src/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, by the GROMACS development team, led by
   5  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   6  * others, as listed in the AUTHORS file in the top-level source
   7  * 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: GeneralizedBorn
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_single
  61                     (t_nblist * gmx_restrict                nlist,
  62                      rvec * gmx_restrict                    xx,
  63                      rvec * gmx_restrict                    ff,
  64                      t_forcerec * gmx_restrict              fr,
  65                      t_mdatoms * gmx_restrict               mdatoms,
  66                      nb_kernel_data_t * gmx_restrict        kernel_data,
  67                      t_nrnb * gmx_restrict                  nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128i          gbitab;
  93     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  94     __m128           minushalf = _mm_set1_ps(-0.5);
  95     real             *invsqrta,*dvda,*gbtab;
  96     __m128i          vfitab;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  99     real             *vftab;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117
 118     invsqrta         = fr->invsqrta;
 119     dvda             = fr->dvda;
 120     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 121     gbtab            = fr->gbtab.data;
 122     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 123
 124     /* Avoid stupid compiler warnings */
 125     jnrA = jnrB = jnrC = jnrD = 0;
 126     j_coord_offsetA = 0;
 127     j_coord_offsetB = 0;
 128     j_coord_offsetC = 0;
 129     j_coord_offsetD = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm_setzero_ps();
 157         fiy0             = _mm_setzero_ps();
 158         fiz0             = _mm_setzero_ps();
 159
 160         /* Load parameters for i particles */
 161         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 162         isai0            = _mm_load1_ps(invsqrta+inr+0);
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_ps();
 166         vgbsum           = _mm_setzero_ps();
 167         dvdasum          = _mm_setzero_ps();
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 171         {
 172
 173             /* Get j neighbor index, and coordinate index */
 174             jnrA             = jjnr[jidx];
 175             jnrB             = jjnr[jidx+1];
 176             jnrC             = jjnr[jidx+2];
 177             jnrD             = jjnr[jidx+3];
 178             j_coord_offsetA  = DIM*jnrA;
 179             j_coord_offsetB  = DIM*jnrB;
 180             j_coord_offsetC  = DIM*jnrC;
 181             j_coord_offsetD  = DIM*jnrD;
 182
 183             /* load j atom coordinates */
 184             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 185                                               x+j_coord_offsetC,x+j_coord_offsetD,
 186                                               &jx0,&jy0,&jz0);
 187
 188             /* Calculate displacement vector */
 189             dx00             = _mm_sub_ps(ix0,jx0);
 190             dy00             = _mm_sub_ps(iy0,jy0);
 191             dz00             = _mm_sub_ps(iz0,jz0);
 192
 193             /* Calculate squared distance and things based on it */
 194             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 195
 196             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 197
 198             /* Load parameters for j particles */
 199             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 200                                                               charge+jnrC+0,charge+jnrD+0);
 201             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 202                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 203
 204             /**************************
 205              * CALCULATE INTERACTIONS *
 206              **************************/
 207
 208             r00              = _mm_mul_ps(rsq00,rinv00);
 209
 210             /* Compute parameters for interactions between i and j atoms */
 211             qq00             = _mm_mul_ps(iq0,jq0);
 212
 213             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 214             isaprod          = _mm_mul_ps(isai0,isaj0);
 215             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 216             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 217
 218             /* Calculate generalized born table index - this is a separate table from the normal one,
 219              * but we use the same procedure by multiplying r with scale and truncating to integer.
 220              */
 221             rt               = _mm_mul_ps(r00,gbscale);
 222             gbitab           = _mm_cvttps_epi32(rt);
 223 #ifdef __XOP__
 224             gbeps            = _mm_frcz_ps(rt);
 225 #else
 226             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 227 #endif
 228             gbitab           = _mm_slli_epi32(gbitab,2);
 229
 230             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 231             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 232             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 233             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 234             _MM_TRANSPOSE4_PS(Y,F,G,H);
 235             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 236             VV               = _mm_macc_ps(gbeps,Fp,Y);
 237             vgb              = _mm_mul_ps(gbqqfactor,VV);
 238
 239             twogbeps         = _mm_add_ps(gbeps,gbeps);
 240             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 241             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 242             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 243             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 244             fjptrA           = dvda+jnrA;
 245             fjptrB           = dvda+jnrB;
 246             fjptrC           = dvda+jnrC;
 247             fjptrD           = dvda+jnrD;
 248             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 249             velec            = _mm_mul_ps(qq00,rinv00);
 250             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 251
 252             /* Update potential sum for this i atom from the interaction with this j atom. */
 253             velecsum         = _mm_add_ps(velecsum,velec);
 254             vgbsum           = _mm_add_ps(vgbsum,vgb);
 255
 256             fscal            = felec;
 257
 258              /* Update vectorial force */
 259             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 260             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 261             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 262
 263             fjptrA             = f+j_coord_offsetA;
 264             fjptrB             = f+j_coord_offsetB;
 265             fjptrC             = f+j_coord_offsetC;
 266             fjptrD             = f+j_coord_offsetD;
 267             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 268                                                    _mm_mul_ps(dx00,fscal),
 269                                                    _mm_mul_ps(dy00,fscal),
 270                                                    _mm_mul_ps(dz00,fscal));
 271
 272             /* Inner loop uses 61 flops */
 273         }
 274
 275         if(jidx<j_index_end)
 276         {
 277
 278             /* Get j neighbor index, and coordinate index */
 279             jnrlistA         = jjnr[jidx];
 280             jnrlistB         = jjnr[jidx+1];
 281             jnrlistC         = jjnr[jidx+2];
 282             jnrlistD         = jjnr[jidx+3];
 283             /* Sign of each element will be negative for non-real atoms.
 284              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 285              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 286              */
 287             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 288             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 289             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 290             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 291             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 292             j_coord_offsetA  = DIM*jnrA;
 293             j_coord_offsetB  = DIM*jnrB;
 294             j_coord_offsetC  = DIM*jnrC;
 295             j_coord_offsetD  = DIM*jnrD;
 296
 297             /* load j atom coordinates */
 298             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 299                                               x+j_coord_offsetC,x+j_coord_offsetD,
 300                                               &jx0,&jy0,&jz0);
 301
 302             /* Calculate displacement vector */
 303             dx00             = _mm_sub_ps(ix0,jx0);
 304             dy00             = _mm_sub_ps(iy0,jy0);
 305             dz00             = _mm_sub_ps(iz0,jz0);
 306
 307             /* Calculate squared distance and things based on it */
 308             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 309
 310             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 311
 312             /* Load parameters for j particles */
 313             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 314                                                               charge+jnrC+0,charge+jnrD+0);
 315             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 316                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 317
 318             /**************************
 319              * CALCULATE INTERACTIONS *
 320              **************************/
 321
 322             r00              = _mm_mul_ps(rsq00,rinv00);
 323             r00              = _mm_andnot_ps(dummy_mask,r00);
 324
 325             /* Compute parameters for interactions between i and j atoms */
 326             qq00             = _mm_mul_ps(iq0,jq0);
 327
 328             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 329             isaprod          = _mm_mul_ps(isai0,isaj0);
 330             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 331             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 332
 333             /* Calculate generalized born table index - this is a separate table from the normal one,
 334              * but we use the same procedure by multiplying r with scale and truncating to integer.
 335              */
 336             rt               = _mm_mul_ps(r00,gbscale);
 337             gbitab           = _mm_cvttps_epi32(rt);
 338 #ifdef __XOP__
 339             gbeps            = _mm_frcz_ps(rt);
 340 #else
 341             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 342 #endif
 343             gbitab           = _mm_slli_epi32(gbitab,2);
 344
 345             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 346             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 347             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 348             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 349             _MM_TRANSPOSE4_PS(Y,F,G,H);
 350             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 351             VV               = _mm_macc_ps(gbeps,Fp,Y);
 352             vgb              = _mm_mul_ps(gbqqfactor,VV);
 353
 354             twogbeps         = _mm_add_ps(gbeps,gbeps);
 355             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 356             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 357             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 358             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 359             /* 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. */
 360             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 361             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 362             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 363             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 364             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 365             velec            = _mm_mul_ps(qq00,rinv00);
 366             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velec            = _mm_andnot_ps(dummy_mask,velec);
 370             velecsum         = _mm_add_ps(velecsum,velec);
 371             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 372             vgbsum           = _mm_add_ps(vgbsum,vgb);
 373
 374             fscal            = felec;
 375
 376             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 377
 378              /* Update vectorial force */
 379             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 380             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 381             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 382
 383             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 384             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 385             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 386             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 387             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 388                                                    _mm_mul_ps(dx00,fscal),
 389                                                    _mm_mul_ps(dy00,fscal),
 390                                                    _mm_mul_ps(dz00,fscal));
 391
 392             /* Inner loop uses 62 flops */
 393         }
 394
 395         /* End of innermost loop */
 396
 397         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 398                                               f+i_coord_offset,fshift+i_shift_offset);
 399
 400         ggid                        = gid[iidx];
 401         /* Update potential energies */
 402         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 403         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 404         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 405         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 406
 407         /* Increment number of inner iterations */
 408         inneriter                  += j_index_end - j_index_start;
 409
 410         /* Outer loop uses 9 flops */
 411     }
 412
 413     /* Increment number of outer iterations */
 414     outeriter        += nri;
 415
 416     /* Update outer/inner flops */
 417
 418     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*62);
 419 }
 420 /*
 421  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 422  * Electrostatics interaction: GeneralizedBorn
 423  * VdW interaction:            None
 424  * Geometry:                   Particle-Particle
 425  * Calculate force/pot:        Force
 426  */
 427 void
 428 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 429                     (t_nblist * gmx_restrict                nlist,
 430                      rvec * gmx_restrict                    xx,
 431                      rvec * gmx_restrict                    ff,
 432                      t_forcerec * gmx_restrict              fr,
 433                      t_mdatoms * gmx_restrict               mdatoms,
 434                      nb_kernel_data_t * gmx_restrict        kernel_data,
 435                      t_nrnb * gmx_restrict                  nrnb)
 436 {
 437     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 438      * just 0 for non-waters.
 439      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 440      * jnr indices corresponding to data put in the four positions in the SIMD register.
 441      */
 442     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 443     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 444     int              jnrA,jnrB,jnrC,jnrD;
 445     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 446     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 447     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 448     real             rcutoff_scalar;
 449     real             *shiftvec,*fshift,*x,*f;
 450     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 451     real             scratch[4*DIM];
 452     __m128           fscal,rcutoff,rcutoff2,jidxall;
 453     int              vdwioffset0;
 454     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 455     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 456     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 457     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 458     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 459     real             *charge;
 460     __m128i          gbitab;
 461     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 462     __m128           minushalf = _mm_set1_ps(-0.5);
 463     real             *invsqrta,*dvda,*gbtab;
 464     __m128i          vfitab;
 465     __m128i          ifour       = _mm_set1_epi32(4);
 466     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 467     real             *vftab;
 468     __m128           dummy_mask,cutoff_mask;
 469     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 470     __m128           one     = _mm_set1_ps(1.0);
 471     __m128           two     = _mm_set1_ps(2.0);
 472     x                = xx[0];
 473     f                = ff[0];
 474
 475     nri              = nlist->nri;
 476     iinr             = nlist->iinr;
 477     jindex           = nlist->jindex;
 478     jjnr             = nlist->jjnr;
 479     shiftidx         = nlist->shift;
 480     gid              = nlist->gid;
 481     shiftvec         = fr->shift_vec[0];
 482     fshift           = fr->fshift[0];
 483     facel            = _mm_set1_ps(fr->epsfac);
 484     charge           = mdatoms->chargeA;
 485
 486     invsqrta         = fr->invsqrta;
 487     dvda             = fr->dvda;
 488     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 489     gbtab            = fr->gbtab.data;
 490     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 491
 492     /* Avoid stupid compiler warnings */
 493     jnrA = jnrB = jnrC = jnrD = 0;
 494     j_coord_offsetA = 0;
 495     j_coord_offsetB = 0;
 496     j_coord_offsetC = 0;
 497     j_coord_offsetD = 0;
 498
 499     outeriter        = 0;
 500     inneriter        = 0;
 501
 502     for(iidx=0;iidx<4*DIM;iidx++)
 503     {
 504         scratch[iidx] = 0.0;
 505     }
 506
 507     /* Start outer loop over neighborlists */
 508     for(iidx=0; iidx<nri; iidx++)
 509     {
 510         /* Load shift vector for this list */
 511         i_shift_offset   = DIM*shiftidx[iidx];
 512
 513         /* Load limits for loop over neighbors */
 514         j_index_start    = jindex[iidx];
 515         j_index_end      = jindex[iidx+1];
 516
 517         /* Get outer coordinate index */
 518         inr              = iinr[iidx];
 519         i_coord_offset   = DIM*inr;
 520
 521         /* Load i particle coords and add shift vector */
 522         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 523
 524         fix0             = _mm_setzero_ps();
 525         fiy0             = _mm_setzero_ps();
 526         fiz0             = _mm_setzero_ps();
 527
 528         /* Load parameters for i particles */
 529         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 530         isai0            = _mm_load1_ps(invsqrta+inr+0);
 531
 532         dvdasum          = _mm_setzero_ps();
 533
 534         /* Start inner kernel loop */
 535         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 536         {
 537
 538             /* Get j neighbor index, and coordinate index */
 539             jnrA             = jjnr[jidx];
 540             jnrB             = jjnr[jidx+1];
 541             jnrC             = jjnr[jidx+2];
 542             jnrD             = jjnr[jidx+3];
 543             j_coord_offsetA  = DIM*jnrA;
 544             j_coord_offsetB  = DIM*jnrB;
 545             j_coord_offsetC  = DIM*jnrC;
 546             j_coord_offsetD  = DIM*jnrD;
 547
 548             /* load j atom coordinates */
 549             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 550                                               x+j_coord_offsetC,x+j_coord_offsetD,
 551                                               &jx0,&jy0,&jz0);
 552
 553             /* Calculate displacement vector */
 554             dx00             = _mm_sub_ps(ix0,jx0);
 555             dy00             = _mm_sub_ps(iy0,jy0);
 556             dz00             = _mm_sub_ps(iz0,jz0);
 557
 558             /* Calculate squared distance and things based on it */
 559             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 560
 561             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 562
 563             /* Load parameters for j particles */
 564             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 565                                                               charge+jnrC+0,charge+jnrD+0);
 566             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 567                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 568
 569             /**************************
 570              * CALCULATE INTERACTIONS *
 571              **************************/
 572
 573             r00              = _mm_mul_ps(rsq00,rinv00);
 574
 575             /* Compute parameters for interactions between i and j atoms */
 576             qq00             = _mm_mul_ps(iq0,jq0);
 577
 578             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 579             isaprod          = _mm_mul_ps(isai0,isaj0);
 580             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 581             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 582
 583             /* Calculate generalized born table index - this is a separate table from the normal one,
 584              * but we use the same procedure by multiplying r with scale and truncating to integer.
 585              */
 586             rt               = _mm_mul_ps(r00,gbscale);
 587             gbitab           = _mm_cvttps_epi32(rt);
 588 #ifdef __XOP__
 589             gbeps            = _mm_frcz_ps(rt);
 590 #else
 591             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 592 #endif
 593             gbitab           = _mm_slli_epi32(gbitab,2);
 594
 595             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 596             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 597             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 598             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 599             _MM_TRANSPOSE4_PS(Y,F,G,H);
 600             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 601             VV               = _mm_macc_ps(gbeps,Fp,Y);
 602             vgb              = _mm_mul_ps(gbqqfactor,VV);
 603
 604             twogbeps         = _mm_add_ps(gbeps,gbeps);
 605             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 606             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 607             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 608             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 609             fjptrA           = dvda+jnrA;
 610             fjptrB           = dvda+jnrB;
 611             fjptrC           = dvda+jnrC;
 612             fjptrD           = dvda+jnrD;
 613             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 614             velec            = _mm_mul_ps(qq00,rinv00);
 615             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 616
 617             fscal            = felec;
 618
 619              /* Update vectorial force */
 620             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 621             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 622             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 623
 624             fjptrA             = f+j_coord_offsetA;
 625             fjptrB             = f+j_coord_offsetB;
 626             fjptrC             = f+j_coord_offsetC;
 627             fjptrD             = f+j_coord_offsetD;
 628             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 629                                                    _mm_mul_ps(dx00,fscal),
 630                                                    _mm_mul_ps(dy00,fscal),
 631                                                    _mm_mul_ps(dz00,fscal));
 632
 633             /* Inner loop uses 59 flops */
 634         }
 635
 636         if(jidx<j_index_end)
 637         {
 638
 639             /* Get j neighbor index, and coordinate index */
 640             jnrlistA         = jjnr[jidx];
 641             jnrlistB         = jjnr[jidx+1];
 642             jnrlistC         = jjnr[jidx+2];
 643             jnrlistD         = jjnr[jidx+3];
 644             /* Sign of each element will be negative for non-real atoms.
 645              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 646              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 647              */
 648             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 649             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 650             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 651             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 652             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 653             j_coord_offsetA  = DIM*jnrA;
 654             j_coord_offsetB  = DIM*jnrB;
 655             j_coord_offsetC  = DIM*jnrC;
 656             j_coord_offsetD  = DIM*jnrD;
 657
 658             /* load j atom coordinates */
 659             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 660                                               x+j_coord_offsetC,x+j_coord_offsetD,
 661                                               &jx0,&jy0,&jz0);
 662
 663             /* Calculate displacement vector */
 664             dx00             = _mm_sub_ps(ix0,jx0);
 665             dy00             = _mm_sub_ps(iy0,jy0);
 666             dz00             = _mm_sub_ps(iz0,jz0);
 667
 668             /* Calculate squared distance and things based on it */
 669             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 670
 671             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 672
 673             /* Load parameters for j particles */
 674             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 675                                                               charge+jnrC+0,charge+jnrD+0);
 676             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 677                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 678
 679             /**************************
 680              * CALCULATE INTERACTIONS *
 681              **************************/
 682
 683             r00              = _mm_mul_ps(rsq00,rinv00);
 684             r00              = _mm_andnot_ps(dummy_mask,r00);
 685
 686             /* Compute parameters for interactions between i and j atoms */
 687             qq00             = _mm_mul_ps(iq0,jq0);
 688
 689             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 690             isaprod          = _mm_mul_ps(isai0,isaj0);
 691             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 692             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 693
 694             /* Calculate generalized born table index - this is a separate table from the normal one,
 695              * but we use the same procedure by multiplying r with scale and truncating to integer.
 696              */
 697             rt               = _mm_mul_ps(r00,gbscale);
 698             gbitab           = _mm_cvttps_epi32(rt);
 699 #ifdef __XOP__
 700             gbeps            = _mm_frcz_ps(rt);
 701 #else
 702             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 703 #endif
 704             gbitab           = _mm_slli_epi32(gbitab,2);
 705
 706             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 707             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 708             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 709             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 710             _MM_TRANSPOSE4_PS(Y,F,G,H);
 711             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 712             VV               = _mm_macc_ps(gbeps,Fp,Y);
 713             vgb              = _mm_mul_ps(gbqqfactor,VV);
 714
 715             twogbeps         = _mm_add_ps(gbeps,gbeps);
 716             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 717             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 718             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 719             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 720             /* 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. */
 721             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 722             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 723             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 724             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 725             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 726             velec            = _mm_mul_ps(qq00,rinv00);
 727             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 728
 729             fscal            = felec;
 730
 731             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 732
 733              /* Update vectorial force */
 734             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 735             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 736             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 737
 738             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 739             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 740             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 741             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 742             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 743                                                    _mm_mul_ps(dx00,fscal),
 744                                                    _mm_mul_ps(dy00,fscal),
 745                                                    _mm_mul_ps(dz00,fscal));
 746
 747             /* Inner loop uses 60 flops */
 748         }
 749
 750         /* End of innermost loop */
 751
 752         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 753                                               f+i_coord_offset,fshift+i_shift_offset);
 754
 755         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 756         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 757
 758         /* Increment number of inner iterations */
 759         inneriter                  += j_index_end - j_index_start;
 760
 761         /* Outer loop uses 7 flops */
 762     }
 763
 764     /* Increment number of outer iterations */
 765     outeriter        += nri;
 766
 767     /* Update outer/inner flops */
 768
 769     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*60);
 770 }