src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_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_ElecEw_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_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     __m128i          ewitab;
 100     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 101     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 102     real             *ewtab;
 103     __m128           dummy_mask,cutoff_mask;
 104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 105     __m128           one     = _mm_set1_ps(1.0);
 106     __m128           two     = _mm_set1_ps(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_ps(fr->ic->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     vftab            = kernel_data->table_vdw->data;
 125     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 126
 127     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 128     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 129     beta2            = _mm_mul_ps(beta,beta);
 130     beta3            = _mm_mul_ps(beta,beta2);
 131     ewtab            = fr->ic->tabq_coul_FDV0;
 132     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 133     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 134
 135     /* Avoid stupid compiler warnings */
 136     jnrA = jnrB = jnrC = jnrD = 0;
 137     j_coord_offsetA = 0;
 138     j_coord_offsetB = 0;
 139     j_coord_offsetC = 0;
 140     j_coord_offsetD = 0;
 141
 142     outeriter        = 0;
 143     inneriter        = 0;
 144
 145     for(iidx=0;iidx<4*DIM;iidx++)
 146     {
 147         scratch[iidx] = 0.0;
 148     }
 149
 150     /* Start outer loop over neighborlists */
 151     for(iidx=0; iidx<nri; iidx++)
 152     {
 153         /* Load shift vector for this list */
 154         i_shift_offset   = DIM*shiftidx[iidx];
 155
 156         /* Load limits for loop over neighbors */
 157         j_index_start    = jindex[iidx];
 158         j_index_end      = jindex[iidx+1];
 159
 160         /* Get outer coordinate index */
 161         inr              = iinr[iidx];
 162         i_coord_offset   = DIM*inr;
 163
 164         /* Load i particle coords and add shift vector */
 165         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 166
 167         fix0             = _mm_setzero_ps();
 168         fiy0             = _mm_setzero_ps();
 169         fiz0             = _mm_setzero_ps();
 170
 171         /* Load parameters for i particles */
 172         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 173         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_ps();
 177         vvdwsum          = _mm_setzero_ps();
 178
 179         /* Start inner kernel loop */
 180         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 181         {
 182
 183             /* Get j neighbor index, and coordinate index */
 184             jnrA             = jjnr[jidx];
 185             jnrB             = jjnr[jidx+1];
 186             jnrC             = jjnr[jidx+2];
 187             jnrD             = jjnr[jidx+3];
 188             j_coord_offsetA  = DIM*jnrA;
 189             j_coord_offsetB  = DIM*jnrB;
 190             j_coord_offsetC  = DIM*jnrC;
 191             j_coord_offsetD  = DIM*jnrD;
 192
 193             /* load j atom coordinates */
 194             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 195                                               x+j_coord_offsetC,x+j_coord_offsetD,
 196                                               &jx0,&jy0,&jz0);
 197
 198             /* Calculate displacement vector */
 199             dx00             = _mm_sub_ps(ix0,jx0);
 200             dy00             = _mm_sub_ps(iy0,jy0);
 201             dz00             = _mm_sub_ps(iz0,jz0);
 202
 203             /* Calculate squared distance and things based on it */
 204             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 205
 206             rinv00           = avx128fma_invsqrt_f(rsq00);
 207
 208             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 209
 210             /* Load parameters for j particles */
 211             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 212                                                               charge+jnrC+0,charge+jnrD+0);
 213             vdwjidx0A        = 2*vdwtype[jnrA+0];
 214             vdwjidx0B        = 2*vdwtype[jnrB+0];
 215             vdwjidx0C        = 2*vdwtype[jnrC+0];
 216             vdwjidx0D        = 2*vdwtype[jnrD+0];
 217
 218             /**************************
 219              * CALCULATE INTERACTIONS *
 220              **************************/
 221
 222             r00              = _mm_mul_ps(rsq00,rinv00);
 223
 224             /* Compute parameters for interactions between i and j atoms */
 225             qq00             = _mm_mul_ps(iq0,jq0);
 226             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 227                                          vdwparam+vdwioffset0+vdwjidx0B,
 228                                          vdwparam+vdwioffset0+vdwjidx0C,
 229                                          vdwparam+vdwioffset0+vdwjidx0D,
 230                                          &c6_00,&c12_00);
 231
 232             /* Calculate table index by multiplying r with table scale and truncate to integer */
 233             rt               = _mm_mul_ps(r00,vftabscale);
 234             vfitab           = _mm_cvttps_epi32(rt);
 235 #ifdef __XOP__
 236             vfeps            = _mm_frcz_ps(rt);
 237 #else
 238             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 239 #endif
 240             twovfeps         = _mm_add_ps(vfeps,vfeps);
 241             vfitab           = _mm_slli_epi32(vfitab,3);
 242
 243             /* EWALD ELECTROSTATICS */
 244
 245             /* Analytical PME correction */
 246             zeta2            = _mm_mul_ps(beta2,rsq00);
 247             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 248             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 249             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 250             felec            = _mm_mul_ps(qq00,felec);
 251             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 252             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 253             velec            = _mm_mul_ps(qq00,velec);
 254
 255             /* CUBIC SPLINE TABLE DISPERSION */
 256             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 257             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 258             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 259             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 260             _MM_TRANSPOSE4_PS(Y,F,G,H);
 261             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 262             VV               = _mm_macc_ps(vfeps,Fp,Y);
 263             vvdw6            = _mm_mul_ps(c6_00,VV);
 264             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 265             fvdw6            = _mm_mul_ps(c6_00,FF);
 266
 267             /* CUBIC SPLINE TABLE REPULSION */
 268             vfitab           = _mm_add_epi32(vfitab,ifour);
 269             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 270             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 271             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 272             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 273             _MM_TRANSPOSE4_PS(Y,F,G,H);
 274             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 275             VV               = _mm_macc_ps(vfeps,Fp,Y);
 276             vvdw12           = _mm_mul_ps(c12_00,VV);
 277             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 278             fvdw12           = _mm_mul_ps(c12_00,FF);
 279             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 280             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 281
 282             /* Update potential sum for this i atom from the interaction with this j atom. */
 283             velecsum         = _mm_add_ps(velecsum,velec);
 284             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 285
 286             fscal            = _mm_add_ps(felec,fvdw);
 287
 288              /* Update vectorial force */
 289             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 290             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 291             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 292
 293             fjptrA             = f+j_coord_offsetA;
 294             fjptrB             = f+j_coord_offsetB;
 295             fjptrC             = f+j_coord_offsetC;
 296             fjptrD             = f+j_coord_offsetD;
 297             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 298                                                    _mm_mul_ps(dx00,fscal),
 299                                                    _mm_mul_ps(dy00,fscal),
 300                                                    _mm_mul_ps(dz00,fscal));
 301
 302             /* Inner loop uses 63 flops */
 303         }
 304
 305         if(jidx<j_index_end)
 306         {
 307
 308             /* Get j neighbor index, and coordinate index */
 309             jnrlistA         = jjnr[jidx];
 310             jnrlistB         = jjnr[jidx+1];
 311             jnrlistC         = jjnr[jidx+2];
 312             jnrlistD         = jjnr[jidx+3];
 313             /* Sign of each element will be negative for non-real atoms.
 314              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 315              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 316              */
 317             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 318             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 319             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 320             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 321             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 322             j_coord_offsetA  = DIM*jnrA;
 323             j_coord_offsetB  = DIM*jnrB;
 324             j_coord_offsetC  = DIM*jnrC;
 325             j_coord_offsetD  = DIM*jnrD;
 326
 327             /* load j atom coordinates */
 328             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 329                                               x+j_coord_offsetC,x+j_coord_offsetD,
 330                                               &jx0,&jy0,&jz0);
 331
 332             /* Calculate displacement vector */
 333             dx00             = _mm_sub_ps(ix0,jx0);
 334             dy00             = _mm_sub_ps(iy0,jy0);
 335             dz00             = _mm_sub_ps(iz0,jz0);
 336
 337             /* Calculate squared distance and things based on it */
 338             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 339
 340             rinv00           = avx128fma_invsqrt_f(rsq00);
 341
 342             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 343
 344             /* Load parameters for j particles */
 345             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 346                                                               charge+jnrC+0,charge+jnrD+0);
 347             vdwjidx0A        = 2*vdwtype[jnrA+0];
 348             vdwjidx0B        = 2*vdwtype[jnrB+0];
 349             vdwjidx0C        = 2*vdwtype[jnrC+0];
 350             vdwjidx0D        = 2*vdwtype[jnrD+0];
 351
 352             /**************************
 353              * CALCULATE INTERACTIONS *
 354              **************************/
 355
 356             r00              = _mm_mul_ps(rsq00,rinv00);
 357             r00              = _mm_andnot_ps(dummy_mask,r00);
 358
 359             /* Compute parameters for interactions between i and j atoms */
 360             qq00             = _mm_mul_ps(iq0,jq0);
 361             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 362                                          vdwparam+vdwioffset0+vdwjidx0B,
 363                                          vdwparam+vdwioffset0+vdwjidx0C,
 364                                          vdwparam+vdwioffset0+vdwjidx0D,
 365                                          &c6_00,&c12_00);
 366
 367             /* Calculate table index by multiplying r with table scale and truncate to integer */
 368             rt               = _mm_mul_ps(r00,vftabscale);
 369             vfitab           = _mm_cvttps_epi32(rt);
 370 #ifdef __XOP__
 371             vfeps            = _mm_frcz_ps(rt);
 372 #else
 373             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 374 #endif
 375             twovfeps         = _mm_add_ps(vfeps,vfeps);
 376             vfitab           = _mm_slli_epi32(vfitab,3);
 377
 378             /* EWALD ELECTROSTATICS */
 379
 380             /* Analytical PME correction */
 381             zeta2            = _mm_mul_ps(beta2,rsq00);
 382             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 383             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 384             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 385             felec            = _mm_mul_ps(qq00,felec);
 386             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 387             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 388             velec            = _mm_mul_ps(qq00,velec);
 389
 390             /* CUBIC SPLINE TABLE DISPERSION */
 391             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 392             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 393             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 394             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 395             _MM_TRANSPOSE4_PS(Y,F,G,H);
 396             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 397             VV               = _mm_macc_ps(vfeps,Fp,Y);
 398             vvdw6            = _mm_mul_ps(c6_00,VV);
 399             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 400             fvdw6            = _mm_mul_ps(c6_00,FF);
 401
 402             /* CUBIC SPLINE TABLE REPULSION */
 403             vfitab           = _mm_add_epi32(vfitab,ifour);
 404             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 405             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 406             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 407             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 408             _MM_TRANSPOSE4_PS(Y,F,G,H);
 409             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 410             VV               = _mm_macc_ps(vfeps,Fp,Y);
 411             vvdw12           = _mm_mul_ps(c12_00,VV);
 412             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 413             fvdw12           = _mm_mul_ps(c12_00,FF);
 414             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 415             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velec            = _mm_andnot_ps(dummy_mask,velec);
 419             velecsum         = _mm_add_ps(velecsum,velec);
 420             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 421             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 422
 423             fscal            = _mm_add_ps(felec,fvdw);
 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             /* Inner loop uses 64 flops */
 442         }
 443
 444         /* End of innermost loop */
 445
 446         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 447                                               f+i_coord_offset,fshift+i_shift_offset);
 448
 449         ggid                        = gid[iidx];
 450         /* Update potential energies */
 451         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 452         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 453
 454         /* Increment number of inner iterations */
 455         inneriter                  += j_index_end - j_index_start;
 456
 457         /* Outer loop uses 9 flops */
 458     }
 459
 460     /* Increment number of outer iterations */
 461     outeriter        += nri;
 462
 463     /* Update outer/inner flops */
 464
 465     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*64);
 466 }
 467 /*
 468  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 469  * Electrostatics interaction: Ewald
 470  * VdW interaction:            CubicSplineTable
 471  * Geometry:                   Particle-Particle
 472  * Calculate force/pot:        Force
 473  */
 474 void
 475 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 476                     (t_nblist                    * gmx_restrict       nlist,
 477                      rvec                        * gmx_restrict          xx,
 478                      rvec                        * gmx_restrict          ff,
 479                      struct t_forcerec           * gmx_restrict          fr,
 480                      t_mdatoms                   * gmx_restrict     mdatoms,
 481                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 482                      t_nrnb                      * gmx_restrict        nrnb)
 483 {
 484     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 485      * just 0 for non-waters.
 486      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 487      * jnr indices corresponding to data put in the four positions in the SIMD register.
 488      */
 489     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 490     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 491     int              jnrA,jnrB,jnrC,jnrD;
 492     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 493     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 494     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 495     real             rcutoff_scalar;
 496     real             *shiftvec,*fshift,*x,*f;
 497     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 498     real             scratch[4*DIM];
 499     __m128           fscal,rcutoff,rcutoff2,jidxall;
 500     int              vdwioffset0;
 501     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 502     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 503     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 504     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 505     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 506     real             *charge;
 507     int              nvdwtype;
 508     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 509     int              *vdwtype;
 510     real             *vdwparam;
 511     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 512     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 513     __m128i          vfitab;
 514     __m128i          ifour       = _mm_set1_epi32(4);
 515     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 516     real             *vftab;
 517     __m128i          ewitab;
 518     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 519     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 520     real             *ewtab;
 521     __m128           dummy_mask,cutoff_mask;
 522     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 523     __m128           one     = _mm_set1_ps(1.0);
 524     __m128           two     = _mm_set1_ps(2.0);
 525     x                = xx[0];
 526     f                = ff[0];
 527
 528     nri              = nlist->nri;
 529     iinr             = nlist->iinr;
 530     jindex           = nlist->jindex;
 531     jjnr             = nlist->jjnr;
 532     shiftidx         = nlist->shift;
 533     gid              = nlist->gid;
 534     shiftvec         = fr->shift_vec[0];
 535     fshift           = fr->fshift[0];
 536     facel            = _mm_set1_ps(fr->ic->epsfac);
 537     charge           = mdatoms->chargeA;
 538     nvdwtype         = fr->ntype;
 539     vdwparam         = fr->nbfp;
 540     vdwtype          = mdatoms->typeA;
 541
 542     vftab            = kernel_data->table_vdw->data;
 543     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 544
 545     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 546     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 547     beta2            = _mm_mul_ps(beta,beta);
 548     beta3            = _mm_mul_ps(beta,beta2);
 549     ewtab            = fr->ic->tabq_coul_F;
 550     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 551     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 552
 553     /* Avoid stupid compiler warnings */
 554     jnrA = jnrB = jnrC = jnrD = 0;
 555     j_coord_offsetA = 0;
 556     j_coord_offsetB = 0;
 557     j_coord_offsetC = 0;
 558     j_coord_offsetD = 0;
 559
 560     outeriter        = 0;
 561     inneriter        = 0;
 562
 563     for(iidx=0;iidx<4*DIM;iidx++)
 564     {
 565         scratch[iidx] = 0.0;
 566     }
 567
 568     /* Start outer loop over neighborlists */
 569     for(iidx=0; iidx<nri; iidx++)
 570     {
 571         /* Load shift vector for this list */
 572         i_shift_offset   = DIM*shiftidx[iidx];
 573
 574         /* Load limits for loop over neighbors */
 575         j_index_start    = jindex[iidx];
 576         j_index_end      = jindex[iidx+1];
 577
 578         /* Get outer coordinate index */
 579         inr              = iinr[iidx];
 580         i_coord_offset   = DIM*inr;
 581
 582         /* Load i particle coords and add shift vector */
 583         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 584
 585         fix0             = _mm_setzero_ps();
 586         fiy0             = _mm_setzero_ps();
 587         fiz0             = _mm_setzero_ps();
 588
 589         /* Load parameters for i particles */
 590         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 591         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 592
 593         /* Start inner kernel loop */
 594         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 595         {
 596
 597             /* Get j neighbor index, and coordinate index */
 598             jnrA             = jjnr[jidx];
 599             jnrB             = jjnr[jidx+1];
 600             jnrC             = jjnr[jidx+2];
 601             jnrD             = jjnr[jidx+3];
 602             j_coord_offsetA  = DIM*jnrA;
 603             j_coord_offsetB  = DIM*jnrB;
 604             j_coord_offsetC  = DIM*jnrC;
 605             j_coord_offsetD  = DIM*jnrD;
 606
 607             /* load j atom coordinates */
 608             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 609                                               x+j_coord_offsetC,x+j_coord_offsetD,
 610                                               &jx0,&jy0,&jz0);
 611
 612             /* Calculate displacement vector */
 613             dx00             = _mm_sub_ps(ix0,jx0);
 614             dy00             = _mm_sub_ps(iy0,jy0);
 615             dz00             = _mm_sub_ps(iz0,jz0);
 616
 617             /* Calculate squared distance and things based on it */
 618             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 619
 620             rinv00           = avx128fma_invsqrt_f(rsq00);
 621
 622             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 623
 624             /* Load parameters for j particles */
 625             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 626                                                               charge+jnrC+0,charge+jnrD+0);
 627             vdwjidx0A        = 2*vdwtype[jnrA+0];
 628             vdwjidx0B        = 2*vdwtype[jnrB+0];
 629             vdwjidx0C        = 2*vdwtype[jnrC+0];
 630             vdwjidx0D        = 2*vdwtype[jnrD+0];
 631
 632             /**************************
 633              * CALCULATE INTERACTIONS *
 634              **************************/
 635
 636             r00              = _mm_mul_ps(rsq00,rinv00);
 637
 638             /* Compute parameters for interactions between i and j atoms */
 639             qq00             = _mm_mul_ps(iq0,jq0);
 640             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 641                                          vdwparam+vdwioffset0+vdwjidx0B,
 642                                          vdwparam+vdwioffset0+vdwjidx0C,
 643                                          vdwparam+vdwioffset0+vdwjidx0D,
 644                                          &c6_00,&c12_00);
 645
 646             /* Calculate table index by multiplying r with table scale and truncate to integer */
 647             rt               = _mm_mul_ps(r00,vftabscale);
 648             vfitab           = _mm_cvttps_epi32(rt);
 649 #ifdef __XOP__
 650             vfeps            = _mm_frcz_ps(rt);
 651 #else
 652             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 653 #endif
 654             twovfeps         = _mm_add_ps(vfeps,vfeps);
 655             vfitab           = _mm_slli_epi32(vfitab,3);
 656
 657             /* EWALD ELECTROSTATICS */
 658
 659             /* Analytical PME correction */
 660             zeta2            = _mm_mul_ps(beta2,rsq00);
 661             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 662             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 663             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 664             felec            = _mm_mul_ps(qq00,felec);
 665
 666             /* CUBIC SPLINE TABLE DISPERSION */
 667             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 668             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 669             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 670             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 671             _MM_TRANSPOSE4_PS(Y,F,G,H);
 672             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 673             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 674             fvdw6            = _mm_mul_ps(c6_00,FF);
 675
 676             /* CUBIC SPLINE TABLE REPULSION */
 677             vfitab           = _mm_add_epi32(vfitab,ifour);
 678             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 679             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 680             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 681             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 682             _MM_TRANSPOSE4_PS(Y,F,G,H);
 683             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 684             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 685             fvdw12           = _mm_mul_ps(c12_00,FF);
 686             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 687
 688             fscal            = _mm_add_ps(felec,fvdw);
 689
 690              /* Update vectorial force */
 691             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 692             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 693             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 694
 695             fjptrA             = f+j_coord_offsetA;
 696             fjptrB             = f+j_coord_offsetB;
 697             fjptrC             = f+j_coord_offsetC;
 698             fjptrD             = f+j_coord_offsetD;
 699             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 700                                                    _mm_mul_ps(dx00,fscal),
 701                                                    _mm_mul_ps(dy00,fscal),
 702                                                    _mm_mul_ps(dz00,fscal));
 703
 704             /* Inner loop uses 54 flops */
 705         }
 706
 707         if(jidx<j_index_end)
 708         {
 709
 710             /* Get j neighbor index, and coordinate index */
 711             jnrlistA         = jjnr[jidx];
 712             jnrlistB         = jjnr[jidx+1];
 713             jnrlistC         = jjnr[jidx+2];
 714             jnrlistD         = jjnr[jidx+3];
 715             /* Sign of each element will be negative for non-real atoms.
 716              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 717              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 718              */
 719             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 720             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 721             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 722             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 723             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 724             j_coord_offsetA  = DIM*jnrA;
 725             j_coord_offsetB  = DIM*jnrB;
 726             j_coord_offsetC  = DIM*jnrC;
 727             j_coord_offsetD  = DIM*jnrD;
 728
 729             /* load j atom coordinates */
 730             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 731                                               x+j_coord_offsetC,x+j_coord_offsetD,
 732                                               &jx0,&jy0,&jz0);
 733
 734             /* Calculate displacement vector */
 735             dx00             = _mm_sub_ps(ix0,jx0);
 736             dy00             = _mm_sub_ps(iy0,jy0);
 737             dz00             = _mm_sub_ps(iz0,jz0);
 738
 739             /* Calculate squared distance and things based on it */
 740             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 741
 742             rinv00           = avx128fma_invsqrt_f(rsq00);
 743
 744             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 745
 746             /* Load parameters for j particles */
 747             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 748                                                               charge+jnrC+0,charge+jnrD+0);
 749             vdwjidx0A        = 2*vdwtype[jnrA+0];
 750             vdwjidx0B        = 2*vdwtype[jnrB+0];
 751             vdwjidx0C        = 2*vdwtype[jnrC+0];
 752             vdwjidx0D        = 2*vdwtype[jnrD+0];
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             r00              = _mm_mul_ps(rsq00,rinv00);
 759             r00              = _mm_andnot_ps(dummy_mask,r00);
 760
 761             /* Compute parameters for interactions between i and j atoms */
 762             qq00             = _mm_mul_ps(iq0,jq0);
 763             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 764                                          vdwparam+vdwioffset0+vdwjidx0B,
 765                                          vdwparam+vdwioffset0+vdwjidx0C,
 766                                          vdwparam+vdwioffset0+vdwjidx0D,
 767                                          &c6_00,&c12_00);
 768
 769             /* Calculate table index by multiplying r with table scale and truncate to integer */
 770             rt               = _mm_mul_ps(r00,vftabscale);
 771             vfitab           = _mm_cvttps_epi32(rt);
 772 #ifdef __XOP__
 773             vfeps            = _mm_frcz_ps(rt);
 774 #else
 775             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 776 #endif
 777             twovfeps         = _mm_add_ps(vfeps,vfeps);
 778             vfitab           = _mm_slli_epi32(vfitab,3);
 779
 780             /* EWALD ELECTROSTATICS */
 781
 782             /* Analytical PME correction */
 783             zeta2            = _mm_mul_ps(beta2,rsq00);
 784             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 785             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 786             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 787             felec            = _mm_mul_ps(qq00,felec);
 788
 789             /* CUBIC SPLINE TABLE DISPERSION */
 790             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 791             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 792             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 793             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 794             _MM_TRANSPOSE4_PS(Y,F,G,H);
 795             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 796             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 797             fvdw6            = _mm_mul_ps(c6_00,FF);
 798
 799             /* CUBIC SPLINE TABLE REPULSION */
 800             vfitab           = _mm_add_epi32(vfitab,ifour);
 801             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 802             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 803             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 804             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 805             _MM_TRANSPOSE4_PS(Y,F,G,H);
 806             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 807             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 808             fvdw12           = _mm_mul_ps(c12_00,FF);
 809             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 810
 811             fscal            = _mm_add_ps(felec,fvdw);
 812
 813             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 814
 815              /* Update vectorial force */
 816             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 817             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 818             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 819
 820             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 821             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 822             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 823             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 824             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 825                                                    _mm_mul_ps(dx00,fscal),
 826                                                    _mm_mul_ps(dy00,fscal),
 827                                                    _mm_mul_ps(dz00,fscal));
 828
 829             /* Inner loop uses 55 flops */
 830         }
 831
 832         /* End of innermost loop */
 833
 834         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 835                                               f+i_coord_offset,fshift+i_shift_offset);
 836
 837         /* Increment number of inner iterations */
 838         inneriter                  += j_index_end - j_index_start;
 839
 840         /* Outer loop uses 7 flops */
 841     }
 842
 843     /* Increment number of outer iterations */
 844     outeriter        += nri;
 845
 846     /* Update outer/inner flops */
 847
 848     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 849 }