src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128i          vfitab;
  99     __m128i          ifour       = _mm_set1_epi32(4);
 100     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 101     real             *vftab;
 102     __m128           dummy_mask,cutoff_mask;
 103     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 104     __m128           one     = _mm_set1_ps(1.0);
 105     __m128           two     = _mm_set1_ps(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_ps(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = _mm_set1_ps(fr->ic->k_rf);
 120     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 121     crf              = _mm_set1_ps(fr->ic->c_rf);
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     vftab            = kernel_data->table_vdw->data;
 127     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 128
 129     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 130     rcutoff_scalar   = fr->rcoulomb;
 131     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 132     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 133
 134     /* Avoid stupid compiler warnings */
 135     jnrA = jnrB = jnrC = jnrD = 0;
 136     j_coord_offsetA = 0;
 137     j_coord_offsetB = 0;
 138     j_coord_offsetC = 0;
 139     j_coord_offsetD = 0;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     for(iidx=0;iidx<4*DIM;iidx++)
 145     {
 146         scratch[iidx] = 0.0;
 147     }
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 165
 166         fix0             = _mm_setzero_ps();
 167         fiy0             = _mm_setzero_ps();
 168         fiz0             = _mm_setzero_ps();
 169
 170         /* Load parameters for i particles */
 171         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 172         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_ps();
 176         vvdwsum          = _mm_setzero_ps();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             jnrC             = jjnr[jidx+2];
 186             jnrD             = jjnr[jidx+3];
 187             j_coord_offsetA  = DIM*jnrA;
 188             j_coord_offsetB  = DIM*jnrB;
 189             j_coord_offsetC  = DIM*jnrC;
 190             j_coord_offsetD  = DIM*jnrD;
 191
 192             /* load j atom coordinates */
 193             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 194                                               x+j_coord_offsetC,x+j_coord_offsetD,
 195                                               &jx0,&jy0,&jz0);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _mm_sub_ps(ix0,jx0);
 199             dy00             = _mm_sub_ps(iy0,jy0);
 200             dz00             = _mm_sub_ps(iz0,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 204
 205             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 206
 207             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 208
 209             /* Load parameters for j particles */
 210             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 211                                                               charge+jnrC+0,charge+jnrD+0);
 212             vdwjidx0A        = 2*vdwtype[jnrA+0];
 213             vdwjidx0B        = 2*vdwtype[jnrB+0];
 214             vdwjidx0C        = 2*vdwtype[jnrC+0];
 215             vdwjidx0D        = 2*vdwtype[jnrD+0];
 216
 217             /**************************
 218              * CALCULATE INTERACTIONS *
 219              **************************/
 220
 221             if (gmx_mm_any_lt(rsq00,rcutoff2))
 222             {
 223
 224             r00              = _mm_mul_ps(rsq00,rinv00);
 225
 226             /* Compute parameters for interactions between i and j atoms */
 227             qq00             = _mm_mul_ps(iq0,jq0);
 228             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 229                                          vdwparam+vdwioffset0+vdwjidx0B,
 230                                          vdwparam+vdwioffset0+vdwjidx0C,
 231                                          vdwparam+vdwioffset0+vdwjidx0D,
 232                                          &c6_00,&c12_00);
 233
 234             /* Calculate table index by multiplying r with table scale and truncate to integer */
 235             rt               = _mm_mul_ps(r00,vftabscale);
 236             vfitab           = _mm_cvttps_epi32(rt);
 237 #ifdef __XOP__
 238             vfeps            = _mm_frcz_ps(rt);
 239 #else
 240             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 241 #endif
 242             twovfeps         = _mm_add_ps(vfeps,vfeps);
 243             vfitab           = _mm_slli_epi32(vfitab,3);
 244
 245             /* REACTION-FIELD ELECTROSTATICS */
 246             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 247             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 248
 249             /* CUBIC SPLINE TABLE DISPERSION */
 250             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 251             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 252             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 253             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 254             _MM_TRANSPOSE4_PS(Y,F,G,H);
 255             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 256             VV               = _mm_macc_ps(vfeps,Fp,Y);
 257             vvdw6            = _mm_mul_ps(c6_00,VV);
 258             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 259             fvdw6            = _mm_mul_ps(c6_00,FF);
 260
 261             /* CUBIC SPLINE TABLE REPULSION */
 262             vfitab           = _mm_add_epi32(vfitab,ifour);
 263             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 264             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 265             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 266             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 267             _MM_TRANSPOSE4_PS(Y,F,G,H);
 268             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 269             VV               = _mm_macc_ps(vfeps,Fp,Y);
 270             vvdw12           = _mm_mul_ps(c12_00,VV);
 271             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 272             fvdw12           = _mm_mul_ps(c12_00,FF);
 273             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 274             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 275
 276             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 277
 278             /* Update potential sum for this i atom from the interaction with this j atom. */
 279             velec            = _mm_and_ps(velec,cutoff_mask);
 280             velecsum         = _mm_add_ps(velecsum,velec);
 281             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 282             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 283
 284             fscal            = _mm_add_ps(felec,fvdw);
 285
 286             fscal            = _mm_and_ps(fscal,cutoff_mask);
 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             }
 303
 304             /* Inner loop uses 75 flops */
 305         }
 306
 307         if(jidx<j_index_end)
 308         {
 309
 310             /* Get j neighbor index, and coordinate index */
 311             jnrlistA         = jjnr[jidx];
 312             jnrlistB         = jjnr[jidx+1];
 313             jnrlistC         = jjnr[jidx+2];
 314             jnrlistD         = jjnr[jidx+3];
 315             /* Sign of each element will be negative for non-real atoms.
 316              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 317              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 318              */
 319             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 320             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 321             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 322             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 323             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 324             j_coord_offsetA  = DIM*jnrA;
 325             j_coord_offsetB  = DIM*jnrB;
 326             j_coord_offsetC  = DIM*jnrC;
 327             j_coord_offsetD  = DIM*jnrD;
 328
 329             /* load j atom coordinates */
 330             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 331                                               x+j_coord_offsetC,x+j_coord_offsetD,
 332                                               &jx0,&jy0,&jz0);
 333
 334             /* Calculate displacement vector */
 335             dx00             = _mm_sub_ps(ix0,jx0);
 336             dy00             = _mm_sub_ps(iy0,jy0);
 337             dz00             = _mm_sub_ps(iz0,jz0);
 338
 339             /* Calculate squared distance and things based on it */
 340             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 341
 342             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 343
 344             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 345
 346             /* Load parameters for j particles */
 347             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 348                                                               charge+jnrC+0,charge+jnrD+0);
 349             vdwjidx0A        = 2*vdwtype[jnrA+0];
 350             vdwjidx0B        = 2*vdwtype[jnrB+0];
 351             vdwjidx0C        = 2*vdwtype[jnrC+0];
 352             vdwjidx0D        = 2*vdwtype[jnrD+0];
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             if (gmx_mm_any_lt(rsq00,rcutoff2))
 359             {
 360
 361             r00              = _mm_mul_ps(rsq00,rinv00);
 362             r00              = _mm_andnot_ps(dummy_mask,r00);
 363
 364             /* Compute parameters for interactions between i and j atoms */
 365             qq00             = _mm_mul_ps(iq0,jq0);
 366             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 367                                          vdwparam+vdwioffset0+vdwjidx0B,
 368                                          vdwparam+vdwioffset0+vdwjidx0C,
 369                                          vdwparam+vdwioffset0+vdwjidx0D,
 370                                          &c6_00,&c12_00);
 371
 372             /* Calculate table index by multiplying r with table scale and truncate to integer */
 373             rt               = _mm_mul_ps(r00,vftabscale);
 374             vfitab           = _mm_cvttps_epi32(rt);
 375 #ifdef __XOP__
 376             vfeps            = _mm_frcz_ps(rt);
 377 #else
 378             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 379 #endif
 380             twovfeps         = _mm_add_ps(vfeps,vfeps);
 381             vfitab           = _mm_slli_epi32(vfitab,3);
 382
 383             /* REACTION-FIELD ELECTROSTATICS */
 384             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 385             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 386
 387             /* CUBIC SPLINE TABLE DISPERSION */
 388             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 389             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 390             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 391             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 392             _MM_TRANSPOSE4_PS(Y,F,G,H);
 393             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 394             VV               = _mm_macc_ps(vfeps,Fp,Y);
 395             vvdw6            = _mm_mul_ps(c6_00,VV);
 396             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 397             fvdw6            = _mm_mul_ps(c6_00,FF);
 398
 399             /* CUBIC SPLINE TABLE REPULSION */
 400             vfitab           = _mm_add_epi32(vfitab,ifour);
 401             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 402             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 403             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 404             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 405             _MM_TRANSPOSE4_PS(Y,F,G,H);
 406             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 407             VV               = _mm_macc_ps(vfeps,Fp,Y);
 408             vvdw12           = _mm_mul_ps(c12_00,VV);
 409             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 410             fvdw12           = _mm_mul_ps(c12_00,FF);
 411             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 412             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 413
 414             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 415
 416             /* Update potential sum for this i atom from the interaction with this j atom. */
 417             velec            = _mm_and_ps(velec,cutoff_mask);
 418             velec            = _mm_andnot_ps(dummy_mask,velec);
 419             velecsum         = _mm_add_ps(velecsum,velec);
 420             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 421             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 422             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 423
 424             fscal            = _mm_add_ps(felec,fvdw);
 425
 426             fscal            = _mm_and_ps(fscal,cutoff_mask);
 427
 428             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 429
 430              /* Update vectorial force */
 431             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 432             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 433             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 434
 435             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 436             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 437             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 438             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 439             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 440                                                    _mm_mul_ps(dx00,fscal),
 441                                                    _mm_mul_ps(dy00,fscal),
 442                                                    _mm_mul_ps(dz00,fscal));
 443
 444             }
 445
 446             /* Inner loop uses 76 flops */
 447         }
 448
 449         /* End of innermost loop */
 450
 451         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 452                                               f+i_coord_offset,fshift+i_shift_offset);
 453
 454         ggid                        = gid[iidx];
 455         /* Update potential energies */
 456         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 457         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 458
 459         /* Increment number of inner iterations */
 460         inneriter                  += j_index_end - j_index_start;
 461
 462         /* Outer loop uses 9 flops */
 463     }
 464
 465     /* Increment number of outer iterations */
 466     outeriter        += nri;
 467
 468     /* Update outer/inner flops */
 469
 470     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
 471 }
 472 /*
 473  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 474  * Electrostatics interaction: ReactionField
 475  * VdW interaction:            CubicSplineTable
 476  * Geometry:                   Particle-Particle
 477  * Calculate force/pot:        Force
 478  */
 479 void
 480 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 481                     (t_nblist                    * gmx_restrict       nlist,
 482                      rvec                        * gmx_restrict          xx,
 483                      rvec                        * gmx_restrict          ff,
 484                      t_forcerec                  * gmx_restrict          fr,
 485                      t_mdatoms                   * gmx_restrict     mdatoms,
 486                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 487                      t_nrnb                      * gmx_restrict        nrnb)
 488 {
 489     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 490      * just 0 for non-waters.
 491      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 492      * jnr indices corresponding to data put in the four positions in the SIMD register.
 493      */
 494     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 495     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 496     int              jnrA,jnrB,jnrC,jnrD;
 497     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 498     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 499     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 500     real             rcutoff_scalar;
 501     real             *shiftvec,*fshift,*x,*f;
 502     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 503     real             scratch[4*DIM];
 504     __m128           fscal,rcutoff,rcutoff2,jidxall;
 505     int              vdwioffset0;
 506     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 507     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 508     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 509     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 510     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 511     real             *charge;
 512     int              nvdwtype;
 513     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 514     int              *vdwtype;
 515     real             *vdwparam;
 516     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 517     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 518     __m128i          vfitab;
 519     __m128i          ifour       = _mm_set1_epi32(4);
 520     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 521     real             *vftab;
 522     __m128           dummy_mask,cutoff_mask;
 523     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 524     __m128           one     = _mm_set1_ps(1.0);
 525     __m128           two     = _mm_set1_ps(2.0);
 526     x                = xx[0];
 527     f                = ff[0];
 528
 529     nri              = nlist->nri;
 530     iinr             = nlist->iinr;
 531     jindex           = nlist->jindex;
 532     jjnr             = nlist->jjnr;
 533     shiftidx         = nlist->shift;
 534     gid              = nlist->gid;
 535     shiftvec         = fr->shift_vec[0];
 536     fshift           = fr->fshift[0];
 537     facel            = _mm_set1_ps(fr->epsfac);
 538     charge           = mdatoms->chargeA;
 539     krf              = _mm_set1_ps(fr->ic->k_rf);
 540     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 541     crf              = _mm_set1_ps(fr->ic->c_rf);
 542     nvdwtype         = fr->ntype;
 543     vdwparam         = fr->nbfp;
 544     vdwtype          = mdatoms->typeA;
 545
 546     vftab            = kernel_data->table_vdw->data;
 547     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 548
 549     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 550     rcutoff_scalar   = fr->rcoulomb;
 551     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 552     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 553
 554     /* Avoid stupid compiler warnings */
 555     jnrA = jnrB = jnrC = jnrD = 0;
 556     j_coord_offsetA = 0;
 557     j_coord_offsetB = 0;
 558     j_coord_offsetC = 0;
 559     j_coord_offsetD = 0;
 560
 561     outeriter        = 0;
 562     inneriter        = 0;
 563
 564     for(iidx=0;iidx<4*DIM;iidx++)
 565     {
 566         scratch[iidx] = 0.0;
 567     }
 568
 569     /* Start outer loop over neighborlists */
 570     for(iidx=0; iidx<nri; iidx++)
 571     {
 572         /* Load shift vector for this list */
 573         i_shift_offset   = DIM*shiftidx[iidx];
 574
 575         /* Load limits for loop over neighbors */
 576         j_index_start    = jindex[iidx];
 577         j_index_end      = jindex[iidx+1];
 578
 579         /* Get outer coordinate index */
 580         inr              = iinr[iidx];
 581         i_coord_offset   = DIM*inr;
 582
 583         /* Load i particle coords and add shift vector */
 584         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 585
 586         fix0             = _mm_setzero_ps();
 587         fiy0             = _mm_setzero_ps();
 588         fiz0             = _mm_setzero_ps();
 589
 590         /* Load parameters for i particles */
 591         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 592         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 593
 594         /* Start inner kernel loop */
 595         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 596         {
 597
 598             /* Get j neighbor index, and coordinate index */
 599             jnrA             = jjnr[jidx];
 600             jnrB             = jjnr[jidx+1];
 601             jnrC             = jjnr[jidx+2];
 602             jnrD             = jjnr[jidx+3];
 603             j_coord_offsetA  = DIM*jnrA;
 604             j_coord_offsetB  = DIM*jnrB;
 605             j_coord_offsetC  = DIM*jnrC;
 606             j_coord_offsetD  = DIM*jnrD;
 607
 608             /* load j atom coordinates */
 609             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 610                                               x+j_coord_offsetC,x+j_coord_offsetD,
 611                                               &jx0,&jy0,&jz0);
 612
 613             /* Calculate displacement vector */
 614             dx00             = _mm_sub_ps(ix0,jx0);
 615             dy00             = _mm_sub_ps(iy0,jy0);
 616             dz00             = _mm_sub_ps(iz0,jz0);
 617
 618             /* Calculate squared distance and things based on it */
 619             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 620
 621             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 622
 623             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 624
 625             /* Load parameters for j particles */
 626             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 627                                                               charge+jnrC+0,charge+jnrD+0);
 628             vdwjidx0A        = 2*vdwtype[jnrA+0];
 629             vdwjidx0B        = 2*vdwtype[jnrB+0];
 630             vdwjidx0C        = 2*vdwtype[jnrC+0];
 631             vdwjidx0D        = 2*vdwtype[jnrD+0];
 632
 633             /**************************
 634              * CALCULATE INTERACTIONS *
 635              **************************/
 636
 637             if (gmx_mm_any_lt(rsq00,rcutoff2))
 638             {
 639
 640             r00              = _mm_mul_ps(rsq00,rinv00);
 641
 642             /* Compute parameters for interactions between i and j atoms */
 643             qq00             = _mm_mul_ps(iq0,jq0);
 644             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 645                                          vdwparam+vdwioffset0+vdwjidx0B,
 646                                          vdwparam+vdwioffset0+vdwjidx0C,
 647                                          vdwparam+vdwioffset0+vdwjidx0D,
 648                                          &c6_00,&c12_00);
 649
 650             /* Calculate table index by multiplying r with table scale and truncate to integer */
 651             rt               = _mm_mul_ps(r00,vftabscale);
 652             vfitab           = _mm_cvttps_epi32(rt);
 653 #ifdef __XOP__
 654             vfeps            = _mm_frcz_ps(rt);
 655 #else
 656             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 657 #endif
 658             twovfeps         = _mm_add_ps(vfeps,vfeps);
 659             vfitab           = _mm_slli_epi32(vfitab,3);
 660
 661             /* REACTION-FIELD ELECTROSTATICS */
 662             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 663
 664             /* CUBIC SPLINE TABLE DISPERSION */
 665             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 666             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 667             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 668             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 669             _MM_TRANSPOSE4_PS(Y,F,G,H);
 670             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 671             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 672             fvdw6            = _mm_mul_ps(c6_00,FF);
 673
 674             /* CUBIC SPLINE TABLE REPULSION */
 675             vfitab           = _mm_add_epi32(vfitab,ifour);
 676             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 677             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 678             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 679             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 680             _MM_TRANSPOSE4_PS(Y,F,G,H);
 681             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 682             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 683             fvdw12           = _mm_mul_ps(c12_00,FF);
 684             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 685
 686             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 687
 688             fscal            = _mm_add_ps(felec,fvdw);
 689
 690             fscal            = _mm_and_ps(fscal,cutoff_mask);
 691
 692              /* Update vectorial force */
 693             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 694             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 695             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 696
 697             fjptrA             = f+j_coord_offsetA;
 698             fjptrB             = f+j_coord_offsetB;
 699             fjptrC             = f+j_coord_offsetC;
 700             fjptrD             = f+j_coord_offsetD;
 701             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 702                                                    _mm_mul_ps(dx00,fscal),
 703                                                    _mm_mul_ps(dy00,fscal),
 704                                                    _mm_mul_ps(dz00,fscal));
 705
 706             }
 707
 708             /* Inner loop uses 60 flops */
 709         }
 710
 711         if(jidx<j_index_end)
 712         {
 713
 714             /* Get j neighbor index, and coordinate index */
 715             jnrlistA         = jjnr[jidx];
 716             jnrlistB         = jjnr[jidx+1];
 717             jnrlistC         = jjnr[jidx+2];
 718             jnrlistD         = jjnr[jidx+3];
 719             /* Sign of each element will be negative for non-real atoms.
 720              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 721              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 722              */
 723             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 724             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 725             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 726             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 727             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 728             j_coord_offsetA  = DIM*jnrA;
 729             j_coord_offsetB  = DIM*jnrB;
 730             j_coord_offsetC  = DIM*jnrC;
 731             j_coord_offsetD  = DIM*jnrD;
 732
 733             /* load j atom coordinates */
 734             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 735                                               x+j_coord_offsetC,x+j_coord_offsetD,
 736                                               &jx0,&jy0,&jz0);
 737
 738             /* Calculate displacement vector */
 739             dx00             = _mm_sub_ps(ix0,jx0);
 740             dy00             = _mm_sub_ps(iy0,jy0);
 741             dz00             = _mm_sub_ps(iz0,jz0);
 742
 743             /* Calculate squared distance and things based on it */
 744             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 745
 746             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 747
 748             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 749
 750             /* Load parameters for j particles */
 751             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 752                                                               charge+jnrC+0,charge+jnrD+0);
 753             vdwjidx0A        = 2*vdwtype[jnrA+0];
 754             vdwjidx0B        = 2*vdwtype[jnrB+0];
 755             vdwjidx0C        = 2*vdwtype[jnrC+0];
 756             vdwjidx0D        = 2*vdwtype[jnrD+0];
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             if (gmx_mm_any_lt(rsq00,rcutoff2))
 763             {
 764
 765             r00              = _mm_mul_ps(rsq00,rinv00);
 766             r00              = _mm_andnot_ps(dummy_mask,r00);
 767
 768             /* Compute parameters for interactions between i and j atoms */
 769             qq00             = _mm_mul_ps(iq0,jq0);
 770             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 771                                          vdwparam+vdwioffset0+vdwjidx0B,
 772                                          vdwparam+vdwioffset0+vdwjidx0C,
 773                                          vdwparam+vdwioffset0+vdwjidx0D,
 774                                          &c6_00,&c12_00);
 775
 776             /* Calculate table index by multiplying r with table scale and truncate to integer */
 777             rt               = _mm_mul_ps(r00,vftabscale);
 778             vfitab           = _mm_cvttps_epi32(rt);
 779 #ifdef __XOP__
 780             vfeps            = _mm_frcz_ps(rt);
 781 #else
 782             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 783 #endif
 784             twovfeps         = _mm_add_ps(vfeps,vfeps);
 785             vfitab           = _mm_slli_epi32(vfitab,3);
 786
 787             /* REACTION-FIELD ELECTROSTATICS */
 788             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 789
 790             /* CUBIC SPLINE TABLE DISPERSION */
 791             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 792             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 793             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 794             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 795             _MM_TRANSPOSE4_PS(Y,F,G,H);
 796             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 797             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 798             fvdw6            = _mm_mul_ps(c6_00,FF);
 799
 800             /* CUBIC SPLINE TABLE REPULSION */
 801             vfitab           = _mm_add_epi32(vfitab,ifour);
 802             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 803             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 804             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 805             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 806             _MM_TRANSPOSE4_PS(Y,F,G,H);
 807             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 808             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 809             fvdw12           = _mm_mul_ps(c12_00,FF);
 810             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 811
 812             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 813
 814             fscal            = _mm_add_ps(felec,fvdw);
 815
 816             fscal            = _mm_and_ps(fscal,cutoff_mask);
 817
 818             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 819
 820              /* Update vectorial force */
 821             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 822             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 823             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 824
 825             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 826             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 827             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 828             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 829             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 830                                                    _mm_mul_ps(dx00,fscal),
 831                                                    _mm_mul_ps(dy00,fscal),
 832                                                    _mm_mul_ps(dz00,fscal));
 833
 834             }
 835
 836             /* Inner loop uses 61 flops */
 837         }
 838
 839         /* End of innermost loop */
 840
 841         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 842                                               f+i_coord_offset,fshift+i_shift_offset);
 843
 844         /* Increment number of inner iterations */
 845         inneriter                  += j_index_end - j_index_start;
 846
 847         /* Outer loop uses 7 flops */
 848     }
 849
 850     /* Increment number of outer iterations */
 851     outeriter        += nri;
 852
 853     /* Update outer/inner flops */
 854
 855     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*61);
 856 }