src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_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_ElecRFCut_VdwCSTab_GeomW4W4_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_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              vdwioffset3;
  76     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  77     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  78     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  79     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  80     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  81     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  82     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  83     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  84     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  85     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  86     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  87     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  88     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  89     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  90     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  91     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  92     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  93     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  94     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  95     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  96     real             *charge;
  97     int              nvdwtype;
  98     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 102     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 103     __m128i          vfitab;
 104     __m128i          ifour       = _mm_set1_epi32(4);
 105     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 106     real             *vftab;
 107     __m128           dummy_mask,cutoff_mask;
 108     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 109     __m128           one     = _mm_set1_ps(1.0);
 110     __m128           two     = _mm_set1_ps(2.0);
 111     x                = xx[0];
 112     f                = ff[0];
 113
 114     nri              = nlist->nri;
 115     iinr             = nlist->iinr;
 116     jindex           = nlist->jindex;
 117     jjnr             = nlist->jjnr;
 118     shiftidx         = nlist->shift;
 119     gid              = nlist->gid;
 120     shiftvec         = fr->shift_vec[0];
 121     fshift           = fr->fshift[0];
 122     facel            = _mm_set1_ps(fr->epsfac);
 123     charge           = mdatoms->chargeA;
 124     krf              = _mm_set1_ps(fr->ic->k_rf);
 125     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 126     crf              = _mm_set1_ps(fr->ic->c_rf);
 127     nvdwtype         = fr->ntype;
 128     vdwparam         = fr->nbfp;
 129     vdwtype          = mdatoms->typeA;
 130
 131     vftab            = kernel_data->table_vdw->data;
 132     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 133
 134     /* Setup water-specific parameters */
 135     inr              = nlist->iinr[0];
 136     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 137     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 138     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 139     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 140
 141     jq1              = _mm_set1_ps(charge[inr+1]);
 142     jq2              = _mm_set1_ps(charge[inr+2]);
 143     jq3              = _mm_set1_ps(charge[inr+3]);
 144     vdwjidx0A        = 2*vdwtype[inr+0];
 145     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 146     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 147     qq11             = _mm_mul_ps(iq1,jq1);
 148     qq12             = _mm_mul_ps(iq1,jq2);
 149     qq13             = _mm_mul_ps(iq1,jq3);
 150     qq21             = _mm_mul_ps(iq2,jq1);
 151     qq22             = _mm_mul_ps(iq2,jq2);
 152     qq23             = _mm_mul_ps(iq2,jq3);
 153     qq31             = _mm_mul_ps(iq3,jq1);
 154     qq32             = _mm_mul_ps(iq3,jq2);
 155     qq33             = _mm_mul_ps(iq3,jq3);
 156
 157     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 158     rcutoff_scalar   = fr->rcoulomb;
 159     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 160     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 161
 162     /* Avoid stupid compiler warnings */
 163     jnrA = jnrB = jnrC = jnrD = 0;
 164     j_coord_offsetA = 0;
 165     j_coord_offsetB = 0;
 166     j_coord_offsetC = 0;
 167     j_coord_offsetD = 0;
 168
 169     outeriter        = 0;
 170     inneriter        = 0;
 171
 172     for(iidx=0;iidx<4*DIM;iidx++)
 173     {
 174         scratch[iidx] = 0.0;
 175     }
 176
 177     /* Start outer loop over neighborlists */
 178     for(iidx=0; iidx<nri; iidx++)
 179     {
 180         /* Load shift vector for this list */
 181         i_shift_offset   = DIM*shiftidx[iidx];
 182
 183         /* Load limits for loop over neighbors */
 184         j_index_start    = jindex[iidx];
 185         j_index_end      = jindex[iidx+1];
 186
 187         /* Get outer coordinate index */
 188         inr              = iinr[iidx];
 189         i_coord_offset   = DIM*inr;
 190
 191         /* Load i particle coords and add shift vector */
 192         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 193                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 194
 195         fix0             = _mm_setzero_ps();
 196         fiy0             = _mm_setzero_ps();
 197         fiz0             = _mm_setzero_ps();
 198         fix1             = _mm_setzero_ps();
 199         fiy1             = _mm_setzero_ps();
 200         fiz1             = _mm_setzero_ps();
 201         fix2             = _mm_setzero_ps();
 202         fiy2             = _mm_setzero_ps();
 203         fiz2             = _mm_setzero_ps();
 204         fix3             = _mm_setzero_ps();
 205         fiy3             = _mm_setzero_ps();
 206         fiz3             = _mm_setzero_ps();
 207
 208         /* Reset potential sums */
 209         velecsum         = _mm_setzero_ps();
 210         vvdwsum          = _mm_setzero_ps();
 211
 212         /* Start inner kernel loop */
 213         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 214         {
 215
 216             /* Get j neighbor index, and coordinate index */
 217             jnrA             = jjnr[jidx];
 218             jnrB             = jjnr[jidx+1];
 219             jnrC             = jjnr[jidx+2];
 220             jnrD             = jjnr[jidx+3];
 221             j_coord_offsetA  = DIM*jnrA;
 222             j_coord_offsetB  = DIM*jnrB;
 223             j_coord_offsetC  = DIM*jnrC;
 224             j_coord_offsetD  = DIM*jnrD;
 225
 226             /* load j atom coordinates */
 227             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 228                                               x+j_coord_offsetC,x+j_coord_offsetD,
 229                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 230                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 231
 232             /* Calculate displacement vector */
 233             dx00             = _mm_sub_ps(ix0,jx0);
 234             dy00             = _mm_sub_ps(iy0,jy0);
 235             dz00             = _mm_sub_ps(iz0,jz0);
 236             dx11             = _mm_sub_ps(ix1,jx1);
 237             dy11             = _mm_sub_ps(iy1,jy1);
 238             dz11             = _mm_sub_ps(iz1,jz1);
 239             dx12             = _mm_sub_ps(ix1,jx2);
 240             dy12             = _mm_sub_ps(iy1,jy2);
 241             dz12             = _mm_sub_ps(iz1,jz2);
 242             dx13             = _mm_sub_ps(ix1,jx3);
 243             dy13             = _mm_sub_ps(iy1,jy3);
 244             dz13             = _mm_sub_ps(iz1,jz3);
 245             dx21             = _mm_sub_ps(ix2,jx1);
 246             dy21             = _mm_sub_ps(iy2,jy1);
 247             dz21             = _mm_sub_ps(iz2,jz1);
 248             dx22             = _mm_sub_ps(ix2,jx2);
 249             dy22             = _mm_sub_ps(iy2,jy2);
 250             dz22             = _mm_sub_ps(iz2,jz2);
 251             dx23             = _mm_sub_ps(ix2,jx3);
 252             dy23             = _mm_sub_ps(iy2,jy3);
 253             dz23             = _mm_sub_ps(iz2,jz3);
 254             dx31             = _mm_sub_ps(ix3,jx1);
 255             dy31             = _mm_sub_ps(iy3,jy1);
 256             dz31             = _mm_sub_ps(iz3,jz1);
 257             dx32             = _mm_sub_ps(ix3,jx2);
 258             dy32             = _mm_sub_ps(iy3,jy2);
 259             dz32             = _mm_sub_ps(iz3,jz2);
 260             dx33             = _mm_sub_ps(ix3,jx3);
 261             dy33             = _mm_sub_ps(iy3,jy3);
 262             dz33             = _mm_sub_ps(iz3,jz3);
 263
 264             /* Calculate squared distance and things based on it */
 265             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 266             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 267             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 268             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 269             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 270             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 271             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 272             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 273             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 274             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 275
 276             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 277             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 278             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 279             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 280             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 281             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 282             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 283             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 284             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 285             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 286
 287             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 288             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 289             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 290             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 291             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 292             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 293             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 294             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 295             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 296
 297             fjx0             = _mm_setzero_ps();
 298             fjy0             = _mm_setzero_ps();
 299             fjz0             = _mm_setzero_ps();
 300             fjx1             = _mm_setzero_ps();
 301             fjy1             = _mm_setzero_ps();
 302             fjz1             = _mm_setzero_ps();
 303             fjx2             = _mm_setzero_ps();
 304             fjy2             = _mm_setzero_ps();
 305             fjz2             = _mm_setzero_ps();
 306             fjx3             = _mm_setzero_ps();
 307             fjy3             = _mm_setzero_ps();
 308             fjz3             = _mm_setzero_ps();
 309
 310             /**************************
 311              * CALCULATE INTERACTIONS *
 312              **************************/
 313
 314             if (gmx_mm_any_lt(rsq00,rcutoff2))
 315             {
 316
 317             r00              = _mm_mul_ps(rsq00,rinv00);
 318
 319             /* Calculate table index by multiplying r with table scale and truncate to integer */
 320             rt               = _mm_mul_ps(r00,vftabscale);
 321             vfitab           = _mm_cvttps_epi32(rt);
 322 #ifdef __XOP__
 323             vfeps            = _mm_frcz_ps(rt);
 324 #else
 325             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 326 #endif
 327             twovfeps         = _mm_add_ps(vfeps,vfeps);
 328             vfitab           = _mm_slli_epi32(vfitab,3);
 329
 330             /* CUBIC SPLINE TABLE DISPERSION */
 331             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 332             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 333             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 334             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 335             _MM_TRANSPOSE4_PS(Y,F,G,H);
 336             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 337             VV               = _mm_macc_ps(vfeps,Fp,Y);
 338             vvdw6            = _mm_mul_ps(c6_00,VV);
 339             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 340             fvdw6            = _mm_mul_ps(c6_00,FF);
 341
 342             /* CUBIC SPLINE TABLE REPULSION */
 343             vfitab           = _mm_add_epi32(vfitab,ifour);
 344             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 345             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 346             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 347             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 348             _MM_TRANSPOSE4_PS(Y,F,G,H);
 349             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 350             VV               = _mm_macc_ps(vfeps,Fp,Y);
 351             vvdw12           = _mm_mul_ps(c12_00,VV);
 352             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 353             fvdw12           = _mm_mul_ps(c12_00,FF);
 354             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 355             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 356
 357             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 361             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 362
 363             fscal            = fvdw;
 364
 365             fscal            = _mm_and_ps(fscal,cutoff_mask);
 366
 367              /* Update vectorial force */
 368             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 369             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 370             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 371
 372             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 373             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 374             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 375
 376             }
 377
 378             /**************************
 379              * CALCULATE INTERACTIONS *
 380              **************************/
 381
 382             if (gmx_mm_any_lt(rsq11,rcutoff2))
 383             {
 384
 385             /* REACTION-FIELD ELECTROSTATICS */
 386             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 387             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 388
 389             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 390
 391             /* Update potential sum for this i atom from the interaction with this j atom. */
 392             velec            = _mm_and_ps(velec,cutoff_mask);
 393             velecsum         = _mm_add_ps(velecsum,velec);
 394
 395             fscal            = felec;
 396
 397             fscal            = _mm_and_ps(fscal,cutoff_mask);
 398
 399              /* Update vectorial force */
 400             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 401             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 402             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 403
 404             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 405             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 406             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 407
 408             }
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             if (gmx_mm_any_lt(rsq12,rcutoff2))
 415             {
 416
 417             /* REACTION-FIELD ELECTROSTATICS */
 418             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 419             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 420
 421             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 422
 423             /* Update potential sum for this i atom from the interaction with this j atom. */
 424             velec            = _mm_and_ps(velec,cutoff_mask);
 425             velecsum         = _mm_add_ps(velecsum,velec);
 426
 427             fscal            = felec;
 428
 429             fscal            = _mm_and_ps(fscal,cutoff_mask);
 430
 431              /* Update vectorial force */
 432             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 433             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 434             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 435
 436             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 437             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 438             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 439
 440             }
 441
 442             /**************************
 443              * CALCULATE INTERACTIONS *
 444              **************************/
 445
 446             if (gmx_mm_any_lt(rsq13,rcutoff2))
 447             {
 448
 449             /* REACTION-FIELD ELECTROSTATICS */
 450             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 451             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 452
 453             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velec            = _mm_and_ps(velec,cutoff_mask);
 457             velecsum         = _mm_add_ps(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             fscal            = _mm_and_ps(fscal,cutoff_mask);
 462
 463              /* Update vectorial force */
 464             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 465             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 466             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 467
 468             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 469             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 470             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 471
 472             }
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             if (gmx_mm_any_lt(rsq21,rcutoff2))
 479             {
 480
 481             /* REACTION-FIELD ELECTROSTATICS */
 482             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 483             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 484
 485             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 486
 487             /* Update potential sum for this i atom from the interaction with this j atom. */
 488             velec            = _mm_and_ps(velec,cutoff_mask);
 489             velecsum         = _mm_add_ps(velecsum,velec);
 490
 491             fscal            = felec;
 492
 493             fscal            = _mm_and_ps(fscal,cutoff_mask);
 494
 495              /* Update vectorial force */
 496             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 497             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 498             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 499
 500             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 501             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 502             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 503
 504             }
 505
 506             /**************************
 507              * CALCULATE INTERACTIONS *
 508              **************************/
 509
 510             if (gmx_mm_any_lt(rsq22,rcutoff2))
 511             {
 512
 513             /* REACTION-FIELD ELECTROSTATICS */
 514             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 515             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 516
 517             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 518
 519             /* Update potential sum for this i atom from the interaction with this j atom. */
 520             velec            = _mm_and_ps(velec,cutoff_mask);
 521             velecsum         = _mm_add_ps(velecsum,velec);
 522
 523             fscal            = felec;
 524
 525             fscal            = _mm_and_ps(fscal,cutoff_mask);
 526
 527              /* Update vectorial force */
 528             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 529             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 530             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 531
 532             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 533             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 534             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 535
 536             }
 537
 538             /**************************
 539              * CALCULATE INTERACTIONS *
 540              **************************/
 541
 542             if (gmx_mm_any_lt(rsq23,rcutoff2))
 543             {
 544
 545             /* REACTION-FIELD ELECTROSTATICS */
 546             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 547             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 548
 549             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 550
 551             /* Update potential sum for this i atom from the interaction with this j atom. */
 552             velec            = _mm_and_ps(velec,cutoff_mask);
 553             velecsum         = _mm_add_ps(velecsum,velec);
 554
 555             fscal            = felec;
 556
 557             fscal            = _mm_and_ps(fscal,cutoff_mask);
 558
 559              /* Update vectorial force */
 560             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 561             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 562             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 563
 564             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 565             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 566             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 567
 568             }
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             if (gmx_mm_any_lt(rsq31,rcutoff2))
 575             {
 576
 577             /* REACTION-FIELD ELECTROSTATICS */
 578             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 579             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 580
 581             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 582
 583             /* Update potential sum for this i atom from the interaction with this j atom. */
 584             velec            = _mm_and_ps(velec,cutoff_mask);
 585             velecsum         = _mm_add_ps(velecsum,velec);
 586
 587             fscal            = felec;
 588
 589             fscal            = _mm_and_ps(fscal,cutoff_mask);
 590
 591              /* Update vectorial force */
 592             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 593             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 594             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 595
 596             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 597             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 598             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 599
 600             }
 601
 602             /**************************
 603              * CALCULATE INTERACTIONS *
 604              **************************/
 605
 606             if (gmx_mm_any_lt(rsq32,rcutoff2))
 607             {
 608
 609             /* REACTION-FIELD ELECTROSTATICS */
 610             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 611             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 612
 613             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 614
 615             /* Update potential sum for this i atom from the interaction with this j atom. */
 616             velec            = _mm_and_ps(velec,cutoff_mask);
 617             velecsum         = _mm_add_ps(velecsum,velec);
 618
 619             fscal            = felec;
 620
 621             fscal            = _mm_and_ps(fscal,cutoff_mask);
 622
 623              /* Update vectorial force */
 624             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 625             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 626             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 627
 628             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 629             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 630             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 631
 632             }
 633
 634             /**************************
 635              * CALCULATE INTERACTIONS *
 636              **************************/
 637
 638             if (gmx_mm_any_lt(rsq33,rcutoff2))
 639             {
 640
 641             /* REACTION-FIELD ELECTROSTATICS */
 642             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 643             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 644
 645             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 646
 647             /* Update potential sum for this i atom from the interaction with this j atom. */
 648             velec            = _mm_and_ps(velec,cutoff_mask);
 649             velecsum         = _mm_add_ps(velecsum,velec);
 650
 651             fscal            = felec;
 652
 653             fscal            = _mm_and_ps(fscal,cutoff_mask);
 654
 655              /* Update vectorial force */
 656             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 657             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 658             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 659
 660             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 661             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 662             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 663
 664             }
 665
 666             fjptrA             = f+j_coord_offsetA;
 667             fjptrB             = f+j_coord_offsetB;
 668             fjptrC             = f+j_coord_offsetC;
 669             fjptrD             = f+j_coord_offsetD;
 670
 671             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 672                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 673                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 674
 675             /* Inner loop uses 417 flops */
 676         }
 677
 678         if(jidx<j_index_end)
 679         {
 680
 681             /* Get j neighbor index, and coordinate index */
 682             jnrlistA         = jjnr[jidx];
 683             jnrlistB         = jjnr[jidx+1];
 684             jnrlistC         = jjnr[jidx+2];
 685             jnrlistD         = jjnr[jidx+3];
 686             /* Sign of each element will be negative for non-real atoms.
 687              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 688              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 689              */
 690             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 691             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 692             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 693             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 694             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 695             j_coord_offsetA  = DIM*jnrA;
 696             j_coord_offsetB  = DIM*jnrB;
 697             j_coord_offsetC  = DIM*jnrC;
 698             j_coord_offsetD  = DIM*jnrD;
 699
 700             /* load j atom coordinates */
 701             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 702                                               x+j_coord_offsetC,x+j_coord_offsetD,
 703                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 704                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 705
 706             /* Calculate displacement vector */
 707             dx00             = _mm_sub_ps(ix0,jx0);
 708             dy00             = _mm_sub_ps(iy0,jy0);
 709             dz00             = _mm_sub_ps(iz0,jz0);
 710             dx11             = _mm_sub_ps(ix1,jx1);
 711             dy11             = _mm_sub_ps(iy1,jy1);
 712             dz11             = _mm_sub_ps(iz1,jz1);
 713             dx12             = _mm_sub_ps(ix1,jx2);
 714             dy12             = _mm_sub_ps(iy1,jy2);
 715             dz12             = _mm_sub_ps(iz1,jz2);
 716             dx13             = _mm_sub_ps(ix1,jx3);
 717             dy13             = _mm_sub_ps(iy1,jy3);
 718             dz13             = _mm_sub_ps(iz1,jz3);
 719             dx21             = _mm_sub_ps(ix2,jx1);
 720             dy21             = _mm_sub_ps(iy2,jy1);
 721             dz21             = _mm_sub_ps(iz2,jz1);
 722             dx22             = _mm_sub_ps(ix2,jx2);
 723             dy22             = _mm_sub_ps(iy2,jy2);
 724             dz22             = _mm_sub_ps(iz2,jz2);
 725             dx23             = _mm_sub_ps(ix2,jx3);
 726             dy23             = _mm_sub_ps(iy2,jy3);
 727             dz23             = _mm_sub_ps(iz2,jz3);
 728             dx31             = _mm_sub_ps(ix3,jx1);
 729             dy31             = _mm_sub_ps(iy3,jy1);
 730             dz31             = _mm_sub_ps(iz3,jz1);
 731             dx32             = _mm_sub_ps(ix3,jx2);
 732             dy32             = _mm_sub_ps(iy3,jy2);
 733             dz32             = _mm_sub_ps(iz3,jz2);
 734             dx33             = _mm_sub_ps(ix3,jx3);
 735             dy33             = _mm_sub_ps(iy3,jy3);
 736             dz33             = _mm_sub_ps(iz3,jz3);
 737
 738             /* Calculate squared distance and things based on it */
 739             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 740             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 741             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 742             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 743             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 744             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 745             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 746             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 747             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 748             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 749
 750             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 751             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 752             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 753             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 754             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 755             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 756             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 757             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 758             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 759             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 760
 761             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 762             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 763             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 764             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 765             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 766             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 767             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 768             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 769             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 770
 771             fjx0             = _mm_setzero_ps();
 772             fjy0             = _mm_setzero_ps();
 773             fjz0             = _mm_setzero_ps();
 774             fjx1             = _mm_setzero_ps();
 775             fjy1             = _mm_setzero_ps();
 776             fjz1             = _mm_setzero_ps();
 777             fjx2             = _mm_setzero_ps();
 778             fjy2             = _mm_setzero_ps();
 779             fjz2             = _mm_setzero_ps();
 780             fjx3             = _mm_setzero_ps();
 781             fjy3             = _mm_setzero_ps();
 782             fjz3             = _mm_setzero_ps();
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             if (gmx_mm_any_lt(rsq00,rcutoff2))
 789             {
 790
 791             r00              = _mm_mul_ps(rsq00,rinv00);
 792             r00              = _mm_andnot_ps(dummy_mask,r00);
 793
 794             /* Calculate table index by multiplying r with table scale and truncate to integer */
 795             rt               = _mm_mul_ps(r00,vftabscale);
 796             vfitab           = _mm_cvttps_epi32(rt);
 797 #ifdef __XOP__
 798             vfeps            = _mm_frcz_ps(rt);
 799 #else
 800             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 801 #endif
 802             twovfeps         = _mm_add_ps(vfeps,vfeps);
 803             vfitab           = _mm_slli_epi32(vfitab,3);
 804
 805             /* CUBIC SPLINE TABLE DISPERSION */
 806             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 807             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 808             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 809             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 810             _MM_TRANSPOSE4_PS(Y,F,G,H);
 811             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 812             VV               = _mm_macc_ps(vfeps,Fp,Y);
 813             vvdw6            = _mm_mul_ps(c6_00,VV);
 814             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 815             fvdw6            = _mm_mul_ps(c6_00,FF);
 816
 817             /* CUBIC SPLINE TABLE REPULSION */
 818             vfitab           = _mm_add_epi32(vfitab,ifour);
 819             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 820             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 821             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 822             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 823             _MM_TRANSPOSE4_PS(Y,F,G,H);
 824             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 825             VV               = _mm_macc_ps(vfeps,Fp,Y);
 826             vvdw12           = _mm_mul_ps(c12_00,VV);
 827             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 828             fvdw12           = _mm_mul_ps(c12_00,FF);
 829             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 830             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 831
 832             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 833
 834             /* Update potential sum for this i atom from the interaction with this j atom. */
 835             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 836             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 837             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 838
 839             fscal            = fvdw;
 840
 841             fscal            = _mm_and_ps(fscal,cutoff_mask);
 842
 843             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 844
 845              /* Update vectorial force */
 846             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 847             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 848             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 849
 850             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 851             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 852             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 853
 854             }
 855
 856             /**************************
 857              * CALCULATE INTERACTIONS *
 858              **************************/
 859
 860             if (gmx_mm_any_lt(rsq11,rcutoff2))
 861             {
 862
 863             /* REACTION-FIELD ELECTROSTATICS */
 864             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 865             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 866
 867             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 868
 869             /* Update potential sum for this i atom from the interaction with this j atom. */
 870             velec            = _mm_and_ps(velec,cutoff_mask);
 871             velec            = _mm_andnot_ps(dummy_mask,velec);
 872             velecsum         = _mm_add_ps(velecsum,velec);
 873
 874             fscal            = felec;
 875
 876             fscal            = _mm_and_ps(fscal,cutoff_mask);
 877
 878             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 879
 880              /* Update vectorial force */
 881             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 882             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 883             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 884
 885             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 886             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 887             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 888
 889             }
 890
 891             /**************************
 892              * CALCULATE INTERACTIONS *
 893              **************************/
 894
 895             if (gmx_mm_any_lt(rsq12,rcutoff2))
 896             {
 897
 898             /* REACTION-FIELD ELECTROSTATICS */
 899             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 900             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 901
 902             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 903
 904             /* Update potential sum for this i atom from the interaction with this j atom. */
 905             velec            = _mm_and_ps(velec,cutoff_mask);
 906             velec            = _mm_andnot_ps(dummy_mask,velec);
 907             velecsum         = _mm_add_ps(velecsum,velec);
 908
 909             fscal            = felec;
 910
 911             fscal            = _mm_and_ps(fscal,cutoff_mask);
 912
 913             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 914
 915              /* Update vectorial force */
 916             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 917             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 918             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 919
 920             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 921             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 922             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 923
 924             }
 925
 926             /**************************
 927              * CALCULATE INTERACTIONS *
 928              **************************/
 929
 930             if (gmx_mm_any_lt(rsq13,rcutoff2))
 931             {
 932
 933             /* REACTION-FIELD ELECTROSTATICS */
 934             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 935             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 936
 937             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 938
 939             /* Update potential sum for this i atom from the interaction with this j atom. */
 940             velec            = _mm_and_ps(velec,cutoff_mask);
 941             velec            = _mm_andnot_ps(dummy_mask,velec);
 942             velecsum         = _mm_add_ps(velecsum,velec);
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_and_ps(fscal,cutoff_mask);
 947
 948             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 949
 950              /* Update vectorial force */
 951             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 952             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 953             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 954
 955             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 956             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 957             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 958
 959             }
 960
 961             /**************************
 962              * CALCULATE INTERACTIONS *
 963              **************************/
 964
 965             if (gmx_mm_any_lt(rsq21,rcutoff2))
 966             {
 967
 968             /* REACTION-FIELD ELECTROSTATICS */
 969             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 970             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 971
 972             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 973
 974             /* Update potential sum for this i atom from the interaction with this j atom. */
 975             velec            = _mm_and_ps(velec,cutoff_mask);
 976             velec            = _mm_andnot_ps(dummy_mask,velec);
 977             velecsum         = _mm_add_ps(velecsum,velec);
 978
 979             fscal            = felec;
 980
 981             fscal            = _mm_and_ps(fscal,cutoff_mask);
 982
 983             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 984
 985              /* Update vectorial force */
 986             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 987             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 988             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 989
 990             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 991             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 992             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 993
 994             }
 995
 996             /**************************
 997              * CALCULATE INTERACTIONS *
 998              **************************/
 999
1000             if (gmx_mm_any_lt(rsq22,rcutoff2))
1001             {
1002
1003             /* REACTION-FIELD ELECTROSTATICS */
1004             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
1005             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1006
1007             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1008
1009             /* Update potential sum for this i atom from the interaction with this j atom. */
1010             velec            = _mm_and_ps(velec,cutoff_mask);
1011             velec            = _mm_andnot_ps(dummy_mask,velec);
1012             velecsum         = _mm_add_ps(velecsum,velec);
1013
1014             fscal            = felec;
1015
1016             fscal            = _mm_and_ps(fscal,cutoff_mask);
1017
1018             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1019
1020              /* Update vectorial force */
1021             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1022             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1023             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1024
1025             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1026             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1027             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1028
1029             }
1030
1031             /**************************
1032              * CALCULATE INTERACTIONS *
1033              **************************/
1034
1035             if (gmx_mm_any_lt(rsq23,rcutoff2))
1036             {
1037
1038             /* REACTION-FIELD ELECTROSTATICS */
1039             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
1040             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1041
1042             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1043
1044             /* Update potential sum for this i atom from the interaction with this j atom. */
1045             velec            = _mm_and_ps(velec,cutoff_mask);
1046             velec            = _mm_andnot_ps(dummy_mask,velec);
1047             velecsum         = _mm_add_ps(velecsum,velec);
1048
1049             fscal            = felec;
1050
1051             fscal            = _mm_and_ps(fscal,cutoff_mask);
1052
1053             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1054
1055              /* Update vectorial force */
1056             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1057             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1058             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1059
1060             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1061             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1062             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1063
1064             }
1065
1066             /**************************
1067              * CALCULATE INTERACTIONS *
1068              **************************/
1069
1070             if (gmx_mm_any_lt(rsq31,rcutoff2))
1071             {
1072
1073             /* REACTION-FIELD ELECTROSTATICS */
1074             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
1075             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1076
1077             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1078
1079             /* Update potential sum for this i atom from the interaction with this j atom. */
1080             velec            = _mm_and_ps(velec,cutoff_mask);
1081             velec            = _mm_andnot_ps(dummy_mask,velec);
1082             velecsum         = _mm_add_ps(velecsum,velec);
1083
1084             fscal            = felec;
1085
1086             fscal            = _mm_and_ps(fscal,cutoff_mask);
1087
1088             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1089
1090              /* Update vectorial force */
1091             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1092             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1093             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1094
1095             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1096             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1097             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1098
1099             }
1100
1101             /**************************
1102              * CALCULATE INTERACTIONS *
1103              **************************/
1104
1105             if (gmx_mm_any_lt(rsq32,rcutoff2))
1106             {
1107
1108             /* REACTION-FIELD ELECTROSTATICS */
1109             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
1110             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1111
1112             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1113
1114             /* Update potential sum for this i atom from the interaction with this j atom. */
1115             velec            = _mm_and_ps(velec,cutoff_mask);
1116             velec            = _mm_andnot_ps(dummy_mask,velec);
1117             velecsum         = _mm_add_ps(velecsum,velec);
1118
1119             fscal            = felec;
1120
1121             fscal            = _mm_and_ps(fscal,cutoff_mask);
1122
1123             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1124
1125              /* Update vectorial force */
1126             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1127             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1128             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1129
1130             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1131             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1132             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1133
1134             }
1135
1136             /**************************
1137              * CALCULATE INTERACTIONS *
1138              **************************/
1139
1140             if (gmx_mm_any_lt(rsq33,rcutoff2))
1141             {
1142
1143             /* REACTION-FIELD ELECTROSTATICS */
1144             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
1145             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1146
1147             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1148
1149             /* Update potential sum for this i atom from the interaction with this j atom. */
1150             velec            = _mm_and_ps(velec,cutoff_mask);
1151             velec            = _mm_andnot_ps(dummy_mask,velec);
1152             velecsum         = _mm_add_ps(velecsum,velec);
1153
1154             fscal            = felec;
1155
1156             fscal            = _mm_and_ps(fscal,cutoff_mask);
1157
1158             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1159
1160              /* Update vectorial force */
1161             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1162             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1163             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1164
1165             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1166             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1167             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1168
1169             }
1170
1171             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1172             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1173             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1174             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1175
1176             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1177                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1178                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1179
1180             /* Inner loop uses 418 flops */
1181         }
1182
1183         /* End of innermost loop */
1184
1185         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1186                                               f+i_coord_offset,fshift+i_shift_offset);
1187
1188         ggid                        = gid[iidx];
1189         /* Update potential energies */
1190         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1191         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1192
1193         /* Increment number of inner iterations */
1194         inneriter                  += j_index_end - j_index_start;
1195
1196         /* Outer loop uses 26 flops */
1197     }
1198
1199     /* Increment number of outer iterations */
1200     outeriter        += nri;
1201
1202     /* Update outer/inner flops */
1203
1204     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*418);
1205 }
1206 /*
1207  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1208  * Electrostatics interaction: ReactionField
1209  * VdW interaction:            CubicSplineTable
1210  * Geometry:                   Water4-Water4
1211  * Calculate force/pot:        Force
1212  */
1213 void
1214 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_single
1215                     (t_nblist * gmx_restrict                nlist,
1216                      rvec * gmx_restrict                    xx,
1217                      rvec * gmx_restrict                    ff,
1218                      t_forcerec * gmx_restrict              fr,
1219                      t_mdatoms * gmx_restrict               mdatoms,
1220                      nb_kernel_data_t * gmx_restrict        kernel_data,
1221                      t_nrnb * gmx_restrict                  nrnb)
1222 {
1223     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1224      * just 0 for non-waters.
1225      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1226      * jnr indices corresponding to data put in the four positions in the SIMD register.
1227      */
1228     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1229     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1230     int              jnrA,jnrB,jnrC,jnrD;
1231     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1232     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1233     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1234     real             rcutoff_scalar;
1235     real             *shiftvec,*fshift,*x,*f;
1236     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1237     real             scratch[4*DIM];
1238     __m128           fscal,rcutoff,rcutoff2,jidxall;
1239     int              vdwioffset0;
1240     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1241     int              vdwioffset1;
1242     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1243     int              vdwioffset2;
1244     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1245     int              vdwioffset3;
1246     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1247     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1248     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1249     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1250     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1251     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1252     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1253     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1254     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1255     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1256     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1257     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1258     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1259     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1260     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1261     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1262     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1263     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1264     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1265     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1266     real             *charge;
1267     int              nvdwtype;
1268     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1269     int              *vdwtype;
1270     real             *vdwparam;
1271     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1272     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1273     __m128i          vfitab;
1274     __m128i          ifour       = _mm_set1_epi32(4);
1275     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
1276     real             *vftab;
1277     __m128           dummy_mask,cutoff_mask;
1278     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1279     __m128           one     = _mm_set1_ps(1.0);
1280     __m128           two     = _mm_set1_ps(2.0);
1281     x                = xx[0];
1282     f                = ff[0];
1283
1284     nri              = nlist->nri;
1285     iinr             = nlist->iinr;
1286     jindex           = nlist->jindex;
1287     jjnr             = nlist->jjnr;
1288     shiftidx         = nlist->shift;
1289     gid              = nlist->gid;
1290     shiftvec         = fr->shift_vec[0];
1291     fshift           = fr->fshift[0];
1292     facel            = _mm_set1_ps(fr->epsfac);
1293     charge           = mdatoms->chargeA;
1294     krf              = _mm_set1_ps(fr->ic->k_rf);
1295     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1296     crf              = _mm_set1_ps(fr->ic->c_rf);
1297     nvdwtype         = fr->ntype;
1298     vdwparam         = fr->nbfp;
1299     vdwtype          = mdatoms->typeA;
1300
1301     vftab            = kernel_data->table_vdw->data;
1302     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1303
1304     /* Setup water-specific parameters */
1305     inr              = nlist->iinr[0];
1306     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1307     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1308     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1309     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1310
1311     jq1              = _mm_set1_ps(charge[inr+1]);
1312     jq2              = _mm_set1_ps(charge[inr+2]);
1313     jq3              = _mm_set1_ps(charge[inr+3]);
1314     vdwjidx0A        = 2*vdwtype[inr+0];
1315     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1316     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1317     qq11             = _mm_mul_ps(iq1,jq1);
1318     qq12             = _mm_mul_ps(iq1,jq2);
1319     qq13             = _mm_mul_ps(iq1,jq3);
1320     qq21             = _mm_mul_ps(iq2,jq1);
1321     qq22             = _mm_mul_ps(iq2,jq2);
1322     qq23             = _mm_mul_ps(iq2,jq3);
1323     qq31             = _mm_mul_ps(iq3,jq1);
1324     qq32             = _mm_mul_ps(iq3,jq2);
1325     qq33             = _mm_mul_ps(iq3,jq3);
1326
1327     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1328     rcutoff_scalar   = fr->rcoulomb;
1329     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1330     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1331
1332     /* Avoid stupid compiler warnings */
1333     jnrA = jnrB = jnrC = jnrD = 0;
1334     j_coord_offsetA = 0;
1335     j_coord_offsetB = 0;
1336     j_coord_offsetC = 0;
1337     j_coord_offsetD = 0;
1338
1339     outeriter        = 0;
1340     inneriter        = 0;
1341
1342     for(iidx=0;iidx<4*DIM;iidx++)
1343     {
1344         scratch[iidx] = 0.0;
1345     }
1346
1347     /* Start outer loop over neighborlists */
1348     for(iidx=0; iidx<nri; iidx++)
1349     {
1350         /* Load shift vector for this list */
1351         i_shift_offset   = DIM*shiftidx[iidx];
1352
1353         /* Load limits for loop over neighbors */
1354         j_index_start    = jindex[iidx];
1355         j_index_end      = jindex[iidx+1];
1356
1357         /* Get outer coordinate index */
1358         inr              = iinr[iidx];
1359         i_coord_offset   = DIM*inr;
1360
1361         /* Load i particle coords and add shift vector */
1362         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1363                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1364
1365         fix0             = _mm_setzero_ps();
1366         fiy0             = _mm_setzero_ps();
1367         fiz0             = _mm_setzero_ps();
1368         fix1             = _mm_setzero_ps();
1369         fiy1             = _mm_setzero_ps();
1370         fiz1             = _mm_setzero_ps();
1371         fix2             = _mm_setzero_ps();
1372         fiy2             = _mm_setzero_ps();
1373         fiz2             = _mm_setzero_ps();
1374         fix3             = _mm_setzero_ps();
1375         fiy3             = _mm_setzero_ps();
1376         fiz3             = _mm_setzero_ps();
1377
1378         /* Start inner kernel loop */
1379         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1380         {
1381
1382             /* Get j neighbor index, and coordinate index */
1383             jnrA             = jjnr[jidx];
1384             jnrB             = jjnr[jidx+1];
1385             jnrC             = jjnr[jidx+2];
1386             jnrD             = jjnr[jidx+3];
1387             j_coord_offsetA  = DIM*jnrA;
1388             j_coord_offsetB  = DIM*jnrB;
1389             j_coord_offsetC  = DIM*jnrC;
1390             j_coord_offsetD  = DIM*jnrD;
1391
1392             /* load j atom coordinates */
1393             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1394                                               x+j_coord_offsetC,x+j_coord_offsetD,
1395                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1396                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1397
1398             /* Calculate displacement vector */
1399             dx00             = _mm_sub_ps(ix0,jx0);
1400             dy00             = _mm_sub_ps(iy0,jy0);
1401             dz00             = _mm_sub_ps(iz0,jz0);
1402             dx11             = _mm_sub_ps(ix1,jx1);
1403             dy11             = _mm_sub_ps(iy1,jy1);
1404             dz11             = _mm_sub_ps(iz1,jz1);
1405             dx12             = _mm_sub_ps(ix1,jx2);
1406             dy12             = _mm_sub_ps(iy1,jy2);
1407             dz12             = _mm_sub_ps(iz1,jz2);
1408             dx13             = _mm_sub_ps(ix1,jx3);
1409             dy13             = _mm_sub_ps(iy1,jy3);
1410             dz13             = _mm_sub_ps(iz1,jz3);
1411             dx21             = _mm_sub_ps(ix2,jx1);
1412             dy21             = _mm_sub_ps(iy2,jy1);
1413             dz21             = _mm_sub_ps(iz2,jz1);
1414             dx22             = _mm_sub_ps(ix2,jx2);
1415             dy22             = _mm_sub_ps(iy2,jy2);
1416             dz22             = _mm_sub_ps(iz2,jz2);
1417             dx23             = _mm_sub_ps(ix2,jx3);
1418             dy23             = _mm_sub_ps(iy2,jy3);
1419             dz23             = _mm_sub_ps(iz2,jz3);
1420             dx31             = _mm_sub_ps(ix3,jx1);
1421             dy31             = _mm_sub_ps(iy3,jy1);
1422             dz31             = _mm_sub_ps(iz3,jz1);
1423             dx32             = _mm_sub_ps(ix3,jx2);
1424             dy32             = _mm_sub_ps(iy3,jy2);
1425             dz32             = _mm_sub_ps(iz3,jz2);
1426             dx33             = _mm_sub_ps(ix3,jx3);
1427             dy33             = _mm_sub_ps(iy3,jy3);
1428             dz33             = _mm_sub_ps(iz3,jz3);
1429
1430             /* Calculate squared distance and things based on it */
1431             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1432             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1433             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1434             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1435             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1436             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1437             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1438             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1439             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1440             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1441
1442             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1443             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1444             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1445             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1446             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1447             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1448             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1449             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1450             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1451             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1452
1453             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1454             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1455             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1456             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1457             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1458             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1459             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1460             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1461             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1462
1463             fjx0             = _mm_setzero_ps();
1464             fjy0             = _mm_setzero_ps();
1465             fjz0             = _mm_setzero_ps();
1466             fjx1             = _mm_setzero_ps();
1467             fjy1             = _mm_setzero_ps();
1468             fjz1             = _mm_setzero_ps();
1469             fjx2             = _mm_setzero_ps();
1470             fjy2             = _mm_setzero_ps();
1471             fjz2             = _mm_setzero_ps();
1472             fjx3             = _mm_setzero_ps();
1473             fjy3             = _mm_setzero_ps();
1474             fjz3             = _mm_setzero_ps();
1475
1476             /**************************
1477              * CALCULATE INTERACTIONS *
1478              **************************/
1479
1480             if (gmx_mm_any_lt(rsq00,rcutoff2))
1481             {
1482
1483             r00              = _mm_mul_ps(rsq00,rinv00);
1484
1485             /* Calculate table index by multiplying r with table scale and truncate to integer */
1486             rt               = _mm_mul_ps(r00,vftabscale);
1487             vfitab           = _mm_cvttps_epi32(rt);
1488 #ifdef __XOP__
1489             vfeps            = _mm_frcz_ps(rt);
1490 #else
1491             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1492 #endif
1493             twovfeps         = _mm_add_ps(vfeps,vfeps);
1494             vfitab           = _mm_slli_epi32(vfitab,3);
1495
1496             /* CUBIC SPLINE TABLE DISPERSION */
1497             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1498             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1499             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1500             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1501             _MM_TRANSPOSE4_PS(Y,F,G,H);
1502             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1503             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1504             fvdw6            = _mm_mul_ps(c6_00,FF);
1505
1506             /* CUBIC SPLINE TABLE REPULSION */
1507             vfitab           = _mm_add_epi32(vfitab,ifour);
1508             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1509             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1510             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1511             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1512             _MM_TRANSPOSE4_PS(Y,F,G,H);
1513             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1514             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1515             fvdw12           = _mm_mul_ps(c12_00,FF);
1516             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1517
1518             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1519
1520             fscal            = fvdw;
1521
1522             fscal            = _mm_and_ps(fscal,cutoff_mask);
1523
1524              /* Update vectorial force */
1525             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1526             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1527             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1528
1529             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1530             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1531             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1532
1533             }
1534
1535             /**************************
1536              * CALCULATE INTERACTIONS *
1537              **************************/
1538
1539             if (gmx_mm_any_lt(rsq11,rcutoff2))
1540             {
1541
1542             /* REACTION-FIELD ELECTROSTATICS */
1543             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1544
1545             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1546
1547             fscal            = felec;
1548
1549             fscal            = _mm_and_ps(fscal,cutoff_mask);
1550
1551              /* Update vectorial force */
1552             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1553             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1554             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1555
1556             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1557             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1558             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1559
1560             }
1561
1562             /**************************
1563              * CALCULATE INTERACTIONS *
1564              **************************/
1565
1566             if (gmx_mm_any_lt(rsq12,rcutoff2))
1567             {
1568
1569             /* REACTION-FIELD ELECTROSTATICS */
1570             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1571
1572             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1573
1574             fscal            = felec;
1575
1576             fscal            = _mm_and_ps(fscal,cutoff_mask);
1577
1578              /* Update vectorial force */
1579             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1580             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1581             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1582
1583             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1584             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1585             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1586
1587             }
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             if (gmx_mm_any_lt(rsq13,rcutoff2))
1594             {
1595
1596             /* REACTION-FIELD ELECTROSTATICS */
1597             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1598
1599             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1600
1601             fscal            = felec;
1602
1603             fscal            = _mm_and_ps(fscal,cutoff_mask);
1604
1605              /* Update vectorial force */
1606             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1607             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1608             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1609
1610             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1611             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1612             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1613
1614             }
1615
1616             /**************************
1617              * CALCULATE INTERACTIONS *
1618              **************************/
1619
1620             if (gmx_mm_any_lt(rsq21,rcutoff2))
1621             {
1622
1623             /* REACTION-FIELD ELECTROSTATICS */
1624             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1625
1626             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1627
1628             fscal            = felec;
1629
1630             fscal            = _mm_and_ps(fscal,cutoff_mask);
1631
1632              /* Update vectorial force */
1633             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1634             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1635             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1636
1637             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1638             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1639             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1640
1641             }
1642
1643             /**************************
1644              * CALCULATE INTERACTIONS *
1645              **************************/
1646
1647             if (gmx_mm_any_lt(rsq22,rcutoff2))
1648             {
1649
1650             /* REACTION-FIELD ELECTROSTATICS */
1651             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1652
1653             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1654
1655             fscal            = felec;
1656
1657             fscal            = _mm_and_ps(fscal,cutoff_mask);
1658
1659              /* Update vectorial force */
1660             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1661             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1662             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1663
1664             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1665             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1666             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1667
1668             }
1669
1670             /**************************
1671              * CALCULATE INTERACTIONS *
1672              **************************/
1673
1674             if (gmx_mm_any_lt(rsq23,rcutoff2))
1675             {
1676
1677             /* REACTION-FIELD ELECTROSTATICS */
1678             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1679
1680             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1681
1682             fscal            = felec;
1683
1684             fscal            = _mm_and_ps(fscal,cutoff_mask);
1685
1686              /* Update vectorial force */
1687             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1688             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1689             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1690
1691             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1692             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1693             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1694
1695             }
1696
1697             /**************************
1698              * CALCULATE INTERACTIONS *
1699              **************************/
1700
1701             if (gmx_mm_any_lt(rsq31,rcutoff2))
1702             {
1703
1704             /* REACTION-FIELD ELECTROSTATICS */
1705             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1706
1707             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1708
1709             fscal            = felec;
1710
1711             fscal            = _mm_and_ps(fscal,cutoff_mask);
1712
1713              /* Update vectorial force */
1714             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1715             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1716             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1717
1718             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1719             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1720             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1721
1722             }
1723
1724             /**************************
1725              * CALCULATE INTERACTIONS *
1726              **************************/
1727
1728             if (gmx_mm_any_lt(rsq32,rcutoff2))
1729             {
1730
1731             /* REACTION-FIELD ELECTROSTATICS */
1732             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1733
1734             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1735
1736             fscal            = felec;
1737
1738             fscal            = _mm_and_ps(fscal,cutoff_mask);
1739
1740              /* Update vectorial force */
1741             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1742             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1743             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1744
1745             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1746             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1747             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1748
1749             }
1750
1751             /**************************
1752              * CALCULATE INTERACTIONS *
1753              **************************/
1754
1755             if (gmx_mm_any_lt(rsq33,rcutoff2))
1756             {
1757
1758             /* REACTION-FIELD ELECTROSTATICS */
1759             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1760
1761             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1762
1763             fscal            = felec;
1764
1765             fscal            = _mm_and_ps(fscal,cutoff_mask);
1766
1767              /* Update vectorial force */
1768             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1769             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1770             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1771
1772             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1773             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1774             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1775
1776             }
1777
1778             fjptrA             = f+j_coord_offsetA;
1779             fjptrB             = f+j_coord_offsetB;
1780             fjptrC             = f+j_coord_offsetC;
1781             fjptrD             = f+j_coord_offsetD;
1782
1783             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1784                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1785                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1786
1787             /* Inner loop uses 354 flops */
1788         }
1789
1790         if(jidx<j_index_end)
1791         {
1792
1793             /* Get j neighbor index, and coordinate index */
1794             jnrlistA         = jjnr[jidx];
1795             jnrlistB         = jjnr[jidx+1];
1796             jnrlistC         = jjnr[jidx+2];
1797             jnrlistD         = jjnr[jidx+3];
1798             /* Sign of each element will be negative for non-real atoms.
1799              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1800              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1801              */
1802             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1803             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1804             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1805             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1806             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1807             j_coord_offsetA  = DIM*jnrA;
1808             j_coord_offsetB  = DIM*jnrB;
1809             j_coord_offsetC  = DIM*jnrC;
1810             j_coord_offsetD  = DIM*jnrD;
1811
1812             /* load j atom coordinates */
1813             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1814                                               x+j_coord_offsetC,x+j_coord_offsetD,
1815                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1816                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1817
1818             /* Calculate displacement vector */
1819             dx00             = _mm_sub_ps(ix0,jx0);
1820             dy00             = _mm_sub_ps(iy0,jy0);
1821             dz00             = _mm_sub_ps(iz0,jz0);
1822             dx11             = _mm_sub_ps(ix1,jx1);
1823             dy11             = _mm_sub_ps(iy1,jy1);
1824             dz11             = _mm_sub_ps(iz1,jz1);
1825             dx12             = _mm_sub_ps(ix1,jx2);
1826             dy12             = _mm_sub_ps(iy1,jy2);
1827             dz12             = _mm_sub_ps(iz1,jz2);
1828             dx13             = _mm_sub_ps(ix1,jx3);
1829             dy13             = _mm_sub_ps(iy1,jy3);
1830             dz13             = _mm_sub_ps(iz1,jz3);
1831             dx21             = _mm_sub_ps(ix2,jx1);
1832             dy21             = _mm_sub_ps(iy2,jy1);
1833             dz21             = _mm_sub_ps(iz2,jz1);
1834             dx22             = _mm_sub_ps(ix2,jx2);
1835             dy22             = _mm_sub_ps(iy2,jy2);
1836             dz22             = _mm_sub_ps(iz2,jz2);
1837             dx23             = _mm_sub_ps(ix2,jx3);
1838             dy23             = _mm_sub_ps(iy2,jy3);
1839             dz23             = _mm_sub_ps(iz2,jz3);
1840             dx31             = _mm_sub_ps(ix3,jx1);
1841             dy31             = _mm_sub_ps(iy3,jy1);
1842             dz31             = _mm_sub_ps(iz3,jz1);
1843             dx32             = _mm_sub_ps(ix3,jx2);
1844             dy32             = _mm_sub_ps(iy3,jy2);
1845             dz32             = _mm_sub_ps(iz3,jz2);
1846             dx33             = _mm_sub_ps(ix3,jx3);
1847             dy33             = _mm_sub_ps(iy3,jy3);
1848             dz33             = _mm_sub_ps(iz3,jz3);
1849
1850             /* Calculate squared distance and things based on it */
1851             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1852             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1853             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1854             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1855             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1856             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1857             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1858             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1859             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1860             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1861
1862             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1863             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1864             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1865             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1866             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1867             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1868             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1869             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1870             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1871             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1872
1873             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1874             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1875             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1876             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1877             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1878             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1879             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1880             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1881             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1882
1883             fjx0             = _mm_setzero_ps();
1884             fjy0             = _mm_setzero_ps();
1885             fjz0             = _mm_setzero_ps();
1886             fjx1             = _mm_setzero_ps();
1887             fjy1             = _mm_setzero_ps();
1888             fjz1             = _mm_setzero_ps();
1889             fjx2             = _mm_setzero_ps();
1890             fjy2             = _mm_setzero_ps();
1891             fjz2             = _mm_setzero_ps();
1892             fjx3             = _mm_setzero_ps();
1893             fjy3             = _mm_setzero_ps();
1894             fjz3             = _mm_setzero_ps();
1895
1896             /**************************
1897              * CALCULATE INTERACTIONS *
1898              **************************/
1899
1900             if (gmx_mm_any_lt(rsq00,rcutoff2))
1901             {
1902
1903             r00              = _mm_mul_ps(rsq00,rinv00);
1904             r00              = _mm_andnot_ps(dummy_mask,r00);
1905
1906             /* Calculate table index by multiplying r with table scale and truncate to integer */
1907             rt               = _mm_mul_ps(r00,vftabscale);
1908             vfitab           = _mm_cvttps_epi32(rt);
1909 #ifdef __XOP__
1910             vfeps            = _mm_frcz_ps(rt);
1911 #else
1912             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1913 #endif
1914             twovfeps         = _mm_add_ps(vfeps,vfeps);
1915             vfitab           = _mm_slli_epi32(vfitab,3);
1916
1917             /* CUBIC SPLINE TABLE DISPERSION */
1918             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1919             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1920             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1921             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1922             _MM_TRANSPOSE4_PS(Y,F,G,H);
1923             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1924             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1925             fvdw6            = _mm_mul_ps(c6_00,FF);
1926
1927             /* CUBIC SPLINE TABLE REPULSION */
1928             vfitab           = _mm_add_epi32(vfitab,ifour);
1929             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1930             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1931             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1932             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1933             _MM_TRANSPOSE4_PS(Y,F,G,H);
1934             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1935             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1936             fvdw12           = _mm_mul_ps(c12_00,FF);
1937             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1938
1939             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1940
1941             fscal            = fvdw;
1942
1943             fscal            = _mm_and_ps(fscal,cutoff_mask);
1944
1945             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1946
1947              /* Update vectorial force */
1948             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1949             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1950             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1951
1952             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1953             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1954             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1955
1956             }
1957
1958             /**************************
1959              * CALCULATE INTERACTIONS *
1960              **************************/
1961
1962             if (gmx_mm_any_lt(rsq11,rcutoff2))
1963             {
1964
1965             /* REACTION-FIELD ELECTROSTATICS */
1966             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1967
1968             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1969
1970             fscal            = felec;
1971
1972             fscal            = _mm_and_ps(fscal,cutoff_mask);
1973
1974             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1975
1976              /* Update vectorial force */
1977             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1978             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1979             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1980
1981             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1982             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1983             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1984
1985             }
1986
1987             /**************************
1988              * CALCULATE INTERACTIONS *
1989              **************************/
1990
1991             if (gmx_mm_any_lt(rsq12,rcutoff2))
1992             {
1993
1994             /* REACTION-FIELD ELECTROSTATICS */
1995             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1996
1997             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1998
1999             fscal            = felec;
2000
2001             fscal            = _mm_and_ps(fscal,cutoff_mask);
2002
2003             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2004
2005              /* Update vectorial force */
2006             fix1             = _mm_macc_ps(dx12,fscal,fix1);
2007             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
2008             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
2009
2010             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
2011             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
2012             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
2013
2014             }
2015
2016             /**************************
2017              * CALCULATE INTERACTIONS *
2018              **************************/
2019
2020             if (gmx_mm_any_lt(rsq13,rcutoff2))
2021             {
2022
2023             /* REACTION-FIELD ELECTROSTATICS */
2024             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
2025
2026             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
2027
2028             fscal            = felec;
2029
2030             fscal            = _mm_and_ps(fscal,cutoff_mask);
2031
2032             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2033
2034              /* Update vectorial force */
2035             fix1             = _mm_macc_ps(dx13,fscal,fix1);
2036             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
2037             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
2038
2039             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
2040             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
2041             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
2042
2043             }
2044
2045             /**************************
2046              * CALCULATE INTERACTIONS *
2047              **************************/
2048
2049             if (gmx_mm_any_lt(rsq21,rcutoff2))
2050             {
2051
2052             /* REACTION-FIELD ELECTROSTATICS */
2053             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
2054
2055             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2056
2057             fscal            = felec;
2058
2059             fscal            = _mm_and_ps(fscal,cutoff_mask);
2060
2061             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2062
2063              /* Update vectorial force */
2064             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2065             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2066             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2067
2068             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2069             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2070             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2071
2072             }
2073
2074             /**************************
2075              * CALCULATE INTERACTIONS *
2076              **************************/
2077
2078             if (gmx_mm_any_lt(rsq22,rcutoff2))
2079             {
2080
2081             /* REACTION-FIELD ELECTROSTATICS */
2082             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
2083
2084             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2085
2086             fscal            = felec;
2087
2088             fscal            = _mm_and_ps(fscal,cutoff_mask);
2089
2090             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2091
2092              /* Update vectorial force */
2093             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2094             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2095             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2096
2097             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2098             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2099             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2100
2101             }
2102
2103             /**************************
2104              * CALCULATE INTERACTIONS *
2105              **************************/
2106
2107             if (gmx_mm_any_lt(rsq23,rcutoff2))
2108             {
2109
2110             /* REACTION-FIELD ELECTROSTATICS */
2111             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
2112
2113             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
2114
2115             fscal            = felec;
2116
2117             fscal            = _mm_and_ps(fscal,cutoff_mask);
2118
2119             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2120
2121              /* Update vectorial force */
2122             fix2             = _mm_macc_ps(dx23,fscal,fix2);
2123             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
2124             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
2125
2126             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
2127             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
2128             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
2129
2130             }
2131
2132             /**************************
2133              * CALCULATE INTERACTIONS *
2134              **************************/
2135
2136             if (gmx_mm_any_lt(rsq31,rcutoff2))
2137             {
2138
2139             /* REACTION-FIELD ELECTROSTATICS */
2140             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
2141
2142             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2143
2144             fscal            = felec;
2145
2146             fscal            = _mm_and_ps(fscal,cutoff_mask);
2147
2148             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2149
2150              /* Update vectorial force */
2151             fix3             = _mm_macc_ps(dx31,fscal,fix3);
2152             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
2153             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
2154
2155             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
2156             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
2157             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
2158
2159             }
2160
2161             /**************************
2162              * CALCULATE INTERACTIONS *
2163              **************************/
2164
2165             if (gmx_mm_any_lt(rsq32,rcutoff2))
2166             {
2167
2168             /* REACTION-FIELD ELECTROSTATICS */
2169             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
2170
2171             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2172
2173             fscal            = felec;
2174
2175             fscal            = _mm_and_ps(fscal,cutoff_mask);
2176
2177             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2178
2179              /* Update vectorial force */
2180             fix3             = _mm_macc_ps(dx32,fscal,fix3);
2181             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
2182             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
2183
2184             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
2185             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
2186             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
2187
2188             }
2189
2190             /**************************
2191              * CALCULATE INTERACTIONS *
2192              **************************/
2193
2194             if (gmx_mm_any_lt(rsq33,rcutoff2))
2195             {
2196
2197             /* REACTION-FIELD ELECTROSTATICS */
2198             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
2199
2200             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2201
2202             fscal            = felec;
2203
2204             fscal            = _mm_and_ps(fscal,cutoff_mask);
2205
2206             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2207
2208              /* Update vectorial force */
2209             fix3             = _mm_macc_ps(dx33,fscal,fix3);
2210             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
2211             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
2212
2213             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
2214             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
2215             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
2216
2217             }
2218
2219             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2220             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2221             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2222             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2223
2224             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2225                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2226                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2227
2228             /* Inner loop uses 355 flops */
2229         }
2230
2231         /* End of innermost loop */
2232
2233         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2234                                               f+i_coord_offset,fshift+i_shift_offset);
2235
2236         /* Increment number of inner iterations */
2237         inneriter                  += j_index_end - j_index_start;
2238
2239         /* Outer loop uses 24 flops */
2240     }
2241
2242     /* Increment number of outer iterations */
2243     outeriter        += nri;
2244
2245     /* Update outer/inner flops */
2246
2247     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*355);
2248 }