src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecEw_VdwLJEw_GeomP1P1_sse4_1_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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 sse4_1_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_sse4_1_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomP1P1_VF_sse4_1_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_sse4_1_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 SSE, 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           tx,ty,tz,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     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  97     real             *vdwgridparam;
  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,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 102     real             *ewtab;
 103     __m128           dummy_mask,cutoff_mask;
 104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 105     __m128           one     = _mm_set1_ps(1.0);
 106     __m128           two     = _mm_set1_ps(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_ps(fr->ic->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123     vdwgridparam     = fr->ljpme_c6grid;
 124     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 125     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 126     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 127
 128     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 129     ewtab            = fr->ic->tabq_coul_FDV0;
 130     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 131     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 132
 133     /* Avoid stupid compiler warnings */
 134     jnrA = jnrB = jnrC = jnrD = 0;
 135     j_coord_offsetA = 0;
 136     j_coord_offsetB = 0;
 137     j_coord_offsetC = 0;
 138     j_coord_offsetD = 0;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     for(iidx=0;iidx<4*DIM;iidx++)
 144     {
 145         scratch[iidx] = 0.0;
 146     }
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 164
 165         fix0             = _mm_setzero_ps();
 166         fiy0             = _mm_setzero_ps();
 167         fiz0             = _mm_setzero_ps();
 168
 169         /* Load parameters for i particles */
 170         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 171         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 172
 173         /* Reset potential sums */
 174         velecsum         = _mm_setzero_ps();
 175         vvdwsum          = _mm_setzero_ps();
 176
 177         /* Start inner kernel loop */
 178         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 179         {
 180
 181             /* Get j neighbor index, and coordinate index */
 182             jnrA             = jjnr[jidx];
 183             jnrB             = jjnr[jidx+1];
 184             jnrC             = jjnr[jidx+2];
 185             jnrD             = jjnr[jidx+3];
 186             j_coord_offsetA  = DIM*jnrA;
 187             j_coord_offsetB  = DIM*jnrB;
 188             j_coord_offsetC  = DIM*jnrC;
 189             j_coord_offsetD  = DIM*jnrD;
 190
 191             /* load j atom coordinates */
 192             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 193                                               x+j_coord_offsetC,x+j_coord_offsetD,
 194                                               &jx0,&jy0,&jz0);
 195
 196             /* Calculate displacement vector */
 197             dx00             = _mm_sub_ps(ix0,jx0);
 198             dy00             = _mm_sub_ps(iy0,jy0);
 199             dz00             = _mm_sub_ps(iz0,jz0);
 200
 201             /* Calculate squared distance and things based on it */
 202             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 203
 204             rinv00           = sse41_invsqrt_f(rsq00);
 205
 206             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 207
 208             /* Load parameters for j particles */
 209             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 210                                                               charge+jnrC+0,charge+jnrD+0);
 211             vdwjidx0A        = 2*vdwtype[jnrA+0];
 212             vdwjidx0B        = 2*vdwtype[jnrB+0];
 213             vdwjidx0C        = 2*vdwtype[jnrC+0];
 214             vdwjidx0D        = 2*vdwtype[jnrD+0];
 215
 216             /**************************
 217              * CALCULATE INTERACTIONS *
 218              **************************/
 219
 220             r00              = _mm_mul_ps(rsq00,rinv00);
 221
 222             /* Compute parameters for interactions between i and j atoms */
 223             qq00             = _mm_mul_ps(iq0,jq0);
 224             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 225                                          vdwparam+vdwioffset0+vdwjidx0B,
 226                                          vdwparam+vdwioffset0+vdwjidx0C,
 227                                          vdwparam+vdwioffset0+vdwjidx0D,
 228                                          &c6_00,&c12_00);
 229
 230             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 231                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 232                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 233                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 234
 235             /* EWALD ELECTROSTATICS */
 236
 237             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 238             ewrt             = _mm_mul_ps(r00,ewtabscale);
 239             ewitab           = _mm_cvttps_epi32(ewrt);
 240             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 241             ewitab           = _mm_slli_epi32(ewitab,2);
 242             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 243             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 244             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 245             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 246             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 247             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 248             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 249             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 250             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 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         = sse41_exp_f(ewcljrsq);
 257             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 258             poly             = _mm_mul_ps(exponent,_mm_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 259             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 260             vvdw6            = _mm_mul_ps(_mm_sub_ps(c6_00,_mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly))),rinvsix);
 261             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 262             vvdw             = _mm_sub_ps(_mm_mul_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_sub_ps(vvdw12,_mm_sub_ps(vvdw6,_mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6)))),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             /* Calculate temporary vectorial force */
 273             tx               = _mm_mul_ps(fscal,dx00);
 274             ty               = _mm_mul_ps(fscal,dy00);
 275             tz               = _mm_mul_ps(fscal,dz00);
 276
 277             /* Update vectorial force */
 278             fix0             = _mm_add_ps(fix0,tx);
 279             fiy0             = _mm_add_ps(fiy0,ty);
 280             fiz0             = _mm_add_ps(fiz0,tz);
 281
 282             fjptrA             = f+j_coord_offsetA;
 283             fjptrB             = f+j_coord_offsetB;
 284             fjptrC             = f+j_coord_offsetC;
 285             fjptrD             = f+j_coord_offsetD;
 286             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 287
 288             /* Inner loop uses 69 flops */
 289         }
 290
 291         if(jidx<j_index_end)
 292         {
 293
 294             /* Get j neighbor index, and coordinate index */
 295             jnrlistA         = jjnr[jidx];
 296             jnrlistB         = jjnr[jidx+1];
 297             jnrlistC         = jjnr[jidx+2];
 298             jnrlistD         = jjnr[jidx+3];
 299             /* Sign of each element will be negative for non-real atoms.
 300              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 301              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 302              */
 303             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 304             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 305             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 306             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 307             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 308             j_coord_offsetA  = DIM*jnrA;
 309             j_coord_offsetB  = DIM*jnrB;
 310             j_coord_offsetC  = DIM*jnrC;
 311             j_coord_offsetD  = DIM*jnrD;
 312
 313             /* load j atom coordinates */
 314             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 315                                               x+j_coord_offsetC,x+j_coord_offsetD,
 316                                               &jx0,&jy0,&jz0);
 317
 318             /* Calculate displacement vector */
 319             dx00             = _mm_sub_ps(ix0,jx0);
 320             dy00             = _mm_sub_ps(iy0,jy0);
 321             dz00             = _mm_sub_ps(iz0,jz0);
 322
 323             /* Calculate squared distance and things based on it */
 324             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 325
 326             rinv00           = sse41_invsqrt_f(rsq00);
 327
 328             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 329
 330             /* Load parameters for j particles */
 331             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 332                                                               charge+jnrC+0,charge+jnrD+0);
 333             vdwjidx0A        = 2*vdwtype[jnrA+0];
 334             vdwjidx0B        = 2*vdwtype[jnrB+0];
 335             vdwjidx0C        = 2*vdwtype[jnrC+0];
 336             vdwjidx0D        = 2*vdwtype[jnrD+0];
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             r00              = _mm_mul_ps(rsq00,rinv00);
 343             r00              = _mm_andnot_ps(dummy_mask,r00);
 344
 345             /* Compute parameters for interactions between i and j atoms */
 346             qq00             = _mm_mul_ps(iq0,jq0);
 347             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 348                                          vdwparam+vdwioffset0+vdwjidx0B,
 349                                          vdwparam+vdwioffset0+vdwjidx0C,
 350                                          vdwparam+vdwioffset0+vdwjidx0D,
 351                                          &c6_00,&c12_00);
 352
 353             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 354                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 355                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 356                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 357
 358             /* EWALD ELECTROSTATICS */
 359
 360             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 361             ewrt             = _mm_mul_ps(r00,ewtabscale);
 362             ewitab           = _mm_cvttps_epi32(ewrt);
 363             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 364             ewitab           = _mm_slli_epi32(ewitab,2);
 365             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 366             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 367             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 368             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 369             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 370             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 371             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 372             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 373             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 374
 375             /* Analytical LJ-PME */
 376             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 377             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 378             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 379             exponent         = sse41_exp_f(ewcljrsq);
 380             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 381             poly             = _mm_mul_ps(exponent,_mm_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 382             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 383             vvdw6            = _mm_mul_ps(_mm_sub_ps(c6_00,_mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly))),rinvsix);
 384             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 385             vvdw             = _mm_sub_ps(_mm_mul_ps(vvdw12,one_twelfth),_mm_mul_ps(vvdw6,one_sixth));
 386             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 387             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,_mm_sub_ps(vvdw6,_mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6)))),rinvsq00);
 388
 389             /* Update potential sum for this i atom from the interaction with this j atom. */
 390             velec            = _mm_andnot_ps(dummy_mask,velec);
 391             velecsum         = _mm_add_ps(velecsum,velec);
 392             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 393             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 394
 395             fscal            = _mm_add_ps(felec,fvdw);
 396
 397             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 398
 399             /* Calculate temporary vectorial force */
 400             tx               = _mm_mul_ps(fscal,dx00);
 401             ty               = _mm_mul_ps(fscal,dy00);
 402             tz               = _mm_mul_ps(fscal,dz00);
 403
 404             /* Update vectorial force */
 405             fix0             = _mm_add_ps(fix0,tx);
 406             fiy0             = _mm_add_ps(fiy0,ty);
 407             fiz0             = _mm_add_ps(fiz0,tz);
 408
 409             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 410             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 411             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 412             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 413             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 414
 415             /* Inner loop uses 70 flops */
 416         }
 417
 418         /* End of innermost loop */
 419
 420         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 421                                               f+i_coord_offset,fshift+i_shift_offset);
 422
 423         ggid                        = gid[iidx];
 424         /* Update potential energies */
 425         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 426         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 427
 428         /* Increment number of inner iterations */
 429         inneriter                  += j_index_end - j_index_start;
 430
 431         /* Outer loop uses 9 flops */
 432     }
 433
 434     /* Increment number of outer iterations */
 435     outeriter        += nri;
 436
 437     /* Update outer/inner flops */
 438
 439     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*70);
 440 }
 441 /*
 442  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_sse4_1_single
 443  * Electrostatics interaction: Ewald
 444  * VdW interaction:            LJEwald
 445  * Geometry:                   Particle-Particle
 446  * Calculate force/pot:        Force
 447  */
 448 void
 449 nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_sse4_1_single
 450                     (t_nblist                    * gmx_restrict       nlist,
 451                      rvec                        * gmx_restrict          xx,
 452                      rvec                        * gmx_restrict          ff,
 453                      struct t_forcerec           * gmx_restrict          fr,
 454                      t_mdatoms                   * gmx_restrict     mdatoms,
 455                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 456                      t_nrnb                      * gmx_restrict        nrnb)
 457 {
 458     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 459      * just 0 for non-waters.
 460      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 461      * jnr indices corresponding to data put in the four positions in the SIMD register.
 462      */
 463     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 464     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 465     int              jnrA,jnrB,jnrC,jnrD;
 466     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 467     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 468     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 469     real             rcutoff_scalar;
 470     real             *shiftvec,*fshift,*x,*f;
 471     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 472     real             scratch[4*DIM];
 473     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 474     int              vdwioffset0;
 475     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 476     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 477     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 478     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 479     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 480     real             *charge;
 481     int              nvdwtype;
 482     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 483     int              *vdwtype;
 484     real             *vdwparam;
 485     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 486     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 487     __m128           c6grid_00;
 488     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 489     real             *vdwgridparam;
 490     __m128           one_half  = _mm_set1_ps(0.5);
 491     __m128           minus_one = _mm_set1_ps(-1.0);
 492     __m128i          ewitab;
 493     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 494     real             *ewtab;
 495     __m128           dummy_mask,cutoff_mask;
 496     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 497     __m128           one     = _mm_set1_ps(1.0);
 498     __m128           two     = _mm_set1_ps(2.0);
 499     x                = xx[0];
 500     f                = ff[0];
 501
 502     nri              = nlist->nri;
 503     iinr             = nlist->iinr;
 504     jindex           = nlist->jindex;
 505     jjnr             = nlist->jjnr;
 506     shiftidx         = nlist->shift;
 507     gid              = nlist->gid;
 508     shiftvec         = fr->shift_vec[0];
 509     fshift           = fr->fshift[0];
 510     facel            = _mm_set1_ps(fr->ic->epsfac);
 511     charge           = mdatoms->chargeA;
 512     nvdwtype         = fr->ntype;
 513     vdwparam         = fr->nbfp;
 514     vdwtype          = mdatoms->typeA;
 515     vdwgridparam     = fr->ljpme_c6grid;
 516     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 517     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 518     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 519
 520     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 521     ewtab            = fr->ic->tabq_coul_F;
 522     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 523     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 524
 525     /* Avoid stupid compiler warnings */
 526     jnrA = jnrB = jnrC = jnrD = 0;
 527     j_coord_offsetA = 0;
 528     j_coord_offsetB = 0;
 529     j_coord_offsetC = 0;
 530     j_coord_offsetD = 0;
 531
 532     outeriter        = 0;
 533     inneriter        = 0;
 534
 535     for(iidx=0;iidx<4*DIM;iidx++)
 536     {
 537         scratch[iidx] = 0.0;
 538     }
 539
 540     /* Start outer loop over neighborlists */
 541     for(iidx=0; iidx<nri; iidx++)
 542     {
 543         /* Load shift vector for this list */
 544         i_shift_offset   = DIM*shiftidx[iidx];
 545
 546         /* Load limits for loop over neighbors */
 547         j_index_start    = jindex[iidx];
 548         j_index_end      = jindex[iidx+1];
 549
 550         /* Get outer coordinate index */
 551         inr              = iinr[iidx];
 552         i_coord_offset   = DIM*inr;
 553
 554         /* Load i particle coords and add shift vector */
 555         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 556
 557         fix0             = _mm_setzero_ps();
 558         fiy0             = _mm_setzero_ps();
 559         fiz0             = _mm_setzero_ps();
 560
 561         /* Load parameters for i particles */
 562         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 563         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 564
 565         /* Start inner kernel loop */
 566         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 567         {
 568
 569             /* Get j neighbor index, and coordinate index */
 570             jnrA             = jjnr[jidx];
 571             jnrB             = jjnr[jidx+1];
 572             jnrC             = jjnr[jidx+2];
 573             jnrD             = jjnr[jidx+3];
 574             j_coord_offsetA  = DIM*jnrA;
 575             j_coord_offsetB  = DIM*jnrB;
 576             j_coord_offsetC  = DIM*jnrC;
 577             j_coord_offsetD  = DIM*jnrD;
 578
 579             /* load j atom coordinates */
 580             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 581                                               x+j_coord_offsetC,x+j_coord_offsetD,
 582                                               &jx0,&jy0,&jz0);
 583
 584             /* Calculate displacement vector */
 585             dx00             = _mm_sub_ps(ix0,jx0);
 586             dy00             = _mm_sub_ps(iy0,jy0);
 587             dz00             = _mm_sub_ps(iz0,jz0);
 588
 589             /* Calculate squared distance and things based on it */
 590             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 591
 592             rinv00           = sse41_invsqrt_f(rsq00);
 593
 594             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 595
 596             /* Load parameters for j particles */
 597             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 598                                                               charge+jnrC+0,charge+jnrD+0);
 599             vdwjidx0A        = 2*vdwtype[jnrA+0];
 600             vdwjidx0B        = 2*vdwtype[jnrB+0];
 601             vdwjidx0C        = 2*vdwtype[jnrC+0];
 602             vdwjidx0D        = 2*vdwtype[jnrD+0];
 603
 604             /**************************
 605              * CALCULATE INTERACTIONS *
 606              **************************/
 607
 608             r00              = _mm_mul_ps(rsq00,rinv00);
 609
 610             /* Compute parameters for interactions between i and j atoms */
 611             qq00             = _mm_mul_ps(iq0,jq0);
 612             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 613                                          vdwparam+vdwioffset0+vdwjidx0B,
 614                                          vdwparam+vdwioffset0+vdwjidx0C,
 615                                          vdwparam+vdwioffset0+vdwjidx0D,
 616                                          &c6_00,&c12_00);
 617
 618             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 619                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 620                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 621                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 622
 623             /* EWALD ELECTROSTATICS */
 624
 625             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 626             ewrt             = _mm_mul_ps(r00,ewtabscale);
 627             ewitab           = _mm_cvttps_epi32(ewrt);
 628             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 629             gmx_mm_load_4pair_swizzle_ps(ewtab + gmx_mm_extract_epi32(ewitab,0),ewtab + gmx_mm_extract_epi32(ewitab,1),
 630                                          ewtab + gmx_mm_extract_epi32(ewitab,2),ewtab + gmx_mm_extract_epi32(ewitab,3),
 631                                          &ewtabF,&ewtabFn);
 632             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
 633             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 634
 635             /* Analytical LJ-PME */
 636             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 637             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 638             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 639             exponent         = sse41_exp_f(ewcljrsq);
 640             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 641             poly             = _mm_mul_ps(exponent,_mm_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 642             /* f6A = 6 * C6grid * (1 - poly) */
 643             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 644             /* f6B = C6grid * exponent * beta^6 */
 645             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 646             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 647             fvdw              = _mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),_mm_sub_ps(c6_00,f6A)),rinvsix),f6B),rinvsq00);
 648
 649             fscal            = _mm_add_ps(felec,fvdw);
 650
 651             /* Calculate temporary vectorial force */
 652             tx               = _mm_mul_ps(fscal,dx00);
 653             ty               = _mm_mul_ps(fscal,dy00);
 654             tz               = _mm_mul_ps(fscal,dz00);
 655
 656             /* Update vectorial force */
 657             fix0             = _mm_add_ps(fix0,tx);
 658             fiy0             = _mm_add_ps(fiy0,ty);
 659             fiz0             = _mm_add_ps(fiz0,tz);
 660
 661             fjptrA             = f+j_coord_offsetA;
 662             fjptrB             = f+j_coord_offsetB;
 663             fjptrC             = f+j_coord_offsetC;
 664             fjptrD             = f+j_coord_offsetD;
 665             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 666
 667             /* Inner loop uses 59 flops */
 668         }
 669
 670         if(jidx<j_index_end)
 671         {
 672
 673             /* Get j neighbor index, and coordinate index */
 674             jnrlistA         = jjnr[jidx];
 675             jnrlistB         = jjnr[jidx+1];
 676             jnrlistC         = jjnr[jidx+2];
 677             jnrlistD         = jjnr[jidx+3];
 678             /* Sign of each element will be negative for non-real atoms.
 679              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 680              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 681              */
 682             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 683             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 684             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 685             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 686             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 687             j_coord_offsetA  = DIM*jnrA;
 688             j_coord_offsetB  = DIM*jnrB;
 689             j_coord_offsetC  = DIM*jnrC;
 690             j_coord_offsetD  = DIM*jnrD;
 691
 692             /* load j atom coordinates */
 693             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 694                                               x+j_coord_offsetC,x+j_coord_offsetD,
 695                                               &jx0,&jy0,&jz0);
 696
 697             /* Calculate displacement vector */
 698             dx00             = _mm_sub_ps(ix0,jx0);
 699             dy00             = _mm_sub_ps(iy0,jy0);
 700             dz00             = _mm_sub_ps(iz0,jz0);
 701
 702             /* Calculate squared distance and things based on it */
 703             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 704
 705             rinv00           = sse41_invsqrt_f(rsq00);
 706
 707             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 708
 709             /* Load parameters for j particles */
 710             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 711                                                               charge+jnrC+0,charge+jnrD+0);
 712             vdwjidx0A        = 2*vdwtype[jnrA+0];
 713             vdwjidx0B        = 2*vdwtype[jnrB+0];
 714             vdwjidx0C        = 2*vdwtype[jnrC+0];
 715             vdwjidx0D        = 2*vdwtype[jnrD+0];
 716
 717             /**************************
 718              * CALCULATE INTERACTIONS *
 719              **************************/
 720
 721             r00              = _mm_mul_ps(rsq00,rinv00);
 722             r00              = _mm_andnot_ps(dummy_mask,r00);
 723
 724             /* Compute parameters for interactions between i and j atoms */
 725             qq00             = _mm_mul_ps(iq0,jq0);
 726             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 727                                          vdwparam+vdwioffset0+vdwjidx0B,
 728                                          vdwparam+vdwioffset0+vdwjidx0C,
 729                                          vdwparam+vdwioffset0+vdwjidx0D,
 730                                          &c6_00,&c12_00);
 731
 732             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 733                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 734                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 735                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 736
 737             /* EWALD ELECTROSTATICS */
 738
 739             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 740             ewrt             = _mm_mul_ps(r00,ewtabscale);
 741             ewitab           = _mm_cvttps_epi32(ewrt);
 742             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 743             gmx_mm_load_4pair_swizzle_ps(ewtab + gmx_mm_extract_epi32(ewitab,0),ewtab + gmx_mm_extract_epi32(ewitab,1),
 744                                          ewtab + gmx_mm_extract_epi32(ewitab,2),ewtab + gmx_mm_extract_epi32(ewitab,3),
 745                                          &ewtabF,&ewtabFn);
 746             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
 747             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 748
 749             /* Analytical LJ-PME */
 750             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 751             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 752             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 753             exponent         = sse41_exp_f(ewcljrsq);
 754             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 755             poly             = _mm_mul_ps(exponent,_mm_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 756             /* f6A = 6 * C6grid * (1 - poly) */
 757             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 758             /* f6B = C6grid * exponent * beta^6 */
 759             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 760             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 761             fvdw              = _mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),_mm_sub_ps(c6_00,f6A)),rinvsix),f6B),rinvsq00);
 762
 763             fscal            = _mm_add_ps(felec,fvdw);
 764
 765             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 766
 767             /* Calculate temporary vectorial force */
 768             tx               = _mm_mul_ps(fscal,dx00);
 769             ty               = _mm_mul_ps(fscal,dy00);
 770             tz               = _mm_mul_ps(fscal,dz00);
 771
 772             /* Update vectorial force */
 773             fix0             = _mm_add_ps(fix0,tx);
 774             fiy0             = _mm_add_ps(fiy0,ty);
 775             fiz0             = _mm_add_ps(fiz0,tz);
 776
 777             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 778             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 779             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 780             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 781             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 782
 783             /* Inner loop uses 60 flops */
 784         }
 785
 786         /* End of innermost loop */
 787
 788         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 789                                               f+i_coord_offset,fshift+i_shift_offset);
 790
 791         /* Increment number of inner iterations */
 792         inneriter                  += j_index_end - j_index_start;
 793
 794         /* Outer loop uses 7 flops */
 795     }
 796
 797     /* Increment number of outer iterations */
 798     outeriter        += nri;
 799
 800     /* Update outer/inner flops */
 801
 802     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*60);
 803 }