src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_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_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_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              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  85     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  94     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  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     krf              = _mm_set1_ps(fr->ic->k_rf);
 117     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 118     crf              = _mm_set1_ps(fr->ic->c_rf);
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_vdw->data;
 124     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 125
 126     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 127     rcutoff_scalar   = fr->ic->rcoulomb;
 128     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 129     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 130
 131     /* Avoid stupid compiler warnings */
 132     jnrA = jnrB = jnrC = jnrD = 0;
 133     j_coord_offsetA = 0;
 134     j_coord_offsetB = 0;
 135     j_coord_offsetC = 0;
 136     j_coord_offsetD = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     for(iidx=0;iidx<4*DIM;iidx++)
 142     {
 143         scratch[iidx] = 0.0;
 144     }
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 162
 163         fix0             = _mm_setzero_ps();
 164         fiy0             = _mm_setzero_ps();
 165         fiz0             = _mm_setzero_ps();
 166
 167         /* Load parameters for i particles */
 168         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 169         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 170
 171         /* Reset potential sums */
 172         velecsum         = _mm_setzero_ps();
 173         vvdwsum          = _mm_setzero_ps();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188
 189             /* load j atom coordinates */
 190             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                               x+j_coord_offsetC,x+j_coord_offsetD,
 192                                               &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm_sub_ps(ix0,jx0);
 196             dy00             = _mm_sub_ps(iy0,jy0);
 197             dz00             = _mm_sub_ps(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 201
 202             rinv00           = avx128fma_invsqrt_f(rsq00);
 203
 204             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 208                                                               charge+jnrC+0,charge+jnrD+0);
 209             vdwjidx0A        = 2*vdwtype[jnrA+0];
 210             vdwjidx0B        = 2*vdwtype[jnrB+0];
 211             vdwjidx0C        = 2*vdwtype[jnrC+0];
 212             vdwjidx0D        = 2*vdwtype[jnrD+0];
 213
 214             /**************************
 215              * CALCULATE INTERACTIONS *
 216              **************************/
 217
 218             if (gmx_mm_any_lt(rsq00,rcutoff2))
 219             {
 220
 221             r00              = _mm_mul_ps(rsq00,rinv00);
 222
 223             /* Compute parameters for interactions between i and j atoms */
 224             qq00             = _mm_mul_ps(iq0,jq0);
 225             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 226                                          vdwparam+vdwioffset0+vdwjidx0B,
 227                                          vdwparam+vdwioffset0+vdwjidx0C,
 228                                          vdwparam+vdwioffset0+vdwjidx0D,
 229                                          &c6_00,&c12_00);
 230
 231             /* Calculate table index by multiplying r with table scale and truncate to integer */
 232             rt               = _mm_mul_ps(r00,vftabscale);
 233             vfitab           = _mm_cvttps_epi32(rt);
 234 #ifdef __XOP__
 235             vfeps            = _mm_frcz_ps(rt);
 236 #else
 237             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 238 #endif
 239             twovfeps         = _mm_add_ps(vfeps,vfeps);
 240             vfitab           = _mm_slli_epi32(vfitab,3);
 241
 242             /* REACTION-FIELD ELECTROSTATICS */
 243             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 244             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 245
 246             /* CUBIC SPLINE TABLE DISPERSION */
 247             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 248             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 249             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 250             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 251             _MM_TRANSPOSE4_PS(Y,F,G,H);
 252             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 253             VV               = _mm_macc_ps(vfeps,Fp,Y);
 254             vvdw6            = _mm_mul_ps(c6_00,VV);
 255             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 256             fvdw6            = _mm_mul_ps(c6_00,FF);
 257
 258             /* CUBIC SPLINE TABLE REPULSION */
 259             vfitab           = _mm_add_epi32(vfitab,ifour);
 260             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 261             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 262             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 263             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 264             _MM_TRANSPOSE4_PS(Y,F,G,H);
 265             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 266             VV               = _mm_macc_ps(vfeps,Fp,Y);
 267             vvdw12           = _mm_mul_ps(c12_00,VV);
 268             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 269             fvdw12           = _mm_mul_ps(c12_00,FF);
 270             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 271             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 272
 273             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 274
 275             /* Update potential sum for this i atom from the interaction with this j atom. */
 276             velec            = _mm_and_ps(velec,cutoff_mask);
 277             velecsum         = _mm_add_ps(velecsum,velec);
 278             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 279             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 280
 281             fscal            = _mm_add_ps(felec,fvdw);
 282
 283             fscal            = _mm_and_ps(fscal,cutoff_mask);
 284
 285              /* Update vectorial force */
 286             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 287             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 288             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 289
 290             fjptrA             = f+j_coord_offsetA;
 291             fjptrB             = f+j_coord_offsetB;
 292             fjptrC             = f+j_coord_offsetC;
 293             fjptrD             = f+j_coord_offsetD;
 294             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 295                                                    _mm_mul_ps(dx00,fscal),
 296                                                    _mm_mul_ps(dy00,fscal),
 297                                                    _mm_mul_ps(dz00,fscal));
 298
 299             }
 300
 301             /* Inner loop uses 75 flops */
 302         }
 303
 304         if(jidx<j_index_end)
 305         {
 306
 307             /* Get j neighbor index, and coordinate index */
 308             jnrlistA         = jjnr[jidx];
 309             jnrlistB         = jjnr[jidx+1];
 310             jnrlistC         = jjnr[jidx+2];
 311             jnrlistD         = jjnr[jidx+3];
 312             /* Sign of each element will be negative for non-real atoms.
 313              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 314              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 315              */
 316             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 317             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 318             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 319             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 320             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 321             j_coord_offsetA  = DIM*jnrA;
 322             j_coord_offsetB  = DIM*jnrB;
 323             j_coord_offsetC  = DIM*jnrC;
 324             j_coord_offsetD  = DIM*jnrD;
 325
 326             /* load j atom coordinates */
 327             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 328                                               x+j_coord_offsetC,x+j_coord_offsetD,
 329                                               &jx0,&jy0,&jz0);
 330
 331             /* Calculate displacement vector */
 332             dx00             = _mm_sub_ps(ix0,jx0);
 333             dy00             = _mm_sub_ps(iy0,jy0);
 334             dz00             = _mm_sub_ps(iz0,jz0);
 335
 336             /* Calculate squared distance and things based on it */
 337             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 338
 339             rinv00           = avx128fma_invsqrt_f(rsq00);
 340
 341             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 342
 343             /* Load parameters for j particles */
 344             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 345                                                               charge+jnrC+0,charge+jnrD+0);
 346             vdwjidx0A        = 2*vdwtype[jnrA+0];
 347             vdwjidx0B        = 2*vdwtype[jnrB+0];
 348             vdwjidx0C        = 2*vdwtype[jnrC+0];
 349             vdwjidx0D        = 2*vdwtype[jnrD+0];
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             if (gmx_mm_any_lt(rsq00,rcutoff2))
 356             {
 357
 358             r00              = _mm_mul_ps(rsq00,rinv00);
 359             r00              = _mm_andnot_ps(dummy_mask,r00);
 360
 361             /* Compute parameters for interactions between i and j atoms */
 362             qq00             = _mm_mul_ps(iq0,jq0);
 363             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 364                                          vdwparam+vdwioffset0+vdwjidx0B,
 365                                          vdwparam+vdwioffset0+vdwjidx0C,
 366                                          vdwparam+vdwioffset0+vdwjidx0D,
 367                                          &c6_00,&c12_00);
 368
 369             /* Calculate table index by multiplying r with table scale and truncate to integer */
 370             rt               = _mm_mul_ps(r00,vftabscale);
 371             vfitab           = _mm_cvttps_epi32(rt);
 372 #ifdef __XOP__
 373             vfeps            = _mm_frcz_ps(rt);
 374 #else
 375             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 376 #endif
 377             twovfeps         = _mm_add_ps(vfeps,vfeps);
 378             vfitab           = _mm_slli_epi32(vfitab,3);
 379
 380             /* REACTION-FIELD ELECTROSTATICS */
 381             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 382             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 383
 384             /* CUBIC SPLINE TABLE DISPERSION */
 385             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 386             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 387             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 388             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 389             _MM_TRANSPOSE4_PS(Y,F,G,H);
 390             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 391             VV               = _mm_macc_ps(vfeps,Fp,Y);
 392             vvdw6            = _mm_mul_ps(c6_00,VV);
 393             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 394             fvdw6            = _mm_mul_ps(c6_00,FF);
 395
 396             /* CUBIC SPLINE TABLE REPULSION */
 397             vfitab           = _mm_add_epi32(vfitab,ifour);
 398             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 399             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 400             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 401             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 402             _MM_TRANSPOSE4_PS(Y,F,G,H);
 403             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 404             VV               = _mm_macc_ps(vfeps,Fp,Y);
 405             vvdw12           = _mm_mul_ps(c12_00,VV);
 406             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 407             fvdw12           = _mm_mul_ps(c12_00,FF);
 408             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 409             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 410
 411             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 412
 413             /* Update potential sum for this i atom from the interaction with this j atom. */
 414             velec            = _mm_and_ps(velec,cutoff_mask);
 415             velec            = _mm_andnot_ps(dummy_mask,velec);
 416             velecsum         = _mm_add_ps(velecsum,velec);
 417             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 418             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 419             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 420
 421             fscal            = _mm_add_ps(felec,fvdw);
 422
 423             fscal            = _mm_and_ps(fscal,cutoff_mask);
 424
 425             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 426
 427              /* Update vectorial force */
 428             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 429             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 430             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 431
 432             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 433             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 434             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 435             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 436             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 437                                                    _mm_mul_ps(dx00,fscal),
 438                                                    _mm_mul_ps(dy00,fscal),
 439                                                    _mm_mul_ps(dz00,fscal));
 440
 441             }
 442
 443             /* Inner loop uses 76 flops */
 444         }
 445
 446         /* End of innermost loop */
 447
 448         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 449                                               f+i_coord_offset,fshift+i_shift_offset);
 450
 451         ggid                        = gid[iidx];
 452         /* Update potential energies */
 453         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 454         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 455
 456         /* Increment number of inner iterations */
 457         inneriter                  += j_index_end - j_index_start;
 458
 459         /* Outer loop uses 9 flops */
 460     }
 461
 462     /* Increment number of outer iterations */
 463     outeriter        += nri;
 464
 465     /* Update outer/inner flops */
 466
 467     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
 468 }
 469 /*
 470  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 471  * Electrostatics interaction: ReactionField
 472  * VdW interaction:            CubicSplineTable
 473  * Geometry:                   Particle-Particle
 474  * Calculate force/pot:        Force
 475  */
 476 void
 477 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 478                     (t_nblist                    * gmx_restrict       nlist,
 479                      rvec                        * gmx_restrict          xx,
 480                      rvec                        * gmx_restrict          ff,
 481                      struct t_forcerec           * gmx_restrict          fr,
 482                      t_mdatoms                   * gmx_restrict     mdatoms,
 483                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 484                      t_nrnb                      * gmx_restrict        nrnb)
 485 {
 486     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 487      * just 0 for non-waters.
 488      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 489      * jnr indices corresponding to data put in the four positions in the SIMD register.
 490      */
 491     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 492     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 493     int              jnrA,jnrB,jnrC,jnrD;
 494     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 495     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 496     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 497     real             rcutoff_scalar;
 498     real             *shiftvec,*fshift,*x,*f;
 499     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 500     real             scratch[4*DIM];
 501     __m128           fscal,rcutoff,rcutoff2,jidxall;
 502     int              vdwioffset0;
 503     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 504     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 505     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 506     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 507     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 508     real             *charge;
 509     int              nvdwtype;
 510     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 511     int              *vdwtype;
 512     real             *vdwparam;
 513     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 514     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 515     __m128i          vfitab;
 516     __m128i          ifour       = _mm_set1_epi32(4);
 517     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 518     real             *vftab;
 519     __m128           dummy_mask,cutoff_mask;
 520     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 521     __m128           one     = _mm_set1_ps(1.0);
 522     __m128           two     = _mm_set1_ps(2.0);
 523     x                = xx[0];
 524     f                = ff[0];
 525
 526     nri              = nlist->nri;
 527     iinr             = nlist->iinr;
 528     jindex           = nlist->jindex;
 529     jjnr             = nlist->jjnr;
 530     shiftidx         = nlist->shift;
 531     gid              = nlist->gid;
 532     shiftvec         = fr->shift_vec[0];
 533     fshift           = fr->fshift[0];
 534     facel            = _mm_set1_ps(fr->ic->epsfac);
 535     charge           = mdatoms->chargeA;
 536     krf              = _mm_set1_ps(fr->ic->k_rf);
 537     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 538     crf              = _mm_set1_ps(fr->ic->c_rf);
 539     nvdwtype         = fr->ntype;
 540     vdwparam         = fr->nbfp;
 541     vdwtype          = mdatoms->typeA;
 542
 543     vftab            = kernel_data->table_vdw->data;
 544     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 545
 546     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 547     rcutoff_scalar   = fr->ic->rcoulomb;
 548     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 549     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 550
 551     /* Avoid stupid compiler warnings */
 552     jnrA = jnrB = jnrC = jnrD = 0;
 553     j_coord_offsetA = 0;
 554     j_coord_offsetB = 0;
 555     j_coord_offsetC = 0;
 556     j_coord_offsetD = 0;
 557
 558     outeriter        = 0;
 559     inneriter        = 0;
 560
 561     for(iidx=0;iidx<4*DIM;iidx++)
 562     {
 563         scratch[iidx] = 0.0;
 564     }
 565
 566     /* Start outer loop over neighborlists */
 567     for(iidx=0; iidx<nri; iidx++)
 568     {
 569         /* Load shift vector for this list */
 570         i_shift_offset   = DIM*shiftidx[iidx];
 571
 572         /* Load limits for loop over neighbors */
 573         j_index_start    = jindex[iidx];
 574         j_index_end      = jindex[iidx+1];
 575
 576         /* Get outer coordinate index */
 577         inr              = iinr[iidx];
 578         i_coord_offset   = DIM*inr;
 579
 580         /* Load i particle coords and add shift vector */
 581         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 582
 583         fix0             = _mm_setzero_ps();
 584         fiy0             = _mm_setzero_ps();
 585         fiz0             = _mm_setzero_ps();
 586
 587         /* Load parameters for i particles */
 588         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 589         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 590
 591         /* Start inner kernel loop */
 592         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 593         {
 594
 595             /* Get j neighbor index, and coordinate index */
 596             jnrA             = jjnr[jidx];
 597             jnrB             = jjnr[jidx+1];
 598             jnrC             = jjnr[jidx+2];
 599             jnrD             = jjnr[jidx+3];
 600             j_coord_offsetA  = DIM*jnrA;
 601             j_coord_offsetB  = DIM*jnrB;
 602             j_coord_offsetC  = DIM*jnrC;
 603             j_coord_offsetD  = DIM*jnrD;
 604
 605             /* load j atom coordinates */
 606             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 607                                               x+j_coord_offsetC,x+j_coord_offsetD,
 608                                               &jx0,&jy0,&jz0);
 609
 610             /* Calculate displacement vector */
 611             dx00             = _mm_sub_ps(ix0,jx0);
 612             dy00             = _mm_sub_ps(iy0,jy0);
 613             dz00             = _mm_sub_ps(iz0,jz0);
 614
 615             /* Calculate squared distance and things based on it */
 616             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 617
 618             rinv00           = avx128fma_invsqrt_f(rsq00);
 619
 620             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 621
 622             /* Load parameters for j particles */
 623             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 624                                                               charge+jnrC+0,charge+jnrD+0);
 625             vdwjidx0A        = 2*vdwtype[jnrA+0];
 626             vdwjidx0B        = 2*vdwtype[jnrB+0];
 627             vdwjidx0C        = 2*vdwtype[jnrC+0];
 628             vdwjidx0D        = 2*vdwtype[jnrD+0];
 629
 630             /**************************
 631              * CALCULATE INTERACTIONS *
 632              **************************/
 633
 634             if (gmx_mm_any_lt(rsq00,rcutoff2))
 635             {
 636
 637             r00              = _mm_mul_ps(rsq00,rinv00);
 638
 639             /* Compute parameters for interactions between i and j atoms */
 640             qq00             = _mm_mul_ps(iq0,jq0);
 641             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 642                                          vdwparam+vdwioffset0+vdwjidx0B,
 643                                          vdwparam+vdwioffset0+vdwjidx0C,
 644                                          vdwparam+vdwioffset0+vdwjidx0D,
 645                                          &c6_00,&c12_00);
 646
 647             /* Calculate table index by multiplying r with table scale and truncate to integer */
 648             rt               = _mm_mul_ps(r00,vftabscale);
 649             vfitab           = _mm_cvttps_epi32(rt);
 650 #ifdef __XOP__
 651             vfeps            = _mm_frcz_ps(rt);
 652 #else
 653             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 654 #endif
 655             twovfeps         = _mm_add_ps(vfeps,vfeps);
 656             vfitab           = _mm_slli_epi32(vfitab,3);
 657
 658             /* REACTION-FIELD ELECTROSTATICS */
 659             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 660
 661             /* CUBIC SPLINE TABLE DISPERSION */
 662             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 663             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 664             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 665             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 666             _MM_TRANSPOSE4_PS(Y,F,G,H);
 667             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 668             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 669             fvdw6            = _mm_mul_ps(c6_00,FF);
 670
 671             /* CUBIC SPLINE TABLE REPULSION */
 672             vfitab           = _mm_add_epi32(vfitab,ifour);
 673             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 674             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 675             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 676             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 677             _MM_TRANSPOSE4_PS(Y,F,G,H);
 678             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 679             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 680             fvdw12           = _mm_mul_ps(c12_00,FF);
 681             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 682
 683             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 684
 685             fscal            = _mm_add_ps(felec,fvdw);
 686
 687             fscal            = _mm_and_ps(fscal,cutoff_mask);
 688
 689              /* Update vectorial force */
 690             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 691             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 692             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 693
 694             fjptrA             = f+j_coord_offsetA;
 695             fjptrB             = f+j_coord_offsetB;
 696             fjptrC             = f+j_coord_offsetC;
 697             fjptrD             = f+j_coord_offsetD;
 698             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 699                                                    _mm_mul_ps(dx00,fscal),
 700                                                    _mm_mul_ps(dy00,fscal),
 701                                                    _mm_mul_ps(dz00,fscal));
 702
 703             }
 704
 705             /* Inner loop uses 60 flops */
 706         }
 707
 708         if(jidx<j_index_end)
 709         {
 710
 711             /* Get j neighbor index, and coordinate index */
 712             jnrlistA         = jjnr[jidx];
 713             jnrlistB         = jjnr[jidx+1];
 714             jnrlistC         = jjnr[jidx+2];
 715             jnrlistD         = jjnr[jidx+3];
 716             /* Sign of each element will be negative for non-real atoms.
 717              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 718              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 719              */
 720             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 721             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 722             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 723             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 724             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 725             j_coord_offsetA  = DIM*jnrA;
 726             j_coord_offsetB  = DIM*jnrB;
 727             j_coord_offsetC  = DIM*jnrC;
 728             j_coord_offsetD  = DIM*jnrD;
 729
 730             /* load j atom coordinates */
 731             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 732                                               x+j_coord_offsetC,x+j_coord_offsetD,
 733                                               &jx0,&jy0,&jz0);
 734
 735             /* Calculate displacement vector */
 736             dx00             = _mm_sub_ps(ix0,jx0);
 737             dy00             = _mm_sub_ps(iy0,jy0);
 738             dz00             = _mm_sub_ps(iz0,jz0);
 739
 740             /* Calculate squared distance and things based on it */
 741             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 742
 743             rinv00           = avx128fma_invsqrt_f(rsq00);
 744
 745             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 746
 747             /* Load parameters for j particles */
 748             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 749                                                               charge+jnrC+0,charge+jnrD+0);
 750             vdwjidx0A        = 2*vdwtype[jnrA+0];
 751             vdwjidx0B        = 2*vdwtype[jnrB+0];
 752             vdwjidx0C        = 2*vdwtype[jnrC+0];
 753             vdwjidx0D        = 2*vdwtype[jnrD+0];
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             if (gmx_mm_any_lt(rsq00,rcutoff2))
 760             {
 761
 762             r00              = _mm_mul_ps(rsq00,rinv00);
 763             r00              = _mm_andnot_ps(dummy_mask,r00);
 764
 765             /* Compute parameters for interactions between i and j atoms */
 766             qq00             = _mm_mul_ps(iq0,jq0);
 767             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 768                                          vdwparam+vdwioffset0+vdwjidx0B,
 769                                          vdwparam+vdwioffset0+vdwjidx0C,
 770                                          vdwparam+vdwioffset0+vdwjidx0D,
 771                                          &c6_00,&c12_00);
 772
 773             /* Calculate table index by multiplying r with table scale and truncate to integer */
 774             rt               = _mm_mul_ps(r00,vftabscale);
 775             vfitab           = _mm_cvttps_epi32(rt);
 776 #ifdef __XOP__
 777             vfeps            = _mm_frcz_ps(rt);
 778 #else
 779             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 780 #endif
 781             twovfeps         = _mm_add_ps(vfeps,vfeps);
 782             vfitab           = _mm_slli_epi32(vfitab,3);
 783
 784             /* REACTION-FIELD ELECTROSTATICS */
 785             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 786
 787             /* CUBIC SPLINE TABLE DISPERSION */
 788             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 789             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 790             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 791             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 792             _MM_TRANSPOSE4_PS(Y,F,G,H);
 793             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 794             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 795             fvdw6            = _mm_mul_ps(c6_00,FF);
 796
 797             /* CUBIC SPLINE TABLE REPULSION */
 798             vfitab           = _mm_add_epi32(vfitab,ifour);
 799             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 800             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 801             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 802             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 803             _MM_TRANSPOSE4_PS(Y,F,G,H);
 804             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 805             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 806             fvdw12           = _mm_mul_ps(c12_00,FF);
 807             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 808
 809             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 810
 811             fscal            = _mm_add_ps(felec,fvdw);
 812
 813             fscal            = _mm_and_ps(fscal,cutoff_mask);
 814
 815             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 816
 817              /* Update vectorial force */
 818             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 819             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 820             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 821
 822             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 823             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 824             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 825             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 826             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 827                                                    _mm_mul_ps(dx00,fscal),
 828                                                    _mm_mul_ps(dy00,fscal),
 829                                                    _mm_mul_ps(dz00,fscal));
 830
 831             }
 832
 833             /* Inner loop uses 61 flops */
 834         }
 835
 836         /* End of innermost loop */
 837
 838         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 839                                               f+i_coord_offset,fshift+i_shift_offset);
 840
 841         /* Increment number of inner iterations */
 842         inneriter                  += j_index_end - j_index_start;
 843
 844         /* Outer loop uses 7 flops */
 845     }
 846
 847     /* Increment number of outer iterations */
 848     outeriter        += nri;
 849
 850     /* Update outer/inner flops */
 851
 852     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*61);
 853 }