src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRF_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 "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "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_ElecRF_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_ElecRF_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     /* Avoid stupid compiler warnings */
 128     jnrA = jnrB = jnrC = jnrD = 0;
 129     j_coord_offsetA = 0;
 130     j_coord_offsetB = 0;
 131     j_coord_offsetC = 0;
 132     j_coord_offsetD = 0;
 133
 134     outeriter        = 0;
 135     inneriter        = 0;
 136
 137     for(iidx=0;iidx<4*DIM;iidx++)
 138     {
 139         scratch[iidx] = 0.0;
 140     }
 141
 142     /* Start outer loop over neighborlists */
 143     for(iidx=0; iidx<nri; iidx++)
 144     {
 145         /* Load shift vector for this list */
 146         i_shift_offset   = DIM*shiftidx[iidx];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 158
 159         fix0             = _mm_setzero_ps();
 160         fiy0             = _mm_setzero_ps();
 161         fiz0             = _mm_setzero_ps();
 162
 163         /* Load parameters for i particles */
 164         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 165         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 166
 167         /* Reset potential sums */
 168         velecsum         = _mm_setzero_ps();
 169         vvdwsum          = _mm_setzero_ps();
 170
 171         /* Start inner kernel loop */
 172         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 173         {
 174
 175             /* Get j neighbor index, and coordinate index */
 176             jnrA             = jjnr[jidx];
 177             jnrB             = jjnr[jidx+1];
 178             jnrC             = jjnr[jidx+2];
 179             jnrD             = jjnr[jidx+3];
 180             j_coord_offsetA  = DIM*jnrA;
 181             j_coord_offsetB  = DIM*jnrB;
 182             j_coord_offsetC  = DIM*jnrC;
 183             j_coord_offsetD  = DIM*jnrD;
 184
 185             /* load j atom coordinates */
 186             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 187                                               x+j_coord_offsetC,x+j_coord_offsetD,
 188                                               &jx0,&jy0,&jz0);
 189
 190             /* Calculate displacement vector */
 191             dx00             = _mm_sub_ps(ix0,jx0);
 192             dy00             = _mm_sub_ps(iy0,jy0);
 193             dz00             = _mm_sub_ps(iz0,jz0);
 194
 195             /* Calculate squared distance and things based on it */
 196             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 197
 198             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 199
 200             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 201
 202             /* Load parameters for j particles */
 203             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 204                                                               charge+jnrC+0,charge+jnrD+0);
 205             vdwjidx0A        = 2*vdwtype[jnrA+0];
 206             vdwjidx0B        = 2*vdwtype[jnrB+0];
 207             vdwjidx0C        = 2*vdwtype[jnrC+0];
 208             vdwjidx0D        = 2*vdwtype[jnrD+0];
 209
 210             /**************************
 211              * CALCULATE INTERACTIONS *
 212              **************************/
 213
 214             r00              = _mm_mul_ps(rsq00,rinv00);
 215
 216             /* Compute parameters for interactions between i and j atoms */
 217             qq00             = _mm_mul_ps(iq0,jq0);
 218             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 219                                          vdwparam+vdwioffset0+vdwjidx0B,
 220                                          vdwparam+vdwioffset0+vdwjidx0C,
 221                                          vdwparam+vdwioffset0+vdwjidx0D,
 222                                          &c6_00,&c12_00);
 223
 224             /* Calculate table index by multiplying r with table scale and truncate to integer */
 225             rt               = _mm_mul_ps(r00,vftabscale);
 226             vfitab           = _mm_cvttps_epi32(rt);
 227 #ifdef __XOP__
 228             vfeps            = _mm_frcz_ps(rt);
 229 #else
 230             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 231 #endif
 232             twovfeps         = _mm_add_ps(vfeps,vfeps);
 233             vfitab           = _mm_slli_epi32(vfitab,3);
 234
 235             /* REACTION-FIELD ELECTROSTATICS */
 236             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 237             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 238
 239             /* CUBIC SPLINE TABLE DISPERSION */
 240             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 241             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 242             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 243             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 244             _MM_TRANSPOSE4_PS(Y,F,G,H);
 245             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 246             VV               = _mm_macc_ps(vfeps,Fp,Y);
 247             vvdw6            = _mm_mul_ps(c6_00,VV);
 248             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 249             fvdw6            = _mm_mul_ps(c6_00,FF);
 250
 251             /* CUBIC SPLINE TABLE REPULSION */
 252             vfitab           = _mm_add_epi32(vfitab,ifour);
 253             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 254             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 255             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 256             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 257             _MM_TRANSPOSE4_PS(Y,F,G,H);
 258             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 259             VV               = _mm_macc_ps(vfeps,Fp,Y);
 260             vvdw12           = _mm_mul_ps(c12_00,VV);
 261             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 262             fvdw12           = _mm_mul_ps(c12_00,FF);
 263             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 264             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 265
 266             /* Update potential sum for this i atom from the interaction with this j atom. */
 267             velecsum         = _mm_add_ps(velecsum,velec);
 268             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 269
 270             fscal            = _mm_add_ps(felec,fvdw);
 271
 272              /* Update vectorial force */
 273             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 274             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 275             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 276
 277             fjptrA             = f+j_coord_offsetA;
 278             fjptrB             = f+j_coord_offsetB;
 279             fjptrC             = f+j_coord_offsetC;
 280             fjptrD             = f+j_coord_offsetD;
 281             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 282                                                    _mm_mul_ps(dx00,fscal),
 283                                                    _mm_mul_ps(dy00,fscal),
 284                                                    _mm_mul_ps(dz00,fscal));
 285
 286             /* Inner loop uses 70 flops */
 287         }
 288
 289         if(jidx<j_index_end)
 290         {
 291
 292             /* Get j neighbor index, and coordinate index */
 293             jnrlistA         = jjnr[jidx];
 294             jnrlistB         = jjnr[jidx+1];
 295             jnrlistC         = jjnr[jidx+2];
 296             jnrlistD         = jjnr[jidx+3];
 297             /* Sign of each element will be negative for non-real atoms.
 298              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 299              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 300              */
 301             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 302             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 303             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 304             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 305             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 306             j_coord_offsetA  = DIM*jnrA;
 307             j_coord_offsetB  = DIM*jnrB;
 308             j_coord_offsetC  = DIM*jnrC;
 309             j_coord_offsetD  = DIM*jnrD;
 310
 311             /* load j atom coordinates */
 312             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 313                                               x+j_coord_offsetC,x+j_coord_offsetD,
 314                                               &jx0,&jy0,&jz0);
 315
 316             /* Calculate displacement vector */
 317             dx00             = _mm_sub_ps(ix0,jx0);
 318             dy00             = _mm_sub_ps(iy0,jy0);
 319             dz00             = _mm_sub_ps(iz0,jz0);
 320
 321             /* Calculate squared distance and things based on it */
 322             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 323
 324             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 325
 326             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 327
 328             /* Load parameters for j particles */
 329             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 330                                                               charge+jnrC+0,charge+jnrD+0);
 331             vdwjidx0A        = 2*vdwtype[jnrA+0];
 332             vdwjidx0B        = 2*vdwtype[jnrB+0];
 333             vdwjidx0C        = 2*vdwtype[jnrC+0];
 334             vdwjidx0D        = 2*vdwtype[jnrD+0];
 335
 336             /**************************
 337              * CALCULATE INTERACTIONS *
 338              **************************/
 339
 340             r00              = _mm_mul_ps(rsq00,rinv00);
 341             r00              = _mm_andnot_ps(dummy_mask,r00);
 342
 343             /* Compute parameters for interactions between i and j atoms */
 344             qq00             = _mm_mul_ps(iq0,jq0);
 345             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 346                                          vdwparam+vdwioffset0+vdwjidx0B,
 347                                          vdwparam+vdwioffset0+vdwjidx0C,
 348                                          vdwparam+vdwioffset0+vdwjidx0D,
 349                                          &c6_00,&c12_00);
 350
 351             /* Calculate table index by multiplying r with table scale and truncate to integer */
 352             rt               = _mm_mul_ps(r00,vftabscale);
 353             vfitab           = _mm_cvttps_epi32(rt);
 354 #ifdef __XOP__
 355             vfeps            = _mm_frcz_ps(rt);
 356 #else
 357             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 358 #endif
 359             twovfeps         = _mm_add_ps(vfeps,vfeps);
 360             vfitab           = _mm_slli_epi32(vfitab,3);
 361
 362             /* REACTION-FIELD ELECTROSTATICS */
 363             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 364             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 365
 366             /* CUBIC SPLINE TABLE DISPERSION */
 367             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 368             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 369             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 370             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 371             _MM_TRANSPOSE4_PS(Y,F,G,H);
 372             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 373             VV               = _mm_macc_ps(vfeps,Fp,Y);
 374             vvdw6            = _mm_mul_ps(c6_00,VV);
 375             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 376             fvdw6            = _mm_mul_ps(c6_00,FF);
 377
 378             /* CUBIC SPLINE TABLE REPULSION */
 379             vfitab           = _mm_add_epi32(vfitab,ifour);
 380             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 381             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 382             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 383             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 384             _MM_TRANSPOSE4_PS(Y,F,G,H);
 385             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 386             VV               = _mm_macc_ps(vfeps,Fp,Y);
 387             vvdw12           = _mm_mul_ps(c12_00,VV);
 388             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 389             fvdw12           = _mm_mul_ps(c12_00,FF);
 390             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 391             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 392
 393             /* Update potential sum for this i atom from the interaction with this j atom. */
 394             velec            = _mm_andnot_ps(dummy_mask,velec);
 395             velecsum         = _mm_add_ps(velecsum,velec);
 396             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 397             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 398
 399             fscal            = _mm_add_ps(felec,fvdw);
 400
 401             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 402
 403              /* Update vectorial force */
 404             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 405             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 406             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 407
 408             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 409             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 410             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 411             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 412             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 413                                                    _mm_mul_ps(dx00,fscal),
 414                                                    _mm_mul_ps(dy00,fscal),
 415                                                    _mm_mul_ps(dz00,fscal));
 416
 417             /* Inner loop uses 71 flops */
 418         }
 419
 420         /* End of innermost loop */
 421
 422         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 423                                               f+i_coord_offset,fshift+i_shift_offset);
 424
 425         ggid                        = gid[iidx];
 426         /* Update potential energies */
 427         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 428         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 429
 430         /* Increment number of inner iterations */
 431         inneriter                  += j_index_end - j_index_start;
 432
 433         /* Outer loop uses 9 flops */
 434     }
 435
 436     /* Increment number of outer iterations */
 437     outeriter        += nri;
 438
 439     /* Update outer/inner flops */
 440
 441     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
 442 }
 443 /*
 444  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 445  * Electrostatics interaction: ReactionField
 446  * VdW interaction:            CubicSplineTable
 447  * Geometry:                   Particle-Particle
 448  * Calculate force/pot:        Force
 449  */
 450 void
 451 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 452                     (t_nblist                    * gmx_restrict       nlist,
 453                      rvec                        * gmx_restrict          xx,
 454                      rvec                        * gmx_restrict          ff,
 455                      t_forcerec                  * gmx_restrict          fr,
 456                      t_mdatoms                   * gmx_restrict     mdatoms,
 457                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 458                      t_nrnb                      * gmx_restrict        nrnb)
 459 {
 460     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 461      * just 0 for non-waters.
 462      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 463      * jnr indices corresponding to data put in the four positions in the SIMD register.
 464      */
 465     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 466     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 467     int              jnrA,jnrB,jnrC,jnrD;
 468     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 469     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 470     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 471     real             rcutoff_scalar;
 472     real             *shiftvec,*fshift,*x,*f;
 473     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 474     real             scratch[4*DIM];
 475     __m128           fscal,rcutoff,rcutoff2,jidxall;
 476     int              vdwioffset0;
 477     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 478     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 479     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 480     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 481     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 482     real             *charge;
 483     int              nvdwtype;
 484     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 485     int              *vdwtype;
 486     real             *vdwparam;
 487     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 488     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 489     __m128i          vfitab;
 490     __m128i          ifour       = _mm_set1_epi32(4);
 491     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 492     real             *vftab;
 493     __m128           dummy_mask,cutoff_mask;
 494     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 495     __m128           one     = _mm_set1_ps(1.0);
 496     __m128           two     = _mm_set1_ps(2.0);
 497     x                = xx[0];
 498     f                = ff[0];
 499
 500     nri              = nlist->nri;
 501     iinr             = nlist->iinr;
 502     jindex           = nlist->jindex;
 503     jjnr             = nlist->jjnr;
 504     shiftidx         = nlist->shift;
 505     gid              = nlist->gid;
 506     shiftvec         = fr->shift_vec[0];
 507     fshift           = fr->fshift[0];
 508     facel            = _mm_set1_ps(fr->epsfac);
 509     charge           = mdatoms->chargeA;
 510     krf              = _mm_set1_ps(fr->ic->k_rf);
 511     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 512     crf              = _mm_set1_ps(fr->ic->c_rf);
 513     nvdwtype         = fr->ntype;
 514     vdwparam         = fr->nbfp;
 515     vdwtype          = mdatoms->typeA;
 516
 517     vftab            = kernel_data->table_vdw->data;
 518     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 519
 520     /* Avoid stupid compiler warnings */
 521     jnrA = jnrB = jnrC = jnrD = 0;
 522     j_coord_offsetA = 0;
 523     j_coord_offsetB = 0;
 524     j_coord_offsetC = 0;
 525     j_coord_offsetD = 0;
 526
 527     outeriter        = 0;
 528     inneriter        = 0;
 529
 530     for(iidx=0;iidx<4*DIM;iidx++)
 531     {
 532         scratch[iidx] = 0.0;
 533     }
 534
 535     /* Start outer loop over neighborlists */
 536     for(iidx=0; iidx<nri; iidx++)
 537     {
 538         /* Load shift vector for this list */
 539         i_shift_offset   = DIM*shiftidx[iidx];
 540
 541         /* Load limits for loop over neighbors */
 542         j_index_start    = jindex[iidx];
 543         j_index_end      = jindex[iidx+1];
 544
 545         /* Get outer coordinate index */
 546         inr              = iinr[iidx];
 547         i_coord_offset   = DIM*inr;
 548
 549         /* Load i particle coords and add shift vector */
 550         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 551
 552         fix0             = _mm_setzero_ps();
 553         fiy0             = _mm_setzero_ps();
 554         fiz0             = _mm_setzero_ps();
 555
 556         /* Load parameters for i particles */
 557         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 558         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 559
 560         /* Start inner kernel loop */
 561         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 562         {
 563
 564             /* Get j neighbor index, and coordinate index */
 565             jnrA             = jjnr[jidx];
 566             jnrB             = jjnr[jidx+1];
 567             jnrC             = jjnr[jidx+2];
 568             jnrD             = jjnr[jidx+3];
 569             j_coord_offsetA  = DIM*jnrA;
 570             j_coord_offsetB  = DIM*jnrB;
 571             j_coord_offsetC  = DIM*jnrC;
 572             j_coord_offsetD  = DIM*jnrD;
 573
 574             /* load j atom coordinates */
 575             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 576                                               x+j_coord_offsetC,x+j_coord_offsetD,
 577                                               &jx0,&jy0,&jz0);
 578
 579             /* Calculate displacement vector */
 580             dx00             = _mm_sub_ps(ix0,jx0);
 581             dy00             = _mm_sub_ps(iy0,jy0);
 582             dz00             = _mm_sub_ps(iz0,jz0);
 583
 584             /* Calculate squared distance and things based on it */
 585             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 586
 587             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 588
 589             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 590
 591             /* Load parameters for j particles */
 592             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 593                                                               charge+jnrC+0,charge+jnrD+0);
 594             vdwjidx0A        = 2*vdwtype[jnrA+0];
 595             vdwjidx0B        = 2*vdwtype[jnrB+0];
 596             vdwjidx0C        = 2*vdwtype[jnrC+0];
 597             vdwjidx0D        = 2*vdwtype[jnrD+0];
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             r00              = _mm_mul_ps(rsq00,rinv00);
 604
 605             /* Compute parameters for interactions between i and j atoms */
 606             qq00             = _mm_mul_ps(iq0,jq0);
 607             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 608                                          vdwparam+vdwioffset0+vdwjidx0B,
 609                                          vdwparam+vdwioffset0+vdwjidx0C,
 610                                          vdwparam+vdwioffset0+vdwjidx0D,
 611                                          &c6_00,&c12_00);
 612
 613             /* Calculate table index by multiplying r with table scale and truncate to integer */
 614             rt               = _mm_mul_ps(r00,vftabscale);
 615             vfitab           = _mm_cvttps_epi32(rt);
 616 #ifdef __XOP__
 617             vfeps            = _mm_frcz_ps(rt);
 618 #else
 619             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 620 #endif
 621             twovfeps         = _mm_add_ps(vfeps,vfeps);
 622             vfitab           = _mm_slli_epi32(vfitab,3);
 623
 624             /* REACTION-FIELD ELECTROSTATICS */
 625             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 626
 627             /* CUBIC SPLINE TABLE DISPERSION */
 628             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 629             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 630             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 631             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 632             _MM_TRANSPOSE4_PS(Y,F,G,H);
 633             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 634             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 635             fvdw6            = _mm_mul_ps(c6_00,FF);
 636
 637             /* CUBIC SPLINE TABLE REPULSION */
 638             vfitab           = _mm_add_epi32(vfitab,ifour);
 639             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 640             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 641             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 642             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 643             _MM_TRANSPOSE4_PS(Y,F,G,H);
 644             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 645             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 646             fvdw12           = _mm_mul_ps(c12_00,FF);
 647             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 648
 649             fscal            = _mm_add_ps(felec,fvdw);
 650
 651              /* Update vectorial force */
 652             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 653             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 654             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 655
 656             fjptrA             = f+j_coord_offsetA;
 657             fjptrB             = f+j_coord_offsetB;
 658             fjptrC             = f+j_coord_offsetC;
 659             fjptrD             = f+j_coord_offsetD;
 660             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 661                                                    _mm_mul_ps(dx00,fscal),
 662                                                    _mm_mul_ps(dy00,fscal),
 663                                                    _mm_mul_ps(dz00,fscal));
 664
 665             /* Inner loop uses 57 flops */
 666         }
 667
 668         if(jidx<j_index_end)
 669         {
 670
 671             /* Get j neighbor index, and coordinate index */
 672             jnrlistA         = jjnr[jidx];
 673             jnrlistB         = jjnr[jidx+1];
 674             jnrlistC         = jjnr[jidx+2];
 675             jnrlistD         = jjnr[jidx+3];
 676             /* Sign of each element will be negative for non-real atoms.
 677              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 678              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 679              */
 680             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 681             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 682             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 683             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 684             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 685             j_coord_offsetA  = DIM*jnrA;
 686             j_coord_offsetB  = DIM*jnrB;
 687             j_coord_offsetC  = DIM*jnrC;
 688             j_coord_offsetD  = DIM*jnrD;
 689
 690             /* load j atom coordinates */
 691             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 692                                               x+j_coord_offsetC,x+j_coord_offsetD,
 693                                               &jx0,&jy0,&jz0);
 694
 695             /* Calculate displacement vector */
 696             dx00             = _mm_sub_ps(ix0,jx0);
 697             dy00             = _mm_sub_ps(iy0,jy0);
 698             dz00             = _mm_sub_ps(iz0,jz0);
 699
 700             /* Calculate squared distance and things based on it */
 701             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 702
 703             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 704
 705             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 706
 707             /* Load parameters for j particles */
 708             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 709                                                               charge+jnrC+0,charge+jnrD+0);
 710             vdwjidx0A        = 2*vdwtype[jnrA+0];
 711             vdwjidx0B        = 2*vdwtype[jnrB+0];
 712             vdwjidx0C        = 2*vdwtype[jnrC+0];
 713             vdwjidx0D        = 2*vdwtype[jnrD+0];
 714
 715             /**************************
 716              * CALCULATE INTERACTIONS *
 717              **************************/
 718
 719             r00              = _mm_mul_ps(rsq00,rinv00);
 720             r00              = _mm_andnot_ps(dummy_mask,r00);
 721
 722             /* Compute parameters for interactions between i and j atoms */
 723             qq00             = _mm_mul_ps(iq0,jq0);
 724             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 725                                          vdwparam+vdwioffset0+vdwjidx0B,
 726                                          vdwparam+vdwioffset0+vdwjidx0C,
 727                                          vdwparam+vdwioffset0+vdwjidx0D,
 728                                          &c6_00,&c12_00);
 729
 730             /* Calculate table index by multiplying r with table scale and truncate to integer */
 731             rt               = _mm_mul_ps(r00,vftabscale);
 732             vfitab           = _mm_cvttps_epi32(rt);
 733 #ifdef __XOP__
 734             vfeps            = _mm_frcz_ps(rt);
 735 #else
 736             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 737 #endif
 738             twovfeps         = _mm_add_ps(vfeps,vfeps);
 739             vfitab           = _mm_slli_epi32(vfitab,3);
 740
 741             /* REACTION-FIELD ELECTROSTATICS */
 742             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 743
 744             /* CUBIC SPLINE TABLE DISPERSION */
 745             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 746             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 747             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 748             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 749             _MM_TRANSPOSE4_PS(Y,F,G,H);
 750             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 751             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 752             fvdw6            = _mm_mul_ps(c6_00,FF);
 753
 754             /* CUBIC SPLINE TABLE REPULSION */
 755             vfitab           = _mm_add_epi32(vfitab,ifour);
 756             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 757             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 758             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 759             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 760             _MM_TRANSPOSE4_PS(Y,F,G,H);
 761             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 762             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 763             fvdw12           = _mm_mul_ps(c12_00,FF);
 764             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 765
 766             fscal            = _mm_add_ps(felec,fvdw);
 767
 768             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 769
 770              /* Update vectorial force */
 771             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 772             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 773             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 774
 775             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 776             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 777             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 778             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 779             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 780                                                    _mm_mul_ps(dx00,fscal),
 781                                                    _mm_mul_ps(dy00,fscal),
 782                                                    _mm_mul_ps(dz00,fscal));
 783
 784             /* Inner loop uses 58 flops */
 785         }
 786
 787         /* End of innermost loop */
 788
 789         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 790                                               f+i_coord_offset,fshift+i_shift_offset);
 791
 792         /* Increment number of inner iterations */
 793         inneriter                  += j_index_end - j_index_start;
 794
 795         /* Outer loop uses 7 flops */
 796     }
 797
 798     /* Increment number of outer iterations */
 799     outeriter        += nri;
 800
 801     /* Update outer/inner flops */
 802
 803     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);
 804 }