src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwCSTab_GeomW4P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW4P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwCSTab_GeomW4P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  91     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  94     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  95     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  96     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     int              nvdwtype;
  99     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 103     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 104     __m128i          vfitab;
 105     __m128i          ifour       = _mm_set1_epi32(4);
 106     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 107     real             *vftab;
 108     __m128i          ewitab;
 109     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 110     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 111     real             *ewtab;
 112     __m128           dummy_mask,cutoff_mask;
 113     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 114     __m128           one     = _mm_set1_ps(1.0);
 115     __m128           two     = _mm_set1_ps(2.0);
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = _mm_set1_ps(fr->ic->epsfac);
 128     charge           = mdatoms->chargeA;
 129     nvdwtype         = fr->ntype;
 130     vdwparam         = fr->nbfp;
 131     vdwtype          = mdatoms->typeA;
 132
 133     vftab            = kernel_data->table_vdw->data;
 134     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 135
 136     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 137     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 138     beta2            = _mm_mul_ps(beta,beta);
 139     beta3            = _mm_mul_ps(beta,beta2);
 140     ewtab            = fr->ic->tabq_coul_FDV0;
 141     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 142     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 143
 144     /* Setup water-specific parameters */
 145     inr              = nlist->iinr[0];
 146     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 147     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 148     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 149     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 150
 151     /* Avoid stupid compiler warnings */
 152     jnrA = jnrB = jnrC = jnrD = 0;
 153     j_coord_offsetA = 0;
 154     j_coord_offsetB = 0;
 155     j_coord_offsetC = 0;
 156     j_coord_offsetD = 0;
 157
 158     outeriter        = 0;
 159     inneriter        = 0;
 160
 161     for(iidx=0;iidx<4*DIM;iidx++)
 162     {
 163         scratch[iidx] = 0.0;
 164     }
 165
 166     /* Start outer loop over neighborlists */
 167     for(iidx=0; iidx<nri; iidx++)
 168     {
 169         /* Load shift vector for this list */
 170         i_shift_offset   = DIM*shiftidx[iidx];
 171
 172         /* Load limits for loop over neighbors */
 173         j_index_start    = jindex[iidx];
 174         j_index_end      = jindex[iidx+1];
 175
 176         /* Get outer coordinate index */
 177         inr              = iinr[iidx];
 178         i_coord_offset   = DIM*inr;
 179
 180         /* Load i particle coords and add shift vector */
 181         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 182                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 183
 184         fix0             = _mm_setzero_ps();
 185         fiy0             = _mm_setzero_ps();
 186         fiz0             = _mm_setzero_ps();
 187         fix1             = _mm_setzero_ps();
 188         fiy1             = _mm_setzero_ps();
 189         fiz1             = _mm_setzero_ps();
 190         fix2             = _mm_setzero_ps();
 191         fiy2             = _mm_setzero_ps();
 192         fiz2             = _mm_setzero_ps();
 193         fix3             = _mm_setzero_ps();
 194         fiy3             = _mm_setzero_ps();
 195         fiz3             = _mm_setzero_ps();
 196
 197         /* Reset potential sums */
 198         velecsum         = _mm_setzero_ps();
 199         vvdwsum          = _mm_setzero_ps();
 200
 201         /* Start inner kernel loop */
 202         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 203         {
 204
 205             /* Get j neighbor index, and coordinate index */
 206             jnrA             = jjnr[jidx];
 207             jnrB             = jjnr[jidx+1];
 208             jnrC             = jjnr[jidx+2];
 209             jnrD             = jjnr[jidx+3];
 210             j_coord_offsetA  = DIM*jnrA;
 211             j_coord_offsetB  = DIM*jnrB;
 212             j_coord_offsetC  = DIM*jnrC;
 213             j_coord_offsetD  = DIM*jnrD;
 214
 215             /* load j atom coordinates */
 216             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 217                                               x+j_coord_offsetC,x+j_coord_offsetD,
 218                                               &jx0,&jy0,&jz0);
 219
 220             /* Calculate displacement vector */
 221             dx00             = _mm_sub_ps(ix0,jx0);
 222             dy00             = _mm_sub_ps(iy0,jy0);
 223             dz00             = _mm_sub_ps(iz0,jz0);
 224             dx10             = _mm_sub_ps(ix1,jx0);
 225             dy10             = _mm_sub_ps(iy1,jy0);
 226             dz10             = _mm_sub_ps(iz1,jz0);
 227             dx20             = _mm_sub_ps(ix2,jx0);
 228             dy20             = _mm_sub_ps(iy2,jy0);
 229             dz20             = _mm_sub_ps(iz2,jz0);
 230             dx30             = _mm_sub_ps(ix3,jx0);
 231             dy30             = _mm_sub_ps(iy3,jy0);
 232             dz30             = _mm_sub_ps(iz3,jz0);
 233
 234             /* Calculate squared distance and things based on it */
 235             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 236             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 237             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 238             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 239
 240             rinv00           = avx128fma_invsqrt_f(rsq00);
 241             rinv10           = avx128fma_invsqrt_f(rsq10);
 242             rinv20           = avx128fma_invsqrt_f(rsq20);
 243             rinv30           = avx128fma_invsqrt_f(rsq30);
 244
 245             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 246             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 247             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 248
 249             /* Load parameters for j particles */
 250             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 251                                                               charge+jnrC+0,charge+jnrD+0);
 252             vdwjidx0A        = 2*vdwtype[jnrA+0];
 253             vdwjidx0B        = 2*vdwtype[jnrB+0];
 254             vdwjidx0C        = 2*vdwtype[jnrC+0];
 255             vdwjidx0D        = 2*vdwtype[jnrD+0];
 256
 257             fjx0             = _mm_setzero_ps();
 258             fjy0             = _mm_setzero_ps();
 259             fjz0             = _mm_setzero_ps();
 260
 261             /**************************
 262              * CALCULATE INTERACTIONS *
 263              **************************/
 264
 265             r00              = _mm_mul_ps(rsq00,rinv00);
 266
 267             /* Compute parameters for interactions between i and j atoms */
 268             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 269                                          vdwparam+vdwioffset0+vdwjidx0B,
 270                                          vdwparam+vdwioffset0+vdwjidx0C,
 271                                          vdwparam+vdwioffset0+vdwjidx0D,
 272                                          &c6_00,&c12_00);
 273
 274             /* Calculate table index by multiplying r with table scale and truncate to integer */
 275             rt               = _mm_mul_ps(r00,vftabscale);
 276             vfitab           = _mm_cvttps_epi32(rt);
 277 #ifdef __XOP__
 278             vfeps            = _mm_frcz_ps(rt);
 279 #else
 280             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 281 #endif
 282             twovfeps         = _mm_add_ps(vfeps,vfeps);
 283             vfitab           = _mm_slli_epi32(vfitab,3);
 284
 285             /* CUBIC SPLINE TABLE DISPERSION */
 286             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 287             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 288             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 289             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 290             _MM_TRANSPOSE4_PS(Y,F,G,H);
 291             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 292             VV               = _mm_macc_ps(vfeps,Fp,Y);
 293             vvdw6            = _mm_mul_ps(c6_00,VV);
 294             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 295             fvdw6            = _mm_mul_ps(c6_00,FF);
 296
 297             /* CUBIC SPLINE TABLE REPULSION */
 298             vfitab           = _mm_add_epi32(vfitab,ifour);
 299             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 300             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 301             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 302             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 303             _MM_TRANSPOSE4_PS(Y,F,G,H);
 304             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 305             VV               = _mm_macc_ps(vfeps,Fp,Y);
 306             vvdw12           = _mm_mul_ps(c12_00,VV);
 307             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 308             fvdw12           = _mm_mul_ps(c12_00,FF);
 309             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 310             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 311
 312             /* Update potential sum for this i atom from the interaction with this j atom. */
 313             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 314
 315             fscal            = fvdw;
 316
 317              /* Update vectorial force */
 318             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 319             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 320             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 321
 322             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 323             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 324             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 325
 326             /**************************
 327              * CALCULATE INTERACTIONS *
 328              **************************/
 329
 330             r10              = _mm_mul_ps(rsq10,rinv10);
 331
 332             /* Compute parameters for interactions between i and j atoms */
 333             qq10             = _mm_mul_ps(iq1,jq0);
 334
 335             /* EWALD ELECTROSTATICS */
 336
 337             /* Analytical PME correction */
 338             zeta2            = _mm_mul_ps(beta2,rsq10);
 339             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 340             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 341             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 342             felec            = _mm_mul_ps(qq10,felec);
 343             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 344             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 345             velec            = _mm_mul_ps(qq10,velec);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velecsum         = _mm_add_ps(velecsum,velec);
 349
 350             fscal            = felec;
 351
 352              /* Update vectorial force */
 353             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 354             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 355             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 356
 357             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 358             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 359             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 360
 361             /**************************
 362              * CALCULATE INTERACTIONS *
 363              **************************/
 364
 365             r20              = _mm_mul_ps(rsq20,rinv20);
 366
 367             /* Compute parameters for interactions between i and j atoms */
 368             qq20             = _mm_mul_ps(iq2,jq0);
 369
 370             /* EWALD ELECTROSTATICS */
 371
 372             /* Analytical PME correction */
 373             zeta2            = _mm_mul_ps(beta2,rsq20);
 374             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 375             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 376             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 377             felec            = _mm_mul_ps(qq20,felec);
 378             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 379             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 380             velec            = _mm_mul_ps(qq20,velec);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velecsum         = _mm_add_ps(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387              /* Update vectorial force */
 388             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 389             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 390             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 391
 392             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 393             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 394             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 395
 396             /**************************
 397              * CALCULATE INTERACTIONS *
 398              **************************/
 399
 400             r30              = _mm_mul_ps(rsq30,rinv30);
 401
 402             /* Compute parameters for interactions between i and j atoms */
 403             qq30             = _mm_mul_ps(iq3,jq0);
 404
 405             /* EWALD ELECTROSTATICS */
 406
 407             /* Analytical PME correction */
 408             zeta2            = _mm_mul_ps(beta2,rsq30);
 409             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
 410             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 411             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 412             felec            = _mm_mul_ps(qq30,felec);
 413             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 414             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv30);
 415             velec            = _mm_mul_ps(qq30,velec);
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velecsum         = _mm_add_ps(velecsum,velec);
 419
 420             fscal            = felec;
 421
 422              /* Update vectorial force */
 423             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 424             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 425             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 426
 427             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 428             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 429             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 430
 431             fjptrA             = f+j_coord_offsetA;
 432             fjptrB             = f+j_coord_offsetB;
 433             fjptrC             = f+j_coord_offsetC;
 434             fjptrD             = f+j_coord_offsetD;
 435
 436             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 437
 438             /* Inner loop uses 146 flops */
 439         }
 440
 441         if(jidx<j_index_end)
 442         {
 443
 444             /* Get j neighbor index, and coordinate index */
 445             jnrlistA         = jjnr[jidx];
 446             jnrlistB         = jjnr[jidx+1];
 447             jnrlistC         = jjnr[jidx+2];
 448             jnrlistD         = jjnr[jidx+3];
 449             /* Sign of each element will be negative for non-real atoms.
 450              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 451              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 452              */
 453             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 454             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 455             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 456             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 457             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 458             j_coord_offsetA  = DIM*jnrA;
 459             j_coord_offsetB  = DIM*jnrB;
 460             j_coord_offsetC  = DIM*jnrC;
 461             j_coord_offsetD  = DIM*jnrD;
 462
 463             /* load j atom coordinates */
 464             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 465                                               x+j_coord_offsetC,x+j_coord_offsetD,
 466                                               &jx0,&jy0,&jz0);
 467
 468             /* Calculate displacement vector */
 469             dx00             = _mm_sub_ps(ix0,jx0);
 470             dy00             = _mm_sub_ps(iy0,jy0);
 471             dz00             = _mm_sub_ps(iz0,jz0);
 472             dx10             = _mm_sub_ps(ix1,jx0);
 473             dy10             = _mm_sub_ps(iy1,jy0);
 474             dz10             = _mm_sub_ps(iz1,jz0);
 475             dx20             = _mm_sub_ps(ix2,jx0);
 476             dy20             = _mm_sub_ps(iy2,jy0);
 477             dz20             = _mm_sub_ps(iz2,jz0);
 478             dx30             = _mm_sub_ps(ix3,jx0);
 479             dy30             = _mm_sub_ps(iy3,jy0);
 480             dz30             = _mm_sub_ps(iz3,jz0);
 481
 482             /* Calculate squared distance and things based on it */
 483             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 484             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 485             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 486             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 487
 488             rinv00           = avx128fma_invsqrt_f(rsq00);
 489             rinv10           = avx128fma_invsqrt_f(rsq10);
 490             rinv20           = avx128fma_invsqrt_f(rsq20);
 491             rinv30           = avx128fma_invsqrt_f(rsq30);
 492
 493             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 494             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 495             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 496
 497             /* Load parameters for j particles */
 498             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 499                                                               charge+jnrC+0,charge+jnrD+0);
 500             vdwjidx0A        = 2*vdwtype[jnrA+0];
 501             vdwjidx0B        = 2*vdwtype[jnrB+0];
 502             vdwjidx0C        = 2*vdwtype[jnrC+0];
 503             vdwjidx0D        = 2*vdwtype[jnrD+0];
 504
 505             fjx0             = _mm_setzero_ps();
 506             fjy0             = _mm_setzero_ps();
 507             fjz0             = _mm_setzero_ps();
 508
 509             /**************************
 510              * CALCULATE INTERACTIONS *
 511              **************************/
 512
 513             r00              = _mm_mul_ps(rsq00,rinv00);
 514             r00              = _mm_andnot_ps(dummy_mask,r00);
 515
 516             /* Compute parameters for interactions between i and j atoms */
 517             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 518                                          vdwparam+vdwioffset0+vdwjidx0B,
 519                                          vdwparam+vdwioffset0+vdwjidx0C,
 520                                          vdwparam+vdwioffset0+vdwjidx0D,
 521                                          &c6_00,&c12_00);
 522
 523             /* Calculate table index by multiplying r with table scale and truncate to integer */
 524             rt               = _mm_mul_ps(r00,vftabscale);
 525             vfitab           = _mm_cvttps_epi32(rt);
 526 #ifdef __XOP__
 527             vfeps            = _mm_frcz_ps(rt);
 528 #else
 529             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 530 #endif
 531             twovfeps         = _mm_add_ps(vfeps,vfeps);
 532             vfitab           = _mm_slli_epi32(vfitab,3);
 533
 534             /* CUBIC SPLINE TABLE DISPERSION */
 535             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 536             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 537             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 538             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 539             _MM_TRANSPOSE4_PS(Y,F,G,H);
 540             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 541             VV               = _mm_macc_ps(vfeps,Fp,Y);
 542             vvdw6            = _mm_mul_ps(c6_00,VV);
 543             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 544             fvdw6            = _mm_mul_ps(c6_00,FF);
 545
 546             /* CUBIC SPLINE TABLE REPULSION */
 547             vfitab           = _mm_add_epi32(vfitab,ifour);
 548             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 549             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 550             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 551             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 552             _MM_TRANSPOSE4_PS(Y,F,G,H);
 553             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 554             VV               = _mm_macc_ps(vfeps,Fp,Y);
 555             vvdw12           = _mm_mul_ps(c12_00,VV);
 556             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 557             fvdw12           = _mm_mul_ps(c12_00,FF);
 558             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 559             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 560
 561             /* Update potential sum for this i atom from the interaction with this j atom. */
 562             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 563             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 564
 565             fscal            = fvdw;
 566
 567             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 568
 569              /* Update vectorial force */
 570             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 571             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 572             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 573
 574             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 575             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 576             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 577
 578             /**************************
 579              * CALCULATE INTERACTIONS *
 580              **************************/
 581
 582             r10              = _mm_mul_ps(rsq10,rinv10);
 583             r10              = _mm_andnot_ps(dummy_mask,r10);
 584
 585             /* Compute parameters for interactions between i and j atoms */
 586             qq10             = _mm_mul_ps(iq1,jq0);
 587
 588             /* EWALD ELECTROSTATICS */
 589
 590             /* Analytical PME correction */
 591             zeta2            = _mm_mul_ps(beta2,rsq10);
 592             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 593             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 594             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 595             felec            = _mm_mul_ps(qq10,felec);
 596             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 597             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 598             velec            = _mm_mul_ps(qq10,velec);
 599
 600             /* Update potential sum for this i atom from the interaction with this j atom. */
 601             velec            = _mm_andnot_ps(dummy_mask,velec);
 602             velecsum         = _mm_add_ps(velecsum,velec);
 603
 604             fscal            = felec;
 605
 606             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 607
 608              /* Update vectorial force */
 609             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 610             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 611             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 612
 613             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 614             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 615             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 616
 617             /**************************
 618              * CALCULATE INTERACTIONS *
 619              **************************/
 620
 621             r20              = _mm_mul_ps(rsq20,rinv20);
 622             r20              = _mm_andnot_ps(dummy_mask,r20);
 623
 624             /* Compute parameters for interactions between i and j atoms */
 625             qq20             = _mm_mul_ps(iq2,jq0);
 626
 627             /* EWALD ELECTROSTATICS */
 628
 629             /* Analytical PME correction */
 630             zeta2            = _mm_mul_ps(beta2,rsq20);
 631             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 632             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 633             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 634             felec            = _mm_mul_ps(qq20,felec);
 635             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 636             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 637             velec            = _mm_mul_ps(qq20,velec);
 638
 639             /* Update potential sum for this i atom from the interaction with this j atom. */
 640             velec            = _mm_andnot_ps(dummy_mask,velec);
 641             velecsum         = _mm_add_ps(velecsum,velec);
 642
 643             fscal            = felec;
 644
 645             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 646
 647              /* Update vectorial force */
 648             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 649             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 650             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 651
 652             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 653             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 654             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 655
 656             /**************************
 657              * CALCULATE INTERACTIONS *
 658              **************************/
 659
 660             r30              = _mm_mul_ps(rsq30,rinv30);
 661             r30              = _mm_andnot_ps(dummy_mask,r30);
 662
 663             /* Compute parameters for interactions between i and j atoms */
 664             qq30             = _mm_mul_ps(iq3,jq0);
 665
 666             /* EWALD ELECTROSTATICS */
 667
 668             /* Analytical PME correction */
 669             zeta2            = _mm_mul_ps(beta2,rsq30);
 670             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
 671             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 672             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 673             felec            = _mm_mul_ps(qq30,felec);
 674             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 675             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv30);
 676             velec            = _mm_mul_ps(qq30,velec);
 677
 678             /* Update potential sum for this i atom from the interaction with this j atom. */
 679             velec            = _mm_andnot_ps(dummy_mask,velec);
 680             velecsum         = _mm_add_ps(velecsum,velec);
 681
 682             fscal            = felec;
 683
 684             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 685
 686              /* Update vectorial force */
 687             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 688             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 689             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 690
 691             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 692             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 693             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 694
 695             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 696             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 697             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 698             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 699
 700             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 701
 702             /* Inner loop uses 150 flops */
 703         }
 704
 705         /* End of innermost loop */
 706
 707         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 708                                               f+i_coord_offset,fshift+i_shift_offset);
 709
 710         ggid                        = gid[iidx];
 711         /* Update potential energies */
 712         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 713         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 714
 715         /* Increment number of inner iterations */
 716         inneriter                  += j_index_end - j_index_start;
 717
 718         /* Outer loop uses 26 flops */
 719     }
 720
 721     /* Increment number of outer iterations */
 722     outeriter        += nri;
 723
 724     /* Update outer/inner flops */
 725
 726     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*150);
 727 }
 728 /*
 729  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW4P1_F_avx_128_fma_single
 730  * Electrostatics interaction: Ewald
 731  * VdW interaction:            CubicSplineTable
 732  * Geometry:                   Water4-Particle
 733  * Calculate force/pot:        Force
 734  */
 735 void
 736 nb_kernel_ElecEw_VdwCSTab_GeomW4P1_F_avx_128_fma_single
 737                     (t_nblist                    * gmx_restrict       nlist,
 738                      rvec                        * gmx_restrict          xx,
 739                      rvec                        * gmx_restrict          ff,
 740                      struct t_forcerec           * gmx_restrict          fr,
 741                      t_mdatoms                   * gmx_restrict     mdatoms,
 742                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 743                      t_nrnb                      * gmx_restrict        nrnb)
 744 {
 745     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 746      * just 0 for non-waters.
 747      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 748      * jnr indices corresponding to data put in the four positions in the SIMD register.
 749      */
 750     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 751     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 752     int              jnrA,jnrB,jnrC,jnrD;
 753     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 754     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 755     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 756     real             rcutoff_scalar;
 757     real             *shiftvec,*fshift,*x,*f;
 758     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 759     real             scratch[4*DIM];
 760     __m128           fscal,rcutoff,rcutoff2,jidxall;
 761     int              vdwioffset0;
 762     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 763     int              vdwioffset1;
 764     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 765     int              vdwioffset2;
 766     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 767     int              vdwioffset3;
 768     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 769     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 770     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 771     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 772     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 773     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 774     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 775     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 776     real             *charge;
 777     int              nvdwtype;
 778     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 779     int              *vdwtype;
 780     real             *vdwparam;
 781     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 782     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 783     __m128i          vfitab;
 784     __m128i          ifour       = _mm_set1_epi32(4);
 785     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 786     real             *vftab;
 787     __m128i          ewitab;
 788     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 789     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 790     real             *ewtab;
 791     __m128           dummy_mask,cutoff_mask;
 792     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 793     __m128           one     = _mm_set1_ps(1.0);
 794     __m128           two     = _mm_set1_ps(2.0);
 795     x                = xx[0];
 796     f                = ff[0];
 797
 798     nri              = nlist->nri;
 799     iinr             = nlist->iinr;
 800     jindex           = nlist->jindex;
 801     jjnr             = nlist->jjnr;
 802     shiftidx         = nlist->shift;
 803     gid              = nlist->gid;
 804     shiftvec         = fr->shift_vec[0];
 805     fshift           = fr->fshift[0];
 806     facel            = _mm_set1_ps(fr->ic->epsfac);
 807     charge           = mdatoms->chargeA;
 808     nvdwtype         = fr->ntype;
 809     vdwparam         = fr->nbfp;
 810     vdwtype          = mdatoms->typeA;
 811
 812     vftab            = kernel_data->table_vdw->data;
 813     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 814
 815     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 816     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 817     beta2            = _mm_mul_ps(beta,beta);
 818     beta3            = _mm_mul_ps(beta,beta2);
 819     ewtab            = fr->ic->tabq_coul_F;
 820     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 821     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 822
 823     /* Setup water-specific parameters */
 824     inr              = nlist->iinr[0];
 825     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 826     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 827     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 828     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 829
 830     /* Avoid stupid compiler warnings */
 831     jnrA = jnrB = jnrC = jnrD = 0;
 832     j_coord_offsetA = 0;
 833     j_coord_offsetB = 0;
 834     j_coord_offsetC = 0;
 835     j_coord_offsetD = 0;
 836
 837     outeriter        = 0;
 838     inneriter        = 0;
 839
 840     for(iidx=0;iidx<4*DIM;iidx++)
 841     {
 842         scratch[iidx] = 0.0;
 843     }
 844
 845     /* Start outer loop over neighborlists */
 846     for(iidx=0; iidx<nri; iidx++)
 847     {
 848         /* Load shift vector for this list */
 849         i_shift_offset   = DIM*shiftidx[iidx];
 850
 851         /* Load limits for loop over neighbors */
 852         j_index_start    = jindex[iidx];
 853         j_index_end      = jindex[iidx+1];
 854
 855         /* Get outer coordinate index */
 856         inr              = iinr[iidx];
 857         i_coord_offset   = DIM*inr;
 858
 859         /* Load i particle coords and add shift vector */
 860         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 861                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 862
 863         fix0             = _mm_setzero_ps();
 864         fiy0             = _mm_setzero_ps();
 865         fiz0             = _mm_setzero_ps();
 866         fix1             = _mm_setzero_ps();
 867         fiy1             = _mm_setzero_ps();
 868         fiz1             = _mm_setzero_ps();
 869         fix2             = _mm_setzero_ps();
 870         fiy2             = _mm_setzero_ps();
 871         fiz2             = _mm_setzero_ps();
 872         fix3             = _mm_setzero_ps();
 873         fiy3             = _mm_setzero_ps();
 874         fiz3             = _mm_setzero_ps();
 875
 876         /* Start inner kernel loop */
 877         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 878         {
 879
 880             /* Get j neighbor index, and coordinate index */
 881             jnrA             = jjnr[jidx];
 882             jnrB             = jjnr[jidx+1];
 883             jnrC             = jjnr[jidx+2];
 884             jnrD             = jjnr[jidx+3];
 885             j_coord_offsetA  = DIM*jnrA;
 886             j_coord_offsetB  = DIM*jnrB;
 887             j_coord_offsetC  = DIM*jnrC;
 888             j_coord_offsetD  = DIM*jnrD;
 889
 890             /* load j atom coordinates */
 891             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 892                                               x+j_coord_offsetC,x+j_coord_offsetD,
 893                                               &jx0,&jy0,&jz0);
 894
 895             /* Calculate displacement vector */
 896             dx00             = _mm_sub_ps(ix0,jx0);
 897             dy00             = _mm_sub_ps(iy0,jy0);
 898             dz00             = _mm_sub_ps(iz0,jz0);
 899             dx10             = _mm_sub_ps(ix1,jx0);
 900             dy10             = _mm_sub_ps(iy1,jy0);
 901             dz10             = _mm_sub_ps(iz1,jz0);
 902             dx20             = _mm_sub_ps(ix2,jx0);
 903             dy20             = _mm_sub_ps(iy2,jy0);
 904             dz20             = _mm_sub_ps(iz2,jz0);
 905             dx30             = _mm_sub_ps(ix3,jx0);
 906             dy30             = _mm_sub_ps(iy3,jy0);
 907             dz30             = _mm_sub_ps(iz3,jz0);
 908
 909             /* Calculate squared distance and things based on it */
 910             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 911             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 912             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 913             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 914
 915             rinv00           = avx128fma_invsqrt_f(rsq00);
 916             rinv10           = avx128fma_invsqrt_f(rsq10);
 917             rinv20           = avx128fma_invsqrt_f(rsq20);
 918             rinv30           = avx128fma_invsqrt_f(rsq30);
 919
 920             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 921             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 922             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 923
 924             /* Load parameters for j particles */
 925             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 926                                                               charge+jnrC+0,charge+jnrD+0);
 927             vdwjidx0A        = 2*vdwtype[jnrA+0];
 928             vdwjidx0B        = 2*vdwtype[jnrB+0];
 929             vdwjidx0C        = 2*vdwtype[jnrC+0];
 930             vdwjidx0D        = 2*vdwtype[jnrD+0];
 931
 932             fjx0             = _mm_setzero_ps();
 933             fjy0             = _mm_setzero_ps();
 934             fjz0             = _mm_setzero_ps();
 935
 936             /**************************
 937              * CALCULATE INTERACTIONS *
 938              **************************/
 939
 940             r00              = _mm_mul_ps(rsq00,rinv00);
 941
 942             /* Compute parameters for interactions between i and j atoms */
 943             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 944                                          vdwparam+vdwioffset0+vdwjidx0B,
 945                                          vdwparam+vdwioffset0+vdwjidx0C,
 946                                          vdwparam+vdwioffset0+vdwjidx0D,
 947                                          &c6_00,&c12_00);
 948
 949             /* Calculate table index by multiplying r with table scale and truncate to integer */
 950             rt               = _mm_mul_ps(r00,vftabscale);
 951             vfitab           = _mm_cvttps_epi32(rt);
 952 #ifdef __XOP__
 953             vfeps            = _mm_frcz_ps(rt);
 954 #else
 955             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 956 #endif
 957             twovfeps         = _mm_add_ps(vfeps,vfeps);
 958             vfitab           = _mm_slli_epi32(vfitab,3);
 959
 960             /* CUBIC SPLINE TABLE DISPERSION */
 961             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 962             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 963             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 964             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 965             _MM_TRANSPOSE4_PS(Y,F,G,H);
 966             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 967             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 968             fvdw6            = _mm_mul_ps(c6_00,FF);
 969
 970             /* CUBIC SPLINE TABLE REPULSION */
 971             vfitab           = _mm_add_epi32(vfitab,ifour);
 972             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 973             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 974             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 975             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 976             _MM_TRANSPOSE4_PS(Y,F,G,H);
 977             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 978             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 979             fvdw12           = _mm_mul_ps(c12_00,FF);
 980             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 981
 982             fscal            = fvdw;
 983
 984              /* Update vectorial force */
 985             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 986             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 987             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 988
 989             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 990             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 991             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 992
 993             /**************************
 994              * CALCULATE INTERACTIONS *
 995              **************************/
 996
 997             r10              = _mm_mul_ps(rsq10,rinv10);
 998
 999             /* Compute parameters for interactions between i and j atoms */
1000             qq10             = _mm_mul_ps(iq1,jq0);
1001
1002             /* EWALD ELECTROSTATICS */
1003
1004             /* Analytical PME correction */
1005             zeta2            = _mm_mul_ps(beta2,rsq10);
1006             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1007             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1008             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1009             felec            = _mm_mul_ps(qq10,felec);
1010
1011             fscal            = felec;
1012
1013              /* Update vectorial force */
1014             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1015             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1016             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1017
1018             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1019             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1020             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1021
1022             /**************************
1023              * CALCULATE INTERACTIONS *
1024              **************************/
1025
1026             r20              = _mm_mul_ps(rsq20,rinv20);
1027
1028             /* Compute parameters for interactions between i and j atoms */
1029             qq20             = _mm_mul_ps(iq2,jq0);
1030
1031             /* EWALD ELECTROSTATICS */
1032
1033             /* Analytical PME correction */
1034             zeta2            = _mm_mul_ps(beta2,rsq20);
1035             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1036             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1037             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1038             felec            = _mm_mul_ps(qq20,felec);
1039
1040             fscal            = felec;
1041
1042              /* Update vectorial force */
1043             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1044             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1045             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1046
1047             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1048             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1049             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1050
1051             /**************************
1052              * CALCULATE INTERACTIONS *
1053              **************************/
1054
1055             r30              = _mm_mul_ps(rsq30,rinv30);
1056
1057             /* Compute parameters for interactions between i and j atoms */
1058             qq30             = _mm_mul_ps(iq3,jq0);
1059
1060             /* EWALD ELECTROSTATICS */
1061
1062             /* Analytical PME correction */
1063             zeta2            = _mm_mul_ps(beta2,rsq30);
1064             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
1065             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1066             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1067             felec            = _mm_mul_ps(qq30,felec);
1068
1069             fscal            = felec;
1070
1071              /* Update vectorial force */
1072             fix3             = _mm_macc_ps(dx30,fscal,fix3);
1073             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
1074             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
1075
1076             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
1077             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
1078             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
1079
1080             fjptrA             = f+j_coord_offsetA;
1081             fjptrB             = f+j_coord_offsetB;
1082             fjptrC             = f+j_coord_offsetC;
1083             fjptrD             = f+j_coord_offsetD;
1084
1085             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1086
1087             /* Inner loop uses 135 flops */
1088         }
1089
1090         if(jidx<j_index_end)
1091         {
1092
1093             /* Get j neighbor index, and coordinate index */
1094             jnrlistA         = jjnr[jidx];
1095             jnrlistB         = jjnr[jidx+1];
1096             jnrlistC         = jjnr[jidx+2];
1097             jnrlistD         = jjnr[jidx+3];
1098             /* Sign of each element will be negative for non-real atoms.
1099              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1100              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1101              */
1102             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1103             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1104             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1105             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1106             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1107             j_coord_offsetA  = DIM*jnrA;
1108             j_coord_offsetB  = DIM*jnrB;
1109             j_coord_offsetC  = DIM*jnrC;
1110             j_coord_offsetD  = DIM*jnrD;
1111
1112             /* load j atom coordinates */
1113             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1114                                               x+j_coord_offsetC,x+j_coord_offsetD,
1115                                               &jx0,&jy0,&jz0);
1116
1117             /* Calculate displacement vector */
1118             dx00             = _mm_sub_ps(ix0,jx0);
1119             dy00             = _mm_sub_ps(iy0,jy0);
1120             dz00             = _mm_sub_ps(iz0,jz0);
1121             dx10             = _mm_sub_ps(ix1,jx0);
1122             dy10             = _mm_sub_ps(iy1,jy0);
1123             dz10             = _mm_sub_ps(iz1,jz0);
1124             dx20             = _mm_sub_ps(ix2,jx0);
1125             dy20             = _mm_sub_ps(iy2,jy0);
1126             dz20             = _mm_sub_ps(iz2,jz0);
1127             dx30             = _mm_sub_ps(ix3,jx0);
1128             dy30             = _mm_sub_ps(iy3,jy0);
1129             dz30             = _mm_sub_ps(iz3,jz0);
1130
1131             /* Calculate squared distance and things based on it */
1132             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1133             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1134             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1135             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
1136
1137             rinv00           = avx128fma_invsqrt_f(rsq00);
1138             rinv10           = avx128fma_invsqrt_f(rsq10);
1139             rinv20           = avx128fma_invsqrt_f(rsq20);
1140             rinv30           = avx128fma_invsqrt_f(rsq30);
1141
1142             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1143             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1144             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
1145
1146             /* Load parameters for j particles */
1147             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1148                                                               charge+jnrC+0,charge+jnrD+0);
1149             vdwjidx0A        = 2*vdwtype[jnrA+0];
1150             vdwjidx0B        = 2*vdwtype[jnrB+0];
1151             vdwjidx0C        = 2*vdwtype[jnrC+0];
1152             vdwjidx0D        = 2*vdwtype[jnrD+0];
1153
1154             fjx0             = _mm_setzero_ps();
1155             fjy0             = _mm_setzero_ps();
1156             fjz0             = _mm_setzero_ps();
1157
1158             /**************************
1159              * CALCULATE INTERACTIONS *
1160              **************************/
1161
1162             r00              = _mm_mul_ps(rsq00,rinv00);
1163             r00              = _mm_andnot_ps(dummy_mask,r00);
1164
1165             /* Compute parameters for interactions between i and j atoms */
1166             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1167                                          vdwparam+vdwioffset0+vdwjidx0B,
1168                                          vdwparam+vdwioffset0+vdwjidx0C,
1169                                          vdwparam+vdwioffset0+vdwjidx0D,
1170                                          &c6_00,&c12_00);
1171
1172             /* Calculate table index by multiplying r with table scale and truncate to integer */
1173             rt               = _mm_mul_ps(r00,vftabscale);
1174             vfitab           = _mm_cvttps_epi32(rt);
1175 #ifdef __XOP__
1176             vfeps            = _mm_frcz_ps(rt);
1177 #else
1178             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1179 #endif
1180             twovfeps         = _mm_add_ps(vfeps,vfeps);
1181             vfitab           = _mm_slli_epi32(vfitab,3);
1182
1183             /* CUBIC SPLINE TABLE DISPERSION */
1184             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1185             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1186             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1187             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1188             _MM_TRANSPOSE4_PS(Y,F,G,H);
1189             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1190             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1191             fvdw6            = _mm_mul_ps(c6_00,FF);
1192
1193             /* CUBIC SPLINE TABLE REPULSION */
1194             vfitab           = _mm_add_epi32(vfitab,ifour);
1195             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1196             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1197             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1198             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1199             _MM_TRANSPOSE4_PS(Y,F,G,H);
1200             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1201             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1202             fvdw12           = _mm_mul_ps(c12_00,FF);
1203             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1204
1205             fscal            = fvdw;
1206
1207             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1208
1209              /* Update vectorial force */
1210             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1211             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1212             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1213
1214             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1215             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1216             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1217
1218             /**************************
1219              * CALCULATE INTERACTIONS *
1220              **************************/
1221
1222             r10              = _mm_mul_ps(rsq10,rinv10);
1223             r10              = _mm_andnot_ps(dummy_mask,r10);
1224
1225             /* Compute parameters for interactions between i and j atoms */
1226             qq10             = _mm_mul_ps(iq1,jq0);
1227
1228             /* EWALD ELECTROSTATICS */
1229
1230             /* Analytical PME correction */
1231             zeta2            = _mm_mul_ps(beta2,rsq10);
1232             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1233             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1234             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1235             felec            = _mm_mul_ps(qq10,felec);
1236
1237             fscal            = felec;
1238
1239             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1240
1241              /* Update vectorial force */
1242             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1243             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1244             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1245
1246             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1247             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1248             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1249
1250             /**************************
1251              * CALCULATE INTERACTIONS *
1252              **************************/
1253
1254             r20              = _mm_mul_ps(rsq20,rinv20);
1255             r20              = _mm_andnot_ps(dummy_mask,r20);
1256
1257             /* Compute parameters for interactions between i and j atoms */
1258             qq20             = _mm_mul_ps(iq2,jq0);
1259
1260             /* EWALD ELECTROSTATICS */
1261
1262             /* Analytical PME correction */
1263             zeta2            = _mm_mul_ps(beta2,rsq20);
1264             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1265             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1266             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1267             felec            = _mm_mul_ps(qq20,felec);
1268
1269             fscal            = felec;
1270
1271             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1272
1273              /* Update vectorial force */
1274             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1275             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1276             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1277
1278             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1279             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1280             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1281
1282             /**************************
1283              * CALCULATE INTERACTIONS *
1284              **************************/
1285
1286             r30              = _mm_mul_ps(rsq30,rinv30);
1287             r30              = _mm_andnot_ps(dummy_mask,r30);
1288
1289             /* Compute parameters for interactions between i and j atoms */
1290             qq30             = _mm_mul_ps(iq3,jq0);
1291
1292             /* EWALD ELECTROSTATICS */
1293
1294             /* Analytical PME correction */
1295             zeta2            = _mm_mul_ps(beta2,rsq30);
1296             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
1297             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1298             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1299             felec            = _mm_mul_ps(qq30,felec);
1300
1301             fscal            = felec;
1302
1303             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1304
1305              /* Update vectorial force */
1306             fix3             = _mm_macc_ps(dx30,fscal,fix3);
1307             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
1308             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
1309
1310             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
1311             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
1312             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
1313
1314             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1315             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1316             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1317             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1318
1319             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1320
1321             /* Inner loop uses 139 flops */
1322         }
1323
1324         /* End of innermost loop */
1325
1326         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1327                                               f+i_coord_offset,fshift+i_shift_offset);
1328
1329         /* Increment number of inner iterations */
1330         inneriter                  += j_index_end - j_index_start;
1331
1332         /* Outer loop uses 24 flops */
1333     }
1334
1335     /* Increment number of outer iterations */
1336     outeriter        += nri;
1337
1338     /* Update outer/inner flops */
1339
1340     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*139);
1341 }