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) 2012,2013,2014, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  95     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  96     __m128i          vfitab;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  99     real             *vftab;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     krf              = _mm_set1_ps(fr->ic->k_rf);
 118     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 119     crf              = _mm_set1_ps(fr->ic->c_rf);
 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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 128     rcutoff_scalar   = fr->rcoulomb;
 129     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 130     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 131
 132     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = jnrC = jnrD = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136     j_coord_offsetC = 0;
 137     j_coord_offsetD = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     for(iidx=0;iidx<4*DIM;iidx++)
 143     {
 144         scratch[iidx] = 0.0;
 145     }
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 163
 164         fix0             = _mm_setzero_ps();
 165         fiy0             = _mm_setzero_ps();
 166         fiz0             = _mm_setzero_ps();
 167
 168         /* Load parameters for i particles */
 169         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 170         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 171
 172         /* Reset potential sums */
 173         velecsum         = _mm_setzero_ps();
 174         vvdwsum          = _mm_setzero_ps();
 175
 176         /* Start inner kernel loop */
 177         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 178         {
 179
 180             /* Get j neighbor index, and coordinate index */
 181             jnrA             = jjnr[jidx];
 182             jnrB             = jjnr[jidx+1];
 183             jnrC             = jjnr[jidx+2];
 184             jnrD             = jjnr[jidx+3];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187             j_coord_offsetC  = DIM*jnrC;
 188             j_coord_offsetD  = DIM*jnrD;
 189
 190             /* load j atom coordinates */
 191             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 192                                               x+j_coord_offsetC,x+j_coord_offsetD,
 193                                               &jx0,&jy0,&jz0);
 194
 195             /* Calculate displacement vector */
 196             dx00             = _mm_sub_ps(ix0,jx0);
 197             dy00             = _mm_sub_ps(iy0,jy0);
 198             dz00             = _mm_sub_ps(iz0,jz0);
 199
 200             /* Calculate squared distance and things based on it */
 201             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 202
 203             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 204
 205             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 206
 207             /* Load parameters for j particles */
 208             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 209                                                               charge+jnrC+0,charge+jnrD+0);
 210             vdwjidx0A        = 2*vdwtype[jnrA+0];
 211             vdwjidx0B        = 2*vdwtype[jnrB+0];
 212             vdwjidx0C        = 2*vdwtype[jnrC+0];
 213             vdwjidx0D        = 2*vdwtype[jnrD+0];
 214
 215             /**************************
 216              * CALCULATE INTERACTIONS *
 217              **************************/
 218
 219             if (gmx_mm_any_lt(rsq00,rcutoff2))
 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             /* REACTION-FIELD ELECTROSTATICS */
 244             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 245             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 246
 247             /* CUBIC SPLINE TABLE DISPERSION */
 248             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 249             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 250             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 251             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 252             _MM_TRANSPOSE4_PS(Y,F,G,H);
 253             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 254             VV               = _mm_macc_ps(vfeps,Fp,Y);
 255             vvdw6            = _mm_mul_ps(c6_00,VV);
 256             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 257             fvdw6            = _mm_mul_ps(c6_00,FF);
 258
 259             /* CUBIC SPLINE TABLE REPULSION */
 260             vfitab           = _mm_add_epi32(vfitab,ifour);
 261             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 262             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 263             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 264             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 265             _MM_TRANSPOSE4_PS(Y,F,G,H);
 266             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 267             VV               = _mm_macc_ps(vfeps,Fp,Y);
 268             vvdw12           = _mm_mul_ps(c12_00,VV);
 269             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 270             fvdw12           = _mm_mul_ps(c12_00,FF);
 271             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 272             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 273
 274             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 275
 276             /* Update potential sum for this i atom from the interaction with this j atom. */
 277             velec            = _mm_and_ps(velec,cutoff_mask);
 278             velecsum         = _mm_add_ps(velecsum,velec);
 279             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 280             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 281
 282             fscal            = _mm_add_ps(felec,fvdw);
 283
 284             fscal            = _mm_and_ps(fscal,cutoff_mask);
 285
 286              /* Update vectorial force */
 287             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 288             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 289             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 290
 291             fjptrA             = f+j_coord_offsetA;
 292             fjptrB             = f+j_coord_offsetB;
 293             fjptrC             = f+j_coord_offsetC;
 294             fjptrD             = f+j_coord_offsetD;
 295             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 296                                                    _mm_mul_ps(dx00,fscal),
 297                                                    _mm_mul_ps(dy00,fscal),
 298                                                    _mm_mul_ps(dz00,fscal));
 299
 300             }
 301
 302             /* Inner loop uses 75 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           = gmx_mm_invsqrt_ps(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             if (gmx_mm_any_lt(rsq00,rcutoff2))
 357             {
 358
 359             r00              = _mm_mul_ps(rsq00,rinv00);
 360             r00              = _mm_andnot_ps(dummy_mask,r00);
 361
 362             /* Compute parameters for interactions between i and j atoms */
 363             qq00             = _mm_mul_ps(iq0,jq0);
 364             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 365                                          vdwparam+vdwioffset0+vdwjidx0B,
 366                                          vdwparam+vdwioffset0+vdwjidx0C,
 367                                          vdwparam+vdwioffset0+vdwjidx0D,
 368                                          &c6_00,&c12_00);
 369
 370             /* Calculate table index by multiplying r with table scale and truncate to integer */
 371             rt               = _mm_mul_ps(r00,vftabscale);
 372             vfitab           = _mm_cvttps_epi32(rt);
 373 #ifdef __XOP__
 374             vfeps            = _mm_frcz_ps(rt);
 375 #else
 376             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 377 #endif
 378             twovfeps         = _mm_add_ps(vfeps,vfeps);
 379             vfitab           = _mm_slli_epi32(vfitab,3);
 380
 381             /* REACTION-FIELD ELECTROSTATICS */
 382             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 383             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 384
 385             /* CUBIC SPLINE TABLE DISPERSION */
 386             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 387             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 388             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 389             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 390             _MM_TRANSPOSE4_PS(Y,F,G,H);
 391             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 392             VV               = _mm_macc_ps(vfeps,Fp,Y);
 393             vvdw6            = _mm_mul_ps(c6_00,VV);
 394             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 395             fvdw6            = _mm_mul_ps(c6_00,FF);
 396
 397             /* CUBIC SPLINE TABLE REPULSION */
 398             vfitab           = _mm_add_epi32(vfitab,ifour);
 399             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 400             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 401             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 402             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 403             _MM_TRANSPOSE4_PS(Y,F,G,H);
 404             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 405             VV               = _mm_macc_ps(vfeps,Fp,Y);
 406             vvdw12           = _mm_mul_ps(c12_00,VV);
 407             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 408             fvdw12           = _mm_mul_ps(c12_00,FF);
 409             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 410             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 411
 412             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velec            = _mm_and_ps(velec,cutoff_mask);
 416             velec            = _mm_andnot_ps(dummy_mask,velec);
 417             velecsum         = _mm_add_ps(velecsum,velec);
 418             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 419             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 420             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 421
 422             fscal            = _mm_add_ps(felec,fvdw);
 423
 424             fscal            = _mm_and_ps(fscal,cutoff_mask);
 425
 426             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 427
 428              /* Update vectorial force */
 429             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 430             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 431             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 432
 433             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 434             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 435             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 436             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 437             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 438                                                    _mm_mul_ps(dx00,fscal),
 439                                                    _mm_mul_ps(dy00,fscal),
 440                                                    _mm_mul_ps(dz00,fscal));
 441
 442             }
 443
 444             /* Inner loop uses 76 flops */
 445         }
 446
 447         /* End of innermost loop */
 448
 449         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 450                                               f+i_coord_offset,fshift+i_shift_offset);
 451
 452         ggid                        = gid[iidx];
 453         /* Update potential energies */
 454         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 455         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 456
 457         /* Increment number of inner iterations */
 458         inneriter                  += j_index_end - j_index_start;
 459
 460         /* Outer loop uses 9 flops */
 461     }
 462
 463     /* Increment number of outer iterations */
 464     outeriter        += nri;
 465
 466     /* Update outer/inner flops */
 467
 468     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
 469 }
 470 /*
 471  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 472  * Electrostatics interaction: ReactionField
 473  * VdW interaction:            CubicSplineTable
 474  * Geometry:                   Particle-Particle
 475  * Calculate force/pot:        Force
 476  */
 477 void
 478 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 479                     (t_nblist                    * gmx_restrict       nlist,
 480                      rvec                        * gmx_restrict          xx,
 481                      rvec                        * gmx_restrict          ff,
 482                      t_forcerec                  * gmx_restrict          fr,
 483                      t_mdatoms                   * gmx_restrict     mdatoms,
 484                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 485                      t_nrnb                      * gmx_restrict        nrnb)
 486 {
 487     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 488      * just 0 for non-waters.
 489      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 490      * jnr indices corresponding to data put in the four positions in the SIMD register.
 491      */
 492     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 493     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 494     int              jnrA,jnrB,jnrC,jnrD;
 495     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 496     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 497     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 498     real             rcutoff_scalar;
 499     real             *shiftvec,*fshift,*x,*f;
 500     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 501     real             scratch[4*DIM];
 502     __m128           fscal,rcutoff,rcutoff2,jidxall;
 503     int              vdwioffset0;
 504     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 505     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 506     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 507     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 508     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 509     real             *charge;
 510     int              nvdwtype;
 511     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 512     int              *vdwtype;
 513     real             *vdwparam;
 514     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 515     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 516     __m128i          vfitab;
 517     __m128i          ifour       = _mm_set1_epi32(4);
 518     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 519     real             *vftab;
 520     __m128           dummy_mask,cutoff_mask;
 521     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 522     __m128           one     = _mm_set1_ps(1.0);
 523     __m128           two     = _mm_set1_ps(2.0);
 524     x                = xx[0];
 525     f                = ff[0];
 526
 527     nri              = nlist->nri;
 528     iinr             = nlist->iinr;
 529     jindex           = nlist->jindex;
 530     jjnr             = nlist->jjnr;
 531     shiftidx         = nlist->shift;
 532     gid              = nlist->gid;
 533     shiftvec         = fr->shift_vec[0];
 534     fshift           = fr->fshift[0];
 535     facel            = _mm_set1_ps(fr->epsfac);
 536     charge           = mdatoms->chargeA;
 537     krf              = _mm_set1_ps(fr->ic->k_rf);
 538     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 539     crf              = _mm_set1_ps(fr->ic->c_rf);
 540     nvdwtype         = fr->ntype;
 541     vdwparam         = fr->nbfp;
 542     vdwtype          = mdatoms->typeA;
 543
 544     vftab            = kernel_data->table_vdw->data;
 545     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 546
 547     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 548     rcutoff_scalar   = fr->rcoulomb;
 549     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 550     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 551
 552     /* Avoid stupid compiler warnings */
 553     jnrA = jnrB = jnrC = jnrD = 0;
 554     j_coord_offsetA = 0;
 555     j_coord_offsetB = 0;
 556     j_coord_offsetC = 0;
 557     j_coord_offsetD = 0;
 558
 559     outeriter        = 0;
 560     inneriter        = 0;
 561
 562     for(iidx=0;iidx<4*DIM;iidx++)
 563     {
 564         scratch[iidx] = 0.0;
 565     }
 566
 567     /* Start outer loop over neighborlists */
 568     for(iidx=0; iidx<nri; iidx++)
 569     {
 570         /* Load shift vector for this list */
 571         i_shift_offset   = DIM*shiftidx[iidx];
 572
 573         /* Load limits for loop over neighbors */
 574         j_index_start    = jindex[iidx];
 575         j_index_end      = jindex[iidx+1];
 576
 577         /* Get outer coordinate index */
 578         inr              = iinr[iidx];
 579         i_coord_offset   = DIM*inr;
 580
 581         /* Load i particle coords and add shift vector */
 582         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 583
 584         fix0             = _mm_setzero_ps();
 585         fiy0             = _mm_setzero_ps();
 586         fiz0             = _mm_setzero_ps();
 587
 588         /* Load parameters for i particles */
 589         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 590         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 591
 592         /* Start inner kernel loop */
 593         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 594         {
 595
 596             /* Get j neighbor index, and coordinate index */
 597             jnrA             = jjnr[jidx];
 598             jnrB             = jjnr[jidx+1];
 599             jnrC             = jjnr[jidx+2];
 600             jnrD             = jjnr[jidx+3];
 601             j_coord_offsetA  = DIM*jnrA;
 602             j_coord_offsetB  = DIM*jnrB;
 603             j_coord_offsetC  = DIM*jnrC;
 604             j_coord_offsetD  = DIM*jnrD;
 605
 606             /* load j atom coordinates */
 607             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 608                                               x+j_coord_offsetC,x+j_coord_offsetD,
 609                                               &jx0,&jy0,&jz0);
 610
 611             /* Calculate displacement vector */
 612             dx00             = _mm_sub_ps(ix0,jx0);
 613             dy00             = _mm_sub_ps(iy0,jy0);
 614             dz00             = _mm_sub_ps(iz0,jz0);
 615
 616             /* Calculate squared distance and things based on it */
 617             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 618
 619             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 620
 621             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 622
 623             /* Load parameters for j particles */
 624             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 625                                                               charge+jnrC+0,charge+jnrD+0);
 626             vdwjidx0A        = 2*vdwtype[jnrA+0];
 627             vdwjidx0B        = 2*vdwtype[jnrB+0];
 628             vdwjidx0C        = 2*vdwtype[jnrC+0];
 629             vdwjidx0D        = 2*vdwtype[jnrD+0];
 630
 631             /**************************
 632              * CALCULATE INTERACTIONS *
 633              **************************/
 634
 635             if (gmx_mm_any_lt(rsq00,rcutoff2))
 636             {
 637
 638             r00              = _mm_mul_ps(rsq00,rinv00);
 639
 640             /* Compute parameters for interactions between i and j atoms */
 641             qq00             = _mm_mul_ps(iq0,jq0);
 642             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 643                                          vdwparam+vdwioffset0+vdwjidx0B,
 644                                          vdwparam+vdwioffset0+vdwjidx0C,
 645                                          vdwparam+vdwioffset0+vdwjidx0D,
 646                                          &c6_00,&c12_00);
 647
 648             /* Calculate table index by multiplying r with table scale and truncate to integer */
 649             rt               = _mm_mul_ps(r00,vftabscale);
 650             vfitab           = _mm_cvttps_epi32(rt);
 651 #ifdef __XOP__
 652             vfeps            = _mm_frcz_ps(rt);
 653 #else
 654             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 655 #endif
 656             twovfeps         = _mm_add_ps(vfeps,vfeps);
 657             vfitab           = _mm_slli_epi32(vfitab,3);
 658
 659             /* REACTION-FIELD ELECTROSTATICS */
 660             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 661
 662             /* CUBIC SPLINE TABLE DISPERSION */
 663             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 664             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 665             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 666             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 667             _MM_TRANSPOSE4_PS(Y,F,G,H);
 668             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 669             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 670             fvdw6            = _mm_mul_ps(c6_00,FF);
 671
 672             /* CUBIC SPLINE TABLE REPULSION */
 673             vfitab           = _mm_add_epi32(vfitab,ifour);
 674             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 675             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 676             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 677             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 678             _MM_TRANSPOSE4_PS(Y,F,G,H);
 679             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 680             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 681             fvdw12           = _mm_mul_ps(c12_00,FF);
 682             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 683
 684             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 685
 686             fscal            = _mm_add_ps(felec,fvdw);
 687
 688             fscal            = _mm_and_ps(fscal,cutoff_mask);
 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             }
 705
 706             /* Inner loop uses 60 flops */
 707         }
 708
 709         if(jidx<j_index_end)
 710         {
 711
 712             /* Get j neighbor index, and coordinate index */
 713             jnrlistA         = jjnr[jidx];
 714             jnrlistB         = jjnr[jidx+1];
 715             jnrlistC         = jjnr[jidx+2];
 716             jnrlistD         = jjnr[jidx+3];
 717             /* Sign of each element will be negative for non-real atoms.
 718              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 719              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 720              */
 721             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 722             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 723             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 724             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 725             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 726             j_coord_offsetA  = DIM*jnrA;
 727             j_coord_offsetB  = DIM*jnrB;
 728             j_coord_offsetC  = DIM*jnrC;
 729             j_coord_offsetD  = DIM*jnrD;
 730
 731             /* load j atom coordinates */
 732             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 733                                               x+j_coord_offsetC,x+j_coord_offsetD,
 734                                               &jx0,&jy0,&jz0);
 735
 736             /* Calculate displacement vector */
 737             dx00             = _mm_sub_ps(ix0,jx0);
 738             dy00             = _mm_sub_ps(iy0,jy0);
 739             dz00             = _mm_sub_ps(iz0,jz0);
 740
 741             /* Calculate squared distance and things based on it */
 742             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 743
 744             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 745
 746             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 747
 748             /* Load parameters for j particles */
 749             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 750                                                               charge+jnrC+0,charge+jnrD+0);
 751             vdwjidx0A        = 2*vdwtype[jnrA+0];
 752             vdwjidx0B        = 2*vdwtype[jnrB+0];
 753             vdwjidx0C        = 2*vdwtype[jnrC+0];
 754             vdwjidx0D        = 2*vdwtype[jnrD+0];
 755
 756             /**************************
 757              * CALCULATE INTERACTIONS *
 758              **************************/
 759
 760             if (gmx_mm_any_lt(rsq00,rcutoff2))
 761             {
 762
 763             r00              = _mm_mul_ps(rsq00,rinv00);
 764             r00              = _mm_andnot_ps(dummy_mask,r00);
 765
 766             /* Compute parameters for interactions between i and j atoms */
 767             qq00             = _mm_mul_ps(iq0,jq0);
 768             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 769                                          vdwparam+vdwioffset0+vdwjidx0B,
 770                                          vdwparam+vdwioffset0+vdwjidx0C,
 771                                          vdwparam+vdwioffset0+vdwjidx0D,
 772                                          &c6_00,&c12_00);
 773
 774             /* Calculate table index by multiplying r with table scale and truncate to integer */
 775             rt               = _mm_mul_ps(r00,vftabscale);
 776             vfitab           = _mm_cvttps_epi32(rt);
 777 #ifdef __XOP__
 778             vfeps            = _mm_frcz_ps(rt);
 779 #else
 780             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 781 #endif
 782             twovfeps         = _mm_add_ps(vfeps,vfeps);
 783             vfitab           = _mm_slli_epi32(vfitab,3);
 784
 785             /* REACTION-FIELD ELECTROSTATICS */
 786             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 787
 788             /* CUBIC SPLINE TABLE DISPERSION */
 789             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 790             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 791             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 792             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 793             _MM_TRANSPOSE4_PS(Y,F,G,H);
 794             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 795             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 796             fvdw6            = _mm_mul_ps(c6_00,FF);
 797
 798             /* CUBIC SPLINE TABLE REPULSION */
 799             vfitab           = _mm_add_epi32(vfitab,ifour);
 800             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 801             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 802             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 803             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 804             _MM_TRANSPOSE4_PS(Y,F,G,H);
 805             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 806             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 807             fvdw12           = _mm_mul_ps(c12_00,FF);
 808             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 809
 810             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 811
 812             fscal            = _mm_add_ps(felec,fvdw);
 813
 814             fscal            = _mm_and_ps(fscal,cutoff_mask);
 815
 816             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 817
 818              /* Update vectorial force */
 819             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 820             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 821             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 822
 823             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 824             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 825             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 826             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 827             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 828                                                    _mm_mul_ps(dx00,fscal),
 829                                                    _mm_mul_ps(dy00,fscal),
 830                                                    _mm_mul_ps(dz00,fscal));
 831
 832             }
 833
 834             /* Inner loop uses 61 flops */
 835         }
 836
 837         /* End of innermost loop */
 838
 839         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 840                                               f+i_coord_offset,fshift+i_shift_offset);
 841
 842         /* Increment number of inner iterations */
 843         inneriter                  += j_index_end - j_index_start;
 844
 845         /* Outer loop uses 7 flops */
 846     }
 847
 848     /* Increment number of outer iterations */
 849     outeriter        += nri;
 850
 851     /* Update outer/inner flops */
 852
 853     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*61);
 854 }