src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_avx_128_fma_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  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     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 113     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 114     __m128i          vfitab;
 115     __m128i          ifour       = _mm_set1_epi32(4);
 116     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 117     real             *vftab;
 118     __m128           dummy_mask,cutoff_mask;
 119     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 120     __m128           one     = _mm_set1_ps(1.0);
 121     __m128           two     = _mm_set1_ps(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_ps(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     krf              = _mm_set1_ps(fr->ic->k_rf);
 136     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 137     crf              = _mm_set1_ps(fr->ic->c_rf);
 138     nvdwtype         = fr->ntype;
 139     vdwparam         = fr->nbfp;
 140     vdwtype          = mdatoms->typeA;
 141
 142     vftab            = kernel_data->table_vdw->data;
 143     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 144
 145     /* Setup water-specific parameters */
 146     inr              = nlist->iinr[0];
 147     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 148     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 149     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 150     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 151
 152     jq0              = _mm_set1_ps(charge[inr+0]);
 153     jq1              = _mm_set1_ps(charge[inr+1]);
 154     jq2              = _mm_set1_ps(charge[inr+2]);
 155     vdwjidx0A        = 2*vdwtype[inr+0];
 156     qq00             = _mm_mul_ps(iq0,jq0);
 157     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 158     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 159     qq01             = _mm_mul_ps(iq0,jq1);
 160     qq02             = _mm_mul_ps(iq0,jq2);
 161     qq10             = _mm_mul_ps(iq1,jq0);
 162     qq11             = _mm_mul_ps(iq1,jq1);
 163     qq12             = _mm_mul_ps(iq1,jq2);
 164     qq20             = _mm_mul_ps(iq2,jq0);
 165     qq21             = _mm_mul_ps(iq2,jq1);
 166     qq22             = _mm_mul_ps(iq2,jq2);
 167
 168     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 169     rcutoff_scalar   = fr->rcoulomb;
 170     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 171     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 172
 173     /* Avoid stupid compiler warnings */
 174     jnrA = jnrB = jnrC = jnrD = 0;
 175     j_coord_offsetA = 0;
 176     j_coord_offsetB = 0;
 177     j_coord_offsetC = 0;
 178     j_coord_offsetD = 0;
 179
 180     outeriter        = 0;
 181     inneriter        = 0;
 182
 183     for(iidx=0;iidx<4*DIM;iidx++)
 184     {
 185         scratch[iidx] = 0.0;
 186     }
 187
 188     /* Start outer loop over neighborlists */
 189     for(iidx=0; iidx<nri; iidx++)
 190     {
 191         /* Load shift vector for this list */
 192         i_shift_offset   = DIM*shiftidx[iidx];
 193
 194         /* Load limits for loop over neighbors */
 195         j_index_start    = jindex[iidx];
 196         j_index_end      = jindex[iidx+1];
 197
 198         /* Get outer coordinate index */
 199         inr              = iinr[iidx];
 200         i_coord_offset   = DIM*inr;
 201
 202         /* Load i particle coords and add shift vector */
 203         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 204                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 205
 206         fix0             = _mm_setzero_ps();
 207         fiy0             = _mm_setzero_ps();
 208         fiz0             = _mm_setzero_ps();
 209         fix1             = _mm_setzero_ps();
 210         fiy1             = _mm_setzero_ps();
 211         fiz1             = _mm_setzero_ps();
 212         fix2             = _mm_setzero_ps();
 213         fiy2             = _mm_setzero_ps();
 214         fiz2             = _mm_setzero_ps();
 215
 216         /* Reset potential sums */
 217         velecsum         = _mm_setzero_ps();
 218         vvdwsum          = _mm_setzero_ps();
 219
 220         /* Start inner kernel loop */
 221         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 222         {
 223
 224             /* Get j neighbor index, and coordinate index */
 225             jnrA             = jjnr[jidx];
 226             jnrB             = jjnr[jidx+1];
 227             jnrC             = jjnr[jidx+2];
 228             jnrD             = jjnr[jidx+3];
 229             j_coord_offsetA  = DIM*jnrA;
 230             j_coord_offsetB  = DIM*jnrB;
 231             j_coord_offsetC  = DIM*jnrC;
 232             j_coord_offsetD  = DIM*jnrD;
 233
 234             /* load j atom coordinates */
 235             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 236                                               x+j_coord_offsetC,x+j_coord_offsetD,
 237                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 238
 239             /* Calculate displacement vector */
 240             dx00             = _mm_sub_ps(ix0,jx0);
 241             dy00             = _mm_sub_ps(iy0,jy0);
 242             dz00             = _mm_sub_ps(iz0,jz0);
 243             dx01             = _mm_sub_ps(ix0,jx1);
 244             dy01             = _mm_sub_ps(iy0,jy1);
 245             dz01             = _mm_sub_ps(iz0,jz1);
 246             dx02             = _mm_sub_ps(ix0,jx2);
 247             dy02             = _mm_sub_ps(iy0,jy2);
 248             dz02             = _mm_sub_ps(iz0,jz2);
 249             dx10             = _mm_sub_ps(ix1,jx0);
 250             dy10             = _mm_sub_ps(iy1,jy0);
 251             dz10             = _mm_sub_ps(iz1,jz0);
 252             dx11             = _mm_sub_ps(ix1,jx1);
 253             dy11             = _mm_sub_ps(iy1,jy1);
 254             dz11             = _mm_sub_ps(iz1,jz1);
 255             dx12             = _mm_sub_ps(ix1,jx2);
 256             dy12             = _mm_sub_ps(iy1,jy2);
 257             dz12             = _mm_sub_ps(iz1,jz2);
 258             dx20             = _mm_sub_ps(ix2,jx0);
 259             dy20             = _mm_sub_ps(iy2,jy0);
 260             dz20             = _mm_sub_ps(iz2,jz0);
 261             dx21             = _mm_sub_ps(ix2,jx1);
 262             dy21             = _mm_sub_ps(iy2,jy1);
 263             dz21             = _mm_sub_ps(iz2,jz1);
 264             dx22             = _mm_sub_ps(ix2,jx2);
 265             dy22             = _mm_sub_ps(iy2,jy2);
 266             dz22             = _mm_sub_ps(iz2,jz2);
 267
 268             /* Calculate squared distance and things based on it */
 269             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 270             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 271             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 272             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 273             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 274             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 275             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 276             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 277             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 278
 279             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 280             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 281             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 282             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 283             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 284             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 285             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 286             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 287             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 288
 289             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 290             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 291             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 292             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 293             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 294             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 295             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 296             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 297             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 298
 299             fjx0             = _mm_setzero_ps();
 300             fjy0             = _mm_setzero_ps();
 301             fjz0             = _mm_setzero_ps();
 302             fjx1             = _mm_setzero_ps();
 303             fjy1             = _mm_setzero_ps();
 304             fjz1             = _mm_setzero_ps();
 305             fjx2             = _mm_setzero_ps();
 306             fjy2             = _mm_setzero_ps();
 307             fjz2             = _mm_setzero_ps();
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             if (gmx_mm_any_lt(rsq00,rcutoff2))
 314             {
 315
 316             r00              = _mm_mul_ps(rsq00,rinv00);
 317
 318             /* Calculate table index by multiplying r with table scale and truncate to integer */
 319             rt               = _mm_mul_ps(r00,vftabscale);
 320             vfitab           = _mm_cvttps_epi32(rt);
 321 #ifdef __XOP__
 322             vfeps            = _mm_frcz_ps(rt);
 323 #else
 324             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 325 #endif
 326             twovfeps         = _mm_add_ps(vfeps,vfeps);
 327             vfitab           = _mm_slli_epi32(vfitab,3);
 328
 329             /* REACTION-FIELD ELECTROSTATICS */
 330             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 331             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 332
 333             /* CUBIC SPLINE TABLE DISPERSION */
 334             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 335             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 336             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 337             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 338             _MM_TRANSPOSE4_PS(Y,F,G,H);
 339             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 340             VV               = _mm_macc_ps(vfeps,Fp,Y);
 341             vvdw6            = _mm_mul_ps(c6_00,VV);
 342             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 343             fvdw6            = _mm_mul_ps(c6_00,FF);
 344
 345             /* CUBIC SPLINE TABLE REPULSION */
 346             vfitab           = _mm_add_epi32(vfitab,ifour);
 347             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 348             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 349             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 350             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 351             _MM_TRANSPOSE4_PS(Y,F,G,H);
 352             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 353             VV               = _mm_macc_ps(vfeps,Fp,Y);
 354             vvdw12           = _mm_mul_ps(c12_00,VV);
 355             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 356             fvdw12           = _mm_mul_ps(c12_00,FF);
 357             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 358             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 359
 360             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velec            = _mm_and_ps(velec,cutoff_mask);
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 366             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 367
 368             fscal            = _mm_add_ps(felec,fvdw);
 369
 370             fscal            = _mm_and_ps(fscal,cutoff_mask);
 371
 372              /* Update vectorial force */
 373             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 374             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 375             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 376
 377             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 378             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 379             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 380
 381             }
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             if (gmx_mm_any_lt(rsq01,rcutoff2))
 388             {
 389
 390             /* REACTION-FIELD ELECTROSTATICS */
 391             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 392             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 393
 394             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velec            = _mm_and_ps(velec,cutoff_mask);
 398             velecsum         = _mm_add_ps(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402             fscal            = _mm_and_ps(fscal,cutoff_mask);
 403
 404              /* Update vectorial force */
 405             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 406             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 407             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 408
 409             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 410             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 411             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 412
 413             }
 414
 415             /**************************
 416              * CALCULATE INTERACTIONS *
 417              **************************/
 418
 419             if (gmx_mm_any_lt(rsq02,rcutoff2))
 420             {
 421
 422             /* REACTION-FIELD ELECTROSTATICS */
 423             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 424             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 425
 426             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 427
 428             /* Update potential sum for this i atom from the interaction with this j atom. */
 429             velec            = _mm_and_ps(velec,cutoff_mask);
 430             velecsum         = _mm_add_ps(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434             fscal            = _mm_and_ps(fscal,cutoff_mask);
 435
 436              /* Update vectorial force */
 437             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 438             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 439             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 440
 441             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 442             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 443             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 444
 445             }
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             if (gmx_mm_any_lt(rsq10,rcutoff2))
 452             {
 453
 454             /* REACTION-FIELD ELECTROSTATICS */
 455             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 456             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 457
 458             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velec            = _mm_and_ps(velec,cutoff_mask);
 462             velecsum         = _mm_add_ps(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466             fscal            = _mm_and_ps(fscal,cutoff_mask);
 467
 468              /* Update vectorial force */
 469             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 470             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 471             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 472
 473             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 474             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 475             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 476
 477             }
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             if (gmx_mm_any_lt(rsq11,rcutoff2))
 484             {
 485
 486             /* REACTION-FIELD ELECTROSTATICS */
 487             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 488             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 489
 490             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 491
 492             /* Update potential sum for this i atom from the interaction with this j atom. */
 493             velec            = _mm_and_ps(velec,cutoff_mask);
 494             velecsum         = _mm_add_ps(velecsum,velec);
 495
 496             fscal            = felec;
 497
 498             fscal            = _mm_and_ps(fscal,cutoff_mask);
 499
 500              /* Update vectorial force */
 501             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 502             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 503             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 504
 505             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 506             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 507             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 508
 509             }
 510
 511             /**************************
 512              * CALCULATE INTERACTIONS *
 513              **************************/
 514
 515             if (gmx_mm_any_lt(rsq12,rcutoff2))
 516             {
 517
 518             /* REACTION-FIELD ELECTROSTATICS */
 519             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 520             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 521
 522             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 523
 524             /* Update potential sum for this i atom from the interaction with this j atom. */
 525             velec            = _mm_and_ps(velec,cutoff_mask);
 526             velecsum         = _mm_add_ps(velecsum,velec);
 527
 528             fscal            = felec;
 529
 530             fscal            = _mm_and_ps(fscal,cutoff_mask);
 531
 532              /* Update vectorial force */
 533             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 534             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 535             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 536
 537             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 538             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 539             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 540
 541             }
 542
 543             /**************************
 544              * CALCULATE INTERACTIONS *
 545              **************************/
 546
 547             if (gmx_mm_any_lt(rsq20,rcutoff2))
 548             {
 549
 550             /* REACTION-FIELD ELECTROSTATICS */
 551             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 552             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 553
 554             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 555
 556             /* Update potential sum for this i atom from the interaction with this j atom. */
 557             velec            = _mm_and_ps(velec,cutoff_mask);
 558             velecsum         = _mm_add_ps(velecsum,velec);
 559
 560             fscal            = felec;
 561
 562             fscal            = _mm_and_ps(fscal,cutoff_mask);
 563
 564              /* Update vectorial force */
 565             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 566             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 567             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 568
 569             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 570             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 571             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 572
 573             }
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             if (gmx_mm_any_lt(rsq21,rcutoff2))
 580             {
 581
 582             /* REACTION-FIELD ELECTROSTATICS */
 583             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 584             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 585
 586             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 587
 588             /* Update potential sum for this i atom from the interaction with this j atom. */
 589             velec            = _mm_and_ps(velec,cutoff_mask);
 590             velecsum         = _mm_add_ps(velecsum,velec);
 591
 592             fscal            = felec;
 593
 594             fscal            = _mm_and_ps(fscal,cutoff_mask);
 595
 596              /* Update vectorial force */
 597             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 598             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 599             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 600
 601             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 602             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 603             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 604
 605             }
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             if (gmx_mm_any_lt(rsq22,rcutoff2))
 612             {
 613
 614             /* REACTION-FIELD ELECTROSTATICS */
 615             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 616             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 617
 618             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 619
 620             /* Update potential sum for this i atom from the interaction with this j atom. */
 621             velec            = _mm_and_ps(velec,cutoff_mask);
 622             velecsum         = _mm_add_ps(velecsum,velec);
 623
 624             fscal            = felec;
 625
 626             fscal            = _mm_and_ps(fscal,cutoff_mask);
 627
 628              /* Update vectorial force */
 629             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 630             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 631             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 632
 633             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 634             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 635             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 636
 637             }
 638
 639             fjptrA             = f+j_coord_offsetA;
 640             fjptrB             = f+j_coord_offsetB;
 641             fjptrC             = f+j_coord_offsetC;
 642             fjptrD             = f+j_coord_offsetD;
 643
 644             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 645                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 646
 647             /* Inner loop uses 387 flops */
 648         }
 649
 650         if(jidx<j_index_end)
 651         {
 652
 653             /* Get j neighbor index, and coordinate index */
 654             jnrlistA         = jjnr[jidx];
 655             jnrlistB         = jjnr[jidx+1];
 656             jnrlistC         = jjnr[jidx+2];
 657             jnrlistD         = jjnr[jidx+3];
 658             /* Sign of each element will be negative for non-real atoms.
 659              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 660              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 661              */
 662             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 663             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 664             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 665             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 666             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 667             j_coord_offsetA  = DIM*jnrA;
 668             j_coord_offsetB  = DIM*jnrB;
 669             j_coord_offsetC  = DIM*jnrC;
 670             j_coord_offsetD  = DIM*jnrD;
 671
 672             /* load j atom coordinates */
 673             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 674                                               x+j_coord_offsetC,x+j_coord_offsetD,
 675                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 676
 677             /* Calculate displacement vector */
 678             dx00             = _mm_sub_ps(ix0,jx0);
 679             dy00             = _mm_sub_ps(iy0,jy0);
 680             dz00             = _mm_sub_ps(iz0,jz0);
 681             dx01             = _mm_sub_ps(ix0,jx1);
 682             dy01             = _mm_sub_ps(iy0,jy1);
 683             dz01             = _mm_sub_ps(iz0,jz1);
 684             dx02             = _mm_sub_ps(ix0,jx2);
 685             dy02             = _mm_sub_ps(iy0,jy2);
 686             dz02             = _mm_sub_ps(iz0,jz2);
 687             dx10             = _mm_sub_ps(ix1,jx0);
 688             dy10             = _mm_sub_ps(iy1,jy0);
 689             dz10             = _mm_sub_ps(iz1,jz0);
 690             dx11             = _mm_sub_ps(ix1,jx1);
 691             dy11             = _mm_sub_ps(iy1,jy1);
 692             dz11             = _mm_sub_ps(iz1,jz1);
 693             dx12             = _mm_sub_ps(ix1,jx2);
 694             dy12             = _mm_sub_ps(iy1,jy2);
 695             dz12             = _mm_sub_ps(iz1,jz2);
 696             dx20             = _mm_sub_ps(ix2,jx0);
 697             dy20             = _mm_sub_ps(iy2,jy0);
 698             dz20             = _mm_sub_ps(iz2,jz0);
 699             dx21             = _mm_sub_ps(ix2,jx1);
 700             dy21             = _mm_sub_ps(iy2,jy1);
 701             dz21             = _mm_sub_ps(iz2,jz1);
 702             dx22             = _mm_sub_ps(ix2,jx2);
 703             dy22             = _mm_sub_ps(iy2,jy2);
 704             dz22             = _mm_sub_ps(iz2,jz2);
 705
 706             /* Calculate squared distance and things based on it */
 707             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 708             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 709             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 710             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 711             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 712             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 713             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 714             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 715             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 716
 717             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 718             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 719             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 720             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 721             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 722             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 723             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 724             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 725             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 726
 727             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 728             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 729             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 730             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 731             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 732             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 733             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 734             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 735             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 736
 737             fjx0             = _mm_setzero_ps();
 738             fjy0             = _mm_setzero_ps();
 739             fjz0             = _mm_setzero_ps();
 740             fjx1             = _mm_setzero_ps();
 741             fjy1             = _mm_setzero_ps();
 742             fjz1             = _mm_setzero_ps();
 743             fjx2             = _mm_setzero_ps();
 744             fjy2             = _mm_setzero_ps();
 745             fjz2             = _mm_setzero_ps();
 746
 747             /**************************
 748              * CALCULATE INTERACTIONS *
 749              **************************/
 750
 751             if (gmx_mm_any_lt(rsq00,rcutoff2))
 752             {
 753
 754             r00              = _mm_mul_ps(rsq00,rinv00);
 755             r00              = _mm_andnot_ps(dummy_mask,r00);
 756
 757             /* Calculate table index by multiplying r with table scale and truncate to integer */
 758             rt               = _mm_mul_ps(r00,vftabscale);
 759             vfitab           = _mm_cvttps_epi32(rt);
 760 #ifdef __XOP__
 761             vfeps            = _mm_frcz_ps(rt);
 762 #else
 763             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 764 #endif
 765             twovfeps         = _mm_add_ps(vfeps,vfeps);
 766             vfitab           = _mm_slli_epi32(vfitab,3);
 767
 768             /* REACTION-FIELD ELECTROSTATICS */
 769             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 770             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 771
 772             /* CUBIC SPLINE TABLE DISPERSION */
 773             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 774             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 775             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 776             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 777             _MM_TRANSPOSE4_PS(Y,F,G,H);
 778             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 779             VV               = _mm_macc_ps(vfeps,Fp,Y);
 780             vvdw6            = _mm_mul_ps(c6_00,VV);
 781             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 782             fvdw6            = _mm_mul_ps(c6_00,FF);
 783
 784             /* CUBIC SPLINE TABLE REPULSION */
 785             vfitab           = _mm_add_epi32(vfitab,ifour);
 786             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 787             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 788             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 789             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 790             _MM_TRANSPOSE4_PS(Y,F,G,H);
 791             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 792             VV               = _mm_macc_ps(vfeps,Fp,Y);
 793             vvdw12           = _mm_mul_ps(c12_00,VV);
 794             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 795             fvdw12           = _mm_mul_ps(c12_00,FF);
 796             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 797             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 798
 799             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 800
 801             /* Update potential sum for this i atom from the interaction with this j atom. */
 802             velec            = _mm_and_ps(velec,cutoff_mask);
 803             velec            = _mm_andnot_ps(dummy_mask,velec);
 804             velecsum         = _mm_add_ps(velecsum,velec);
 805             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 806             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 807             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 808
 809             fscal            = _mm_add_ps(felec,fvdw);
 810
 811             fscal            = _mm_and_ps(fscal,cutoff_mask);
 812
 813             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 814
 815              /* Update vectorial force */
 816             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 817             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 818             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 819
 820             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 821             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 822             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 823
 824             }
 825
 826             /**************************
 827              * CALCULATE INTERACTIONS *
 828              **************************/
 829
 830             if (gmx_mm_any_lt(rsq01,rcutoff2))
 831             {
 832
 833             /* REACTION-FIELD ELECTROSTATICS */
 834             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 835             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 836
 837             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 838
 839             /* Update potential sum for this i atom from the interaction with this j atom. */
 840             velec            = _mm_and_ps(velec,cutoff_mask);
 841             velec            = _mm_andnot_ps(dummy_mask,velec);
 842             velecsum         = _mm_add_ps(velecsum,velec);
 843
 844             fscal            = felec;
 845
 846             fscal            = _mm_and_ps(fscal,cutoff_mask);
 847
 848             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 849
 850              /* Update vectorial force */
 851             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 852             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 853             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 854
 855             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 856             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 857             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 858
 859             }
 860
 861             /**************************
 862              * CALCULATE INTERACTIONS *
 863              **************************/
 864
 865             if (gmx_mm_any_lt(rsq02,rcutoff2))
 866             {
 867
 868             /* REACTION-FIELD ELECTROSTATICS */
 869             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 870             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 871
 872             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 873
 874             /* Update potential sum for this i atom from the interaction with this j atom. */
 875             velec            = _mm_and_ps(velec,cutoff_mask);
 876             velec            = _mm_andnot_ps(dummy_mask,velec);
 877             velecsum         = _mm_add_ps(velecsum,velec);
 878
 879             fscal            = felec;
 880
 881             fscal            = _mm_and_ps(fscal,cutoff_mask);
 882
 883             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 884
 885              /* Update vectorial force */
 886             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 887             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 888             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 889
 890             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 891             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 892             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 893
 894             }
 895
 896             /**************************
 897              * CALCULATE INTERACTIONS *
 898              **************************/
 899
 900             if (gmx_mm_any_lt(rsq10,rcutoff2))
 901             {
 902
 903             /* REACTION-FIELD ELECTROSTATICS */
 904             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 905             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 906
 907             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 908
 909             /* Update potential sum for this i atom from the interaction with this j atom. */
 910             velec            = _mm_and_ps(velec,cutoff_mask);
 911             velec            = _mm_andnot_ps(dummy_mask,velec);
 912             velecsum         = _mm_add_ps(velecsum,velec);
 913
 914             fscal            = felec;
 915
 916             fscal            = _mm_and_ps(fscal,cutoff_mask);
 917
 918             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 919
 920              /* Update vectorial force */
 921             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 922             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 923             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 924
 925             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 926             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 927             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 928
 929             }
 930
 931             /**************************
 932              * CALCULATE INTERACTIONS *
 933              **************************/
 934
 935             if (gmx_mm_any_lt(rsq11,rcutoff2))
 936             {
 937
 938             /* REACTION-FIELD ELECTROSTATICS */
 939             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 940             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 941
 942             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 943
 944             /* Update potential sum for this i atom from the interaction with this j atom. */
 945             velec            = _mm_and_ps(velec,cutoff_mask);
 946             velec            = _mm_andnot_ps(dummy_mask,velec);
 947             velecsum         = _mm_add_ps(velecsum,velec);
 948
 949             fscal            = felec;
 950
 951             fscal            = _mm_and_ps(fscal,cutoff_mask);
 952
 953             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 954
 955              /* Update vectorial force */
 956             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 957             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 958             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 959
 960             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 961             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 962             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 963
 964             }
 965
 966             /**************************
 967              * CALCULATE INTERACTIONS *
 968              **************************/
 969
 970             if (gmx_mm_any_lt(rsq12,rcutoff2))
 971             {
 972
 973             /* REACTION-FIELD ELECTROSTATICS */
 974             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 975             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 976
 977             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 978
 979             /* Update potential sum for this i atom from the interaction with this j atom. */
 980             velec            = _mm_and_ps(velec,cutoff_mask);
 981             velec            = _mm_andnot_ps(dummy_mask,velec);
 982             velecsum         = _mm_add_ps(velecsum,velec);
 983
 984             fscal            = felec;
 985
 986             fscal            = _mm_and_ps(fscal,cutoff_mask);
 987
 988             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 989
 990              /* Update vectorial force */
 991             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 992             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 993             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 994
 995             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 996             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 997             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 998
 999             }
1000
1001             /**************************
1002              * CALCULATE INTERACTIONS *
1003              **************************/
1004
1005             if (gmx_mm_any_lt(rsq20,rcutoff2))
1006             {
1007
1008             /* REACTION-FIELD ELECTROSTATICS */
1009             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
1010             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1011
1012             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1013
1014             /* Update potential sum for this i atom from the interaction with this j atom. */
1015             velec            = _mm_and_ps(velec,cutoff_mask);
1016             velec            = _mm_andnot_ps(dummy_mask,velec);
1017             velecsum         = _mm_add_ps(velecsum,velec);
1018
1019             fscal            = felec;
1020
1021             fscal            = _mm_and_ps(fscal,cutoff_mask);
1022
1023             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1024
1025              /* Update vectorial force */
1026             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1027             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1028             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1029
1030             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1031             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1032             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1033
1034             }
1035
1036             /**************************
1037              * CALCULATE INTERACTIONS *
1038              **************************/
1039
1040             if (gmx_mm_any_lt(rsq21,rcutoff2))
1041             {
1042
1043             /* REACTION-FIELD ELECTROSTATICS */
1044             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
1045             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1046
1047             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1048
1049             /* Update potential sum for this i atom from the interaction with this j atom. */
1050             velec            = _mm_and_ps(velec,cutoff_mask);
1051             velec            = _mm_andnot_ps(dummy_mask,velec);
1052             velecsum         = _mm_add_ps(velecsum,velec);
1053
1054             fscal            = felec;
1055
1056             fscal            = _mm_and_ps(fscal,cutoff_mask);
1057
1058             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1059
1060              /* Update vectorial force */
1061             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1062             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1063             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1064
1065             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1066             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1067             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1068
1069             }
1070
1071             /**************************
1072              * CALCULATE INTERACTIONS *
1073              **************************/
1074
1075             if (gmx_mm_any_lt(rsq22,rcutoff2))
1076             {
1077
1078             /* REACTION-FIELD ELECTROSTATICS */
1079             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
1080             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1081
1082             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1083
1084             /* Update potential sum for this i atom from the interaction with this j atom. */
1085             velec            = _mm_and_ps(velec,cutoff_mask);
1086             velec            = _mm_andnot_ps(dummy_mask,velec);
1087             velecsum         = _mm_add_ps(velecsum,velec);
1088
1089             fscal            = felec;
1090
1091             fscal            = _mm_and_ps(fscal,cutoff_mask);
1092
1093             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1094
1095              /* Update vectorial force */
1096             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1097             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1098             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1099
1100             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1101             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1102             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1103
1104             }
1105
1106             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1107             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1108             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1109             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1110
1111             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1112                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1113
1114             /* Inner loop uses 388 flops */
1115         }
1116
1117         /* End of innermost loop */
1118
1119         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1120                                               f+i_coord_offset,fshift+i_shift_offset);
1121
1122         ggid                        = gid[iidx];
1123         /* Update potential energies */
1124         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1125         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1126
1127         /* Increment number of inner iterations */
1128         inneriter                  += j_index_end - j_index_start;
1129
1130         /* Outer loop uses 20 flops */
1131     }
1132
1133     /* Increment number of outer iterations */
1134     outeriter        += nri;
1135
1136     /* Update outer/inner flops */
1137
1138     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*388);
1139 }
1140 /*
1141  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_avx_128_fma_single
1142  * Electrostatics interaction: ReactionField
1143  * VdW interaction:            CubicSplineTable
1144  * Geometry:                   Water3-Water3
1145  * Calculate force/pot:        Force
1146  */
1147 void
1148 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_avx_128_fma_single
1149                     (t_nblist                    * gmx_restrict       nlist,
1150                      rvec                        * gmx_restrict          xx,
1151                      rvec                        * gmx_restrict          ff,
1152                      t_forcerec                  * gmx_restrict          fr,
1153                      t_mdatoms                   * gmx_restrict     mdatoms,
1154                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1155                      t_nrnb                      * gmx_restrict        nrnb)
1156 {
1157     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1158      * just 0 for non-waters.
1159      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1160      * jnr indices corresponding to data put in the four positions in the SIMD register.
1161      */
1162     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1163     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1164     int              jnrA,jnrB,jnrC,jnrD;
1165     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1166     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1167     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1168     real             rcutoff_scalar;
1169     real             *shiftvec,*fshift,*x,*f;
1170     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1171     real             scratch[4*DIM];
1172     __m128           fscal,rcutoff,rcutoff2,jidxall;
1173     int              vdwioffset0;
1174     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1175     int              vdwioffset1;
1176     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1177     int              vdwioffset2;
1178     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1179     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1180     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1181     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1182     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1183     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1184     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1185     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1186     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1187     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1188     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1189     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1190     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1191     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1192     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1193     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1194     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1195     real             *charge;
1196     int              nvdwtype;
1197     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1198     int              *vdwtype;
1199     real             *vdwparam;
1200     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1201     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1202     __m128i          vfitab;
1203     __m128i          ifour       = _mm_set1_epi32(4);
1204     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1205     real             *vftab;
1206     __m128           dummy_mask,cutoff_mask;
1207     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1208     __m128           one     = _mm_set1_ps(1.0);
1209     __m128           two     = _mm_set1_ps(2.0);
1210     x                = xx[0];
1211     f                = ff[0];
1212
1213     nri              = nlist->nri;
1214     iinr             = nlist->iinr;
1215     jindex           = nlist->jindex;
1216     jjnr             = nlist->jjnr;
1217     shiftidx         = nlist->shift;
1218     gid              = nlist->gid;
1219     shiftvec         = fr->shift_vec[0];
1220     fshift           = fr->fshift[0];
1221     facel            = _mm_set1_ps(fr->epsfac);
1222     charge           = mdatoms->chargeA;
1223     krf              = _mm_set1_ps(fr->ic->k_rf);
1224     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1225     crf              = _mm_set1_ps(fr->ic->c_rf);
1226     nvdwtype         = fr->ntype;
1227     vdwparam         = fr->nbfp;
1228     vdwtype          = mdatoms->typeA;
1229
1230     vftab            = kernel_data->table_vdw->data;
1231     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1232
1233     /* Setup water-specific parameters */
1234     inr              = nlist->iinr[0];
1235     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1236     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1237     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1238     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1239
1240     jq0              = _mm_set1_ps(charge[inr+0]);
1241     jq1              = _mm_set1_ps(charge[inr+1]);
1242     jq2              = _mm_set1_ps(charge[inr+2]);
1243     vdwjidx0A        = 2*vdwtype[inr+0];
1244     qq00             = _mm_mul_ps(iq0,jq0);
1245     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1246     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1247     qq01             = _mm_mul_ps(iq0,jq1);
1248     qq02             = _mm_mul_ps(iq0,jq2);
1249     qq10             = _mm_mul_ps(iq1,jq0);
1250     qq11             = _mm_mul_ps(iq1,jq1);
1251     qq12             = _mm_mul_ps(iq1,jq2);
1252     qq20             = _mm_mul_ps(iq2,jq0);
1253     qq21             = _mm_mul_ps(iq2,jq1);
1254     qq22             = _mm_mul_ps(iq2,jq2);
1255
1256     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1257     rcutoff_scalar   = fr->rcoulomb;
1258     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1259     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1260
1261     /* Avoid stupid compiler warnings */
1262     jnrA = jnrB = jnrC = jnrD = 0;
1263     j_coord_offsetA = 0;
1264     j_coord_offsetB = 0;
1265     j_coord_offsetC = 0;
1266     j_coord_offsetD = 0;
1267
1268     outeriter        = 0;
1269     inneriter        = 0;
1270
1271     for(iidx=0;iidx<4*DIM;iidx++)
1272     {
1273         scratch[iidx] = 0.0;
1274     }
1275
1276     /* Start outer loop over neighborlists */
1277     for(iidx=0; iidx<nri; iidx++)
1278     {
1279         /* Load shift vector for this list */
1280         i_shift_offset   = DIM*shiftidx[iidx];
1281
1282         /* Load limits for loop over neighbors */
1283         j_index_start    = jindex[iidx];
1284         j_index_end      = jindex[iidx+1];
1285
1286         /* Get outer coordinate index */
1287         inr              = iinr[iidx];
1288         i_coord_offset   = DIM*inr;
1289
1290         /* Load i particle coords and add shift vector */
1291         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1292                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1293
1294         fix0             = _mm_setzero_ps();
1295         fiy0             = _mm_setzero_ps();
1296         fiz0             = _mm_setzero_ps();
1297         fix1             = _mm_setzero_ps();
1298         fiy1             = _mm_setzero_ps();
1299         fiz1             = _mm_setzero_ps();
1300         fix2             = _mm_setzero_ps();
1301         fiy2             = _mm_setzero_ps();
1302         fiz2             = _mm_setzero_ps();
1303
1304         /* Start inner kernel loop */
1305         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1306         {
1307
1308             /* Get j neighbor index, and coordinate index */
1309             jnrA             = jjnr[jidx];
1310             jnrB             = jjnr[jidx+1];
1311             jnrC             = jjnr[jidx+2];
1312             jnrD             = jjnr[jidx+3];
1313             j_coord_offsetA  = DIM*jnrA;
1314             j_coord_offsetB  = DIM*jnrB;
1315             j_coord_offsetC  = DIM*jnrC;
1316             j_coord_offsetD  = DIM*jnrD;
1317
1318             /* load j atom coordinates */
1319             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1320                                               x+j_coord_offsetC,x+j_coord_offsetD,
1321                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1322
1323             /* Calculate displacement vector */
1324             dx00             = _mm_sub_ps(ix0,jx0);
1325             dy00             = _mm_sub_ps(iy0,jy0);
1326             dz00             = _mm_sub_ps(iz0,jz0);
1327             dx01             = _mm_sub_ps(ix0,jx1);
1328             dy01             = _mm_sub_ps(iy0,jy1);
1329             dz01             = _mm_sub_ps(iz0,jz1);
1330             dx02             = _mm_sub_ps(ix0,jx2);
1331             dy02             = _mm_sub_ps(iy0,jy2);
1332             dz02             = _mm_sub_ps(iz0,jz2);
1333             dx10             = _mm_sub_ps(ix1,jx0);
1334             dy10             = _mm_sub_ps(iy1,jy0);
1335             dz10             = _mm_sub_ps(iz1,jz0);
1336             dx11             = _mm_sub_ps(ix1,jx1);
1337             dy11             = _mm_sub_ps(iy1,jy1);
1338             dz11             = _mm_sub_ps(iz1,jz1);
1339             dx12             = _mm_sub_ps(ix1,jx2);
1340             dy12             = _mm_sub_ps(iy1,jy2);
1341             dz12             = _mm_sub_ps(iz1,jz2);
1342             dx20             = _mm_sub_ps(ix2,jx0);
1343             dy20             = _mm_sub_ps(iy2,jy0);
1344             dz20             = _mm_sub_ps(iz2,jz0);
1345             dx21             = _mm_sub_ps(ix2,jx1);
1346             dy21             = _mm_sub_ps(iy2,jy1);
1347             dz21             = _mm_sub_ps(iz2,jz1);
1348             dx22             = _mm_sub_ps(ix2,jx2);
1349             dy22             = _mm_sub_ps(iy2,jy2);
1350             dz22             = _mm_sub_ps(iz2,jz2);
1351
1352             /* Calculate squared distance and things based on it */
1353             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1354             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1355             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1356             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1357             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1358             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1359             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1360             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1361             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1362
1363             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1364             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1365             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1366             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1367             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1368             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1369             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1370             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1371             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1372
1373             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1374             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1375             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1376             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1377             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1378             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1379             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1380             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1381             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1382
1383             fjx0             = _mm_setzero_ps();
1384             fjy0             = _mm_setzero_ps();
1385             fjz0             = _mm_setzero_ps();
1386             fjx1             = _mm_setzero_ps();
1387             fjy1             = _mm_setzero_ps();
1388             fjz1             = _mm_setzero_ps();
1389             fjx2             = _mm_setzero_ps();
1390             fjy2             = _mm_setzero_ps();
1391             fjz2             = _mm_setzero_ps();
1392
1393             /**************************
1394              * CALCULATE INTERACTIONS *
1395              **************************/
1396
1397             if (gmx_mm_any_lt(rsq00,rcutoff2))
1398             {
1399
1400             r00              = _mm_mul_ps(rsq00,rinv00);
1401
1402             /* Calculate table index by multiplying r with table scale and truncate to integer */
1403             rt               = _mm_mul_ps(r00,vftabscale);
1404             vfitab           = _mm_cvttps_epi32(rt);
1405 #ifdef __XOP__
1406             vfeps            = _mm_frcz_ps(rt);
1407 #else
1408             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1409 #endif
1410             twovfeps         = _mm_add_ps(vfeps,vfeps);
1411             vfitab           = _mm_slli_epi32(vfitab,3);
1412
1413             /* REACTION-FIELD ELECTROSTATICS */
1414             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1415
1416             /* CUBIC SPLINE TABLE DISPERSION */
1417             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1418             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1419             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1420             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1421             _MM_TRANSPOSE4_PS(Y,F,G,H);
1422             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1423             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1424             fvdw6            = _mm_mul_ps(c6_00,FF);
1425
1426             /* CUBIC SPLINE TABLE REPULSION */
1427             vfitab           = _mm_add_epi32(vfitab,ifour);
1428             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1429             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1430             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1431             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1432             _MM_TRANSPOSE4_PS(Y,F,G,H);
1433             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1434             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1435             fvdw12           = _mm_mul_ps(c12_00,FF);
1436             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1437
1438             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1439
1440             fscal            = _mm_add_ps(felec,fvdw);
1441
1442             fscal            = _mm_and_ps(fscal,cutoff_mask);
1443
1444              /* Update vectorial force */
1445             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1446             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1447             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1448
1449             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1450             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1451             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1452
1453             }
1454
1455             /**************************
1456              * CALCULATE INTERACTIONS *
1457              **************************/
1458
1459             if (gmx_mm_any_lt(rsq01,rcutoff2))
1460             {
1461
1462             /* REACTION-FIELD ELECTROSTATICS */
1463             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1464
1465             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1466
1467             fscal            = felec;
1468
1469             fscal            = _mm_and_ps(fscal,cutoff_mask);
1470
1471              /* Update vectorial force */
1472             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1473             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1474             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1475
1476             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1477             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1478             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1479
1480             }
1481
1482             /**************************
1483              * CALCULATE INTERACTIONS *
1484              **************************/
1485
1486             if (gmx_mm_any_lt(rsq02,rcutoff2))
1487             {
1488
1489             /* REACTION-FIELD ELECTROSTATICS */
1490             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1491
1492             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1493
1494             fscal            = felec;
1495
1496             fscal            = _mm_and_ps(fscal,cutoff_mask);
1497
1498              /* Update vectorial force */
1499             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1500             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1501             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1502
1503             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1504             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1505             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1506
1507             }
1508
1509             /**************************
1510              * CALCULATE INTERACTIONS *
1511              **************************/
1512
1513             if (gmx_mm_any_lt(rsq10,rcutoff2))
1514             {
1515
1516             /* REACTION-FIELD ELECTROSTATICS */
1517             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1518
1519             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1520
1521             fscal            = felec;
1522
1523             fscal            = _mm_and_ps(fscal,cutoff_mask);
1524
1525              /* Update vectorial force */
1526             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1527             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1528             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1529
1530             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1531             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1532             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1533
1534             }
1535
1536             /**************************
1537              * CALCULATE INTERACTIONS *
1538              **************************/
1539
1540             if (gmx_mm_any_lt(rsq11,rcutoff2))
1541             {
1542
1543             /* REACTION-FIELD ELECTROSTATICS */
1544             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1545
1546             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1547
1548             fscal            = felec;
1549
1550             fscal            = _mm_and_ps(fscal,cutoff_mask);
1551
1552              /* Update vectorial force */
1553             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1554             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1555             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1556
1557             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1558             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1559             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1560
1561             }
1562
1563             /**************************
1564              * CALCULATE INTERACTIONS *
1565              **************************/
1566
1567             if (gmx_mm_any_lt(rsq12,rcutoff2))
1568             {
1569
1570             /* REACTION-FIELD ELECTROSTATICS */
1571             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1572
1573             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1574
1575             fscal            = felec;
1576
1577             fscal            = _mm_and_ps(fscal,cutoff_mask);
1578
1579              /* Update vectorial force */
1580             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1581             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1582             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1583
1584             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1585             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1586             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1587
1588             }
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             if (gmx_mm_any_lt(rsq20,rcutoff2))
1595             {
1596
1597             /* REACTION-FIELD ELECTROSTATICS */
1598             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1599
1600             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1601
1602             fscal            = felec;
1603
1604             fscal            = _mm_and_ps(fscal,cutoff_mask);
1605
1606              /* Update vectorial force */
1607             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1608             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1609             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1610
1611             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1612             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1613             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1614
1615             }
1616
1617             /**************************
1618              * CALCULATE INTERACTIONS *
1619              **************************/
1620
1621             if (gmx_mm_any_lt(rsq21,rcutoff2))
1622             {
1623
1624             /* REACTION-FIELD ELECTROSTATICS */
1625             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1626
1627             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1628
1629             fscal            = felec;
1630
1631             fscal            = _mm_and_ps(fscal,cutoff_mask);
1632
1633              /* Update vectorial force */
1634             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1635             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1636             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1637
1638             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1639             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1640             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1641
1642             }
1643
1644             /**************************
1645              * CALCULATE INTERACTIONS *
1646              **************************/
1647
1648             if (gmx_mm_any_lt(rsq22,rcutoff2))
1649             {
1650
1651             /* REACTION-FIELD ELECTROSTATICS */
1652             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1653
1654             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1655
1656             fscal            = felec;
1657
1658             fscal            = _mm_and_ps(fscal,cutoff_mask);
1659
1660              /* Update vectorial force */
1661             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1662             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1663             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1664
1665             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1666             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1667             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1668
1669             }
1670
1671             fjptrA             = f+j_coord_offsetA;
1672             fjptrB             = f+j_coord_offsetB;
1673             fjptrC             = f+j_coord_offsetC;
1674             fjptrD             = f+j_coord_offsetD;
1675
1676             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1677                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1678
1679             /* Inner loop uses 324 flops */
1680         }
1681
1682         if(jidx<j_index_end)
1683         {
1684
1685             /* Get j neighbor index, and coordinate index */
1686             jnrlistA         = jjnr[jidx];
1687             jnrlistB         = jjnr[jidx+1];
1688             jnrlistC         = jjnr[jidx+2];
1689             jnrlistD         = jjnr[jidx+3];
1690             /* Sign of each element will be negative for non-real atoms.
1691              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1692              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1693              */
1694             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1695             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1696             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1697             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1698             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1699             j_coord_offsetA  = DIM*jnrA;
1700             j_coord_offsetB  = DIM*jnrB;
1701             j_coord_offsetC  = DIM*jnrC;
1702             j_coord_offsetD  = DIM*jnrD;
1703
1704             /* load j atom coordinates */
1705             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1706                                               x+j_coord_offsetC,x+j_coord_offsetD,
1707                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1708
1709             /* Calculate displacement vector */
1710             dx00             = _mm_sub_ps(ix0,jx0);
1711             dy00             = _mm_sub_ps(iy0,jy0);
1712             dz00             = _mm_sub_ps(iz0,jz0);
1713             dx01             = _mm_sub_ps(ix0,jx1);
1714             dy01             = _mm_sub_ps(iy0,jy1);
1715             dz01             = _mm_sub_ps(iz0,jz1);
1716             dx02             = _mm_sub_ps(ix0,jx2);
1717             dy02             = _mm_sub_ps(iy0,jy2);
1718             dz02             = _mm_sub_ps(iz0,jz2);
1719             dx10             = _mm_sub_ps(ix1,jx0);
1720             dy10             = _mm_sub_ps(iy1,jy0);
1721             dz10             = _mm_sub_ps(iz1,jz0);
1722             dx11             = _mm_sub_ps(ix1,jx1);
1723             dy11             = _mm_sub_ps(iy1,jy1);
1724             dz11             = _mm_sub_ps(iz1,jz1);
1725             dx12             = _mm_sub_ps(ix1,jx2);
1726             dy12             = _mm_sub_ps(iy1,jy2);
1727             dz12             = _mm_sub_ps(iz1,jz2);
1728             dx20             = _mm_sub_ps(ix2,jx0);
1729             dy20             = _mm_sub_ps(iy2,jy0);
1730             dz20             = _mm_sub_ps(iz2,jz0);
1731             dx21             = _mm_sub_ps(ix2,jx1);
1732             dy21             = _mm_sub_ps(iy2,jy1);
1733             dz21             = _mm_sub_ps(iz2,jz1);
1734             dx22             = _mm_sub_ps(ix2,jx2);
1735             dy22             = _mm_sub_ps(iy2,jy2);
1736             dz22             = _mm_sub_ps(iz2,jz2);
1737
1738             /* Calculate squared distance and things based on it */
1739             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1740             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1741             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1742             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1743             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1744             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1745             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1746             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1747             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1748
1749             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1750             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1751             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1752             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1753             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1754             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1755             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1756             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1757             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1758
1759             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1760             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1761             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1762             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1763             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1764             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1765             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1766             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1767             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1768
1769             fjx0             = _mm_setzero_ps();
1770             fjy0             = _mm_setzero_ps();
1771             fjz0             = _mm_setzero_ps();
1772             fjx1             = _mm_setzero_ps();
1773             fjy1             = _mm_setzero_ps();
1774             fjz1             = _mm_setzero_ps();
1775             fjx2             = _mm_setzero_ps();
1776             fjy2             = _mm_setzero_ps();
1777             fjz2             = _mm_setzero_ps();
1778
1779             /**************************
1780              * CALCULATE INTERACTIONS *
1781              **************************/
1782
1783             if (gmx_mm_any_lt(rsq00,rcutoff2))
1784             {
1785
1786             r00              = _mm_mul_ps(rsq00,rinv00);
1787             r00              = _mm_andnot_ps(dummy_mask,r00);
1788
1789             /* Calculate table index by multiplying r with table scale and truncate to integer */
1790             rt               = _mm_mul_ps(r00,vftabscale);
1791             vfitab           = _mm_cvttps_epi32(rt);
1792 #ifdef __XOP__
1793             vfeps            = _mm_frcz_ps(rt);
1794 #else
1795             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1796 #endif
1797             twovfeps         = _mm_add_ps(vfeps,vfeps);
1798             vfitab           = _mm_slli_epi32(vfitab,3);
1799
1800             /* REACTION-FIELD ELECTROSTATICS */
1801             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1802
1803             /* CUBIC SPLINE TABLE DISPERSION */
1804             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1805             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1806             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1807             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1808             _MM_TRANSPOSE4_PS(Y,F,G,H);
1809             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1810             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1811             fvdw6            = _mm_mul_ps(c6_00,FF);
1812
1813             /* CUBIC SPLINE TABLE REPULSION */
1814             vfitab           = _mm_add_epi32(vfitab,ifour);
1815             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1816             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1817             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1818             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1819             _MM_TRANSPOSE4_PS(Y,F,G,H);
1820             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1821             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1822             fvdw12           = _mm_mul_ps(c12_00,FF);
1823             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1824
1825             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1826
1827             fscal            = _mm_add_ps(felec,fvdw);
1828
1829             fscal            = _mm_and_ps(fscal,cutoff_mask);
1830
1831             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1832
1833              /* Update vectorial force */
1834             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1835             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1836             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1837
1838             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1839             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1840             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1841
1842             }
1843
1844             /**************************
1845              * CALCULATE INTERACTIONS *
1846              **************************/
1847
1848             if (gmx_mm_any_lt(rsq01,rcutoff2))
1849             {
1850
1851             /* REACTION-FIELD ELECTROSTATICS */
1852             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1853
1854             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1855
1856             fscal            = felec;
1857
1858             fscal            = _mm_and_ps(fscal,cutoff_mask);
1859
1860             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1861
1862              /* Update vectorial force */
1863             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1864             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1865             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1866
1867             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1868             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1869             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1870
1871             }
1872
1873             /**************************
1874              * CALCULATE INTERACTIONS *
1875              **************************/
1876
1877             if (gmx_mm_any_lt(rsq02,rcutoff2))
1878             {
1879
1880             /* REACTION-FIELD ELECTROSTATICS */
1881             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1882
1883             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1884
1885             fscal            = felec;
1886
1887             fscal            = _mm_and_ps(fscal,cutoff_mask);
1888
1889             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1890
1891              /* Update vectorial force */
1892             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1893             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1894             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1895
1896             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1897             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1898             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1899
1900             }
1901
1902             /**************************
1903              * CALCULATE INTERACTIONS *
1904              **************************/
1905
1906             if (gmx_mm_any_lt(rsq10,rcutoff2))
1907             {
1908
1909             /* REACTION-FIELD ELECTROSTATICS */
1910             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1911
1912             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1913
1914             fscal            = felec;
1915
1916             fscal            = _mm_and_ps(fscal,cutoff_mask);
1917
1918             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1919
1920              /* Update vectorial force */
1921             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1922             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1923             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1924
1925             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1926             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1927             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1928
1929             }
1930
1931             /**************************
1932              * CALCULATE INTERACTIONS *
1933              **************************/
1934
1935             if (gmx_mm_any_lt(rsq11,rcutoff2))
1936             {
1937
1938             /* REACTION-FIELD ELECTROSTATICS */
1939             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1940
1941             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1942
1943             fscal            = felec;
1944
1945             fscal            = _mm_and_ps(fscal,cutoff_mask);
1946
1947             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1948
1949              /* Update vectorial force */
1950             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1951             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1952             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1953
1954             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1955             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1956             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1957
1958             }
1959
1960             /**************************
1961              * CALCULATE INTERACTIONS *
1962              **************************/
1963
1964             if (gmx_mm_any_lt(rsq12,rcutoff2))
1965             {
1966
1967             /* REACTION-FIELD ELECTROSTATICS */
1968             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1969
1970             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1971
1972             fscal            = felec;
1973
1974             fscal            = _mm_and_ps(fscal,cutoff_mask);
1975
1976             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1977
1978              /* Update vectorial force */
1979             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1980             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1981             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1982
1983             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1984             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1985             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1986
1987             }
1988
1989             /**************************
1990              * CALCULATE INTERACTIONS *
1991              **************************/
1992
1993             if (gmx_mm_any_lt(rsq20,rcutoff2))
1994             {
1995
1996             /* REACTION-FIELD ELECTROSTATICS */
1997             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1998
1999             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2000
2001             fscal            = felec;
2002
2003             fscal            = _mm_and_ps(fscal,cutoff_mask);
2004
2005             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2006
2007              /* Update vectorial force */
2008             fix2             = _mm_macc_ps(dx20,fscal,fix2);
2009             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
2010             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
2011
2012             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
2013             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
2014             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
2015
2016             }
2017
2018             /**************************
2019              * CALCULATE INTERACTIONS *
2020              **************************/
2021
2022             if (gmx_mm_any_lt(rsq21,rcutoff2))
2023             {
2024
2025             /* REACTION-FIELD ELECTROSTATICS */
2026             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
2027
2028             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2029
2030             fscal            = felec;
2031
2032             fscal            = _mm_and_ps(fscal,cutoff_mask);
2033
2034             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2035
2036              /* Update vectorial force */
2037             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2038             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2039             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2040
2041             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2042             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2043             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2044
2045             }
2046
2047             /**************************
2048              * CALCULATE INTERACTIONS *
2049              **************************/
2050
2051             if (gmx_mm_any_lt(rsq22,rcutoff2))
2052             {
2053
2054             /* REACTION-FIELD ELECTROSTATICS */
2055             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
2056
2057             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2058
2059             fscal            = felec;
2060
2061             fscal            = _mm_and_ps(fscal,cutoff_mask);
2062
2063             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2064
2065              /* Update vectorial force */
2066             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2067             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2068             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2069
2070             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2071             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2072             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2073
2074             }
2075
2076             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2077             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2078             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2079             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2080
2081             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2082                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2083
2084             /* Inner loop uses 325 flops */
2085         }
2086
2087         /* End of innermost loop */
2088
2089         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2090                                               f+i_coord_offset,fshift+i_shift_offset);
2091
2092         /* Increment number of inner iterations */
2093         inneriter                  += j_index_end - j_index_start;
2094
2095         /* Outer loop uses 18 flops */
2096     }
2097
2098     /* Increment number of outer iterations */
2099     outeriter        += nri;
2100
2101     /* Update outer/inner flops */
2102
2103     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*325);
2104 }