src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_avx_128_fma_single
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  98     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  99     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 100     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 101     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 102     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 103     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 104     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 105     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 106     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 107     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 108     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 109     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 110     real             *charge;
 111     int              nvdwtype;
 112     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 113     int              *vdwtype;
 114     real             *vdwparam;
 115     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 116     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 117     __m128           dummy_mask,cutoff_mask;
 118     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 119     __m128           one     = _mm_set1_ps(1.0);
 120     __m128           two     = _mm_set1_ps(2.0);
 121     x                = xx[0];
 122     f                = ff[0];
 123
 124     nri              = nlist->nri;
 125     iinr             = nlist->iinr;
 126     jindex           = nlist->jindex;
 127     jjnr             = nlist->jjnr;
 128     shiftidx         = nlist->shift;
 129     gid              = nlist->gid;
 130     shiftvec         = fr->shift_vec[0];
 131     fshift           = fr->fshift[0];
 132     facel            = _mm_set1_ps(fr->epsfac);
 133     charge           = mdatoms->chargeA;
 134     nvdwtype         = fr->ntype;
 135     vdwparam         = fr->nbfp;
 136     vdwtype          = mdatoms->typeA;
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 141     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 142     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     jq1              = _mm_set1_ps(charge[inr+1]);
 146     jq2              = _mm_set1_ps(charge[inr+2]);
 147     jq3              = _mm_set1_ps(charge[inr+3]);
 148     vdwjidx0A        = 2*vdwtype[inr+0];
 149     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 150     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 151     qq11             = _mm_mul_ps(iq1,jq1);
 152     qq12             = _mm_mul_ps(iq1,jq2);
 153     qq13             = _mm_mul_ps(iq1,jq3);
 154     qq21             = _mm_mul_ps(iq2,jq1);
 155     qq22             = _mm_mul_ps(iq2,jq2);
 156     qq23             = _mm_mul_ps(iq2,jq3);
 157     qq31             = _mm_mul_ps(iq3,jq1);
 158     qq32             = _mm_mul_ps(iq3,jq2);
 159     qq33             = _mm_mul_ps(iq3,jq3);
 160
 161     /* Avoid stupid compiler warnings */
 162     jnrA = jnrB = jnrC = jnrD = 0;
 163     j_coord_offsetA = 0;
 164     j_coord_offsetB = 0;
 165     j_coord_offsetC = 0;
 166     j_coord_offsetD = 0;
 167
 168     outeriter        = 0;
 169     inneriter        = 0;
 170
 171     for(iidx=0;iidx<4*DIM;iidx++)
 172     {
 173         scratch[iidx] = 0.0;
 174     }
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 193
 194         fix0             = _mm_setzero_ps();
 195         fiy0             = _mm_setzero_ps();
 196         fiz0             = _mm_setzero_ps();
 197         fix1             = _mm_setzero_ps();
 198         fiy1             = _mm_setzero_ps();
 199         fiz1             = _mm_setzero_ps();
 200         fix2             = _mm_setzero_ps();
 201         fiy2             = _mm_setzero_ps();
 202         fiz2             = _mm_setzero_ps();
 203         fix3             = _mm_setzero_ps();
 204         fiy3             = _mm_setzero_ps();
 205         fiz3             = _mm_setzero_ps();
 206
 207         /* Reset potential sums */
 208         velecsum         = _mm_setzero_ps();
 209         vvdwsum          = _mm_setzero_ps();
 210
 211         /* Start inner kernel loop */
 212         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 213         {
 214
 215             /* Get j neighbor index, and coordinate index */
 216             jnrA             = jjnr[jidx];
 217             jnrB             = jjnr[jidx+1];
 218             jnrC             = jjnr[jidx+2];
 219             jnrD             = jjnr[jidx+3];
 220             j_coord_offsetA  = DIM*jnrA;
 221             j_coord_offsetB  = DIM*jnrB;
 222             j_coord_offsetC  = DIM*jnrC;
 223             j_coord_offsetD  = DIM*jnrD;
 224
 225             /* load j atom coordinates */
 226             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 227                                               x+j_coord_offsetC,x+j_coord_offsetD,
 228                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 229                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 230
 231             /* Calculate displacement vector */
 232             dx00             = _mm_sub_ps(ix0,jx0);
 233             dy00             = _mm_sub_ps(iy0,jy0);
 234             dz00             = _mm_sub_ps(iz0,jz0);
 235             dx11             = _mm_sub_ps(ix1,jx1);
 236             dy11             = _mm_sub_ps(iy1,jy1);
 237             dz11             = _mm_sub_ps(iz1,jz1);
 238             dx12             = _mm_sub_ps(ix1,jx2);
 239             dy12             = _mm_sub_ps(iy1,jy2);
 240             dz12             = _mm_sub_ps(iz1,jz2);
 241             dx13             = _mm_sub_ps(ix1,jx3);
 242             dy13             = _mm_sub_ps(iy1,jy3);
 243             dz13             = _mm_sub_ps(iz1,jz3);
 244             dx21             = _mm_sub_ps(ix2,jx1);
 245             dy21             = _mm_sub_ps(iy2,jy1);
 246             dz21             = _mm_sub_ps(iz2,jz1);
 247             dx22             = _mm_sub_ps(ix2,jx2);
 248             dy22             = _mm_sub_ps(iy2,jy2);
 249             dz22             = _mm_sub_ps(iz2,jz2);
 250             dx23             = _mm_sub_ps(ix2,jx3);
 251             dy23             = _mm_sub_ps(iy2,jy3);
 252             dz23             = _mm_sub_ps(iz2,jz3);
 253             dx31             = _mm_sub_ps(ix3,jx1);
 254             dy31             = _mm_sub_ps(iy3,jy1);
 255             dz31             = _mm_sub_ps(iz3,jz1);
 256             dx32             = _mm_sub_ps(ix3,jx2);
 257             dy32             = _mm_sub_ps(iy3,jy2);
 258             dz32             = _mm_sub_ps(iz3,jz2);
 259             dx33             = _mm_sub_ps(ix3,jx3);
 260             dy33             = _mm_sub_ps(iy3,jy3);
 261             dz33             = _mm_sub_ps(iz3,jz3);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 265             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 266             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 267             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 268             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 269             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 270             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 271             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 272             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 273             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 274
 275             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 276             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 277             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 278             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 279             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 280             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 281             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 282             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 283             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 284
 285             rinvsq00         = gmx_mm_inv_ps(rsq00);
 286             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 287             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 288             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 289             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 290             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 291             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 292             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 293             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 294             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 295
 296             fjx0             = _mm_setzero_ps();
 297             fjy0             = _mm_setzero_ps();
 298             fjz0             = _mm_setzero_ps();
 299             fjx1             = _mm_setzero_ps();
 300             fjy1             = _mm_setzero_ps();
 301             fjz1             = _mm_setzero_ps();
 302             fjx2             = _mm_setzero_ps();
 303             fjy2             = _mm_setzero_ps();
 304             fjz2             = _mm_setzero_ps();
 305             fjx3             = _mm_setzero_ps();
 306             fjy3             = _mm_setzero_ps();
 307             fjz3             = _mm_setzero_ps();
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             /* LENNARD-JONES DISPERSION/REPULSION */
 314
 315             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 316             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 317             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 318             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 319             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 320
 321             /* Update potential sum for this i atom from the interaction with this j atom. */
 322             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 323
 324             fscal            = fvdw;
 325
 326              /* Update vectorial force */
 327             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 328             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 329             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 330
 331             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 332             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 333             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             /* COULOMB ELECTROSTATICS */
 340             velec            = _mm_mul_ps(qq11,rinv11);
 341             felec            = _mm_mul_ps(velec,rinvsq11);
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             velecsum         = _mm_add_ps(velecsum,velec);
 345
 346             fscal            = felec;
 347
 348              /* Update vectorial force */
 349             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 350             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 351             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 352
 353             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 354             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 355             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             /* COULOMB ELECTROSTATICS */
 362             velec            = _mm_mul_ps(qq12,rinv12);
 363             felec            = _mm_mul_ps(velec,rinvsq12);
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             velecsum         = _mm_add_ps(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370              /* Update vectorial force */
 371             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 372             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 373             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 374
 375             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 376             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 377             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             /* COULOMB ELECTROSTATICS */
 384             velec            = _mm_mul_ps(qq13,rinv13);
 385             felec            = _mm_mul_ps(velec,rinvsq13);
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velecsum         = _mm_add_ps(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392              /* Update vectorial force */
 393             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 394             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 395             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 396
 397             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 398             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 399             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             /* COULOMB ELECTROSTATICS */
 406             velec            = _mm_mul_ps(qq21,rinv21);
 407             felec            = _mm_mul_ps(velec,rinvsq21);
 408
 409             /* Update potential sum for this i atom from the interaction with this j atom. */
 410             velecsum         = _mm_add_ps(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414              /* Update vectorial force */
 415             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 416             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 417             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 418
 419             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 420             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 421             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             /* COULOMB ELECTROSTATICS */
 428             velec            = _mm_mul_ps(qq22,rinv22);
 429             felec            = _mm_mul_ps(velec,rinvsq22);
 430
 431             /* Update potential sum for this i atom from the interaction with this j atom. */
 432             velecsum         = _mm_add_ps(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436              /* Update vectorial force */
 437             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 438             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 439             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 440
 441             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 442             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 443             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* COULOMB ELECTROSTATICS */
 450             velec            = _mm_mul_ps(qq23,rinv23);
 451             felec            = _mm_mul_ps(velec,rinvsq23);
 452
 453             /* Update potential sum for this i atom from the interaction with this j atom. */
 454             velecsum         = _mm_add_ps(velecsum,velec);
 455
 456             fscal            = felec;
 457
 458              /* Update vectorial force */
 459             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 460             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 461             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 462
 463             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 464             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 465             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 466
 467             /**************************
 468              * CALCULATE INTERACTIONS *
 469              **************************/
 470
 471             /* COULOMB ELECTROSTATICS */
 472             velec            = _mm_mul_ps(qq31,rinv31);
 473             felec            = _mm_mul_ps(velec,rinvsq31);
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velecsum         = _mm_add_ps(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480              /* Update vectorial force */
 481             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 482             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 483             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 484
 485             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 486             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 487             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 488
 489             /**************************
 490              * CALCULATE INTERACTIONS *
 491              **************************/
 492
 493             /* COULOMB ELECTROSTATICS */
 494             velec            = _mm_mul_ps(qq32,rinv32);
 495             felec            = _mm_mul_ps(velec,rinvsq32);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velecsum         = _mm_add_ps(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502              /* Update vectorial force */
 503             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 504             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 505             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 506
 507             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 508             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 509             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 510
 511             /**************************
 512              * CALCULATE INTERACTIONS *
 513              **************************/
 514
 515             /* COULOMB ELECTROSTATICS */
 516             velec            = _mm_mul_ps(qq33,rinv33);
 517             felec            = _mm_mul_ps(velec,rinvsq33);
 518
 519             /* Update potential sum for this i atom from the interaction with this j atom. */
 520             velecsum         = _mm_add_ps(velecsum,velec);
 521
 522             fscal            = felec;
 523
 524              /* Update vectorial force */
 525             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 526             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 527             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 528
 529             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 530             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 531             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 532
 533             fjptrA             = f+j_coord_offsetA;
 534             fjptrB             = f+j_coord_offsetB;
 535             fjptrC             = f+j_coord_offsetC;
 536             fjptrD             = f+j_coord_offsetD;
 537
 538             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 539                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 540                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 541
 542             /* Inner loop uses 317 flops */
 543         }
 544
 545         if(jidx<j_index_end)
 546         {
 547
 548             /* Get j neighbor index, and coordinate index */
 549             jnrlistA         = jjnr[jidx];
 550             jnrlistB         = jjnr[jidx+1];
 551             jnrlistC         = jjnr[jidx+2];
 552             jnrlistD         = jjnr[jidx+3];
 553             /* Sign of each element will be negative for non-real atoms.
 554              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 555              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 556              */
 557             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 558             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 559             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 560             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 561             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 562             j_coord_offsetA  = DIM*jnrA;
 563             j_coord_offsetB  = DIM*jnrB;
 564             j_coord_offsetC  = DIM*jnrC;
 565             j_coord_offsetD  = DIM*jnrD;
 566
 567             /* load j atom coordinates */
 568             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 569                                               x+j_coord_offsetC,x+j_coord_offsetD,
 570                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 571                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 572
 573             /* Calculate displacement vector */
 574             dx00             = _mm_sub_ps(ix0,jx0);
 575             dy00             = _mm_sub_ps(iy0,jy0);
 576             dz00             = _mm_sub_ps(iz0,jz0);
 577             dx11             = _mm_sub_ps(ix1,jx1);
 578             dy11             = _mm_sub_ps(iy1,jy1);
 579             dz11             = _mm_sub_ps(iz1,jz1);
 580             dx12             = _mm_sub_ps(ix1,jx2);
 581             dy12             = _mm_sub_ps(iy1,jy2);
 582             dz12             = _mm_sub_ps(iz1,jz2);
 583             dx13             = _mm_sub_ps(ix1,jx3);
 584             dy13             = _mm_sub_ps(iy1,jy3);
 585             dz13             = _mm_sub_ps(iz1,jz3);
 586             dx21             = _mm_sub_ps(ix2,jx1);
 587             dy21             = _mm_sub_ps(iy2,jy1);
 588             dz21             = _mm_sub_ps(iz2,jz1);
 589             dx22             = _mm_sub_ps(ix2,jx2);
 590             dy22             = _mm_sub_ps(iy2,jy2);
 591             dz22             = _mm_sub_ps(iz2,jz2);
 592             dx23             = _mm_sub_ps(ix2,jx3);
 593             dy23             = _mm_sub_ps(iy2,jy3);
 594             dz23             = _mm_sub_ps(iz2,jz3);
 595             dx31             = _mm_sub_ps(ix3,jx1);
 596             dy31             = _mm_sub_ps(iy3,jy1);
 597             dz31             = _mm_sub_ps(iz3,jz1);
 598             dx32             = _mm_sub_ps(ix3,jx2);
 599             dy32             = _mm_sub_ps(iy3,jy2);
 600             dz32             = _mm_sub_ps(iz3,jz2);
 601             dx33             = _mm_sub_ps(ix3,jx3);
 602             dy33             = _mm_sub_ps(iy3,jy3);
 603             dz33             = _mm_sub_ps(iz3,jz3);
 604
 605             /* Calculate squared distance and things based on it */
 606             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 607             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 608             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 609             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 610             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 611             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 612             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 613             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 614             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 615             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 616
 617             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 618             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 619             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 620             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 621             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 622             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 623             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 624             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 625             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 626
 627             rinvsq00         = gmx_mm_inv_ps(rsq00);
 628             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 629             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 630             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 631             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 632             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 633             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 634             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 635             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 636             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 637
 638             fjx0             = _mm_setzero_ps();
 639             fjy0             = _mm_setzero_ps();
 640             fjz0             = _mm_setzero_ps();
 641             fjx1             = _mm_setzero_ps();
 642             fjy1             = _mm_setzero_ps();
 643             fjz1             = _mm_setzero_ps();
 644             fjx2             = _mm_setzero_ps();
 645             fjy2             = _mm_setzero_ps();
 646             fjz2             = _mm_setzero_ps();
 647             fjx3             = _mm_setzero_ps();
 648             fjy3             = _mm_setzero_ps();
 649             fjz3             = _mm_setzero_ps();
 650
 651             /**************************
 652              * CALCULATE INTERACTIONS *
 653              **************************/
 654
 655             /* LENNARD-JONES DISPERSION/REPULSION */
 656
 657             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 658             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 659             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 660             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 661             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 662
 663             /* Update potential sum for this i atom from the interaction with this j atom. */
 664             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 665             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 666
 667             fscal            = fvdw;
 668
 669             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 670
 671              /* Update vectorial force */
 672             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 673             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 674             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 675
 676             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 677             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 678             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 679
 680             /**************************
 681              * CALCULATE INTERACTIONS *
 682              **************************/
 683
 684             /* COULOMB ELECTROSTATICS */
 685             velec            = _mm_mul_ps(qq11,rinv11);
 686             felec            = _mm_mul_ps(velec,rinvsq11);
 687
 688             /* Update potential sum for this i atom from the interaction with this j atom. */
 689             velec            = _mm_andnot_ps(dummy_mask,velec);
 690             velecsum         = _mm_add_ps(velecsum,velec);
 691
 692             fscal            = felec;
 693
 694             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 695
 696              /* Update vectorial force */
 697             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 698             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 699             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 700
 701             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 702             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 703             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 704
 705             /**************************
 706              * CALCULATE INTERACTIONS *
 707              **************************/
 708
 709             /* COULOMB ELECTROSTATICS */
 710             velec            = _mm_mul_ps(qq12,rinv12);
 711             felec            = _mm_mul_ps(velec,rinvsq12);
 712
 713             /* Update potential sum for this i atom from the interaction with this j atom. */
 714             velec            = _mm_andnot_ps(dummy_mask,velec);
 715             velecsum         = _mm_add_ps(velecsum,velec);
 716
 717             fscal            = felec;
 718
 719             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 720
 721              /* Update vectorial force */
 722             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 723             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 724             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 725
 726             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 727             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 728             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 729
 730             /**************************
 731              * CALCULATE INTERACTIONS *
 732              **************************/
 733
 734             /* COULOMB ELECTROSTATICS */
 735             velec            = _mm_mul_ps(qq13,rinv13);
 736             felec            = _mm_mul_ps(velec,rinvsq13);
 737
 738             /* Update potential sum for this i atom from the interaction with this j atom. */
 739             velec            = _mm_andnot_ps(dummy_mask,velec);
 740             velecsum         = _mm_add_ps(velecsum,velec);
 741
 742             fscal            = felec;
 743
 744             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 745
 746              /* Update vectorial force */
 747             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 748             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 749             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 750
 751             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 752             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 753             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             /* COULOMB ELECTROSTATICS */
 760             velec            = _mm_mul_ps(qq21,rinv21);
 761             felec            = _mm_mul_ps(velec,rinvsq21);
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _mm_andnot_ps(dummy_mask,velec);
 765             velecsum         = _mm_add_ps(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 770
 771              /* Update vectorial force */
 772             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 773             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 774             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 775
 776             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 777             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 778             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 779
 780             /**************************
 781              * CALCULATE INTERACTIONS *
 782              **************************/
 783
 784             /* COULOMB ELECTROSTATICS */
 785             velec            = _mm_mul_ps(qq22,rinv22);
 786             felec            = _mm_mul_ps(velec,rinvsq22);
 787
 788             /* Update potential sum for this i atom from the interaction with this j atom. */
 789             velec            = _mm_andnot_ps(dummy_mask,velec);
 790             velecsum         = _mm_add_ps(velecsum,velec);
 791
 792             fscal            = felec;
 793
 794             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 795
 796              /* Update vectorial force */
 797             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 798             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 799             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 800
 801             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 802             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 803             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 804
 805             /**************************
 806              * CALCULATE INTERACTIONS *
 807              **************************/
 808
 809             /* COULOMB ELECTROSTATICS */
 810             velec            = _mm_mul_ps(qq23,rinv23);
 811             felec            = _mm_mul_ps(velec,rinvsq23);
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _mm_andnot_ps(dummy_mask,velec);
 815             velecsum         = _mm_add_ps(velecsum,velec);
 816
 817             fscal            = felec;
 818
 819             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 820
 821              /* Update vectorial force */
 822             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 823             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 824             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 825
 826             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 827             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 828             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             /* COULOMB ELECTROSTATICS */
 835             velec            = _mm_mul_ps(qq31,rinv31);
 836             felec            = _mm_mul_ps(velec,rinvsq31);
 837
 838             /* Update potential sum for this i atom from the interaction with this j atom. */
 839             velec            = _mm_andnot_ps(dummy_mask,velec);
 840             velecsum         = _mm_add_ps(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 845
 846              /* Update vectorial force */
 847             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 848             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 849             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 850
 851             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 852             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 853             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 854
 855             /**************************
 856              * CALCULATE INTERACTIONS *
 857              **************************/
 858
 859             /* COULOMB ELECTROSTATICS */
 860             velec            = _mm_mul_ps(qq32,rinv32);
 861             felec            = _mm_mul_ps(velec,rinvsq32);
 862
 863             /* Update potential sum for this i atom from the interaction with this j atom. */
 864             velec            = _mm_andnot_ps(dummy_mask,velec);
 865             velecsum         = _mm_add_ps(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 870
 871              /* Update vectorial force */
 872             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 873             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 874             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 875
 876             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 877             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 878             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 879
 880             /**************************
 881              * CALCULATE INTERACTIONS *
 882              **************************/
 883
 884             /* COULOMB ELECTROSTATICS */
 885             velec            = _mm_mul_ps(qq33,rinv33);
 886             felec            = _mm_mul_ps(velec,rinvsq33);
 887
 888             /* Update potential sum for this i atom from the interaction with this j atom. */
 889             velec            = _mm_andnot_ps(dummy_mask,velec);
 890             velecsum         = _mm_add_ps(velecsum,velec);
 891
 892             fscal            = felec;
 893
 894             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 895
 896              /* Update vectorial force */
 897             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 898             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 899             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 900
 901             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 902             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 903             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 904
 905             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 906             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 907             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 908             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 909
 910             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 911                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 912                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 913
 914             /* Inner loop uses 317 flops */
 915         }
 916
 917         /* End of innermost loop */
 918
 919         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 920                                               f+i_coord_offset,fshift+i_shift_offset);
 921
 922         ggid                        = gid[iidx];
 923         /* Update potential energies */
 924         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 925         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 926
 927         /* Increment number of inner iterations */
 928         inneriter                  += j_index_end - j_index_start;
 929
 930         /* Outer loop uses 26 flops */
 931     }
 932
 933     /* Increment number of outer iterations */
 934     outeriter        += nri;
 935
 936     /* Update outer/inner flops */
 937
 938     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*317);
 939 }
 940 /*
 941  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_single
 942  * Electrostatics interaction: Coulomb
 943  * VdW interaction:            LennardJones
 944  * Geometry:                   Water4-Water4
 945  * Calculate force/pot:        Force
 946  */
 947 void
 948 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_single
 949                     (t_nblist                    * gmx_restrict       nlist,
 950                      rvec                        * gmx_restrict          xx,
 951                      rvec                        * gmx_restrict          ff,
 952                      t_forcerec                  * gmx_restrict          fr,
 953                      t_mdatoms                   * gmx_restrict     mdatoms,
 954                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 955                      t_nrnb                      * gmx_restrict        nrnb)
 956 {
 957     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 958      * just 0 for non-waters.
 959      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 960      * jnr indices corresponding to data put in the four positions in the SIMD register.
 961      */
 962     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 963     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 964     int              jnrA,jnrB,jnrC,jnrD;
 965     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 966     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 967     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 968     real             rcutoff_scalar;
 969     real             *shiftvec,*fshift,*x,*f;
 970     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 971     real             scratch[4*DIM];
 972     __m128           fscal,rcutoff,rcutoff2,jidxall;
 973     int              vdwioffset0;
 974     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 975     int              vdwioffset1;
 976     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 977     int              vdwioffset2;
 978     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 979     int              vdwioffset3;
 980     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 981     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 982     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 983     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 984     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 985     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 986     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 987     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 988     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 989     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 990     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 991     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 992     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 993     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 994     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 995     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 996     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 997     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 998     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 999     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1000     real             *charge;
1001     int              nvdwtype;
1002     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1003     int              *vdwtype;
1004     real             *vdwparam;
1005     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1006     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1007     __m128           dummy_mask,cutoff_mask;
1008     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1009     __m128           one     = _mm_set1_ps(1.0);
1010     __m128           two     = _mm_set1_ps(2.0);
1011     x                = xx[0];
1012     f                = ff[0];
1013
1014     nri              = nlist->nri;
1015     iinr             = nlist->iinr;
1016     jindex           = nlist->jindex;
1017     jjnr             = nlist->jjnr;
1018     shiftidx         = nlist->shift;
1019     gid              = nlist->gid;
1020     shiftvec         = fr->shift_vec[0];
1021     fshift           = fr->fshift[0];
1022     facel            = _mm_set1_ps(fr->epsfac);
1023     charge           = mdatoms->chargeA;
1024     nvdwtype         = fr->ntype;
1025     vdwparam         = fr->nbfp;
1026     vdwtype          = mdatoms->typeA;
1027
1028     /* Setup water-specific parameters */
1029     inr              = nlist->iinr[0];
1030     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1031     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1032     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1033     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1034
1035     jq1              = _mm_set1_ps(charge[inr+1]);
1036     jq2              = _mm_set1_ps(charge[inr+2]);
1037     jq3              = _mm_set1_ps(charge[inr+3]);
1038     vdwjidx0A        = 2*vdwtype[inr+0];
1039     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1040     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1041     qq11             = _mm_mul_ps(iq1,jq1);
1042     qq12             = _mm_mul_ps(iq1,jq2);
1043     qq13             = _mm_mul_ps(iq1,jq3);
1044     qq21             = _mm_mul_ps(iq2,jq1);
1045     qq22             = _mm_mul_ps(iq2,jq2);
1046     qq23             = _mm_mul_ps(iq2,jq3);
1047     qq31             = _mm_mul_ps(iq3,jq1);
1048     qq32             = _mm_mul_ps(iq3,jq2);
1049     qq33             = _mm_mul_ps(iq3,jq3);
1050
1051     /* Avoid stupid compiler warnings */
1052     jnrA = jnrB = jnrC = jnrD = 0;
1053     j_coord_offsetA = 0;
1054     j_coord_offsetB = 0;
1055     j_coord_offsetC = 0;
1056     j_coord_offsetD = 0;
1057
1058     outeriter        = 0;
1059     inneriter        = 0;
1060
1061     for(iidx=0;iidx<4*DIM;iidx++)
1062     {
1063         scratch[iidx] = 0.0;
1064     }
1065
1066     /* Start outer loop over neighborlists */
1067     for(iidx=0; iidx<nri; iidx++)
1068     {
1069         /* Load shift vector for this list */
1070         i_shift_offset   = DIM*shiftidx[iidx];
1071
1072         /* Load limits for loop over neighbors */
1073         j_index_start    = jindex[iidx];
1074         j_index_end      = jindex[iidx+1];
1075
1076         /* Get outer coordinate index */
1077         inr              = iinr[iidx];
1078         i_coord_offset   = DIM*inr;
1079
1080         /* Load i particle coords and add shift vector */
1081         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1082                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1083
1084         fix0             = _mm_setzero_ps();
1085         fiy0             = _mm_setzero_ps();
1086         fiz0             = _mm_setzero_ps();
1087         fix1             = _mm_setzero_ps();
1088         fiy1             = _mm_setzero_ps();
1089         fiz1             = _mm_setzero_ps();
1090         fix2             = _mm_setzero_ps();
1091         fiy2             = _mm_setzero_ps();
1092         fiz2             = _mm_setzero_ps();
1093         fix3             = _mm_setzero_ps();
1094         fiy3             = _mm_setzero_ps();
1095         fiz3             = _mm_setzero_ps();
1096
1097         /* Start inner kernel loop */
1098         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1099         {
1100
1101             /* Get j neighbor index, and coordinate index */
1102             jnrA             = jjnr[jidx];
1103             jnrB             = jjnr[jidx+1];
1104             jnrC             = jjnr[jidx+2];
1105             jnrD             = jjnr[jidx+3];
1106             j_coord_offsetA  = DIM*jnrA;
1107             j_coord_offsetB  = DIM*jnrB;
1108             j_coord_offsetC  = DIM*jnrC;
1109             j_coord_offsetD  = DIM*jnrD;
1110
1111             /* load j atom coordinates */
1112             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1113                                               x+j_coord_offsetC,x+j_coord_offsetD,
1114                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1115                                               &jy2,&jz2,&jx3,&jy3,&jz3);
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             dx11             = _mm_sub_ps(ix1,jx1);
1122             dy11             = _mm_sub_ps(iy1,jy1);
1123             dz11             = _mm_sub_ps(iz1,jz1);
1124             dx12             = _mm_sub_ps(ix1,jx2);
1125             dy12             = _mm_sub_ps(iy1,jy2);
1126             dz12             = _mm_sub_ps(iz1,jz2);
1127             dx13             = _mm_sub_ps(ix1,jx3);
1128             dy13             = _mm_sub_ps(iy1,jy3);
1129             dz13             = _mm_sub_ps(iz1,jz3);
1130             dx21             = _mm_sub_ps(ix2,jx1);
1131             dy21             = _mm_sub_ps(iy2,jy1);
1132             dz21             = _mm_sub_ps(iz2,jz1);
1133             dx22             = _mm_sub_ps(ix2,jx2);
1134             dy22             = _mm_sub_ps(iy2,jy2);
1135             dz22             = _mm_sub_ps(iz2,jz2);
1136             dx23             = _mm_sub_ps(ix2,jx3);
1137             dy23             = _mm_sub_ps(iy2,jy3);
1138             dz23             = _mm_sub_ps(iz2,jz3);
1139             dx31             = _mm_sub_ps(ix3,jx1);
1140             dy31             = _mm_sub_ps(iy3,jy1);
1141             dz31             = _mm_sub_ps(iz3,jz1);
1142             dx32             = _mm_sub_ps(ix3,jx2);
1143             dy32             = _mm_sub_ps(iy3,jy2);
1144             dz32             = _mm_sub_ps(iz3,jz2);
1145             dx33             = _mm_sub_ps(ix3,jx3);
1146             dy33             = _mm_sub_ps(iy3,jy3);
1147             dz33             = _mm_sub_ps(iz3,jz3);
1148
1149             /* Calculate squared distance and things based on it */
1150             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1151             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1152             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1153             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1154             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1155             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1156             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1157             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1158             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1159             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1160
1161             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1162             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1163             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1164             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1165             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1166             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1167             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1168             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1169             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1170
1171             rinvsq00         = gmx_mm_inv_ps(rsq00);
1172             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1173             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1174             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1175             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1176             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1177             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1178             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1179             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1180             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1181
1182             fjx0             = _mm_setzero_ps();
1183             fjy0             = _mm_setzero_ps();
1184             fjz0             = _mm_setzero_ps();
1185             fjx1             = _mm_setzero_ps();
1186             fjy1             = _mm_setzero_ps();
1187             fjz1             = _mm_setzero_ps();
1188             fjx2             = _mm_setzero_ps();
1189             fjy2             = _mm_setzero_ps();
1190             fjz2             = _mm_setzero_ps();
1191             fjx3             = _mm_setzero_ps();
1192             fjy3             = _mm_setzero_ps();
1193             fjz3             = _mm_setzero_ps();
1194
1195             /**************************
1196              * CALCULATE INTERACTIONS *
1197              **************************/
1198
1199             /* LENNARD-JONES DISPERSION/REPULSION */
1200
1201             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1202             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1203
1204             fscal            = fvdw;
1205
1206              /* Update vectorial force */
1207             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1208             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1209             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1210
1211             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1212             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1213             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1214
1215             /**************************
1216              * CALCULATE INTERACTIONS *
1217              **************************/
1218
1219             /* COULOMB ELECTROSTATICS */
1220             velec            = _mm_mul_ps(qq11,rinv11);
1221             felec            = _mm_mul_ps(velec,rinvsq11);
1222
1223             fscal            = felec;
1224
1225              /* Update vectorial force */
1226             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1227             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1228             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1229
1230             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1231             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1232             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1233
1234             /**************************
1235              * CALCULATE INTERACTIONS *
1236              **************************/
1237
1238             /* COULOMB ELECTROSTATICS */
1239             velec            = _mm_mul_ps(qq12,rinv12);
1240             felec            = _mm_mul_ps(velec,rinvsq12);
1241
1242             fscal            = felec;
1243
1244              /* Update vectorial force */
1245             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1246             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1247             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1248
1249             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1250             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1251             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1252
1253             /**************************
1254              * CALCULATE INTERACTIONS *
1255              **************************/
1256
1257             /* COULOMB ELECTROSTATICS */
1258             velec            = _mm_mul_ps(qq13,rinv13);
1259             felec            = _mm_mul_ps(velec,rinvsq13);
1260
1261             fscal            = felec;
1262
1263              /* Update vectorial force */
1264             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1265             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1266             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1267
1268             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1269             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1270             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1271
1272             /**************************
1273              * CALCULATE INTERACTIONS *
1274              **************************/
1275
1276             /* COULOMB ELECTROSTATICS */
1277             velec            = _mm_mul_ps(qq21,rinv21);
1278             felec            = _mm_mul_ps(velec,rinvsq21);
1279
1280             fscal            = felec;
1281
1282              /* Update vectorial force */
1283             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1284             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1285             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1286
1287             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1288             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1289             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1290
1291             /**************************
1292              * CALCULATE INTERACTIONS *
1293              **************************/
1294
1295             /* COULOMB ELECTROSTATICS */
1296             velec            = _mm_mul_ps(qq22,rinv22);
1297             felec            = _mm_mul_ps(velec,rinvsq22);
1298
1299             fscal            = felec;
1300
1301              /* Update vectorial force */
1302             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1303             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1304             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1305
1306             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1307             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1308             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1309
1310             /**************************
1311              * CALCULATE INTERACTIONS *
1312              **************************/
1313
1314             /* COULOMB ELECTROSTATICS */
1315             velec            = _mm_mul_ps(qq23,rinv23);
1316             felec            = _mm_mul_ps(velec,rinvsq23);
1317
1318             fscal            = felec;
1319
1320              /* Update vectorial force */
1321             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1322             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1323             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1324
1325             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1326             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1327             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1328
1329             /**************************
1330              * CALCULATE INTERACTIONS *
1331              **************************/
1332
1333             /* COULOMB ELECTROSTATICS */
1334             velec            = _mm_mul_ps(qq31,rinv31);
1335             felec            = _mm_mul_ps(velec,rinvsq31);
1336
1337             fscal            = felec;
1338
1339              /* Update vectorial force */
1340             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1341             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1342             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1343
1344             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1345             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1346             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* COULOMB ELECTROSTATICS */
1353             velec            = _mm_mul_ps(qq32,rinv32);
1354             felec            = _mm_mul_ps(velec,rinvsq32);
1355
1356             fscal            = felec;
1357
1358              /* Update vectorial force */
1359             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1360             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1361             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1362
1363             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1364             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1365             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1366
1367             /**************************
1368              * CALCULATE INTERACTIONS *
1369              **************************/
1370
1371             /* COULOMB ELECTROSTATICS */
1372             velec            = _mm_mul_ps(qq33,rinv33);
1373             felec            = _mm_mul_ps(velec,rinvsq33);
1374
1375             fscal            = felec;
1376
1377              /* Update vectorial force */
1378             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1379             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1380             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1381
1382             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1383             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1384             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1385
1386             fjptrA             = f+j_coord_offsetA;
1387             fjptrB             = f+j_coord_offsetB;
1388             fjptrC             = f+j_coord_offsetC;
1389             fjptrD             = f+j_coord_offsetD;
1390
1391             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1392                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1393                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1394
1395             /* Inner loop uses 303 flops */
1396         }
1397
1398         if(jidx<j_index_end)
1399         {
1400
1401             /* Get j neighbor index, and coordinate index */
1402             jnrlistA         = jjnr[jidx];
1403             jnrlistB         = jjnr[jidx+1];
1404             jnrlistC         = jjnr[jidx+2];
1405             jnrlistD         = jjnr[jidx+3];
1406             /* Sign of each element will be negative for non-real atoms.
1407              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1408              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1409              */
1410             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1411             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1412             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1413             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1414             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1415             j_coord_offsetA  = DIM*jnrA;
1416             j_coord_offsetB  = DIM*jnrB;
1417             j_coord_offsetC  = DIM*jnrC;
1418             j_coord_offsetD  = DIM*jnrD;
1419
1420             /* load j atom coordinates */
1421             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1422                                               x+j_coord_offsetC,x+j_coord_offsetD,
1423                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1424                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1425
1426             /* Calculate displacement vector */
1427             dx00             = _mm_sub_ps(ix0,jx0);
1428             dy00             = _mm_sub_ps(iy0,jy0);
1429             dz00             = _mm_sub_ps(iz0,jz0);
1430             dx11             = _mm_sub_ps(ix1,jx1);
1431             dy11             = _mm_sub_ps(iy1,jy1);
1432             dz11             = _mm_sub_ps(iz1,jz1);
1433             dx12             = _mm_sub_ps(ix1,jx2);
1434             dy12             = _mm_sub_ps(iy1,jy2);
1435             dz12             = _mm_sub_ps(iz1,jz2);
1436             dx13             = _mm_sub_ps(ix1,jx3);
1437             dy13             = _mm_sub_ps(iy1,jy3);
1438             dz13             = _mm_sub_ps(iz1,jz3);
1439             dx21             = _mm_sub_ps(ix2,jx1);
1440             dy21             = _mm_sub_ps(iy2,jy1);
1441             dz21             = _mm_sub_ps(iz2,jz1);
1442             dx22             = _mm_sub_ps(ix2,jx2);
1443             dy22             = _mm_sub_ps(iy2,jy2);
1444             dz22             = _mm_sub_ps(iz2,jz2);
1445             dx23             = _mm_sub_ps(ix2,jx3);
1446             dy23             = _mm_sub_ps(iy2,jy3);
1447             dz23             = _mm_sub_ps(iz2,jz3);
1448             dx31             = _mm_sub_ps(ix3,jx1);
1449             dy31             = _mm_sub_ps(iy3,jy1);
1450             dz31             = _mm_sub_ps(iz3,jz1);
1451             dx32             = _mm_sub_ps(ix3,jx2);
1452             dy32             = _mm_sub_ps(iy3,jy2);
1453             dz32             = _mm_sub_ps(iz3,jz2);
1454             dx33             = _mm_sub_ps(ix3,jx3);
1455             dy33             = _mm_sub_ps(iy3,jy3);
1456             dz33             = _mm_sub_ps(iz3,jz3);
1457
1458             /* Calculate squared distance and things based on it */
1459             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1460             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1461             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1462             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1463             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1464             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1465             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1466             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1467             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1468             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1469
1470             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1471             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1472             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1473             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1474             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1475             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1476             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1477             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1478             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1479
1480             rinvsq00         = gmx_mm_inv_ps(rsq00);
1481             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1482             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1483             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1484             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1485             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1486             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1487             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1488             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1489             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1490
1491             fjx0             = _mm_setzero_ps();
1492             fjy0             = _mm_setzero_ps();
1493             fjz0             = _mm_setzero_ps();
1494             fjx1             = _mm_setzero_ps();
1495             fjy1             = _mm_setzero_ps();
1496             fjz1             = _mm_setzero_ps();
1497             fjx2             = _mm_setzero_ps();
1498             fjy2             = _mm_setzero_ps();
1499             fjz2             = _mm_setzero_ps();
1500             fjx3             = _mm_setzero_ps();
1501             fjy3             = _mm_setzero_ps();
1502             fjz3             = _mm_setzero_ps();
1503
1504             /**************************
1505              * CALCULATE INTERACTIONS *
1506              **************************/
1507
1508             /* LENNARD-JONES DISPERSION/REPULSION */
1509
1510             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1511             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1512
1513             fscal            = fvdw;
1514
1515             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1516
1517              /* Update vectorial force */
1518             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1519             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1520             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1521
1522             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1523             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1524             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1525
1526             /**************************
1527              * CALCULATE INTERACTIONS *
1528              **************************/
1529
1530             /* COULOMB ELECTROSTATICS */
1531             velec            = _mm_mul_ps(qq11,rinv11);
1532             felec            = _mm_mul_ps(velec,rinvsq11);
1533
1534             fscal            = felec;
1535
1536             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1537
1538              /* Update vectorial force */
1539             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1540             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1541             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1542
1543             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1544             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1545             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1546
1547             /**************************
1548              * CALCULATE INTERACTIONS *
1549              **************************/
1550
1551             /* COULOMB ELECTROSTATICS */
1552             velec            = _mm_mul_ps(qq12,rinv12);
1553             felec            = _mm_mul_ps(velec,rinvsq12);
1554
1555             fscal            = felec;
1556
1557             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1558
1559              /* Update vectorial force */
1560             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1561             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1562             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1563
1564             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1565             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1566             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1567
1568             /**************************
1569              * CALCULATE INTERACTIONS *
1570              **************************/
1571
1572             /* COULOMB ELECTROSTATICS */
1573             velec            = _mm_mul_ps(qq13,rinv13);
1574             felec            = _mm_mul_ps(velec,rinvsq13);
1575
1576             fscal            = felec;
1577
1578             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1579
1580              /* Update vectorial force */
1581             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1582             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1583             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1584
1585             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1586             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1587             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             /* COULOMB ELECTROSTATICS */
1594             velec            = _mm_mul_ps(qq21,rinv21);
1595             felec            = _mm_mul_ps(velec,rinvsq21);
1596
1597             fscal            = felec;
1598
1599             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1600
1601              /* Update vectorial force */
1602             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1603             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1604             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1605
1606             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1607             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1608             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1609
1610             /**************************
1611              * CALCULATE INTERACTIONS *
1612              **************************/
1613
1614             /* COULOMB ELECTROSTATICS */
1615             velec            = _mm_mul_ps(qq22,rinv22);
1616             felec            = _mm_mul_ps(velec,rinvsq22);
1617
1618             fscal            = felec;
1619
1620             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1621
1622              /* Update vectorial force */
1623             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1624             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1625             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1626
1627             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1628             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1629             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1630
1631             /**************************
1632              * CALCULATE INTERACTIONS *
1633              **************************/
1634
1635             /* COULOMB ELECTROSTATICS */
1636             velec            = _mm_mul_ps(qq23,rinv23);
1637             felec            = _mm_mul_ps(velec,rinvsq23);
1638
1639             fscal            = felec;
1640
1641             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1642
1643              /* Update vectorial force */
1644             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1645             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1646             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1647
1648             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1649             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1650             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1651
1652             /**************************
1653              * CALCULATE INTERACTIONS *
1654              **************************/
1655
1656             /* COULOMB ELECTROSTATICS */
1657             velec            = _mm_mul_ps(qq31,rinv31);
1658             felec            = _mm_mul_ps(velec,rinvsq31);
1659
1660             fscal            = felec;
1661
1662             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1663
1664              /* Update vectorial force */
1665             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1666             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1667             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1668
1669             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1670             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1671             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1672
1673             /**************************
1674              * CALCULATE INTERACTIONS *
1675              **************************/
1676
1677             /* COULOMB ELECTROSTATICS */
1678             velec            = _mm_mul_ps(qq32,rinv32);
1679             felec            = _mm_mul_ps(velec,rinvsq32);
1680
1681             fscal            = felec;
1682
1683             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1684
1685              /* Update vectorial force */
1686             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1687             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1688             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1689
1690             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1691             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1692             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1693
1694             /**************************
1695              * CALCULATE INTERACTIONS *
1696              **************************/
1697
1698             /* COULOMB ELECTROSTATICS */
1699             velec            = _mm_mul_ps(qq33,rinv33);
1700             felec            = _mm_mul_ps(velec,rinvsq33);
1701
1702             fscal            = felec;
1703
1704             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1705
1706              /* Update vectorial force */
1707             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1708             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1709             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1710
1711             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1712             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1713             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1714
1715             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1716             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1717             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1718             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1719
1720             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1721                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1722                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1723
1724             /* Inner loop uses 303 flops */
1725         }
1726
1727         /* End of innermost loop */
1728
1729         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1730                                               f+i_coord_offset,fshift+i_shift_offset);
1731
1732         /* Increment number of inner iterations */
1733         inneriter                  += j_index_end - j_index_start;
1734
1735         /* Outer loop uses 24 flops */
1736     }
1737
1738     /* Increment number of outer iterations */
1739     outeriter        += nri;
1740
1741     /* Update outer/inner flops */
1742
1743     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1744 }