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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEw_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     __m128i          ewitab;
 101     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 102     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 103     real             *ewtab;
 104     __m128           dummy_mask,cutoff_mask;
 105     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 106     __m128           one     = _mm_set1_ps(1.0);
 107     __m128           two     = _mm_set1_ps(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_ps(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     vftab            = kernel_data->table_vdw->data;
 126     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 127
 128     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 129     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 130     beta2            = _mm_mul_ps(beta,beta);
 131     beta3            = _mm_mul_ps(beta,beta2);
 132     ewtab            = fr->ic->tabq_coul_FDV0;
 133     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 134     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 135
 136     /* Avoid stupid compiler warnings */
 137     jnrA = jnrB = jnrC = jnrD = 0;
 138     j_coord_offsetA = 0;
 139     j_coord_offsetB = 0;
 140     j_coord_offsetC = 0;
 141     j_coord_offsetD = 0;
 142
 143     outeriter        = 0;
 144     inneriter        = 0;
 145
 146     for(iidx=0;iidx<4*DIM;iidx++)
 147     {
 148         scratch[iidx] = 0.0;
 149     }
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 167
 168         fix0             = _mm_setzero_ps();
 169         fiy0             = _mm_setzero_ps();
 170         fiz0             = _mm_setzero_ps();
 171
 172         /* Load parameters for i particles */
 173         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 174         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 175
 176         /* Reset potential sums */
 177         velecsum         = _mm_setzero_ps();
 178         vvdwsum          = _mm_setzero_ps();
 179
 180         /* Start inner kernel loop */
 181         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 182         {
 183
 184             /* Get j neighbor index, and coordinate index */
 185             jnrA             = jjnr[jidx];
 186             jnrB             = jjnr[jidx+1];
 187             jnrC             = jjnr[jidx+2];
 188             jnrD             = jjnr[jidx+3];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191             j_coord_offsetC  = DIM*jnrC;
 192             j_coord_offsetD  = DIM*jnrD;
 193
 194             /* load j atom coordinates */
 195             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 196                                               x+j_coord_offsetC,x+j_coord_offsetD,
 197                                               &jx0,&jy0,&jz0);
 198
 199             /* Calculate displacement vector */
 200             dx00             = _mm_sub_ps(ix0,jx0);
 201             dy00             = _mm_sub_ps(iy0,jy0);
 202             dz00             = _mm_sub_ps(iz0,jz0);
 203
 204             /* Calculate squared distance and things based on it */
 205             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 206
 207             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 208
 209             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 210
 211             /* Load parameters for j particles */
 212             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 213                                                               charge+jnrC+0,charge+jnrD+0);
 214             vdwjidx0A        = 2*vdwtype[jnrA+0];
 215             vdwjidx0B        = 2*vdwtype[jnrB+0];
 216             vdwjidx0C        = 2*vdwtype[jnrC+0];
 217             vdwjidx0D        = 2*vdwtype[jnrD+0];
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             r00              = _mm_mul_ps(rsq00,rinv00);
 224
 225             /* Compute parameters for interactions between i and j atoms */
 226             qq00             = _mm_mul_ps(iq0,jq0);
 227             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 228                                          vdwparam+vdwioffset0+vdwjidx0B,
 229                                          vdwparam+vdwioffset0+vdwjidx0C,
 230                                          vdwparam+vdwioffset0+vdwjidx0D,
 231                                          &c6_00,&c12_00);
 232
 233             /* Calculate table index by multiplying r with table scale and truncate to integer */
 234             rt               = _mm_mul_ps(r00,vftabscale);
 235             vfitab           = _mm_cvttps_epi32(rt);
 236 #ifdef __XOP__
 237             vfeps            = _mm_frcz_ps(rt);
 238 #else
 239             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 240 #endif
 241             twovfeps         = _mm_add_ps(vfeps,vfeps);
 242             vfitab           = _mm_slli_epi32(vfitab,3);
 243
 244             /* EWALD ELECTROSTATICS */
 245
 246             /* Analytical PME correction */
 247             zeta2            = _mm_mul_ps(beta2,rsq00);
 248             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 249             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 250             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 251             felec            = _mm_mul_ps(qq00,felec);
 252             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 253             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 254             velec            = _mm_mul_ps(qq00,velec);
 255
 256             /* CUBIC SPLINE TABLE DISPERSION */
 257             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 258             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 259             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 260             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 261             _MM_TRANSPOSE4_PS(Y,F,G,H);
 262             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 263             VV               = _mm_macc_ps(vfeps,Fp,Y);
 264             vvdw6            = _mm_mul_ps(c6_00,VV);
 265             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 266             fvdw6            = _mm_mul_ps(c6_00,FF);
 267
 268             /* CUBIC SPLINE TABLE REPULSION */
 269             vfitab           = _mm_add_epi32(vfitab,ifour);
 270             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 271             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 272             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 273             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 274             _MM_TRANSPOSE4_PS(Y,F,G,H);
 275             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 276             VV               = _mm_macc_ps(vfeps,Fp,Y);
 277             vvdw12           = _mm_mul_ps(c12_00,VV);
 278             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 279             fvdw12           = _mm_mul_ps(c12_00,FF);
 280             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 281             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 282
 283             /* Update potential sum for this i atom from the interaction with this j atom. */
 284             velecsum         = _mm_add_ps(velecsum,velec);
 285             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 286
 287             fscal            = _mm_add_ps(felec,fvdw);
 288
 289              /* Update vectorial force */
 290             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 291             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 292             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 293
 294             fjptrA             = f+j_coord_offsetA;
 295             fjptrB             = f+j_coord_offsetB;
 296             fjptrC             = f+j_coord_offsetC;
 297             fjptrD             = f+j_coord_offsetD;
 298             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 299                                                    _mm_mul_ps(dx00,fscal),
 300                                                    _mm_mul_ps(dy00,fscal),
 301                                                    _mm_mul_ps(dz00,fscal));
 302
 303             /* Inner loop uses 63 flops */
 304         }
 305
 306         if(jidx<j_index_end)
 307         {
 308
 309             /* Get j neighbor index, and coordinate index */
 310             jnrlistA         = jjnr[jidx];
 311             jnrlistB         = jjnr[jidx+1];
 312             jnrlistC         = jjnr[jidx+2];
 313             jnrlistD         = jjnr[jidx+3];
 314             /* Sign of each element will be negative for non-real atoms.
 315              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 316              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 317              */
 318             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 319             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 320             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 321             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 322             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 323             j_coord_offsetA  = DIM*jnrA;
 324             j_coord_offsetB  = DIM*jnrB;
 325             j_coord_offsetC  = DIM*jnrC;
 326             j_coord_offsetD  = DIM*jnrD;
 327
 328             /* load j atom coordinates */
 329             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 330                                               x+j_coord_offsetC,x+j_coord_offsetD,
 331                                               &jx0,&jy0,&jz0);
 332
 333             /* Calculate displacement vector */
 334             dx00             = _mm_sub_ps(ix0,jx0);
 335             dy00             = _mm_sub_ps(iy0,jy0);
 336             dz00             = _mm_sub_ps(iz0,jz0);
 337
 338             /* Calculate squared distance and things based on it */
 339             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 340
 341             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 342
 343             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 344
 345             /* Load parameters for j particles */
 346             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 347                                                               charge+jnrC+0,charge+jnrD+0);
 348             vdwjidx0A        = 2*vdwtype[jnrA+0];
 349             vdwjidx0B        = 2*vdwtype[jnrB+0];
 350             vdwjidx0C        = 2*vdwtype[jnrC+0];
 351             vdwjidx0D        = 2*vdwtype[jnrD+0];
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             r00              = _mm_mul_ps(rsq00,rinv00);
 358             r00              = _mm_andnot_ps(dummy_mask,r00);
 359
 360             /* Compute parameters for interactions between i and j atoms */
 361             qq00             = _mm_mul_ps(iq0,jq0);
 362             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 363                                          vdwparam+vdwioffset0+vdwjidx0B,
 364                                          vdwparam+vdwioffset0+vdwjidx0C,
 365                                          vdwparam+vdwioffset0+vdwjidx0D,
 366                                          &c6_00,&c12_00);
 367
 368             /* Calculate table index by multiplying r with table scale and truncate to integer */
 369             rt               = _mm_mul_ps(r00,vftabscale);
 370             vfitab           = _mm_cvttps_epi32(rt);
 371 #ifdef __XOP__
 372             vfeps            = _mm_frcz_ps(rt);
 373 #else
 374             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 375 #endif
 376             twovfeps         = _mm_add_ps(vfeps,vfeps);
 377             vfitab           = _mm_slli_epi32(vfitab,3);
 378
 379             /* EWALD ELECTROSTATICS */
 380
 381             /* Analytical PME correction */
 382             zeta2            = _mm_mul_ps(beta2,rsq00);
 383             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 384             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 385             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 386             felec            = _mm_mul_ps(qq00,felec);
 387             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 388             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 389             velec            = _mm_mul_ps(qq00,velec);
 390
 391             /* CUBIC SPLINE TABLE DISPERSION */
 392             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 393             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 394             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 395             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 396             _MM_TRANSPOSE4_PS(Y,F,G,H);
 397             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 398             VV               = _mm_macc_ps(vfeps,Fp,Y);
 399             vvdw6            = _mm_mul_ps(c6_00,VV);
 400             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 401             fvdw6            = _mm_mul_ps(c6_00,FF);
 402
 403             /* CUBIC SPLINE TABLE REPULSION */
 404             vfitab           = _mm_add_epi32(vfitab,ifour);
 405             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 406             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 407             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 408             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 409             _MM_TRANSPOSE4_PS(Y,F,G,H);
 410             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 411             VV               = _mm_macc_ps(vfeps,Fp,Y);
 412             vvdw12           = _mm_mul_ps(c12_00,VV);
 413             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 414             fvdw12           = _mm_mul_ps(c12_00,FF);
 415             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 416             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velec            = _mm_andnot_ps(dummy_mask,velec);
 420             velecsum         = _mm_add_ps(velecsum,velec);
 421             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 422             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 423
 424             fscal            = _mm_add_ps(felec,fvdw);
 425
 426             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 427
 428              /* Update vectorial force */
 429             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 430             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 431             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 432
 433             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 434             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 435             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 436             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 437             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 438                                                    _mm_mul_ps(dx00,fscal),
 439                                                    _mm_mul_ps(dy00,fscal),
 440                                                    _mm_mul_ps(dz00,fscal));
 441
 442             /* Inner loop uses 64 flops */
 443         }
 444
 445         /* End of innermost loop */
 446
 447         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 448                                               f+i_coord_offset,fshift+i_shift_offset);
 449
 450         ggid                        = gid[iidx];
 451         /* Update potential energies */
 452         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 453         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 454
 455         /* Increment number of inner iterations */
 456         inneriter                  += j_index_end - j_index_start;
 457
 458         /* Outer loop uses 9 flops */
 459     }
 460
 461     /* Increment number of outer iterations */
 462     outeriter        += nri;
 463
 464     /* Update outer/inner flops */
 465
 466     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*64);
 467 }
 468 /*
 469  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 470  * Electrostatics interaction: Ewald
 471  * VdW interaction:            CubicSplineTable
 472  * Geometry:                   Particle-Particle
 473  * Calculate force/pot:        Force
 474  */
 475 void
 476 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 477                     (t_nblist                    * gmx_restrict       nlist,
 478                      rvec                        * gmx_restrict          xx,
 479                      rvec                        * gmx_restrict          ff,
 480                      t_forcerec                  * gmx_restrict          fr,
 481                      t_mdatoms                   * gmx_restrict     mdatoms,
 482                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 483                      t_nrnb                      * gmx_restrict        nrnb)
 484 {
 485     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 486      * just 0 for non-waters.
 487      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 488      * jnr indices corresponding to data put in the four positions in the SIMD register.
 489      */
 490     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 491     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 492     int              jnrA,jnrB,jnrC,jnrD;
 493     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 494     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 495     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 496     real             rcutoff_scalar;
 497     real             *shiftvec,*fshift,*x,*f;
 498     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 499     real             scratch[4*DIM];
 500     __m128           fscal,rcutoff,rcutoff2,jidxall;
 501     int              vdwioffset0;
 502     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 503     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 504     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 505     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 506     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 507     real             *charge;
 508     int              nvdwtype;
 509     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 510     int              *vdwtype;
 511     real             *vdwparam;
 512     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 513     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 514     __m128i          vfitab;
 515     __m128i          ifour       = _mm_set1_epi32(4);
 516     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 517     real             *vftab;
 518     __m128i          ewitab;
 519     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 520     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 521     real             *ewtab;
 522     __m128           dummy_mask,cutoff_mask;
 523     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 524     __m128           one     = _mm_set1_ps(1.0);
 525     __m128           two     = _mm_set1_ps(2.0);
 526     x                = xx[0];
 527     f                = ff[0];
 528
 529     nri              = nlist->nri;
 530     iinr             = nlist->iinr;
 531     jindex           = nlist->jindex;
 532     jjnr             = nlist->jjnr;
 533     shiftidx         = nlist->shift;
 534     gid              = nlist->gid;
 535     shiftvec         = fr->shift_vec[0];
 536     fshift           = fr->fshift[0];
 537     facel            = _mm_set1_ps(fr->epsfac);
 538     charge           = mdatoms->chargeA;
 539     nvdwtype         = fr->ntype;
 540     vdwparam         = fr->nbfp;
 541     vdwtype          = mdatoms->typeA;
 542
 543     vftab            = kernel_data->table_vdw->data;
 544     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 545
 546     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 547     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 548     beta2            = _mm_mul_ps(beta,beta);
 549     beta3            = _mm_mul_ps(beta,beta2);
 550     ewtab            = fr->ic->tabq_coul_F;
 551     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 552     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 553
 554     /* Avoid stupid compiler warnings */
 555     jnrA = jnrB = jnrC = jnrD = 0;
 556     j_coord_offsetA = 0;
 557     j_coord_offsetB = 0;
 558     j_coord_offsetC = 0;
 559     j_coord_offsetD = 0;
 560
 561     outeriter        = 0;
 562     inneriter        = 0;
 563
 564     for(iidx=0;iidx<4*DIM;iidx++)
 565     {
 566         scratch[iidx] = 0.0;
 567     }
 568
 569     /* Start outer loop over neighborlists */
 570     for(iidx=0; iidx<nri; iidx++)
 571     {
 572         /* Load shift vector for this list */
 573         i_shift_offset   = DIM*shiftidx[iidx];
 574
 575         /* Load limits for loop over neighbors */
 576         j_index_start    = jindex[iidx];
 577         j_index_end      = jindex[iidx+1];
 578
 579         /* Get outer coordinate index */
 580         inr              = iinr[iidx];
 581         i_coord_offset   = DIM*inr;
 582
 583         /* Load i particle coords and add shift vector */
 584         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 585
 586         fix0             = _mm_setzero_ps();
 587         fiy0             = _mm_setzero_ps();
 588         fiz0             = _mm_setzero_ps();
 589
 590         /* Load parameters for i particles */
 591         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 592         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 593
 594         /* Start inner kernel loop */
 595         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 596         {
 597
 598             /* Get j neighbor index, and coordinate index */
 599             jnrA             = jjnr[jidx];
 600             jnrB             = jjnr[jidx+1];
 601             jnrC             = jjnr[jidx+2];
 602             jnrD             = jjnr[jidx+3];
 603             j_coord_offsetA  = DIM*jnrA;
 604             j_coord_offsetB  = DIM*jnrB;
 605             j_coord_offsetC  = DIM*jnrC;
 606             j_coord_offsetD  = DIM*jnrD;
 607
 608             /* load j atom coordinates */
 609             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 610                                               x+j_coord_offsetC,x+j_coord_offsetD,
 611                                               &jx0,&jy0,&jz0);
 612
 613             /* Calculate displacement vector */
 614             dx00             = _mm_sub_ps(ix0,jx0);
 615             dy00             = _mm_sub_ps(iy0,jy0);
 616             dz00             = _mm_sub_ps(iz0,jz0);
 617
 618             /* Calculate squared distance and things based on it */
 619             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 620
 621             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 622
 623             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 624
 625             /* Load parameters for j particles */
 626             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 627                                                               charge+jnrC+0,charge+jnrD+0);
 628             vdwjidx0A        = 2*vdwtype[jnrA+0];
 629             vdwjidx0B        = 2*vdwtype[jnrB+0];
 630             vdwjidx0C        = 2*vdwtype[jnrC+0];
 631             vdwjidx0D        = 2*vdwtype[jnrD+0];
 632
 633             /**************************
 634              * CALCULATE INTERACTIONS *
 635              **************************/
 636
 637             r00              = _mm_mul_ps(rsq00,rinv00);
 638
 639             /* Compute parameters for interactions between i and j atoms */
 640             qq00             = _mm_mul_ps(iq0,jq0);
 641             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 642                                          vdwparam+vdwioffset0+vdwjidx0B,
 643                                          vdwparam+vdwioffset0+vdwjidx0C,
 644                                          vdwparam+vdwioffset0+vdwjidx0D,
 645                                          &c6_00,&c12_00);
 646
 647             /* Calculate table index by multiplying r with table scale and truncate to integer */
 648             rt               = _mm_mul_ps(r00,vftabscale);
 649             vfitab           = _mm_cvttps_epi32(rt);
 650 #ifdef __XOP__
 651             vfeps            = _mm_frcz_ps(rt);
 652 #else
 653             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 654 #endif
 655             twovfeps         = _mm_add_ps(vfeps,vfeps);
 656             vfitab           = _mm_slli_epi32(vfitab,3);
 657
 658             /* EWALD ELECTROSTATICS */
 659
 660             /* Analytical PME correction */
 661             zeta2            = _mm_mul_ps(beta2,rsq00);
 662             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 663             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 664             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 665             felec            = _mm_mul_ps(qq00,felec);
 666
 667             /* CUBIC SPLINE TABLE DISPERSION */
 668             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 669             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 670             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 671             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 672             _MM_TRANSPOSE4_PS(Y,F,G,H);
 673             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 674             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 675             fvdw6            = _mm_mul_ps(c6_00,FF);
 676
 677             /* CUBIC SPLINE TABLE REPULSION */
 678             vfitab           = _mm_add_epi32(vfitab,ifour);
 679             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 680             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 681             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 682             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 683             _MM_TRANSPOSE4_PS(Y,F,G,H);
 684             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 685             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 686             fvdw12           = _mm_mul_ps(c12_00,FF);
 687             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 688
 689             fscal            = _mm_add_ps(felec,fvdw);
 690
 691              /* Update vectorial force */
 692             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 693             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 694             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 695
 696             fjptrA             = f+j_coord_offsetA;
 697             fjptrB             = f+j_coord_offsetB;
 698             fjptrC             = f+j_coord_offsetC;
 699             fjptrD             = f+j_coord_offsetD;
 700             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 701                                                    _mm_mul_ps(dx00,fscal),
 702                                                    _mm_mul_ps(dy00,fscal),
 703                                                    _mm_mul_ps(dz00,fscal));
 704
 705             /* Inner loop uses 54 flops */
 706         }
 707
 708         if(jidx<j_index_end)
 709         {
 710
 711             /* Get j neighbor index, and coordinate index */
 712             jnrlistA         = jjnr[jidx];
 713             jnrlistB         = jjnr[jidx+1];
 714             jnrlistC         = jjnr[jidx+2];
 715             jnrlistD         = jjnr[jidx+3];
 716             /* Sign of each element will be negative for non-real atoms.
 717              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 718              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 719              */
 720             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 721             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 722             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 723             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 724             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 725             j_coord_offsetA  = DIM*jnrA;
 726             j_coord_offsetB  = DIM*jnrB;
 727             j_coord_offsetC  = DIM*jnrC;
 728             j_coord_offsetD  = DIM*jnrD;
 729
 730             /* load j atom coordinates */
 731             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 732                                               x+j_coord_offsetC,x+j_coord_offsetD,
 733                                               &jx0,&jy0,&jz0);
 734
 735             /* Calculate displacement vector */
 736             dx00             = _mm_sub_ps(ix0,jx0);
 737             dy00             = _mm_sub_ps(iy0,jy0);
 738             dz00             = _mm_sub_ps(iz0,jz0);
 739
 740             /* Calculate squared distance and things based on it */
 741             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 742
 743             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 744
 745             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 746
 747             /* Load parameters for j particles */
 748             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 749                                                               charge+jnrC+0,charge+jnrD+0);
 750             vdwjidx0A        = 2*vdwtype[jnrA+0];
 751             vdwjidx0B        = 2*vdwtype[jnrB+0];
 752             vdwjidx0C        = 2*vdwtype[jnrC+0];
 753             vdwjidx0D        = 2*vdwtype[jnrD+0];
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             r00              = _mm_mul_ps(rsq00,rinv00);
 760             r00              = _mm_andnot_ps(dummy_mask,r00);
 761
 762             /* Compute parameters for interactions between i and j atoms */
 763             qq00             = _mm_mul_ps(iq0,jq0);
 764             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 765                                          vdwparam+vdwioffset0+vdwjidx0B,
 766                                          vdwparam+vdwioffset0+vdwjidx0C,
 767                                          vdwparam+vdwioffset0+vdwjidx0D,
 768                                          &c6_00,&c12_00);
 769
 770             /* Calculate table index by multiplying r with table scale and truncate to integer */
 771             rt               = _mm_mul_ps(r00,vftabscale);
 772             vfitab           = _mm_cvttps_epi32(rt);
 773 #ifdef __XOP__
 774             vfeps            = _mm_frcz_ps(rt);
 775 #else
 776             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 777 #endif
 778             twovfeps         = _mm_add_ps(vfeps,vfeps);
 779             vfitab           = _mm_slli_epi32(vfitab,3);
 780
 781             /* EWALD ELECTROSTATICS */
 782
 783             /* Analytical PME correction */
 784             zeta2            = _mm_mul_ps(beta2,rsq00);
 785             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 786             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 787             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 788             felec            = _mm_mul_ps(qq00,felec);
 789
 790             /* CUBIC SPLINE TABLE DISPERSION */
 791             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 792             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 793             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 794             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 795             _MM_TRANSPOSE4_PS(Y,F,G,H);
 796             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 797             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 798             fvdw6            = _mm_mul_ps(c6_00,FF);
 799
 800             /* CUBIC SPLINE TABLE REPULSION */
 801             vfitab           = _mm_add_epi32(vfitab,ifour);
 802             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 803             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 804             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 805             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 806             _MM_TRANSPOSE4_PS(Y,F,G,H);
 807             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 808             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 809             fvdw12           = _mm_mul_ps(c12_00,FF);
 810             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 811
 812             fscal            = _mm_add_ps(felec,fvdw);
 813
 814             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 815
 816              /* Update vectorial force */
 817             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 818             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 819             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 820
 821             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 822             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 823             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 824             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 825             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 826                                                    _mm_mul_ps(dx00,fscal),
 827                                                    _mm_mul_ps(dy00,fscal),
 828                                                    _mm_mul_ps(dz00,fscal));
 829
 830             /* Inner loop uses 55 flops */
 831         }
 832
 833         /* End of innermost loop */
 834
 835         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 836                                               f+i_coord_offset,fshift+i_shift_offset);
 837
 838         /* Increment number of inner iterations */
 839         inneriter                  += j_index_end - j_index_start;
 840
 841         /* Outer loop uses 7 flops */
 842     }
 843
 844     /* Increment number of outer iterations */
 845     outeriter        += nri;
 846
 847     /* Update outer/inner flops */
 848
 849     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 850 }