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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/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_ElecRF_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_ElecRF_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     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = jnrC = jnrD = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133     j_coord_offsetC = 0;
 134     j_coord_offsetD = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     for(iidx=0;iidx<4*DIM;iidx++)
 140     {
 141         scratch[iidx] = 0.0;
 142     }
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 160
 161         fix0             = _mm_setzero_ps();
 162         fiy0             = _mm_setzero_ps();
 163         fiz0             = _mm_setzero_ps();
 164
 165         /* Load parameters for i particles */
 166         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 167         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 168
 169         /* Reset potential sums */
 170         velecsum         = _mm_setzero_ps();
 171         vvdwsum          = _mm_setzero_ps();
 172
 173         /* Start inner kernel loop */
 174         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 175         {
 176
 177             /* Get j neighbor index, and coordinate index */
 178             jnrA             = jjnr[jidx];
 179             jnrB             = jjnr[jidx+1];
 180             jnrC             = jjnr[jidx+2];
 181             jnrD             = jjnr[jidx+3];
 182             j_coord_offsetA  = DIM*jnrA;
 183             j_coord_offsetB  = DIM*jnrB;
 184             j_coord_offsetC  = DIM*jnrC;
 185             j_coord_offsetD  = DIM*jnrD;
 186
 187             /* load j atom coordinates */
 188             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                               x+j_coord_offsetC,x+j_coord_offsetD,
 190                                               &jx0,&jy0,&jz0);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm_sub_ps(ix0,jx0);
 194             dy00             = _mm_sub_ps(iy0,jy0);
 195             dz00             = _mm_sub_ps(iz0,jz0);
 196
 197             /* Calculate squared distance and things based on it */
 198             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 199
 200             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 201
 202             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 203
 204             /* Load parameters for j particles */
 205             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 206                                                               charge+jnrC+0,charge+jnrD+0);
 207             vdwjidx0A        = 2*vdwtype[jnrA+0];
 208             vdwjidx0B        = 2*vdwtype[jnrB+0];
 209             vdwjidx0C        = 2*vdwtype[jnrC+0];
 210             vdwjidx0D        = 2*vdwtype[jnrD+0];
 211
 212             /**************************
 213              * CALCULATE INTERACTIONS *
 214              **************************/
 215
 216             r00              = _mm_mul_ps(rsq00,rinv00);
 217
 218             /* Compute parameters for interactions between i and j atoms */
 219             qq00             = _mm_mul_ps(iq0,jq0);
 220             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 221                                          vdwparam+vdwioffset0+vdwjidx0B,
 222                                          vdwparam+vdwioffset0+vdwjidx0C,
 223                                          vdwparam+vdwioffset0+vdwjidx0D,
 224                                          &c6_00,&c12_00);
 225
 226             /* Calculate table index by multiplying r with table scale and truncate to integer */
 227             rt               = _mm_mul_ps(r00,vftabscale);
 228             vfitab           = _mm_cvttps_epi32(rt);
 229 #ifdef __XOP__
 230             vfeps            = _mm_frcz_ps(rt);
 231 #else
 232             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 233 #endif
 234             twovfeps         = _mm_add_ps(vfeps,vfeps);
 235             vfitab           = _mm_slli_epi32(vfitab,3);
 236
 237             /* REACTION-FIELD ELECTROSTATICS */
 238             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 239             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 240
 241             /* CUBIC SPLINE TABLE DISPERSION */
 242             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 243             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 244             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 245             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 246             _MM_TRANSPOSE4_PS(Y,F,G,H);
 247             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 248             VV               = _mm_macc_ps(vfeps,Fp,Y);
 249             vvdw6            = _mm_mul_ps(c6_00,VV);
 250             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 251             fvdw6            = _mm_mul_ps(c6_00,FF);
 252
 253             /* CUBIC SPLINE TABLE REPULSION */
 254             vfitab           = _mm_add_epi32(vfitab,ifour);
 255             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 256             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 257             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 258             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 259             _MM_TRANSPOSE4_PS(Y,F,G,H);
 260             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 261             VV               = _mm_macc_ps(vfeps,Fp,Y);
 262             vvdw12           = _mm_mul_ps(c12_00,VV);
 263             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 264             fvdw12           = _mm_mul_ps(c12_00,FF);
 265             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 266             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 267
 268             /* Update potential sum for this i atom from the interaction with this j atom. */
 269             velecsum         = _mm_add_ps(velecsum,velec);
 270             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 271
 272             fscal            = _mm_add_ps(felec,fvdw);
 273
 274              /* Update vectorial force */
 275             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 276             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 277             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 278
 279             fjptrA             = f+j_coord_offsetA;
 280             fjptrB             = f+j_coord_offsetB;
 281             fjptrC             = f+j_coord_offsetC;
 282             fjptrD             = f+j_coord_offsetD;
 283             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 284                                                    _mm_mul_ps(dx00,fscal),
 285                                                    _mm_mul_ps(dy00,fscal),
 286                                                    _mm_mul_ps(dz00,fscal));
 287
 288             /* Inner loop uses 70 flops */
 289         }
 290
 291         if(jidx<j_index_end)
 292         {
 293
 294             /* Get j neighbor index, and coordinate index */
 295             jnrlistA         = jjnr[jidx];
 296             jnrlistB         = jjnr[jidx+1];
 297             jnrlistC         = jjnr[jidx+2];
 298             jnrlistD         = jjnr[jidx+3];
 299             /* Sign of each element will be negative for non-real atoms.
 300              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 301              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 302              */
 303             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 304             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 305             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 306             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 307             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 308             j_coord_offsetA  = DIM*jnrA;
 309             j_coord_offsetB  = DIM*jnrB;
 310             j_coord_offsetC  = DIM*jnrC;
 311             j_coord_offsetD  = DIM*jnrD;
 312
 313             /* load j atom coordinates */
 314             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 315                                               x+j_coord_offsetC,x+j_coord_offsetD,
 316                                               &jx0,&jy0,&jz0);
 317
 318             /* Calculate displacement vector */
 319             dx00             = _mm_sub_ps(ix0,jx0);
 320             dy00             = _mm_sub_ps(iy0,jy0);
 321             dz00             = _mm_sub_ps(iz0,jz0);
 322
 323             /* Calculate squared distance and things based on it */
 324             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 325
 326             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 327
 328             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 329
 330             /* Load parameters for j particles */
 331             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 332                                                               charge+jnrC+0,charge+jnrD+0);
 333             vdwjidx0A        = 2*vdwtype[jnrA+0];
 334             vdwjidx0B        = 2*vdwtype[jnrB+0];
 335             vdwjidx0C        = 2*vdwtype[jnrC+0];
 336             vdwjidx0D        = 2*vdwtype[jnrD+0];
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             r00              = _mm_mul_ps(rsq00,rinv00);
 343             r00              = _mm_andnot_ps(dummy_mask,r00);
 344
 345             /* Compute parameters for interactions between i and j atoms */
 346             qq00             = _mm_mul_ps(iq0,jq0);
 347             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 348                                          vdwparam+vdwioffset0+vdwjidx0B,
 349                                          vdwparam+vdwioffset0+vdwjidx0C,
 350                                          vdwparam+vdwioffset0+vdwjidx0D,
 351                                          &c6_00,&c12_00);
 352
 353             /* Calculate table index by multiplying r with table scale and truncate to integer */
 354             rt               = _mm_mul_ps(r00,vftabscale);
 355             vfitab           = _mm_cvttps_epi32(rt);
 356 #ifdef __XOP__
 357             vfeps            = _mm_frcz_ps(rt);
 358 #else
 359             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 360 #endif
 361             twovfeps         = _mm_add_ps(vfeps,vfeps);
 362             vfitab           = _mm_slli_epi32(vfitab,3);
 363
 364             /* REACTION-FIELD ELECTROSTATICS */
 365             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 366             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 367
 368             /* CUBIC SPLINE TABLE DISPERSION */
 369             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 370             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 371             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 372             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 373             _MM_TRANSPOSE4_PS(Y,F,G,H);
 374             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 375             VV               = _mm_macc_ps(vfeps,Fp,Y);
 376             vvdw6            = _mm_mul_ps(c6_00,VV);
 377             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 378             fvdw6            = _mm_mul_ps(c6_00,FF);
 379
 380             /* CUBIC SPLINE TABLE REPULSION */
 381             vfitab           = _mm_add_epi32(vfitab,ifour);
 382             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 383             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 384             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 385             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 386             _MM_TRANSPOSE4_PS(Y,F,G,H);
 387             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 388             VV               = _mm_macc_ps(vfeps,Fp,Y);
 389             vvdw12           = _mm_mul_ps(c12_00,VV);
 390             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 391             fvdw12           = _mm_mul_ps(c12_00,FF);
 392             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 393             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velec            = _mm_andnot_ps(dummy_mask,velec);
 397             velecsum         = _mm_add_ps(velecsum,velec);
 398             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 399             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 400
 401             fscal            = _mm_add_ps(felec,fvdw);
 402
 403             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 404
 405              /* Update vectorial force */
 406             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 407             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 408             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 409
 410             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 411             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 412             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 413             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 414             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 415                                                    _mm_mul_ps(dx00,fscal),
 416                                                    _mm_mul_ps(dy00,fscal),
 417                                                    _mm_mul_ps(dz00,fscal));
 418
 419             /* Inner loop uses 71 flops */
 420         }
 421
 422         /* End of innermost loop */
 423
 424         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 425                                               f+i_coord_offset,fshift+i_shift_offset);
 426
 427         ggid                        = gid[iidx];
 428         /* Update potential energies */
 429         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 430         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 431
 432         /* Increment number of inner iterations */
 433         inneriter                  += j_index_end - j_index_start;
 434
 435         /* Outer loop uses 9 flops */
 436     }
 437
 438     /* Increment number of outer iterations */
 439     outeriter        += nri;
 440
 441     /* Update outer/inner flops */
 442
 443     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
 444 }
 445 /*
 446  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 447  * Electrostatics interaction: ReactionField
 448  * VdW interaction:            CubicSplineTable
 449  * Geometry:                   Particle-Particle
 450  * Calculate force/pot:        Force
 451  */
 452 void
 453 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 454                     (t_nblist                    * gmx_restrict       nlist,
 455                      rvec                        * gmx_restrict          xx,
 456                      rvec                        * gmx_restrict          ff,
 457                      t_forcerec                  * gmx_restrict          fr,
 458                      t_mdatoms                   * gmx_restrict     mdatoms,
 459                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 460                      t_nrnb                      * gmx_restrict        nrnb)
 461 {
 462     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 463      * just 0 for non-waters.
 464      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 465      * jnr indices corresponding to data put in the four positions in the SIMD register.
 466      */
 467     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 468     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 469     int              jnrA,jnrB,jnrC,jnrD;
 470     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 471     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 472     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 473     real             rcutoff_scalar;
 474     real             *shiftvec,*fshift,*x,*f;
 475     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 476     real             scratch[4*DIM];
 477     __m128           fscal,rcutoff,rcutoff2,jidxall;
 478     int              vdwioffset0;
 479     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 480     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 481     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 482     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 483     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 484     real             *charge;
 485     int              nvdwtype;
 486     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 487     int              *vdwtype;
 488     real             *vdwparam;
 489     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 490     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 491     __m128i          vfitab;
 492     __m128i          ifour       = _mm_set1_epi32(4);
 493     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 494     real             *vftab;
 495     __m128           dummy_mask,cutoff_mask;
 496     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 497     __m128           one     = _mm_set1_ps(1.0);
 498     __m128           two     = _mm_set1_ps(2.0);
 499     x                = xx[0];
 500     f                = ff[0];
 501
 502     nri              = nlist->nri;
 503     iinr             = nlist->iinr;
 504     jindex           = nlist->jindex;
 505     jjnr             = nlist->jjnr;
 506     shiftidx         = nlist->shift;
 507     gid              = nlist->gid;
 508     shiftvec         = fr->shift_vec[0];
 509     fshift           = fr->fshift[0];
 510     facel            = _mm_set1_ps(fr->epsfac);
 511     charge           = mdatoms->chargeA;
 512     krf              = _mm_set1_ps(fr->ic->k_rf);
 513     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 514     crf              = _mm_set1_ps(fr->ic->c_rf);
 515     nvdwtype         = fr->ntype;
 516     vdwparam         = fr->nbfp;
 517     vdwtype          = mdatoms->typeA;
 518
 519     vftab            = kernel_data->table_vdw->data;
 520     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 521
 522     /* Avoid stupid compiler warnings */
 523     jnrA = jnrB = jnrC = jnrD = 0;
 524     j_coord_offsetA = 0;
 525     j_coord_offsetB = 0;
 526     j_coord_offsetC = 0;
 527     j_coord_offsetD = 0;
 528
 529     outeriter        = 0;
 530     inneriter        = 0;
 531
 532     for(iidx=0;iidx<4*DIM;iidx++)
 533     {
 534         scratch[iidx] = 0.0;
 535     }
 536
 537     /* Start outer loop over neighborlists */
 538     for(iidx=0; iidx<nri; iidx++)
 539     {
 540         /* Load shift vector for this list */
 541         i_shift_offset   = DIM*shiftidx[iidx];
 542
 543         /* Load limits for loop over neighbors */
 544         j_index_start    = jindex[iidx];
 545         j_index_end      = jindex[iidx+1];
 546
 547         /* Get outer coordinate index */
 548         inr              = iinr[iidx];
 549         i_coord_offset   = DIM*inr;
 550
 551         /* Load i particle coords and add shift vector */
 552         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 553
 554         fix0             = _mm_setzero_ps();
 555         fiy0             = _mm_setzero_ps();
 556         fiz0             = _mm_setzero_ps();
 557
 558         /* Load parameters for i particles */
 559         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 560         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 561
 562         /* Start inner kernel loop */
 563         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 564         {
 565
 566             /* Get j neighbor index, and coordinate index */
 567             jnrA             = jjnr[jidx];
 568             jnrB             = jjnr[jidx+1];
 569             jnrC             = jjnr[jidx+2];
 570             jnrD             = jjnr[jidx+3];
 571             j_coord_offsetA  = DIM*jnrA;
 572             j_coord_offsetB  = DIM*jnrB;
 573             j_coord_offsetC  = DIM*jnrC;
 574             j_coord_offsetD  = DIM*jnrD;
 575
 576             /* load j atom coordinates */
 577             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 578                                               x+j_coord_offsetC,x+j_coord_offsetD,
 579                                               &jx0,&jy0,&jz0);
 580
 581             /* Calculate displacement vector */
 582             dx00             = _mm_sub_ps(ix0,jx0);
 583             dy00             = _mm_sub_ps(iy0,jy0);
 584             dz00             = _mm_sub_ps(iz0,jz0);
 585
 586             /* Calculate squared distance and things based on it */
 587             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 588
 589             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 590
 591             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 592
 593             /* Load parameters for j particles */
 594             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 595                                                               charge+jnrC+0,charge+jnrD+0);
 596             vdwjidx0A        = 2*vdwtype[jnrA+0];
 597             vdwjidx0B        = 2*vdwtype[jnrB+0];
 598             vdwjidx0C        = 2*vdwtype[jnrC+0];
 599             vdwjidx0D        = 2*vdwtype[jnrD+0];
 600
 601             /**************************
 602              * CALCULATE INTERACTIONS *
 603              **************************/
 604
 605             r00              = _mm_mul_ps(rsq00,rinv00);
 606
 607             /* Compute parameters for interactions between i and j atoms */
 608             qq00             = _mm_mul_ps(iq0,jq0);
 609             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 610                                          vdwparam+vdwioffset0+vdwjidx0B,
 611                                          vdwparam+vdwioffset0+vdwjidx0C,
 612                                          vdwparam+vdwioffset0+vdwjidx0D,
 613                                          &c6_00,&c12_00);
 614
 615             /* Calculate table index by multiplying r with table scale and truncate to integer */
 616             rt               = _mm_mul_ps(r00,vftabscale);
 617             vfitab           = _mm_cvttps_epi32(rt);
 618 #ifdef __XOP__
 619             vfeps            = _mm_frcz_ps(rt);
 620 #else
 621             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 622 #endif
 623             twovfeps         = _mm_add_ps(vfeps,vfeps);
 624             vfitab           = _mm_slli_epi32(vfitab,3);
 625
 626             /* REACTION-FIELD ELECTROSTATICS */
 627             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 628
 629             /* CUBIC SPLINE TABLE DISPERSION */
 630             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 631             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 632             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 633             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 634             _MM_TRANSPOSE4_PS(Y,F,G,H);
 635             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 636             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 637             fvdw6            = _mm_mul_ps(c6_00,FF);
 638
 639             /* CUBIC SPLINE TABLE REPULSION */
 640             vfitab           = _mm_add_epi32(vfitab,ifour);
 641             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 642             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 643             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 644             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 645             _MM_TRANSPOSE4_PS(Y,F,G,H);
 646             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 647             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 648             fvdw12           = _mm_mul_ps(c12_00,FF);
 649             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 650
 651             fscal            = _mm_add_ps(felec,fvdw);
 652
 653              /* Update vectorial force */
 654             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 655             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 656             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 657
 658             fjptrA             = f+j_coord_offsetA;
 659             fjptrB             = f+j_coord_offsetB;
 660             fjptrC             = f+j_coord_offsetC;
 661             fjptrD             = f+j_coord_offsetD;
 662             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 663                                                    _mm_mul_ps(dx00,fscal),
 664                                                    _mm_mul_ps(dy00,fscal),
 665                                                    _mm_mul_ps(dz00,fscal));
 666
 667             /* Inner loop uses 57 flops */
 668         }
 669
 670         if(jidx<j_index_end)
 671         {
 672
 673             /* Get j neighbor index, and coordinate index */
 674             jnrlistA         = jjnr[jidx];
 675             jnrlistB         = jjnr[jidx+1];
 676             jnrlistC         = jjnr[jidx+2];
 677             jnrlistD         = jjnr[jidx+3];
 678             /* Sign of each element will be negative for non-real atoms.
 679              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 680              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 681              */
 682             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 683             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 684             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 685             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 686             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 687             j_coord_offsetA  = DIM*jnrA;
 688             j_coord_offsetB  = DIM*jnrB;
 689             j_coord_offsetC  = DIM*jnrC;
 690             j_coord_offsetD  = DIM*jnrD;
 691
 692             /* load j atom coordinates */
 693             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 694                                               x+j_coord_offsetC,x+j_coord_offsetD,
 695                                               &jx0,&jy0,&jz0);
 696
 697             /* Calculate displacement vector */
 698             dx00             = _mm_sub_ps(ix0,jx0);
 699             dy00             = _mm_sub_ps(iy0,jy0);
 700             dz00             = _mm_sub_ps(iz0,jz0);
 701
 702             /* Calculate squared distance and things based on it */
 703             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 704
 705             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 706
 707             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 708
 709             /* Load parameters for j particles */
 710             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 711                                                               charge+jnrC+0,charge+jnrD+0);
 712             vdwjidx0A        = 2*vdwtype[jnrA+0];
 713             vdwjidx0B        = 2*vdwtype[jnrB+0];
 714             vdwjidx0C        = 2*vdwtype[jnrC+0];
 715             vdwjidx0D        = 2*vdwtype[jnrD+0];
 716
 717             /**************************
 718              * CALCULATE INTERACTIONS *
 719              **************************/
 720
 721             r00              = _mm_mul_ps(rsq00,rinv00);
 722             r00              = _mm_andnot_ps(dummy_mask,r00);
 723
 724             /* Compute parameters for interactions between i and j atoms */
 725             qq00             = _mm_mul_ps(iq0,jq0);
 726             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 727                                          vdwparam+vdwioffset0+vdwjidx0B,
 728                                          vdwparam+vdwioffset0+vdwjidx0C,
 729                                          vdwparam+vdwioffset0+vdwjidx0D,
 730                                          &c6_00,&c12_00);
 731
 732             /* Calculate table index by multiplying r with table scale and truncate to integer */
 733             rt               = _mm_mul_ps(r00,vftabscale);
 734             vfitab           = _mm_cvttps_epi32(rt);
 735 #ifdef __XOP__
 736             vfeps            = _mm_frcz_ps(rt);
 737 #else
 738             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 739 #endif
 740             twovfeps         = _mm_add_ps(vfeps,vfeps);
 741             vfitab           = _mm_slli_epi32(vfitab,3);
 742
 743             /* REACTION-FIELD ELECTROSTATICS */
 744             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 745
 746             /* CUBIC SPLINE TABLE DISPERSION */
 747             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 748             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 749             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 750             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 751             _MM_TRANSPOSE4_PS(Y,F,G,H);
 752             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 753             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 754             fvdw6            = _mm_mul_ps(c6_00,FF);
 755
 756             /* CUBIC SPLINE TABLE REPULSION */
 757             vfitab           = _mm_add_epi32(vfitab,ifour);
 758             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 759             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 760             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 761             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 762             _MM_TRANSPOSE4_PS(Y,F,G,H);
 763             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 764             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 765             fvdw12           = _mm_mul_ps(c12_00,FF);
 766             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 767
 768             fscal            = _mm_add_ps(felec,fvdw);
 769
 770             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 771
 772              /* Update vectorial force */
 773             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 774             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 775             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 776
 777             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 778             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 779             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 780             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 781             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 782                                                    _mm_mul_ps(dx00,fscal),
 783                                                    _mm_mul_ps(dy00,fscal),
 784                                                    _mm_mul_ps(dz00,fscal));
 785
 786             /* Inner loop uses 58 flops */
 787         }
 788
 789         /* End of innermost loop */
 790
 791         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 792                                               f+i_coord_offset,fshift+i_shift_offset);
 793
 794         /* Increment number of inner iterations */
 795         inneriter                  += j_index_end - j_index_start;
 796
 797         /* Outer loop uses 7 flops */
 798     }
 799
 800     /* Increment number of outer iterations */
 801     outeriter        += nri;
 802
 803     /* Update outer/inner flops */
 804
 805     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);
 806 }