src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCSTab_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_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: CubicSplineTable
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128i          vfitab;
  99     __m128i          ifour       = _mm_set1_epi32(4);
 100     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 101     real             *vftab;
 102     __m128           dummy_mask,cutoff_mask;
 103     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 104     __m128           one     = _mm_set1_ps(1.0);
 105     __m128           two     = _mm_set1_ps(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_ps(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_elec_vdw->data;
 124     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = jnrC = jnrD = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130     j_coord_offsetC = 0;
 131     j_coord_offsetD = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     for(iidx=0;iidx<4*DIM;iidx++)
 137     {
 138         scratch[iidx] = 0.0;
 139     }
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 157
 158         fix0             = _mm_setzero_ps();
 159         fiy0             = _mm_setzero_ps();
 160         fiz0             = _mm_setzero_ps();
 161
 162         /* Load parameters for i particles */
 163         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 164         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 165
 166         /* Reset potential sums */
 167         velecsum         = _mm_setzero_ps();
 168         vvdwsum          = _mm_setzero_ps();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             jnrC             = jjnr[jidx+2];
 178             jnrD             = jjnr[jidx+3];
 179             j_coord_offsetA  = DIM*jnrA;
 180             j_coord_offsetB  = DIM*jnrB;
 181             j_coord_offsetC  = DIM*jnrC;
 182             j_coord_offsetD  = DIM*jnrD;
 183
 184             /* load j atom coordinates */
 185             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 186                                               x+j_coord_offsetC,x+j_coord_offsetD,
 187                                               &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _mm_sub_ps(ix0,jx0);
 191             dy00             = _mm_sub_ps(iy0,jy0);
 192             dz00             = _mm_sub_ps(iz0,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 196
 197             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 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(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 231
 232             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 233             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 234             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 235             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 236             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 237             _MM_TRANSPOSE4_PS(Y,F,G,H);
 238             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 239             VV               = _mm_macc_ps(vfeps,Fp,Y);
 240             velec            = _mm_mul_ps(qq00,VV);
 241             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 242             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 243
 244             /* CUBIC SPLINE TABLE DISPERSION */
 245             vfitab           = _mm_add_epi32(vfitab,ifour);
 246             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 247             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 248             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 249             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 250             _MM_TRANSPOSE4_PS(Y,F,G,H);
 251             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 252             VV               = _mm_macc_ps(vfeps,Fp,Y);
 253             vvdw6            = _mm_mul_ps(c6_00,VV);
 254             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 255             fvdw6            = _mm_mul_ps(c6_00,FF);
 256
 257             /* CUBIC SPLINE TABLE REPULSION */
 258             vfitab           = _mm_add_epi32(vfitab,ifour);
 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             vvdw12           = _mm_mul_ps(c12_00,VV);
 267             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 268             fvdw12           = _mm_mul_ps(c12_00,FF);
 269             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 270             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 271
 272             /* Update potential sum for this i atom from the interaction with this j atom. */
 273             velecsum         = _mm_add_ps(velecsum,velec);
 274             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 275
 276             fscal            = _mm_add_ps(felec,fvdw);
 277
 278              /* Update vectorial force */
 279             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 280             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 281             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 282
 283             fjptrA             = f+j_coord_offsetA;
 284             fjptrB             = f+j_coord_offsetB;
 285             fjptrC             = f+j_coord_offsetC;
 286             fjptrD             = f+j_coord_offsetD;
 287             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 288                                                    _mm_mul_ps(dx00,fscal),
 289                                                    _mm_mul_ps(dy00,fscal),
 290                                                    _mm_mul_ps(dz00,fscal));
 291
 292             /* Inner loop uses 76 flops */
 293         }
 294
 295         if(jidx<j_index_end)
 296         {
 297
 298             /* Get j neighbor index, and coordinate index */
 299             jnrlistA         = jjnr[jidx];
 300             jnrlistB         = jjnr[jidx+1];
 301             jnrlistC         = jjnr[jidx+2];
 302             jnrlistD         = jjnr[jidx+3];
 303             /* Sign of each element will be negative for non-real atoms.
 304              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 305              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 306              */
 307             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 308             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 309             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 310             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 311             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 312             j_coord_offsetA  = DIM*jnrA;
 313             j_coord_offsetB  = DIM*jnrB;
 314             j_coord_offsetC  = DIM*jnrC;
 315             j_coord_offsetD  = DIM*jnrD;
 316
 317             /* load j atom coordinates */
 318             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 319                                               x+j_coord_offsetC,x+j_coord_offsetD,
 320                                               &jx0,&jy0,&jz0);
 321
 322             /* Calculate displacement vector */
 323             dx00             = _mm_sub_ps(ix0,jx0);
 324             dy00             = _mm_sub_ps(iy0,jy0);
 325             dz00             = _mm_sub_ps(iz0,jz0);
 326
 327             /* Calculate squared distance and things based on it */
 328             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 329
 330             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 331
 332             /* Load parameters for j particles */
 333             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 334                                                               charge+jnrC+0,charge+jnrD+0);
 335             vdwjidx0A        = 2*vdwtype[jnrA+0];
 336             vdwjidx0B        = 2*vdwtype[jnrB+0];
 337             vdwjidx0C        = 2*vdwtype[jnrC+0];
 338             vdwjidx0D        = 2*vdwtype[jnrD+0];
 339
 340             /**************************
 341              * CALCULATE INTERACTIONS *
 342              **************************/
 343
 344             r00              = _mm_mul_ps(rsq00,rinv00);
 345             r00              = _mm_andnot_ps(dummy_mask,r00);
 346
 347             /* Compute parameters for interactions between i and j atoms */
 348             qq00             = _mm_mul_ps(iq0,jq0);
 349             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 350                                          vdwparam+vdwioffset0+vdwjidx0B,
 351                                          vdwparam+vdwioffset0+vdwjidx0C,
 352                                          vdwparam+vdwioffset0+vdwjidx0D,
 353                                          &c6_00,&c12_00);
 354
 355             /* Calculate table index by multiplying r with table scale and truncate to integer */
 356             rt               = _mm_mul_ps(r00,vftabscale);
 357             vfitab           = _mm_cvttps_epi32(rt);
 358 #ifdef __XOP__
 359             vfeps            = _mm_frcz_ps(rt);
 360 #else
 361             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 362 #endif
 363             twovfeps         = _mm_add_ps(vfeps,vfeps);
 364             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 365
 366             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 367             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 368             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 369             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 370             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 371             _MM_TRANSPOSE4_PS(Y,F,G,H);
 372             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 373             VV               = _mm_macc_ps(vfeps,Fp,Y);
 374             velec            = _mm_mul_ps(qq00,VV);
 375             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 376             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 377
 378             /* CUBIC SPLINE TABLE DISPERSION */
 379             vfitab           = _mm_add_epi32(vfitab,ifour);
 380             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 381             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 382             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 383             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 384             _MM_TRANSPOSE4_PS(Y,F,G,H);
 385             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 386             VV               = _mm_macc_ps(vfeps,Fp,Y);
 387             vvdw6            = _mm_mul_ps(c6_00,VV);
 388             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 389             fvdw6            = _mm_mul_ps(c6_00,FF);
 390
 391             /* CUBIC SPLINE TABLE REPULSION */
 392             vfitab           = _mm_add_epi32(vfitab,ifour);
 393             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 394             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 395             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 396             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 397             _MM_TRANSPOSE4_PS(Y,F,G,H);
 398             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 399             VV               = _mm_macc_ps(vfeps,Fp,Y);
 400             vvdw12           = _mm_mul_ps(c12_00,VV);
 401             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 402             fvdw12           = _mm_mul_ps(c12_00,FF);
 403             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 404             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velec            = _mm_andnot_ps(dummy_mask,velec);
 408             velecsum         = _mm_add_ps(velecsum,velec);
 409             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 410             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 411
 412             fscal            = _mm_add_ps(felec,fvdw);
 413
 414             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 415
 416              /* Update vectorial force */
 417             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 418             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 419             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 420
 421             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 422             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 423             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 424             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 425             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 426                                                    _mm_mul_ps(dx00,fscal),
 427                                                    _mm_mul_ps(dy00,fscal),
 428                                                    _mm_mul_ps(dz00,fscal));
 429
 430             /* Inner loop uses 77 flops */
 431         }
 432
 433         /* End of innermost loop */
 434
 435         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 436                                               f+i_coord_offset,fshift+i_shift_offset);
 437
 438         ggid                        = gid[iidx];
 439         /* Update potential energies */
 440         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 441         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 442
 443         /* Increment number of inner iterations */
 444         inneriter                  += j_index_end - j_index_start;
 445
 446         /* Outer loop uses 9 flops */
 447     }
 448
 449     /* Increment number of outer iterations */
 450     outeriter        += nri;
 451
 452     /* Update outer/inner flops */
 453
 454     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*77);
 455 }
 456 /*
 457  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 458  * Electrostatics interaction: CubicSplineTable
 459  * VdW interaction:            CubicSplineTable
 460  * Geometry:                   Particle-Particle
 461  * Calculate force/pot:        Force
 462  */
 463 void
 464 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 465                     (t_nblist                    * gmx_restrict       nlist,
 466                      rvec                        * gmx_restrict          xx,
 467                      rvec                        * gmx_restrict          ff,
 468                      t_forcerec                  * gmx_restrict          fr,
 469                      t_mdatoms                   * gmx_restrict     mdatoms,
 470                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 471                      t_nrnb                      * gmx_restrict        nrnb)
 472 {
 473     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 474      * just 0 for non-waters.
 475      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 476      * jnr indices corresponding to data put in the four positions in the SIMD register.
 477      */
 478     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 479     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 480     int              jnrA,jnrB,jnrC,jnrD;
 481     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 482     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 483     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 484     real             rcutoff_scalar;
 485     real             *shiftvec,*fshift,*x,*f;
 486     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 487     real             scratch[4*DIM];
 488     __m128           fscal,rcutoff,rcutoff2,jidxall;
 489     int              vdwioffset0;
 490     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 491     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 492     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 493     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 494     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 495     real             *charge;
 496     int              nvdwtype;
 497     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 498     int              *vdwtype;
 499     real             *vdwparam;
 500     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 501     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 502     __m128i          vfitab;
 503     __m128i          ifour       = _mm_set1_epi32(4);
 504     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 505     real             *vftab;
 506     __m128           dummy_mask,cutoff_mask;
 507     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 508     __m128           one     = _mm_set1_ps(1.0);
 509     __m128           two     = _mm_set1_ps(2.0);
 510     x                = xx[0];
 511     f                = ff[0];
 512
 513     nri              = nlist->nri;
 514     iinr             = nlist->iinr;
 515     jindex           = nlist->jindex;
 516     jjnr             = nlist->jjnr;
 517     shiftidx         = nlist->shift;
 518     gid              = nlist->gid;
 519     shiftvec         = fr->shift_vec[0];
 520     fshift           = fr->fshift[0];
 521     facel            = _mm_set1_ps(fr->epsfac);
 522     charge           = mdatoms->chargeA;
 523     nvdwtype         = fr->ntype;
 524     vdwparam         = fr->nbfp;
 525     vdwtype          = mdatoms->typeA;
 526
 527     vftab            = kernel_data->table_elec_vdw->data;
 528     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 529
 530     /* Avoid stupid compiler warnings */
 531     jnrA = jnrB = jnrC = jnrD = 0;
 532     j_coord_offsetA = 0;
 533     j_coord_offsetB = 0;
 534     j_coord_offsetC = 0;
 535     j_coord_offsetD = 0;
 536
 537     outeriter        = 0;
 538     inneriter        = 0;
 539
 540     for(iidx=0;iidx<4*DIM;iidx++)
 541     {
 542         scratch[iidx] = 0.0;
 543     }
 544
 545     /* Start outer loop over neighborlists */
 546     for(iidx=0; iidx<nri; iidx++)
 547     {
 548         /* Load shift vector for this list */
 549         i_shift_offset   = DIM*shiftidx[iidx];
 550
 551         /* Load limits for loop over neighbors */
 552         j_index_start    = jindex[iidx];
 553         j_index_end      = jindex[iidx+1];
 554
 555         /* Get outer coordinate index */
 556         inr              = iinr[iidx];
 557         i_coord_offset   = DIM*inr;
 558
 559         /* Load i particle coords and add shift vector */
 560         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 561
 562         fix0             = _mm_setzero_ps();
 563         fiy0             = _mm_setzero_ps();
 564         fiz0             = _mm_setzero_ps();
 565
 566         /* Load parameters for i particles */
 567         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 568         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 569
 570         /* Start inner kernel loop */
 571         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 572         {
 573
 574             /* Get j neighbor index, and coordinate index */
 575             jnrA             = jjnr[jidx];
 576             jnrB             = jjnr[jidx+1];
 577             jnrC             = jjnr[jidx+2];
 578             jnrD             = jjnr[jidx+3];
 579             j_coord_offsetA  = DIM*jnrA;
 580             j_coord_offsetB  = DIM*jnrB;
 581             j_coord_offsetC  = DIM*jnrC;
 582             j_coord_offsetD  = DIM*jnrD;
 583
 584             /* load j atom coordinates */
 585             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 586                                               x+j_coord_offsetC,x+j_coord_offsetD,
 587                                               &jx0,&jy0,&jz0);
 588
 589             /* Calculate displacement vector */
 590             dx00             = _mm_sub_ps(ix0,jx0);
 591             dy00             = _mm_sub_ps(iy0,jy0);
 592             dz00             = _mm_sub_ps(iz0,jz0);
 593
 594             /* Calculate squared distance and things based on it */
 595             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 596
 597             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 598
 599             /* Load parameters for j particles */
 600             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 601                                                               charge+jnrC+0,charge+jnrD+0);
 602             vdwjidx0A        = 2*vdwtype[jnrA+0];
 603             vdwjidx0B        = 2*vdwtype[jnrB+0];
 604             vdwjidx0C        = 2*vdwtype[jnrC+0];
 605             vdwjidx0D        = 2*vdwtype[jnrD+0];
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             r00              = _mm_mul_ps(rsq00,rinv00);
 612
 613             /* Compute parameters for interactions between i and j atoms */
 614             qq00             = _mm_mul_ps(iq0,jq0);
 615             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 616                                          vdwparam+vdwioffset0+vdwjidx0B,
 617                                          vdwparam+vdwioffset0+vdwjidx0C,
 618                                          vdwparam+vdwioffset0+vdwjidx0D,
 619                                          &c6_00,&c12_00);
 620
 621             /* Calculate table index by multiplying r with table scale and truncate to integer */
 622             rt               = _mm_mul_ps(r00,vftabscale);
 623             vfitab           = _mm_cvttps_epi32(rt);
 624 #ifdef __XOP__
 625             vfeps            = _mm_frcz_ps(rt);
 626 #else
 627             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 628 #endif
 629             twovfeps         = _mm_add_ps(vfeps,vfeps);
 630             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 631
 632             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 633             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 634             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 635             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 636             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 637             _MM_TRANSPOSE4_PS(Y,F,G,H);
 638             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 639             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 640             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 641
 642             /* CUBIC SPLINE TABLE DISPERSION */
 643             vfitab           = _mm_add_epi32(vfitab,ifour);
 644             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 645             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 646             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 647             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 648             _MM_TRANSPOSE4_PS(Y,F,G,H);
 649             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 650             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 651             fvdw6            = _mm_mul_ps(c6_00,FF);
 652
 653             /* CUBIC SPLINE TABLE REPULSION */
 654             vfitab           = _mm_add_epi32(vfitab,ifour);
 655             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 656             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 657             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 658             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 659             _MM_TRANSPOSE4_PS(Y,F,G,H);
 660             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 661             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 662             fvdw12           = _mm_mul_ps(c12_00,FF);
 663             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 664
 665             fscal            = _mm_add_ps(felec,fvdw);
 666
 667              /* Update vectorial force */
 668             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 669             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 670             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 671
 672             fjptrA             = f+j_coord_offsetA;
 673             fjptrB             = f+j_coord_offsetB;
 674             fjptrC             = f+j_coord_offsetC;
 675             fjptrD             = f+j_coord_offsetD;
 676             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 677                                                    _mm_mul_ps(dx00,fscal),
 678                                                    _mm_mul_ps(dy00,fscal),
 679                                                    _mm_mul_ps(dz00,fscal));
 680
 681             /* Inner loop uses 64 flops */
 682         }
 683
 684         if(jidx<j_index_end)
 685         {
 686
 687             /* Get j neighbor index, and coordinate index */
 688             jnrlistA         = jjnr[jidx];
 689             jnrlistB         = jjnr[jidx+1];
 690             jnrlistC         = jjnr[jidx+2];
 691             jnrlistD         = jjnr[jidx+3];
 692             /* Sign of each element will be negative for non-real atoms.
 693              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 694              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 695              */
 696             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 697             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 698             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 699             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 700             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 701             j_coord_offsetA  = DIM*jnrA;
 702             j_coord_offsetB  = DIM*jnrB;
 703             j_coord_offsetC  = DIM*jnrC;
 704             j_coord_offsetD  = DIM*jnrD;
 705
 706             /* load j atom coordinates */
 707             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 708                                               x+j_coord_offsetC,x+j_coord_offsetD,
 709                                               &jx0,&jy0,&jz0);
 710
 711             /* Calculate displacement vector */
 712             dx00             = _mm_sub_ps(ix0,jx0);
 713             dy00             = _mm_sub_ps(iy0,jy0);
 714             dz00             = _mm_sub_ps(iz0,jz0);
 715
 716             /* Calculate squared distance and things based on it */
 717             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 718
 719             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 720
 721             /* Load parameters for j particles */
 722             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 723                                                               charge+jnrC+0,charge+jnrD+0);
 724             vdwjidx0A        = 2*vdwtype[jnrA+0];
 725             vdwjidx0B        = 2*vdwtype[jnrB+0];
 726             vdwjidx0C        = 2*vdwtype[jnrC+0];
 727             vdwjidx0D        = 2*vdwtype[jnrD+0];
 728
 729             /**************************
 730              * CALCULATE INTERACTIONS *
 731              **************************/
 732
 733             r00              = _mm_mul_ps(rsq00,rinv00);
 734             r00              = _mm_andnot_ps(dummy_mask,r00);
 735
 736             /* Compute parameters for interactions between i and j atoms */
 737             qq00             = _mm_mul_ps(iq0,jq0);
 738             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 739                                          vdwparam+vdwioffset0+vdwjidx0B,
 740                                          vdwparam+vdwioffset0+vdwjidx0C,
 741                                          vdwparam+vdwioffset0+vdwjidx0D,
 742                                          &c6_00,&c12_00);
 743
 744             /* Calculate table index by multiplying r with table scale and truncate to integer */
 745             rt               = _mm_mul_ps(r00,vftabscale);
 746             vfitab           = _mm_cvttps_epi32(rt);
 747 #ifdef __XOP__
 748             vfeps            = _mm_frcz_ps(rt);
 749 #else
 750             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 751 #endif
 752             twovfeps         = _mm_add_ps(vfeps,vfeps);
 753             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 754
 755             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 756             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 757             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 758             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 759             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 760             _MM_TRANSPOSE4_PS(Y,F,G,H);
 761             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 762             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 763             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 764
 765             /* CUBIC SPLINE TABLE DISPERSION */
 766             vfitab           = _mm_add_epi32(vfitab,ifour);
 767             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 768             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 769             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 770             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 771             _MM_TRANSPOSE4_PS(Y,F,G,H);
 772             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 773             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 774             fvdw6            = _mm_mul_ps(c6_00,FF);
 775
 776             /* CUBIC SPLINE TABLE REPULSION */
 777             vfitab           = _mm_add_epi32(vfitab,ifour);
 778             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 779             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 780             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 781             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 782             _MM_TRANSPOSE4_PS(Y,F,G,H);
 783             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 784             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 785             fvdw12           = _mm_mul_ps(c12_00,FF);
 786             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 787
 788             fscal            = _mm_add_ps(felec,fvdw);
 789
 790             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 791
 792              /* Update vectorial force */
 793             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 794             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 795             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 796
 797             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 798             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 799             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 800             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 801             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 802                                                    _mm_mul_ps(dx00,fscal),
 803                                                    _mm_mul_ps(dy00,fscal),
 804                                                    _mm_mul_ps(dz00,fscal));
 805
 806             /* Inner loop uses 65 flops */
 807         }
 808
 809         /* End of innermost loop */
 810
 811         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 812                                               f+i_coord_offset,fshift+i_shift_offset);
 813
 814         /* Increment number of inner iterations */
 815         inneriter                  += j_index_end - j_index_start;
 816
 817         /* Outer loop uses 7 flops */
 818     }
 819
 820     /* Increment number of outer iterations */
 821     outeriter        += nri;
 822
 823     /* Update outer/inner flops */
 824
 825     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*65);
 826 }