src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCSTab_VdwNone_GeomW3P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomW3P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: CubicSplineTable
  52  * VdW interaction:            None
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCSTab_VdwNone_GeomW3P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  92     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  93     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     __m128i          vfitab;
  96     __m128i          ifour       = _mm_set1_epi32(4);
  97     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  98     real             *vftab;
  99     __m128           dummy_mask,cutoff_mask;
 100     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 101     __m128           one     = _mm_set1_ps(1.0);
 102     __m128           two     = _mm_set1_ps(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm_set1_ps(fr->ic->epsfac);
 115     charge           = mdatoms->chargeA;
 116
 117     vftab            = kernel_data->table_elec->data;
 118     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 119
 120     /* Setup water-specific parameters */
 121     inr              = nlist->iinr[0];
 122     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 123     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 124     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = jnrC = jnrD = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130     j_coord_offsetC = 0;
 131     j_coord_offsetD = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     for(iidx=0;iidx<4*DIM;iidx++)
 137     {
 138         scratch[iidx] = 0.0;
 139     }
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 157                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 158
 159         fix0             = _mm_setzero_ps();
 160         fiy0             = _mm_setzero_ps();
 161         fiz0             = _mm_setzero_ps();
 162         fix1             = _mm_setzero_ps();
 163         fiy1             = _mm_setzero_ps();
 164         fiz1             = _mm_setzero_ps();
 165         fix2             = _mm_setzero_ps();
 166         fiy2             = _mm_setzero_ps();
 167         fiz2             = _mm_setzero_ps();
 168
 169         /* Reset potential sums */
 170         velecsum         = _mm_setzero_ps();
 171
 172         /* Start inner kernel loop */
 173         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 174         {
 175
 176             /* Get j neighbor index, and coordinate index */
 177             jnrA             = jjnr[jidx];
 178             jnrB             = jjnr[jidx+1];
 179             jnrC             = jjnr[jidx+2];
 180             jnrD             = jjnr[jidx+3];
 181             j_coord_offsetA  = DIM*jnrA;
 182             j_coord_offsetB  = DIM*jnrB;
 183             j_coord_offsetC  = DIM*jnrC;
 184             j_coord_offsetD  = DIM*jnrD;
 185
 186             /* load j atom coordinates */
 187             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 188                                               x+j_coord_offsetC,x+j_coord_offsetD,
 189                                               &jx0,&jy0,&jz0);
 190
 191             /* Calculate displacement vector */
 192             dx00             = _mm_sub_ps(ix0,jx0);
 193             dy00             = _mm_sub_ps(iy0,jy0);
 194             dz00             = _mm_sub_ps(iz0,jz0);
 195             dx10             = _mm_sub_ps(ix1,jx0);
 196             dy10             = _mm_sub_ps(iy1,jy0);
 197             dz10             = _mm_sub_ps(iz1,jz0);
 198             dx20             = _mm_sub_ps(ix2,jx0);
 199             dy20             = _mm_sub_ps(iy2,jy0);
 200             dz20             = _mm_sub_ps(iz2,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 204             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 205             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 206
 207             rinv00           = avx128fma_invsqrt_f(rsq00);
 208             rinv10           = avx128fma_invsqrt_f(rsq10);
 209             rinv20           = avx128fma_invsqrt_f(rsq20);
 210
 211             /* Load parameters for j particles */
 212             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 213                                                               charge+jnrC+0,charge+jnrD+0);
 214
 215             fjx0             = _mm_setzero_ps();
 216             fjy0             = _mm_setzero_ps();
 217             fjz0             = _mm_setzero_ps();
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             r00              = _mm_mul_ps(rsq00,rinv00);
 224
 225             /* Compute parameters for interactions between i and j atoms */
 226             qq00             = _mm_mul_ps(iq0,jq0);
 227
 228             /* Calculate table index by multiplying r with table scale and truncate to integer */
 229             rt               = _mm_mul_ps(r00,vftabscale);
 230             vfitab           = _mm_cvttps_epi32(rt);
 231 #ifdef __XOP__
 232             vfeps            = _mm_frcz_ps(rt);
 233 #else
 234             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 235 #endif
 236             twovfeps         = _mm_add_ps(vfeps,vfeps);
 237             vfitab           = _mm_slli_epi32(vfitab,2);
 238
 239             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 240             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 241             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 242             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 243             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 244             _MM_TRANSPOSE4_PS(Y,F,G,H);
 245             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 246             VV               = _mm_macc_ps(vfeps,Fp,Y);
 247             velec            = _mm_mul_ps(qq00,VV);
 248             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 249             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 250
 251             /* Update potential sum for this i atom from the interaction with this j atom. */
 252             velecsum         = _mm_add_ps(velecsum,velec);
 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             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 262             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 263             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 264
 265             /**************************
 266              * CALCULATE INTERACTIONS *
 267              **************************/
 268
 269             r10              = _mm_mul_ps(rsq10,rinv10);
 270
 271             /* Compute parameters for interactions between i and j atoms */
 272             qq10             = _mm_mul_ps(iq1,jq0);
 273
 274             /* Calculate table index by multiplying r with table scale and truncate to integer */
 275             rt               = _mm_mul_ps(r10,vftabscale);
 276             vfitab           = _mm_cvttps_epi32(rt);
 277 #ifdef __XOP__
 278             vfeps            = _mm_frcz_ps(rt);
 279 #else
 280             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 281 #endif
 282             twovfeps         = _mm_add_ps(vfeps,vfeps);
 283             vfitab           = _mm_slli_epi32(vfitab,2);
 284
 285             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 286             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 287             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 288             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 289             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 290             _MM_TRANSPOSE4_PS(Y,F,G,H);
 291             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 292             VV               = _mm_macc_ps(vfeps,Fp,Y);
 293             velec            = _mm_mul_ps(qq10,VV);
 294             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 295             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq10,FF),_mm_mul_ps(vftabscale,rinv10)));
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _mm_add_ps(velecsum,velec);
 299
 300             fscal            = felec;
 301
 302              /* Update vectorial force */
 303             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 304             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 305             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 306
 307             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 308             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 309             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             r20              = _mm_mul_ps(rsq20,rinv20);
 316
 317             /* Compute parameters for interactions between i and j atoms */
 318             qq20             = _mm_mul_ps(iq2,jq0);
 319
 320             /* Calculate table index by multiplying r with table scale and truncate to integer */
 321             rt               = _mm_mul_ps(r20,vftabscale);
 322             vfitab           = _mm_cvttps_epi32(rt);
 323 #ifdef __XOP__
 324             vfeps            = _mm_frcz_ps(rt);
 325 #else
 326             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 327 #endif
 328             twovfeps         = _mm_add_ps(vfeps,vfeps);
 329             vfitab           = _mm_slli_epi32(vfitab,2);
 330
 331             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 332             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 333             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 334             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 335             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 336             _MM_TRANSPOSE4_PS(Y,F,G,H);
 337             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 338             VV               = _mm_macc_ps(vfeps,Fp,Y);
 339             velec            = _mm_mul_ps(qq20,VV);
 340             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 341             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq20,FF),_mm_mul_ps(vftabscale,rinv20)));
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             velecsum         = _mm_add_ps(velecsum,velec);
 345
 346             fscal            = felec;
 347
 348              /* Update vectorial force */
 349             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 350             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 351             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 352
 353             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 354             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 355             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 356
 357             fjptrA             = f+j_coord_offsetA;
 358             fjptrB             = f+j_coord_offsetB;
 359             fjptrC             = f+j_coord_offsetC;
 360             fjptrD             = f+j_coord_offsetD;
 361
 362             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 363
 364             /* Inner loop uses 138 flops */
 365         }
 366
 367         if(jidx<j_index_end)
 368         {
 369
 370             /* Get j neighbor index, and coordinate index */
 371             jnrlistA         = jjnr[jidx];
 372             jnrlistB         = jjnr[jidx+1];
 373             jnrlistC         = jjnr[jidx+2];
 374             jnrlistD         = jjnr[jidx+3];
 375             /* Sign of each element will be negative for non-real atoms.
 376              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 377              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 378              */
 379             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 380             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 381             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 382             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 383             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 384             j_coord_offsetA  = DIM*jnrA;
 385             j_coord_offsetB  = DIM*jnrB;
 386             j_coord_offsetC  = DIM*jnrC;
 387             j_coord_offsetD  = DIM*jnrD;
 388
 389             /* load j atom coordinates */
 390             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 391                                               x+j_coord_offsetC,x+j_coord_offsetD,
 392                                               &jx0,&jy0,&jz0);
 393
 394             /* Calculate displacement vector */
 395             dx00             = _mm_sub_ps(ix0,jx0);
 396             dy00             = _mm_sub_ps(iy0,jy0);
 397             dz00             = _mm_sub_ps(iz0,jz0);
 398             dx10             = _mm_sub_ps(ix1,jx0);
 399             dy10             = _mm_sub_ps(iy1,jy0);
 400             dz10             = _mm_sub_ps(iz1,jz0);
 401             dx20             = _mm_sub_ps(ix2,jx0);
 402             dy20             = _mm_sub_ps(iy2,jy0);
 403             dz20             = _mm_sub_ps(iz2,jz0);
 404
 405             /* Calculate squared distance and things based on it */
 406             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 407             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 408             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 409
 410             rinv00           = avx128fma_invsqrt_f(rsq00);
 411             rinv10           = avx128fma_invsqrt_f(rsq10);
 412             rinv20           = avx128fma_invsqrt_f(rsq20);
 413
 414             /* Load parameters for j particles */
 415             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 416                                                               charge+jnrC+0,charge+jnrD+0);
 417
 418             fjx0             = _mm_setzero_ps();
 419             fjy0             = _mm_setzero_ps();
 420             fjz0             = _mm_setzero_ps();
 421
 422             /**************************
 423              * CALCULATE INTERACTIONS *
 424              **************************/
 425
 426             r00              = _mm_mul_ps(rsq00,rinv00);
 427             r00              = _mm_andnot_ps(dummy_mask,r00);
 428
 429             /* Compute parameters for interactions between i and j atoms */
 430             qq00             = _mm_mul_ps(iq0,jq0);
 431
 432             /* Calculate table index by multiplying r with table scale and truncate to integer */
 433             rt               = _mm_mul_ps(r00,vftabscale);
 434             vfitab           = _mm_cvttps_epi32(rt);
 435 #ifdef __XOP__
 436             vfeps            = _mm_frcz_ps(rt);
 437 #else
 438             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 439 #endif
 440             twovfeps         = _mm_add_ps(vfeps,vfeps);
 441             vfitab           = _mm_slli_epi32(vfitab,2);
 442
 443             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 444             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 445             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 446             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 447             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 448             _MM_TRANSPOSE4_PS(Y,F,G,H);
 449             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 450             VV               = _mm_macc_ps(vfeps,Fp,Y);
 451             velec            = _mm_mul_ps(qq00,VV);
 452             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 453             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velec            = _mm_andnot_ps(dummy_mask,velec);
 457             velecsum         = _mm_add_ps(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 462
 463              /* Update vectorial force */
 464             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 465             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 466             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 467
 468             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 469             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 470             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             r10              = _mm_mul_ps(rsq10,rinv10);
 477             r10              = _mm_andnot_ps(dummy_mask,r10);
 478
 479             /* Compute parameters for interactions between i and j atoms */
 480             qq10             = _mm_mul_ps(iq1,jq0);
 481
 482             /* Calculate table index by multiplying r with table scale and truncate to integer */
 483             rt               = _mm_mul_ps(r10,vftabscale);
 484             vfitab           = _mm_cvttps_epi32(rt);
 485 #ifdef __XOP__
 486             vfeps            = _mm_frcz_ps(rt);
 487 #else
 488             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 489 #endif
 490             twovfeps         = _mm_add_ps(vfeps,vfeps);
 491             vfitab           = _mm_slli_epi32(vfitab,2);
 492
 493             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 494             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 495             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 496             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 497             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 498             _MM_TRANSPOSE4_PS(Y,F,G,H);
 499             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 500             VV               = _mm_macc_ps(vfeps,Fp,Y);
 501             velec            = _mm_mul_ps(qq10,VV);
 502             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 503             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq10,FF),_mm_mul_ps(vftabscale,rinv10)));
 504
 505             /* Update potential sum for this i atom from the interaction with this j atom. */
 506             velec            = _mm_andnot_ps(dummy_mask,velec);
 507             velecsum         = _mm_add_ps(velecsum,velec);
 508
 509             fscal            = felec;
 510
 511             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 512
 513              /* Update vectorial force */
 514             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 515             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 516             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 517
 518             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 519             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 520             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 521
 522             /**************************
 523              * CALCULATE INTERACTIONS *
 524              **************************/
 525
 526             r20              = _mm_mul_ps(rsq20,rinv20);
 527             r20              = _mm_andnot_ps(dummy_mask,r20);
 528
 529             /* Compute parameters for interactions between i and j atoms */
 530             qq20             = _mm_mul_ps(iq2,jq0);
 531
 532             /* Calculate table index by multiplying r with table scale and truncate to integer */
 533             rt               = _mm_mul_ps(r20,vftabscale);
 534             vfitab           = _mm_cvttps_epi32(rt);
 535 #ifdef __XOP__
 536             vfeps            = _mm_frcz_ps(rt);
 537 #else
 538             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 539 #endif
 540             twovfeps         = _mm_add_ps(vfeps,vfeps);
 541             vfitab           = _mm_slli_epi32(vfitab,2);
 542
 543             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 544             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 545             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 546             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 547             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 548             _MM_TRANSPOSE4_PS(Y,F,G,H);
 549             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 550             VV               = _mm_macc_ps(vfeps,Fp,Y);
 551             velec            = _mm_mul_ps(qq20,VV);
 552             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 553             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq20,FF),_mm_mul_ps(vftabscale,rinv20)));
 554
 555             /* Update potential sum for this i atom from the interaction with this j atom. */
 556             velec            = _mm_andnot_ps(dummy_mask,velec);
 557             velecsum         = _mm_add_ps(velecsum,velec);
 558
 559             fscal            = felec;
 560
 561             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 562
 563              /* Update vectorial force */
 564             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 565             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 566             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 567
 568             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 569             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 570             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 571
 572             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 573             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 574             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 575             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 576
 577             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 578
 579             /* Inner loop uses 141 flops */
 580         }
 581
 582         /* End of innermost loop */
 583
 584         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 585                                               f+i_coord_offset,fshift+i_shift_offset);
 586
 587         ggid                        = gid[iidx];
 588         /* Update potential energies */
 589         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 590
 591         /* Increment number of inner iterations */
 592         inneriter                  += j_index_end - j_index_start;
 593
 594         /* Outer loop uses 19 flops */
 595     }
 596
 597     /* Increment number of outer iterations */
 598     outeriter        += nri;
 599
 600     /* Update outer/inner flops */
 601
 602     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_VF,outeriter*19 + inneriter*141);
 603 }
 604 /*
 605  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomW3P1_F_avx_128_fma_single
 606  * Electrostatics interaction: CubicSplineTable
 607  * VdW interaction:            None
 608  * Geometry:                   Water3-Particle
 609  * Calculate force/pot:        Force
 610  */
 611 void
 612 nb_kernel_ElecCSTab_VdwNone_GeomW3P1_F_avx_128_fma_single
 613                     (t_nblist                    * gmx_restrict       nlist,
 614                      rvec                        * gmx_restrict          xx,
 615                      rvec                        * gmx_restrict          ff,
 616                      struct t_forcerec           * gmx_restrict          fr,
 617                      t_mdatoms                   * gmx_restrict     mdatoms,
 618                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 619                      t_nrnb                      * gmx_restrict        nrnb)
 620 {
 621     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 622      * just 0 for non-waters.
 623      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 624      * jnr indices corresponding to data put in the four positions in the SIMD register.
 625      */
 626     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 627     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 628     int              jnrA,jnrB,jnrC,jnrD;
 629     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 630     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 631     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 632     real             rcutoff_scalar;
 633     real             *shiftvec,*fshift,*x,*f;
 634     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 635     real             scratch[4*DIM];
 636     __m128           fscal,rcutoff,rcutoff2,jidxall;
 637     int              vdwioffset0;
 638     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 639     int              vdwioffset1;
 640     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 641     int              vdwioffset2;
 642     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 643     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 644     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 645     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 646     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 647     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 648     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 649     real             *charge;
 650     __m128i          vfitab;
 651     __m128i          ifour       = _mm_set1_epi32(4);
 652     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 653     real             *vftab;
 654     __m128           dummy_mask,cutoff_mask;
 655     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 656     __m128           one     = _mm_set1_ps(1.0);
 657     __m128           two     = _mm_set1_ps(2.0);
 658     x                = xx[0];
 659     f                = ff[0];
 660
 661     nri              = nlist->nri;
 662     iinr             = nlist->iinr;
 663     jindex           = nlist->jindex;
 664     jjnr             = nlist->jjnr;
 665     shiftidx         = nlist->shift;
 666     gid              = nlist->gid;
 667     shiftvec         = fr->shift_vec[0];
 668     fshift           = fr->fshift[0];
 669     facel            = _mm_set1_ps(fr->ic->epsfac);
 670     charge           = mdatoms->chargeA;
 671
 672     vftab            = kernel_data->table_elec->data;
 673     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 674
 675     /* Setup water-specific parameters */
 676     inr              = nlist->iinr[0];
 677     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 678     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 679     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 680
 681     /* Avoid stupid compiler warnings */
 682     jnrA = jnrB = jnrC = jnrD = 0;
 683     j_coord_offsetA = 0;
 684     j_coord_offsetB = 0;
 685     j_coord_offsetC = 0;
 686     j_coord_offsetD = 0;
 687
 688     outeriter        = 0;
 689     inneriter        = 0;
 690
 691     for(iidx=0;iidx<4*DIM;iidx++)
 692     {
 693         scratch[iidx] = 0.0;
 694     }
 695
 696     /* Start outer loop over neighborlists */
 697     for(iidx=0; iidx<nri; iidx++)
 698     {
 699         /* Load shift vector for this list */
 700         i_shift_offset   = DIM*shiftidx[iidx];
 701
 702         /* Load limits for loop over neighbors */
 703         j_index_start    = jindex[iidx];
 704         j_index_end      = jindex[iidx+1];
 705
 706         /* Get outer coordinate index */
 707         inr              = iinr[iidx];
 708         i_coord_offset   = DIM*inr;
 709
 710         /* Load i particle coords and add shift vector */
 711         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 712                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 713
 714         fix0             = _mm_setzero_ps();
 715         fiy0             = _mm_setzero_ps();
 716         fiz0             = _mm_setzero_ps();
 717         fix1             = _mm_setzero_ps();
 718         fiy1             = _mm_setzero_ps();
 719         fiz1             = _mm_setzero_ps();
 720         fix2             = _mm_setzero_ps();
 721         fiy2             = _mm_setzero_ps();
 722         fiz2             = _mm_setzero_ps();
 723
 724         /* Start inner kernel loop */
 725         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 726         {
 727
 728             /* Get j neighbor index, and coordinate index */
 729             jnrA             = jjnr[jidx];
 730             jnrB             = jjnr[jidx+1];
 731             jnrC             = jjnr[jidx+2];
 732             jnrD             = jjnr[jidx+3];
 733             j_coord_offsetA  = DIM*jnrA;
 734             j_coord_offsetB  = DIM*jnrB;
 735             j_coord_offsetC  = DIM*jnrC;
 736             j_coord_offsetD  = DIM*jnrD;
 737
 738             /* load j atom coordinates */
 739             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 740                                               x+j_coord_offsetC,x+j_coord_offsetD,
 741                                               &jx0,&jy0,&jz0);
 742
 743             /* Calculate displacement vector */
 744             dx00             = _mm_sub_ps(ix0,jx0);
 745             dy00             = _mm_sub_ps(iy0,jy0);
 746             dz00             = _mm_sub_ps(iz0,jz0);
 747             dx10             = _mm_sub_ps(ix1,jx0);
 748             dy10             = _mm_sub_ps(iy1,jy0);
 749             dz10             = _mm_sub_ps(iz1,jz0);
 750             dx20             = _mm_sub_ps(ix2,jx0);
 751             dy20             = _mm_sub_ps(iy2,jy0);
 752             dz20             = _mm_sub_ps(iz2,jz0);
 753
 754             /* Calculate squared distance and things based on it */
 755             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 756             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 757             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 758
 759             rinv00           = avx128fma_invsqrt_f(rsq00);
 760             rinv10           = avx128fma_invsqrt_f(rsq10);
 761             rinv20           = avx128fma_invsqrt_f(rsq20);
 762
 763             /* Load parameters for j particles */
 764             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 765                                                               charge+jnrC+0,charge+jnrD+0);
 766
 767             fjx0             = _mm_setzero_ps();
 768             fjy0             = _mm_setzero_ps();
 769             fjz0             = _mm_setzero_ps();
 770
 771             /**************************
 772              * CALCULATE INTERACTIONS *
 773              **************************/
 774
 775             r00              = _mm_mul_ps(rsq00,rinv00);
 776
 777             /* Compute parameters for interactions between i and j atoms */
 778             qq00             = _mm_mul_ps(iq0,jq0);
 779
 780             /* Calculate table index by multiplying r with table scale and truncate to integer */
 781             rt               = _mm_mul_ps(r00,vftabscale);
 782             vfitab           = _mm_cvttps_epi32(rt);
 783 #ifdef __XOP__
 784             vfeps            = _mm_frcz_ps(rt);
 785 #else
 786             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 787 #endif
 788             twovfeps         = _mm_add_ps(vfeps,vfeps);
 789             vfitab           = _mm_slli_epi32(vfitab,2);
 790
 791             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 792             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 793             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 794             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 795             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 796             _MM_TRANSPOSE4_PS(Y,F,G,H);
 797             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 798             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 799             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 800
 801             fscal            = felec;
 802
 803              /* Update vectorial force */
 804             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 805             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 806             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 807
 808             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 809             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 810             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 811
 812             /**************************
 813              * CALCULATE INTERACTIONS *
 814              **************************/
 815
 816             r10              = _mm_mul_ps(rsq10,rinv10);
 817
 818             /* Compute parameters for interactions between i and j atoms */
 819             qq10             = _mm_mul_ps(iq1,jq0);
 820
 821             /* Calculate table index by multiplying r with table scale and truncate to integer */
 822             rt               = _mm_mul_ps(r10,vftabscale);
 823             vfitab           = _mm_cvttps_epi32(rt);
 824 #ifdef __XOP__
 825             vfeps            = _mm_frcz_ps(rt);
 826 #else
 827             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 828 #endif
 829             twovfeps         = _mm_add_ps(vfeps,vfeps);
 830             vfitab           = _mm_slli_epi32(vfitab,2);
 831
 832             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 833             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 834             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 835             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 836             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 837             _MM_TRANSPOSE4_PS(Y,F,G,H);
 838             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 839             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 840             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq10,FF),_mm_mul_ps(vftabscale,rinv10)));
 841
 842             fscal            = felec;
 843
 844              /* Update vectorial force */
 845             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 846             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 847             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 848
 849             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 850             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 851             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 852
 853             /**************************
 854              * CALCULATE INTERACTIONS *
 855              **************************/
 856
 857             r20              = _mm_mul_ps(rsq20,rinv20);
 858
 859             /* Compute parameters for interactions between i and j atoms */
 860             qq20             = _mm_mul_ps(iq2,jq0);
 861
 862             /* Calculate table index by multiplying r with table scale and truncate to integer */
 863             rt               = _mm_mul_ps(r20,vftabscale);
 864             vfitab           = _mm_cvttps_epi32(rt);
 865 #ifdef __XOP__
 866             vfeps            = _mm_frcz_ps(rt);
 867 #else
 868             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 869 #endif
 870             twovfeps         = _mm_add_ps(vfeps,vfeps);
 871             vfitab           = _mm_slli_epi32(vfitab,2);
 872
 873             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 874             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 875             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 876             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 877             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 878             _MM_TRANSPOSE4_PS(Y,F,G,H);
 879             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 880             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 881             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq20,FF),_mm_mul_ps(vftabscale,rinv20)));
 882
 883             fscal            = felec;
 884
 885              /* Update vectorial force */
 886             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 887             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 888             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 889
 890             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 891             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 892             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 893
 894             fjptrA             = f+j_coord_offsetA;
 895             fjptrB             = f+j_coord_offsetB;
 896             fjptrC             = f+j_coord_offsetC;
 897             fjptrD             = f+j_coord_offsetD;
 898
 899             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 900
 901             /* Inner loop uses 126 flops */
 902         }
 903
 904         if(jidx<j_index_end)
 905         {
 906
 907             /* Get j neighbor index, and coordinate index */
 908             jnrlistA         = jjnr[jidx];
 909             jnrlistB         = jjnr[jidx+1];
 910             jnrlistC         = jjnr[jidx+2];
 911             jnrlistD         = jjnr[jidx+3];
 912             /* Sign of each element will be negative for non-real atoms.
 913              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 914              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 915              */
 916             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 917             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 918             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 919             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 920             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 921             j_coord_offsetA  = DIM*jnrA;
 922             j_coord_offsetB  = DIM*jnrB;
 923             j_coord_offsetC  = DIM*jnrC;
 924             j_coord_offsetD  = DIM*jnrD;
 925
 926             /* load j atom coordinates */
 927             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 928                                               x+j_coord_offsetC,x+j_coord_offsetD,
 929                                               &jx0,&jy0,&jz0);
 930
 931             /* Calculate displacement vector */
 932             dx00             = _mm_sub_ps(ix0,jx0);
 933             dy00             = _mm_sub_ps(iy0,jy0);
 934             dz00             = _mm_sub_ps(iz0,jz0);
 935             dx10             = _mm_sub_ps(ix1,jx0);
 936             dy10             = _mm_sub_ps(iy1,jy0);
 937             dz10             = _mm_sub_ps(iz1,jz0);
 938             dx20             = _mm_sub_ps(ix2,jx0);
 939             dy20             = _mm_sub_ps(iy2,jy0);
 940             dz20             = _mm_sub_ps(iz2,jz0);
 941
 942             /* Calculate squared distance and things based on it */
 943             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 944             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 945             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 946
 947             rinv00           = avx128fma_invsqrt_f(rsq00);
 948             rinv10           = avx128fma_invsqrt_f(rsq10);
 949             rinv20           = avx128fma_invsqrt_f(rsq20);
 950
 951             /* Load parameters for j particles */
 952             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 953                                                               charge+jnrC+0,charge+jnrD+0);
 954
 955             fjx0             = _mm_setzero_ps();
 956             fjy0             = _mm_setzero_ps();
 957             fjz0             = _mm_setzero_ps();
 958
 959             /**************************
 960              * CALCULATE INTERACTIONS *
 961              **************************/
 962
 963             r00              = _mm_mul_ps(rsq00,rinv00);
 964             r00              = _mm_andnot_ps(dummy_mask,r00);
 965
 966             /* Compute parameters for interactions between i and j atoms */
 967             qq00             = _mm_mul_ps(iq0,jq0);
 968
 969             /* Calculate table index by multiplying r with table scale and truncate to integer */
 970             rt               = _mm_mul_ps(r00,vftabscale);
 971             vfitab           = _mm_cvttps_epi32(rt);
 972 #ifdef __XOP__
 973             vfeps            = _mm_frcz_ps(rt);
 974 #else
 975             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 976 #endif
 977             twovfeps         = _mm_add_ps(vfeps,vfeps);
 978             vfitab           = _mm_slli_epi32(vfitab,2);
 979
 980             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 981             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 982             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 983             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 984             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 985             _MM_TRANSPOSE4_PS(Y,F,G,H);
 986             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 987             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 988             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 989
 990             fscal            = felec;
 991
 992             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 993
 994              /* Update vectorial force */
 995             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 996             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 997             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 998
 999             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1000             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1001             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1002
1003             /**************************
1004              * CALCULATE INTERACTIONS *
1005              **************************/
1006
1007             r10              = _mm_mul_ps(rsq10,rinv10);
1008             r10              = _mm_andnot_ps(dummy_mask,r10);
1009
1010             /* Compute parameters for interactions between i and j atoms */
1011             qq10             = _mm_mul_ps(iq1,jq0);
1012
1013             /* Calculate table index by multiplying r with table scale and truncate to integer */
1014             rt               = _mm_mul_ps(r10,vftabscale);
1015             vfitab           = _mm_cvttps_epi32(rt);
1016 #ifdef __XOP__
1017             vfeps            = _mm_frcz_ps(rt);
1018 #else
1019             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1020 #endif
1021             twovfeps         = _mm_add_ps(vfeps,vfeps);
1022             vfitab           = _mm_slli_epi32(vfitab,2);
1023
1024             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1025             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1026             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1027             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1028             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1029             _MM_TRANSPOSE4_PS(Y,F,G,H);
1030             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1031             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1032             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq10,FF),_mm_mul_ps(vftabscale,rinv10)));
1033
1034             fscal            = felec;
1035
1036             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1037
1038              /* Update vectorial force */
1039             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1040             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1041             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1042
1043             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1044             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1045             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1046
1047             /**************************
1048              * CALCULATE INTERACTIONS *
1049              **************************/
1050
1051             r20              = _mm_mul_ps(rsq20,rinv20);
1052             r20              = _mm_andnot_ps(dummy_mask,r20);
1053
1054             /* Compute parameters for interactions between i and j atoms */
1055             qq20             = _mm_mul_ps(iq2,jq0);
1056
1057             /* Calculate table index by multiplying r with table scale and truncate to integer */
1058             rt               = _mm_mul_ps(r20,vftabscale);
1059             vfitab           = _mm_cvttps_epi32(rt);
1060 #ifdef __XOP__
1061             vfeps            = _mm_frcz_ps(rt);
1062 #else
1063             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1064 #endif
1065             twovfeps         = _mm_add_ps(vfeps,vfeps);
1066             vfitab           = _mm_slli_epi32(vfitab,2);
1067
1068             /* CUBIC SPLINE TABLE ELECTROSTATICS */
1069             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1070             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1071             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1072             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1073             _MM_TRANSPOSE4_PS(Y,F,G,H);
1074             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1075             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1076             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq20,FF),_mm_mul_ps(vftabscale,rinv20)));
1077
1078             fscal            = felec;
1079
1080             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1081
1082              /* Update vectorial force */
1083             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1084             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1085             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1086
1087             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1088             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1089             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1090
1091             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1092             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1093             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1094             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1095
1096             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1097
1098             /* Inner loop uses 129 flops */
1099         }
1100
1101         /* End of innermost loop */
1102
1103         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1104                                               f+i_coord_offset,fshift+i_shift_offset);
1105
1106         /* Increment number of inner iterations */
1107         inneriter                  += j_index_end - j_index_start;
1108
1109         /* Outer loop uses 18 flops */
1110     }
1111
1112     /* Increment number of outer iterations */
1113     outeriter        += nri;
1114
1115     /* Update outer/inner flops */
1116
1117     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_F,outeriter*18 + inneriter*129);
1118 }