src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwNone_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_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: GeneralizedBorn
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecGB_VdwNone_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     __m128i          gbitab;
  90     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  91     __m128           minushalf = _mm_set1_ps(-0.5);
  92     real             *invsqrta,*dvda,*gbtab;
  93     __m128i          vfitab;
  94     __m128i          ifour       = _mm_set1_epi32(4);
  95     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  96     real             *vftab;
  97     __m128           dummy_mask,cutoff_mask;
  98     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  99     __m128           one     = _mm_set1_ps(1.0);
 100     __m128           two     = _mm_set1_ps(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = _mm_set1_ps(fr->ic->epsfac);
 113     charge           = mdatoms->chargeA;
 114
 115     invsqrta         = fr->invsqrta;
 116     dvda             = fr->dvda;
 117     gbtabscale       = _mm_set1_ps(fr->gbtab->scale);
 118     gbtab            = fr->gbtab->data;
 119     gbinvepsdiff     = _mm_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 120
 121     /* Avoid stupid compiler warnings */
 122     jnrA = jnrB = jnrC = jnrD = 0;
 123     j_coord_offsetA = 0;
 124     j_coord_offsetB = 0;
 125     j_coord_offsetC = 0;
 126     j_coord_offsetD = 0;
 127
 128     outeriter        = 0;
 129     inneriter        = 0;
 130
 131     for(iidx=0;iidx<4*DIM;iidx++)
 132     {
 133         scratch[iidx] = 0.0;
 134     }
 135
 136     /* Start outer loop over neighborlists */
 137     for(iidx=0; iidx<nri; iidx++)
 138     {
 139         /* Load shift vector for this list */
 140         i_shift_offset   = DIM*shiftidx[iidx];
 141
 142         /* Load limits for loop over neighbors */
 143         j_index_start    = jindex[iidx];
 144         j_index_end      = jindex[iidx+1];
 145
 146         /* Get outer coordinate index */
 147         inr              = iinr[iidx];
 148         i_coord_offset   = DIM*inr;
 149
 150         /* Load i particle coords and add shift vector */
 151         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 152
 153         fix0             = _mm_setzero_ps();
 154         fiy0             = _mm_setzero_ps();
 155         fiz0             = _mm_setzero_ps();
 156
 157         /* Load parameters for i particles */
 158         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 159         isai0            = _mm_load1_ps(invsqrta+inr+0);
 160
 161         /* Reset potential sums */
 162         velecsum         = _mm_setzero_ps();
 163         vgbsum           = _mm_setzero_ps();
 164         dvdasum          = _mm_setzero_ps();
 165
 166         /* Start inner kernel loop */
 167         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 168         {
 169
 170             /* Get j neighbor index, and coordinate index */
 171             jnrA             = jjnr[jidx];
 172             jnrB             = jjnr[jidx+1];
 173             jnrC             = jjnr[jidx+2];
 174             jnrD             = jjnr[jidx+3];
 175             j_coord_offsetA  = DIM*jnrA;
 176             j_coord_offsetB  = DIM*jnrB;
 177             j_coord_offsetC  = DIM*jnrC;
 178             j_coord_offsetD  = DIM*jnrD;
 179
 180             /* load j atom coordinates */
 181             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 182                                               x+j_coord_offsetC,x+j_coord_offsetD,
 183                                               &jx0,&jy0,&jz0);
 184
 185             /* Calculate displacement vector */
 186             dx00             = _mm_sub_ps(ix0,jx0);
 187             dy00             = _mm_sub_ps(iy0,jy0);
 188             dz00             = _mm_sub_ps(iz0,jz0);
 189
 190             /* Calculate squared distance and things based on it */
 191             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 192
 193             rinv00           = avx128fma_invsqrt_f(rsq00);
 194
 195             /* Load parameters for j particles */
 196             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 197                                                               charge+jnrC+0,charge+jnrD+0);
 198             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 199                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 200
 201             /**************************
 202              * CALCULATE INTERACTIONS *
 203              **************************/
 204
 205             r00              = _mm_mul_ps(rsq00,rinv00);
 206
 207             /* Compute parameters for interactions between i and j atoms */
 208             qq00             = _mm_mul_ps(iq0,jq0);
 209
 210             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 211             isaprod          = _mm_mul_ps(isai0,isaj0);
 212             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 213             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 214
 215             /* Calculate generalized born table index - this is a separate table from the normal one,
 216              * but we use the same procedure by multiplying r with scale and truncating to integer.
 217              */
 218             rt               = _mm_mul_ps(r00,gbscale);
 219             gbitab           = _mm_cvttps_epi32(rt);
 220 #ifdef __XOP__
 221             gbeps            = _mm_frcz_ps(rt);
 222 #else
 223             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 224 #endif
 225             gbitab           = _mm_slli_epi32(gbitab,2);
 226
 227             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 228             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 229             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 230             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 231             _MM_TRANSPOSE4_PS(Y,F,G,H);
 232             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 233             VV               = _mm_macc_ps(gbeps,Fp,Y);
 234             vgb              = _mm_mul_ps(gbqqfactor,VV);
 235
 236             twogbeps         = _mm_add_ps(gbeps,gbeps);
 237             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 238             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 239             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 240             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 241             fjptrA           = dvda+jnrA;
 242             fjptrB           = dvda+jnrB;
 243             fjptrC           = dvda+jnrC;
 244             fjptrD           = dvda+jnrD;
 245             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 246             velec            = _mm_mul_ps(qq00,rinv00);
 247             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 248
 249             /* Update potential sum for this i atom from the interaction with this j atom. */
 250             velecsum         = _mm_add_ps(velecsum,velec);
 251             vgbsum           = _mm_add_ps(vgbsum,vgb);
 252
 253             fscal            = felec;
 254
 255              /* Update vectorial force */
 256             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 257             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 258             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 259
 260             fjptrA             = f+j_coord_offsetA;
 261             fjptrB             = f+j_coord_offsetB;
 262             fjptrC             = f+j_coord_offsetC;
 263             fjptrD             = f+j_coord_offsetD;
 264             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 265                                                    _mm_mul_ps(dx00,fscal),
 266                                                    _mm_mul_ps(dy00,fscal),
 267                                                    _mm_mul_ps(dz00,fscal));
 268
 269             /* Inner loop uses 61 flops */
 270         }
 271
 272         if(jidx<j_index_end)
 273         {
 274
 275             /* Get j neighbor index, and coordinate index */
 276             jnrlistA         = jjnr[jidx];
 277             jnrlistB         = jjnr[jidx+1];
 278             jnrlistC         = jjnr[jidx+2];
 279             jnrlistD         = jjnr[jidx+3];
 280             /* Sign of each element will be negative for non-real atoms.
 281              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 282              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 283              */
 284             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 285             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 286             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 287             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 288             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 289             j_coord_offsetA  = DIM*jnrA;
 290             j_coord_offsetB  = DIM*jnrB;
 291             j_coord_offsetC  = DIM*jnrC;
 292             j_coord_offsetD  = DIM*jnrD;
 293
 294             /* load j atom coordinates */
 295             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 296                                               x+j_coord_offsetC,x+j_coord_offsetD,
 297                                               &jx0,&jy0,&jz0);
 298
 299             /* Calculate displacement vector */
 300             dx00             = _mm_sub_ps(ix0,jx0);
 301             dy00             = _mm_sub_ps(iy0,jy0);
 302             dz00             = _mm_sub_ps(iz0,jz0);
 303
 304             /* Calculate squared distance and things based on it */
 305             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 306
 307             rinv00           = avx128fma_invsqrt_f(rsq00);
 308
 309             /* Load parameters for j particles */
 310             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 311                                                               charge+jnrC+0,charge+jnrD+0);
 312             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 313                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 314
 315             /**************************
 316              * CALCULATE INTERACTIONS *
 317              **************************/
 318
 319             r00              = _mm_mul_ps(rsq00,rinv00);
 320             r00              = _mm_andnot_ps(dummy_mask,r00);
 321
 322             /* Compute parameters for interactions between i and j atoms */
 323             qq00             = _mm_mul_ps(iq0,jq0);
 324
 325             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 326             isaprod          = _mm_mul_ps(isai0,isaj0);
 327             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 328             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 329
 330             /* Calculate generalized born table index - this is a separate table from the normal one,
 331              * but we use the same procedure by multiplying r with scale and truncating to integer.
 332              */
 333             rt               = _mm_mul_ps(r00,gbscale);
 334             gbitab           = _mm_cvttps_epi32(rt);
 335 #ifdef __XOP__
 336             gbeps            = _mm_frcz_ps(rt);
 337 #else
 338             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 339 #endif
 340             gbitab           = _mm_slli_epi32(gbitab,2);
 341
 342             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 343             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 344             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 345             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 346             _MM_TRANSPOSE4_PS(Y,F,G,H);
 347             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 348             VV               = _mm_macc_ps(gbeps,Fp,Y);
 349             vgb              = _mm_mul_ps(gbqqfactor,VV);
 350
 351             twogbeps         = _mm_add_ps(gbeps,gbeps);
 352             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 353             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 354             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 355             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 356             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 357             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 358             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 359             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 360             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 361             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 362             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 363             velec            = _mm_mul_ps(qq00,rinv00);
 364             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 365
 366             /* Update potential sum for this i atom from the interaction with this j atom. */
 367             velec            = _mm_andnot_ps(dummy_mask,velec);
 368             velecsum         = _mm_add_ps(velecsum,velec);
 369             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 370             vgbsum           = _mm_add_ps(vgbsum,vgb);
 371
 372             fscal            = felec;
 373
 374             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 375
 376              /* Update vectorial force */
 377             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 378             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 379             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 380
 381             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 382             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 383             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 384             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 385             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 386                                                    _mm_mul_ps(dx00,fscal),
 387                                                    _mm_mul_ps(dy00,fscal),
 388                                                    _mm_mul_ps(dz00,fscal));
 389
 390             /* Inner loop uses 62 flops */
 391         }
 392
 393         /* End of innermost loop */
 394
 395         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 396                                               f+i_coord_offset,fshift+i_shift_offset);
 397
 398         ggid                        = gid[iidx];
 399         /* Update potential energies */
 400         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 401         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 402         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 403         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 404
 405         /* Increment number of inner iterations */
 406         inneriter                  += j_index_end - j_index_start;
 407
 408         /* Outer loop uses 9 flops */
 409     }
 410
 411     /* Increment number of outer iterations */
 412     outeriter        += nri;
 413
 414     /* Update outer/inner flops */
 415
 416     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*62);
 417 }
 418 /*
 419  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 420  * Electrostatics interaction: GeneralizedBorn
 421  * VdW interaction:            None
 422  * Geometry:                   Particle-Particle
 423  * Calculate force/pot:        Force
 424  */
 425 void
 426 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 427                     (t_nblist                    * gmx_restrict       nlist,
 428                      rvec                        * gmx_restrict          xx,
 429                      rvec                        * gmx_restrict          ff,
 430                      struct t_forcerec           * gmx_restrict          fr,
 431                      t_mdatoms                   * gmx_restrict     mdatoms,
 432                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 433                      t_nrnb                      * gmx_restrict        nrnb)
 434 {
 435     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 436      * just 0 for non-waters.
 437      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 438      * jnr indices corresponding to data put in the four positions in the SIMD register.
 439      */
 440     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 441     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 442     int              jnrA,jnrB,jnrC,jnrD;
 443     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 444     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 445     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 446     real             rcutoff_scalar;
 447     real             *shiftvec,*fshift,*x,*f;
 448     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 449     real             scratch[4*DIM];
 450     __m128           fscal,rcutoff,rcutoff2,jidxall;
 451     int              vdwioffset0;
 452     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 453     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 454     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 455     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 456     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 457     real             *charge;
 458     __m128i          gbitab;
 459     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 460     __m128           minushalf = _mm_set1_ps(-0.5);
 461     real             *invsqrta,*dvda,*gbtab;
 462     __m128i          vfitab;
 463     __m128i          ifour       = _mm_set1_epi32(4);
 464     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 465     real             *vftab;
 466     __m128           dummy_mask,cutoff_mask;
 467     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 468     __m128           one     = _mm_set1_ps(1.0);
 469     __m128           two     = _mm_set1_ps(2.0);
 470     x                = xx[0];
 471     f                = ff[0];
 472
 473     nri              = nlist->nri;
 474     iinr             = nlist->iinr;
 475     jindex           = nlist->jindex;
 476     jjnr             = nlist->jjnr;
 477     shiftidx         = nlist->shift;
 478     gid              = nlist->gid;
 479     shiftvec         = fr->shift_vec[0];
 480     fshift           = fr->fshift[0];
 481     facel            = _mm_set1_ps(fr->ic->epsfac);
 482     charge           = mdatoms->chargeA;
 483
 484     invsqrta         = fr->invsqrta;
 485     dvda             = fr->dvda;
 486     gbtabscale       = _mm_set1_ps(fr->gbtab->scale);
 487     gbtab            = fr->gbtab->data;
 488     gbinvepsdiff     = _mm_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 489
 490     /* Avoid stupid compiler warnings */
 491     jnrA = jnrB = jnrC = jnrD = 0;
 492     j_coord_offsetA = 0;
 493     j_coord_offsetB = 0;
 494     j_coord_offsetC = 0;
 495     j_coord_offsetD = 0;
 496
 497     outeriter        = 0;
 498     inneriter        = 0;
 499
 500     for(iidx=0;iidx<4*DIM;iidx++)
 501     {
 502         scratch[iidx] = 0.0;
 503     }
 504
 505     /* Start outer loop over neighborlists */
 506     for(iidx=0; iidx<nri; iidx++)
 507     {
 508         /* Load shift vector for this list */
 509         i_shift_offset   = DIM*shiftidx[iidx];
 510
 511         /* Load limits for loop over neighbors */
 512         j_index_start    = jindex[iidx];
 513         j_index_end      = jindex[iidx+1];
 514
 515         /* Get outer coordinate index */
 516         inr              = iinr[iidx];
 517         i_coord_offset   = DIM*inr;
 518
 519         /* Load i particle coords and add shift vector */
 520         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 521
 522         fix0             = _mm_setzero_ps();
 523         fiy0             = _mm_setzero_ps();
 524         fiz0             = _mm_setzero_ps();
 525
 526         /* Load parameters for i particles */
 527         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 528         isai0            = _mm_load1_ps(invsqrta+inr+0);
 529
 530         dvdasum          = _mm_setzero_ps();
 531
 532         /* Start inner kernel loop */
 533         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 534         {
 535
 536             /* Get j neighbor index, and coordinate index */
 537             jnrA             = jjnr[jidx];
 538             jnrB             = jjnr[jidx+1];
 539             jnrC             = jjnr[jidx+2];
 540             jnrD             = jjnr[jidx+3];
 541             j_coord_offsetA  = DIM*jnrA;
 542             j_coord_offsetB  = DIM*jnrB;
 543             j_coord_offsetC  = DIM*jnrC;
 544             j_coord_offsetD  = DIM*jnrD;
 545
 546             /* load j atom coordinates */
 547             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 548                                               x+j_coord_offsetC,x+j_coord_offsetD,
 549                                               &jx0,&jy0,&jz0);
 550
 551             /* Calculate displacement vector */
 552             dx00             = _mm_sub_ps(ix0,jx0);
 553             dy00             = _mm_sub_ps(iy0,jy0);
 554             dz00             = _mm_sub_ps(iz0,jz0);
 555
 556             /* Calculate squared distance and things based on it */
 557             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 558
 559             rinv00           = avx128fma_invsqrt_f(rsq00);
 560
 561             /* Load parameters for j particles */
 562             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 563                                                               charge+jnrC+0,charge+jnrD+0);
 564             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 565                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 566
 567             /**************************
 568              * CALCULATE INTERACTIONS *
 569              **************************/
 570
 571             r00              = _mm_mul_ps(rsq00,rinv00);
 572
 573             /* Compute parameters for interactions between i and j atoms */
 574             qq00             = _mm_mul_ps(iq0,jq0);
 575
 576             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 577             isaprod          = _mm_mul_ps(isai0,isaj0);
 578             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 579             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 580
 581             /* Calculate generalized born table index - this is a separate table from the normal one,
 582              * but we use the same procedure by multiplying r with scale and truncating to integer.
 583              */
 584             rt               = _mm_mul_ps(r00,gbscale);
 585             gbitab           = _mm_cvttps_epi32(rt);
 586 #ifdef __XOP__
 587             gbeps            = _mm_frcz_ps(rt);
 588 #else
 589             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 590 #endif
 591             gbitab           = _mm_slli_epi32(gbitab,2);
 592
 593             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 594             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 595             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 596             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 597             _MM_TRANSPOSE4_PS(Y,F,G,H);
 598             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 599             VV               = _mm_macc_ps(gbeps,Fp,Y);
 600             vgb              = _mm_mul_ps(gbqqfactor,VV);
 601
 602             twogbeps         = _mm_add_ps(gbeps,gbeps);
 603             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 604             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 605             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 606             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 607             fjptrA           = dvda+jnrA;
 608             fjptrB           = dvda+jnrB;
 609             fjptrC           = dvda+jnrC;
 610             fjptrD           = dvda+jnrD;
 611             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 612             velec            = _mm_mul_ps(qq00,rinv00);
 613             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 614
 615             fscal            = felec;
 616
 617              /* Update vectorial force */
 618             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 619             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 620             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 621
 622             fjptrA             = f+j_coord_offsetA;
 623             fjptrB             = f+j_coord_offsetB;
 624             fjptrC             = f+j_coord_offsetC;
 625             fjptrD             = f+j_coord_offsetD;
 626             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 627                                                    _mm_mul_ps(dx00,fscal),
 628                                                    _mm_mul_ps(dy00,fscal),
 629                                                    _mm_mul_ps(dz00,fscal));
 630
 631             /* Inner loop uses 59 flops */
 632         }
 633
 634         if(jidx<j_index_end)
 635         {
 636
 637             /* Get j neighbor index, and coordinate index */
 638             jnrlistA         = jjnr[jidx];
 639             jnrlistB         = jjnr[jidx+1];
 640             jnrlistC         = jjnr[jidx+2];
 641             jnrlistD         = jjnr[jidx+3];
 642             /* Sign of each element will be negative for non-real atoms.
 643              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 644              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 645              */
 646             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 647             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 648             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 649             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 650             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 651             j_coord_offsetA  = DIM*jnrA;
 652             j_coord_offsetB  = DIM*jnrB;
 653             j_coord_offsetC  = DIM*jnrC;
 654             j_coord_offsetD  = DIM*jnrD;
 655
 656             /* load j atom coordinates */
 657             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 658                                               x+j_coord_offsetC,x+j_coord_offsetD,
 659                                               &jx0,&jy0,&jz0);
 660
 661             /* Calculate displacement vector */
 662             dx00             = _mm_sub_ps(ix0,jx0);
 663             dy00             = _mm_sub_ps(iy0,jy0);
 664             dz00             = _mm_sub_ps(iz0,jz0);
 665
 666             /* Calculate squared distance and things based on it */
 667             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 668
 669             rinv00           = avx128fma_invsqrt_f(rsq00);
 670
 671             /* Load parameters for j particles */
 672             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 673                                                               charge+jnrC+0,charge+jnrD+0);
 674             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 675                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             r00              = _mm_mul_ps(rsq00,rinv00);
 682             r00              = _mm_andnot_ps(dummy_mask,r00);
 683
 684             /* Compute parameters for interactions between i and j atoms */
 685             qq00             = _mm_mul_ps(iq0,jq0);
 686
 687             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 688             isaprod          = _mm_mul_ps(isai0,isaj0);
 689             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 690             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 691
 692             /* Calculate generalized born table index - this is a separate table from the normal one,
 693              * but we use the same procedure by multiplying r with scale and truncating to integer.
 694              */
 695             rt               = _mm_mul_ps(r00,gbscale);
 696             gbitab           = _mm_cvttps_epi32(rt);
 697 #ifdef __XOP__
 698             gbeps            = _mm_frcz_ps(rt);
 699 #else
 700             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 701 #endif
 702             gbitab           = _mm_slli_epi32(gbitab,2);
 703
 704             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 705             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 706             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 707             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 708             _MM_TRANSPOSE4_PS(Y,F,G,H);
 709             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 710             VV               = _mm_macc_ps(gbeps,Fp,Y);
 711             vgb              = _mm_mul_ps(gbqqfactor,VV);
 712
 713             twogbeps         = _mm_add_ps(gbeps,gbeps);
 714             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 715             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 716             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 717             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 718             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 719             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 720             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 721             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 722             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 723             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 724             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 725             velec            = _mm_mul_ps(qq00,rinv00);
 726             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 727
 728             fscal            = felec;
 729
 730             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 731
 732              /* Update vectorial force */
 733             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 734             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 735             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 736
 737             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 738             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 739             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 740             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 741             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 742                                                    _mm_mul_ps(dx00,fscal),
 743                                                    _mm_mul_ps(dy00,fscal),
 744                                                    _mm_mul_ps(dz00,fscal));
 745
 746             /* Inner loop uses 60 flops */
 747         }
 748
 749         /* End of innermost loop */
 750
 751         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 752                                               f+i_coord_offset,fshift+i_shift_offset);
 753
 754         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 755         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 756
 757         /* Increment number of inner iterations */
 758         inneriter                  += j_index_end - j_index_start;
 759
 760         /* Outer loop uses 7 flops */
 761     }
 762
 763     /* Increment number of outer iterations */
 764     outeriter        += nri;
 765
 766     /* Update outer/inner flops */
 767
 768     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*60);
 769 }