src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRF_VdwCSTab_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     int              nvdwtype;
  77     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  78     int              *vdwtype;
  79     real             *vdwparam;
  80     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  81     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  82     __m128i          vfitab;
  83     __m128i          ifour       = _mm_set1_epi32(4);
  84     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  85     real             *vftab;
  86     __m128           dummy_mask,cutoff_mask;
  87     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  88     __m128           one     = _mm_set1_ps(1.0);
  89     __m128           two     = _mm_set1_ps(2.0);
  90     x                = xx[0];
  91     f                = ff[0];
  92
  93     nri              = nlist->nri;
  94     iinr             = nlist->iinr;
  95     jindex           = nlist->jindex;
  96     jjnr             = nlist->jjnr;
  97     shiftidx         = nlist->shift;
  98     gid              = nlist->gid;
  99     shiftvec         = fr->shift_vec[0];
 100     fshift           = fr->fshift[0];
 101     facel            = _mm_set1_ps(fr->epsfac);
 102     charge           = mdatoms->chargeA;
 103     krf              = _mm_set1_ps(fr->ic->k_rf);
 104     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 105     crf              = _mm_set1_ps(fr->ic->c_rf);
 106     nvdwtype         = fr->ntype;
 107     vdwparam         = fr->nbfp;
 108     vdwtype          = mdatoms->typeA;
 109
 110     vftab            = kernel_data->table_vdw->data;
 111     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 112
 113     /* Avoid stupid compiler warnings */
 114     jnrA = jnrB = jnrC = jnrD = 0;
 115     j_coord_offsetA = 0;
 116     j_coord_offsetB = 0;
 117     j_coord_offsetC = 0;
 118     j_coord_offsetD = 0;
 119
 120     outeriter        = 0;
 121     inneriter        = 0;
 122
 123     for(iidx=0;iidx<4*DIM;iidx++)
 124     {
 125         scratch[iidx] = 0.0;
 126     }
 127
 128     /* Start outer loop over neighborlists */
 129     for(iidx=0; iidx<nri; iidx++)
 130     {
 131         /* Load shift vector for this list */
 132         i_shift_offset   = DIM*shiftidx[iidx];
 133
 134         /* Load limits for loop over neighbors */
 135         j_index_start    = jindex[iidx];
 136         j_index_end      = jindex[iidx+1];
 137
 138         /* Get outer coordinate index */
 139         inr              = iinr[iidx];
 140         i_coord_offset   = DIM*inr;
 141
 142         /* Load i particle coords and add shift vector */
 143         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 144
 145         fix0             = _mm_setzero_ps();
 146         fiy0             = _mm_setzero_ps();
 147         fiz0             = _mm_setzero_ps();
 148
 149         /* Load parameters for i particles */
 150         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 151         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 152
 153         /* Reset potential sums */
 154         velecsum         = _mm_setzero_ps();
 155         vvdwsum          = _mm_setzero_ps();
 156
 157         /* Start inner kernel loop */
 158         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 159         {
 160
 161             /* Get j neighbor index, and coordinate index */
 162             jnrA             = jjnr[jidx];
 163             jnrB             = jjnr[jidx+1];
 164             jnrC             = jjnr[jidx+2];
 165             jnrD             = jjnr[jidx+3];
 166             j_coord_offsetA  = DIM*jnrA;
 167             j_coord_offsetB  = DIM*jnrB;
 168             j_coord_offsetC  = DIM*jnrC;
 169             j_coord_offsetD  = DIM*jnrD;
 170
 171             /* load j atom coordinates */
 172             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 173                                               x+j_coord_offsetC,x+j_coord_offsetD,
 174                                               &jx0,&jy0,&jz0);
 175
 176             /* Calculate displacement vector */
 177             dx00             = _mm_sub_ps(ix0,jx0);
 178             dy00             = _mm_sub_ps(iy0,jy0);
 179             dz00             = _mm_sub_ps(iz0,jz0);
 180
 181             /* Calculate squared distance and things based on it */
 182             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 183
 184             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 185
 186             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 187
 188             /* Load parameters for j particles */
 189             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 190                                                               charge+jnrC+0,charge+jnrD+0);
 191             vdwjidx0A        = 2*vdwtype[jnrA+0];
 192             vdwjidx0B        = 2*vdwtype[jnrB+0];
 193             vdwjidx0C        = 2*vdwtype[jnrC+0];
 194             vdwjidx0D        = 2*vdwtype[jnrD+0];
 195
 196             /**************************
 197              * CALCULATE INTERACTIONS *
 198              **************************/
 199
 200             r00              = _mm_mul_ps(rsq00,rinv00);
 201
 202             /* Compute parameters for interactions between i and j atoms */
 203             qq00             = _mm_mul_ps(iq0,jq0);
 204             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 205                                          vdwparam+vdwioffset0+vdwjidx0B,
 206                                          vdwparam+vdwioffset0+vdwjidx0C,
 207                                          vdwparam+vdwioffset0+vdwjidx0D,
 208                                          &c6_00,&c12_00);
 209
 210             /* Calculate table index by multiplying r with table scale and truncate to integer */
 211             rt               = _mm_mul_ps(r00,vftabscale);
 212             vfitab           = _mm_cvttps_epi32(rt);
 213 #ifdef __XOP__
 214             vfeps            = _mm_frcz_ps(rt);
 215 #else
 216             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 217 #endif
 218             twovfeps         = _mm_add_ps(vfeps,vfeps);
 219             vfitab           = _mm_slli_epi32(vfitab,3);
 220
 221             /* REACTION-FIELD ELECTROSTATICS */
 222             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 223             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 224
 225             /* CUBIC SPLINE TABLE DISPERSION */
 226             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 227             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 228             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 229             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 230             _MM_TRANSPOSE4_PS(Y,F,G,H);
 231             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 232             VV               = _mm_macc_ps(vfeps,Fp,Y);
 233             vvdw6            = _mm_mul_ps(c6_00,VV);
 234             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 235             fvdw6            = _mm_mul_ps(c6_00,FF);
 236
 237             /* CUBIC SPLINE TABLE REPULSION */
 238             vfitab           = _mm_add_epi32(vfitab,ifour);
 239             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 240             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 241             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 242             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 243             _MM_TRANSPOSE4_PS(Y,F,G,H);
 244             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 245             VV               = _mm_macc_ps(vfeps,Fp,Y);
 246             vvdw12           = _mm_mul_ps(c12_00,VV);
 247             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 248             fvdw12           = _mm_mul_ps(c12_00,FF);
 249             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 250             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 251
 252             /* Update potential sum for this i atom from the interaction with this j atom. */
 253             velecsum         = _mm_add_ps(velecsum,velec);
 254             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 255
 256             fscal            = _mm_add_ps(felec,fvdw);
 257
 258              /* Update vectorial force */
 259             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 260             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 261             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 262
 263             fjptrA             = f+j_coord_offsetA;
 264             fjptrB             = f+j_coord_offsetB;
 265             fjptrC             = f+j_coord_offsetC;
 266             fjptrD             = f+j_coord_offsetD;
 267             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 268                                                    _mm_mul_ps(dx00,fscal),
 269                                                    _mm_mul_ps(dy00,fscal),
 270                                                    _mm_mul_ps(dz00,fscal));
 271
 272             /* Inner loop uses 70 flops */
 273         }
 274
 275         if(jidx<j_index_end)
 276         {
 277
 278             /* Get j neighbor index, and coordinate index */
 279             jnrlistA         = jjnr[jidx];
 280             jnrlistB         = jjnr[jidx+1];
 281             jnrlistC         = jjnr[jidx+2];
 282             jnrlistD         = jjnr[jidx+3];
 283             /* Sign of each element will be negative for non-real atoms.
 284              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 285              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 286              */
 287             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 288             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 289             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 290             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 291             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 292             j_coord_offsetA  = DIM*jnrA;
 293             j_coord_offsetB  = DIM*jnrB;
 294             j_coord_offsetC  = DIM*jnrC;
 295             j_coord_offsetD  = DIM*jnrD;
 296
 297             /* load j atom coordinates */
 298             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 299                                               x+j_coord_offsetC,x+j_coord_offsetD,
 300                                               &jx0,&jy0,&jz0);
 301
 302             /* Calculate displacement vector */
 303             dx00             = _mm_sub_ps(ix0,jx0);
 304             dy00             = _mm_sub_ps(iy0,jy0);
 305             dz00             = _mm_sub_ps(iz0,jz0);
 306
 307             /* Calculate squared distance and things based on it */
 308             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 309
 310             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 311
 312             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 313
 314             /* Load parameters for j particles */
 315             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 316                                                               charge+jnrC+0,charge+jnrD+0);
 317             vdwjidx0A        = 2*vdwtype[jnrA+0];
 318             vdwjidx0B        = 2*vdwtype[jnrB+0];
 319             vdwjidx0C        = 2*vdwtype[jnrC+0];
 320             vdwjidx0D        = 2*vdwtype[jnrD+0];
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             r00              = _mm_mul_ps(rsq00,rinv00);
 327             r00              = _mm_andnot_ps(dummy_mask,r00);
 328
 329             /* Compute parameters for interactions between i and j atoms */
 330             qq00             = _mm_mul_ps(iq0,jq0);
 331             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 332                                          vdwparam+vdwioffset0+vdwjidx0B,
 333                                          vdwparam+vdwioffset0+vdwjidx0C,
 334                                          vdwparam+vdwioffset0+vdwjidx0D,
 335                                          &c6_00,&c12_00);
 336
 337             /* Calculate table index by multiplying r with table scale and truncate to integer */
 338             rt               = _mm_mul_ps(r00,vftabscale);
 339             vfitab           = _mm_cvttps_epi32(rt);
 340 #ifdef __XOP__
 341             vfeps            = _mm_frcz_ps(rt);
 342 #else
 343             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 344 #endif
 345             twovfeps         = _mm_add_ps(vfeps,vfeps);
 346             vfitab           = _mm_slli_epi32(vfitab,3);
 347
 348             /* REACTION-FIELD ELECTROSTATICS */
 349             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 350             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 351
 352             /* CUBIC SPLINE TABLE DISPERSION */
 353             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 354             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 355             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 356             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 357             _MM_TRANSPOSE4_PS(Y,F,G,H);
 358             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 359             VV               = _mm_macc_ps(vfeps,Fp,Y);
 360             vvdw6            = _mm_mul_ps(c6_00,VV);
 361             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 362             fvdw6            = _mm_mul_ps(c6_00,FF);
 363
 364             /* CUBIC SPLINE TABLE REPULSION */
 365             vfitab           = _mm_add_epi32(vfitab,ifour);
 366             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 367             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 368             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 369             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 370             _MM_TRANSPOSE4_PS(Y,F,G,H);
 371             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 372             VV               = _mm_macc_ps(vfeps,Fp,Y);
 373             vvdw12           = _mm_mul_ps(c12_00,VV);
 374             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 375             fvdw12           = _mm_mul_ps(c12_00,FF);
 376             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 377             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm_andnot_ps(dummy_mask,velec);
 381             velecsum         = _mm_add_ps(velecsum,velec);
 382             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 383             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 384
 385             fscal            = _mm_add_ps(felec,fvdw);
 386
 387             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 388
 389              /* Update vectorial force */
 390             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 391             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 392             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 393
 394             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 395             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 396             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 397             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 398             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 399                                                    _mm_mul_ps(dx00,fscal),
 400                                                    _mm_mul_ps(dy00,fscal),
 401                                                    _mm_mul_ps(dz00,fscal));
 402
 403             /* Inner loop uses 71 flops */
 404         }
 405
 406         /* End of innermost loop */
 407
 408         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 409                                               f+i_coord_offset,fshift+i_shift_offset);
 410
 411         ggid                        = gid[iidx];
 412         /* Update potential energies */
 413         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 414         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 415
 416         /* Increment number of inner iterations */
 417         inneriter                  += j_index_end - j_index_start;
 418
 419         /* Outer loop uses 9 flops */
 420     }
 421
 422     /* Increment number of outer iterations */
 423     outeriter        += nri;
 424
 425     /* Update outer/inner flops */
 426
 427     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
 428 }
 429 /*
 430  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 431  * Electrostatics interaction: ReactionField
 432  * VdW interaction:            CubicSplineTable
 433  * Geometry:                   Particle-Particle
 434  * Calculate force/pot:        Force
 435  */
 436 void
 437 nb_kernel_ElecRF_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 438                     (t_nblist * gmx_restrict                nlist,
 439                      rvec * gmx_restrict                    xx,
 440                      rvec * gmx_restrict                    ff,
 441                      t_forcerec * gmx_restrict              fr,
 442                      t_mdatoms * gmx_restrict               mdatoms,
 443                      nb_kernel_data_t * gmx_restrict        kernel_data,
 444                      t_nrnb * gmx_restrict                  nrnb)
 445 {
 446     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 447      * just 0 for non-waters.
 448      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 449      * jnr indices corresponding to data put in the four positions in the SIMD register.
 450      */
 451     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 452     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 453     int              jnrA,jnrB,jnrC,jnrD;
 454     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 455     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 456     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 457     real             rcutoff_scalar;
 458     real             *shiftvec,*fshift,*x,*f;
 459     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 460     real             scratch[4*DIM];
 461     __m128           fscal,rcutoff,rcutoff2,jidxall;
 462     int              vdwioffset0;
 463     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 464     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 465     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 466     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 467     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 468     real             *charge;
 469     int              nvdwtype;
 470     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 471     int              *vdwtype;
 472     real             *vdwparam;
 473     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 474     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 475     __m128i          vfitab;
 476     __m128i          ifour       = _mm_set1_epi32(4);
 477     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 478     real             *vftab;
 479     __m128           dummy_mask,cutoff_mask;
 480     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 481     __m128           one     = _mm_set1_ps(1.0);
 482     __m128           two     = _mm_set1_ps(2.0);
 483     x                = xx[0];
 484     f                = ff[0];
 485
 486     nri              = nlist->nri;
 487     iinr             = nlist->iinr;
 488     jindex           = nlist->jindex;
 489     jjnr             = nlist->jjnr;
 490     shiftidx         = nlist->shift;
 491     gid              = nlist->gid;
 492     shiftvec         = fr->shift_vec[0];
 493     fshift           = fr->fshift[0];
 494     facel            = _mm_set1_ps(fr->epsfac);
 495     charge           = mdatoms->chargeA;
 496     krf              = _mm_set1_ps(fr->ic->k_rf);
 497     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 498     crf              = _mm_set1_ps(fr->ic->c_rf);
 499     nvdwtype         = fr->ntype;
 500     vdwparam         = fr->nbfp;
 501     vdwtype          = mdatoms->typeA;
 502
 503     vftab            = kernel_data->table_vdw->data;
 504     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 505
 506     /* Avoid stupid compiler warnings */
 507     jnrA = jnrB = jnrC = jnrD = 0;
 508     j_coord_offsetA = 0;
 509     j_coord_offsetB = 0;
 510     j_coord_offsetC = 0;
 511     j_coord_offsetD = 0;
 512
 513     outeriter        = 0;
 514     inneriter        = 0;
 515
 516     for(iidx=0;iidx<4*DIM;iidx++)
 517     {
 518         scratch[iidx] = 0.0;
 519     }
 520
 521     /* Start outer loop over neighborlists */
 522     for(iidx=0; iidx<nri; iidx++)
 523     {
 524         /* Load shift vector for this list */
 525         i_shift_offset   = DIM*shiftidx[iidx];
 526
 527         /* Load limits for loop over neighbors */
 528         j_index_start    = jindex[iidx];
 529         j_index_end      = jindex[iidx+1];
 530
 531         /* Get outer coordinate index */
 532         inr              = iinr[iidx];
 533         i_coord_offset   = DIM*inr;
 534
 535         /* Load i particle coords and add shift vector */
 536         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 537
 538         fix0             = _mm_setzero_ps();
 539         fiy0             = _mm_setzero_ps();
 540         fiz0             = _mm_setzero_ps();
 541
 542         /* Load parameters for i particles */
 543         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 544         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 545
 546         /* Start inner kernel loop */
 547         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 548         {
 549
 550             /* Get j neighbor index, and coordinate index */
 551             jnrA             = jjnr[jidx];
 552             jnrB             = jjnr[jidx+1];
 553             jnrC             = jjnr[jidx+2];
 554             jnrD             = jjnr[jidx+3];
 555             j_coord_offsetA  = DIM*jnrA;
 556             j_coord_offsetB  = DIM*jnrB;
 557             j_coord_offsetC  = DIM*jnrC;
 558             j_coord_offsetD  = DIM*jnrD;
 559
 560             /* load j atom coordinates */
 561             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 562                                               x+j_coord_offsetC,x+j_coord_offsetD,
 563                                               &jx0,&jy0,&jz0);
 564
 565             /* Calculate displacement vector */
 566             dx00             = _mm_sub_ps(ix0,jx0);
 567             dy00             = _mm_sub_ps(iy0,jy0);
 568             dz00             = _mm_sub_ps(iz0,jz0);
 569
 570             /* Calculate squared distance and things based on it */
 571             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 572
 573             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 574
 575             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 576
 577             /* Load parameters for j particles */
 578             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 579                                                               charge+jnrC+0,charge+jnrD+0);
 580             vdwjidx0A        = 2*vdwtype[jnrA+0];
 581             vdwjidx0B        = 2*vdwtype[jnrB+0];
 582             vdwjidx0C        = 2*vdwtype[jnrC+0];
 583             vdwjidx0D        = 2*vdwtype[jnrD+0];
 584
 585             /**************************
 586              * CALCULATE INTERACTIONS *
 587              **************************/
 588
 589             r00              = _mm_mul_ps(rsq00,rinv00);
 590
 591             /* Compute parameters for interactions between i and j atoms */
 592             qq00             = _mm_mul_ps(iq0,jq0);
 593             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 594                                          vdwparam+vdwioffset0+vdwjidx0B,
 595                                          vdwparam+vdwioffset0+vdwjidx0C,
 596                                          vdwparam+vdwioffset0+vdwjidx0D,
 597                                          &c6_00,&c12_00);
 598
 599             /* Calculate table index by multiplying r with table scale and truncate to integer */
 600             rt               = _mm_mul_ps(r00,vftabscale);
 601             vfitab           = _mm_cvttps_epi32(rt);
 602 #ifdef __XOP__
 603             vfeps            = _mm_frcz_ps(rt);
 604 #else
 605             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 606 #endif
 607             twovfeps         = _mm_add_ps(vfeps,vfeps);
 608             vfitab           = _mm_slli_epi32(vfitab,3);
 609
 610             /* REACTION-FIELD ELECTROSTATICS */
 611             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 612
 613             /* CUBIC SPLINE TABLE DISPERSION */
 614             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 615             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 616             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 617             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 618             _MM_TRANSPOSE4_PS(Y,F,G,H);
 619             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 620             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 621             fvdw6            = _mm_mul_ps(c6_00,FF);
 622
 623             /* CUBIC SPLINE TABLE REPULSION */
 624             vfitab           = _mm_add_epi32(vfitab,ifour);
 625             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 626             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 627             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 628             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 629             _MM_TRANSPOSE4_PS(Y,F,G,H);
 630             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 631             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 632             fvdw12           = _mm_mul_ps(c12_00,FF);
 633             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 634
 635             fscal            = _mm_add_ps(felec,fvdw);
 636
 637              /* Update vectorial force */
 638             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 639             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 640             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 641
 642             fjptrA             = f+j_coord_offsetA;
 643             fjptrB             = f+j_coord_offsetB;
 644             fjptrC             = f+j_coord_offsetC;
 645             fjptrD             = f+j_coord_offsetD;
 646             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 647                                                    _mm_mul_ps(dx00,fscal),
 648                                                    _mm_mul_ps(dy00,fscal),
 649                                                    _mm_mul_ps(dz00,fscal));
 650
 651             /* Inner loop uses 57 flops */
 652         }
 653
 654         if(jidx<j_index_end)
 655         {
 656
 657             /* Get j neighbor index, and coordinate index */
 658             jnrlistA         = jjnr[jidx];
 659             jnrlistB         = jjnr[jidx+1];
 660             jnrlistC         = jjnr[jidx+2];
 661             jnrlistD         = jjnr[jidx+3];
 662             /* Sign of each element will be negative for non-real atoms.
 663              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 664              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 665              */
 666             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 667             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 668             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 669             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 670             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 671             j_coord_offsetA  = DIM*jnrA;
 672             j_coord_offsetB  = DIM*jnrB;
 673             j_coord_offsetC  = DIM*jnrC;
 674             j_coord_offsetD  = DIM*jnrD;
 675
 676             /* load j atom coordinates */
 677             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 678                                               x+j_coord_offsetC,x+j_coord_offsetD,
 679                                               &jx0,&jy0,&jz0);
 680
 681             /* Calculate displacement vector */
 682             dx00             = _mm_sub_ps(ix0,jx0);
 683             dy00             = _mm_sub_ps(iy0,jy0);
 684             dz00             = _mm_sub_ps(iz0,jz0);
 685
 686             /* Calculate squared distance and things based on it */
 687             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 688
 689             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 690
 691             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 692
 693             /* Load parameters for j particles */
 694             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 695                                                               charge+jnrC+0,charge+jnrD+0);
 696             vdwjidx0A        = 2*vdwtype[jnrA+0];
 697             vdwjidx0B        = 2*vdwtype[jnrB+0];
 698             vdwjidx0C        = 2*vdwtype[jnrC+0];
 699             vdwjidx0D        = 2*vdwtype[jnrD+0];
 700
 701             /**************************
 702              * CALCULATE INTERACTIONS *
 703              **************************/
 704
 705             r00              = _mm_mul_ps(rsq00,rinv00);
 706             r00              = _mm_andnot_ps(dummy_mask,r00);
 707
 708             /* Compute parameters for interactions between i and j atoms */
 709             qq00             = _mm_mul_ps(iq0,jq0);
 710             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 711                                          vdwparam+vdwioffset0+vdwjidx0B,
 712                                          vdwparam+vdwioffset0+vdwjidx0C,
 713                                          vdwparam+vdwioffset0+vdwjidx0D,
 714                                          &c6_00,&c12_00);
 715
 716             /* Calculate table index by multiplying r with table scale and truncate to integer */
 717             rt               = _mm_mul_ps(r00,vftabscale);
 718             vfitab           = _mm_cvttps_epi32(rt);
 719 #ifdef __XOP__
 720             vfeps            = _mm_frcz_ps(rt);
 721 #else
 722             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 723 #endif
 724             twovfeps         = _mm_add_ps(vfeps,vfeps);
 725             vfitab           = _mm_slli_epi32(vfitab,3);
 726
 727             /* REACTION-FIELD ELECTROSTATICS */
 728             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 729
 730             /* CUBIC SPLINE TABLE DISPERSION */
 731             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 732             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 733             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 734             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 735             _MM_TRANSPOSE4_PS(Y,F,G,H);
 736             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 737             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 738             fvdw6            = _mm_mul_ps(c6_00,FF);
 739
 740             /* CUBIC SPLINE TABLE REPULSION */
 741             vfitab           = _mm_add_epi32(vfitab,ifour);
 742             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 743             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 744             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 745             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 746             _MM_TRANSPOSE4_PS(Y,F,G,H);
 747             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 748             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 749             fvdw12           = _mm_mul_ps(c12_00,FF);
 750             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 751
 752             fscal            = _mm_add_ps(felec,fvdw);
 753
 754             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 755
 756              /* Update vectorial force */
 757             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 758             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 759             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 760
 761             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 762             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 763             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 764             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 765             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 766                                                    _mm_mul_ps(dx00,fscal),
 767                                                    _mm_mul_ps(dy00,fscal),
 768                                                    _mm_mul_ps(dz00,fscal));
 769
 770             /* Inner loop uses 58 flops */
 771         }
 772
 773         /* End of innermost loop */
 774
 775         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 776                                               f+i_coord_offset,fshift+i_shift_offset);
 777
 778         /* Increment number of inner iterations */
 779         inneriter                  += j_index_end - j_index_start;
 780
 781         /* Outer loop uses 7 flops */
 782     }
 783
 784     /* Increment number of outer iterations */
 785     outeriter        += nri;
 786
 787     /* Update outer/inner flops */
 788
 789     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);
 790 }