src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_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_ElecEw_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_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     __m128i          ewitab;
  87     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  88     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  89     real             *ewtab;
  90     __m128           dummy_mask,cutoff_mask;
  91     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  92     __m128           one     = _mm_set1_ps(1.0);
  93     __m128           two     = _mm_set1_ps(2.0);
  94     x                = xx[0];
  95     f                = ff[0];
  96
  97     nri              = nlist->nri;
  98     iinr             = nlist->iinr;
  99     jindex           = nlist->jindex;
 100     jjnr             = nlist->jjnr;
 101     shiftidx         = nlist->shift;
 102     gid              = nlist->gid;
 103     shiftvec         = fr->shift_vec[0];
 104     fshift           = fr->fshift[0];
 105     facel            = _mm_set1_ps(fr->epsfac);
 106     charge           = mdatoms->chargeA;
 107     nvdwtype         = fr->ntype;
 108     vdwparam         = fr->nbfp;
 109     vdwtype          = mdatoms->typeA;
 110
 111     vftab            = kernel_data->table_vdw->data;
 112     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 113
 114     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 115     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 116     beta2            = _mm_mul_ps(beta,beta);
 117     beta3            = _mm_mul_ps(beta,beta2);
 118     ewtab            = fr->ic->tabq_coul_FDV0;
 119     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 120     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm_setzero_ps();
 155         fiy0             = _mm_setzero_ps();
 156         fiz0             = _mm_setzero_ps();
 157
 158         /* Load parameters for i particles */
 159         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 160         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm_setzero_ps();
 164         vvdwsum          = _mm_setzero_ps();
 165
 166         /* Start inner kernel loop */
 167         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 168         {
 169
 170             /* Get j neighbor index, and coordinate index */
 171             jnrA             = jjnr[jidx];
 172             jnrB             = jjnr[jidx+1];
 173             jnrC             = jjnr[jidx+2];
 174             jnrD             = jjnr[jidx+3];
 175             j_coord_offsetA  = DIM*jnrA;
 176             j_coord_offsetB  = DIM*jnrB;
 177             j_coord_offsetC  = DIM*jnrC;
 178             j_coord_offsetD  = DIM*jnrD;
 179
 180             /* load j atom coordinates */
 181             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 182                                               x+j_coord_offsetC,x+j_coord_offsetD,
 183                                               &jx0,&jy0,&jz0);
 184
 185             /* Calculate displacement vector */
 186             dx00             = _mm_sub_ps(ix0,jx0);
 187             dy00             = _mm_sub_ps(iy0,jy0);
 188             dz00             = _mm_sub_ps(iz0,jz0);
 189
 190             /* Calculate squared distance and things based on it */
 191             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 192
 193             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 194
 195             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 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(vfitab,3);
 229
 230             /* EWALD ELECTROSTATICS */
 231
 232             /* Analytical PME correction */
 233             zeta2            = _mm_mul_ps(beta2,rsq00);
 234             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 235             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 236             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 237             felec            = _mm_mul_ps(qq00,felec);
 238             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 239             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 240             velec            = _mm_mul_ps(qq00,velec);
 241
 242             /* CUBIC SPLINE TABLE DISPERSION */
 243             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 244             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 245             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 246             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 247             _MM_TRANSPOSE4_PS(Y,F,G,H);
 248             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 249             VV               = _mm_macc_ps(vfeps,Fp,Y);
 250             vvdw6            = _mm_mul_ps(c6_00,VV);
 251             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 252             fvdw6            = _mm_mul_ps(c6_00,FF);
 253
 254             /* CUBIC SPLINE TABLE REPULSION */
 255             vfitab           = _mm_add_epi32(vfitab,ifour);
 256             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 257             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 258             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 259             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 260             _MM_TRANSPOSE4_PS(Y,F,G,H);
 261             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 262             VV               = _mm_macc_ps(vfeps,Fp,Y);
 263             vvdw12           = _mm_mul_ps(c12_00,VV);
 264             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 265             fvdw12           = _mm_mul_ps(c12_00,FF);
 266             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 267             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 268
 269             /* Update potential sum for this i atom from the interaction with this j atom. */
 270             velecsum         = _mm_add_ps(velecsum,velec);
 271             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 272
 273             fscal            = _mm_add_ps(felec,fvdw);
 274
 275              /* Update vectorial force */
 276             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 277             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 278             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 279
 280             fjptrA             = f+j_coord_offsetA;
 281             fjptrB             = f+j_coord_offsetB;
 282             fjptrC             = f+j_coord_offsetC;
 283             fjptrD             = f+j_coord_offsetD;
 284             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 285                                                    _mm_mul_ps(dx00,fscal),
 286                                                    _mm_mul_ps(dy00,fscal),
 287                                                    _mm_mul_ps(dz00,fscal));
 288
 289             /* Inner loop uses 63 flops */
 290         }
 291
 292         if(jidx<j_index_end)
 293         {
 294
 295             /* Get j neighbor index, and coordinate index */
 296             jnrlistA         = jjnr[jidx];
 297             jnrlistB         = jjnr[jidx+1];
 298             jnrlistC         = jjnr[jidx+2];
 299             jnrlistD         = jjnr[jidx+3];
 300             /* Sign of each element will be negative for non-real atoms.
 301              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 302              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 303              */
 304             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 305             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 306             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 307             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 308             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 309             j_coord_offsetA  = DIM*jnrA;
 310             j_coord_offsetB  = DIM*jnrB;
 311             j_coord_offsetC  = DIM*jnrC;
 312             j_coord_offsetD  = DIM*jnrD;
 313
 314             /* load j atom coordinates */
 315             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 316                                               x+j_coord_offsetC,x+j_coord_offsetD,
 317                                               &jx0,&jy0,&jz0);
 318
 319             /* Calculate displacement vector */
 320             dx00             = _mm_sub_ps(ix0,jx0);
 321             dy00             = _mm_sub_ps(iy0,jy0);
 322             dz00             = _mm_sub_ps(iz0,jz0);
 323
 324             /* Calculate squared distance and things based on it */
 325             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 326
 327             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 328
 329             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 330
 331             /* Load parameters for j particles */
 332             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 333                                                               charge+jnrC+0,charge+jnrD+0);
 334             vdwjidx0A        = 2*vdwtype[jnrA+0];
 335             vdwjidx0B        = 2*vdwtype[jnrB+0];
 336             vdwjidx0C        = 2*vdwtype[jnrC+0];
 337             vdwjidx0D        = 2*vdwtype[jnrD+0];
 338
 339             /**************************
 340              * CALCULATE INTERACTIONS *
 341              **************************/
 342
 343             r00              = _mm_mul_ps(rsq00,rinv00);
 344             r00              = _mm_andnot_ps(dummy_mask,r00);
 345
 346             /* Compute parameters for interactions between i and j atoms */
 347             qq00             = _mm_mul_ps(iq0,jq0);
 348             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 349                                          vdwparam+vdwioffset0+vdwjidx0B,
 350                                          vdwparam+vdwioffset0+vdwjidx0C,
 351                                          vdwparam+vdwioffset0+vdwjidx0D,
 352                                          &c6_00,&c12_00);
 353
 354             /* Calculate table index by multiplying r with table scale and truncate to integer */
 355             rt               = _mm_mul_ps(r00,vftabscale);
 356             vfitab           = _mm_cvttps_epi32(rt);
 357 #ifdef __XOP__
 358             vfeps            = _mm_frcz_ps(rt);
 359 #else
 360             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 361 #endif
 362             twovfeps         = _mm_add_ps(vfeps,vfeps);
 363             vfitab           = _mm_slli_epi32(vfitab,3);
 364
 365             /* EWALD ELECTROSTATICS */
 366
 367             /* Analytical PME correction */
 368             zeta2            = _mm_mul_ps(beta2,rsq00);
 369             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 370             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 371             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 372             felec            = _mm_mul_ps(qq00,felec);
 373             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 374             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 375             velec            = _mm_mul_ps(qq00,velec);
 376
 377             /* CUBIC SPLINE TABLE DISPERSION */
 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 64 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*64);
 453 }
 454 /*
 455  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 456  * Electrostatics interaction: Ewald
 457  * VdW interaction:            CubicSplineTable
 458  * Geometry:                   Particle-Particle
 459  * Calculate force/pot:        Force
 460  */
 461 void
 462 nb_kernel_ElecEw_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_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     __m128i          ewitab;
 505     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 506     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 507     real             *ewtab;
 508     __m128           dummy_mask,cutoff_mask;
 509     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 510     __m128           one     = _mm_set1_ps(1.0);
 511     __m128           two     = _mm_set1_ps(2.0);
 512     x                = xx[0];
 513     f                = ff[0];
 514
 515     nri              = nlist->nri;
 516     iinr             = nlist->iinr;
 517     jindex           = nlist->jindex;
 518     jjnr             = nlist->jjnr;
 519     shiftidx         = nlist->shift;
 520     gid              = nlist->gid;
 521     shiftvec         = fr->shift_vec[0];
 522     fshift           = fr->fshift[0];
 523     facel            = _mm_set1_ps(fr->epsfac);
 524     charge           = mdatoms->chargeA;
 525     nvdwtype         = fr->ntype;
 526     vdwparam         = fr->nbfp;
 527     vdwtype          = mdatoms->typeA;
 528
 529     vftab            = kernel_data->table_vdw->data;
 530     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 531
 532     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 533     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 534     beta2            = _mm_mul_ps(beta,beta);
 535     beta3            = _mm_mul_ps(beta,beta2);
 536     ewtab            = fr->ic->tabq_coul_F;
 537     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 538     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 539
 540     /* Avoid stupid compiler warnings */
 541     jnrA = jnrB = jnrC = jnrD = 0;
 542     j_coord_offsetA = 0;
 543     j_coord_offsetB = 0;
 544     j_coord_offsetC = 0;
 545     j_coord_offsetD = 0;
 546
 547     outeriter        = 0;
 548     inneriter        = 0;
 549
 550     for(iidx=0;iidx<4*DIM;iidx++)
 551     {
 552         scratch[iidx] = 0.0;
 553     }
 554
 555     /* Start outer loop over neighborlists */
 556     for(iidx=0; iidx<nri; iidx++)
 557     {
 558         /* Load shift vector for this list */
 559         i_shift_offset   = DIM*shiftidx[iidx];
 560
 561         /* Load limits for loop over neighbors */
 562         j_index_start    = jindex[iidx];
 563         j_index_end      = jindex[iidx+1];
 564
 565         /* Get outer coordinate index */
 566         inr              = iinr[iidx];
 567         i_coord_offset   = DIM*inr;
 568
 569         /* Load i particle coords and add shift vector */
 570         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 571
 572         fix0             = _mm_setzero_ps();
 573         fiy0             = _mm_setzero_ps();
 574         fiz0             = _mm_setzero_ps();
 575
 576         /* Load parameters for i particles */
 577         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 578         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 579
 580         /* Start inner kernel loop */
 581         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 582         {
 583
 584             /* Get j neighbor index, and coordinate index */
 585             jnrA             = jjnr[jidx];
 586             jnrB             = jjnr[jidx+1];
 587             jnrC             = jjnr[jidx+2];
 588             jnrD             = jjnr[jidx+3];
 589             j_coord_offsetA  = DIM*jnrA;
 590             j_coord_offsetB  = DIM*jnrB;
 591             j_coord_offsetC  = DIM*jnrC;
 592             j_coord_offsetD  = DIM*jnrD;
 593
 594             /* load j atom coordinates */
 595             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 596                                               x+j_coord_offsetC,x+j_coord_offsetD,
 597                                               &jx0,&jy0,&jz0);
 598
 599             /* Calculate displacement vector */
 600             dx00             = _mm_sub_ps(ix0,jx0);
 601             dy00             = _mm_sub_ps(iy0,jy0);
 602             dz00             = _mm_sub_ps(iz0,jz0);
 603
 604             /* Calculate squared distance and things based on it */
 605             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 606
 607             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 608
 609             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 610
 611             /* Load parameters for j particles */
 612             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 613                                                               charge+jnrC+0,charge+jnrD+0);
 614             vdwjidx0A        = 2*vdwtype[jnrA+0];
 615             vdwjidx0B        = 2*vdwtype[jnrB+0];
 616             vdwjidx0C        = 2*vdwtype[jnrC+0];
 617             vdwjidx0D        = 2*vdwtype[jnrD+0];
 618
 619             /**************************
 620              * CALCULATE INTERACTIONS *
 621              **************************/
 622
 623             r00              = _mm_mul_ps(rsq00,rinv00);
 624
 625             /* Compute parameters for interactions between i and j atoms */
 626             qq00             = _mm_mul_ps(iq0,jq0);
 627             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 628                                          vdwparam+vdwioffset0+vdwjidx0B,
 629                                          vdwparam+vdwioffset0+vdwjidx0C,
 630                                          vdwparam+vdwioffset0+vdwjidx0D,
 631                                          &c6_00,&c12_00);
 632
 633             /* Calculate table index by multiplying r with table scale and truncate to integer */
 634             rt               = _mm_mul_ps(r00,vftabscale);
 635             vfitab           = _mm_cvttps_epi32(rt);
 636 #ifdef __XOP__
 637             vfeps            = _mm_frcz_ps(rt);
 638 #else
 639             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 640 #endif
 641             twovfeps         = _mm_add_ps(vfeps,vfeps);
 642             vfitab           = _mm_slli_epi32(vfitab,3);
 643
 644             /* EWALD ELECTROSTATICS */
 645
 646             /* Analytical PME correction */
 647             zeta2            = _mm_mul_ps(beta2,rsq00);
 648             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 649             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 650             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 651             felec            = _mm_mul_ps(qq00,felec);
 652
 653             /* CUBIC SPLINE TABLE DISPERSION */
 654             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 655             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 656             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 657             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 658             _MM_TRANSPOSE4_PS(Y,F,G,H);
 659             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 660             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 661             fvdw6            = _mm_mul_ps(c6_00,FF);
 662
 663             /* CUBIC SPLINE TABLE REPULSION */
 664             vfitab           = _mm_add_epi32(vfitab,ifour);
 665             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 666             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 667             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 668             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 669             _MM_TRANSPOSE4_PS(Y,F,G,H);
 670             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 671             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 672             fvdw12           = _mm_mul_ps(c12_00,FF);
 673             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 674
 675             fscal            = _mm_add_ps(felec,fvdw);
 676
 677              /* Update vectorial force */
 678             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 679             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 680             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 681
 682             fjptrA             = f+j_coord_offsetA;
 683             fjptrB             = f+j_coord_offsetB;
 684             fjptrC             = f+j_coord_offsetC;
 685             fjptrD             = f+j_coord_offsetD;
 686             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 687                                                    _mm_mul_ps(dx00,fscal),
 688                                                    _mm_mul_ps(dy00,fscal),
 689                                                    _mm_mul_ps(dz00,fscal));
 690
 691             /* Inner loop uses 54 flops */
 692         }
 693
 694         if(jidx<j_index_end)
 695         {
 696
 697             /* Get j neighbor index, and coordinate index */
 698             jnrlistA         = jjnr[jidx];
 699             jnrlistB         = jjnr[jidx+1];
 700             jnrlistC         = jjnr[jidx+2];
 701             jnrlistD         = jjnr[jidx+3];
 702             /* Sign of each element will be negative for non-real atoms.
 703              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 704              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 705              */
 706             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 707             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 708             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 709             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 710             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 711             j_coord_offsetA  = DIM*jnrA;
 712             j_coord_offsetB  = DIM*jnrB;
 713             j_coord_offsetC  = DIM*jnrC;
 714             j_coord_offsetD  = DIM*jnrD;
 715
 716             /* load j atom coordinates */
 717             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 718                                               x+j_coord_offsetC,x+j_coord_offsetD,
 719                                               &jx0,&jy0,&jz0);
 720
 721             /* Calculate displacement vector */
 722             dx00             = _mm_sub_ps(ix0,jx0);
 723             dy00             = _mm_sub_ps(iy0,jy0);
 724             dz00             = _mm_sub_ps(iz0,jz0);
 725
 726             /* Calculate squared distance and things based on it */
 727             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 728
 729             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 730
 731             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 732
 733             /* Load parameters for j particles */
 734             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 735                                                               charge+jnrC+0,charge+jnrD+0);
 736             vdwjidx0A        = 2*vdwtype[jnrA+0];
 737             vdwjidx0B        = 2*vdwtype[jnrB+0];
 738             vdwjidx0C        = 2*vdwtype[jnrC+0];
 739             vdwjidx0D        = 2*vdwtype[jnrD+0];
 740
 741             /**************************
 742              * CALCULATE INTERACTIONS *
 743              **************************/
 744
 745             r00              = _mm_mul_ps(rsq00,rinv00);
 746             r00              = _mm_andnot_ps(dummy_mask,r00);
 747
 748             /* Compute parameters for interactions between i and j atoms */
 749             qq00             = _mm_mul_ps(iq0,jq0);
 750             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 751                                          vdwparam+vdwioffset0+vdwjidx0B,
 752                                          vdwparam+vdwioffset0+vdwjidx0C,
 753                                          vdwparam+vdwioffset0+vdwjidx0D,
 754                                          &c6_00,&c12_00);
 755
 756             /* Calculate table index by multiplying r with table scale and truncate to integer */
 757             rt               = _mm_mul_ps(r00,vftabscale);
 758             vfitab           = _mm_cvttps_epi32(rt);
 759 #ifdef __XOP__
 760             vfeps            = _mm_frcz_ps(rt);
 761 #else
 762             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 763 #endif
 764             twovfeps         = _mm_add_ps(vfeps,vfeps);
 765             vfitab           = _mm_slli_epi32(vfitab,3);
 766
 767             /* EWALD ELECTROSTATICS */
 768
 769             /* Analytical PME correction */
 770             zeta2            = _mm_mul_ps(beta2,rsq00);
 771             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 772             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 773             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 774             felec            = _mm_mul_ps(qq00,felec);
 775
 776             /* CUBIC SPLINE TABLE DISPERSION */
 777             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 778             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 779             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 780             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 781             _MM_TRANSPOSE4_PS(Y,F,G,H);
 782             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 783             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 784             fvdw6            = _mm_mul_ps(c6_00,FF);
 785
 786             /* CUBIC SPLINE TABLE REPULSION */
 787             vfitab           = _mm_add_epi32(vfitab,ifour);
 788             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 789             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 790             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 791             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 792             _MM_TRANSPOSE4_PS(Y,F,G,H);
 793             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 794             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 795             fvdw12           = _mm_mul_ps(c12_00,FF);
 796             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 797
 798             fscal            = _mm_add_ps(felec,fvdw);
 799
 800             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 801
 802              /* Update vectorial force */
 803             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 804             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 805             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 806
 807             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 808             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 809             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 810             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 811             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 812                                                    _mm_mul_ps(dx00,fscal),
 813                                                    _mm_mul_ps(dy00,fscal),
 814                                                    _mm_mul_ps(dz00,fscal));
 815
 816             /* Inner loop uses 55 flops */
 817         }
 818
 819         /* End of innermost loop */
 820
 821         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 822                                               f+i_coord_offset,fshift+i_shift_offset);
 823
 824         /* Increment number of inner iterations */
 825         inneriter                  += j_index_end - j_index_start;
 826
 827         /* Outer loop uses 7 flops */
 828     }
 829
 830     /* Increment number of outer iterations */
 831     outeriter        += nri;
 832
 833     /* Update outer/inner flops */
 834
 835     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*55);
 836 }