src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRF_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_ElecRF_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_ElecRF_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     /* Avoid stupid compiler warnings */
 169     jnrA = jnrB = jnrC = jnrD = 0;
 170     j_coord_offsetA = 0;
 171     j_coord_offsetB = 0;
 172     j_coord_offsetC = 0;
 173     j_coord_offsetD = 0;
 174
 175     outeriter        = 0;
 176     inneriter        = 0;
 177
 178     for(iidx=0;iidx<4*DIM;iidx++)
 179     {
 180         scratch[iidx] = 0.0;
 181     }
 182
 183     /* Start outer loop over neighborlists */
 184     for(iidx=0; iidx<nri; iidx++)
 185     {
 186         /* Load shift vector for this list */
 187         i_shift_offset   = DIM*shiftidx[iidx];
 188
 189         /* Load limits for loop over neighbors */
 190         j_index_start    = jindex[iidx];
 191         j_index_end      = jindex[iidx+1];
 192
 193         /* Get outer coordinate index */
 194         inr              = iinr[iidx];
 195         i_coord_offset   = DIM*inr;
 196
 197         /* Load i particle coords and add shift vector */
 198         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 199                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 200
 201         fix0             = _mm_setzero_ps();
 202         fiy0             = _mm_setzero_ps();
 203         fiz0             = _mm_setzero_ps();
 204         fix1             = _mm_setzero_ps();
 205         fiy1             = _mm_setzero_ps();
 206         fiz1             = _mm_setzero_ps();
 207         fix2             = _mm_setzero_ps();
 208         fiy2             = _mm_setzero_ps();
 209         fiz2             = _mm_setzero_ps();
 210
 211         /* Reset potential sums */
 212         velecsum         = _mm_setzero_ps();
 213         vvdwsum          = _mm_setzero_ps();
 214
 215         /* Start inner kernel loop */
 216         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 217         {
 218
 219             /* Get j neighbor index, and coordinate index */
 220             jnrA             = jjnr[jidx];
 221             jnrB             = jjnr[jidx+1];
 222             jnrC             = jjnr[jidx+2];
 223             jnrD             = jjnr[jidx+3];
 224             j_coord_offsetA  = DIM*jnrA;
 225             j_coord_offsetB  = DIM*jnrB;
 226             j_coord_offsetC  = DIM*jnrC;
 227             j_coord_offsetD  = DIM*jnrD;
 228
 229             /* load j atom coordinates */
 230             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 231                                               x+j_coord_offsetC,x+j_coord_offsetD,
 232                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 233
 234             /* Calculate displacement vector */
 235             dx00             = _mm_sub_ps(ix0,jx0);
 236             dy00             = _mm_sub_ps(iy0,jy0);
 237             dz00             = _mm_sub_ps(iz0,jz0);
 238             dx01             = _mm_sub_ps(ix0,jx1);
 239             dy01             = _mm_sub_ps(iy0,jy1);
 240             dz01             = _mm_sub_ps(iz0,jz1);
 241             dx02             = _mm_sub_ps(ix0,jx2);
 242             dy02             = _mm_sub_ps(iy0,jy2);
 243             dz02             = _mm_sub_ps(iz0,jz2);
 244             dx10             = _mm_sub_ps(ix1,jx0);
 245             dy10             = _mm_sub_ps(iy1,jy0);
 246             dz10             = _mm_sub_ps(iz1,jz0);
 247             dx11             = _mm_sub_ps(ix1,jx1);
 248             dy11             = _mm_sub_ps(iy1,jy1);
 249             dz11             = _mm_sub_ps(iz1,jz1);
 250             dx12             = _mm_sub_ps(ix1,jx2);
 251             dy12             = _mm_sub_ps(iy1,jy2);
 252             dz12             = _mm_sub_ps(iz1,jz2);
 253             dx20             = _mm_sub_ps(ix2,jx0);
 254             dy20             = _mm_sub_ps(iy2,jy0);
 255             dz20             = _mm_sub_ps(iz2,jz0);
 256             dx21             = _mm_sub_ps(ix2,jx1);
 257             dy21             = _mm_sub_ps(iy2,jy1);
 258             dz21             = _mm_sub_ps(iz2,jz1);
 259             dx22             = _mm_sub_ps(ix2,jx2);
 260             dy22             = _mm_sub_ps(iy2,jy2);
 261             dz22             = _mm_sub_ps(iz2,jz2);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 265             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 266             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 267             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 268             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 269             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 270             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 271             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 272             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 273
 274             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 275             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 276             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 277             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 278             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 279             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 280             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 281             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 282             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 283
 284             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 285             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 286             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 287             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 288             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 289             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 290             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 291             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 292             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 293
 294             fjx0             = _mm_setzero_ps();
 295             fjy0             = _mm_setzero_ps();
 296             fjz0             = _mm_setzero_ps();
 297             fjx1             = _mm_setzero_ps();
 298             fjy1             = _mm_setzero_ps();
 299             fjz1             = _mm_setzero_ps();
 300             fjx2             = _mm_setzero_ps();
 301             fjy2             = _mm_setzero_ps();
 302             fjz2             = _mm_setzero_ps();
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             r00              = _mm_mul_ps(rsq00,rinv00);
 309
 310             /* Calculate table index by multiplying r with table scale and truncate to integer */
 311             rt               = _mm_mul_ps(r00,vftabscale);
 312             vfitab           = _mm_cvttps_epi32(rt);
 313 #ifdef __XOP__
 314             vfeps            = _mm_frcz_ps(rt);
 315 #else
 316             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 317 #endif
 318             twovfeps         = _mm_add_ps(vfeps,vfeps);
 319             vfitab           = _mm_slli_epi32(vfitab,3);
 320
 321             /* REACTION-FIELD ELECTROSTATICS */
 322             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 323             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 324
 325             /* CUBIC SPLINE TABLE DISPERSION */
 326             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 327             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 328             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 329             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 330             _MM_TRANSPOSE4_PS(Y,F,G,H);
 331             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 332             VV               = _mm_macc_ps(vfeps,Fp,Y);
 333             vvdw6            = _mm_mul_ps(c6_00,VV);
 334             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 335             fvdw6            = _mm_mul_ps(c6_00,FF);
 336
 337             /* CUBIC SPLINE TABLE REPULSION */
 338             vfitab           = _mm_add_epi32(vfitab,ifour);
 339             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 340             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 341             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 342             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 343             _MM_TRANSPOSE4_PS(Y,F,G,H);
 344             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 345             VV               = _mm_macc_ps(vfeps,Fp,Y);
 346             vvdw12           = _mm_mul_ps(c12_00,VV);
 347             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 348             fvdw12           = _mm_mul_ps(c12_00,FF);
 349             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 350             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             velecsum         = _mm_add_ps(velecsum,velec);
 354             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 355
 356             fscal            = _mm_add_ps(felec,fvdw);
 357
 358              /* Update vectorial force */
 359             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 360             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 361             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 362
 363             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 364             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 365             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             /* REACTION-FIELD ELECTROSTATICS */
 372             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 373             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 374
 375             /* Update potential sum for this i atom from the interaction with this j atom. */
 376             velecsum         = _mm_add_ps(velecsum,velec);
 377
 378             fscal            = felec;
 379
 380              /* Update vectorial force */
 381             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 382             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 383             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 384
 385             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 386             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 387             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 388
 389             /**************************
 390              * CALCULATE INTERACTIONS *
 391              **************************/
 392
 393             /* REACTION-FIELD ELECTROSTATICS */
 394             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 395             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velecsum         = _mm_add_ps(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402              /* Update vectorial force */
 403             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 404             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 405             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 406
 407             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 408             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 409             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             /* REACTION-FIELD ELECTROSTATICS */
 416             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 417             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velecsum         = _mm_add_ps(velecsum,velec);
 421
 422             fscal            = felec;
 423
 424              /* Update vectorial force */
 425             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 426             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 427             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 428
 429             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 430             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 431             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 439             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 440
 441             /* Update potential sum for this i atom from the interaction with this j atom. */
 442             velecsum         = _mm_add_ps(velecsum,velec);
 443
 444             fscal            = felec;
 445
 446              /* Update vectorial force */
 447             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 448             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 449             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 450
 451             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 452             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 453             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 454
 455             /**************************
 456              * CALCULATE INTERACTIONS *
 457              **************************/
 458
 459             /* REACTION-FIELD ELECTROSTATICS */
 460             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 461             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 462
 463             /* Update potential sum for this i atom from the interaction with this j atom. */
 464             velecsum         = _mm_add_ps(velecsum,velec);
 465
 466             fscal            = felec;
 467
 468              /* Update vectorial force */
 469             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 470             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 471             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 472
 473             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 474             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 475             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 476
 477             /**************************
 478              * CALCULATE INTERACTIONS *
 479              **************************/
 480
 481             /* REACTION-FIELD ELECTROSTATICS */
 482             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 483             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 484
 485             /* Update potential sum for this i atom from the interaction with this j atom. */
 486             velecsum         = _mm_add_ps(velecsum,velec);
 487
 488             fscal            = felec;
 489
 490              /* Update vectorial force */
 491             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 492             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 493             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 494
 495             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 496             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 497             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 498
 499             /**************************
 500              * CALCULATE INTERACTIONS *
 501              **************************/
 502
 503             /* REACTION-FIELD ELECTROSTATICS */
 504             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 505             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velecsum         = _mm_add_ps(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512              /* Update vectorial force */
 513             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 514             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 515             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 516
 517             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 518             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 519             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 520
 521             /**************************
 522              * CALCULATE INTERACTIONS *
 523              **************************/
 524
 525             /* REACTION-FIELD ELECTROSTATICS */
 526             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 527             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 528
 529             /* Update potential sum for this i atom from the interaction with this j atom. */
 530             velecsum         = _mm_add_ps(velecsum,velec);
 531
 532             fscal            = felec;
 533
 534              /* Update vectorial force */
 535             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 536             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 537             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 538
 539             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 540             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 541             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 542
 543             fjptrA             = f+j_coord_offsetA;
 544             fjptrB             = f+j_coord_offsetB;
 545             fjptrC             = f+j_coord_offsetC;
 546             fjptrD             = f+j_coord_offsetD;
 547
 548             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 549                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 550
 551             /* Inner loop uses 350 flops */
 552         }
 553
 554         if(jidx<j_index_end)
 555         {
 556
 557             /* Get j neighbor index, and coordinate index */
 558             jnrlistA         = jjnr[jidx];
 559             jnrlistB         = jjnr[jidx+1];
 560             jnrlistC         = jjnr[jidx+2];
 561             jnrlistD         = jjnr[jidx+3];
 562             /* Sign of each element will be negative for non-real atoms.
 563              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 564              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 565              */
 566             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 567             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 568             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 569             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 570             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 571             j_coord_offsetA  = DIM*jnrA;
 572             j_coord_offsetB  = DIM*jnrB;
 573             j_coord_offsetC  = DIM*jnrC;
 574             j_coord_offsetD  = DIM*jnrD;
 575
 576             /* load j atom coordinates */
 577             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 578                                               x+j_coord_offsetC,x+j_coord_offsetD,
 579                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 580
 581             /* Calculate displacement vector */
 582             dx00             = _mm_sub_ps(ix0,jx0);
 583             dy00             = _mm_sub_ps(iy0,jy0);
 584             dz00             = _mm_sub_ps(iz0,jz0);
 585             dx01             = _mm_sub_ps(ix0,jx1);
 586             dy01             = _mm_sub_ps(iy0,jy1);
 587             dz01             = _mm_sub_ps(iz0,jz1);
 588             dx02             = _mm_sub_ps(ix0,jx2);
 589             dy02             = _mm_sub_ps(iy0,jy2);
 590             dz02             = _mm_sub_ps(iz0,jz2);
 591             dx10             = _mm_sub_ps(ix1,jx0);
 592             dy10             = _mm_sub_ps(iy1,jy0);
 593             dz10             = _mm_sub_ps(iz1,jz0);
 594             dx11             = _mm_sub_ps(ix1,jx1);
 595             dy11             = _mm_sub_ps(iy1,jy1);
 596             dz11             = _mm_sub_ps(iz1,jz1);
 597             dx12             = _mm_sub_ps(ix1,jx2);
 598             dy12             = _mm_sub_ps(iy1,jy2);
 599             dz12             = _mm_sub_ps(iz1,jz2);
 600             dx20             = _mm_sub_ps(ix2,jx0);
 601             dy20             = _mm_sub_ps(iy2,jy0);
 602             dz20             = _mm_sub_ps(iz2,jz0);
 603             dx21             = _mm_sub_ps(ix2,jx1);
 604             dy21             = _mm_sub_ps(iy2,jy1);
 605             dz21             = _mm_sub_ps(iz2,jz1);
 606             dx22             = _mm_sub_ps(ix2,jx2);
 607             dy22             = _mm_sub_ps(iy2,jy2);
 608             dz22             = _mm_sub_ps(iz2,jz2);
 609
 610             /* Calculate squared distance and things based on it */
 611             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 612             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 613             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 614             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 615             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 616             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 617             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 618             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 619             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 620
 621             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 622             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 623             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 624             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 625             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 626             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 627             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 628             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 629             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 630
 631             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 632             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 633             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 634             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 635             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 636             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 637             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 638             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 639             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 640
 641             fjx0             = _mm_setzero_ps();
 642             fjy0             = _mm_setzero_ps();
 643             fjz0             = _mm_setzero_ps();
 644             fjx1             = _mm_setzero_ps();
 645             fjy1             = _mm_setzero_ps();
 646             fjz1             = _mm_setzero_ps();
 647             fjx2             = _mm_setzero_ps();
 648             fjy2             = _mm_setzero_ps();
 649             fjz2             = _mm_setzero_ps();
 650
 651             /**************************
 652              * CALCULATE INTERACTIONS *
 653              **************************/
 654
 655             r00              = _mm_mul_ps(rsq00,rinv00);
 656             r00              = _mm_andnot_ps(dummy_mask,r00);
 657
 658             /* Calculate table index by multiplying r with table scale and truncate to integer */
 659             rt               = _mm_mul_ps(r00,vftabscale);
 660             vfitab           = _mm_cvttps_epi32(rt);
 661 #ifdef __XOP__
 662             vfeps            = _mm_frcz_ps(rt);
 663 #else
 664             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 665 #endif
 666             twovfeps         = _mm_add_ps(vfeps,vfeps);
 667             vfitab           = _mm_slli_epi32(vfitab,3);
 668
 669             /* REACTION-FIELD ELECTROSTATICS */
 670             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 671             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 672
 673             /* CUBIC SPLINE TABLE DISPERSION */
 674             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 675             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 676             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 677             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 678             _MM_TRANSPOSE4_PS(Y,F,G,H);
 679             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 680             VV               = _mm_macc_ps(vfeps,Fp,Y);
 681             vvdw6            = _mm_mul_ps(c6_00,VV);
 682             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 683             fvdw6            = _mm_mul_ps(c6_00,FF);
 684
 685             /* CUBIC SPLINE TABLE REPULSION */
 686             vfitab           = _mm_add_epi32(vfitab,ifour);
 687             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 688             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 689             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 690             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 691             _MM_TRANSPOSE4_PS(Y,F,G,H);
 692             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 693             VV               = _mm_macc_ps(vfeps,Fp,Y);
 694             vvdw12           = _mm_mul_ps(c12_00,VV);
 695             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 696             fvdw12           = _mm_mul_ps(c12_00,FF);
 697             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 698             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 699
 700             /* Update potential sum for this i atom from the interaction with this j atom. */
 701             velec            = _mm_andnot_ps(dummy_mask,velec);
 702             velecsum         = _mm_add_ps(velecsum,velec);
 703             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 704             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 705
 706             fscal            = _mm_add_ps(felec,fvdw);
 707
 708             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 709
 710              /* Update vectorial force */
 711             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 712             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 713             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 714
 715             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 716             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 717             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 725             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 726
 727             /* Update potential sum for this i atom from the interaction with this j atom. */
 728             velec            = _mm_andnot_ps(dummy_mask,velec);
 729             velecsum         = _mm_add_ps(velecsum,velec);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 734
 735              /* Update vectorial force */
 736             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 737             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 738             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 739
 740             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 741             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 742             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 743
 744             /**************************
 745              * CALCULATE INTERACTIONS *
 746              **************************/
 747
 748             /* REACTION-FIELD ELECTROSTATICS */
 749             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 750             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 751
 752             /* Update potential sum for this i atom from the interaction with this j atom. */
 753             velec            = _mm_andnot_ps(dummy_mask,velec);
 754             velecsum         = _mm_add_ps(velecsum,velec);
 755
 756             fscal            = felec;
 757
 758             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 759
 760              /* Update vectorial force */
 761             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 762             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 763             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 764
 765             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 766             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 767             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* REACTION-FIELD ELECTROSTATICS */
 774             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 775             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_andnot_ps(dummy_mask,velec);
 779             velecsum         = _mm_add_ps(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 784
 785              /* Update vectorial force */
 786             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 787             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 788             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 789
 790             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 791             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 792             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 800             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_andnot_ps(dummy_mask,velec);
 804             velecsum         = _mm_add_ps(velecsum,velec);
 805
 806             fscal            = felec;
 807
 808             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 809
 810              /* Update vectorial force */
 811             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 812             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 813             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 814
 815             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 816             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 817             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             /* REACTION-FIELD ELECTROSTATICS */
 824             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 825             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 826
 827             /* Update potential sum for this i atom from the interaction with this j atom. */
 828             velec            = _mm_andnot_ps(dummy_mask,velec);
 829             velecsum         = _mm_add_ps(velecsum,velec);
 830
 831             fscal            = felec;
 832
 833             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 834
 835              /* Update vectorial force */
 836             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 837             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 838             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 839
 840             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 841             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 842             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* REACTION-FIELD ELECTROSTATICS */
 849             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 850             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 851
 852             /* Update potential sum for this i atom from the interaction with this j atom. */
 853             velec            = _mm_andnot_ps(dummy_mask,velec);
 854             velecsum         = _mm_add_ps(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 859
 860              /* Update vectorial force */
 861             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 862             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 863             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 864
 865             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 866             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 867             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 868
 869             /**************************
 870              * CALCULATE INTERACTIONS *
 871              **************************/
 872
 873             /* REACTION-FIELD ELECTROSTATICS */
 874             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 875             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 876
 877             /* Update potential sum for this i atom from the interaction with this j atom. */
 878             velec            = _mm_andnot_ps(dummy_mask,velec);
 879             velecsum         = _mm_add_ps(velecsum,velec);
 880
 881             fscal            = felec;
 882
 883             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 884
 885              /* Update vectorial force */
 886             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 887             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 888             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 889
 890             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 891             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 892             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             /* REACTION-FIELD ELECTROSTATICS */
 899             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 900             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 901
 902             /* Update potential sum for this i atom from the interaction with this j atom. */
 903             velec            = _mm_andnot_ps(dummy_mask,velec);
 904             velecsum         = _mm_add_ps(velecsum,velec);
 905
 906             fscal            = felec;
 907
 908             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 909
 910              /* Update vectorial force */
 911             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 912             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 913             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 914
 915             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 916             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 917             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 918
 919             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 920             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 921             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 922             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 923
 924             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 925                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 926
 927             /* Inner loop uses 351 flops */
 928         }
 929
 930         /* End of innermost loop */
 931
 932         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 933                                               f+i_coord_offset,fshift+i_shift_offset);
 934
 935         ggid                        = gid[iidx];
 936         /* Update potential energies */
 937         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 938         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 939
 940         /* Increment number of inner iterations */
 941         inneriter                  += j_index_end - j_index_start;
 942
 943         /* Outer loop uses 20 flops */
 944     }
 945
 946     /* Increment number of outer iterations */
 947     outeriter        += nri;
 948
 949     /* Update outer/inner flops */
 950
 951     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*351);
 952 }
 953 /*
 954  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_avx_128_fma_single
 955  * Electrostatics interaction: ReactionField
 956  * VdW interaction:            CubicSplineTable
 957  * Geometry:                   Water3-Water3
 958  * Calculate force/pot:        Force
 959  */
 960 void
 961 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_avx_128_fma_single
 962                     (t_nblist                    * gmx_restrict       nlist,
 963                      rvec                        * gmx_restrict          xx,
 964                      rvec                        * gmx_restrict          ff,
 965                      t_forcerec                  * gmx_restrict          fr,
 966                      t_mdatoms                   * gmx_restrict     mdatoms,
 967                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 968                      t_nrnb                      * gmx_restrict        nrnb)
 969 {
 970     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 971      * just 0 for non-waters.
 972      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 973      * jnr indices corresponding to data put in the four positions in the SIMD register.
 974      */
 975     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 976     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 977     int              jnrA,jnrB,jnrC,jnrD;
 978     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 979     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 980     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 981     real             rcutoff_scalar;
 982     real             *shiftvec,*fshift,*x,*f;
 983     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 984     real             scratch[4*DIM];
 985     __m128           fscal,rcutoff,rcutoff2,jidxall;
 986     int              vdwioffset0;
 987     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 988     int              vdwioffset1;
 989     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 990     int              vdwioffset2;
 991     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 992     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 993     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 994     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 995     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 996     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 997     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 998     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 999     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1000     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1001     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1002     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1003     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1004     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1005     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1006     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1007     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1008     real             *charge;
1009     int              nvdwtype;
1010     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1011     int              *vdwtype;
1012     real             *vdwparam;
1013     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1014     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1015     __m128i          vfitab;
1016     __m128i          ifour       = _mm_set1_epi32(4);
1017     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1018     real             *vftab;
1019     __m128           dummy_mask,cutoff_mask;
1020     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1021     __m128           one     = _mm_set1_ps(1.0);
1022     __m128           two     = _mm_set1_ps(2.0);
1023     x                = xx[0];
1024     f                = ff[0];
1025
1026     nri              = nlist->nri;
1027     iinr             = nlist->iinr;
1028     jindex           = nlist->jindex;
1029     jjnr             = nlist->jjnr;
1030     shiftidx         = nlist->shift;
1031     gid              = nlist->gid;
1032     shiftvec         = fr->shift_vec[0];
1033     fshift           = fr->fshift[0];
1034     facel            = _mm_set1_ps(fr->epsfac);
1035     charge           = mdatoms->chargeA;
1036     krf              = _mm_set1_ps(fr->ic->k_rf);
1037     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1038     crf              = _mm_set1_ps(fr->ic->c_rf);
1039     nvdwtype         = fr->ntype;
1040     vdwparam         = fr->nbfp;
1041     vdwtype          = mdatoms->typeA;
1042
1043     vftab            = kernel_data->table_vdw->data;
1044     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1045
1046     /* Setup water-specific parameters */
1047     inr              = nlist->iinr[0];
1048     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1049     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1050     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1051     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1052
1053     jq0              = _mm_set1_ps(charge[inr+0]);
1054     jq1              = _mm_set1_ps(charge[inr+1]);
1055     jq2              = _mm_set1_ps(charge[inr+2]);
1056     vdwjidx0A        = 2*vdwtype[inr+0];
1057     qq00             = _mm_mul_ps(iq0,jq0);
1058     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1059     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1060     qq01             = _mm_mul_ps(iq0,jq1);
1061     qq02             = _mm_mul_ps(iq0,jq2);
1062     qq10             = _mm_mul_ps(iq1,jq0);
1063     qq11             = _mm_mul_ps(iq1,jq1);
1064     qq12             = _mm_mul_ps(iq1,jq2);
1065     qq20             = _mm_mul_ps(iq2,jq0);
1066     qq21             = _mm_mul_ps(iq2,jq1);
1067     qq22             = _mm_mul_ps(iq2,jq2);
1068
1069     /* Avoid stupid compiler warnings */
1070     jnrA = jnrB = jnrC = jnrD = 0;
1071     j_coord_offsetA = 0;
1072     j_coord_offsetB = 0;
1073     j_coord_offsetC = 0;
1074     j_coord_offsetD = 0;
1075
1076     outeriter        = 0;
1077     inneriter        = 0;
1078
1079     for(iidx=0;iidx<4*DIM;iidx++)
1080     {
1081         scratch[iidx] = 0.0;
1082     }
1083
1084     /* Start outer loop over neighborlists */
1085     for(iidx=0; iidx<nri; iidx++)
1086     {
1087         /* Load shift vector for this list */
1088         i_shift_offset   = DIM*shiftidx[iidx];
1089
1090         /* Load limits for loop over neighbors */
1091         j_index_start    = jindex[iidx];
1092         j_index_end      = jindex[iidx+1];
1093
1094         /* Get outer coordinate index */
1095         inr              = iinr[iidx];
1096         i_coord_offset   = DIM*inr;
1097
1098         /* Load i particle coords and add shift vector */
1099         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1100                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1101
1102         fix0             = _mm_setzero_ps();
1103         fiy0             = _mm_setzero_ps();
1104         fiz0             = _mm_setzero_ps();
1105         fix1             = _mm_setzero_ps();
1106         fiy1             = _mm_setzero_ps();
1107         fiz1             = _mm_setzero_ps();
1108         fix2             = _mm_setzero_ps();
1109         fiy2             = _mm_setzero_ps();
1110         fiz2             = _mm_setzero_ps();
1111
1112         /* Start inner kernel loop */
1113         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1114         {
1115
1116             /* Get j neighbor index, and coordinate index */
1117             jnrA             = jjnr[jidx];
1118             jnrB             = jjnr[jidx+1];
1119             jnrC             = jjnr[jidx+2];
1120             jnrD             = jjnr[jidx+3];
1121             j_coord_offsetA  = DIM*jnrA;
1122             j_coord_offsetB  = DIM*jnrB;
1123             j_coord_offsetC  = DIM*jnrC;
1124             j_coord_offsetD  = DIM*jnrD;
1125
1126             /* load j atom coordinates */
1127             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1128                                               x+j_coord_offsetC,x+j_coord_offsetD,
1129                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1130
1131             /* Calculate displacement vector */
1132             dx00             = _mm_sub_ps(ix0,jx0);
1133             dy00             = _mm_sub_ps(iy0,jy0);
1134             dz00             = _mm_sub_ps(iz0,jz0);
1135             dx01             = _mm_sub_ps(ix0,jx1);
1136             dy01             = _mm_sub_ps(iy0,jy1);
1137             dz01             = _mm_sub_ps(iz0,jz1);
1138             dx02             = _mm_sub_ps(ix0,jx2);
1139             dy02             = _mm_sub_ps(iy0,jy2);
1140             dz02             = _mm_sub_ps(iz0,jz2);
1141             dx10             = _mm_sub_ps(ix1,jx0);
1142             dy10             = _mm_sub_ps(iy1,jy0);
1143             dz10             = _mm_sub_ps(iz1,jz0);
1144             dx11             = _mm_sub_ps(ix1,jx1);
1145             dy11             = _mm_sub_ps(iy1,jy1);
1146             dz11             = _mm_sub_ps(iz1,jz1);
1147             dx12             = _mm_sub_ps(ix1,jx2);
1148             dy12             = _mm_sub_ps(iy1,jy2);
1149             dz12             = _mm_sub_ps(iz1,jz2);
1150             dx20             = _mm_sub_ps(ix2,jx0);
1151             dy20             = _mm_sub_ps(iy2,jy0);
1152             dz20             = _mm_sub_ps(iz2,jz0);
1153             dx21             = _mm_sub_ps(ix2,jx1);
1154             dy21             = _mm_sub_ps(iy2,jy1);
1155             dz21             = _mm_sub_ps(iz2,jz1);
1156             dx22             = _mm_sub_ps(ix2,jx2);
1157             dy22             = _mm_sub_ps(iy2,jy2);
1158             dz22             = _mm_sub_ps(iz2,jz2);
1159
1160             /* Calculate squared distance and things based on it */
1161             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1162             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1163             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1164             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1165             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1166             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1167             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1168             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1169             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1170
1171             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1172             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1173             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1174             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1175             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1176             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1177             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1178             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1179             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1180
1181             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1182             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1183             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1184             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1185             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1186             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1187             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1188             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1189             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1190
1191             fjx0             = _mm_setzero_ps();
1192             fjy0             = _mm_setzero_ps();
1193             fjz0             = _mm_setzero_ps();
1194             fjx1             = _mm_setzero_ps();
1195             fjy1             = _mm_setzero_ps();
1196             fjz1             = _mm_setzero_ps();
1197             fjx2             = _mm_setzero_ps();
1198             fjy2             = _mm_setzero_ps();
1199             fjz2             = _mm_setzero_ps();
1200
1201             /**************************
1202              * CALCULATE INTERACTIONS *
1203              **************************/
1204
1205             r00              = _mm_mul_ps(rsq00,rinv00);
1206
1207             /* Calculate table index by multiplying r with table scale and truncate to integer */
1208             rt               = _mm_mul_ps(r00,vftabscale);
1209             vfitab           = _mm_cvttps_epi32(rt);
1210 #ifdef __XOP__
1211             vfeps            = _mm_frcz_ps(rt);
1212 #else
1213             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1214 #endif
1215             twovfeps         = _mm_add_ps(vfeps,vfeps);
1216             vfitab           = _mm_slli_epi32(vfitab,3);
1217
1218             /* REACTION-FIELD ELECTROSTATICS */
1219             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1220
1221             /* CUBIC SPLINE TABLE DISPERSION */
1222             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1223             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1224             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1225             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1226             _MM_TRANSPOSE4_PS(Y,F,G,H);
1227             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1228             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1229             fvdw6            = _mm_mul_ps(c6_00,FF);
1230
1231             /* CUBIC SPLINE TABLE REPULSION */
1232             vfitab           = _mm_add_epi32(vfitab,ifour);
1233             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1234             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1235             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1236             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1237             _MM_TRANSPOSE4_PS(Y,F,G,H);
1238             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1239             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1240             fvdw12           = _mm_mul_ps(c12_00,FF);
1241             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1242
1243             fscal            = _mm_add_ps(felec,fvdw);
1244
1245              /* Update vectorial force */
1246             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1247             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1248             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1249
1250             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1251             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1252             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1253
1254             /**************************
1255              * CALCULATE INTERACTIONS *
1256              **************************/
1257
1258             /* REACTION-FIELD ELECTROSTATICS */
1259             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1260
1261             fscal            = felec;
1262
1263              /* Update vectorial force */
1264             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1265             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1266             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1267
1268             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1269             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1270             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1271
1272             /**************************
1273              * CALCULATE INTERACTIONS *
1274              **************************/
1275
1276             /* REACTION-FIELD ELECTROSTATICS */
1277             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1278
1279             fscal            = felec;
1280
1281              /* Update vectorial force */
1282             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1283             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1284             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1285
1286             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1287             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1288             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             /* REACTION-FIELD ELECTROSTATICS */
1295             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1296
1297             fscal            = felec;
1298
1299              /* Update vectorial force */
1300             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1301             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1302             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1303
1304             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1305             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1306             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1307
1308             /**************************
1309              * CALCULATE INTERACTIONS *
1310              **************************/
1311
1312             /* REACTION-FIELD ELECTROSTATICS */
1313             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1314
1315             fscal            = felec;
1316
1317              /* Update vectorial force */
1318             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1319             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1320             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1321
1322             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1323             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1324             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1325
1326             /**************************
1327              * CALCULATE INTERACTIONS *
1328              **************************/
1329
1330             /* REACTION-FIELD ELECTROSTATICS */
1331             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1332
1333             fscal            = felec;
1334
1335              /* Update vectorial force */
1336             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1337             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1338             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1339
1340             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1341             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1342             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             /* REACTION-FIELD ELECTROSTATICS */
1349             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1350
1351             fscal            = felec;
1352
1353              /* Update vectorial force */
1354             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1355             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1356             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1357
1358             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1359             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1360             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1361
1362             /**************************
1363              * CALCULATE INTERACTIONS *
1364              **************************/
1365
1366             /* REACTION-FIELD ELECTROSTATICS */
1367             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1368
1369             fscal            = felec;
1370
1371              /* Update vectorial force */
1372             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1373             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1374             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1375
1376             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1377             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1378             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1379
1380             /**************************
1381              * CALCULATE INTERACTIONS *
1382              **************************/
1383
1384             /* REACTION-FIELD ELECTROSTATICS */
1385             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1386
1387             fscal            = felec;
1388
1389              /* Update vectorial force */
1390             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1391             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1392             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1393
1394             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1395             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1396             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1397
1398             fjptrA             = f+j_coord_offsetA;
1399             fjptrB             = f+j_coord_offsetB;
1400             fjptrC             = f+j_coord_offsetC;
1401             fjptrD             = f+j_coord_offsetD;
1402
1403             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1404                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1405
1406             /* Inner loop uses 297 flops */
1407         }
1408
1409         if(jidx<j_index_end)
1410         {
1411
1412             /* Get j neighbor index, and coordinate index */
1413             jnrlistA         = jjnr[jidx];
1414             jnrlistB         = jjnr[jidx+1];
1415             jnrlistC         = jjnr[jidx+2];
1416             jnrlistD         = jjnr[jidx+3];
1417             /* Sign of each element will be negative for non-real atoms.
1418              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1419              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1420              */
1421             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1422             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1423             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1424             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1425             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1426             j_coord_offsetA  = DIM*jnrA;
1427             j_coord_offsetB  = DIM*jnrB;
1428             j_coord_offsetC  = DIM*jnrC;
1429             j_coord_offsetD  = DIM*jnrD;
1430
1431             /* load j atom coordinates */
1432             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1433                                               x+j_coord_offsetC,x+j_coord_offsetD,
1434                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1435
1436             /* Calculate displacement vector */
1437             dx00             = _mm_sub_ps(ix0,jx0);
1438             dy00             = _mm_sub_ps(iy0,jy0);
1439             dz00             = _mm_sub_ps(iz0,jz0);
1440             dx01             = _mm_sub_ps(ix0,jx1);
1441             dy01             = _mm_sub_ps(iy0,jy1);
1442             dz01             = _mm_sub_ps(iz0,jz1);
1443             dx02             = _mm_sub_ps(ix0,jx2);
1444             dy02             = _mm_sub_ps(iy0,jy2);
1445             dz02             = _mm_sub_ps(iz0,jz2);
1446             dx10             = _mm_sub_ps(ix1,jx0);
1447             dy10             = _mm_sub_ps(iy1,jy0);
1448             dz10             = _mm_sub_ps(iz1,jz0);
1449             dx11             = _mm_sub_ps(ix1,jx1);
1450             dy11             = _mm_sub_ps(iy1,jy1);
1451             dz11             = _mm_sub_ps(iz1,jz1);
1452             dx12             = _mm_sub_ps(ix1,jx2);
1453             dy12             = _mm_sub_ps(iy1,jy2);
1454             dz12             = _mm_sub_ps(iz1,jz2);
1455             dx20             = _mm_sub_ps(ix2,jx0);
1456             dy20             = _mm_sub_ps(iy2,jy0);
1457             dz20             = _mm_sub_ps(iz2,jz0);
1458             dx21             = _mm_sub_ps(ix2,jx1);
1459             dy21             = _mm_sub_ps(iy2,jy1);
1460             dz21             = _mm_sub_ps(iz2,jz1);
1461             dx22             = _mm_sub_ps(ix2,jx2);
1462             dy22             = _mm_sub_ps(iy2,jy2);
1463             dz22             = _mm_sub_ps(iz2,jz2);
1464
1465             /* Calculate squared distance and things based on it */
1466             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1467             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1468             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1469             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1470             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1471             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1472             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1473             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1474             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1475
1476             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1477             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1478             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1479             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1480             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1481             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1482             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1483             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1484             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1485
1486             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1487             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1488             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1489             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1490             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1491             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1492             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1493             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1494             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1495
1496             fjx0             = _mm_setzero_ps();
1497             fjy0             = _mm_setzero_ps();
1498             fjz0             = _mm_setzero_ps();
1499             fjx1             = _mm_setzero_ps();
1500             fjy1             = _mm_setzero_ps();
1501             fjz1             = _mm_setzero_ps();
1502             fjx2             = _mm_setzero_ps();
1503             fjy2             = _mm_setzero_ps();
1504             fjz2             = _mm_setzero_ps();
1505
1506             /**************************
1507              * CALCULATE INTERACTIONS *
1508              **************************/
1509
1510             r00              = _mm_mul_ps(rsq00,rinv00);
1511             r00              = _mm_andnot_ps(dummy_mask,r00);
1512
1513             /* Calculate table index by multiplying r with table scale and truncate to integer */
1514             rt               = _mm_mul_ps(r00,vftabscale);
1515             vfitab           = _mm_cvttps_epi32(rt);
1516 #ifdef __XOP__
1517             vfeps            = _mm_frcz_ps(rt);
1518 #else
1519             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1520 #endif
1521             twovfeps         = _mm_add_ps(vfeps,vfeps);
1522             vfitab           = _mm_slli_epi32(vfitab,3);
1523
1524             /* REACTION-FIELD ELECTROSTATICS */
1525             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1526
1527             /* CUBIC SPLINE TABLE DISPERSION */
1528             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1529             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1530             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1531             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1532             _MM_TRANSPOSE4_PS(Y,F,G,H);
1533             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1534             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1535             fvdw6            = _mm_mul_ps(c6_00,FF);
1536
1537             /* CUBIC SPLINE TABLE REPULSION */
1538             vfitab           = _mm_add_epi32(vfitab,ifour);
1539             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1540             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1541             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1542             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1543             _MM_TRANSPOSE4_PS(Y,F,G,H);
1544             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1545             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1546             fvdw12           = _mm_mul_ps(c12_00,FF);
1547             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1548
1549             fscal            = _mm_add_ps(felec,fvdw);
1550
1551             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1552
1553              /* Update vectorial force */
1554             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1555             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1556             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1557
1558             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1559             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1560             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1561
1562             /**************************
1563              * CALCULATE INTERACTIONS *
1564              **************************/
1565
1566             /* REACTION-FIELD ELECTROSTATICS */
1567             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1568
1569             fscal            = felec;
1570
1571             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1572
1573              /* Update vectorial force */
1574             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1575             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1576             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1577
1578             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1579             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1580             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1581
1582             /**************************
1583              * CALCULATE INTERACTIONS *
1584              **************************/
1585
1586             /* REACTION-FIELD ELECTROSTATICS */
1587             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1588
1589             fscal            = felec;
1590
1591             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1592
1593              /* Update vectorial force */
1594             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1595             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1596             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1597
1598             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1599             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1600             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1601
1602             /**************************
1603              * CALCULATE INTERACTIONS *
1604              **************************/
1605
1606             /* REACTION-FIELD ELECTROSTATICS */
1607             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1608
1609             fscal            = felec;
1610
1611             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1612
1613              /* Update vectorial force */
1614             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1615             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1616             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1617
1618             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1619             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1620             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1621
1622             /**************************
1623              * CALCULATE INTERACTIONS *
1624              **************************/
1625
1626             /* REACTION-FIELD ELECTROSTATICS */
1627             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1628
1629             fscal            = felec;
1630
1631             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1632
1633              /* Update vectorial force */
1634             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1635             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1636             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1637
1638             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1639             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1640             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1641
1642             /**************************
1643              * CALCULATE INTERACTIONS *
1644              **************************/
1645
1646             /* REACTION-FIELD ELECTROSTATICS */
1647             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1648
1649             fscal            = felec;
1650
1651             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1652
1653              /* Update vectorial force */
1654             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1655             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1656             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1657
1658             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1659             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1660             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1661
1662             /**************************
1663              * CALCULATE INTERACTIONS *
1664              **************************/
1665
1666             /* REACTION-FIELD ELECTROSTATICS */
1667             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1668
1669             fscal            = felec;
1670
1671             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1672
1673              /* Update vectorial force */
1674             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1675             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1676             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1677
1678             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1679             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1680             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1681
1682             /**************************
1683              * CALCULATE INTERACTIONS *
1684              **************************/
1685
1686             /* REACTION-FIELD ELECTROSTATICS */
1687             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1688
1689             fscal            = felec;
1690
1691             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1692
1693              /* Update vectorial force */
1694             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1695             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1696             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1697
1698             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1699             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1700             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1701
1702             /**************************
1703              * CALCULATE INTERACTIONS *
1704              **************************/
1705
1706             /* REACTION-FIELD ELECTROSTATICS */
1707             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1708
1709             fscal            = felec;
1710
1711             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1712
1713              /* Update vectorial force */
1714             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1715             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1716             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1717
1718             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1719             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1720             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1721
1722             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1723             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1724             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1725             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1726
1727             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1728                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1729
1730             /* Inner loop uses 298 flops */
1731         }
1732
1733         /* End of innermost loop */
1734
1735         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1736                                               f+i_coord_offset,fshift+i_shift_offset);
1737
1738         /* Increment number of inner iterations */
1739         inneriter                  += j_index_end - j_index_start;
1740
1741         /* Outer loop uses 18 flops */
1742     }
1743
1744     /* Increment number of outer iterations */
1745     outeriter        += nri;
1746
1747     /* Update outer/inner flops */
1748
1749     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*298);
1750 }