src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwNone_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     __m128i          gbitab;
  77     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  78     __m128           minushalf = _mm_set1_ps(-0.5);
  79     real             *invsqrta,*dvda,*gbtab;
  80     __m128i          vfitab;
  81     __m128i          ifour       = _mm_set1_epi32(4);
  82     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  83     real             *vftab;
  84     __m128           dummy_mask,cutoff_mask;
  85     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  86     __m128           one     = _mm_set1_ps(1.0);
  87     __m128           two     = _mm_set1_ps(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     facel            = _mm_set1_ps(fr->epsfac);
 100     charge           = mdatoms->chargeA;
 101
 102     invsqrta         = fr->invsqrta;
 103     dvda             = fr->dvda;
 104     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 105     gbtab            = fr->gbtab.data;
 106     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 107
 108     /* Avoid stupid compiler warnings */
 109     jnrA = jnrB = jnrC = jnrD = 0;
 110     j_coord_offsetA = 0;
 111     j_coord_offsetB = 0;
 112     j_coord_offsetC = 0;
 113     j_coord_offsetD = 0;
 114
 115     outeriter        = 0;
 116     inneriter        = 0;
 117
 118     for(iidx=0;iidx<4*DIM;iidx++)
 119     {
 120         scratch[iidx] = 0.0;
 121     }
 122
 123     /* Start outer loop over neighborlists */
 124     for(iidx=0; iidx<nri; iidx++)
 125     {
 126         /* Load shift vector for this list */
 127         i_shift_offset   = DIM*shiftidx[iidx];
 128
 129         /* Load limits for loop over neighbors */
 130         j_index_start    = jindex[iidx];
 131         j_index_end      = jindex[iidx+1];
 132
 133         /* Get outer coordinate index */
 134         inr              = iinr[iidx];
 135         i_coord_offset   = DIM*inr;
 136
 137         /* Load i particle coords and add shift vector */
 138         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 139
 140         fix0             = _mm_setzero_ps();
 141         fiy0             = _mm_setzero_ps();
 142         fiz0             = _mm_setzero_ps();
 143
 144         /* Load parameters for i particles */
 145         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 146         isai0            = _mm_load1_ps(invsqrta+inr+0);
 147
 148         /* Reset potential sums */
 149         velecsum         = _mm_setzero_ps();
 150         vgbsum           = _mm_setzero_ps();
 151         dvdasum          = _mm_setzero_ps();
 152
 153         /* Start inner kernel loop */
 154         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 155         {
 156
 157             /* Get j neighbor index, and coordinate index */
 158             jnrA             = jjnr[jidx];
 159             jnrB             = jjnr[jidx+1];
 160             jnrC             = jjnr[jidx+2];
 161             jnrD             = jjnr[jidx+3];
 162             j_coord_offsetA  = DIM*jnrA;
 163             j_coord_offsetB  = DIM*jnrB;
 164             j_coord_offsetC  = DIM*jnrC;
 165             j_coord_offsetD  = DIM*jnrD;
 166
 167             /* load j atom coordinates */
 168             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 169                                               x+j_coord_offsetC,x+j_coord_offsetD,
 170                                               &jx0,&jy0,&jz0);
 171
 172             /* Calculate displacement vector */
 173             dx00             = _mm_sub_ps(ix0,jx0);
 174             dy00             = _mm_sub_ps(iy0,jy0);
 175             dz00             = _mm_sub_ps(iz0,jz0);
 176
 177             /* Calculate squared distance and things based on it */
 178             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 179
 180             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 181
 182             /* Load parameters for j particles */
 183             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 184                                                               charge+jnrC+0,charge+jnrD+0);
 185             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 186                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 187
 188             /**************************
 189              * CALCULATE INTERACTIONS *
 190              **************************/
 191
 192             r00              = _mm_mul_ps(rsq00,rinv00);
 193
 194             /* Compute parameters for interactions between i and j atoms */
 195             qq00             = _mm_mul_ps(iq0,jq0);
 196
 197             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 198             isaprod          = _mm_mul_ps(isai0,isaj0);
 199             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 200             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 201
 202             /* Calculate generalized born table index - this is a separate table from the normal one,
 203              * but we use the same procedure by multiplying r with scale and truncating to integer.
 204              */
 205             rt               = _mm_mul_ps(r00,gbscale);
 206             gbitab           = _mm_cvttps_epi32(rt);
 207 #ifdef __XOP__
 208             gbeps            = _mm_frcz_ps(rt);
 209 #else
 210             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 211 #endif
 212             gbitab           = _mm_slli_epi32(gbitab,2);
 213
 214             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 215             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 216             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 217             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 218             _MM_TRANSPOSE4_PS(Y,F,G,H);
 219             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 220             VV               = _mm_macc_ps(gbeps,Fp,Y);
 221             vgb              = _mm_mul_ps(gbqqfactor,VV);
 222
 223             twogbeps         = _mm_add_ps(gbeps,gbeps);
 224             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 225             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 226             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 227             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 228             fjptrA           = dvda+jnrA;
 229             fjptrB           = dvda+jnrB;
 230             fjptrC           = dvda+jnrC;
 231             fjptrD           = dvda+jnrD;
 232             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 233             velec            = _mm_mul_ps(qq00,rinv00);
 234             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 235
 236             /* Update potential sum for this i atom from the interaction with this j atom. */
 237             velecsum         = _mm_add_ps(velecsum,velec);
 238             vgbsum           = _mm_add_ps(vgbsum,vgb);
 239
 240             fscal            = felec;
 241
 242              /* Update vectorial force */
 243             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 244             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 245             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 246
 247             fjptrA             = f+j_coord_offsetA;
 248             fjptrB             = f+j_coord_offsetB;
 249             fjptrC             = f+j_coord_offsetC;
 250             fjptrD             = f+j_coord_offsetD;
 251             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 252                                                    _mm_mul_ps(dx00,fscal),
 253                                                    _mm_mul_ps(dy00,fscal),
 254                                                    _mm_mul_ps(dz00,fscal));
 255
 256             /* Inner loop uses 61 flops */
 257         }
 258
 259         if(jidx<j_index_end)
 260         {
 261
 262             /* Get j neighbor index, and coordinate index */
 263             jnrlistA         = jjnr[jidx];
 264             jnrlistB         = jjnr[jidx+1];
 265             jnrlistC         = jjnr[jidx+2];
 266             jnrlistD         = jjnr[jidx+3];
 267             /* Sign of each element will be negative for non-real atoms.
 268              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 269              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 270              */
 271             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 272             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 273             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 274             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 275             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 276             j_coord_offsetA  = DIM*jnrA;
 277             j_coord_offsetB  = DIM*jnrB;
 278             j_coord_offsetC  = DIM*jnrC;
 279             j_coord_offsetD  = DIM*jnrD;
 280
 281             /* load j atom coordinates */
 282             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 283                                               x+j_coord_offsetC,x+j_coord_offsetD,
 284                                               &jx0,&jy0,&jz0);
 285
 286             /* Calculate displacement vector */
 287             dx00             = _mm_sub_ps(ix0,jx0);
 288             dy00             = _mm_sub_ps(iy0,jy0);
 289             dz00             = _mm_sub_ps(iz0,jz0);
 290
 291             /* Calculate squared distance and things based on it */
 292             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 293
 294             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 295
 296             /* Load parameters for j particles */
 297             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 298                                                               charge+jnrC+0,charge+jnrD+0);
 299             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 300                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             r00              = _mm_mul_ps(rsq00,rinv00);
 307             r00              = _mm_andnot_ps(dummy_mask,r00);
 308
 309             /* Compute parameters for interactions between i and j atoms */
 310             qq00             = _mm_mul_ps(iq0,jq0);
 311
 312             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 313             isaprod          = _mm_mul_ps(isai0,isaj0);
 314             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 315             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 316
 317             /* Calculate generalized born table index - this is a separate table from the normal one,
 318              * but we use the same procedure by multiplying r with scale and truncating to integer.
 319              */
 320             rt               = _mm_mul_ps(r00,gbscale);
 321             gbitab           = _mm_cvttps_epi32(rt);
 322 #ifdef __XOP__
 323             gbeps            = _mm_frcz_ps(rt);
 324 #else
 325             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 326 #endif
 327             gbitab           = _mm_slli_epi32(gbitab,2);
 328
 329             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 330             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 331             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 332             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 333             _MM_TRANSPOSE4_PS(Y,F,G,H);
 334             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 335             VV               = _mm_macc_ps(gbeps,Fp,Y);
 336             vgb              = _mm_mul_ps(gbqqfactor,VV);
 337
 338             twogbeps         = _mm_add_ps(gbeps,gbeps);
 339             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 340             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 341             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 342             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 343             /* 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. */
 344             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 345             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 346             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 347             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 348             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 349             velec            = _mm_mul_ps(qq00,rinv00);
 350             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             velec            = _mm_andnot_ps(dummy_mask,velec);
 354             velecsum         = _mm_add_ps(velecsum,velec);
 355             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 356             vgbsum           = _mm_add_ps(vgbsum,vgb);
 357
 358             fscal            = felec;
 359
 360             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 361
 362              /* Update vectorial force */
 363             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 364             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 365             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 366
 367             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 368             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 369             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 370             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 371             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 372                                                    _mm_mul_ps(dx00,fscal),
 373                                                    _mm_mul_ps(dy00,fscal),
 374                                                    _mm_mul_ps(dz00,fscal));
 375
 376             /* Inner loop uses 62 flops */
 377         }
 378
 379         /* End of innermost loop */
 380
 381         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 382                                               f+i_coord_offset,fshift+i_shift_offset);
 383
 384         ggid                        = gid[iidx];
 385         /* Update potential energies */
 386         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 387         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 388         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 389         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 390
 391         /* Increment number of inner iterations */
 392         inneriter                  += j_index_end - j_index_start;
 393
 394         /* Outer loop uses 9 flops */
 395     }
 396
 397     /* Increment number of outer iterations */
 398     outeriter        += nri;
 399
 400     /* Update outer/inner flops */
 401
 402     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*62);
 403 }
 404 /*
 405  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 406  * Electrostatics interaction: GeneralizedBorn
 407  * VdW interaction:            None
 408  * Geometry:                   Particle-Particle
 409  * Calculate force/pot:        Force
 410  */
 411 void
 412 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 413                     (t_nblist * gmx_restrict                nlist,
 414                      rvec * gmx_restrict                    xx,
 415                      rvec * gmx_restrict                    ff,
 416                      t_forcerec * gmx_restrict              fr,
 417                      t_mdatoms * gmx_restrict               mdatoms,
 418                      nb_kernel_data_t * gmx_restrict        kernel_data,
 419                      t_nrnb * gmx_restrict                  nrnb)
 420 {
 421     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 422      * just 0 for non-waters.
 423      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 424      * jnr indices corresponding to data put in the four positions in the SIMD register.
 425      */
 426     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 427     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 428     int              jnrA,jnrB,jnrC,jnrD;
 429     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 430     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 431     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 432     real             rcutoff_scalar;
 433     real             *shiftvec,*fshift,*x,*f;
 434     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 435     real             scratch[4*DIM];
 436     __m128           fscal,rcutoff,rcutoff2,jidxall;
 437     int              vdwioffset0;
 438     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 439     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 440     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 441     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 442     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 443     real             *charge;
 444     __m128i          gbitab;
 445     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 446     __m128           minushalf = _mm_set1_ps(-0.5);
 447     real             *invsqrta,*dvda,*gbtab;
 448     __m128i          vfitab;
 449     __m128i          ifour       = _mm_set1_epi32(4);
 450     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 451     real             *vftab;
 452     __m128           dummy_mask,cutoff_mask;
 453     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 454     __m128           one     = _mm_set1_ps(1.0);
 455     __m128           two     = _mm_set1_ps(2.0);
 456     x                = xx[0];
 457     f                = ff[0];
 458
 459     nri              = nlist->nri;
 460     iinr             = nlist->iinr;
 461     jindex           = nlist->jindex;
 462     jjnr             = nlist->jjnr;
 463     shiftidx         = nlist->shift;
 464     gid              = nlist->gid;
 465     shiftvec         = fr->shift_vec[0];
 466     fshift           = fr->fshift[0];
 467     facel            = _mm_set1_ps(fr->epsfac);
 468     charge           = mdatoms->chargeA;
 469
 470     invsqrta         = fr->invsqrta;
 471     dvda             = fr->dvda;
 472     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 473     gbtab            = fr->gbtab.data;
 474     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 475
 476     /* Avoid stupid compiler warnings */
 477     jnrA = jnrB = jnrC = jnrD = 0;
 478     j_coord_offsetA = 0;
 479     j_coord_offsetB = 0;
 480     j_coord_offsetC = 0;
 481     j_coord_offsetD = 0;
 482
 483     outeriter        = 0;
 484     inneriter        = 0;
 485
 486     for(iidx=0;iidx<4*DIM;iidx++)
 487     {
 488         scratch[iidx] = 0.0;
 489     }
 490
 491     /* Start outer loop over neighborlists */
 492     for(iidx=0; iidx<nri; iidx++)
 493     {
 494         /* Load shift vector for this list */
 495         i_shift_offset   = DIM*shiftidx[iidx];
 496
 497         /* Load limits for loop over neighbors */
 498         j_index_start    = jindex[iidx];
 499         j_index_end      = jindex[iidx+1];
 500
 501         /* Get outer coordinate index */
 502         inr              = iinr[iidx];
 503         i_coord_offset   = DIM*inr;
 504
 505         /* Load i particle coords and add shift vector */
 506         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 507
 508         fix0             = _mm_setzero_ps();
 509         fiy0             = _mm_setzero_ps();
 510         fiz0             = _mm_setzero_ps();
 511
 512         /* Load parameters for i particles */
 513         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 514         isai0            = _mm_load1_ps(invsqrta+inr+0);
 515
 516         dvdasum          = _mm_setzero_ps();
 517
 518         /* Start inner kernel loop */
 519         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 520         {
 521
 522             /* Get j neighbor index, and coordinate index */
 523             jnrA             = jjnr[jidx];
 524             jnrB             = jjnr[jidx+1];
 525             jnrC             = jjnr[jidx+2];
 526             jnrD             = jjnr[jidx+3];
 527             j_coord_offsetA  = DIM*jnrA;
 528             j_coord_offsetB  = DIM*jnrB;
 529             j_coord_offsetC  = DIM*jnrC;
 530             j_coord_offsetD  = DIM*jnrD;
 531
 532             /* load j atom coordinates */
 533             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 534                                               x+j_coord_offsetC,x+j_coord_offsetD,
 535                                               &jx0,&jy0,&jz0);
 536
 537             /* Calculate displacement vector */
 538             dx00             = _mm_sub_ps(ix0,jx0);
 539             dy00             = _mm_sub_ps(iy0,jy0);
 540             dz00             = _mm_sub_ps(iz0,jz0);
 541
 542             /* Calculate squared distance and things based on it */
 543             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 544
 545             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 546
 547             /* Load parameters for j particles */
 548             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 549                                                               charge+jnrC+0,charge+jnrD+0);
 550             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 551                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 552
 553             /**************************
 554              * CALCULATE INTERACTIONS *
 555              **************************/
 556
 557             r00              = _mm_mul_ps(rsq00,rinv00);
 558
 559             /* Compute parameters for interactions between i and j atoms */
 560             qq00             = _mm_mul_ps(iq0,jq0);
 561
 562             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 563             isaprod          = _mm_mul_ps(isai0,isaj0);
 564             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 565             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 566
 567             /* Calculate generalized born table index - this is a separate table from the normal one,
 568              * but we use the same procedure by multiplying r with scale and truncating to integer.
 569              */
 570             rt               = _mm_mul_ps(r00,gbscale);
 571             gbitab           = _mm_cvttps_epi32(rt);
 572 #ifdef __XOP__
 573             gbeps            = _mm_frcz_ps(rt);
 574 #else
 575             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 576 #endif
 577             gbitab           = _mm_slli_epi32(gbitab,2);
 578
 579             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 580             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 581             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 582             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 583             _MM_TRANSPOSE4_PS(Y,F,G,H);
 584             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 585             VV               = _mm_macc_ps(gbeps,Fp,Y);
 586             vgb              = _mm_mul_ps(gbqqfactor,VV);
 587
 588             twogbeps         = _mm_add_ps(gbeps,gbeps);
 589             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 590             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 591             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 592             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 593             fjptrA           = dvda+jnrA;
 594             fjptrB           = dvda+jnrB;
 595             fjptrC           = dvda+jnrC;
 596             fjptrD           = dvda+jnrD;
 597             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 598             velec            = _mm_mul_ps(qq00,rinv00);
 599             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 600
 601             fscal            = felec;
 602
 603              /* Update vectorial force */
 604             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 605             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 606             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 607
 608             fjptrA             = f+j_coord_offsetA;
 609             fjptrB             = f+j_coord_offsetB;
 610             fjptrC             = f+j_coord_offsetC;
 611             fjptrD             = f+j_coord_offsetD;
 612             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 613                                                    _mm_mul_ps(dx00,fscal),
 614                                                    _mm_mul_ps(dy00,fscal),
 615                                                    _mm_mul_ps(dz00,fscal));
 616
 617             /* Inner loop uses 59 flops */
 618         }
 619
 620         if(jidx<j_index_end)
 621         {
 622
 623             /* Get j neighbor index, and coordinate index */
 624             jnrlistA         = jjnr[jidx];
 625             jnrlistB         = jjnr[jidx+1];
 626             jnrlistC         = jjnr[jidx+2];
 627             jnrlistD         = jjnr[jidx+3];
 628             /* Sign of each element will be negative for non-real atoms.
 629              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 630              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 631              */
 632             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 633             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 634             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 635             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 636             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 637             j_coord_offsetA  = DIM*jnrA;
 638             j_coord_offsetB  = DIM*jnrB;
 639             j_coord_offsetC  = DIM*jnrC;
 640             j_coord_offsetD  = DIM*jnrD;
 641
 642             /* load j atom coordinates */
 643             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 644                                               x+j_coord_offsetC,x+j_coord_offsetD,
 645                                               &jx0,&jy0,&jz0);
 646
 647             /* Calculate displacement vector */
 648             dx00             = _mm_sub_ps(ix0,jx0);
 649             dy00             = _mm_sub_ps(iy0,jy0);
 650             dz00             = _mm_sub_ps(iz0,jz0);
 651
 652             /* Calculate squared distance and things based on it */
 653             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 654
 655             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 656
 657             /* Load parameters for j particles */
 658             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 659                                                               charge+jnrC+0,charge+jnrD+0);
 660             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 661                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 662
 663             /**************************
 664              * CALCULATE INTERACTIONS *
 665              **************************/
 666
 667             r00              = _mm_mul_ps(rsq00,rinv00);
 668             r00              = _mm_andnot_ps(dummy_mask,r00);
 669
 670             /* Compute parameters for interactions between i and j atoms */
 671             qq00             = _mm_mul_ps(iq0,jq0);
 672
 673             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 674             isaprod          = _mm_mul_ps(isai0,isaj0);
 675             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 676             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 677
 678             /* Calculate generalized born table index - this is a separate table from the normal one,
 679              * but we use the same procedure by multiplying r with scale and truncating to integer.
 680              */
 681             rt               = _mm_mul_ps(r00,gbscale);
 682             gbitab           = _mm_cvttps_epi32(rt);
 683 #ifdef __XOP__
 684             gbeps            = _mm_frcz_ps(rt);
 685 #else
 686             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 687 #endif
 688             gbitab           = _mm_slli_epi32(gbitab,2);
 689
 690             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 691             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 692             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 693             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 694             _MM_TRANSPOSE4_PS(Y,F,G,H);
 695             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 696             VV               = _mm_macc_ps(gbeps,Fp,Y);
 697             vgb              = _mm_mul_ps(gbqqfactor,VV);
 698
 699             twogbeps         = _mm_add_ps(gbeps,gbeps);
 700             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 701             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 702             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 703             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 704             /* 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. */
 705             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 706             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 707             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 708             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 709             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 710             velec            = _mm_mul_ps(qq00,rinv00);
 711             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 712
 713             fscal            = felec;
 714
 715             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 716
 717              /* Update vectorial force */
 718             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 719             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 720             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 721
 722             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 723             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 724             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 725             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 726             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 727                                                    _mm_mul_ps(dx00,fscal),
 728                                                    _mm_mul_ps(dy00,fscal),
 729                                                    _mm_mul_ps(dz00,fscal));
 730
 731             /* Inner loop uses 60 flops */
 732         }
 733
 734         /* End of innermost loop */
 735
 736         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 737                                               f+i_coord_offset,fshift+i_shift_offset);
 738
 739         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 740         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 741
 742         /* Increment number of inner iterations */
 743         inneriter                  += j_index_end - j_index_start;
 744
 745         /* Outer loop uses 7 flops */
 746     }
 747
 748     /* Increment number of outer iterations */
 749     outeriter        += nri;
 750
 751     /* Update outer/inner flops */
 752
 753     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*60);
 754 }