src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSw_VdwLJSw_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_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSw_VdwLJSw_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          ewitab;
  83     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  84     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  85     real             *ewtab;
  86     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  87     real             rswitch_scalar,d_scalar;
  88     __m128           dummy_mask,cutoff_mask;
  89     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  90     __m128           one     = _mm_set1_ps(1.0);
  91     __m128           two     = _mm_set1_ps(2.0);
  92     x                = xx[0];
  93     f                = ff[0];
  94
  95     nri              = nlist->nri;
  96     iinr             = nlist->iinr;
  97     jindex           = nlist->jindex;
  98     jjnr             = nlist->jjnr;
  99     shiftidx         = nlist->shift;
 100     gid              = nlist->gid;
 101     shiftvec         = fr->shift_vec[0];
 102     fshift           = fr->fshift[0];
 103     facel            = _mm_set1_ps(fr->epsfac);
 104     charge           = mdatoms->chargeA;
 105     nvdwtype         = fr->ntype;
 106     vdwparam         = fr->nbfp;
 107     vdwtype          = mdatoms->typeA;
 108
 109     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 110     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 111     beta2            = _mm_mul_ps(beta,beta);
 112     beta3            = _mm_mul_ps(beta,beta2);
 113     ewtab            = fr->ic->tabq_coul_FDV0;
 114     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 115     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 116
 117     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 118     rcutoff_scalar   = fr->rcoulomb;
 119     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 120     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 121
 122     rswitch_scalar   = fr->rcoulomb_switch;
 123     rswitch          = _mm_set1_ps(rswitch_scalar);
 124     /* Setup switch parameters */
 125     d_scalar         = rcutoff_scalar-rswitch_scalar;
 126     d                = _mm_set1_ps(d_scalar);
 127     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 128     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 129     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 130     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 131     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 132     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 133
 134     /* Avoid stupid compiler warnings */
 135     jnrA = jnrB = jnrC = jnrD = 0;
 136     j_coord_offsetA = 0;
 137     j_coord_offsetB = 0;
 138     j_coord_offsetC = 0;
 139     j_coord_offsetD = 0;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     for(iidx=0;iidx<4*DIM;iidx++)
 145     {
 146         scratch[iidx] = 0.0;
 147     }
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 165
 166         fix0             = _mm_setzero_ps();
 167         fiy0             = _mm_setzero_ps();
 168         fiz0             = _mm_setzero_ps();
 169
 170         /* Load parameters for i particles */
 171         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 172         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_ps();
 176         vvdwsum          = _mm_setzero_ps();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             jnrC             = jjnr[jidx+2];
 186             jnrD             = jjnr[jidx+3];
 187             j_coord_offsetA  = DIM*jnrA;
 188             j_coord_offsetB  = DIM*jnrB;
 189             j_coord_offsetC  = DIM*jnrC;
 190             j_coord_offsetD  = DIM*jnrD;
 191
 192             /* load j atom coordinates */
 193             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 194                                               x+j_coord_offsetC,x+j_coord_offsetD,
 195                                               &jx0,&jy0,&jz0);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _mm_sub_ps(ix0,jx0);
 199             dy00             = _mm_sub_ps(iy0,jy0);
 200             dz00             = _mm_sub_ps(iz0,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 204
 205             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 206
 207             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 208
 209             /* Load parameters for j particles */
 210             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 211                                                               charge+jnrC+0,charge+jnrD+0);
 212             vdwjidx0A        = 2*vdwtype[jnrA+0];
 213             vdwjidx0B        = 2*vdwtype[jnrB+0];
 214             vdwjidx0C        = 2*vdwtype[jnrC+0];
 215             vdwjidx0D        = 2*vdwtype[jnrD+0];
 216
 217             /**************************
 218              * CALCULATE INTERACTIONS *
 219              **************************/
 220
 221             if (gmx_mm_any_lt(rsq00,rcutoff2))
 222             {
 223
 224             r00              = _mm_mul_ps(rsq00,rinv00);
 225
 226             /* Compute parameters for interactions between i and j atoms */
 227             qq00             = _mm_mul_ps(iq0,jq0);
 228             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 229                                          vdwparam+vdwioffset0+vdwjidx0B,
 230                                          vdwparam+vdwioffset0+vdwjidx0C,
 231                                          vdwparam+vdwioffset0+vdwjidx0D,
 232                                          &c6_00,&c12_00);
 233
 234             /* EWALD ELECTROSTATICS */
 235
 236             /* Analytical PME correction */
 237             zeta2            = _mm_mul_ps(beta2,rsq00);
 238             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 239             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 240             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 241             felec            = _mm_mul_ps(qq00,felec);
 242             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 243             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 244             velec            = _mm_mul_ps(qq00,velec);
 245
 246             /* LENNARD-JONES DISPERSION/REPULSION */
 247
 248             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 249             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 250             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 251             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 252             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 253
 254             d                = _mm_sub_ps(r00,rswitch);
 255             d                = _mm_max_ps(d,_mm_setzero_ps());
 256             d2               = _mm_mul_ps(d,d);
 257             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 258
 259             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 260
 261             /* Evaluate switch function */
 262             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 263             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 264             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 265             velec            = _mm_mul_ps(velec,sw);
 266             vvdw             = _mm_mul_ps(vvdw,sw);
 267             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 268
 269             /* Update potential sum for this i atom from the interaction with this j atom. */
 270             velec            = _mm_and_ps(velec,cutoff_mask);
 271             velecsum         = _mm_add_ps(velecsum,velec);
 272             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 273             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 274
 275             fscal            = _mm_add_ps(felec,fvdw);
 276
 277             fscal            = _mm_and_ps(fscal,cutoff_mask);
 278
 279              /* Update vectorial force */
 280             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 281             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 282             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 283
 284             fjptrA             = f+j_coord_offsetA;
 285             fjptrB             = f+j_coord_offsetB;
 286             fjptrC             = f+j_coord_offsetC;
 287             fjptrD             = f+j_coord_offsetD;
 288             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 289                                                    _mm_mul_ps(dx00,fscal),
 290                                                    _mm_mul_ps(dy00,fscal),
 291                                                    _mm_mul_ps(dz00,fscal));
 292
 293             }
 294
 295             /* Inner loop uses 71 flops */
 296         }
 297
 298         if(jidx<j_index_end)
 299         {
 300
 301             /* Get j neighbor index, and coordinate index */
 302             jnrlistA         = jjnr[jidx];
 303             jnrlistB         = jjnr[jidx+1];
 304             jnrlistC         = jjnr[jidx+2];
 305             jnrlistD         = jjnr[jidx+3];
 306             /* Sign of each element will be negative for non-real atoms.
 307              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 308              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 309              */
 310             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 311             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 312             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 313             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 314             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 315             j_coord_offsetA  = DIM*jnrA;
 316             j_coord_offsetB  = DIM*jnrB;
 317             j_coord_offsetC  = DIM*jnrC;
 318             j_coord_offsetD  = DIM*jnrD;
 319
 320             /* load j atom coordinates */
 321             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 322                                               x+j_coord_offsetC,x+j_coord_offsetD,
 323                                               &jx0,&jy0,&jz0);
 324
 325             /* Calculate displacement vector */
 326             dx00             = _mm_sub_ps(ix0,jx0);
 327             dy00             = _mm_sub_ps(iy0,jy0);
 328             dz00             = _mm_sub_ps(iz0,jz0);
 329
 330             /* Calculate squared distance and things based on it */
 331             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 332
 333             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 334
 335             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 336
 337             /* Load parameters for j particles */
 338             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 339                                                               charge+jnrC+0,charge+jnrD+0);
 340             vdwjidx0A        = 2*vdwtype[jnrA+0];
 341             vdwjidx0B        = 2*vdwtype[jnrB+0];
 342             vdwjidx0C        = 2*vdwtype[jnrC+0];
 343             vdwjidx0D        = 2*vdwtype[jnrD+0];
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             if (gmx_mm_any_lt(rsq00,rcutoff2))
 350             {
 351
 352             r00              = _mm_mul_ps(rsq00,rinv00);
 353             r00              = _mm_andnot_ps(dummy_mask,r00);
 354
 355             /* Compute parameters for interactions between i and j atoms */
 356             qq00             = _mm_mul_ps(iq0,jq0);
 357             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 358                                          vdwparam+vdwioffset0+vdwjidx0B,
 359                                          vdwparam+vdwioffset0+vdwjidx0C,
 360                                          vdwparam+vdwioffset0+vdwjidx0D,
 361                                          &c6_00,&c12_00);
 362
 363             /* EWALD ELECTROSTATICS */
 364
 365             /* Analytical PME correction */
 366             zeta2            = _mm_mul_ps(beta2,rsq00);
 367             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 368             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 369             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 370             felec            = _mm_mul_ps(qq00,felec);
 371             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 372             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 373             velec            = _mm_mul_ps(qq00,velec);
 374
 375             /* LENNARD-JONES DISPERSION/REPULSION */
 376
 377             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 378             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 379             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 380             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 381             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 382
 383             d                = _mm_sub_ps(r00,rswitch);
 384             d                = _mm_max_ps(d,_mm_setzero_ps());
 385             d2               = _mm_mul_ps(d,d);
 386             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 387
 388             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 389
 390             /* Evaluate switch function */
 391             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 392             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 393             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 394             velec            = _mm_mul_ps(velec,sw);
 395             vvdw             = _mm_mul_ps(vvdw,sw);
 396             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 397
 398             /* Update potential sum for this i atom from the interaction with this j atom. */
 399             velec            = _mm_and_ps(velec,cutoff_mask);
 400             velec            = _mm_andnot_ps(dummy_mask,velec);
 401             velecsum         = _mm_add_ps(velecsum,velec);
 402             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 403             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 404             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 405
 406             fscal            = _mm_add_ps(felec,fvdw);
 407
 408             fscal            = _mm_and_ps(fscal,cutoff_mask);
 409
 410             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 411
 412              /* Update vectorial force */
 413             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 414             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 415             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 416
 417             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 418             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 419             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 420             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 421             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 422                                                    _mm_mul_ps(dx00,fscal),
 423                                                    _mm_mul_ps(dy00,fscal),
 424                                                    _mm_mul_ps(dz00,fscal));
 425
 426             }
 427
 428             /* Inner loop uses 72 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*72);
 453 }
 454 /*
 455  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 456  * Electrostatics interaction: Ewald
 457  * VdW interaction:            LennardJones
 458  * Geometry:                   Particle-Particle
 459  * Calculate force/pot:        Force
 460  */
 461 void
 462 nb_kernel_ElecEwSw_VdwLJSw_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          ewitab;
 501     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 502     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 503     real             *ewtab;
 504     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 505     real             rswitch_scalar,d_scalar;
 506     __m128           dummy_mask,cutoff_mask;
 507     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 508     __m128           one     = _mm_set1_ps(1.0);
 509     __m128           two     = _mm_set1_ps(2.0);
 510     x                = xx[0];
 511     f                = ff[0];
 512
 513     nri              = nlist->nri;
 514     iinr             = nlist->iinr;
 515     jindex           = nlist->jindex;
 516     jjnr             = nlist->jjnr;
 517     shiftidx         = nlist->shift;
 518     gid              = nlist->gid;
 519     shiftvec         = fr->shift_vec[0];
 520     fshift           = fr->fshift[0];
 521     facel            = _mm_set1_ps(fr->epsfac);
 522     charge           = mdatoms->chargeA;
 523     nvdwtype         = fr->ntype;
 524     vdwparam         = fr->nbfp;
 525     vdwtype          = mdatoms->typeA;
 526
 527     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 528     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 529     beta2            = _mm_mul_ps(beta,beta);
 530     beta3            = _mm_mul_ps(beta,beta2);
 531     ewtab            = fr->ic->tabq_coul_FDV0;
 532     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 533     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 534
 535     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 536     rcutoff_scalar   = fr->rcoulomb;
 537     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 538     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 539
 540     rswitch_scalar   = fr->rcoulomb_switch;
 541     rswitch          = _mm_set1_ps(rswitch_scalar);
 542     /* Setup switch parameters */
 543     d_scalar         = rcutoff_scalar-rswitch_scalar;
 544     d                = _mm_set1_ps(d_scalar);
 545     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 546     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 547     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 548     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 549     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 550     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 551
 552     /* Avoid stupid compiler warnings */
 553     jnrA = jnrB = jnrC = jnrD = 0;
 554     j_coord_offsetA = 0;
 555     j_coord_offsetB = 0;
 556     j_coord_offsetC = 0;
 557     j_coord_offsetD = 0;
 558
 559     outeriter        = 0;
 560     inneriter        = 0;
 561
 562     for(iidx=0;iidx<4*DIM;iidx++)
 563     {
 564         scratch[iidx] = 0.0;
 565     }
 566
 567     /* Start outer loop over neighborlists */
 568     for(iidx=0; iidx<nri; iidx++)
 569     {
 570         /* Load shift vector for this list */
 571         i_shift_offset   = DIM*shiftidx[iidx];
 572
 573         /* Load limits for loop over neighbors */
 574         j_index_start    = jindex[iidx];
 575         j_index_end      = jindex[iidx+1];
 576
 577         /* Get outer coordinate index */
 578         inr              = iinr[iidx];
 579         i_coord_offset   = DIM*inr;
 580
 581         /* Load i particle coords and add shift vector */
 582         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 583
 584         fix0             = _mm_setzero_ps();
 585         fiy0             = _mm_setzero_ps();
 586         fiz0             = _mm_setzero_ps();
 587
 588         /* Load parameters for i particles */
 589         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 590         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 591
 592         /* Start inner kernel loop */
 593         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 594         {
 595
 596             /* Get j neighbor index, and coordinate index */
 597             jnrA             = jjnr[jidx];
 598             jnrB             = jjnr[jidx+1];
 599             jnrC             = jjnr[jidx+2];
 600             jnrD             = jjnr[jidx+3];
 601             j_coord_offsetA  = DIM*jnrA;
 602             j_coord_offsetB  = DIM*jnrB;
 603             j_coord_offsetC  = DIM*jnrC;
 604             j_coord_offsetD  = DIM*jnrD;
 605
 606             /* load j atom coordinates */
 607             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 608                                               x+j_coord_offsetC,x+j_coord_offsetD,
 609                                               &jx0,&jy0,&jz0);
 610
 611             /* Calculate displacement vector */
 612             dx00             = _mm_sub_ps(ix0,jx0);
 613             dy00             = _mm_sub_ps(iy0,jy0);
 614             dz00             = _mm_sub_ps(iz0,jz0);
 615
 616             /* Calculate squared distance and things based on it */
 617             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 618
 619             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 620
 621             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 622
 623             /* Load parameters for j particles */
 624             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 625                                                               charge+jnrC+0,charge+jnrD+0);
 626             vdwjidx0A        = 2*vdwtype[jnrA+0];
 627             vdwjidx0B        = 2*vdwtype[jnrB+0];
 628             vdwjidx0C        = 2*vdwtype[jnrC+0];
 629             vdwjidx0D        = 2*vdwtype[jnrD+0];
 630
 631             /**************************
 632              * CALCULATE INTERACTIONS *
 633              **************************/
 634
 635             if (gmx_mm_any_lt(rsq00,rcutoff2))
 636             {
 637
 638             r00              = _mm_mul_ps(rsq00,rinv00);
 639
 640             /* Compute parameters for interactions between i and j atoms */
 641             qq00             = _mm_mul_ps(iq0,jq0);
 642             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 643                                          vdwparam+vdwioffset0+vdwjidx0B,
 644                                          vdwparam+vdwioffset0+vdwjidx0C,
 645                                          vdwparam+vdwioffset0+vdwjidx0D,
 646                                          &c6_00,&c12_00);
 647
 648             /* EWALD ELECTROSTATICS */
 649
 650             /* Analytical PME correction */
 651             zeta2            = _mm_mul_ps(beta2,rsq00);
 652             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 653             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 654             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 655             felec            = _mm_mul_ps(qq00,felec);
 656             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 657             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 658             velec            = _mm_mul_ps(qq00,velec);
 659
 660             /* LENNARD-JONES DISPERSION/REPULSION */
 661
 662             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 663             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 664             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 665             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 666             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 667
 668             d                = _mm_sub_ps(r00,rswitch);
 669             d                = _mm_max_ps(d,_mm_setzero_ps());
 670             d2               = _mm_mul_ps(d,d);
 671             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 672
 673             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 674
 675             /* Evaluate switch function */
 676             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 677             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 678             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 679             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 680
 681             fscal            = _mm_add_ps(felec,fvdw);
 682
 683             fscal            = _mm_and_ps(fscal,cutoff_mask);
 684
 685              /* Update vectorial force */
 686             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 687             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 688             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 689
 690             fjptrA             = f+j_coord_offsetA;
 691             fjptrB             = f+j_coord_offsetB;
 692             fjptrC             = f+j_coord_offsetC;
 693             fjptrD             = f+j_coord_offsetD;
 694             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 695                                                    _mm_mul_ps(dx00,fscal),
 696                                                    _mm_mul_ps(dy00,fscal),
 697                                                    _mm_mul_ps(dz00,fscal));
 698
 699             }
 700
 701             /* Inner loop uses 65 flops */
 702         }
 703
 704         if(jidx<j_index_end)
 705         {
 706
 707             /* Get j neighbor index, and coordinate index */
 708             jnrlistA         = jjnr[jidx];
 709             jnrlistB         = jjnr[jidx+1];
 710             jnrlistC         = jjnr[jidx+2];
 711             jnrlistD         = jjnr[jidx+3];
 712             /* Sign of each element will be negative for non-real atoms.
 713              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 714              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 715              */
 716             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 717             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 718             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 719             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 720             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 721             j_coord_offsetA  = DIM*jnrA;
 722             j_coord_offsetB  = DIM*jnrB;
 723             j_coord_offsetC  = DIM*jnrC;
 724             j_coord_offsetD  = DIM*jnrD;
 725
 726             /* load j atom coordinates */
 727             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 728                                               x+j_coord_offsetC,x+j_coord_offsetD,
 729                                               &jx0,&jy0,&jz0);
 730
 731             /* Calculate displacement vector */
 732             dx00             = _mm_sub_ps(ix0,jx0);
 733             dy00             = _mm_sub_ps(iy0,jy0);
 734             dz00             = _mm_sub_ps(iz0,jz0);
 735
 736             /* Calculate squared distance and things based on it */
 737             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 738
 739             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 740
 741             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 742
 743             /* Load parameters for j particles */
 744             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 745                                                               charge+jnrC+0,charge+jnrD+0);
 746             vdwjidx0A        = 2*vdwtype[jnrA+0];
 747             vdwjidx0B        = 2*vdwtype[jnrB+0];
 748             vdwjidx0C        = 2*vdwtype[jnrC+0];
 749             vdwjidx0D        = 2*vdwtype[jnrD+0];
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             if (gmx_mm_any_lt(rsq00,rcutoff2))
 756             {
 757
 758             r00              = _mm_mul_ps(rsq00,rinv00);
 759             r00              = _mm_andnot_ps(dummy_mask,r00);
 760
 761             /* Compute parameters for interactions between i and j atoms */
 762             qq00             = _mm_mul_ps(iq0,jq0);
 763             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 764                                          vdwparam+vdwioffset0+vdwjidx0B,
 765                                          vdwparam+vdwioffset0+vdwjidx0C,
 766                                          vdwparam+vdwioffset0+vdwjidx0D,
 767                                          &c6_00,&c12_00);
 768
 769             /* EWALD ELECTROSTATICS */
 770
 771             /* Analytical PME correction */
 772             zeta2            = _mm_mul_ps(beta2,rsq00);
 773             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 774             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 775             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 776             felec            = _mm_mul_ps(qq00,felec);
 777             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 778             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 779             velec            = _mm_mul_ps(qq00,velec);
 780
 781             /* LENNARD-JONES DISPERSION/REPULSION */
 782
 783             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 784             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 785             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 786             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 787             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 788
 789             d                = _mm_sub_ps(r00,rswitch);
 790             d                = _mm_max_ps(d,_mm_setzero_ps());
 791             d2               = _mm_mul_ps(d,d);
 792             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 793
 794             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 795
 796             /* Evaluate switch function */
 797             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 798             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 799             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 800             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 801
 802             fscal            = _mm_add_ps(felec,fvdw);
 803
 804             fscal            = _mm_and_ps(fscal,cutoff_mask);
 805
 806             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 807
 808              /* Update vectorial force */
 809             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 810             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 811             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 812
 813             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 814             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 815             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 816             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 817             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 818                                                    _mm_mul_ps(dx00,fscal),
 819                                                    _mm_mul_ps(dy00,fscal),
 820                                                    _mm_mul_ps(dz00,fscal));
 821
 822             }
 823
 824             /* Inner loop uses 66 flops */
 825         }
 826
 827         /* End of innermost loop */
 828
 829         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 830                                               f+i_coord_offset,fshift+i_shift_offset);
 831
 832         /* Increment number of inner iterations */
 833         inneriter                  += j_index_end - j_index_start;
 834
 835         /* Outer loop uses 7 flops */
 836     }
 837
 838     /* Increment number of outer iterations */
 839     outeriter        += nri;
 840
 841     /* Update outer/inner flops */
 842
 843     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*66);
 844 }