0);
 582             dz00             = _mm_sub_ps(iz0,jz0);
 583
 584             /* Calculate squared distance and things based on it */
 585             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 586
 587             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 588
 589             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 590
 591             /* Load parameters for j particles */
 592             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 593                                                               charge+jnrC+0,charge+jnrD+0);
 594             vdwjidx0A        = 2*vdwtype[jnrA+0];
 595             vdwjidx0B        = 2*vdwtype[jnrB+0];
 596             vdwjidx0C        = 2*vdwtype[jnrC+0];
 597             vdwjidx0D        = 2*vdwtype[jnrD+0];
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             r00              = _mm_mul_ps(rsq00,rinv00);
 604
 605             /* Compute parameters for interactions between i and j atoms */
 606             qq00             = _mm_mul_ps(iq0,jq0);
 607             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 608                                          vdwparam+vdwioffset0+vdwjidx0B,
 609                                          vdwparam+vdwioffset0+vdwjidx0C,
 610                                          vdwparam+vdwioffset0+vdwjidx0D,
 611                                          &c6_00,&c12_00);
 612
 613             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 614                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 615                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 616                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 617
 618             /* EWALD ELECTROSTATICS */
 619
 620             /* Analytical PME correction */
 621             zeta2            = _mm_mul_ps(beta2,rsq00);
 622             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 623             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 624             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 625             felec            = _mm_mul_ps(qq00,felec);
 626
 627             /* Analytical LJ-PME */
 628             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 629             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 630             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 631             exponent         = gmx_simd_exp_r(ewcljrsq);
 632             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 633             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 634             /* f6A = 6 * C6grid * (1 - poly) */
 635             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 636             /* f6B = C6grid * exponent * beta^6 */
 637             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 638             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 639             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 640
 641             fscal            = _mm_add_ps(felec,fvdw);
 642
 643              /* Update vectorial force */
 644             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 645             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 646             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 647
 648             fjptrA             = f+j_coord_offsetA;
 649             fjptrB             = f+j_coord_offsetB;
 650             fjptrC             = f+j_coord_offsetC;
 651             fjptrD             = f+j_coord_offsetD;
 652             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 653                                                    _mm_mul_ps(dx00,fscal),
 654                                                    _mm_mul_ps(dy00,fscal),
 655                                                    _mm_mul_ps(dz00,fscal));
 656
 657             /* Inner loop uses 49 flops */
 658         }
 659
 660         if(jidx<j_index_end)
 661         {
 662
 663             /* Get j neighbor index, and coordinate index */
 664             jnrlistA         = jjnr[jidx];
 665             jnrlistB         = jjnr[jidx+1];
 666             jnrlistC         = jjnr[jidx+2];
 667             jnrlistD         = jjnr[jidx+3];
 668             /* Sign of each element will be negative for non-real atoms.
 669              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 670              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 671              */
 672             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 673             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 674             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 675             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 676             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 677             j_coord_offsetA  = DIM*jnrA;
 678             j_coord_offsetB  = DIM*jnrB;
 679             j_coord_offsetC  = DIM*jnrC;
 680             j_coord_offsetD  = DIM*jnrD;
 681
 682             /* load j atom coordinates */
 683             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 684                                               x+j_coord_offsetC,x+j_coord_offsetD,
 685                                               &jx0,&jy0,&jz0);
 686
 687             /* Calculate displacement vector */
 688             dx00             = _mm_sub_ps(ix0,jx0);
 689             dy00             = _mm_sub_ps(iy0,jy0);
 690             dz00             = _mm_sub_ps(iz0,jz0);
 691
 692             /* Calculate squared distance and things based on it */
 693             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 694
 695             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 696
 697             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 698
 699             /* Load parameters for j particles */
 700             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 701                                                               charge+jnrC+0,charge+jnrD+0);
 702             vdwjidx0A        = 2*vdwtype[jnrA+0];
 703             vdwjidx0B        = 2*vdwtype[jnrB+0];
 704             vdwjidx0C        = 2*vdwtype[jnrC+0];
 705             vdwjidx0D        = 2*vdwtype[jnrD+0];
 706
 707             /**************************
 708              * CALCULATE INTERACTIONS *
 709              **************************/
 710
 711             r00              = _mm_mul_ps(rsq00,rinv00);
 712             r00              = _mm_andnot_ps(dummy_mask,r00);
 713
 714             /* Compute parameters for interactions between i and j atoms */
 715             qq00             = _mm_mul_ps(iq0,jq0);
 716             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 717                                          vdwparam+vdwioffset0+vdwjidx0B,
 718                                          vdwparam+vdwioffset0+vdwjidx0C,
 719                                          vdwparam+vdwioffset0+vdwjidx0D,
 720                                          &c6_00,&c12_00);
 721
 722             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 723                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 724                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 725                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 726
 727             /* EWALD ELECTROSTATICS */
 728
 729             /* Analytical PME correction */
 730             zeta2            = _mm_mul_ps(beta2,rsq00);
 731             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 732             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 733             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 734             felec            = _mm_mul_ps(qq00,felec);
 735
 736             /* Analytical LJ-PME */
 737             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 738             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 739             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 740             exponent         = gmx_simd_exp_r(ewcljrsq);
 741             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 742             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 743             /* f6A = 6 * C6grid * (1 - poly) */
 744             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 745             /* f6B = C6grid * exponent * beta^6 */
 746             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 747             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 748             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 749
 750             fscal            = _mm_add_ps(felec,fvdw);
 751
 752             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 753
 754              /* Update vectorial force */
 755             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 756             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 757             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 758
 759             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 760             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 761             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 762             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 763             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 764                                                    _mm_mul_ps(dx00,fscal),
 765                                                    _mm_mul_ps(dy00,fscal),
 766                                                    _mm_mul_ps(dz00,fscal));
 767
 768             /* Inner loop uses 50 flops */
 769         }
 770
 771         /* End of innermost loop */
 772
 773         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 774                                               f+i_coord_offset,fshift+i_shift_offset);
 775
 776         /* Increment number of inner iterations */
 777         inneriter                  += j_index_end - j_index_start;
 778
 779         /* Outer loop uses 7 flops */
 780     }
 781
 782     /* Increment number of outer iterations */
 783     outeriter        += nri;
 784
 785     /* Update outer/inner flops */
 786
 787     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*50);
 788 }