src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwCSTab_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW3W3_VF_avx_128_fma_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_VdwCSTab_GeomW3W3_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128i          vfitab;
  99     __m128i          ifour       = _mm_set1_epi32(4);
 100     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 101     real             *vftab;
 102     __m128i          ewitab;
 103     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 104     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 105     real             *ewtab;
 106     __m128           dummy_mask,cutoff_mask;
 107     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 108     __m128           one     = _mm_set1_ps(1.0);
 109     __m128           two     = _mm_set1_ps(2.0);
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = _mm_set1_ps(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     vftab            = kernel_data->table_vdw->data;
 128     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 129
 130     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 131     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 132     beta2            = _mm_mul_ps(beta,beta);
 133     beta3            = _mm_mul_ps(beta,beta2);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 136     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 141     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 142     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     jq0              = _mm_set1_ps(charge[inr+0]);
 146     jq1              = _mm_set1_ps(charge[inr+1]);
 147     jq2              = _mm_set1_ps(charge[inr+2]);
 148     vdwjidx0A        = 2*vdwtype[inr+0];
 149     qq00             = _mm_mul_ps(iq0,jq0);
 150     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 151     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 152     qq01             = _mm_mul_ps(iq0,jq1);
 153     qq02             = _mm_mul_ps(iq0,jq2);
 154     qq10             = _mm_mul_ps(iq1,jq0);
 155     qq11             = _mm_mul_ps(iq1,jq1);
 156     qq12             = _mm_mul_ps(iq1,jq2);
 157     qq20             = _mm_mul_ps(iq2,jq0);
 158     qq21             = _mm_mul_ps(iq2,jq1);
 159     qq22             = _mm_mul_ps(iq2,jq2);
 160
 161     /* Avoid stupid compiler warnings */
 162     jnrA = jnrB = jnrC = jnrD = 0;
 163     j_coord_offsetA = 0;
 164     j_coord_offsetB = 0;
 165     j_coord_offsetC = 0;
 166     j_coord_offsetD = 0;
 167
 168     outeriter        = 0;
 169     inneriter        = 0;
 170
 171     for(iidx=0;iidx<4*DIM;iidx++)
 172     {
 173         scratch[iidx] = 0.0;
 174     }
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 193
 194         fix0             = _mm_setzero_ps();
 195         fiy0             = _mm_setzero_ps();
 196         fiz0             = _mm_setzero_ps();
 197         fix1             = _mm_setzero_ps();
 198         fiy1             = _mm_setzero_ps();
 199         fiz1             = _mm_setzero_ps();
 200         fix2             = _mm_setzero_ps();
 201         fiy2             = _mm_setzero_ps();
 202         fiz2             = _mm_setzero_ps();
 203
 204         /* Reset potential sums */
 205         velecsum         = _mm_setzero_ps();
 206         vvdwsum          = _mm_setzero_ps();
 207
 208         /* Start inner kernel loop */
 209         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 210         {
 211
 212             /* Get j neighbor index, and coordinate index */
 213             jnrA             = jjnr[jidx];
 214             jnrB             = jjnr[jidx+1];
 215             jnrC             = jjnr[jidx+2];
 216             jnrD             = jjnr[jidx+3];
 217             j_coord_offsetA  = DIM*jnrA;
 218             j_coord_offsetB  = DIM*jnrB;
 219             j_coord_offsetC  = DIM*jnrC;
 220             j_coord_offsetD  = DIM*jnrD;
 221
 222             /* load j atom coordinates */
 223             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 224                                               x+j_coord_offsetC,x+j_coord_offsetD,
 225                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 226
 227             /* Calculate displacement vector */
 228             dx00             = _mm_sub_ps(ix0,jx0);
 229             dy00             = _mm_sub_ps(iy0,jy0);
 230             dz00             = _mm_sub_ps(iz0,jz0);
 231             dx01             = _mm_sub_ps(ix0,jx1);
 232             dy01             = _mm_sub_ps(iy0,jy1);
 233             dz01             = _mm_sub_ps(iz0,jz1);
 234             dx02             = _mm_sub_ps(ix0,jx2);
 235             dy02             = _mm_sub_ps(iy0,jy2);
 236             dz02             = _mm_sub_ps(iz0,jz2);
 237             dx10             = _mm_sub_ps(ix1,jx0);
 238             dy10             = _mm_sub_ps(iy1,jy0);
 239             dz10             = _mm_sub_ps(iz1,jz0);
 240             dx11             = _mm_sub_ps(ix1,jx1);
 241             dy11             = _mm_sub_ps(iy1,jy1);
 242             dz11             = _mm_sub_ps(iz1,jz1);
 243             dx12             = _mm_sub_ps(ix1,jx2);
 244             dy12             = _mm_sub_ps(iy1,jy2);
 245             dz12             = _mm_sub_ps(iz1,jz2);
 246             dx20             = _mm_sub_ps(ix2,jx0);
 247             dy20             = _mm_sub_ps(iy2,jy0);
 248             dz20             = _mm_sub_ps(iz2,jz0);
 249             dx21             = _mm_sub_ps(ix2,jx1);
 250             dy21             = _mm_sub_ps(iy2,jy1);
 251             dz21             = _mm_sub_ps(iz2,jz1);
 252             dx22             = _mm_sub_ps(ix2,jx2);
 253             dy22             = _mm_sub_ps(iy2,jy2);
 254             dz22             = _mm_sub_ps(iz2,jz2);
 255
 256             /* Calculate squared distance and things based on it */
 257             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 258             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 259             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 260             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 261             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 262             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 263             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 264             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 265             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 266
 267             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 268             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 269             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 270             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 271             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 272             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 273             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 274             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 275             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 276
 277             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 278             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 279             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 280             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 281             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 282             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 283             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 284             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 285             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 286
 287             fjx0             = _mm_setzero_ps();
 288             fjy0             = _mm_setzero_ps();
 289             fjz0             = _mm_setzero_ps();
 290             fjx1             = _mm_setzero_ps();
 291             fjy1             = _mm_setzero_ps();
 292             fjz1             = _mm_setzero_ps();
 293             fjx2             = _mm_setzero_ps();
 294             fjy2             = _mm_setzero_ps();
 295             fjz2             = _mm_setzero_ps();
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             r00              = _mm_mul_ps(rsq00,rinv00);
 302
 303             /* Calculate table index by multiplying r with table scale and truncate to integer */
 304             rt               = _mm_mul_ps(r00,vftabscale);
 305             vfitab           = _mm_cvttps_epi32(rt);
 306 #ifdef __XOP__
 307             vfeps            = _mm_frcz_ps(rt);
 308 #else
 309             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 310 #endif
 311             twovfeps         = _mm_add_ps(vfeps,vfeps);
 312             vfitab           = _mm_slli_epi32(vfitab,3);
 313
 314             /* EWALD ELECTROSTATICS */
 315
 316             /* Analytical PME correction */
 317             zeta2            = _mm_mul_ps(beta2,rsq00);
 318             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 319             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 320             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 321             felec            = _mm_mul_ps(qq00,felec);
 322             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 323             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 324             velec            = _mm_mul_ps(qq00,velec);
 325
 326             /* CUBIC SPLINE TABLE DISPERSION */
 327             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 328             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 329             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 330             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 331             _MM_TRANSPOSE4_PS(Y,F,G,H);
 332             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 333             VV               = _mm_macc_ps(vfeps,Fp,Y);
 334             vvdw6            = _mm_mul_ps(c6_00,VV);
 335             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 336             fvdw6            = _mm_mul_ps(c6_00,FF);
 337
 338             /* CUBIC SPLINE TABLE REPULSION */
 339             vfitab           = _mm_add_epi32(vfitab,ifour);
 340             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 341             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 342             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 343             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 344             _MM_TRANSPOSE4_PS(Y,F,G,H);
 345             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 346             VV               = _mm_macc_ps(vfeps,Fp,Y);
 347             vvdw12           = _mm_mul_ps(c12_00,VV);
 348             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 349             fvdw12           = _mm_mul_ps(c12_00,FF);
 350             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 351             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velecsum         = _mm_add_ps(velecsum,velec);
 355             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 356
 357             fscal            = _mm_add_ps(felec,fvdw);
 358
 359              /* Update vectorial force */
 360             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 361             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 362             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 363
 364             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 365             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 366             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             r01              = _mm_mul_ps(rsq01,rinv01);
 373
 374             /* EWALD ELECTROSTATICS */
 375
 376             /* Analytical PME correction */
 377             zeta2            = _mm_mul_ps(beta2,rsq01);
 378             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 379             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 380             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 381             felec            = _mm_mul_ps(qq01,felec);
 382             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 383             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 384             velec            = _mm_mul_ps(qq01,velec);
 385
 386             /* Update potential sum for this i atom from the interaction with this j atom. */
 387             velecsum         = _mm_add_ps(velecsum,velec);
 388
 389             fscal            = felec;
 390
 391              /* Update vectorial force */
 392             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 393             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 394             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 395
 396             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 397             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 398             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 399
 400             /**************************
 401              * CALCULATE INTERACTIONS *
 402              **************************/
 403
 404             r02              = _mm_mul_ps(rsq02,rinv02);
 405
 406             /* EWALD ELECTROSTATICS */
 407
 408             /* Analytical PME correction */
 409             zeta2            = _mm_mul_ps(beta2,rsq02);
 410             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 411             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 412             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 413             felec            = _mm_mul_ps(qq02,felec);
 414             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 415             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 416             velec            = _mm_mul_ps(qq02,velec);
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velecsum         = _mm_add_ps(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423              /* Update vectorial force */
 424             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 425             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 426             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 427
 428             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 429             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 430             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             r10              = _mm_mul_ps(rsq10,rinv10);
 437
 438             /* EWALD ELECTROSTATICS */
 439
 440             /* Analytical PME correction */
 441             zeta2            = _mm_mul_ps(beta2,rsq10);
 442             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 443             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 444             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 445             felec            = _mm_mul_ps(qq10,felec);
 446             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 447             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 448             velec            = _mm_mul_ps(qq10,velec);
 449
 450             /* Update potential sum for this i atom from the interaction with this j atom. */
 451             velecsum         = _mm_add_ps(velecsum,velec);
 452
 453             fscal            = felec;
 454
 455              /* Update vectorial force */
 456             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 457             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 458             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 459
 460             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 461             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 462             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             r11              = _mm_mul_ps(rsq11,rinv11);
 469
 470             /* EWALD ELECTROSTATICS */
 471
 472             /* Analytical PME correction */
 473             zeta2            = _mm_mul_ps(beta2,rsq11);
 474             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 475             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 476             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 477             felec            = _mm_mul_ps(qq11,felec);
 478             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 479             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 480             velec            = _mm_mul_ps(qq11,velec);
 481
 482             /* Update potential sum for this i atom from the interaction with this j atom. */
 483             velecsum         = _mm_add_ps(velecsum,velec);
 484
 485             fscal            = felec;
 486
 487              /* Update vectorial force */
 488             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 489             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 490             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 491
 492             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 493             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 494             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 495
 496             /**************************
 497              * CALCULATE INTERACTIONS *
 498              **************************/
 499
 500             r12              = _mm_mul_ps(rsq12,rinv12);
 501
 502             /* EWALD ELECTROSTATICS */
 503
 504             /* Analytical PME correction */
 505             zeta2            = _mm_mul_ps(beta2,rsq12);
 506             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 507             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 508             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 509             felec            = _mm_mul_ps(qq12,felec);
 510             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 511             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 512             velec            = _mm_mul_ps(qq12,velec);
 513
 514             /* Update potential sum for this i atom from the interaction with this j atom. */
 515             velecsum         = _mm_add_ps(velecsum,velec);
 516
 517             fscal            = felec;
 518
 519              /* Update vectorial force */
 520             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 521             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 522             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 523
 524             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 525             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 526             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             r20              = _mm_mul_ps(rsq20,rinv20);
 533
 534             /* EWALD ELECTROSTATICS */
 535
 536             /* Analytical PME correction */
 537             zeta2            = _mm_mul_ps(beta2,rsq20);
 538             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 539             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 540             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 541             felec            = _mm_mul_ps(qq20,felec);
 542             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 543             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 544             velec            = _mm_mul_ps(qq20,velec);
 545
 546             /* Update potential sum for this i atom from the interaction with this j atom. */
 547             velecsum         = _mm_add_ps(velecsum,velec);
 548
 549             fscal            = felec;
 550
 551              /* Update vectorial force */
 552             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 553             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 554             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 555
 556             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 557             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 558             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 559
 560             /**************************
 561              * CALCULATE INTERACTIONS *
 562              **************************/
 563
 564             r21              = _mm_mul_ps(rsq21,rinv21);
 565
 566             /* EWALD ELECTROSTATICS */
 567
 568             /* Analytical PME correction */
 569             zeta2            = _mm_mul_ps(beta2,rsq21);
 570             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 571             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 572             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 573             felec            = _mm_mul_ps(qq21,felec);
 574             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 575             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 576             velec            = _mm_mul_ps(qq21,velec);
 577
 578             /* Update potential sum for this i atom from the interaction with this j atom. */
 579             velecsum         = _mm_add_ps(velecsum,velec);
 580
 581             fscal            = felec;
 582
 583              /* Update vectorial force */
 584             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 585             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 586             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 587
 588             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 589             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 590             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             r22              = _mm_mul_ps(rsq22,rinv22);
 597
 598             /* EWALD ELECTROSTATICS */
 599
 600             /* Analytical PME correction */
 601             zeta2            = _mm_mul_ps(beta2,rsq22);
 602             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 603             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 604             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 605             felec            = _mm_mul_ps(qq22,felec);
 606             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 607             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 608             velec            = _mm_mul_ps(qq22,velec);
 609
 610             /* Update potential sum for this i atom from the interaction with this j atom. */
 611             velecsum         = _mm_add_ps(velecsum,velec);
 612
 613             fscal            = felec;
 614
 615              /* Update vectorial force */
 616             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 617             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 618             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 619
 620             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 621             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 622             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 623
 624             fjptrA             = f+j_coord_offsetA;
 625             fjptrB             = f+j_coord_offsetB;
 626             fjptrC             = f+j_coord_offsetC;
 627             fjptrD             = f+j_coord_offsetD;
 628
 629             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 630                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 631
 632             /* Inner loop uses 295 flops */
 633         }
 634
 635         if(jidx<j_index_end)
 636         {
 637
 638             /* Get j neighbor index, and coordinate index */
 639             jnrlistA         = jjnr[jidx];
 640             jnrlistB         = jjnr[jidx+1];
 641             jnrlistC         = jjnr[jidx+2];
 642             jnrlistD         = jjnr[jidx+3];
 643             /* Sign of each element will be negative for non-real atoms.
 644              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 645              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 646              */
 647             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 648             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 649             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 650             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 651             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 652             j_coord_offsetA  = DIM*jnrA;
 653             j_coord_offsetB  = DIM*jnrB;
 654             j_coord_offsetC  = DIM*jnrC;
 655             j_coord_offsetD  = DIM*jnrD;
 656
 657             /* load j atom coordinates */
 658             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 659                                               x+j_coord_offsetC,x+j_coord_offsetD,
 660                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 661
 662             /* Calculate displacement vector */
 663             dx00             = _mm_sub_ps(ix0,jx0);
 664             dy00             = _mm_sub_ps(iy0,jy0);
 665             dz00             = _mm_sub_ps(iz0,jz0);
 666             dx01             = _mm_sub_ps(ix0,jx1);
 667             dy01             = _mm_sub_ps(iy0,jy1);
 668             dz01             = _mm_sub_ps(iz0,jz1);
 669             dx02             = _mm_sub_ps(ix0,jx2);
 670             dy02             = _mm_sub_ps(iy0,jy2);
 671             dz02             = _mm_sub_ps(iz0,jz2);
 672             dx10             = _mm_sub_ps(ix1,jx0);
 673             dy10             = _mm_sub_ps(iy1,jy0);
 674             dz10             = _mm_sub_ps(iz1,jz0);
 675             dx11             = _mm_sub_ps(ix1,jx1);
 676             dy11             = _mm_sub_ps(iy1,jy1);
 677             dz11             = _mm_sub_ps(iz1,jz1);
 678             dx12             = _mm_sub_ps(ix1,jx2);
 679             dy12             = _mm_sub_ps(iy1,jy2);
 680             dz12             = _mm_sub_ps(iz1,jz2);
 681             dx20             = _mm_sub_ps(ix2,jx0);
 682             dy20             = _mm_sub_ps(iy2,jy0);
 683             dz20             = _mm_sub_ps(iz2,jz0);
 684             dx21             = _mm_sub_ps(ix2,jx1);
 685             dy21             = _mm_sub_ps(iy2,jy1);
 686             dz21             = _mm_sub_ps(iz2,jz1);
 687             dx22             = _mm_sub_ps(ix2,jx2);
 688             dy22             = _mm_sub_ps(iy2,jy2);
 689             dz22             = _mm_sub_ps(iz2,jz2);
 690
 691             /* Calculate squared distance and things based on it */
 692             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 693             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 694             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 695             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 696             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 697             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 698             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 699             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 700             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 701
 702             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 703             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 704             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 705             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 706             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 707             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 708             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 709             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 710             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 711
 712             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 713             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 714             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 715             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 716             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 717             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 718             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 719             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 720             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 721
 722             fjx0             = _mm_setzero_ps();
 723             fjy0             = _mm_setzero_ps();
 724             fjz0             = _mm_setzero_ps();
 725             fjx1             = _mm_setzero_ps();
 726             fjy1             = _mm_setzero_ps();
 727             fjz1             = _mm_setzero_ps();
 728             fjx2             = _mm_setzero_ps();
 729             fjy2             = _mm_setzero_ps();
 730             fjz2             = _mm_setzero_ps();
 731
 732             /**************************
 733              * CALCULATE INTERACTIONS *
 734              **************************/
 735
 736             r00              = _mm_mul_ps(rsq00,rinv00);
 737             r00              = _mm_andnot_ps(dummy_mask,r00);
 738
 739             /* Calculate table index by multiplying r with table scale and truncate to integer */
 740             rt               = _mm_mul_ps(r00,vftabscale);
 741             vfitab           = _mm_cvttps_epi32(rt);
 742 #ifdef __XOP__
 743             vfeps            = _mm_frcz_ps(rt);
 744 #else
 745             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 746 #endif
 747             twovfeps         = _mm_add_ps(vfeps,vfeps);
 748             vfitab           = _mm_slli_epi32(vfitab,3);
 749
 750             /* EWALD ELECTROSTATICS */
 751
 752             /* Analytical PME correction */
 753             zeta2            = _mm_mul_ps(beta2,rsq00);
 754             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 755             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 756             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 757             felec            = _mm_mul_ps(qq00,felec);
 758             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 759             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 760             velec            = _mm_mul_ps(qq00,velec);
 761
 762             /* CUBIC SPLINE TABLE DISPERSION */
 763             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 764             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 765             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 766             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 767             _MM_TRANSPOSE4_PS(Y,F,G,H);
 768             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 769             VV               = _mm_macc_ps(vfeps,Fp,Y);
 770             vvdw6            = _mm_mul_ps(c6_00,VV);
 771             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 772             fvdw6            = _mm_mul_ps(c6_00,FF);
 773
 774             /* CUBIC SPLINE TABLE REPULSION */
 775             vfitab           = _mm_add_epi32(vfitab,ifour);
 776             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 777             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 778             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 779             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 780             _MM_TRANSPOSE4_PS(Y,F,G,H);
 781             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 782             VV               = _mm_macc_ps(vfeps,Fp,Y);
 783             vvdw12           = _mm_mul_ps(c12_00,VV);
 784             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 785             fvdw12           = _mm_mul_ps(c12_00,FF);
 786             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 787             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 788
 789             /* Update potential sum for this i atom from the interaction with this j atom. */
 790             velec            = _mm_andnot_ps(dummy_mask,velec);
 791             velecsum         = _mm_add_ps(velecsum,velec);
 792             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 793             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 794
 795             fscal            = _mm_add_ps(felec,fvdw);
 796
 797             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 798
 799              /* Update vectorial force */
 800             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 801             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 802             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 803
 804             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 805             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 806             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 807
 808             /**************************
 809              * CALCULATE INTERACTIONS *
 810              **************************/
 811
 812             r01              = _mm_mul_ps(rsq01,rinv01);
 813             r01              = _mm_andnot_ps(dummy_mask,r01);
 814
 815             /* EWALD ELECTROSTATICS */
 816
 817             /* Analytical PME correction */
 818             zeta2            = _mm_mul_ps(beta2,rsq01);
 819             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 820             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 821             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 822             felec            = _mm_mul_ps(qq01,felec);
 823             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 824             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 825             velec            = _mm_mul_ps(qq01,velec);
 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             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 837             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 838             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 839
 840             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 841             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 842             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             r02              = _mm_mul_ps(rsq02,rinv02);
 849             r02              = _mm_andnot_ps(dummy_mask,r02);
 850
 851             /* EWALD ELECTROSTATICS */
 852
 853             /* Analytical PME correction */
 854             zeta2            = _mm_mul_ps(beta2,rsq02);
 855             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 856             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 857             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 858             felec            = _mm_mul_ps(qq02,felec);
 859             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 860             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 861             velec            = _mm_mul_ps(qq02,velec);
 862
 863             /* Update potential sum for this i atom from the interaction with this j atom. */
 864             velec            = _mm_andnot_ps(dummy_mask,velec);
 865             velecsum         = _mm_add_ps(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 870
 871              /* Update vectorial force */
 872             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 873             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 874             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 875
 876             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 877             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 878             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 879
 880             /**************************
 881              * CALCULATE INTERACTIONS *
 882              **************************/
 883
 884             r10              = _mm_mul_ps(rsq10,rinv10);
 885             r10              = _mm_andnot_ps(dummy_mask,r10);
 886
 887             /* EWALD ELECTROSTATICS */
 888
 889             /* Analytical PME correction */
 890             zeta2            = _mm_mul_ps(beta2,rsq10);
 891             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 892             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 893             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 894             felec            = _mm_mul_ps(qq10,felec);
 895             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 896             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 897             velec            = _mm_mul_ps(qq10,velec);
 898
 899             /* Update potential sum for this i atom from the interaction with this j atom. */
 900             velec            = _mm_andnot_ps(dummy_mask,velec);
 901             velecsum         = _mm_add_ps(velecsum,velec);
 902
 903             fscal            = felec;
 904
 905             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 906
 907              /* Update vectorial force */
 908             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 909             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 910             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 911
 912             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 913             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 914             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 915
 916             /**************************
 917              * CALCULATE INTERACTIONS *
 918              **************************/
 919
 920             r11              = _mm_mul_ps(rsq11,rinv11);
 921             r11              = _mm_andnot_ps(dummy_mask,r11);
 922
 923             /* EWALD ELECTROSTATICS */
 924
 925             /* Analytical PME correction */
 926             zeta2            = _mm_mul_ps(beta2,rsq11);
 927             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 928             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 929             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 930             felec            = _mm_mul_ps(qq11,felec);
 931             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 932             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 933             velec            = _mm_mul_ps(qq11,velec);
 934
 935             /* Update potential sum for this i atom from the interaction with this j atom. */
 936             velec            = _mm_andnot_ps(dummy_mask,velec);
 937             velecsum         = _mm_add_ps(velecsum,velec);
 938
 939             fscal            = felec;
 940
 941             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 942
 943              /* Update vectorial force */
 944             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 945             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 946             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 947
 948             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 949             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 950             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 951
 952             /**************************
 953              * CALCULATE INTERACTIONS *
 954              **************************/
 955
 956             r12              = _mm_mul_ps(rsq12,rinv12);
 957             r12              = _mm_andnot_ps(dummy_mask,r12);
 958
 959             /* EWALD ELECTROSTATICS */
 960
 961             /* Analytical PME correction */
 962             zeta2            = _mm_mul_ps(beta2,rsq12);
 963             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 964             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 965             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 966             felec            = _mm_mul_ps(qq12,felec);
 967             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 968             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 969             velec            = _mm_mul_ps(qq12,velec);
 970
 971             /* Update potential sum for this i atom from the interaction with this j atom. */
 972             velec            = _mm_andnot_ps(dummy_mask,velec);
 973             velecsum         = _mm_add_ps(velecsum,velec);
 974
 975             fscal            = felec;
 976
 977             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 978
 979              /* Update vectorial force */
 980             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 981             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 982             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 983
 984             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 985             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 986             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 987
 988             /**************************
 989              * CALCULATE INTERACTIONS *
 990              **************************/
 991
 992             r20              = _mm_mul_ps(rsq20,rinv20);
 993             r20              = _mm_andnot_ps(dummy_mask,r20);
 994
 995             /* EWALD ELECTROSTATICS */
 996
 997             /* Analytical PME correction */
 998             zeta2            = _mm_mul_ps(beta2,rsq20);
 999             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1000             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1001             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1002             felec            = _mm_mul_ps(qq20,felec);
1003             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1004             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
1005             velec            = _mm_mul_ps(qq20,velec);
1006
1007             /* Update potential sum for this i atom from the interaction with this j atom. */
1008             velec            = _mm_andnot_ps(dummy_mask,velec);
1009             velecsum         = _mm_add_ps(velecsum,velec);
1010
1011             fscal            = felec;
1012
1013             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1014
1015              /* Update vectorial force */
1016             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1017             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1018             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1019
1020             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1021             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1022             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1023
1024             /**************************
1025              * CALCULATE INTERACTIONS *
1026              **************************/
1027
1028             r21              = _mm_mul_ps(rsq21,rinv21);
1029             r21              = _mm_andnot_ps(dummy_mask,r21);
1030
1031             /* EWALD ELECTROSTATICS */
1032
1033             /* Analytical PME correction */
1034             zeta2            = _mm_mul_ps(beta2,rsq21);
1035             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1036             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1037             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1038             felec            = _mm_mul_ps(qq21,felec);
1039             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1040             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
1041             velec            = _mm_mul_ps(qq21,velec);
1042
1043             /* Update potential sum for this i atom from the interaction with this j atom. */
1044             velec            = _mm_andnot_ps(dummy_mask,velec);
1045             velecsum         = _mm_add_ps(velecsum,velec);
1046
1047             fscal            = felec;
1048
1049             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1050
1051              /* Update vectorial force */
1052             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1053             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1054             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1055
1056             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1057             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1058             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1059
1060             /**************************
1061              * CALCULATE INTERACTIONS *
1062              **************************/
1063
1064             r22              = _mm_mul_ps(rsq22,rinv22);
1065             r22              = _mm_andnot_ps(dummy_mask,r22);
1066
1067             /* EWALD ELECTROSTATICS */
1068
1069             /* Analytical PME correction */
1070             zeta2            = _mm_mul_ps(beta2,rsq22);
1071             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1072             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1073             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1074             felec            = _mm_mul_ps(qq22,felec);
1075             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1076             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
1077             velec            = _mm_mul_ps(qq22,velec);
1078
1079             /* Update potential sum for this i atom from the interaction with this j atom. */
1080             velec            = _mm_andnot_ps(dummy_mask,velec);
1081             velecsum         = _mm_add_ps(velecsum,velec);
1082
1083             fscal            = felec;
1084
1085             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1086
1087              /* Update vectorial force */
1088             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1089             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1090             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1091
1092             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1093             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1094             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1095
1096             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1097             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1098             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1099             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1100
1101             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1102                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1103
1104             /* Inner loop uses 304 flops */
1105         }
1106
1107         /* End of innermost loop */
1108
1109         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1110                                               f+i_coord_offset,fshift+i_shift_offset);
1111
1112         ggid                        = gid[iidx];
1113         /* Update potential energies */
1114         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1115         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1116
1117         /* Increment number of inner iterations */
1118         inneriter                  += j_index_end - j_index_start;
1119
1120         /* Outer loop uses 20 flops */
1121     }
1122
1123     /* Increment number of outer iterations */
1124     outeriter        += nri;
1125
1126     /* Update outer/inner flops */
1127
1128     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*304);
1129 }
1130 /*
1131  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW3W3_F_avx_128_fma_single
1132  * Electrostatics interaction: Ewald
1133  * VdW interaction:            CubicSplineTable
1134  * Geometry:                   Water3-Water3
1135  * Calculate force/pot:        Force
1136  */
1137 void
1138 nb_kernel_ElecEw_VdwCSTab_GeomW3W3_F_avx_128_fma_single
1139                     (t_nblist * gmx_restrict                nlist,
1140                      rvec * gmx_restrict                    xx,
1141                      rvec * gmx_restrict                    ff,
1142                      t_forcerec * gmx_restrict              fr,
1143                      t_mdatoms * gmx_restrict               mdatoms,
1144                      nb_kernel_data_t * gmx_restrict        kernel_data,
1145                      t_nrnb * gmx_restrict                  nrnb)
1146 {
1147     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1148      * just 0 for non-waters.
1149      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1150      * jnr indices corresponding to data put in the four positions in the SIMD register.
1151      */
1152     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1153     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1154     int              jnrA,jnrB,jnrC,jnrD;
1155     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1156     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1157     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1158     real             rcutoff_scalar;
1159     real             *shiftvec,*fshift,*x,*f;
1160     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1161     real             scratch[4*DIM];
1162     __m128           fscal,rcutoff,rcutoff2,jidxall;
1163     int              vdwioffset0;
1164     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1165     int              vdwioffset1;
1166     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1167     int              vdwioffset2;
1168     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1169     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1170     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1171     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1172     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1173     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1174     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1175     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1176     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1177     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1178     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1179     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1180     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1181     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1182     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1183     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1184     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1185     real             *charge;
1186     int              nvdwtype;
1187     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1188     int              *vdwtype;
1189     real             *vdwparam;
1190     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1191     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1192     __m128i          vfitab;
1193     __m128i          ifour       = _mm_set1_epi32(4);
1194     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1195     real             *vftab;
1196     __m128i          ewitab;
1197     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1198     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1199     real             *ewtab;
1200     __m128           dummy_mask,cutoff_mask;
1201     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1202     __m128           one     = _mm_set1_ps(1.0);
1203     __m128           two     = _mm_set1_ps(2.0);
1204     x                = xx[0];
1205     f                = ff[0];
1206
1207     nri              = nlist->nri;
1208     iinr             = nlist->iinr;
1209     jindex           = nlist->jindex;
1210     jjnr             = nlist->jjnr;
1211     shiftidx         = nlist->shift;
1212     gid              = nlist->gid;
1213     shiftvec         = fr->shift_vec[0];
1214     fshift           = fr->fshift[0];
1215     facel            = _mm_set1_ps(fr->epsfac);
1216     charge           = mdatoms->chargeA;
1217     nvdwtype         = fr->ntype;
1218     vdwparam         = fr->nbfp;
1219     vdwtype          = mdatoms->typeA;
1220
1221     vftab            = kernel_data->table_vdw->data;
1222     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1223
1224     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1225     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
1226     beta2            = _mm_mul_ps(beta,beta);
1227     beta3            = _mm_mul_ps(beta,beta2);
1228     ewtab            = fr->ic->tabq_coul_F;
1229     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1230     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1231
1232     /* Setup water-specific parameters */
1233     inr              = nlist->iinr[0];
1234     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1235     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1236     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1237     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1238
1239     jq0              = _mm_set1_ps(charge[inr+0]);
1240     jq1              = _mm_set1_ps(charge[inr+1]);
1241     jq2              = _mm_set1_ps(charge[inr+2]);
1242     vdwjidx0A        = 2*vdwtype[inr+0];
1243     qq00             = _mm_mul_ps(iq0,jq0);
1244     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1245     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1246     qq01             = _mm_mul_ps(iq0,jq1);
1247     qq02             = _mm_mul_ps(iq0,jq2);
1248     qq10             = _mm_mul_ps(iq1,jq0);
1249     qq11             = _mm_mul_ps(iq1,jq1);
1250     qq12             = _mm_mul_ps(iq1,jq2);
1251     qq20             = _mm_mul_ps(iq2,jq0);
1252     qq21             = _mm_mul_ps(iq2,jq1);
1253     qq22             = _mm_mul_ps(iq2,jq2);
1254
1255     /* Avoid stupid compiler warnings */
1256     jnrA = jnrB = jnrC = jnrD = 0;
1257     j_coord_offsetA = 0;
1258     j_coord_offsetB = 0;
1259     j_coord_offsetC = 0;
1260     j_coord_offsetD = 0;
1261
1262     outeriter        = 0;
1263     inneriter        = 0;
1264
1265     for(iidx=0;iidx<4*DIM;iidx++)
1266     {
1267         scratch[iidx] = 0.0;
1268     }
1269
1270     /* Start outer loop over neighborlists */
1271     for(iidx=0; iidx<nri; iidx++)
1272     {
1273         /* Load shift vector for this list */
1274         i_shift_offset   = DIM*shiftidx[iidx];
1275
1276         /* Load limits for loop over neighbors */
1277         j_index_start    = jindex[iidx];
1278         j_index_end      = jindex[iidx+1];
1279
1280         /* Get outer coordinate index */
1281         inr              = iinr[iidx];
1282         i_coord_offset   = DIM*inr;
1283
1284         /* Load i particle coords and add shift vector */
1285         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1286                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1287
1288         fix0             = _mm_setzero_ps();
1289         fiy0             = _mm_setzero_ps();
1290         fiz0             = _mm_setzero_ps();
1291         fix1             = _mm_setzero_ps();
1292         fiy1             = _mm_setzero_ps();
1293         fiz1             = _mm_setzero_ps();
1294         fix2             = _mm_setzero_ps();
1295         fiy2             = _mm_setzero_ps();
1296         fiz2             = _mm_setzero_ps();
1297
1298         /* Start inner kernel loop */
1299         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1300         {
1301
1302             /* Get j neighbor index, and coordinate index */
1303             jnrA             = jjnr[jidx];
1304             jnrB             = jjnr[jidx+1];
1305             jnrC             = jjnr[jidx+2];
1306             jnrD             = jjnr[jidx+3];
1307             j_coord_offsetA  = DIM*jnrA;
1308             j_coord_offsetB  = DIM*jnrB;
1309             j_coord_offsetC  = DIM*jnrC;
1310             j_coord_offsetD  = DIM*jnrD;
1311
1312             /* load j atom coordinates */
1313             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1314                                               x+j_coord_offsetC,x+j_coord_offsetD,
1315                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1316
1317             /* Calculate displacement vector */
1318             dx00             = _mm_sub_ps(ix0,jx0);
1319             dy00             = _mm_sub_ps(iy0,jy0);
1320             dz00             = _mm_sub_ps(iz0,jz0);
1321             dx01             = _mm_sub_ps(ix0,jx1);
1322             dy01             = _mm_sub_ps(iy0,jy1);
1323             dz01             = _mm_sub_ps(iz0,jz1);
1324             dx02             = _mm_sub_ps(ix0,jx2);
1325             dy02             = _mm_sub_ps(iy0,jy2);
1326             dz02             = _mm_sub_ps(iz0,jz2);
1327             dx10             = _mm_sub_ps(ix1,jx0);
1328             dy10             = _mm_sub_ps(iy1,jy0);
1329             dz10             = _mm_sub_ps(iz1,jz0);
1330             dx11             = _mm_sub_ps(ix1,jx1);
1331             dy11             = _mm_sub_ps(iy1,jy1);
1332             dz11             = _mm_sub_ps(iz1,jz1);
1333             dx12             = _mm_sub_ps(ix1,jx2);
1334             dy12             = _mm_sub_ps(iy1,jy2);
1335             dz12             = _mm_sub_ps(iz1,jz2);
1336             dx20             = _mm_sub_ps(ix2,jx0);
1337             dy20             = _mm_sub_ps(iy2,jy0);
1338             dz20             = _mm_sub_ps(iz2,jz0);
1339             dx21             = _mm_sub_ps(ix2,jx1);
1340             dy21             = _mm_sub_ps(iy2,jy1);
1341             dz21             = _mm_sub_ps(iz2,jz1);
1342             dx22             = _mm_sub_ps(ix2,jx2);
1343             dy22             = _mm_sub_ps(iy2,jy2);
1344             dz22             = _mm_sub_ps(iz2,jz2);
1345
1346             /* Calculate squared distance and things based on it */
1347             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1348             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1349             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1350             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1351             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1352             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1353             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1354             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1355             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1356
1357             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1358             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1359             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1360             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1361             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1362             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1363             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1364             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1365             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1366
1367             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1368             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1369             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1370             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1371             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1372             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1373             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1374             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1375             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1376
1377             fjx0             = _mm_setzero_ps();
1378             fjy0             = _mm_setzero_ps();
1379             fjz0             = _mm_setzero_ps();
1380             fjx1             = _mm_setzero_ps();
1381             fjy1             = _mm_setzero_ps();
1382             fjz1             = _mm_setzero_ps();
1383             fjx2             = _mm_setzero_ps();
1384             fjy2             = _mm_setzero_ps();
1385             fjz2             = _mm_setzero_ps();
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             r00              = _mm_mul_ps(rsq00,rinv00);
1392
1393             /* Calculate table index by multiplying r with table scale and truncate to integer */
1394             rt               = _mm_mul_ps(r00,vftabscale);
1395             vfitab           = _mm_cvttps_epi32(rt);
1396 #ifdef __XOP__
1397             vfeps            = _mm_frcz_ps(rt);
1398 #else
1399             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1400 #endif
1401             twovfeps         = _mm_add_ps(vfeps,vfeps);
1402             vfitab           = _mm_slli_epi32(vfitab,3);
1403
1404             /* EWALD ELECTROSTATICS */
1405
1406             /* Analytical PME correction */
1407             zeta2            = _mm_mul_ps(beta2,rsq00);
1408             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1409             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1410             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1411             felec            = _mm_mul_ps(qq00,felec);
1412
1413             /* CUBIC SPLINE TABLE DISPERSION */
1414             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1415             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1416             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1417             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1418             _MM_TRANSPOSE4_PS(Y,F,G,H);
1419             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1420             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1421             fvdw6            = _mm_mul_ps(c6_00,FF);
1422
1423             /* CUBIC SPLINE TABLE REPULSION */
1424             vfitab           = _mm_add_epi32(vfitab,ifour);
1425             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1426             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1427             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1428             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1429             _MM_TRANSPOSE4_PS(Y,F,G,H);
1430             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1431             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1432             fvdw12           = _mm_mul_ps(c12_00,FF);
1433             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1434
1435             fscal            = _mm_add_ps(felec,fvdw);
1436
1437              /* Update vectorial force */
1438             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1439             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1440             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1441
1442             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1443             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1444             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1445
1446             /**************************
1447              * CALCULATE INTERACTIONS *
1448              **************************/
1449
1450             r01              = _mm_mul_ps(rsq01,rinv01);
1451
1452             /* EWALD ELECTROSTATICS */
1453
1454             /* Analytical PME correction */
1455             zeta2            = _mm_mul_ps(beta2,rsq01);
1456             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1457             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1458             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1459             felec            = _mm_mul_ps(qq01,felec);
1460
1461             fscal            = felec;
1462
1463              /* Update vectorial force */
1464             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1465             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1466             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1467
1468             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1469             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1470             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1471
1472             /**************************
1473              * CALCULATE INTERACTIONS *
1474              **************************/
1475
1476             r02              = _mm_mul_ps(rsq02,rinv02);
1477
1478             /* EWALD ELECTROSTATICS */
1479
1480             /* Analytical PME correction */
1481             zeta2            = _mm_mul_ps(beta2,rsq02);
1482             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1483             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1484             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1485             felec            = _mm_mul_ps(qq02,felec);
1486
1487             fscal            = felec;
1488
1489              /* Update vectorial force */
1490             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1491             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1492             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1493
1494             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1495             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1496             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1497
1498             /**************************
1499              * CALCULATE INTERACTIONS *
1500              **************************/
1501
1502             r10              = _mm_mul_ps(rsq10,rinv10);
1503
1504             /* EWALD ELECTROSTATICS */
1505
1506             /* Analytical PME correction */
1507             zeta2            = _mm_mul_ps(beta2,rsq10);
1508             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1509             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1510             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1511             felec            = _mm_mul_ps(qq10,felec);
1512
1513             fscal            = felec;
1514
1515              /* Update vectorial force */
1516             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1517             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1518             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1519
1520             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1521             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1522             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1523
1524             /**************************
1525              * CALCULATE INTERACTIONS *
1526              **************************/
1527
1528             r11              = _mm_mul_ps(rsq11,rinv11);
1529
1530             /* EWALD ELECTROSTATICS */
1531
1532             /* Analytical PME correction */
1533             zeta2            = _mm_mul_ps(beta2,rsq11);
1534             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1535             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1536             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1537             felec            = _mm_mul_ps(qq11,felec);
1538
1539             fscal            = felec;
1540
1541              /* Update vectorial force */
1542             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1543             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1544             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1545
1546             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1547             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1548             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1549
1550             /**************************
1551              * CALCULATE INTERACTIONS *
1552              **************************/
1553
1554             r12              = _mm_mul_ps(rsq12,rinv12);
1555
1556             /* EWALD ELECTROSTATICS */
1557
1558             /* Analytical PME correction */
1559             zeta2            = _mm_mul_ps(beta2,rsq12);
1560             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1561             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1562             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1563             felec            = _mm_mul_ps(qq12,felec);
1564
1565             fscal            = felec;
1566
1567              /* Update vectorial force */
1568             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1569             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1570             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1571
1572             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1573             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1574             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1575
1576             /**************************
1577              * CALCULATE INTERACTIONS *
1578              **************************/
1579
1580             r20              = _mm_mul_ps(rsq20,rinv20);
1581
1582             /* EWALD ELECTROSTATICS */
1583
1584             /* Analytical PME correction */
1585             zeta2            = _mm_mul_ps(beta2,rsq20);
1586             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1587             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1588             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1589             felec            = _mm_mul_ps(qq20,felec);
1590
1591             fscal            = felec;
1592
1593              /* Update vectorial force */
1594             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1595             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1596             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1597
1598             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1599             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1600             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1601
1602             /**************************
1603              * CALCULATE INTERACTIONS *
1604              **************************/
1605
1606             r21              = _mm_mul_ps(rsq21,rinv21);
1607
1608             /* EWALD ELECTROSTATICS */
1609
1610             /* Analytical PME correction */
1611             zeta2            = _mm_mul_ps(beta2,rsq21);
1612             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1613             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1614             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1615             felec            = _mm_mul_ps(qq21,felec);
1616
1617             fscal            = felec;
1618
1619              /* Update vectorial force */
1620             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1621             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1622             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1623
1624             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1625             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1626             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1627
1628             /**************************
1629              * CALCULATE INTERACTIONS *
1630              **************************/
1631
1632             r22              = _mm_mul_ps(rsq22,rinv22);
1633
1634             /* EWALD ELECTROSTATICS */
1635
1636             /* Analytical PME correction */
1637             zeta2            = _mm_mul_ps(beta2,rsq22);
1638             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1639             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1640             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1641             felec            = _mm_mul_ps(qq22,felec);
1642
1643             fscal            = felec;
1644
1645              /* Update vectorial force */
1646             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1647             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1648             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1649
1650             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1651             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1652             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1653
1654             fjptrA             = f+j_coord_offsetA;
1655             fjptrB             = f+j_coord_offsetB;
1656             fjptrC             = f+j_coord_offsetC;
1657             fjptrD             = f+j_coord_offsetD;
1658
1659             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1660                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1661
1662             /* Inner loop uses 278 flops */
1663         }
1664
1665         if(jidx<j_index_end)
1666         {
1667
1668             /* Get j neighbor index, and coordinate index */
1669             jnrlistA         = jjnr[jidx];
1670             jnrlistB         = jjnr[jidx+1];
1671             jnrlistC         = jjnr[jidx+2];
1672             jnrlistD         = jjnr[jidx+3];
1673             /* Sign of each element will be negative for non-real atoms.
1674              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1675              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1676              */
1677             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1678             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1679             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1680             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1681             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1682             j_coord_offsetA  = DIM*jnrA;
1683             j_coord_offsetB  = DIM*jnrB;
1684             j_coord_offsetC  = DIM*jnrC;
1685             j_coord_offsetD  = DIM*jnrD;
1686
1687             /* load j atom coordinates */
1688             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1689                                               x+j_coord_offsetC,x+j_coord_offsetD,
1690                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1691
1692             /* Calculate displacement vector */
1693             dx00             = _mm_sub_ps(ix0,jx0);
1694             dy00             = _mm_sub_ps(iy0,jy0);
1695             dz00             = _mm_sub_ps(iz0,jz0);
1696             dx01             = _mm_sub_ps(ix0,jx1);
1697             dy01             = _mm_sub_ps(iy0,jy1);
1698             dz01             = _mm_sub_ps(iz0,jz1);
1699             dx02             = _mm_sub_ps(ix0,jx2);
1700             dy02             = _mm_sub_ps(iy0,jy2);
1701             dz02             = _mm_sub_ps(iz0,jz2);
1702             dx10             = _mm_sub_ps(ix1,jx0);
1703             dy10             = _mm_sub_ps(iy1,jy0);
1704             dz10             = _mm_sub_ps(iz1,jz0);
1705             dx11             = _mm_sub_ps(ix1,jx1);
1706             dy11             = _mm_sub_ps(iy1,jy1);
1707             dz11             = _mm_sub_ps(iz1,jz1);
1708             dx12             = _mm_sub_ps(ix1,jx2);
1709             dy12             = _mm_sub_ps(iy1,jy2);
1710             dz12             = _mm_sub_ps(iz1,jz2);
1711             dx20             = _mm_sub_ps(ix2,jx0);
1712             dy20             = _mm_sub_ps(iy2,jy0);
1713             dz20             = _mm_sub_ps(iz2,jz0);
1714             dx21             = _mm_sub_ps(ix2,jx1);
1715             dy21             = _mm_sub_ps(iy2,jy1);
1716             dz21             = _mm_sub_ps(iz2,jz1);
1717             dx22             = _mm_sub_ps(ix2,jx2);
1718             dy22             = _mm_sub_ps(iy2,jy2);
1719             dz22             = _mm_sub_ps(iz2,jz2);
1720
1721             /* Calculate squared distance and things based on it */
1722             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1723             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1724             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1725             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1726             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1727             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1728             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1729             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1730             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1731
1732             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1733             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1734             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1735             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1736             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1737             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1738             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1739             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1740             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1741
1742             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1743             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1744             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1745             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1746             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1747             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1748             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1749             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1750             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1751
1752             fjx0             = _mm_setzero_ps();
1753             fjy0             = _mm_setzero_ps();
1754             fjz0             = _mm_setzero_ps();
1755             fjx1             = _mm_setzero_ps();
1756             fjy1             = _mm_setzero_ps();
1757             fjz1             = _mm_setzero_ps();
1758             fjx2             = _mm_setzero_ps();
1759             fjy2             = _mm_setzero_ps();
1760             fjz2             = _mm_setzero_ps();
1761
1762             /**************************
1763              * CALCULATE INTERACTIONS *
1764              **************************/
1765
1766             r00              = _mm_mul_ps(rsq00,rinv00);
1767             r00              = _mm_andnot_ps(dummy_mask,r00);
1768
1769             /* Calculate table index by multiplying r with table scale and truncate to integer */
1770             rt               = _mm_mul_ps(r00,vftabscale);
1771             vfitab           = _mm_cvttps_epi32(rt);
1772 #ifdef __XOP__
1773             vfeps            = _mm_frcz_ps(rt);
1774 #else
1775             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1776 #endif
1777             twovfeps         = _mm_add_ps(vfeps,vfeps);
1778             vfitab           = _mm_slli_epi32(vfitab,3);
1779
1780             /* EWALD ELECTROSTATICS */
1781
1782             /* Analytical PME correction */
1783             zeta2            = _mm_mul_ps(beta2,rsq00);
1784             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1785             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1786             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1787             felec            = _mm_mul_ps(qq00,felec);
1788
1789             /* CUBIC SPLINE TABLE DISPERSION */
1790             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1791             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1792             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1793             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1794             _MM_TRANSPOSE4_PS(Y,F,G,H);
1795             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1796             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1797             fvdw6            = _mm_mul_ps(c6_00,FF);
1798
1799             /* CUBIC SPLINE TABLE REPULSION */
1800             vfitab           = _mm_add_epi32(vfitab,ifour);
1801             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1802             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1803             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1804             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1805             _MM_TRANSPOSE4_PS(Y,F,G,H);
1806             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1807             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1808             fvdw12           = _mm_mul_ps(c12_00,FF);
1809             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1810
1811             fscal            = _mm_add_ps(felec,fvdw);
1812
1813             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1814
1815              /* Update vectorial force */
1816             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1817             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1818             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1819
1820             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1821             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1822             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1823
1824             /**************************
1825              * CALCULATE INTERACTIONS *
1826              **************************/
1827
1828             r01              = _mm_mul_ps(rsq01,rinv01);
1829             r01              = _mm_andnot_ps(dummy_mask,r01);
1830
1831             /* EWALD ELECTROSTATICS */
1832
1833             /* Analytical PME correction */
1834             zeta2            = _mm_mul_ps(beta2,rsq01);
1835             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1836             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1837             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1838             felec            = _mm_mul_ps(qq01,felec);
1839
1840             fscal            = felec;
1841
1842             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1843
1844              /* Update vectorial force */
1845             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1846             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1847             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1848
1849             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1850             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1851             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1852
1853             /**************************
1854              * CALCULATE INTERACTIONS *
1855              **************************/
1856
1857             r02              = _mm_mul_ps(rsq02,rinv02);
1858             r02              = _mm_andnot_ps(dummy_mask,r02);
1859
1860             /* EWALD ELECTROSTATICS */
1861
1862             /* Analytical PME correction */
1863             zeta2            = _mm_mul_ps(beta2,rsq02);
1864             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1865             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1866             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1867             felec            = _mm_mul_ps(qq02,felec);
1868
1869             fscal            = felec;
1870
1871             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1872
1873              /* Update vectorial force */
1874             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1875             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1876             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1877
1878             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1879             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1880             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1881
1882             /**************************
1883              * CALCULATE INTERACTIONS *
1884              **************************/
1885
1886             r10              = _mm_mul_ps(rsq10,rinv10);
1887             r10              = _mm_andnot_ps(dummy_mask,r10);
1888
1889             /* EWALD ELECTROSTATICS */
1890
1891             /* Analytical PME correction */
1892             zeta2            = _mm_mul_ps(beta2,rsq10);
1893             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1894             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1895             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1896             felec            = _mm_mul_ps(qq10,felec);
1897
1898             fscal            = felec;
1899
1900             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1901
1902              /* Update vectorial force */
1903             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1904             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1905             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1906
1907             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1908             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1909             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1910
1911             /**************************
1912              * CALCULATE INTERACTIONS *
1913              **************************/
1914
1915             r11              = _mm_mul_ps(rsq11,rinv11);
1916             r11              = _mm_andnot_ps(dummy_mask,r11);
1917
1918             /* EWALD ELECTROSTATICS */
1919
1920             /* Analytical PME correction */
1921             zeta2            = _mm_mul_ps(beta2,rsq11);
1922             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1923             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1924             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1925             felec            = _mm_mul_ps(qq11,felec);
1926
1927             fscal            = felec;
1928
1929             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1930
1931              /* Update vectorial force */
1932             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1933             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1934             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1935
1936             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1937             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1938             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1939
1940             /**************************
1941              * CALCULATE INTERACTIONS *
1942              **************************/
1943
1944             r12              = _mm_mul_ps(rsq12,rinv12);
1945             r12              = _mm_andnot_ps(dummy_mask,r12);
1946
1947             /* EWALD ELECTROSTATICS */
1948
1949             /* Analytical PME correction */
1950             zeta2            = _mm_mul_ps(beta2,rsq12);
1951             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1952             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1953             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1954             felec            = _mm_mul_ps(qq12,felec);
1955
1956             fscal            = felec;
1957
1958             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1959
1960              /* Update vectorial force */
1961             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1962             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1963             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1964
1965             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1966             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1967             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1968
1969             /**************************
1970              * CALCULATE INTERACTIONS *
1971              **************************/
1972
1973             r20              = _mm_mul_ps(rsq20,rinv20);
1974             r20              = _mm_andnot_ps(dummy_mask,r20);
1975
1976             /* EWALD ELECTROSTATICS */
1977
1978             /* Analytical PME correction */
1979             zeta2            = _mm_mul_ps(beta2,rsq20);
1980             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1981             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1982             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1983             felec            = _mm_mul_ps(qq20,felec);
1984
1985             fscal            = felec;
1986
1987             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1988
1989              /* Update vectorial force */
1990             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1991             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1992             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1993
1994             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1995             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1996             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1997
1998             /**************************
1999              * CALCULATE INTERACTIONS *
2000              **************************/
2001
2002             r21              = _mm_mul_ps(rsq21,rinv21);
2003             r21              = _mm_andnot_ps(dummy_mask,r21);
2004
2005             /* EWALD ELECTROSTATICS */
2006
2007             /* Analytical PME correction */
2008             zeta2            = _mm_mul_ps(beta2,rsq21);
2009             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
2010             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2011             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2012             felec            = _mm_mul_ps(qq21,felec);
2013
2014             fscal            = felec;
2015
2016             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2017
2018              /* Update vectorial force */
2019             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2020             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2021             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2022
2023             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2024             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2025             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2026
2027             /**************************
2028              * CALCULATE INTERACTIONS *
2029              **************************/
2030
2031             r22              = _mm_mul_ps(rsq22,rinv22);
2032             r22              = _mm_andnot_ps(dummy_mask,r22);
2033
2034             /* EWALD ELECTROSTATICS */
2035
2036             /* Analytical PME correction */
2037             zeta2            = _mm_mul_ps(beta2,rsq22);
2038             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
2039             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2040             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2041             felec            = _mm_mul_ps(qq22,felec);
2042
2043             fscal            = felec;
2044
2045             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2046
2047              /* Update vectorial force */
2048             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2049             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2050             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2051
2052             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2053             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2054             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2055
2056             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2057             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2058             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2059             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2060
2061             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2062                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2063
2064             /* Inner loop uses 287 flops */
2065         }
2066
2067         /* End of innermost loop */
2068
2069         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2070                                               f+i_coord_offset,fshift+i_shift_offset);
2071
2072         /* Increment number of inner iterations */
2073         inneriter                  += j_index_end - j_index_start;
2074
2075         /* Outer loop uses 18 flops */
2076     }
2077
2078     /* Increment number of outer iterations */
2079     outeriter        += nri;
2080
2081     /* Update outer/inner flops */
2082
2083     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*287);
2084 }