src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_double 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_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_double
  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_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwjidx0A,vdwjidx0B;
  69     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  70     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  71     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  72     real             *charge;
  73     int              nvdwtype;
  74     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  75     int              *vdwtype;
  76     real             *vdwparam;
  77     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  78     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  79     __m128i          ewitab;
  80     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  81     real             *ewtab;
  82     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  83     real             rswitch_scalar,d_scalar;
  84     __m128d          dummy_mask,cutoff_mask;
  85     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  86     __m128d          one     = _mm_set1_pd(1.0);
  87     __m128d          two     = _mm_set1_pd(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     facel            = _mm_set1_pd(fr->epsfac);
 100     charge           = mdatoms->chargeA;
 101     nvdwtype         = fr->ntype;
 102     vdwparam         = fr->nbfp;
 103     vdwtype          = mdatoms->typeA;
 104
 105     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 106     ewtab            = fr->ic->tabq_coul_FDV0;
 107     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 108     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 109
 110     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 111     rcutoff_scalar   = fr->rcoulomb;
 112     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 113     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 114
 115     rswitch_scalar   = fr->rcoulomb_switch;
 116     rswitch          = _mm_set1_pd(rswitch_scalar);
 117     /* Setup switch parameters */
 118     d_scalar         = rcutoff_scalar-rswitch_scalar;
 119     d                = _mm_set1_pd(d_scalar);
 120     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 121     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 122     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 123     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 124     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 125     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 126
 127     /* Avoid stupid compiler warnings */
 128     jnrA = jnrB = 0;
 129     j_coord_offsetA = 0;
 130     j_coord_offsetB = 0;
 131
 132     outeriter        = 0;
 133     inneriter        = 0;
 134
 135     /* Start outer loop over neighborlists */
 136     for(iidx=0; iidx<nri; iidx++)
 137     {
 138         /* Load shift vector for this list */
 139         i_shift_offset   = DIM*shiftidx[iidx];
 140
 141         /* Load limits for loop over neighbors */
 142         j_index_start    = jindex[iidx];
 143         j_index_end      = jindex[iidx+1];
 144
 145         /* Get outer coordinate index */
 146         inr              = iinr[iidx];
 147         i_coord_offset   = DIM*inr;
 148
 149         /* Load i particle coords and add shift vector */
 150         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 151
 152         fix0             = _mm_setzero_pd();
 153         fiy0             = _mm_setzero_pd();
 154         fiz0             = _mm_setzero_pd();
 155
 156         /* Load parameters for i particles */
 157         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 158         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 159
 160         /* Reset potential sums */
 161         velecsum         = _mm_setzero_pd();
 162         vvdwsum          = _mm_setzero_pd();
 163
 164         /* Start inner kernel loop */
 165         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 166         {
 167
 168             /* Get j neighbor index, and coordinate index */
 169             jnrA             = jjnr[jidx];
 170             jnrB             = jjnr[jidx+1];
 171             j_coord_offsetA  = DIM*jnrA;
 172             j_coord_offsetB  = DIM*jnrB;
 173
 174             /* load j atom coordinates */
 175             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 176                                               &jx0,&jy0,&jz0);
 177
 178             /* Calculate displacement vector */
 179             dx00             = _mm_sub_pd(ix0,jx0);
 180             dy00             = _mm_sub_pd(iy0,jy0);
 181             dz00             = _mm_sub_pd(iz0,jz0);
 182
 183             /* Calculate squared distance and things based on it */
 184             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 185
 186             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 187
 188             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 189
 190             /* Load parameters for j particles */
 191             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 192             vdwjidx0A        = 2*vdwtype[jnrA+0];
 193             vdwjidx0B        = 2*vdwtype[jnrB+0];
 194
 195             /**************************
 196              * CALCULATE INTERACTIONS *
 197              **************************/
 198
 199             if (gmx_mm_any_lt(rsq00,rcutoff2))
 200             {
 201
 202             r00              = _mm_mul_pd(rsq00,rinv00);
 203
 204             /* Compute parameters for interactions between i and j atoms */
 205             qq00             = _mm_mul_pd(iq0,jq0);
 206             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 207                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 208
 209             /* EWALD ELECTROSTATICS */
 210
 211             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 212             ewrt             = _mm_mul_pd(r00,ewtabscale);
 213             ewitab           = _mm_cvttpd_epi32(ewrt);
 214 #ifdef __XOP__
 215             eweps            = _mm_frcz_pd(ewrt);
 216 #else
 217             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 218 #endif
 219             twoeweps         = _mm_add_pd(eweps,eweps);
 220             ewitab           = _mm_slli_epi32(ewitab,2);
 221             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 222             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 223             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 224             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 225             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 226             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 227             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 228             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 229             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 230             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 231
 232             /* LENNARD-JONES DISPERSION/REPULSION */
 233
 234             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 235             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 236             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 237             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 238             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 239
 240             d                = _mm_sub_pd(r00,rswitch);
 241             d                = _mm_max_pd(d,_mm_setzero_pd());
 242             d2               = _mm_mul_pd(d,d);
 243             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 244
 245             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 246
 247             /* Evaluate switch function */
 248             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 249             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 250             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 251             velec            = _mm_mul_pd(velec,sw);
 252             vvdw             = _mm_mul_pd(vvdw,sw);
 253             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 254
 255             /* Update potential sum for this i atom from the interaction with this j atom. */
 256             velec            = _mm_and_pd(velec,cutoff_mask);
 257             velecsum         = _mm_add_pd(velecsum,velec);
 258             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 259             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 260
 261             fscal            = _mm_add_pd(felec,fvdw);
 262
 263             fscal            = _mm_and_pd(fscal,cutoff_mask);
 264
 265             /* Update vectorial force */
 266             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 267             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 268             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 269
 270             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 271                                                    _mm_mul_pd(dx00,fscal),
 272                                                    _mm_mul_pd(dy00,fscal),
 273                                                    _mm_mul_pd(dz00,fscal));
 274
 275             }
 276
 277             /* Inner loop uses 86 flops */
 278         }
 279
 280         if(jidx<j_index_end)
 281         {
 282
 283             jnrA             = jjnr[jidx];
 284             j_coord_offsetA  = DIM*jnrA;
 285
 286             /* load j atom coordinates */
 287             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 288                                               &jx0,&jy0,&jz0);
 289
 290             /* Calculate displacement vector */
 291             dx00             = _mm_sub_pd(ix0,jx0);
 292             dy00             = _mm_sub_pd(iy0,jy0);
 293             dz00             = _mm_sub_pd(iz0,jz0);
 294
 295             /* Calculate squared distance and things based on it */
 296             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 297
 298             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 299
 300             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 301
 302             /* Load parameters for j particles */
 303             jq0              = _mm_load_sd(charge+jnrA+0);
 304             vdwjidx0A        = 2*vdwtype[jnrA+0];
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             if (gmx_mm_any_lt(rsq00,rcutoff2))
 311             {
 312
 313             r00              = _mm_mul_pd(rsq00,rinv00);
 314
 315             /* Compute parameters for interactions between i and j atoms */
 316             qq00             = _mm_mul_pd(iq0,jq0);
 317             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 318
 319             /* EWALD ELECTROSTATICS */
 320
 321             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 322             ewrt             = _mm_mul_pd(r00,ewtabscale);
 323             ewitab           = _mm_cvttpd_epi32(ewrt);
 324 #ifdef __XOP__
 325             eweps            = _mm_frcz_pd(ewrt);
 326 #else
 327             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 328 #endif
 329             twoeweps         = _mm_add_pd(eweps,eweps);
 330             ewitab           = _mm_slli_epi32(ewitab,2);
 331             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 332             ewtabD           = _mm_setzero_pd();
 333             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 334             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 335             ewtabFn          = _mm_setzero_pd();
 336             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 337             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 338             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 339             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 340             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 341
 342             /* LENNARD-JONES DISPERSION/REPULSION */
 343
 344             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 345             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 346             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 347             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 348             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 349
 350             d                = _mm_sub_pd(r00,rswitch);
 351             d                = _mm_max_pd(d,_mm_setzero_pd());
 352             d2               = _mm_mul_pd(d,d);
 353             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 354
 355             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 356
 357             /* Evaluate switch function */
 358             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 359             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 360             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 361             velec            = _mm_mul_pd(velec,sw);
 362             vvdw             = _mm_mul_pd(vvdw,sw);
 363             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             velec            = _mm_and_pd(velec,cutoff_mask);
 367             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 368             velecsum         = _mm_add_pd(velecsum,velec);
 369             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 370             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 371             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 372
 373             fscal            = _mm_add_pd(felec,fvdw);
 374
 375             fscal            = _mm_and_pd(fscal,cutoff_mask);
 376
 377             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 378
 379             /* Update vectorial force */
 380             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 381             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 382             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 383
 384             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 385                                                    _mm_mul_pd(dx00,fscal),
 386                                                    _mm_mul_pd(dy00,fscal),
 387                                                    _mm_mul_pd(dz00,fscal));
 388
 389             }
 390
 391             /* Inner loop uses 86 flops */
 392         }
 393
 394         /* End of innermost loop */
 395
 396         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 397                                               f+i_coord_offset,fshift+i_shift_offset);
 398
 399         ggid                        = gid[iidx];
 400         /* Update potential energies */
 401         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 402         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 403
 404         /* Increment number of inner iterations */
 405         inneriter                  += j_index_end - j_index_start;
 406
 407         /* Outer loop uses 9 flops */
 408     }
 409
 410     /* Increment number of outer iterations */
 411     outeriter        += nri;
 412
 413     /* Update outer/inner flops */
 414
 415     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*86);
 416 }
 417 /*
 418  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 419  * Electrostatics interaction: Ewald
 420  * VdW interaction:            LennardJones
 421  * Geometry:                   Particle-Particle
 422  * Calculate force/pot:        Force
 423  */
 424 void
 425 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 426                     (t_nblist * gmx_restrict                nlist,
 427                      rvec * gmx_restrict                    xx,
 428                      rvec * gmx_restrict                    ff,
 429                      t_forcerec * gmx_restrict              fr,
 430                      t_mdatoms * gmx_restrict               mdatoms,
 431                      nb_kernel_data_t * gmx_restrict        kernel_data,
 432                      t_nrnb * gmx_restrict                  nrnb)
 433 {
 434     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 435      * just 0 for non-waters.
 436      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 437      * jnr indices corresponding to data put in the four positions in the SIMD register.
 438      */
 439     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 440     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 441     int              jnrA,jnrB;
 442     int              j_coord_offsetA,j_coord_offsetB;
 443     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 444     real             rcutoff_scalar;
 445     real             *shiftvec,*fshift,*x,*f;
 446     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 447     int              vdwioffset0;
 448     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 449     int              vdwjidx0A,vdwjidx0B;
 450     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 451     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 452     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 453     real             *charge;
 454     int              nvdwtype;
 455     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 456     int              *vdwtype;
 457     real             *vdwparam;
 458     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 459     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 460     __m128i          ewitab;
 461     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 462     real             *ewtab;
 463     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 464     real             rswitch_scalar,d_scalar;
 465     __m128d          dummy_mask,cutoff_mask;
 466     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 467     __m128d          one     = _mm_set1_pd(1.0);
 468     __m128d          two     = _mm_set1_pd(2.0);
 469     x                = xx[0];
 470     f                = ff[0];
 471
 472     nri              = nlist->nri;
 473     iinr             = nlist->iinr;
 474     jindex           = nlist->jindex;
 475     jjnr             = nlist->jjnr;
 476     shiftidx         = nlist->shift;
 477     gid              = nlist->gid;
 478     shiftvec         = fr->shift_vec[0];
 479     fshift           = fr->fshift[0];
 480     facel            = _mm_set1_pd(fr->epsfac);
 481     charge           = mdatoms->chargeA;
 482     nvdwtype         = fr->ntype;
 483     vdwparam         = fr->nbfp;
 484     vdwtype          = mdatoms->typeA;
 485
 486     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 487     ewtab            = fr->ic->tabq_coul_FDV0;
 488     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 489     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 490
 491     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 492     rcutoff_scalar   = fr->rcoulomb;
 493     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 494     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 495
 496     rswitch_scalar   = fr->rcoulomb_switch;
 497     rswitch          = _mm_set1_pd(rswitch_scalar);
 498     /* Setup switch parameters */
 499     d_scalar         = rcutoff_scalar-rswitch_scalar;
 500     d                = _mm_set1_pd(d_scalar);
 501     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 502     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 503     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 504     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 505     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 506     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 507
 508     /* Avoid stupid compiler warnings */
 509     jnrA = jnrB = 0;
 510     j_coord_offsetA = 0;
 511     j_coord_offsetB = 0;
 512
 513     outeriter        = 0;
 514     inneriter        = 0;
 515
 516     /* Start outer loop over neighborlists */
 517     for(iidx=0; iidx<nri; iidx++)
 518     {
 519         /* Load shift vector for this list */
 520         i_shift_offset   = DIM*shiftidx[iidx];
 521
 522         /* Load limits for loop over neighbors */
 523         j_index_start    = jindex[iidx];
 524         j_index_end      = jindex[iidx+1];
 525
 526         /* Get outer coordinate index */
 527         inr              = iinr[iidx];
 528         i_coord_offset   = DIM*inr;
 529
 530         /* Load i particle coords and add shift vector */
 531         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 532
 533         fix0             = _mm_setzero_pd();
 534         fiy0             = _mm_setzero_pd();
 535         fiz0             = _mm_setzero_pd();
 536
 537         /* Load parameters for i particles */
 538         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 539         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 540
 541         /* Start inner kernel loop */
 542         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 543         {
 544
 545             /* Get j neighbor index, and coordinate index */
 546             jnrA             = jjnr[jidx];
 547             jnrB             = jjnr[jidx+1];
 548             j_coord_offsetA  = DIM*jnrA;
 549             j_coord_offsetB  = DIM*jnrB;
 550
 551             /* load j atom coordinates */
 552             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 553                                               &jx0,&jy0,&jz0);
 554
 555             /* Calculate displacement vector */
 556             dx00             = _mm_sub_pd(ix0,jx0);
 557             dy00             = _mm_sub_pd(iy0,jy0);
 558             dz00             = _mm_sub_pd(iz0,jz0);
 559
 560             /* Calculate squared distance and things based on it */
 561             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 562
 563             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 564
 565             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 566
 567             /* Load parameters for j particles */
 568             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 569             vdwjidx0A        = 2*vdwtype[jnrA+0];
 570             vdwjidx0B        = 2*vdwtype[jnrB+0];
 571
 572             /**************************
 573              * CALCULATE INTERACTIONS *
 574              **************************/
 575
 576             if (gmx_mm_any_lt(rsq00,rcutoff2))
 577             {
 578
 579             r00              = _mm_mul_pd(rsq00,rinv00);
 580
 581             /* Compute parameters for interactions between i and j atoms */
 582             qq00             = _mm_mul_pd(iq0,jq0);
 583             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 584                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 585
 586             /* EWALD ELECTROSTATICS */
 587
 588             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 589             ewrt             = _mm_mul_pd(r00,ewtabscale);
 590             ewitab           = _mm_cvttpd_epi32(ewrt);
 591 #ifdef __XOP__
 592             eweps            = _mm_frcz_pd(ewrt);
 593 #else
 594             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 595 #endif
 596             twoeweps         = _mm_add_pd(eweps,eweps);
 597             ewitab           = _mm_slli_epi32(ewitab,2);
 598             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 599             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 600             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 601             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 602             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 603             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 604             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 605             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 606             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 607             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 608
 609             /* LENNARD-JONES DISPERSION/REPULSION */
 610
 611             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 612             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 613             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 614             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 615             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 616
 617             d                = _mm_sub_pd(r00,rswitch);
 618             d                = _mm_max_pd(d,_mm_setzero_pd());
 619             d2               = _mm_mul_pd(d,d);
 620             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 621
 622             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 623
 624             /* Evaluate switch function */
 625             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 626             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 627             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 628             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 629
 630             fscal            = _mm_add_pd(felec,fvdw);
 631
 632             fscal            = _mm_and_pd(fscal,cutoff_mask);
 633
 634             /* Update vectorial force */
 635             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 636             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 637             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 638
 639             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 640                                                    _mm_mul_pd(dx00,fscal),
 641                                                    _mm_mul_pd(dy00,fscal),
 642                                                    _mm_mul_pd(dz00,fscal));
 643
 644             }
 645
 646             /* Inner loop uses 80 flops */
 647         }
 648
 649         if(jidx<j_index_end)
 650         {
 651
 652             jnrA             = jjnr[jidx];
 653             j_coord_offsetA  = DIM*jnrA;
 654
 655             /* load j atom coordinates */
 656             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 657                                               &jx0,&jy0,&jz0);
 658
 659             /* Calculate displacement vector */
 660             dx00             = _mm_sub_pd(ix0,jx0);
 661             dy00             = _mm_sub_pd(iy0,jy0);
 662             dz00             = _mm_sub_pd(iz0,jz0);
 663
 664             /* Calculate squared distance and things based on it */
 665             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 666
 667             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 668
 669             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 670
 671             /* Load parameters for j particles */
 672             jq0              = _mm_load_sd(charge+jnrA+0);
 673             vdwjidx0A        = 2*vdwtype[jnrA+0];
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             if (gmx_mm_any_lt(rsq00,rcutoff2))
 680             {
 681
 682             r00              = _mm_mul_pd(rsq00,rinv00);
 683
 684             /* Compute parameters for interactions between i and j atoms */
 685             qq00             = _mm_mul_pd(iq0,jq0);
 686             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 687
 688             /* EWALD ELECTROSTATICS */
 689
 690             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 691             ewrt             = _mm_mul_pd(r00,ewtabscale);
 692             ewitab           = _mm_cvttpd_epi32(ewrt);
 693 #ifdef __XOP__
 694             eweps            = _mm_frcz_pd(ewrt);
 695 #else
 696             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 697 #endif
 698             twoeweps         = _mm_add_pd(eweps,eweps);
 699             ewitab           = _mm_slli_epi32(ewitab,2);
 700             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 701             ewtabD           = _mm_setzero_pd();
 702             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 703             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 704             ewtabFn          = _mm_setzero_pd();
 705             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 706             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 707             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 708             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 709             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 710
 711             /* LENNARD-JONES DISPERSION/REPULSION */
 712
 713             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 714             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 715             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 716             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 717             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 718
 719             d                = _mm_sub_pd(r00,rswitch);
 720             d                = _mm_max_pd(d,_mm_setzero_pd());
 721             d2               = _mm_mul_pd(d,d);
 722             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 723
 724             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 725
 726             /* Evaluate switch function */
 727             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 728             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 729             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 730             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 731
 732             fscal            = _mm_add_pd(felec,fvdw);
 733
 734             fscal            = _mm_and_pd(fscal,cutoff_mask);
 735
 736             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 737
 738             /* Update vectorial force */
 739             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 740             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 741             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 742
 743             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 744                                                    _mm_mul_pd(dx00,fscal),
 745                                                    _mm_mul_pd(dy00,fscal),
 746                                                    _mm_mul_pd(dz00,fscal));
 747
 748             }
 749
 750             /* Inner loop uses 80 flops */
 751         }
 752
 753         /* End of innermost loop */
 754
 755         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 756                                               f+i_coord_offset,fshift+i_shift_offset);
 757
 758         /* Increment number of inner iterations */
 759         inneriter                  += j_index_end - j_index_start;
 760
 761         /* Outer loop uses 7 flops */
 762     }
 763
 764     /* Increment number of outer iterations */
 765     outeriter        += nri;
 766
 767     /* Update outer/inner flops */
 768
 769     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*80);
 770 }