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, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: GeneralizedBorn
  53  * VdW interaction:            None
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     __m128i          gbitab;
  91     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  92     __m128           minushalf = _mm_set1_ps(-0.5);
  93     real             *invsqrta,*dvda,*gbtab;
  94     __m128i          vfitab;
  95     __m128i          ifour       = _mm_set1_epi32(4);
  96     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  97     real             *vftab;
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm_set1_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115
 116     invsqrta         = fr->invsqrta;
 117     dvda             = fr->dvda;
 118     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 119     gbtab            = fr->gbtab.data;
 120     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm_setzero_ps();
 155         fiy0             = _mm_setzero_ps();
 156         fiz0             = _mm_setzero_ps();
 157
 158         /* Load parameters for i particles */
 159         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 160         isai0            = _mm_load1_ps(invsqrta+inr+0);
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm_setzero_ps();
 164         vgbsum           = _mm_setzero_ps();
 165         dvdasum          = _mm_setzero_ps();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178             j_coord_offsetC  = DIM*jnrC;
 179             j_coord_offsetD  = DIM*jnrD;
 180
 181             /* load j atom coordinates */
 182             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 183                                               x+j_coord_offsetC,x+j_coord_offsetD,
 184                                               &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm_sub_ps(ix0,jx0);
 188             dy00             = _mm_sub_ps(iy0,jy0);
 189             dz00             = _mm_sub_ps(iz0,jz0);
 190
 191             /* Calculate squared distance and things based on it */
 192             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 193
 194             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 195
 196             /* Load parameters for j particles */
 197             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 198                                                               charge+jnrC+0,charge+jnrD+0);
 199             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 200                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 201
 202             /**************************
 203              * CALCULATE INTERACTIONS *
 204              **************************/
 205
 206             r00              = _mm_mul_ps(rsq00,rinv00);
 207
 208             /* Compute parameters for interactions between i and j atoms */
 209             qq00             = _mm_mul_ps(iq0,jq0);
 210
 211             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 212             isaprod          = _mm_mul_ps(isai0,isaj0);
 213             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 214             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 215
 216             /* Calculate generalized born table index - this is a separate table from the normal one,
 217              * but we use the same procedure by multiplying r with scale and truncating to integer.
 218              */
 219             rt               = _mm_mul_ps(r00,gbscale);
 220             gbitab           = _mm_cvttps_epi32(rt);
 221 #ifdef __XOP__
 222             gbeps            = _mm_frcz_ps(rt);
 223 #else
 224             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 225 #endif
 226             gbitab           = _mm_slli_epi32(gbitab,2);
 227
 228             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 229             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 230             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 231             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 232             _MM_TRANSPOSE4_PS(Y,F,G,H);
 233             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 234             VV               = _mm_macc_ps(gbeps,Fp,Y);
 235             vgb              = _mm_mul_ps(gbqqfactor,VV);
 236
 237             twogbeps         = _mm_add_ps(gbeps,gbeps);
 238             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 239             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 240             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 241             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 242             fjptrA           = dvda+jnrA;
 243             fjptrB           = dvda+jnrB;
 244             fjptrC           = dvda+jnrC;
 245             fjptrD           = dvda+jnrD;
 246             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 247             velec            = _mm_mul_ps(qq00,rinv00);
 248             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 249
 250             /* Update potential sum for this i atom from the interaction with this j atom. */
 251             velecsum         = _mm_add_ps(velecsum,velec);
 252             vgbsum           = _mm_add_ps(vgbsum,vgb);
 253
 254             fscal            = felec;
 255
 256              /* Update vectorial force */
 257             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 258             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 259             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 260
 261             fjptrA             = f+j_coord_offsetA;
 262             fjptrB             = f+j_coord_offsetB;
 263             fjptrC             = f+j_coord_offsetC;
 264             fjptrD             = f+j_coord_offsetD;
 265             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 266                                                    _mm_mul_ps(dx00,fscal),
 267                                                    _mm_mul_ps(dy00,fscal),
 268                                                    _mm_mul_ps(dz00,fscal));
 269
 270             /* Inner loop uses 61 flops */
 271         }
 272
 273         if(jidx<j_index_end)
 274         {
 275
 276             /* Get j neighbor index, and coordinate index */
 277             jnrlistA         = jjnr[jidx];
 278             jnrlistB         = jjnr[jidx+1];
 279             jnrlistC         = jjnr[jidx+2];
 280             jnrlistD         = jjnr[jidx+3];
 281             /* Sign of each element will be negative for non-real atoms.
 282              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 283              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 284              */
 285             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 286             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 287             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 288             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 289             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 290             j_coord_offsetA  = DIM*jnrA;
 291             j_coord_offsetB  = DIM*jnrB;
 292             j_coord_offsetC  = DIM*jnrC;
 293             j_coord_offsetD  = DIM*jnrD;
 294
 295             /* load j atom coordinates */
 296             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 297                                               x+j_coord_offsetC,x+j_coord_offsetD,
 298                                               &jx0,&jy0,&jz0);
 299
 300             /* Calculate displacement vector */
 301             dx00             = _mm_sub_ps(ix0,jx0);
 302             dy00             = _mm_sub_ps(iy0,jy0);
 303             dz00             = _mm_sub_ps(iz0,jz0);
 304
 305             /* Calculate squared distance and things based on it */
 306             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 307
 308             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 309
 310             /* Load parameters for j particles */
 311             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 312                                                               charge+jnrC+0,charge+jnrD+0);
 313             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 314                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             r00              = _mm_mul_ps(rsq00,rinv00);
 321             r00              = _mm_andnot_ps(dummy_mask,r00);
 322
 323             /* Compute parameters for interactions between i and j atoms */
 324             qq00             = _mm_mul_ps(iq0,jq0);
 325
 326             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 327             isaprod          = _mm_mul_ps(isai0,isaj0);
 328             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 329             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 330
 331             /* Calculate generalized born table index - this is a separate table from the normal one,
 332              * but we use the same procedure by multiplying r with scale and truncating to integer.
 333              */
 334             rt               = _mm_mul_ps(r00,gbscale);
 335             gbitab           = _mm_cvttps_epi32(rt);
 336 #ifdef __XOP__
 337             gbeps            = _mm_frcz_ps(rt);
 338 #else
 339             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 340 #endif
 341             gbitab           = _mm_slli_epi32(gbitab,2);
 342
 343             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 344             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 345             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 346             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 347             _MM_TRANSPOSE4_PS(Y,F,G,H);
 348             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 349             VV               = _mm_macc_ps(gbeps,Fp,Y);
 350             vgb              = _mm_mul_ps(gbqqfactor,VV);
 351
 352             twogbeps         = _mm_add_ps(gbeps,gbeps);
 353             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 354             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 355             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 356             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 357             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 358             /* 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. */
 359             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 360             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 361             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 362             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 363             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 364             velec            = _mm_mul_ps(qq00,rinv00);
 365             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velec            = _mm_andnot_ps(dummy_mask,velec);
 369             velecsum         = _mm_add_ps(velecsum,velec);
 370             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 371             vgbsum           = _mm_add_ps(vgbsum,vgb);
 372
 373             fscal            = felec;
 374
 375             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 376
 377              /* Update vectorial force */
 378             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 379             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 380             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 381
 382             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 383             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 384             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 385             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 386             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 387                                                    _mm_mul_ps(dx00,fscal),
 388                                                    _mm_mul_ps(dy00,fscal),
 389                                                    _mm_mul_ps(dz00,fscal));
 390
 391             /* Inner loop uses 62 flops */
 392         }
 393
 394         /* End of innermost loop */
 395
 396         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 397                                               f+i_coord_offset,fshift+i_shift_offset);
 398
 399         ggid                        = gid[iidx];
 400         /* Update potential energies */
 401         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 402         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 403         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 404         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 405
 406         /* Increment number of inner iterations */
 407         inneriter                  += j_index_end - j_index_start;
 408
 409         /* Outer loop uses 9 flops */
 410     }
 411
 412     /* Increment number of outer iterations */
 413     outeriter        += nri;
 414
 415     /* Update outer/inner flops */
 416
 417     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*62);
 418 }
 419 /*
 420  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 421  * Electrostatics interaction: GeneralizedBorn
 422  * VdW interaction:            None
 423  * Geometry:                   Particle-Particle
 424  * Calculate force/pot:        Force
 425  */
 426 void
 427 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 428                     (t_nblist                    * gmx_restrict       nlist,
 429                      rvec                        * gmx_restrict          xx,
 430                      rvec                        * gmx_restrict          ff,
 431                      t_forcerec                  * gmx_restrict          fr,
 432                      t_mdatoms                   * gmx_restrict     mdatoms,
 433                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 434                      t_nrnb                      * gmx_restrict        nrnb)
 435 {
 436     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 437      * just 0 for non-waters.
 438      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 439      * jnr indices corresponding to data put in the four positions in the SIMD register.
 440      */
 441     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 442     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 443     int              jnrA,jnrB,jnrC,jnrD;
 444     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 445     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 446     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 447     real             rcutoff_scalar;
 448     real             *shiftvec,*fshift,*x,*f;
 449     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 450     real             scratch[4*DIM];
 451     __m128           fscal,rcutoff,rcutoff2,jidxall;
 452     int              vdwioffset0;
 453     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 454     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 455     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 456     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 457     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 458     real             *charge;
 459     __m128i          gbitab;
 460     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 461     __m128           minushalf = _mm_set1_ps(-0.5);
 462     real             *invsqrta,*dvda,*gbtab;
 463     __m128i          vfitab;
 464     __m128i          ifour       = _mm_set1_epi32(4);
 465     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 466     real             *vftab;
 467     __m128           dummy_mask,cutoff_mask;
 468     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 469     __m128           one     = _mm_set1_ps(1.0);
 470     __m128           two     = _mm_set1_ps(2.0);
 471     x                = xx[0];
 472     f                = ff[0];
 473
 474     nri              = nlist->nri;
 475     iinr             = nlist->iinr;
 476     jindex           = nlist->jindex;
 477     jjnr             = nlist->jjnr;
 478     shiftidx         = nlist->shift;
 479     gid              = nlist->gid;
 480     shiftvec         = fr->shift_vec[0];
 481     fshift           = fr->fshift[0];
 482     facel            = _mm_set1_ps(fr->epsfac);
 483     charge           = mdatoms->chargeA;
 484
 485     invsqrta         = fr->invsqrta;
 486     dvda             = fr->dvda;
 487     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 488     gbtab            = fr->gbtab.data;
 489     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 490
 491     /* Avoid stupid compiler warnings */
 492     jnrA = jnrB = jnrC = jnrD = 0;
 493     j_coord_offsetA = 0;
 494     j_coord_offsetB = 0;
 495     j_coord_offsetC = 0;
 496     j_coord_offsetD = 0;
 497
 498     outeriter        = 0;
 499     inneriter        = 0;
 500
 501     for(iidx=0;iidx<4*DIM;iidx++)
 502     {
 503         scratch[iidx] = 0.0;
 504     }
 505
 506     /* Start outer loop over neighborlists */
 507     for(iidx=0; iidx<nri; iidx++)
 508     {
 509         /* Load shift vector for this list */
 510         i_shift_offset   = DIM*shiftidx[iidx];
 511
 512         /* Load limits for loop over neighbors */
 513         j_index_start    = jindex[iidx];
 514         j_index_end      = jindex[iidx+1];
 515
 516         /* Get outer coordinate index */
 517         inr              = iinr[iidx];
 518         i_coord_offset   = DIM*inr;
 519
 520         /* Load i particle coords and add shift vector */
 521         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 522
 523         fix0             = _mm_setzero_ps();
 524         fiy0             = _mm_setzero_ps();
 525         fiz0             = _mm_setzero_ps();
 526
 527         /* Load parameters for i particles */
 528         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 529         isai0            = _mm_load1_ps(invsqrta+inr+0);
 530
 531         dvdasum          = _mm_setzero_ps();
 532
 533         /* Start inner kernel loop */
 534         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 535         {
 536
 537             /* Get j neighbor index, and coordinate index */
 538             jnrA             = jjnr[jidx];
 539             jnrB             = jjnr[jidx+1];
 540             jnrC             = jjnr[jidx+2];
 541             jnrD             = jjnr[jidx+3];
 542             j_coord_offsetA  = DIM*jnrA;
 543             j_coord_offsetB  = DIM*jnrB;
 544             j_coord_offsetC  = DIM*jnrC;
 545             j_coord_offsetD  = DIM*jnrD;
 546
 547             /* load j atom coordinates */
 548             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 549                                               x+j_coord_offsetC,x+j_coord_offsetD,
 550                                               &jx0,&jy0,&jz0);
 551
 552             /* Calculate displacement vector */
 553             dx00             = _mm_sub_ps(ix0,jx0);
 554             dy00             = _mm_sub_ps(iy0,jy0);
 555             dz00             = _mm_sub_ps(iz0,jz0);
 556
 557             /* Calculate squared distance and things based on it */
 558             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 559
 560             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 561
 562             /* Load parameters for j particles */
 563             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 564                                                               charge+jnrC+0,charge+jnrD+0);
 565             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 566                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             r00              = _mm_mul_ps(rsq00,rinv00);
 573
 574             /* Compute parameters for interactions between i and j atoms */
 575             qq00             = _mm_mul_ps(iq0,jq0);
 576
 577             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 578             isaprod          = _mm_mul_ps(isai0,isaj0);
 579             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 580             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 581
 582             /* Calculate generalized born table index - this is a separate table from the normal one,
 583              * but we use the same procedure by multiplying r with scale and truncating to integer.
 584              */
 585             rt               = _mm_mul_ps(r00,gbscale);
 586             gbitab           = _mm_cvttps_epi32(rt);
 587 #ifdef __XOP__
 588             gbeps            = _mm_frcz_ps(rt);
 589 #else
 590             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 591 #endif
 592             gbitab           = _mm_slli_epi32(gbitab,2);
 593
 594             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 595             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 596             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 597             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 598             _MM_TRANSPOSE4_PS(Y,F,G,H);
 599             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 600             VV               = _mm_macc_ps(gbeps,Fp,Y);
 601             vgb              = _mm_mul_ps(gbqqfactor,VV);
 602
 603             twogbeps         = _mm_add_ps(gbeps,gbeps);
 604             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 605             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 606             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 607             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 608             fjptrA           = dvda+jnrA;
 609             fjptrB           = dvda+jnrB;
 610             fjptrC           = dvda+jnrC;
 611             fjptrD           = dvda+jnrD;
 612             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 613             velec            = _mm_mul_ps(qq00,rinv00);
 614             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 615
 616             fscal            = felec;
 617
 618              /* Update vectorial force */
 619             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 620             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 621             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 622
 623             fjptrA             = f+j_coord_offsetA;
 624             fjptrB             = f+j_coord_offsetB;
 625             fjptrC             = f+j_coord_offsetC;
 626             fjptrD             = f+j_coord_offsetD;
 627             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 628                                                    _mm_mul_ps(dx00,fscal),
 629                                                    _mm_mul_ps(dy00,fscal),
 630                                                    _mm_mul_ps(dz00,fscal));
 631
 632             /* Inner loop uses 59 flops */
 633         }
 634
 635         if(jidx<j_index_end)
 636         {
 637
 638             /* Get j neighbor index, and coordinate index */
 639             jnrlistA         = jjnr[jidx];
 640             jnrlistB         = jjnr[jidx+1];
 641             jnrlistC         = jjnr[jidx+2];
 642             jnrlistD         = jjnr[jidx+3];
 643             /* Sign of each element will be negative for non-real atoms.
 644              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 645              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 646              */
 647             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 648             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 649             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 650             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 651             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 652             j_coord_offsetA  = DIM*jnrA;
 653             j_coord_offsetB  = DIM*jnrB;
 654             j_coord_offsetC  = DIM*jnrC;
 655             j_coord_offsetD  = DIM*jnrD;
 656
 657             /* load j atom coordinates */
 658             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 659                                               x+j_coord_offsetC,x+j_coord_offsetD,
 660                                               &jx0,&jy0,&jz0);
 661
 662             /* Calculate displacement vector */
 663             dx00             = _mm_sub_ps(ix0,jx0);
 664             dy00             = _mm_sub_ps(iy0,jy0);
 665             dz00             = _mm_sub_ps(iz0,jz0);
 666
 667             /* Calculate squared distance and things based on it */
 668             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 669
 670             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 671
 672             /* Load parameters for j particles */
 673             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 674                                                               charge+jnrC+0,charge+jnrD+0);
 675             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 676                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             r00              = _mm_mul_ps(rsq00,rinv00);
 683             r00              = _mm_andnot_ps(dummy_mask,r00);
 684
 685             /* Compute parameters for interactions between i and j atoms */
 686             qq00             = _mm_mul_ps(iq0,jq0);
 687
 688             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 689             isaprod          = _mm_mul_ps(isai0,isaj0);
 690             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 691             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 692
 693             /* Calculate generalized born table index - this is a separate table from the normal one,
 694              * but we use the same procedure by multiplying r with scale and truncating to integer.
 695              */
 696             rt               = _mm_mul_ps(r00,gbscale);
 697             gbitab           = _mm_cvttps_epi32(rt);
 698 #ifdef __XOP__
 699             gbeps            = _mm_frcz_ps(rt);
 700 #else
 701             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 702 #endif
 703             gbitab           = _mm_slli_epi32(gbitab,2);
 704
 705             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 706             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 707             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 708             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 709             _MM_TRANSPOSE4_PS(Y,F,G,H);
 710             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 711             VV               = _mm_macc_ps(gbeps,Fp,Y);
 712             vgb              = _mm_mul_ps(gbqqfactor,VV);
 713
 714             twogbeps         = _mm_add_ps(gbeps,gbeps);
 715             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 716             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 717             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 718             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 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 }