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, 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 #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 "gromacs/simd/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_unused * 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             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 359             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 360             /* 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. */
 361             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 362             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 363             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 364             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 365             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 366             velec            = _mm_mul_ps(qq00,rinv00);
 367             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm_andnot_ps(dummy_mask,velec);
 371             velecsum         = _mm_add_ps(velecsum,velec);
 372             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 373             vgbsum           = _mm_add_ps(vgbsum,vgb);
 374
 375             fscal            = felec;
 376
 377             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 378
 379              /* Update vectorial force */
 380             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 381             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 382             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 383
 384             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 385             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 386             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 387             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 388             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 389                                                    _mm_mul_ps(dx00,fscal),
 390                                                    _mm_mul_ps(dy00,fscal),
 391                                                    _mm_mul_ps(dz00,fscal));
 392
 393             /* Inner loop uses 62 flops */
 394         }
 395
 396         /* End of innermost loop */
 397
 398         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 399                                               f+i_coord_offset,fshift+i_shift_offset);
 400
 401         ggid                        = gid[iidx];
 402         /* Update potential energies */
 403         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 404         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 405         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 406         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 407
 408         /* Increment number of inner iterations */
 409         inneriter                  += j_index_end - j_index_start;
 410
 411         /* Outer loop uses 9 flops */
 412     }
 413
 414     /* Increment number of outer iterations */
 415     outeriter        += nri;
 416
 417     /* Update outer/inner flops */
 418
 419     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*62);
 420 }
 421 /*
 422  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 423  * Electrostatics interaction: GeneralizedBorn
 424  * VdW interaction:            None
 425  * Geometry:                   Particle-Particle
 426  * Calculate force/pot:        Force
 427  */
 428 void
 429 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 430                     (t_nblist                    * gmx_restrict       nlist,
 431                      rvec                        * gmx_restrict          xx,
 432                      rvec                        * gmx_restrict          ff,
 433                      t_forcerec                  * gmx_restrict          fr,
 434                      t_mdatoms                   * gmx_restrict     mdatoms,
 435                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 436                      t_nrnb                      * gmx_restrict        nrnb)
 437 {
 438     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 439      * just 0 for non-waters.
 440      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 441      * jnr indices corresponding to data put in the four positions in the SIMD register.
 442      */
 443     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 444     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 445     int              jnrA,jnrB,jnrC,jnrD;
 446     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 447     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 448     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 449     real             rcutoff_scalar;
 450     real             *shiftvec,*fshift,*x,*f;
 451     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 452     real             scratch[4*DIM];
 453     __m128           fscal,rcutoff,rcutoff2,jidxall;
 454     int              vdwioffset0;
 455     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 456     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 457     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 458     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 459     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 460     real             *charge;
 461     __m128i          gbitab;
 462     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 463     __m128           minushalf = _mm_set1_ps(-0.5);
 464     real             *invsqrta,*dvda,*gbtab;
 465     __m128i          vfitab;
 466     __m128i          ifour       = _mm_set1_epi32(4);
 467     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 468     real             *vftab;
 469     __m128           dummy_mask,cutoff_mask;
 470     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 471     __m128           one     = _mm_set1_ps(1.0);
 472     __m128           two     = _mm_set1_ps(2.0);
 473     x                = xx[0];
 474     f                = ff[0];
 475
 476     nri              = nlist->nri;
 477     iinr             = nlist->iinr;
 478     jindex           = nlist->jindex;
 479     jjnr             = nlist->jjnr;
 480     shiftidx         = nlist->shift;
 481     gid              = nlist->gid;
 482     shiftvec         = fr->shift_vec[0];
 483     fshift           = fr->fshift[0];
 484     facel            = _mm_set1_ps(fr->epsfac);
 485     charge           = mdatoms->chargeA;
 486
 487     invsqrta         = fr->invsqrta;
 488     dvda             = fr->dvda;
 489     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 490     gbtab            = fr->gbtab.data;
 491     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 492
 493     /* Avoid stupid compiler warnings */
 494     jnrA = jnrB = jnrC = jnrD = 0;
 495     j_coord_offsetA = 0;
 496     j_coord_offsetB = 0;
 497     j_coord_offsetC = 0;
 498     j_coord_offsetD = 0;
 499
 500     outeriter        = 0;
 501     inneriter        = 0;
 502
 503     for(iidx=0;iidx<4*DIM;iidx++)
 504     {
 505         scratch[iidx] = 0.0;
 506     }
 507
 508     /* Start outer loop over neighborlists */
 509     for(iidx=0; iidx<nri; iidx++)
 510     {
 511         /* Load shift vector for this list */
 512         i_shift_offset   = DIM*shiftidx[iidx];
 513
 514         /* Load limits for loop over neighbors */
 515         j_index_start    = jindex[iidx];
 516         j_index_end      = jindex[iidx+1];
 517
 518         /* Get outer coordinate index */
 519         inr              = iinr[iidx];
 520         i_coord_offset   = DIM*inr;
 521
 522         /* Load i particle coords and add shift vector */
 523         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 524
 525         fix0             = _mm_setzero_ps();
 526         fiy0             = _mm_setzero_ps();
 527         fiz0             = _mm_setzero_ps();
 528
 529         /* Load parameters for i particles */
 530         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 531         isai0            = _mm_load1_ps(invsqrta+inr+0);
 532
 533         dvdasum          = _mm_setzero_ps();
 534
 535         /* Start inner kernel loop */
 536         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 537         {
 538
 539             /* Get j neighbor index, and coordinate index */
 540             jnrA             = jjnr[jidx];
 541             jnrB             = jjnr[jidx+1];
 542             jnrC             = jjnr[jidx+2];
 543             jnrD             = jjnr[jidx+3];
 544             j_coord_offsetA  = DIM*jnrA;
 545             j_coord_offsetB  = DIM*jnrB;
 546             j_coord_offsetC  = DIM*jnrC;
 547             j_coord_offsetD  = DIM*jnrD;
 548
 549             /* load j atom coordinates */
 550             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 551                                               x+j_coord_offsetC,x+j_coord_offsetD,
 552                                               &jx0,&jy0,&jz0);
 553
 554             /* Calculate displacement vector */
 555             dx00             = _mm_sub_ps(ix0,jx0);
 556             dy00             = _mm_sub_ps(iy0,jy0);
 557             dz00             = _mm_sub_ps(iz0,jz0);
 558
 559             /* Calculate squared distance and things based on it */
 560             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 561
 562             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 563
 564             /* Load parameters for j particles */
 565             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 566                                                               charge+jnrC+0,charge+jnrD+0);
 567             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 568                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             r00              = _mm_mul_ps(rsq00,rinv00);
 575
 576             /* Compute parameters for interactions between i and j atoms */
 577             qq00             = _mm_mul_ps(iq0,jq0);
 578
 579             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 580             isaprod          = _mm_mul_ps(isai0,isaj0);
 581             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 582             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 583
 584             /* Calculate generalized born table index - this is a separate table from the normal one,
 585              * but we use the same procedure by multiplying r with scale and truncating to integer.
 586              */
 587             rt               = _mm_mul_ps(r00,gbscale);
 588             gbitab           = _mm_cvttps_epi32(rt);
 589 #ifdef __XOP__
 590             gbeps            = _mm_frcz_ps(rt);
 591 #else
 592             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 593 #endif
 594             gbitab           = _mm_slli_epi32(gbitab,2);
 595
 596             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 597             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 598             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 599             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 600             _MM_TRANSPOSE4_PS(Y,F,G,H);
 601             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 602             VV               = _mm_macc_ps(gbeps,Fp,Y);
 603             vgb              = _mm_mul_ps(gbqqfactor,VV);
 604
 605             twogbeps         = _mm_add_ps(gbeps,gbeps);
 606             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 607             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 608             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 609             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 610             fjptrA           = dvda+jnrA;
 611             fjptrB           = dvda+jnrB;
 612             fjptrC           = dvda+jnrC;
 613             fjptrD           = dvda+jnrD;
 614             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 615             velec            = _mm_mul_ps(qq00,rinv00);
 616             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 617
 618             fscal            = felec;
 619
 620              /* Update vectorial force */
 621             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 622             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 623             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 624
 625             fjptrA             = f+j_coord_offsetA;
 626             fjptrB             = f+j_coord_offsetB;
 627             fjptrC             = f+j_coord_offsetC;
 628             fjptrD             = f+j_coord_offsetD;
 629             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 630                                                    _mm_mul_ps(dx00,fscal),
 631                                                    _mm_mul_ps(dy00,fscal),
 632                                                    _mm_mul_ps(dz00,fscal));
 633
 634             /* Inner loop uses 59 flops */
 635         }
 636
 637         if(jidx<j_index_end)
 638         {
 639
 640             /* Get j neighbor index, and coordinate index */
 641             jnrlistA         = jjnr[jidx];
 642             jnrlistB         = jjnr[jidx+1];
 643             jnrlistC         = jjnr[jidx+2];
 644             jnrlistD         = jjnr[jidx+3];
 645             /* Sign of each element will be negative for non-real atoms.
 646              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 647              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 648              */
 649             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 650             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 651             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 652             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 653             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 654             j_coord_offsetA  = DIM*jnrA;
 655             j_coord_offsetB  = DIM*jnrB;
 656             j_coord_offsetC  = DIM*jnrC;
 657             j_coord_offsetD  = DIM*jnrD;
 658
 659             /* load j atom coordinates */
 660             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 661                                               x+j_coord_offsetC,x+j_coord_offsetD,
 662                                               &jx0,&jy0,&jz0);
 663
 664             /* Calculate displacement vector */
 665             dx00             = _mm_sub_ps(ix0,jx0);
 666             dy00             = _mm_sub_ps(iy0,jy0);
 667             dz00             = _mm_sub_ps(iz0,jz0);
 668
 669             /* Calculate squared distance and things based on it */
 670             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 671
 672             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 673
 674             /* Load parameters for j particles */
 675             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 676                                                               charge+jnrC+0,charge+jnrD+0);
 677             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 678                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 679
 680             /**************************
 681              * CALCULATE INTERACTIONS *
 682              **************************/
 683
 684             r00              = _mm_mul_ps(rsq00,rinv00);
 685             r00              = _mm_andnot_ps(dummy_mask,r00);
 686
 687             /* Compute parameters for interactions between i and j atoms */
 688             qq00             = _mm_mul_ps(iq0,jq0);
 689
 690             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 691             isaprod          = _mm_mul_ps(isai0,isaj0);
 692             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 693             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 694
 695             /* Calculate generalized born table index - this is a separate table from the normal one,
 696              * but we use the same procedure by multiplying r with scale and truncating to integer.
 697              */
 698             rt               = _mm_mul_ps(r00,gbscale);
 699             gbitab           = _mm_cvttps_epi32(rt);
 700 #ifdef __XOP__
 701             gbeps            = _mm_frcz_ps(rt);
 702 #else
 703             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 704 #endif
 705             gbitab           = _mm_slli_epi32(gbitab,2);
 706
 707             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 708             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 709             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 710             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 711             _MM_TRANSPOSE4_PS(Y,F,G,H);
 712             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 713             VV               = _mm_macc_ps(gbeps,Fp,Y);
 714             vgb              = _mm_mul_ps(gbqqfactor,VV);
 715
 716             twogbeps         = _mm_add_ps(gbeps,gbeps);
 717             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 718             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 719             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 720             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 721             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 722             /* 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. */
 723             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 724             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 725             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 726             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 727             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 728             velec            = _mm_mul_ps(qq00,rinv00);
 729             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 734
 735              /* Update vectorial force */
 736             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 737             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 738             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 739
 740             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 741             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 742             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 743             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 744             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 745                                                    _mm_mul_ps(dx00,fscal),
 746                                                    _mm_mul_ps(dy00,fscal),
 747                                                    _mm_mul_ps(dz00,fscal));
 748
 749             /* Inner loop uses 60 flops */
 750         }
 751
 752         /* End of innermost loop */
 753
 754         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 755                                               f+i_coord_offset,fshift+i_shift_offset);
 756
 757         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 758         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 759
 760         /* Increment number of inner iterations */
 761         inneriter                  += j_index_end - j_index_start;
 762
 763         /* Outer loop uses 7 flops */
 764     }
 765
 766     /* Increment number of outer iterations */
 767     outeriter        += nri;
 768
 769     /* Update outer/inner flops */
 770
 771     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*60);
 772 }