src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_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_ElecEw_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEw_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     __m128i          ewitab;
 103     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 104     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 105     real             *ewtab;
 106     __m128           dummy_mask,cutoff_mask;
 107     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 108     __m128           one     = _mm_set1_ps(1.0);
 109     __m128           two     = _mm_set1_ps(2.0);
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = _mm_set1_ps(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     vftab            = kernel_data->table_vdw->data;
 128     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 129
 130     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 131     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 132     beta2            = _mm_mul_ps(beta,beta);
 133     beta3            = _mm_mul_ps(beta,beta2);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 136     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 137
 138     /* Avoid stupid compiler warnings */
 139     jnrA = jnrB = jnrC = jnrD = 0;
 140     j_coord_offsetA = 0;
 141     j_coord_offsetB = 0;
 142     j_coord_offsetC = 0;
 143     j_coord_offsetD = 0;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     for(iidx=0;iidx<4*DIM;iidx++)
 149     {
 150         scratch[iidx] = 0.0;
 151     }
 152
 153     /* Start outer loop over neighborlists */
 154     for(iidx=0; iidx<nri; iidx++)
 155     {
 156         /* Load shift vector for this list */
 157         i_shift_offset   = DIM*shiftidx[iidx];
 158
 159         /* Load limits for loop over neighbors */
 160         j_index_start    = jindex[iidx];
 161         j_index_end      = jindex[iidx+1];
 162
 163         /* Get outer coordinate index */
 164         inr              = iinr[iidx];
 165         i_coord_offset   = DIM*inr;
 166
 167         /* Load i particle coords and add shift vector */
 168         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 169
 170         fix0             = _mm_setzero_ps();
 171         fiy0             = _mm_setzero_ps();
 172         fiz0             = _mm_setzero_ps();
 173
 174         /* Load parameters for i particles */
 175         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 176         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 177
 178         /* Reset potential sums */
 179         velecsum         = _mm_setzero_ps();
 180         vvdwsum          = _mm_setzero_ps();
 181
 182         /* Start inner kernel loop */
 183         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 184         {
 185
 186             /* Get j neighbor index, and coordinate index */
 187             jnrA             = jjnr[jidx];
 188             jnrB             = jjnr[jidx+1];
 189             jnrC             = jjnr[jidx+2];
 190             jnrD             = jjnr[jidx+3];
 191             j_coord_offsetA  = DIM*jnrA;
 192             j_coord_offsetB  = DIM*jnrB;
 193             j_coord_offsetC  = DIM*jnrC;
 194             j_coord_offsetD  = DIM*jnrD;
 195
 196             /* load j atom coordinates */
 197             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 198                                               x+j_coord_offsetC,x+j_coord_offsetD,
 199                                               &jx0,&jy0,&jz0);
 200
 201             /* Calculate displacement vector */
 202             dx00             = _mm_sub_ps(ix0,jx0);
 203             dy00             = _mm_sub_ps(iy0,jy0);
 204             dz00             = _mm_sub_ps(iz0,jz0);
 205
 206             /* Calculate squared distance and things based on it */
 207             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 208
 209             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 210
 211             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 212
 213             /* Load parameters for j particles */
 214             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 215                                                               charge+jnrC+0,charge+jnrD+0);
 216             vdwjidx0A        = 2*vdwtype[jnrA+0];
 217             vdwjidx0B        = 2*vdwtype[jnrB+0];
 218             vdwjidx0C        = 2*vdwtype[jnrC+0];
 219             vdwjidx0D        = 2*vdwtype[jnrD+0];
 220
 221             /**************************
 222              * CALCULATE INTERACTIONS *
 223              **************************/
 224
 225             r00              = _mm_mul_ps(rsq00,rinv00);
 226
 227             /* Compute parameters for interactions between i and j atoms */
 228             qq00             = _mm_mul_ps(iq0,jq0);
 229             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 230                                          vdwparam+vdwioffset0+vdwjidx0B,
 231                                          vdwparam+vdwioffset0+vdwjidx0C,
 232                                          vdwparam+vdwioffset0+vdwjidx0D,
 233                                          &c6_00,&c12_00);
 234
 235             /* Calculate table index by multiplying r with table scale and truncate to integer */
 236             rt               = _mm_mul_ps(r00,vftabscale);
 237             vfitab           = _mm_cvttps_epi32(rt);
 238 #ifdef __XOP__
 239             vfeps            = _mm_frcz_ps(rt);
 240 #else
 241             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 242 #endif
 243             twovfeps         = _mm_add_ps(vfeps,vfeps);
 244             vfitab           = _mm_slli_epi32(vfitab,3);
 245
 246             /* EWALD ELECTROSTATICS */
 247
 248             /* Analytical PME correction */
 249             zeta2            = _mm_mul_ps(beta2,rsq00);
 250             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 251             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 252             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 253             felec            = _mm_mul_ps(qq00,felec);
 254             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 255             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 256             velec            = _mm_mul_ps(qq00,velec);
 257
 258             /* CUBIC SPLINE TABLE DISPERSION */
 259             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 260             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 261             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 262             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 263             _MM_TRANSPOSE4_PS(Y,F,G,H);
 264             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 265             VV               = _mm_macc_ps(vfeps,Fp,Y);
 266             vvdw6            = _mm_mul_ps(c6_00,VV);
 267             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 268             fvdw6            = _mm_mul_ps(c6_00,FF);
 269
 270             /* CUBIC SPLINE TABLE REPULSION */
 271             vfitab           = _mm_add_epi32(vfitab,ifour);
 272             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 273             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 274             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 275             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 276             _MM_TRANSPOSE4_PS(Y,F,G,H);
 277             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 278             VV               = _mm_macc_ps(vfeps,Fp,Y);
 279             vvdw12           = _mm_mul_ps(c12_00,VV);
 280             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 281             fvdw12           = _mm_mul_ps(c12_00,FF);
 282             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 283             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 284
 285             /* Update potential sum for this i atom from the interaction with this j atom. */
 286             velecsum         = _mm_add_ps(velecsum,velec);
 287             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 288
 289             fscal            = _mm_add_ps(felec,fvdw);
 290
 291              /* Update vectorial force */
 292             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 293             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 294             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 295
 296             fjptrA             = f+j_coord_offsetA;
 297             fjptrB             = f+j_coord_offsetB;
 298             fjptrC             = f+j_coord_offsetC;
 299             fjptrD             = f+j_coord_offsetD;
 300             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 301                                                    _mm_mul_ps(dx00,fscal),
 302                                                    _mm_mul_ps(dy00,fscal),
 303                                                    _mm_mul_ps(dz00,fscal));
 304
 305             /* Inner loop uses 63 flops */
 306         }
 307
 308         if(jidx<j_index_end)
 309         {
 310
 311             /* Get j neighbor index, and coordinate index */
 312             jnrlistA         = jjnr[jidx];
 313             jnrlistB         = jjnr[jidx+1];
 314             jnrlistC         = jjnr[jidx+2];
 315             jnrlistD         = jjnr[jidx+3];
 316             /* Sign of each element will be negative for non-real atoms.
 317              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 318              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 319              */
 320             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 321             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 322             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 323             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 324             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 325             j_coord_offsetA  = DIM*jnrA;
 326             j_coord_offsetB  = DIM*jnrB;
 327             j_coord_offsetC  = DIM*jnrC;
 328             j_coord_offsetD  = DIM*jnrD;
 329
 330             /* load j atom coordinates */
 331             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 332                                               x+j_coord_offsetC,x+j_coord_offsetD,
 333                                               &jx0,&jy0,&jz0);
 334
 335             /* Calculate displacement vector */
 336             dx00             = _mm_sub_ps(ix0,jx0);
 337             dy00             = _mm_sub_ps(iy0,jy0);
 338             dz00             = _mm_sub_ps(iz0,jz0);
 339
 340             /* Calculate squared distance and things based on it */
 341             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 342
 343             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 344
 345             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 346
 347             /* Load parameters for j particles */
 348             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 349                                                               charge+jnrC+0,charge+jnrD+0);
 350             vdwjidx0A        = 2*vdwtype[jnrA+0];
 351             vdwjidx0B        = 2*vdwtype[jnrB+0];
 352             vdwjidx0C        = 2*vdwtype[jnrC+0];
 353             vdwjidx0D        = 2*vdwtype[jnrD+0];
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             r00              = _mm_mul_ps(rsq00,rinv00);
 360             r00              = _mm_andnot_ps(dummy_mask,r00);
 361
 362             /* Compute parameters for interactions between i and j atoms */
 363             qq00             = _mm_mul_ps(iq0,jq0);
 364             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 365                                          vdwparam+vdwioffset0+vdwjidx0B,
 366                                          vdwparam+vdwioffset0+vdwjidx0C,
 367                                          vdwparam+vdwioffset0+vdwjidx0D,
 368                                          &c6_00,&c12_00);
 369
 370             /* Calculate table index by multiplying r with table scale and truncate to integer */
 371             rt               = _mm_mul_ps(r00,vftabscale);
 372             vfitab           = _mm_cvttps_epi32(rt);
 373 #ifdef __XOP__
 374             vfeps            = _mm_frcz_ps(rt);
 375 #else
 376             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 377 #endif
 378             twovfeps         = _mm_add_ps(vfeps,vfeps);
 379             vfitab           = _mm_slli_epi32(vfitab,3);
 380
 381             /* EWALD ELECTROSTATICS */
 382
 383             /* Analytical PME correction */
 384             zeta2            = _mm_mul_ps(beta2,rsq00);
 385             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 386             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 387             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 388             felec            = _mm_mul_ps(qq00,felec);
 389             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 390             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 391             velec            = _mm_mul_ps(qq00,velec);
 392
 393             /* CUBIC SPLINE TABLE DISPERSION */
 394             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 395             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 396             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 397             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 398             _MM_TRANSPOSE4_PS(Y,F,G,H);
 399             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 400             VV               = _mm_macc_ps(vfeps,Fp,Y);
 401             vvdw6            = _mm_mul_ps(c6_00,VV);
 402             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 403             fvdw6            = _mm_mul_ps(c6_00,FF);
 404
 405             /* CUBIC SPLINE TABLE REPULSION */
 406             vfitab           = _mm_add_epi32(vfitab,ifour);
 407             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 408             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 409             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 410             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 411             _MM_TRANSPOSE4_PS(Y,F,G,H);
 412             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 413             VV               = _mm_macc_ps(vfeps,Fp,Y);
 414             vvdw12           = _mm_mul_ps(c12_00,VV);
 415             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 416             fvdw12           = _mm_mul_ps(c12_00,FF);
 417             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 418             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 419
 420             /* Update potential sum for this i atom from the interaction with this j atom. */
 421             velec            = _mm_andnot_ps(dummy_mask,velec);
 422             velecsum         = _mm_add_ps(velecsum,velec);
 423             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 424             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 425
 426             fscal            = _mm_add_ps(felec,fvdw);
 427
 428             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 429
 430              /* Update vectorial force */
 431             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 432             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 433             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 434
 435             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 436             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 437             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 438             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 439             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 440                                                    _mm_mul_ps(dx00,fscal),
 441                                                    _mm_mul_ps(dy00,fscal),
 442                                                    _mm_mul_ps(dz00,fscal));
 443
 444             /* Inner loop uses 64 flops */
 445         }
 446
 447         /* End of innermost loop */
 448
 449         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 450                                               f+i_coord_offset,fshift+i_shift_offset);
 451
 452         ggid                        = gid[iidx];
 453         /* Update potential energies */
 454         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 455         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 456
 457         /* Increment number of inner iterations */
 458         inneriter                  += j_index_end - j_index_start;
 459
 460         /* Outer loop uses 9 flops */
 461     }
 462
 463     /* Increment number of outer iterations */
 464     outeriter        += nri;
 465
 466     /* Update outer/inner flops */
 467
 468     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*64);
 469 }
 470 /*
 471  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 472  * Electrostatics interaction: Ewald
 473  * VdW interaction:            CubicSplineTable
 474  * Geometry:                   Particle-Particle
 475  * Calculate force/pot:        Force
 476  */
 477 void
 478 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 479                     (t_nblist                    * gmx_restrict       nlist,
 480                      rvec                        * gmx_restrict          xx,
 481                      rvec                        * gmx_restrict          ff,
 482                      t_forcerec                  * gmx_restrict          fr,
 483                      t_mdatoms                   * gmx_restrict     mdatoms,
 484                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 485                      t_nrnb                      * gmx_restrict        nrnb)
 486 {
 487     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 488      * just 0 for non-waters.
 489      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 490      * jnr indices corresponding to data put in the four positions in the SIMD register.
 491      */
 492     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 493     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 494     int              jnrA,jnrB,jnrC,jnrD;
 495     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 496     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 497     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 498     real             rcutoff_scalar;
 499     real             *shiftvec,*fshift,*x,*f;
 500     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 501     real             scratch[4*DIM];
 502     __m128           fscal,rcutoff,rcutoff2,jidxall;
 503     int              vdwioffset0;
 504     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 505     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 506     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 507     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 508     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 509     real             *charge;
 510     int              nvdwtype;
 511     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 512     int              *vdwtype;
 513     real             *vdwparam;
 514     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 515     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 516     __m128i          vfitab;
 517     __m128i          ifour       = _mm_set1_epi32(4);
 518     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 519     real             *vftab;
 520     __m128i          ewitab;
 521     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 522     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 523     real             *ewtab;
 524     __m128           dummy_mask,cutoff_mask;
 525     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 526     __m128           one     = _mm_set1_ps(1.0);
 527     __m128           two     = _mm_set1_ps(2.0);
 528     x                = xx[0];
 529     f                = ff[0];
 530
 531     nri              = nlist->nri;
 532     iinr             = nlist->iinr;
 533     jindex           = nlist->jindex;
 534     jjnr             = nlist->jjnr;
 535     shiftidx         = nlist->shift;
 536     gid              = nlist->gid;
 537     shiftvec         = fr->shift_vec[0];
 538     fshift           = fr->fshift[0];
 539     facel            = _mm_set1_ps(fr->epsfac);
 540     charge           = mdatoms->chargeA;
 541     nvdwtype         = fr->ntype;
 542     vdwparam         = fr->nbfp;
 543     vdwtype          = mdatoms->typeA;
 544
 545     vftab            = kernel_data->table_vdw->data;
 546     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 547
 548     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 549     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 550     beta2            = _mm_mul_ps(beta,beta);
 551     beta3            = _mm_mul_ps(beta,beta2);
 552     ewtab            = fr->ic->tabq_coul_F;
 553     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 554     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 555
 556     /* Avoid stupid compiler warnings */
 557     jnrA = jnrB = jnrC = jnrD = 0;
 558     j_coord_offsetA = 0;
 559     j_coord_offsetB = 0;
 560     j_coord_offsetC = 0;
 561     j_coord_offsetD = 0;
 562
 563     outeriter        = 0;
 564     inneriter        = 0;
 565
 566     for(iidx=0;iidx<4*DIM;iidx++)
 567     {
 568         scratch[iidx] = 0.0;
 569     }
 570
 571     /* Start outer loop over neighborlists */
 572     for(iidx=0; iidx<nri; iidx++)
 573     {
 574         /* Load shift vector for this list */
 575         i_shift_offset   = DIM*shiftidx[iidx];
 576
 577         /* Load limits for loop over neighbors */
 578         j_index_start    = jindex[iidx];
 579         j_index_end      = jindex[iidx+1];
 580
 581         /* Get outer coordinate index */
 582         inr              = iinr[iidx];
 583         i_coord_offset   = DIM*inr;
 584
 585         /* Load i particle coords and add shift vector */
 586         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 587
 588         fix0             = _mm_setzero_ps();
 589         fiy0             = _mm_setzero_ps();
 590         fiz0             = _mm_setzero_ps();
 591
 592         /* Load parameters for i particles */
 593         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 594         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 595
 596         /* Start inner kernel loop */
 597         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 598         {
 599
 600             /* Get j neighbor index, and coordinate index */
 601             jnrA             = jjnr[jidx];
 602             jnrB             = jjnr[jidx+1];
 603             jnrC             = jjnr[jidx+2];
 604             jnrD             = jjnr[jidx+3];
 605             j_coord_offsetA  = DIM*jnrA;
 606             j_coord_offsetB  = DIM*jnrB;
 607             j_coord_offsetC  = DIM*jnrC;
 608             j_coord_offsetD  = DIM*jnrD;
 609
 610             /* load j atom coordinates */
 611             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 612                                               x+j_coord_offsetC,x+j_coord_offsetD,
 613                                               &jx0,&jy0,&jz0);
 614
 615             /* Calculate displacement vector */
 616             dx00             = _mm_sub_ps(ix0,jx0);
 617             dy00             = _mm_sub_ps(iy0,jy0);
 618             dz00             = _mm_sub_ps(iz0,jz0);
 619
 620             /* Calculate squared distance and things based on it */
 621             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 622
 623             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 624
 625             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 626
 627             /* Load parameters for j particles */
 628             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 629                                                               charge+jnrC+0,charge+jnrD+0);
 630             vdwjidx0A        = 2*vdwtype[jnrA+0];
 631             vdwjidx0B        = 2*vdwtype[jnrB+0];
 632             vdwjidx0C        = 2*vdwtype[jnrC+0];
 633             vdwjidx0D        = 2*vdwtype[jnrD+0];
 634
 635             /**************************
 636              * CALCULATE INTERACTIONS *
 637              **************************/
 638
 639             r00              = _mm_mul_ps(rsq00,rinv00);
 640
 641             /* Compute parameters for interactions between i and j atoms */
 642             qq00             = _mm_mul_ps(iq0,jq0);
 643             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 644                                          vdwparam+vdwioffset0+vdwjidx0B,
 645                                          vdwparam+vdwioffset0+vdwjidx0C,
 646                                          vdwparam+vdwioffset0+vdwjidx0D,
 647                                          &c6_00,&c12_00);
 648
 649             /* Calculate table index by multiplying r with table scale and truncate to integer */
 650             rt               = _mm_mul_ps(r00,vftabscale);
 651             vfitab           = _mm_cvttps_epi32(rt);
 652 #ifdef __XOP__
 653             vfeps            = _mm_frcz_ps(rt);
 654 #else
 655             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 656 #endif
 657             twovfeps         = _mm_add_ps(vfeps,vfeps);
 658             vfitab           = _mm_slli_epi32(vfitab,3);
 659
 660             /* EWALD ELECTROSTATICS */
 661
 662             /* Analytical PME correction */
 663             zeta2            = _mm_mul_ps(beta2,rsq00);
 664             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 665             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 666             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 667             felec            = _mm_mul_ps(qq00,felec);
 668
 669             /* CUBIC SPLINE TABLE DISPERSION */
 670             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 671             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 672             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 673             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 674             _MM_TRANSPOSE4_PS(Y,F,G,H);
 675             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 676             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 677             fvdw6            = _mm_mul_ps(c6_00,FF);
 678
 679             /* CUBIC SPLINE TABLE REPULSION */
 680             vfitab           = _mm_add_epi32(vfitab,ifour);
 681             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 682             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 683             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 684             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 685             _MM_TRANSPOSE4_PS(Y,F,G,H);
 686             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 687             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 688             fvdw12           = _mm_mul_ps(c12_00,FF);
 689             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 690
 691             fscal            = _mm_add_ps(felec,fvdw);
 692
 693              /* Update vectorial force */
 694             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 695             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 696             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 697
 698             fjptrA             = f+j_coord_offsetA;
 699             fjptrB             = f+j_coord_offsetB;
 700             fjptrC             = f+j_coord_offsetC;
 701             fjptrD             = f+j_coord_offsetD;
 702             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 703                                                    _mm_mul_ps(dx00,fscal),
 704                                                    _mm_mul_ps(dy00,fscal),
 705                                                    _mm_mul_ps(dz00,fscal));
 706
 707             /* Inner loop uses 54 flops */
 708         }
 709
 710         if(jidx<j_index_end)
 711         {
 712
 713             /* Get j neighbor index, and coordinate index */
 714             jnrlistA         = jjnr[jidx];
 715             jnrlistB         = jjnr[jidx+1];
 716             jnrlistC         = jjnr[jidx+2];
 717             jnrlistD         = jjnr[jidx+3];
 718             /* Sign of each element will be negative for non-real atoms.
 719              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 720              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 721              */
 722             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 723             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 724             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 725             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 726             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 727             j_coord_offsetA  = DIM*jnrA;
 728             j_coord_offsetB  = DIM*jnrB;
 729             j_coord_offsetC  = DIM*jnrC;
 730             j_coord_offsetD  = DIM*jnrD;
 731
 732             /* load j atom coordinates */
 733             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 734                                               x+j_coord_offsetC,x+j_coord_offsetD,
 735                                               &jx0,&jy0,&jz0);
 736
 737             /* Calculate displacement vector */
 738             dx00             = _mm_sub_ps(ix0,jx0);
 739             dy00             = _mm_sub_ps(iy0,jy0);
 740             dz00             = _mm_sub_ps(iz0,jz0);
 741
 742             /* Calculate squared distance and things based on it */
 743             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 744
 745             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 746
 747             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 748
 749             /* Load parameters for j particles */
 750             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 751                                                               charge+jnrC+0,charge+jnrD+0);
 752             vdwjidx0A        = 2*vdwtype[jnrA+0];
 753             vdwjidx0B        = 2*vdwtype[jnrB+0];
 754             vdwjidx0C        = 2*vdwtype[jnrC+0];
 755             vdwjidx0D        = 2*vdwtype[jnrD+0];
 756
 757             /**************************
 758              * CALCULATE INTERACTIONS *
 759              **************************/
 760
 761             r00              = _mm_mul_ps(rsq00,rinv00);
 762             r00              = _mm_andnot_ps(dummy_mask,r00);
 763
 764             /* Compute parameters for interactions between i and j atoms */
 765             qq00             = _mm_mul_ps(iq0,jq0);
 766             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 767                                          vdwparam+vdwioffset0+vdwjidx0B,
 768                                          vdwparam+vdwioffset0+vdwjidx0C,
 769                                          vdwparam+vdwioffset0+vdwjidx0D,
 770                                          &c6_00,&c12_00);
 771
 772             /* Calculate table index by multiplying r with table scale and truncate to integer */
 773             rt               = _mm_mul_ps(r00,vftabscale);
 774             vfitab           = _mm_cvttps_epi32(rt);
 775 #ifdef __XOP__
 776             vfeps            = _mm_frcz_ps(rt);
 777 #else
 778             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 779 #endif
 780             twovfeps         = _mm_add_ps(vfeps,vfeps);
 781             vfitab           = _mm_slli_epi32(vfitab,3);
 782
 783             /* EWALD ELECTROSTATICS */
 784
 785             /* Analytical PME correction */
 786             zeta2            = _mm_mul_ps(beta2,rsq00);
 787             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 788             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 789             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 790             felec            = _mm_mul_ps(qq00,felec);
 791
 792             /* CUBIC SPLINE TABLE DISPERSION */
 793             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 794             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 795             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 796             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 797             _MM_TRANSPOSE4_PS(Y,F,G,H);
 798             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 799             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 800             fvdw6            = _mm_mul_ps(c6_00,FF);
 801
 802             /* CUBIC SPLINE TABLE REPULSION */
 803             vfitab           = _mm_add_epi32(vfitab,ifour);
 804             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 805             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 806             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 807             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 808             _MM_TRANSPOSE4_PS(Y,F,G,H);
 809             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 810             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 811             fvdw12           = _mm_mul_ps(c12_00,FF);
 812             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 813
 814             fscal            = _mm_add_ps(felec,fvdw);
 815
 816             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 817
 818              /* Update vectorial force */
 819             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 820             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 821             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 822
 823             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 824             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 825             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 826             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 827             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 828                                                    _mm_mul_ps(dx00,fscal),
 829                                                    _mm_mul_ps(dy00,fscal),
 830                                                    _mm_mul_ps(dz00,fscal));
 831
 832             /* Inner loop uses 55 flops */
 833         }
 834
 835         /* End of innermost loop */
 836
 837         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 838                                               f+i_coord_offset,fshift+i_shift_offset);
 839
 840         /* Increment number of inner iterations */
 841         inneriter                  += j_index_end - j_index_start;
 842
 843         /* Outer loop uses 7 flops */
 844     }
 845
 846     /* Increment number of outer iterations */
 847     outeriter        += nri;
 848
 849     /* Update outer/inner flops */
 850
 851     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 852 }