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 "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_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: CubicSplineTable
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  95     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  96     __m128i          vfitab;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  99     real             *vftab;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     vftab            = kernel_data->table_elec_vdw->data;
 122     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 123
 124     /* Avoid stupid compiler warnings */
 125     jnrA = jnrB = jnrC = jnrD = 0;
 126     j_coord_offsetA = 0;
 127     j_coord_offsetB = 0;
 128     j_coord_offsetC = 0;
 129     j_coord_offsetD = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm_setzero_ps();
 157         fiy0             = _mm_setzero_ps();
 158         fiz0             = _mm_setzero_ps();
 159
 160         /* Load parameters for i particles */
 161         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 162         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_ps();
 166         vvdwsum          = _mm_setzero_ps();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             jnrC             = jjnr[jidx+2];
 176             jnrD             = jjnr[jidx+3];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179             j_coord_offsetC  = DIM*jnrC;
 180             j_coord_offsetD  = DIM*jnrD;
 181
 182             /* load j atom coordinates */
 183             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 184                                               x+j_coord_offsetC,x+j_coord_offsetD,
 185                                               &jx0,&jy0,&jz0);
 186
 187             /* Calculate displacement vector */
 188             dx00             = _mm_sub_ps(ix0,jx0);
 189             dy00             = _mm_sub_ps(iy0,jy0);
 190             dz00             = _mm_sub_ps(iz0,jz0);
 191
 192             /* Calculate squared distance and things based on it */
 193             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 194
 195             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 196
 197             /* Load parameters for j particles */
 198             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 199                                                               charge+jnrC+0,charge+jnrD+0);
 200             vdwjidx0A        = 2*vdwtype[jnrA+0];
 201             vdwjidx0B        = 2*vdwtype[jnrB+0];
 202             vdwjidx0C        = 2*vdwtype[jnrC+0];
 203             vdwjidx0D        = 2*vdwtype[jnrD+0];
 204
 205             /**************************
 206              * CALCULATE INTERACTIONS *
 207              **************************/
 208
 209             r00              = _mm_mul_ps(rsq00,rinv00);
 210
 211             /* Compute parameters for interactions between i and j atoms */
 212             qq00             = _mm_mul_ps(iq0,jq0);
 213             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 214                                          vdwparam+vdwioffset0+vdwjidx0B,
 215                                          vdwparam+vdwioffset0+vdwjidx0C,
 216                                          vdwparam+vdwioffset0+vdwjidx0D,
 217                                          &c6_00,&c12_00);
 218
 219             /* Calculate table index by multiplying r with table scale and truncate to integer */
 220             rt               = _mm_mul_ps(r00,vftabscale);
 221             vfitab           = _mm_cvttps_epi32(rt);
 222 #ifdef __XOP__
 223             vfeps            = _mm_frcz_ps(rt);
 224 #else
 225             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 226 #endif
 227             twovfeps         = _mm_add_ps(vfeps,vfeps);
 228             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 229
 230             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 231             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 232             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 233             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 234             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 235             _MM_TRANSPOSE4_PS(Y,F,G,H);
 236             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 237             VV               = _mm_macc_ps(vfeps,Fp,Y);
 238             velec            = _mm_mul_ps(qq00,VV);
 239             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 240             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 241
 242             /* CUBIC SPLINE TABLE DISPERSION */
 243             vfitab           = _mm_add_epi32(vfitab,ifour);
 244             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 245             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 246             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 247             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 248             _MM_TRANSPOSE4_PS(Y,F,G,H);
 249             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 250             VV               = _mm_macc_ps(vfeps,Fp,Y);
 251             vvdw6            = _mm_mul_ps(c6_00,VV);
 252             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 253             fvdw6            = _mm_mul_ps(c6_00,FF);
 254
 255             /* CUBIC SPLINE TABLE REPULSION */
 256             vfitab           = _mm_add_epi32(vfitab,ifour);
 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             vvdw12           = _mm_mul_ps(c12_00,VV);
 265             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 266             fvdw12           = _mm_mul_ps(c12_00,FF);
 267             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 268             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 269
 270             /* Update potential sum for this i atom from the interaction with this j atom. */
 271             velecsum         = _mm_add_ps(velecsum,velec);
 272             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 273
 274             fscal            = _mm_add_ps(felec,fvdw);
 275
 276              /* Update vectorial force */
 277             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 278             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 279             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 280
 281             fjptrA             = f+j_coord_offsetA;
 282             fjptrB             = f+j_coord_offsetB;
 283             fjptrC             = f+j_coord_offsetC;
 284             fjptrD             = f+j_coord_offsetD;
 285             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 286                                                    _mm_mul_ps(dx00,fscal),
 287                                                    _mm_mul_ps(dy00,fscal),
 288                                                    _mm_mul_ps(dz00,fscal));
 289
 290             /* Inner loop uses 76 flops */
 291         }
 292
 293         if(jidx<j_index_end)
 294         {
 295
 296             /* Get j neighbor index, and coordinate index */
 297             jnrlistA         = jjnr[jidx];
 298             jnrlistB         = jjnr[jidx+1];
 299             jnrlistC         = jjnr[jidx+2];
 300             jnrlistD         = jjnr[jidx+3];
 301             /* Sign of each element will be negative for non-real atoms.
 302              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 303              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 304              */
 305             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 306             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 307             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 308             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 309             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 310             j_coord_offsetA  = DIM*jnrA;
 311             j_coord_offsetB  = DIM*jnrB;
 312             j_coord_offsetC  = DIM*jnrC;
 313             j_coord_offsetD  = DIM*jnrD;
 314
 315             /* load j atom coordinates */
 316             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 317                                               x+j_coord_offsetC,x+j_coord_offsetD,
 318                                               &jx0,&jy0,&jz0);
 319
 320             /* Calculate displacement vector */
 321             dx00             = _mm_sub_ps(ix0,jx0);
 322             dy00             = _mm_sub_ps(iy0,jy0);
 323             dz00             = _mm_sub_ps(iz0,jz0);
 324
 325             /* Calculate squared distance and things based on it */
 326             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 327
 328             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 329
 330             /* Load parameters for j particles */
 331             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 332                                                               charge+jnrC+0,charge+jnrD+0);
 333             vdwjidx0A        = 2*vdwtype[jnrA+0];
 334             vdwjidx0B        = 2*vdwtype[jnrB+0];
 335             vdwjidx0C        = 2*vdwtype[jnrC+0];
 336             vdwjidx0D        = 2*vdwtype[jnrD+0];
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             r00              = _mm_mul_ps(rsq00,rinv00);
 343             r00              = _mm_andnot_ps(dummy_mask,r00);
 344
 345             /* Compute parameters for interactions between i and j atoms */
 346             qq00             = _mm_mul_ps(iq0,jq0);
 347             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 348                                          vdwparam+vdwioffset0+vdwjidx0B,
 349                                          vdwparam+vdwioffset0+vdwjidx0C,
 350                                          vdwparam+vdwioffset0+vdwjidx0D,
 351                                          &c6_00,&c12_00);
 352
 353             /* Calculate table index by multiplying r with table scale and truncate to integer */
 354             rt               = _mm_mul_ps(r00,vftabscale);
 355             vfitab           = _mm_cvttps_epi32(rt);
 356 #ifdef __XOP__
 357             vfeps            = _mm_frcz_ps(rt);
 358 #else
 359             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 360 #endif
 361             twovfeps         = _mm_add_ps(vfeps,vfeps);
 362             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 363
 364             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 365             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 366             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 367             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 368             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 369             _MM_TRANSPOSE4_PS(Y,F,G,H);
 370             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 371             VV               = _mm_macc_ps(vfeps,Fp,Y);
 372             velec            = _mm_mul_ps(qq00,VV);
 373             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 374             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 375
 376             /* CUBIC SPLINE TABLE DISPERSION */
 377             vfitab           = _mm_add_epi32(vfitab,ifour);
 378             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 379             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 380             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 381             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 382             _MM_TRANSPOSE4_PS(Y,F,G,H);
 383             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 384             VV               = _mm_macc_ps(vfeps,Fp,Y);
 385             vvdw6            = _mm_mul_ps(c6_00,VV);
 386             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 387             fvdw6            = _mm_mul_ps(c6_00,FF);
 388
 389             /* CUBIC SPLINE TABLE REPULSION */
 390             vfitab           = _mm_add_epi32(vfitab,ifour);
 391             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 392             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 393             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 394             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 395             _MM_TRANSPOSE4_PS(Y,F,G,H);
 396             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 397             VV               = _mm_macc_ps(vfeps,Fp,Y);
 398             vvdw12           = _mm_mul_ps(c12_00,VV);
 399             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 400             fvdw12           = _mm_mul_ps(c12_00,FF);
 401             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 402             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velec            = _mm_andnot_ps(dummy_mask,velec);
 406             velecsum         = _mm_add_ps(velecsum,velec);
 407             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 408             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 409
 410             fscal            = _mm_add_ps(felec,fvdw);
 411
 412             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 413
 414              /* Update vectorial force */
 415             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 416             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 417             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 418
 419             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 420             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 421             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 422             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 423             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 424                                                    _mm_mul_ps(dx00,fscal),
 425                                                    _mm_mul_ps(dy00,fscal),
 426                                                    _mm_mul_ps(dz00,fscal));
 427
 428             /* Inner loop uses 77 flops */
 429         }
 430
 431         /* End of innermost loop */
 432
 433         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 434                                               f+i_coord_offset,fshift+i_shift_offset);
 435
 436         ggid                        = gid[iidx];
 437         /* Update potential energies */
 438         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 439         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 440
 441         /* Increment number of inner iterations */
 442         inneriter                  += j_index_end - j_index_start;
 443
 444         /* Outer loop uses 9 flops */
 445     }
 446
 447     /* Increment number of outer iterations */
 448     outeriter        += nri;
 449
 450     /* Update outer/inner flops */
 451
 452     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*77);
 453 }
 454 /*
 455  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 456  * Electrostatics interaction: CubicSplineTable
 457  * VdW interaction:            CubicSplineTable
 458  * Geometry:                   Particle-Particle
 459  * Calculate force/pot:        Force
 460  */
 461 void
 462 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 463                     (t_nblist                    * gmx_restrict       nlist,
 464                      rvec                        * gmx_restrict          xx,
 465                      rvec                        * gmx_restrict          ff,
 466                      t_forcerec                  * gmx_restrict          fr,
 467                      t_mdatoms                   * gmx_restrict     mdatoms,
 468                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 469                      t_nrnb                      * gmx_restrict        nrnb)
 470 {
 471     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 472      * just 0 for non-waters.
 473      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 474      * jnr indices corresponding to data put in the four positions in the SIMD register.
 475      */
 476     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 477     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 478     int              jnrA,jnrB,jnrC,jnrD;
 479     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 480     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 481     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 482     real             rcutoff_scalar;
 483     real             *shiftvec,*fshift,*x,*f;
 484     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 485     real             scratch[4*DIM];
 486     __m128           fscal,rcutoff,rcutoff2,jidxall;
 487     int              vdwioffset0;
 488     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 489     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 490     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 491     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 492     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 493     real             *charge;
 494     int              nvdwtype;
 495     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 496     int              *vdwtype;
 497     real             *vdwparam;
 498     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 499     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 500     __m128i          vfitab;
 501     __m128i          ifour       = _mm_set1_epi32(4);
 502     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 503     real             *vftab;
 504     __m128           dummy_mask,cutoff_mask;
 505     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 506     __m128           one     = _mm_set1_ps(1.0);
 507     __m128           two     = _mm_set1_ps(2.0);
 508     x                = xx[0];
 509     f                = ff[0];
 510
 511     nri              = nlist->nri;
 512     iinr             = nlist->iinr;
 513     jindex           = nlist->jindex;
 514     jjnr             = nlist->jjnr;
 515     shiftidx         = nlist->shift;
 516     gid              = nlist->gid;
 517     shiftvec         = fr->shift_vec[0];
 518     fshift           = fr->fshift[0];
 519     facel            = _mm_set1_ps(fr->epsfac);
 520     charge           = mdatoms->chargeA;
 521     nvdwtype         = fr->ntype;
 522     vdwparam         = fr->nbfp;
 523     vdwtype          = mdatoms->typeA;
 524
 525     vftab            = kernel_data->table_elec_vdw->data;
 526     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 527
 528     /* Avoid stupid compiler warnings */
 529     jnrA = jnrB = jnrC = jnrD = 0;
 530     j_coord_offsetA = 0;
 531     j_coord_offsetB = 0;
 532     j_coord_offsetC = 0;
 533     j_coord_offsetD = 0;
 534
 535     outeriter        = 0;
 536     inneriter        = 0;
 537
 538     for(iidx=0;iidx<4*DIM;iidx++)
 539     {
 540         scratch[iidx] = 0.0;
 541     }
 542
 543     /* Start outer loop over neighborlists */
 544     for(iidx=0; iidx<nri; iidx++)
 545     {
 546         /* Load shift vector for this list */
 547         i_shift_offset   = DIM*shiftidx[iidx];
 548
 549         /* Load limits for loop over neighbors */
 550         j_index_start    = jindex[iidx];
 551         j_index_end      = jindex[iidx+1];
 552
 553         /* Get outer coordinate index */
 554         inr              = iinr[iidx];
 555         i_coord_offset   = DIM*inr;
 556
 557         /* Load i particle coords and add shift vector */
 558         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 559
 560         fix0             = _mm_setzero_ps();
 561         fiy0             = _mm_setzero_ps();
 562         fiz0             = _mm_setzero_ps();
 563
 564         /* Load parameters for i particles */
 565         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 566         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 567
 568         /* Start inner kernel loop */
 569         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 570         {
 571
 572             /* Get j neighbor index, and coordinate index */
 573             jnrA             = jjnr[jidx];
 574             jnrB             = jjnr[jidx+1];
 575             jnrC             = jjnr[jidx+2];
 576             jnrD             = jjnr[jidx+3];
 577             j_coord_offsetA  = DIM*jnrA;
 578             j_coord_offsetB  = DIM*jnrB;
 579             j_coord_offsetC  = DIM*jnrC;
 580             j_coord_offsetD  = DIM*jnrD;
 581
 582             /* load j atom coordinates */
 583             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 584                                               x+j_coord_offsetC,x+j_coord_offsetD,
 585                                               &jx0,&jy0,&jz0);
 586
 587             /* Calculate displacement vector */
 588             dx00             = _mm_sub_ps(ix0,jx0);
 589             dy00             = _mm_sub_ps(iy0,jy0);
 590             dz00             = _mm_sub_ps(iz0,jz0);
 591
 592             /* Calculate squared distance and things based on it */
 593             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 594
 595             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 596
 597             /* Load parameters for j particles */
 598             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 599                                                               charge+jnrC+0,charge+jnrD+0);
 600             vdwjidx0A        = 2*vdwtype[jnrA+0];
 601             vdwjidx0B        = 2*vdwtype[jnrB+0];
 602             vdwjidx0C        = 2*vdwtype[jnrC+0];
 603             vdwjidx0D        = 2*vdwtype[jnrD+0];
 604
 605             /**************************
 606              * CALCULATE INTERACTIONS *
 607              **************************/
 608
 609             r00              = _mm_mul_ps(rsq00,rinv00);
 610
 611             /* Compute parameters for interactions between i and j atoms */
 612             qq00             = _mm_mul_ps(iq0,jq0);
 613             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 614                                          vdwparam+vdwioffset0+vdwjidx0B,
 615                                          vdwparam+vdwioffset0+vdwjidx0C,
 616                                          vdwparam+vdwioffset0+vdwjidx0D,
 617                                          &c6_00,&c12_00);
 618
 619             /* Calculate table index by multiplying r with table scale and truncate to integer */
 620             rt               = _mm_mul_ps(r00,vftabscale);
 621             vfitab           = _mm_cvttps_epi32(rt);
 622 #ifdef __XOP__
 623             vfeps            = _mm_frcz_ps(rt);
 624 #else
 625             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 626 #endif
 627             twovfeps         = _mm_add_ps(vfeps,vfeps);
 628             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 629
 630             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 631             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 632             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 633             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 634             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 635             _MM_TRANSPOSE4_PS(Y,F,G,H);
 636             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 637             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 638             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 639
 640             /* CUBIC SPLINE TABLE DISPERSION */
 641             vfitab           = _mm_add_epi32(vfitab,ifour);
 642             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 643             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 644             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 645             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 646             _MM_TRANSPOSE4_PS(Y,F,G,H);
 647             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 648             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 649             fvdw6            = _mm_mul_ps(c6_00,FF);
 650
 651             /* CUBIC SPLINE TABLE REPULSION */
 652             vfitab           = _mm_add_epi32(vfitab,ifour);
 653             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 654             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 655             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 656             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 657             _MM_TRANSPOSE4_PS(Y,F,G,H);
 658             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 659             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 660             fvdw12           = _mm_mul_ps(c12_00,FF);
 661             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 662
 663             fscal            = _mm_add_ps(felec,fvdw);
 664
 665              /* Update vectorial force */
 666             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 667             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 668             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 669
 670             fjptrA             = f+j_coord_offsetA;
 671             fjptrB             = f+j_coord_offsetB;
 672             fjptrC             = f+j_coord_offsetC;
 673             fjptrD             = f+j_coord_offsetD;
 674             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 675                                                    _mm_mul_ps(dx00,fscal),
 676                                                    _mm_mul_ps(dy00,fscal),
 677                                                    _mm_mul_ps(dz00,fscal));
 678
 679             /* Inner loop uses 64 flops */
 680         }
 681
 682         if(jidx<j_index_end)
 683         {
 684
 685             /* Get j neighbor index, and coordinate index */
 686             jnrlistA         = jjnr[jidx];
 687             jnrlistB         = jjnr[jidx+1];
 688             jnrlistC         = jjnr[jidx+2];
 689             jnrlistD         = jjnr[jidx+3];
 690             /* Sign of each element will be negative for non-real atoms.
 691              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 692              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 693              */
 694             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 695             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 696             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 697             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 698             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 699             j_coord_offsetA  = DIM*jnrA;
 700             j_coord_offsetB  = DIM*jnrB;
 701             j_coord_offsetC  = DIM*jnrC;
 702             j_coord_offsetD  = DIM*jnrD;
 703
 704             /* load j atom coordinates */
 705             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 706                                               x+j_coord_offsetC,x+j_coord_offsetD,
 707                                               &jx0,&jy0,&jz0);
 708
 709             /* Calculate displacement vector */
 710             dx00             = _mm_sub_ps(ix0,jx0);
 711             dy00             = _mm_sub_ps(iy0,jy0);
 712             dz00             = _mm_sub_ps(iz0,jz0);
 713
 714             /* Calculate squared distance and things based on it */
 715             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 716
 717             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 718
 719             /* Load parameters for j particles */
 720             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 721                                                               charge+jnrC+0,charge+jnrD+0);
 722             vdwjidx0A        = 2*vdwtype[jnrA+0];
 723             vdwjidx0B        = 2*vdwtype[jnrB+0];
 724             vdwjidx0C        = 2*vdwtype[jnrC+0];
 725             vdwjidx0D        = 2*vdwtype[jnrD+0];
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             r00              = _mm_mul_ps(rsq00,rinv00);
 732             r00              = _mm_andnot_ps(dummy_mask,r00);
 733
 734             /* Compute parameters for interactions between i and j atoms */
 735             qq00             = _mm_mul_ps(iq0,jq0);
 736             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 737                                          vdwparam+vdwioffset0+vdwjidx0B,
 738                                          vdwparam+vdwioffset0+vdwjidx0C,
 739                                          vdwparam+vdwioffset0+vdwjidx0D,
 740                                          &c6_00,&c12_00);
 741
 742             /* Calculate table index by multiplying r with table scale and truncate to integer */
 743             rt               = _mm_mul_ps(r00,vftabscale);
 744             vfitab           = _mm_cvttps_epi32(rt);
 745 #ifdef __XOP__
 746             vfeps            = _mm_frcz_ps(rt);
 747 #else
 748             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 749 #endif
 750             twovfeps         = _mm_add_ps(vfeps,vfeps);
 751             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 752
 753             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 754             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 755             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 756             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 757             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 758             _MM_TRANSPOSE4_PS(Y,F,G,H);
 759             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 760             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 761             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 762
 763             /* CUBIC SPLINE TABLE DISPERSION */
 764             vfitab           = _mm_add_epi32(vfitab,ifour);
 765             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 766             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 767             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 768             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 769             _MM_TRANSPOSE4_PS(Y,F,G,H);
 770             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 771             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 772             fvdw6            = _mm_mul_ps(c6_00,FF);
 773
 774             /* CUBIC SPLINE TABLE REPULSION */
 775             vfitab           = _mm_add_epi32(vfitab,ifour);
 776             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 777             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 778             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 779             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 780             _MM_TRANSPOSE4_PS(Y,F,G,H);
 781             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 782             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 783             fvdw12           = _mm_mul_ps(c12_00,FF);
 784             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 785
 786             fscal            = _mm_add_ps(felec,fvdw);
 787
 788             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 789
 790              /* Update vectorial force */
 791             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 792             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 793             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 794
 795             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 796             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 797             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 798             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 799             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 800                                                    _mm_mul_ps(dx00,fscal),
 801                                                    _mm_mul_ps(dy00,fscal),
 802                                                    _mm_mul_ps(dz00,fscal));
 803
 804             /* Inner loop uses 65 flops */
 805         }
 806
 807         /* End of innermost loop */
 808
 809         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 810                                               f+i_coord_offset,fshift+i_shift_offset);
 811
 812         /* Increment number of inner iterations */
 813         inneriter                  += j_index_end - j_index_start;
 814
 815         /* Outer loop uses 7 flops */
 816     }
 817
 818     /* Increment number of outer iterations */
 819     outeriter        += nri;
 820
 821     /* Update outer/inner flops */
 822
 823     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*65);
 824 }