src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_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_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_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     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     vftab            = kernel_data->table_vdw->data;
 122     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 123
 124     /* Avoid stupid compiler warnings */
 125     jnrA = jnrB = jnrC = jnrD = 0;
 126     j_coord_offsetA = 0;
 127     j_coord_offsetB = 0;
 128     j_coord_offsetC = 0;
 129     j_coord_offsetD = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm_setzero_ps();
 157         fiy0             = _mm_setzero_ps();
 158         fiz0             = _mm_setzero_ps();
 159
 160         /* Load parameters for i particles */
 161         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 162         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_ps();
 166         vvdwsum          = _mm_setzero_ps();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             jnrC             = jjnr[jidx+2];
 176             jnrD             = jjnr[jidx+3];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179             j_coord_offsetC  = DIM*jnrC;
 180             j_coord_offsetD  = DIM*jnrD;
 181
 182             /* load j atom coordinates */
 183             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 184                                               x+j_coord_offsetC,x+j_coord_offsetD,
 185                                               &jx0,&jy0,&jz0);
 186
 187             /* Calculate displacement vector */
 188             dx00             = _mm_sub_ps(ix0,jx0);
 189             dy00             = _mm_sub_ps(iy0,jy0);
 190             dz00             = _mm_sub_ps(iz0,jz0);
 191
 192             /* Calculate squared distance and things based on it */
 193             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 194
 195             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 196
 197             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 198
 199             /* Load parameters for j particles */
 200             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 201                                                               charge+jnrC+0,charge+jnrD+0);
 202             vdwjidx0A        = 2*vdwtype[jnrA+0];
 203             vdwjidx0B        = 2*vdwtype[jnrB+0];
 204             vdwjidx0C        = 2*vdwtype[jnrC+0];
 205             vdwjidx0D        = 2*vdwtype[jnrD+0];
 206
 207             /**************************
 208              * CALCULATE INTERACTIONS *
 209              **************************/
 210
 211             r00              = _mm_mul_ps(rsq00,rinv00);
 212
 213             /* Compute parameters for interactions between i and j atoms */
 214             qq00             = _mm_mul_ps(iq0,jq0);
 215             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 216                                          vdwparam+vdwioffset0+vdwjidx0B,
 217                                          vdwparam+vdwioffset0+vdwjidx0C,
 218                                          vdwparam+vdwioffset0+vdwjidx0D,
 219                                          &c6_00,&c12_00);
 220
 221             /* Calculate table index by multiplying r with table scale and truncate to integer */
 222             rt               = _mm_mul_ps(r00,vftabscale);
 223             vfitab           = _mm_cvttps_epi32(rt);
 224 #ifdef __XOP__
 225             vfeps            = _mm_frcz_ps(rt);
 226 #else
 227             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 228 #endif
 229             twovfeps         = _mm_add_ps(vfeps,vfeps);
 230             vfitab           = _mm_slli_epi32(vfitab,3);
 231
 232             /* COULOMB ELECTROSTATICS */
 233             velec            = _mm_mul_ps(qq00,rinv00);
 234             felec            = _mm_mul_ps(velec,rinvsq00);
 235
 236             /* CUBIC SPLINE TABLE DISPERSION */
 237             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 238             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 239             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 240             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 241             _MM_TRANSPOSE4_PS(Y,F,G,H);
 242             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 243             VV               = _mm_macc_ps(vfeps,Fp,Y);
 244             vvdw6            = _mm_mul_ps(c6_00,VV);
 245             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 246             fvdw6            = _mm_mul_ps(c6_00,FF);
 247
 248             /* CUBIC SPLINE TABLE REPULSION */
 249             vfitab           = _mm_add_epi32(vfitab,ifour);
 250             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 251             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 252             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 253             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 254             _MM_TRANSPOSE4_PS(Y,F,G,H);
 255             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 256             VV               = _mm_macc_ps(vfeps,Fp,Y);
 257             vvdw12           = _mm_mul_ps(c12_00,VV);
 258             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 259             fvdw12           = _mm_mul_ps(c12_00,FF);
 260             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 261             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 262
 263             /* Update potential sum for this i atom from the interaction with this j atom. */
 264             velecsum         = _mm_add_ps(velecsum,velec);
 265             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 266
 267             fscal            = _mm_add_ps(felec,fvdw);
 268
 269              /* Update vectorial force */
 270             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 271             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 272             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 273
 274             fjptrA             = f+j_coord_offsetA;
 275             fjptrB             = f+j_coord_offsetB;
 276             fjptrC             = f+j_coord_offsetC;
 277             fjptrD             = f+j_coord_offsetD;
 278             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 279                                                    _mm_mul_ps(dx00,fscal),
 280                                                    _mm_mul_ps(dy00,fscal),
 281                                                    _mm_mul_ps(dz00,fscal));
 282
 283             /* Inner loop uses 66 flops */
 284         }
 285
 286         if(jidx<j_index_end)
 287         {
 288
 289             /* Get j neighbor index, and coordinate index */
 290             jnrlistA         = jjnr[jidx];
 291             jnrlistB         = jjnr[jidx+1];
 292             jnrlistC         = jjnr[jidx+2];
 293             jnrlistD         = jjnr[jidx+3];
 294             /* Sign of each element will be negative for non-real atoms.
 295              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 296              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 297              */
 298             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 299             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 300             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 301             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 302             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 303             j_coord_offsetA  = DIM*jnrA;
 304             j_coord_offsetB  = DIM*jnrB;
 305             j_coord_offsetC  = DIM*jnrC;
 306             j_coord_offsetD  = DIM*jnrD;
 307
 308             /* load j atom coordinates */
 309             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 310                                               x+j_coord_offsetC,x+j_coord_offsetD,
 311                                               &jx0,&jy0,&jz0);
 312
 313             /* Calculate displacement vector */
 314             dx00             = _mm_sub_ps(ix0,jx0);
 315             dy00             = _mm_sub_ps(iy0,jy0);
 316             dz00             = _mm_sub_ps(iz0,jz0);
 317
 318             /* Calculate squared distance and things based on it */
 319             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 320
 321             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 322
 323             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 324
 325             /* Load parameters for j particles */
 326             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 327                                                               charge+jnrC+0,charge+jnrD+0);
 328             vdwjidx0A        = 2*vdwtype[jnrA+0];
 329             vdwjidx0B        = 2*vdwtype[jnrB+0];
 330             vdwjidx0C        = 2*vdwtype[jnrC+0];
 331             vdwjidx0D        = 2*vdwtype[jnrD+0];
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             r00              = _mm_mul_ps(rsq00,rinv00);
 338             r00              = _mm_andnot_ps(dummy_mask,r00);
 339
 340             /* Compute parameters for interactions between i and j atoms */
 341             qq00             = _mm_mul_ps(iq0,jq0);
 342             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 343                                          vdwparam+vdwioffset0+vdwjidx0B,
 344                                          vdwparam+vdwioffset0+vdwjidx0C,
 345                                          vdwparam+vdwioffset0+vdwjidx0D,
 346                                          &c6_00,&c12_00);
 347
 348             /* Calculate table index by multiplying r with table scale and truncate to integer */
 349             rt               = _mm_mul_ps(r00,vftabscale);
 350             vfitab           = _mm_cvttps_epi32(rt);
 351 #ifdef __XOP__
 352             vfeps            = _mm_frcz_ps(rt);
 353 #else
 354             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 355 #endif
 356             twovfeps         = _mm_add_ps(vfeps,vfeps);
 357             vfitab           = _mm_slli_epi32(vfitab,3);
 358
 359             /* COULOMB ELECTROSTATICS */
 360             velec            = _mm_mul_ps(qq00,rinv00);
 361             felec            = _mm_mul_ps(velec,rinvsq00);
 362
 363             /* CUBIC SPLINE TABLE DISPERSION */
 364             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 365             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 366             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 367             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 368             _MM_TRANSPOSE4_PS(Y,F,G,H);
 369             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 370             VV               = _mm_macc_ps(vfeps,Fp,Y);
 371             vvdw6            = _mm_mul_ps(c6_00,VV);
 372             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 373             fvdw6            = _mm_mul_ps(c6_00,FF);
 374
 375             /* CUBIC SPLINE TABLE REPULSION */
 376             vfitab           = _mm_add_epi32(vfitab,ifour);
 377             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 378             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 379             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 380             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 381             _MM_TRANSPOSE4_PS(Y,F,G,H);
 382             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 383             VV               = _mm_macc_ps(vfeps,Fp,Y);
 384             vvdw12           = _mm_mul_ps(c12_00,VV);
 385             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 386             fvdw12           = _mm_mul_ps(c12_00,FF);
 387             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 388             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 389
 390             /* Update potential sum for this i atom from the interaction with this j atom. */
 391             velec            = _mm_andnot_ps(dummy_mask,velec);
 392             velecsum         = _mm_add_ps(velecsum,velec);
 393             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 394             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 395
 396             fscal            = _mm_add_ps(felec,fvdw);
 397
 398             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 399
 400              /* Update vectorial force */
 401             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 402             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 403             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 404
 405             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 406             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 407             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 408             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 409             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 410                                                    _mm_mul_ps(dx00,fscal),
 411                                                    _mm_mul_ps(dy00,fscal),
 412                                                    _mm_mul_ps(dz00,fscal));
 413
 414             /* Inner loop uses 67 flops */
 415         }
 416
 417         /* End of innermost loop */
 418
 419         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 420                                               f+i_coord_offset,fshift+i_shift_offset);
 421
 422         ggid                        = gid[iidx];
 423         /* Update potential energies */
 424         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 425         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 426
 427         /* Increment number of inner iterations */
 428         inneriter                  += j_index_end - j_index_start;
 429
 430         /* Outer loop uses 9 flops */
 431     }
 432
 433     /* Increment number of outer iterations */
 434     outeriter        += nri;
 435
 436     /* Update outer/inner flops */
 437
 438     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*67);
 439 }
 440 /*
 441  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 442  * Electrostatics interaction: Coulomb
 443  * VdW interaction:            CubicSplineTable
 444  * Geometry:                   Particle-Particle
 445  * Calculate force/pot:        Force
 446  */
 447 void
 448 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 449                     (t_nblist                    * gmx_restrict       nlist,
 450                      rvec                        * gmx_restrict          xx,
 451                      rvec                        * gmx_restrict          ff,
 452                      t_forcerec                  * gmx_restrict          fr,
 453                      t_mdatoms                   * gmx_restrict     mdatoms,
 454                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 455                      t_nrnb                      * gmx_restrict        nrnb)
 456 {
 457     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 458      * just 0 for non-waters.
 459      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 460      * jnr indices corresponding to data put in the four positions in the SIMD register.
 461      */
 462     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 463     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 464     int              jnrA,jnrB,jnrC,jnrD;
 465     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 466     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 467     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 468     real             rcutoff_scalar;
 469     real             *shiftvec,*fshift,*x,*f;
 470     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 471     real             scratch[4*DIM];
 472     __m128           fscal,rcutoff,rcutoff2,jidxall;
 473     int              vdwioffset0;
 474     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 475     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 476     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 477     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 478     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 479     real             *charge;
 480     int              nvdwtype;
 481     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 482     int              *vdwtype;
 483     real             *vdwparam;
 484     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 485     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 486     __m128i          vfitab;
 487     __m128i          ifour       = _mm_set1_epi32(4);
 488     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 489     real             *vftab;
 490     __m128           dummy_mask,cutoff_mask;
 491     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 492     __m128           one     = _mm_set1_ps(1.0);
 493     __m128           two     = _mm_set1_ps(2.0);
 494     x                = xx[0];
 495     f                = ff[0];
 496
 497     nri              = nlist->nri;
 498     iinr             = nlist->iinr;
 499     jindex           = nlist->jindex;
 500     jjnr             = nlist->jjnr;
 501     shiftidx         = nlist->shift;
 502     gid              = nlist->gid;
 503     shiftvec         = fr->shift_vec[0];
 504     fshift           = fr->fshift[0];
 505     facel            = _mm_set1_ps(fr->epsfac);
 506     charge           = mdatoms->chargeA;
 507     nvdwtype         = fr->ntype;
 508     vdwparam         = fr->nbfp;
 509     vdwtype          = mdatoms->typeA;
 510
 511     vftab            = kernel_data->table_vdw->data;
 512     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 513
 514     /* Avoid stupid compiler warnings */
 515     jnrA = jnrB = jnrC = jnrD = 0;
 516     j_coord_offsetA = 0;
 517     j_coord_offsetB = 0;
 518     j_coord_offsetC = 0;
 519     j_coord_offsetD = 0;
 520
 521     outeriter        = 0;
 522     inneriter        = 0;
 523
 524     for(iidx=0;iidx<4*DIM;iidx++)
 525     {
 526         scratch[iidx] = 0.0;
 527     }
 528
 529     /* Start outer loop over neighborlists */
 530     for(iidx=0; iidx<nri; iidx++)
 531     {
 532         /* Load shift vector for this list */
 533         i_shift_offset   = DIM*shiftidx[iidx];
 534
 535         /* Load limits for loop over neighbors */
 536         j_index_start    = jindex[iidx];
 537         j_index_end      = jindex[iidx+1];
 538
 539         /* Get outer coordinate index */
 540         inr              = iinr[iidx];
 541         i_coord_offset   = DIM*inr;
 542
 543         /* Load i particle coords and add shift vector */
 544         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 545
 546         fix0             = _mm_setzero_ps();
 547         fiy0             = _mm_setzero_ps();
 548         fiz0             = _mm_setzero_ps();
 549
 550         /* Load parameters for i particles */
 551         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 552         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 553
 554         /* Start inner kernel loop */
 555         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 556         {
 557
 558             /* Get j neighbor index, and coordinate index */
 559             jnrA             = jjnr[jidx];
 560             jnrB             = jjnr[jidx+1];
 561             jnrC             = jjnr[jidx+2];
 562             jnrD             = jjnr[jidx+3];
 563             j_coord_offsetA  = DIM*jnrA;
 564             j_coord_offsetB  = DIM*jnrB;
 565             j_coord_offsetC  = DIM*jnrC;
 566             j_coord_offsetD  = DIM*jnrD;
 567
 568             /* load j atom coordinates */
 569             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 570                                               x+j_coord_offsetC,x+j_coord_offsetD,
 571                                               &jx0,&jy0,&jz0);
 572
 573             /* Calculate displacement vector */
 574             dx00             = _mm_sub_ps(ix0,jx0);
 575             dy00             = _mm_sub_ps(iy0,jy0);
 576             dz00             = _mm_sub_ps(iz0,jz0);
 577
 578             /* Calculate squared distance and things based on it */
 579             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 580
 581             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 582
 583             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 584
 585             /* Load parameters for j particles */
 586             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 587                                                               charge+jnrC+0,charge+jnrD+0);
 588             vdwjidx0A        = 2*vdwtype[jnrA+0];
 589             vdwjidx0B        = 2*vdwtype[jnrB+0];
 590             vdwjidx0C        = 2*vdwtype[jnrC+0];
 591             vdwjidx0D        = 2*vdwtype[jnrD+0];
 592
 593             /**************************
 594              * CALCULATE INTERACTIONS *
 595              **************************/
 596
 597             r00              = _mm_mul_ps(rsq00,rinv00);
 598
 599             /* Compute parameters for interactions between i and j atoms */
 600             qq00             = _mm_mul_ps(iq0,jq0);
 601             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 602                                          vdwparam+vdwioffset0+vdwjidx0B,
 603                                          vdwparam+vdwioffset0+vdwjidx0C,
 604                                          vdwparam+vdwioffset0+vdwjidx0D,
 605                                          &c6_00,&c12_00);
 606
 607             /* Calculate table index by multiplying r with table scale and truncate to integer */
 608             rt               = _mm_mul_ps(r00,vftabscale);
 609             vfitab           = _mm_cvttps_epi32(rt);
 610 #ifdef __XOP__
 611             vfeps            = _mm_frcz_ps(rt);
 612 #else
 613             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 614 #endif
 615             twovfeps         = _mm_add_ps(vfeps,vfeps);
 616             vfitab           = _mm_slli_epi32(vfitab,3);
 617
 618             /* COULOMB ELECTROSTATICS */
 619             velec            = _mm_mul_ps(qq00,rinv00);
 620             felec            = _mm_mul_ps(velec,rinvsq00);
 621
 622             /* CUBIC SPLINE TABLE DISPERSION */
 623             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 624             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 625             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 626             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 627             _MM_TRANSPOSE4_PS(Y,F,G,H);
 628             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 629             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 630             fvdw6            = _mm_mul_ps(c6_00,FF);
 631
 632             /* CUBIC SPLINE TABLE REPULSION */
 633             vfitab           = _mm_add_epi32(vfitab,ifour);
 634             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 635             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 636             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 637             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 638             _MM_TRANSPOSE4_PS(Y,F,G,H);
 639             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 640             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 641             fvdw12           = _mm_mul_ps(c12_00,FF);
 642             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 643
 644             fscal            = _mm_add_ps(felec,fvdw);
 645
 646              /* Update vectorial force */
 647             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 648             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 649             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 650
 651             fjptrA             = f+j_coord_offsetA;
 652             fjptrB             = f+j_coord_offsetB;
 653             fjptrC             = f+j_coord_offsetC;
 654             fjptrD             = f+j_coord_offsetD;
 655             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 656                                                    _mm_mul_ps(dx00,fscal),
 657                                                    _mm_mul_ps(dy00,fscal),
 658                                                    _mm_mul_ps(dz00,fscal));
 659
 660             /* Inner loop uses 57 flops */
 661         }
 662
 663         if(jidx<j_index_end)
 664         {
 665
 666             /* Get j neighbor index, and coordinate index */
 667             jnrlistA         = jjnr[jidx];
 668             jnrlistB         = jjnr[jidx+1];
 669             jnrlistC         = jjnr[jidx+2];
 670             jnrlistD         = jjnr[jidx+3];
 671             /* Sign of each element will be negative for non-real atoms.
 672              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 673              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 674              */
 675             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 676             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 677             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 678             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 679             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 680             j_coord_offsetA  = DIM*jnrA;
 681             j_coord_offsetB  = DIM*jnrB;
 682             j_coord_offsetC  = DIM*jnrC;
 683             j_coord_offsetD  = DIM*jnrD;
 684
 685             /* load j atom coordinates */
 686             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 687                                               x+j_coord_offsetC,x+j_coord_offsetD,
 688                                               &jx0,&jy0,&jz0);
 689
 690             /* Calculate displacement vector */
 691             dx00             = _mm_sub_ps(ix0,jx0);
 692             dy00             = _mm_sub_ps(iy0,jy0);
 693             dz00             = _mm_sub_ps(iz0,jz0);
 694
 695             /* Calculate squared distance and things based on it */
 696             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 697
 698             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 699
 700             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 701
 702             /* Load parameters for j particles */
 703             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 704                                                               charge+jnrC+0,charge+jnrD+0);
 705             vdwjidx0A        = 2*vdwtype[jnrA+0];
 706             vdwjidx0B        = 2*vdwtype[jnrB+0];
 707             vdwjidx0C        = 2*vdwtype[jnrC+0];
 708             vdwjidx0D        = 2*vdwtype[jnrD+0];
 709
 710             /**************************
 711              * CALCULATE INTERACTIONS *
 712              **************************/
 713
 714             r00              = _mm_mul_ps(rsq00,rinv00);
 715             r00              = _mm_andnot_ps(dummy_mask,r00);
 716
 717             /* Compute parameters for interactions between i and j atoms */
 718             qq00             = _mm_mul_ps(iq0,jq0);
 719             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 720                                          vdwparam+vdwioffset0+vdwjidx0B,
 721                                          vdwparam+vdwioffset0+vdwjidx0C,
 722                                          vdwparam+vdwioffset0+vdwjidx0D,
 723                                          &c6_00,&c12_00);
 724
 725             /* Calculate table index by multiplying r with table scale and truncate to integer */
 726             rt               = _mm_mul_ps(r00,vftabscale);
 727             vfitab           = _mm_cvttps_epi32(rt);
 728 #ifdef __XOP__
 729             vfeps            = _mm_frcz_ps(rt);
 730 #else
 731             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 732 #endif
 733             twovfeps         = _mm_add_ps(vfeps,vfeps);
 734             vfitab           = _mm_slli_epi32(vfitab,3);
 735
 736             /* COULOMB ELECTROSTATICS */
 737             velec            = _mm_mul_ps(qq00,rinv00);
 738             felec            = _mm_mul_ps(velec,rinvsq00);
 739
 740             /* CUBIC SPLINE TABLE DISPERSION */
 741             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 742             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 743             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 744             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 745             _MM_TRANSPOSE4_PS(Y,F,G,H);
 746             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 747             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 748             fvdw6            = _mm_mul_ps(c6_00,FF);
 749
 750             /* CUBIC SPLINE TABLE REPULSION */
 751             vfitab           = _mm_add_epi32(vfitab,ifour);
 752             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 753             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 754             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 755             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 756             _MM_TRANSPOSE4_PS(Y,F,G,H);
 757             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 758             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 759             fvdw12           = _mm_mul_ps(c12_00,FF);
 760             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 761
 762             fscal            = _mm_add_ps(felec,fvdw);
 763
 764             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 765
 766              /* Update vectorial force */
 767             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 768             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 769             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 770
 771             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 772             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 773             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 774             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 775             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 776                                                    _mm_mul_ps(dx00,fscal),
 777                                                    _mm_mul_ps(dy00,fscal),
 778                                                    _mm_mul_ps(dz00,fscal));
 779
 780             /* Inner loop uses 58 flops */
 781         }
 782
 783         /* End of innermost loop */
 784
 785         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 786                                               f+i_coord_offset,fshift+i_shift_offset);
 787
 788         /* Increment number of inner iterations */
 789         inneriter                  += j_index_end - j_index_start;
 790
 791         /* Outer loop uses 7 flops */
 792     }
 793
 794     /* Increment number of outer iterations */
 795     outeriter        += nri;
 796
 797     /* Update outer/inner flops */
 798
 799     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);
 800 }