src/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_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_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse4_1_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse4_1_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 SSE, 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           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128i          vfitab;
  99     __m128i          ifour       = _mm_set1_epi32(4);
 100     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 101     real             *vftab;
 102     __m128           dummy_mask,cutoff_mask;
 103     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 104     __m128           one     = _mm_set1_ps(1.0);
 105     __m128           two     = _mm_set1_ps(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_ps(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = _mm_set1_ps(fr->ic->k_rf);
 120     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 121     crf              = _mm_set1_ps(fr->ic->c_rf);
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     vftab            = kernel_data->table_vdw->data;
 127     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 132     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 133     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 134     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 135
 136     jq0              = _mm_set1_ps(charge[inr+0]);
 137     jq1              = _mm_set1_ps(charge[inr+1]);
 138     jq2              = _mm_set1_ps(charge[inr+2]);
 139     vdwjidx0A        = 2*vdwtype[inr+0];
 140     qq00             = _mm_mul_ps(iq0,jq0);
 141     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 142     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 143     qq01             = _mm_mul_ps(iq0,jq1);
 144     qq02             = _mm_mul_ps(iq0,jq2);
 145     qq10             = _mm_mul_ps(iq1,jq0);
 146     qq11             = _mm_mul_ps(iq1,jq1);
 147     qq12             = _mm_mul_ps(iq1,jq2);
 148     qq20             = _mm_mul_ps(iq2,jq0);
 149     qq21             = _mm_mul_ps(iq2,jq1);
 150     qq22             = _mm_mul_ps(iq2,jq2);
 151
 152     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 153     rcutoff_scalar   = fr->rcoulomb;
 154     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 155     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 156
 157     /* Avoid stupid compiler warnings */
 158     jnrA = jnrB = jnrC = jnrD = 0;
 159     j_coord_offsetA = 0;
 160     j_coord_offsetB = 0;
 161     j_coord_offsetC = 0;
 162     j_coord_offsetD = 0;
 163
 164     outeriter        = 0;
 165     inneriter        = 0;
 166
 167     for(iidx=0;iidx<4*DIM;iidx++)
 168     {
 169         scratch[iidx] = 0.0;
 170     }
 171
 172     /* Start outer loop over neighborlists */
 173     for(iidx=0; iidx<nri; iidx++)
 174     {
 175         /* Load shift vector for this list */
 176         i_shift_offset   = DIM*shiftidx[iidx];
 177
 178         /* Load limits for loop over neighbors */
 179         j_index_start    = jindex[iidx];
 180         j_index_end      = jindex[iidx+1];
 181
 182         /* Get outer coordinate index */
 183         inr              = iinr[iidx];
 184         i_coord_offset   = DIM*inr;
 185
 186         /* Load i particle coords and add shift vector */
 187         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 188                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 189
 190         fix0             = _mm_setzero_ps();
 191         fiy0             = _mm_setzero_ps();
 192         fiz0             = _mm_setzero_ps();
 193         fix1             = _mm_setzero_ps();
 194         fiy1             = _mm_setzero_ps();
 195         fiz1             = _mm_setzero_ps();
 196         fix2             = _mm_setzero_ps();
 197         fiy2             = _mm_setzero_ps();
 198         fiz2             = _mm_setzero_ps();
 199
 200         /* Reset potential sums */
 201         velecsum         = _mm_setzero_ps();
 202         vvdwsum          = _mm_setzero_ps();
 203
 204         /* Start inner kernel loop */
 205         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 206         {
 207
 208             /* Get j neighbor index, and coordinate index */
 209             jnrA             = jjnr[jidx];
 210             jnrB             = jjnr[jidx+1];
 211             jnrC             = jjnr[jidx+2];
 212             jnrD             = jjnr[jidx+3];
 213             j_coord_offsetA  = DIM*jnrA;
 214             j_coord_offsetB  = DIM*jnrB;
 215             j_coord_offsetC  = DIM*jnrC;
 216             j_coord_offsetD  = DIM*jnrD;
 217
 218             /* load j atom coordinates */
 219             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 220                                               x+j_coord_offsetC,x+j_coord_offsetD,
 221                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 222
 223             /* Calculate displacement vector */
 224             dx00             = _mm_sub_ps(ix0,jx0);
 225             dy00             = _mm_sub_ps(iy0,jy0);
 226             dz00             = _mm_sub_ps(iz0,jz0);
 227             dx01             = _mm_sub_ps(ix0,jx1);
 228             dy01             = _mm_sub_ps(iy0,jy1);
 229             dz01             = _mm_sub_ps(iz0,jz1);
 230             dx02             = _mm_sub_ps(ix0,jx2);
 231             dy02             = _mm_sub_ps(iy0,jy2);
 232             dz02             = _mm_sub_ps(iz0,jz2);
 233             dx10             = _mm_sub_ps(ix1,jx0);
 234             dy10             = _mm_sub_ps(iy1,jy0);
 235             dz10             = _mm_sub_ps(iz1,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             dx20             = _mm_sub_ps(ix2,jx0);
 243             dy20             = _mm_sub_ps(iy2,jy0);
 244             dz20             = _mm_sub_ps(iz2,jz0);
 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
 252             /* Calculate squared distance and things based on it */
 253             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 254             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 255             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 256             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 257             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 258             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 259             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 260             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 261             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 262
 263             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 264             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 265             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 266             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 267             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 268             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 269             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 270             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 271             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 272
 273             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 274             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 275             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 276             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 277             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 278             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 279             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 280             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 281             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 282
 283             fjx0             = _mm_setzero_ps();
 284             fjy0             = _mm_setzero_ps();
 285             fjz0             = _mm_setzero_ps();
 286             fjx1             = _mm_setzero_ps();
 287             fjy1             = _mm_setzero_ps();
 288             fjz1             = _mm_setzero_ps();
 289             fjx2             = _mm_setzero_ps();
 290             fjy2             = _mm_setzero_ps();
 291             fjz2             = _mm_setzero_ps();
 292
 293             /**************************
 294              * CALCULATE INTERACTIONS *
 295              **************************/
 296
 297             if (gmx_mm_any_lt(rsq00,rcutoff2))
 298             {
 299
 300             r00              = _mm_mul_ps(rsq00,rinv00);
 301
 302             /* Calculate table index by multiplying r with table scale and truncate to integer */
 303             rt               = _mm_mul_ps(r00,vftabscale);
 304             vfitab           = _mm_cvttps_epi32(rt);
 305             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 306             vfitab           = _mm_slli_epi32(vfitab,3);
 307
 308             /* REACTION-FIELD ELECTROSTATICS */
 309             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 310             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 311
 312             /* CUBIC SPLINE TABLE DISPERSION */
 313             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 314             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 315             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 316             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 317             _MM_TRANSPOSE4_PS(Y,F,G,H);
 318             Heps             = _mm_mul_ps(vfeps,H);
 319             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 320             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 321             vvdw6            = _mm_mul_ps(c6_00,VV);
 322             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 323             fvdw6            = _mm_mul_ps(c6_00,FF);
 324
 325             /* CUBIC SPLINE TABLE REPULSION */
 326             vfitab           = _mm_add_epi32(vfitab,ifour);
 327             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 328             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 329             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 330             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 331             _MM_TRANSPOSE4_PS(Y,F,G,H);
 332             Heps             = _mm_mul_ps(vfeps,H);
 333             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 334             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 335             vvdw12           = _mm_mul_ps(c12_00,VV);
 336             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 337             fvdw12           = _mm_mul_ps(c12_00,FF);
 338             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 339             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 340
 341             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             velec            = _mm_and_ps(velec,cutoff_mask);
 345             velecsum         = _mm_add_ps(velecsum,velec);
 346             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 347             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 348
 349             fscal            = _mm_add_ps(felec,fvdw);
 350
 351             fscal            = _mm_and_ps(fscal,cutoff_mask);
 352
 353             /* Calculate temporary vectorial force */
 354             tx               = _mm_mul_ps(fscal,dx00);
 355             ty               = _mm_mul_ps(fscal,dy00);
 356             tz               = _mm_mul_ps(fscal,dz00);
 357
 358             /* Update vectorial force */
 359             fix0             = _mm_add_ps(fix0,tx);
 360             fiy0             = _mm_add_ps(fiy0,ty);
 361             fiz0             = _mm_add_ps(fiz0,tz);
 362
 363             fjx0             = _mm_add_ps(fjx0,tx);
 364             fjy0             = _mm_add_ps(fjy0,ty);
 365             fjz0             = _mm_add_ps(fjz0,tz);
 366
 367             }
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             if (gmx_mm_any_lt(rsq01,rcutoff2))
 374             {
 375
 376             /* REACTION-FIELD ELECTROSTATICS */
 377             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 378             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 379
 380             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velec            = _mm_and_ps(velec,cutoff_mask);
 384             velecsum         = _mm_add_ps(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             fscal            = _mm_and_ps(fscal,cutoff_mask);
 389
 390             /* Calculate temporary vectorial force */
 391             tx               = _mm_mul_ps(fscal,dx01);
 392             ty               = _mm_mul_ps(fscal,dy01);
 393             tz               = _mm_mul_ps(fscal,dz01);
 394
 395             /* Update vectorial force */
 396             fix0             = _mm_add_ps(fix0,tx);
 397             fiy0             = _mm_add_ps(fiy0,ty);
 398             fiz0             = _mm_add_ps(fiz0,tz);
 399
 400             fjx1             = _mm_add_ps(fjx1,tx);
 401             fjy1             = _mm_add_ps(fjy1,ty);
 402             fjz1             = _mm_add_ps(fjz1,tz);
 403
 404             }
 405
 406             /**************************
 407              * CALCULATE INTERACTIONS *
 408              **************************/
 409
 410             if (gmx_mm_any_lt(rsq02,rcutoff2))
 411             {
 412
 413             /* REACTION-FIELD ELECTROSTATICS */
 414             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 415             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 416
 417             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velec            = _mm_and_ps(velec,cutoff_mask);
 421             velecsum         = _mm_add_ps(velecsum,velec);
 422
 423             fscal            = felec;
 424
 425             fscal            = _mm_and_ps(fscal,cutoff_mask);
 426
 427             /* Calculate temporary vectorial force */
 428             tx               = _mm_mul_ps(fscal,dx02);
 429             ty               = _mm_mul_ps(fscal,dy02);
 430             tz               = _mm_mul_ps(fscal,dz02);
 431
 432             /* Update vectorial force */
 433             fix0             = _mm_add_ps(fix0,tx);
 434             fiy0             = _mm_add_ps(fiy0,ty);
 435             fiz0             = _mm_add_ps(fiz0,tz);
 436
 437             fjx2             = _mm_add_ps(fjx2,tx);
 438             fjy2             = _mm_add_ps(fjy2,ty);
 439             fjz2             = _mm_add_ps(fjz2,tz);
 440
 441             }
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             if (gmx_mm_any_lt(rsq10,rcutoff2))
 448             {
 449
 450             /* REACTION-FIELD ELECTROSTATICS */
 451             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 452             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 453
 454             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             velec            = _mm_and_ps(velec,cutoff_mask);
 458             velecsum         = _mm_add_ps(velecsum,velec);
 459
 460             fscal            = felec;
 461
 462             fscal            = _mm_and_ps(fscal,cutoff_mask);
 463
 464             /* Calculate temporary vectorial force */
 465             tx               = _mm_mul_ps(fscal,dx10);
 466             ty               = _mm_mul_ps(fscal,dy10);
 467             tz               = _mm_mul_ps(fscal,dz10);
 468
 469             /* Update vectorial force */
 470             fix1             = _mm_add_ps(fix1,tx);
 471             fiy1             = _mm_add_ps(fiy1,ty);
 472             fiz1             = _mm_add_ps(fiz1,tz);
 473
 474             fjx0             = _mm_add_ps(fjx0,tx);
 475             fjy0             = _mm_add_ps(fjy0,ty);
 476             fjz0             = _mm_add_ps(fjz0,tz);
 477
 478             }
 479
 480             /**************************
 481              * CALCULATE INTERACTIONS *
 482              **************************/
 483
 484             if (gmx_mm_any_lt(rsq11,rcutoff2))
 485             {
 486
 487             /* REACTION-FIELD ELECTROSTATICS */
 488             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 489             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 490
 491             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 492
 493             /* Update potential sum for this i atom from the interaction with this j atom. */
 494             velec            = _mm_and_ps(velec,cutoff_mask);
 495             velecsum         = _mm_add_ps(velecsum,velec);
 496
 497             fscal            = felec;
 498
 499             fscal            = _mm_and_ps(fscal,cutoff_mask);
 500
 501             /* Calculate temporary vectorial force */
 502             tx               = _mm_mul_ps(fscal,dx11);
 503             ty               = _mm_mul_ps(fscal,dy11);
 504             tz               = _mm_mul_ps(fscal,dz11);
 505
 506             /* Update vectorial force */
 507             fix1             = _mm_add_ps(fix1,tx);
 508             fiy1             = _mm_add_ps(fiy1,ty);
 509             fiz1             = _mm_add_ps(fiz1,tz);
 510
 511             fjx1             = _mm_add_ps(fjx1,tx);
 512             fjy1             = _mm_add_ps(fjy1,ty);
 513             fjz1             = _mm_add_ps(fjz1,tz);
 514
 515             }
 516
 517             /**************************
 518              * CALCULATE INTERACTIONS *
 519              **************************/
 520
 521             if (gmx_mm_any_lt(rsq12,rcutoff2))
 522             {
 523
 524             /* REACTION-FIELD ELECTROSTATICS */
 525             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 526             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 527
 528             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 529
 530             /* Update potential sum for this i atom from the interaction with this j atom. */
 531             velec            = _mm_and_ps(velec,cutoff_mask);
 532             velecsum         = _mm_add_ps(velecsum,velec);
 533
 534             fscal            = felec;
 535
 536             fscal            = _mm_and_ps(fscal,cutoff_mask);
 537
 538             /* Calculate temporary vectorial force */
 539             tx               = _mm_mul_ps(fscal,dx12);
 540             ty               = _mm_mul_ps(fscal,dy12);
 541             tz               = _mm_mul_ps(fscal,dz12);
 542
 543             /* Update vectorial force */
 544             fix1             = _mm_add_ps(fix1,tx);
 545             fiy1             = _mm_add_ps(fiy1,ty);
 546             fiz1             = _mm_add_ps(fiz1,tz);
 547
 548             fjx2             = _mm_add_ps(fjx2,tx);
 549             fjy2             = _mm_add_ps(fjy2,ty);
 550             fjz2             = _mm_add_ps(fjz2,tz);
 551
 552             }
 553
 554             /**************************
 555              * CALCULATE INTERACTIONS *
 556              **************************/
 557
 558             if (gmx_mm_any_lt(rsq20,rcutoff2))
 559             {
 560
 561             /* REACTION-FIELD ELECTROSTATICS */
 562             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 563             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 564
 565             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 566
 567             /* Update potential sum for this i atom from the interaction with this j atom. */
 568             velec            = _mm_and_ps(velec,cutoff_mask);
 569             velecsum         = _mm_add_ps(velecsum,velec);
 570
 571             fscal            = felec;
 572
 573             fscal            = _mm_and_ps(fscal,cutoff_mask);
 574
 575             /* Calculate temporary vectorial force */
 576             tx               = _mm_mul_ps(fscal,dx20);
 577             ty               = _mm_mul_ps(fscal,dy20);
 578             tz               = _mm_mul_ps(fscal,dz20);
 579
 580             /* Update vectorial force */
 581             fix2             = _mm_add_ps(fix2,tx);
 582             fiy2             = _mm_add_ps(fiy2,ty);
 583             fiz2             = _mm_add_ps(fiz2,tz);
 584
 585             fjx0             = _mm_add_ps(fjx0,tx);
 586             fjy0             = _mm_add_ps(fjy0,ty);
 587             fjz0             = _mm_add_ps(fjz0,tz);
 588
 589             }
 590
 591             /**************************
 592              * CALCULATE INTERACTIONS *
 593              **************************/
 594
 595             if (gmx_mm_any_lt(rsq21,rcutoff2))
 596             {
 597
 598             /* REACTION-FIELD ELECTROSTATICS */
 599             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 600             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 601
 602             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 603
 604             /* Update potential sum for this i atom from the interaction with this j atom. */
 605             velec            = _mm_and_ps(velec,cutoff_mask);
 606             velecsum         = _mm_add_ps(velecsum,velec);
 607
 608             fscal            = felec;
 609
 610             fscal            = _mm_and_ps(fscal,cutoff_mask);
 611
 612             /* Calculate temporary vectorial force */
 613             tx               = _mm_mul_ps(fscal,dx21);
 614             ty               = _mm_mul_ps(fscal,dy21);
 615             tz               = _mm_mul_ps(fscal,dz21);
 616
 617             /* Update vectorial force */
 618             fix2             = _mm_add_ps(fix2,tx);
 619             fiy2             = _mm_add_ps(fiy2,ty);
 620             fiz2             = _mm_add_ps(fiz2,tz);
 621
 622             fjx1             = _mm_add_ps(fjx1,tx);
 623             fjy1             = _mm_add_ps(fjy1,ty);
 624             fjz1             = _mm_add_ps(fjz1,tz);
 625
 626             }
 627
 628             /**************************
 629              * CALCULATE INTERACTIONS *
 630              **************************/
 631
 632             if (gmx_mm_any_lt(rsq22,rcutoff2))
 633             {
 634
 635             /* REACTION-FIELD ELECTROSTATICS */
 636             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 637             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 638
 639             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 640
 641             /* Update potential sum for this i atom from the interaction with this j atom. */
 642             velec            = _mm_and_ps(velec,cutoff_mask);
 643             velecsum         = _mm_add_ps(velecsum,velec);
 644
 645             fscal            = felec;
 646
 647             fscal            = _mm_and_ps(fscal,cutoff_mask);
 648
 649             /* Calculate temporary vectorial force */
 650             tx               = _mm_mul_ps(fscal,dx22);
 651             ty               = _mm_mul_ps(fscal,dy22);
 652             tz               = _mm_mul_ps(fscal,dz22);
 653
 654             /* Update vectorial force */
 655             fix2             = _mm_add_ps(fix2,tx);
 656             fiy2             = _mm_add_ps(fiy2,ty);
 657             fiz2             = _mm_add_ps(fiz2,tz);
 658
 659             fjx2             = _mm_add_ps(fjx2,tx);
 660             fjy2             = _mm_add_ps(fjy2,ty);
 661             fjz2             = _mm_add_ps(fjz2,tz);
 662
 663             }
 664
 665             fjptrA             = f+j_coord_offsetA;
 666             fjptrB             = f+j_coord_offsetB;
 667             fjptrC             = f+j_coord_offsetC;
 668             fjptrD             = f+j_coord_offsetD;
 669
 670             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 671                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 672
 673             /* Inner loop uses 360 flops */
 674         }
 675
 676         if(jidx<j_index_end)
 677         {
 678
 679             /* Get j neighbor index, and coordinate index */
 680             jnrlistA         = jjnr[jidx];
 681             jnrlistB         = jjnr[jidx+1];
 682             jnrlistC         = jjnr[jidx+2];
 683             jnrlistD         = jjnr[jidx+3];
 684             /* Sign of each element will be negative for non-real atoms.
 685              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 686              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 687              */
 688             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 689             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 690             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 691             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 692             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 693             j_coord_offsetA  = DIM*jnrA;
 694             j_coord_offsetB  = DIM*jnrB;
 695             j_coord_offsetC  = DIM*jnrC;
 696             j_coord_offsetD  = DIM*jnrD;
 697
 698             /* load j atom coordinates */
 699             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 700                                               x+j_coord_offsetC,x+j_coord_offsetD,
 701                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 702
 703             /* Calculate displacement vector */
 704             dx00             = _mm_sub_ps(ix0,jx0);
 705             dy00             = _mm_sub_ps(iy0,jy0);
 706             dz00             = _mm_sub_ps(iz0,jz0);
 707             dx01             = _mm_sub_ps(ix0,jx1);
 708             dy01             = _mm_sub_ps(iy0,jy1);
 709             dz01             = _mm_sub_ps(iz0,jz1);
 710             dx02             = _mm_sub_ps(ix0,jx2);
 711             dy02             = _mm_sub_ps(iy0,jy2);
 712             dz02             = _mm_sub_ps(iz0,jz2);
 713             dx10             = _mm_sub_ps(ix1,jx0);
 714             dy10             = _mm_sub_ps(iy1,jy0);
 715             dz10             = _mm_sub_ps(iz1,jz0);
 716             dx11             = _mm_sub_ps(ix1,jx1);
 717             dy11             = _mm_sub_ps(iy1,jy1);
 718             dz11             = _mm_sub_ps(iz1,jz1);
 719             dx12             = _mm_sub_ps(ix1,jx2);
 720             dy12             = _mm_sub_ps(iy1,jy2);
 721             dz12             = _mm_sub_ps(iz1,jz2);
 722             dx20             = _mm_sub_ps(ix2,jx0);
 723             dy20             = _mm_sub_ps(iy2,jy0);
 724             dz20             = _mm_sub_ps(iz2,jz0);
 725             dx21             = _mm_sub_ps(ix2,jx1);
 726             dy21             = _mm_sub_ps(iy2,jy1);
 727             dz21             = _mm_sub_ps(iz2,jz1);
 728             dx22             = _mm_sub_ps(ix2,jx2);
 729             dy22             = _mm_sub_ps(iy2,jy2);
 730             dz22             = _mm_sub_ps(iz2,jz2);
 731
 732             /* Calculate squared distance and things based on it */
 733             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 734             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 735             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 736             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 737             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 738             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 739             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 740             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 741             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 742
 743             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 744             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 745             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 746             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 747             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 748             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 749             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 750             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 751             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 752
 753             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 754             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 755             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 756             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 757             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 758             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 759             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 760             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 761             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 762
 763             fjx0             = _mm_setzero_ps();
 764             fjy0             = _mm_setzero_ps();
 765             fjz0             = _mm_setzero_ps();
 766             fjx1             = _mm_setzero_ps();
 767             fjy1             = _mm_setzero_ps();
 768             fjz1             = _mm_setzero_ps();
 769             fjx2             = _mm_setzero_ps();
 770             fjy2             = _mm_setzero_ps();
 771             fjz2             = _mm_setzero_ps();
 772
 773             /**************************
 774              * CALCULATE INTERACTIONS *
 775              **************************/
 776
 777             if (gmx_mm_any_lt(rsq00,rcutoff2))
 778             {
 779
 780             r00              = _mm_mul_ps(rsq00,rinv00);
 781             r00              = _mm_andnot_ps(dummy_mask,r00);
 782
 783             /* Calculate table index by multiplying r with table scale and truncate to integer */
 784             rt               = _mm_mul_ps(r00,vftabscale);
 785             vfitab           = _mm_cvttps_epi32(rt);
 786             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 787             vfitab           = _mm_slli_epi32(vfitab,3);
 788
 789             /* REACTION-FIELD ELECTROSTATICS */
 790             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 791             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 792
 793             /* CUBIC SPLINE TABLE DISPERSION */
 794             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 795             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 796             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 797             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 798             _MM_TRANSPOSE4_PS(Y,F,G,H);
 799             Heps             = _mm_mul_ps(vfeps,H);
 800             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 801             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 802             vvdw6            = _mm_mul_ps(c6_00,VV);
 803             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 804             fvdw6            = _mm_mul_ps(c6_00,FF);
 805
 806             /* CUBIC SPLINE TABLE REPULSION */
 807             vfitab           = _mm_add_epi32(vfitab,ifour);
 808             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 809             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 810             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 811             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 812             _MM_TRANSPOSE4_PS(Y,F,G,H);
 813             Heps             = _mm_mul_ps(vfeps,H);
 814             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 815             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 816             vvdw12           = _mm_mul_ps(c12_00,VV);
 817             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 818             fvdw12           = _mm_mul_ps(c12_00,FF);
 819             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 820             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 821
 822             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 823
 824             /* Update potential sum for this i atom from the interaction with this j atom. */
 825             velec            = _mm_and_ps(velec,cutoff_mask);
 826             velec            = _mm_andnot_ps(dummy_mask,velec);
 827             velecsum         = _mm_add_ps(velecsum,velec);
 828             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 829             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 830             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 831
 832             fscal            = _mm_add_ps(felec,fvdw);
 833
 834             fscal            = _mm_and_ps(fscal,cutoff_mask);
 835
 836             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 837
 838             /* Calculate temporary vectorial force */
 839             tx               = _mm_mul_ps(fscal,dx00);
 840             ty               = _mm_mul_ps(fscal,dy00);
 841             tz               = _mm_mul_ps(fscal,dz00);
 842
 843             /* Update vectorial force */
 844             fix0             = _mm_add_ps(fix0,tx);
 845             fiy0             = _mm_add_ps(fiy0,ty);
 846             fiz0             = _mm_add_ps(fiz0,tz);
 847
 848             fjx0             = _mm_add_ps(fjx0,tx);
 849             fjy0             = _mm_add_ps(fjy0,ty);
 850             fjz0             = _mm_add_ps(fjz0,tz);
 851
 852             }
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             if (gmx_mm_any_lt(rsq01,rcutoff2))
 859             {
 860
 861             /* REACTION-FIELD ELECTROSTATICS */
 862             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 863             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 864
 865             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 866
 867             /* Update potential sum for this i atom from the interaction with this j atom. */
 868             velec            = _mm_and_ps(velec,cutoff_mask);
 869             velec            = _mm_andnot_ps(dummy_mask,velec);
 870             velecsum         = _mm_add_ps(velecsum,velec);
 871
 872             fscal            = felec;
 873
 874             fscal            = _mm_and_ps(fscal,cutoff_mask);
 875
 876             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 877
 878             /* Calculate temporary vectorial force */
 879             tx               = _mm_mul_ps(fscal,dx01);
 880             ty               = _mm_mul_ps(fscal,dy01);
 881             tz               = _mm_mul_ps(fscal,dz01);
 882
 883             /* Update vectorial force */
 884             fix0             = _mm_add_ps(fix0,tx);
 885             fiy0             = _mm_add_ps(fiy0,ty);
 886             fiz0             = _mm_add_ps(fiz0,tz);
 887
 888             fjx1             = _mm_add_ps(fjx1,tx);
 889             fjy1             = _mm_add_ps(fjy1,ty);
 890             fjz1             = _mm_add_ps(fjz1,tz);
 891
 892             }
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             if (gmx_mm_any_lt(rsq02,rcutoff2))
 899             {
 900
 901             /* REACTION-FIELD ELECTROSTATICS */
 902             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 903             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 904
 905             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 906
 907             /* Update potential sum for this i atom from the interaction with this j atom. */
 908             velec            = _mm_and_ps(velec,cutoff_mask);
 909             velec            = _mm_andnot_ps(dummy_mask,velec);
 910             velecsum         = _mm_add_ps(velecsum,velec);
 911
 912             fscal            = felec;
 913
 914             fscal            = _mm_and_ps(fscal,cutoff_mask);
 915
 916             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 917
 918             /* Calculate temporary vectorial force */
 919             tx               = _mm_mul_ps(fscal,dx02);
 920             ty               = _mm_mul_ps(fscal,dy02);
 921             tz               = _mm_mul_ps(fscal,dz02);
 922
 923             /* Update vectorial force */
 924             fix0             = _mm_add_ps(fix0,tx);
 925             fiy0             = _mm_add_ps(fiy0,ty);
 926             fiz0             = _mm_add_ps(fiz0,tz);
 927
 928             fjx2             = _mm_add_ps(fjx2,tx);
 929             fjy2             = _mm_add_ps(fjy2,ty);
 930             fjz2             = _mm_add_ps(fjz2,tz);
 931
 932             }
 933
 934             /**************************
 935              * CALCULATE INTERACTIONS *
 936              **************************/
 937
 938             if (gmx_mm_any_lt(rsq10,rcutoff2))
 939             {
 940
 941             /* REACTION-FIELD ELECTROSTATICS */
 942             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 943             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 944
 945             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 946
 947             /* Update potential sum for this i atom from the interaction with this j atom. */
 948             velec            = _mm_and_ps(velec,cutoff_mask);
 949             velec            = _mm_andnot_ps(dummy_mask,velec);
 950             velecsum         = _mm_add_ps(velecsum,velec);
 951
 952             fscal            = felec;
 953
 954             fscal            = _mm_and_ps(fscal,cutoff_mask);
 955
 956             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 957
 958             /* Calculate temporary vectorial force */
 959             tx               = _mm_mul_ps(fscal,dx10);
 960             ty               = _mm_mul_ps(fscal,dy10);
 961             tz               = _mm_mul_ps(fscal,dz10);
 962
 963             /* Update vectorial force */
 964             fix1             = _mm_add_ps(fix1,tx);
 965             fiy1             = _mm_add_ps(fiy1,ty);
 966             fiz1             = _mm_add_ps(fiz1,tz);
 967
 968             fjx0             = _mm_add_ps(fjx0,tx);
 969             fjy0             = _mm_add_ps(fjy0,ty);
 970             fjz0             = _mm_add_ps(fjz0,tz);
 971
 972             }
 973
 974             /**************************
 975              * CALCULATE INTERACTIONS *
 976              **************************/
 977
 978             if (gmx_mm_any_lt(rsq11,rcutoff2))
 979             {
 980
 981             /* REACTION-FIELD ELECTROSTATICS */
 982             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 983             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 984
 985             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 986
 987             /* Update potential sum for this i atom from the interaction with this j atom. */
 988             velec            = _mm_and_ps(velec,cutoff_mask);
 989             velec            = _mm_andnot_ps(dummy_mask,velec);
 990             velecsum         = _mm_add_ps(velecsum,velec);
 991
 992             fscal            = felec;
 993
 994             fscal            = _mm_and_ps(fscal,cutoff_mask);
 995
 996             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 997
 998             /* Calculate temporary vectorial force */
 999             tx               = _mm_mul_ps(fscal,dx11);
1000             ty               = _mm_mul_ps(fscal,dy11);
1001             tz               = _mm_mul_ps(fscal,dz11);
1002
1003             /* Update vectorial force */
1004             fix1             = _mm_add_ps(fix1,tx);
1005             fiy1             = _mm_add_ps(fiy1,ty);
1006             fiz1             = _mm_add_ps(fiz1,tz);
1007
1008             fjx1             = _mm_add_ps(fjx1,tx);
1009             fjy1             = _mm_add_ps(fjy1,ty);
1010             fjz1             = _mm_add_ps(fjz1,tz);
1011
1012             }
1013
1014             /**************************
1015              * CALCULATE INTERACTIONS *
1016              **************************/
1017
1018             if (gmx_mm_any_lt(rsq12,rcutoff2))
1019             {
1020
1021             /* REACTION-FIELD ELECTROSTATICS */
1022             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
1023             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1024
1025             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1026
1027             /* Update potential sum for this i atom from the interaction with this j atom. */
1028             velec            = _mm_and_ps(velec,cutoff_mask);
1029             velec            = _mm_andnot_ps(dummy_mask,velec);
1030             velecsum         = _mm_add_ps(velecsum,velec);
1031
1032             fscal            = felec;
1033
1034             fscal            = _mm_and_ps(fscal,cutoff_mask);
1035
1036             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1037
1038             /* Calculate temporary vectorial force */
1039             tx               = _mm_mul_ps(fscal,dx12);
1040             ty               = _mm_mul_ps(fscal,dy12);
1041             tz               = _mm_mul_ps(fscal,dz12);
1042
1043             /* Update vectorial force */
1044             fix1             = _mm_add_ps(fix1,tx);
1045             fiy1             = _mm_add_ps(fiy1,ty);
1046             fiz1             = _mm_add_ps(fiz1,tz);
1047
1048             fjx2             = _mm_add_ps(fjx2,tx);
1049             fjy2             = _mm_add_ps(fjy2,ty);
1050             fjz2             = _mm_add_ps(fjz2,tz);
1051
1052             }
1053
1054             /**************************
1055              * CALCULATE INTERACTIONS *
1056              **************************/
1057
1058             if (gmx_mm_any_lt(rsq20,rcutoff2))
1059             {
1060
1061             /* REACTION-FIELD ELECTROSTATICS */
1062             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
1063             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1064
1065             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1066
1067             /* Update potential sum for this i atom from the interaction with this j atom. */
1068             velec            = _mm_and_ps(velec,cutoff_mask);
1069             velec            = _mm_andnot_ps(dummy_mask,velec);
1070             velecsum         = _mm_add_ps(velecsum,velec);
1071
1072             fscal            = felec;
1073
1074             fscal            = _mm_and_ps(fscal,cutoff_mask);
1075
1076             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1077
1078             /* Calculate temporary vectorial force */
1079             tx               = _mm_mul_ps(fscal,dx20);
1080             ty               = _mm_mul_ps(fscal,dy20);
1081             tz               = _mm_mul_ps(fscal,dz20);
1082
1083             /* Update vectorial force */
1084             fix2             = _mm_add_ps(fix2,tx);
1085             fiy2             = _mm_add_ps(fiy2,ty);
1086             fiz2             = _mm_add_ps(fiz2,tz);
1087
1088             fjx0             = _mm_add_ps(fjx0,tx);
1089             fjy0             = _mm_add_ps(fjy0,ty);
1090             fjz0             = _mm_add_ps(fjz0,tz);
1091
1092             }
1093
1094             /**************************
1095              * CALCULATE INTERACTIONS *
1096              **************************/
1097
1098             if (gmx_mm_any_lt(rsq21,rcutoff2))
1099             {
1100
1101             /* REACTION-FIELD ELECTROSTATICS */
1102             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
1103             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1104
1105             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1106
1107             /* Update potential sum for this i atom from the interaction with this j atom. */
1108             velec            = _mm_and_ps(velec,cutoff_mask);
1109             velec            = _mm_andnot_ps(dummy_mask,velec);
1110             velecsum         = _mm_add_ps(velecsum,velec);
1111
1112             fscal            = felec;
1113
1114             fscal            = _mm_and_ps(fscal,cutoff_mask);
1115
1116             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1117
1118             /* Calculate temporary vectorial force */
1119             tx               = _mm_mul_ps(fscal,dx21);
1120             ty               = _mm_mul_ps(fscal,dy21);
1121             tz               = _mm_mul_ps(fscal,dz21);
1122
1123             /* Update vectorial force */
1124             fix2             = _mm_add_ps(fix2,tx);
1125             fiy2             = _mm_add_ps(fiy2,ty);
1126             fiz2             = _mm_add_ps(fiz2,tz);
1127
1128             fjx1             = _mm_add_ps(fjx1,tx);
1129             fjy1             = _mm_add_ps(fjy1,ty);
1130             fjz1             = _mm_add_ps(fjz1,tz);
1131
1132             }
1133
1134             /**************************
1135              * CALCULATE INTERACTIONS *
1136              **************************/
1137
1138             if (gmx_mm_any_lt(rsq22,rcutoff2))
1139             {
1140
1141             /* REACTION-FIELD ELECTROSTATICS */
1142             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1143             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1144
1145             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1146
1147             /* Update potential sum for this i atom from the interaction with this j atom. */
1148             velec            = _mm_and_ps(velec,cutoff_mask);
1149             velec            = _mm_andnot_ps(dummy_mask,velec);
1150             velecsum         = _mm_add_ps(velecsum,velec);
1151
1152             fscal            = felec;
1153
1154             fscal            = _mm_and_ps(fscal,cutoff_mask);
1155
1156             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1157
1158             /* Calculate temporary vectorial force */
1159             tx               = _mm_mul_ps(fscal,dx22);
1160             ty               = _mm_mul_ps(fscal,dy22);
1161             tz               = _mm_mul_ps(fscal,dz22);
1162
1163             /* Update vectorial force */
1164             fix2             = _mm_add_ps(fix2,tx);
1165             fiy2             = _mm_add_ps(fiy2,ty);
1166             fiz2             = _mm_add_ps(fiz2,tz);
1167
1168             fjx2             = _mm_add_ps(fjx2,tx);
1169             fjy2             = _mm_add_ps(fjy2,ty);
1170             fjz2             = _mm_add_ps(fjz2,tz);
1171
1172             }
1173
1174             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1175             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1176             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1177             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1178
1179             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1180                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1181
1182             /* Inner loop uses 361 flops */
1183         }
1184
1185         /* End of innermost loop */
1186
1187         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1188                                               f+i_coord_offset,fshift+i_shift_offset);
1189
1190         ggid                        = gid[iidx];
1191         /* Update potential energies */
1192         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1193         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1194
1195         /* Increment number of inner iterations */
1196         inneriter                  += j_index_end - j_index_start;
1197
1198         /* Outer loop uses 20 flops */
1199     }
1200
1201     /* Increment number of outer iterations */
1202     outeriter        += nri;
1203
1204     /* Update outer/inner flops */
1205
1206     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*361);
1207 }
1208 /*
1209  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse4_1_single
1210  * Electrostatics interaction: ReactionField
1211  * VdW interaction:            CubicSplineTable
1212  * Geometry:                   Water3-Water3
1213  * Calculate force/pot:        Force
1214  */
1215 void
1216 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse4_1_single
1217                     (t_nblist * gmx_restrict                nlist,
1218                      rvec * gmx_restrict                    xx,
1219                      rvec * gmx_restrict                    ff,
1220                      t_forcerec * gmx_restrict              fr,
1221                      t_mdatoms * gmx_restrict               mdatoms,
1222                      nb_kernel_data_t * gmx_restrict        kernel_data,
1223                      t_nrnb * gmx_restrict                  nrnb)
1224 {
1225     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1226      * just 0 for non-waters.
1227      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1228      * jnr indices corresponding to data put in the four positions in the SIMD register.
1229      */
1230     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1231     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1232     int              jnrA,jnrB,jnrC,jnrD;
1233     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1234     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1235     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1236     real             rcutoff_scalar;
1237     real             *shiftvec,*fshift,*x,*f;
1238     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1239     real             scratch[4*DIM];
1240     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1241     int              vdwioffset0;
1242     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1243     int              vdwioffset1;
1244     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1245     int              vdwioffset2;
1246     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
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     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1254     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1255     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1256     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1257     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1258     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1259     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1260     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1261     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1262     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1263     real             *charge;
1264     int              nvdwtype;
1265     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1266     int              *vdwtype;
1267     real             *vdwparam;
1268     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1269     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1270     __m128i          vfitab;
1271     __m128i          ifour       = _mm_set1_epi32(4);
1272     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1273     real             *vftab;
1274     __m128           dummy_mask,cutoff_mask;
1275     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1276     __m128           one     = _mm_set1_ps(1.0);
1277     __m128           two     = _mm_set1_ps(2.0);
1278     x                = xx[0];
1279     f                = ff[0];
1280
1281     nri              = nlist->nri;
1282     iinr             = nlist->iinr;
1283     jindex           = nlist->jindex;
1284     jjnr             = nlist->jjnr;
1285     shiftidx         = nlist->shift;
1286     gid              = nlist->gid;
1287     shiftvec         = fr->shift_vec[0];
1288     fshift           = fr->fshift[0];
1289     facel            = _mm_set1_ps(fr->epsfac);
1290     charge           = mdatoms->chargeA;
1291     krf              = _mm_set1_ps(fr->ic->k_rf);
1292     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1293     crf              = _mm_set1_ps(fr->ic->c_rf);
1294     nvdwtype         = fr->ntype;
1295     vdwparam         = fr->nbfp;
1296     vdwtype          = mdatoms->typeA;
1297
1298     vftab            = kernel_data->table_vdw->data;
1299     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
1300
1301     /* Setup water-specific parameters */
1302     inr              = nlist->iinr[0];
1303     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1304     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1305     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1306     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1307
1308     jq0              = _mm_set1_ps(charge[inr+0]);
1309     jq1              = _mm_set1_ps(charge[inr+1]);
1310     jq2              = _mm_set1_ps(charge[inr+2]);
1311     vdwjidx0A        = 2*vdwtype[inr+0];
1312     qq00             = _mm_mul_ps(iq0,jq0);
1313     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1314     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1315     qq01             = _mm_mul_ps(iq0,jq1);
1316     qq02             = _mm_mul_ps(iq0,jq2);
1317     qq10             = _mm_mul_ps(iq1,jq0);
1318     qq11             = _mm_mul_ps(iq1,jq1);
1319     qq12             = _mm_mul_ps(iq1,jq2);
1320     qq20             = _mm_mul_ps(iq2,jq0);
1321     qq21             = _mm_mul_ps(iq2,jq1);
1322     qq22             = _mm_mul_ps(iq2,jq2);
1323
1324     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1325     rcutoff_scalar   = fr->rcoulomb;
1326     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1327     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1328
1329     /* Avoid stupid compiler warnings */
1330     jnrA = jnrB = jnrC = jnrD = 0;
1331     j_coord_offsetA = 0;
1332     j_coord_offsetB = 0;
1333     j_coord_offsetC = 0;
1334     j_coord_offsetD = 0;
1335
1336     outeriter        = 0;
1337     inneriter        = 0;
1338
1339     for(iidx=0;iidx<4*DIM;iidx++)
1340     {
1341         scratch[iidx] = 0.0;
1342     }
1343
1344     /* Start outer loop over neighborlists */
1345     for(iidx=0; iidx<nri; iidx++)
1346     {
1347         /* Load shift vector for this list */
1348         i_shift_offset   = DIM*shiftidx[iidx];
1349
1350         /* Load limits for loop over neighbors */
1351         j_index_start    = jindex[iidx];
1352         j_index_end      = jindex[iidx+1];
1353
1354         /* Get outer coordinate index */
1355         inr              = iinr[iidx];
1356         i_coord_offset   = DIM*inr;
1357
1358         /* Load i particle coords and add shift vector */
1359         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1360                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1361
1362         fix0             = _mm_setzero_ps();
1363         fiy0             = _mm_setzero_ps();
1364         fiz0             = _mm_setzero_ps();
1365         fix1             = _mm_setzero_ps();
1366         fiy1             = _mm_setzero_ps();
1367         fiz1             = _mm_setzero_ps();
1368         fix2             = _mm_setzero_ps();
1369         fiy2             = _mm_setzero_ps();
1370         fiz2             = _mm_setzero_ps();
1371
1372         /* Start inner kernel loop */
1373         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1374         {
1375
1376             /* Get j neighbor index, and coordinate index */
1377             jnrA             = jjnr[jidx];
1378             jnrB             = jjnr[jidx+1];
1379             jnrC             = jjnr[jidx+2];
1380             jnrD             = jjnr[jidx+3];
1381             j_coord_offsetA  = DIM*jnrA;
1382             j_coord_offsetB  = DIM*jnrB;
1383             j_coord_offsetC  = DIM*jnrC;
1384             j_coord_offsetD  = DIM*jnrD;
1385
1386             /* load j atom coordinates */
1387             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1388                                               x+j_coord_offsetC,x+j_coord_offsetD,
1389                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1390
1391             /* Calculate displacement vector */
1392             dx00             = _mm_sub_ps(ix0,jx0);
1393             dy00             = _mm_sub_ps(iy0,jy0);
1394             dz00             = _mm_sub_ps(iz0,jz0);
1395             dx01             = _mm_sub_ps(ix0,jx1);
1396             dy01             = _mm_sub_ps(iy0,jy1);
1397             dz01             = _mm_sub_ps(iz0,jz1);
1398             dx02             = _mm_sub_ps(ix0,jx2);
1399             dy02             = _mm_sub_ps(iy0,jy2);
1400             dz02             = _mm_sub_ps(iz0,jz2);
1401             dx10             = _mm_sub_ps(ix1,jx0);
1402             dy10             = _mm_sub_ps(iy1,jy0);
1403             dz10             = _mm_sub_ps(iz1,jz0);
1404             dx11             = _mm_sub_ps(ix1,jx1);
1405             dy11             = _mm_sub_ps(iy1,jy1);
1406             dz11             = _mm_sub_ps(iz1,jz1);
1407             dx12             = _mm_sub_ps(ix1,jx2);
1408             dy12             = _mm_sub_ps(iy1,jy2);
1409             dz12             = _mm_sub_ps(iz1,jz2);
1410             dx20             = _mm_sub_ps(ix2,jx0);
1411             dy20             = _mm_sub_ps(iy2,jy0);
1412             dz20             = _mm_sub_ps(iz2,jz0);
1413             dx21             = _mm_sub_ps(ix2,jx1);
1414             dy21             = _mm_sub_ps(iy2,jy1);
1415             dz21             = _mm_sub_ps(iz2,jz1);
1416             dx22             = _mm_sub_ps(ix2,jx2);
1417             dy22             = _mm_sub_ps(iy2,jy2);
1418             dz22             = _mm_sub_ps(iz2,jz2);
1419
1420             /* Calculate squared distance and things based on it */
1421             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1422             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1423             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1424             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1425             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1426             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1427             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1428             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1429             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1430
1431             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1432             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1433             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1434             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1435             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1436             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1437             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1438             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1439             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1440
1441             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1442             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1443             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1444             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1445             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1446             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1447             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1448             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1449             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1450
1451             fjx0             = _mm_setzero_ps();
1452             fjy0             = _mm_setzero_ps();
1453             fjz0             = _mm_setzero_ps();
1454             fjx1             = _mm_setzero_ps();
1455             fjy1             = _mm_setzero_ps();
1456             fjz1             = _mm_setzero_ps();
1457             fjx2             = _mm_setzero_ps();
1458             fjy2             = _mm_setzero_ps();
1459             fjz2             = _mm_setzero_ps();
1460
1461             /**************************
1462              * CALCULATE INTERACTIONS *
1463              **************************/
1464
1465             if (gmx_mm_any_lt(rsq00,rcutoff2))
1466             {
1467
1468             r00              = _mm_mul_ps(rsq00,rinv00);
1469
1470             /* Calculate table index by multiplying r with table scale and truncate to integer */
1471             rt               = _mm_mul_ps(r00,vftabscale);
1472             vfitab           = _mm_cvttps_epi32(rt);
1473             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1474             vfitab           = _mm_slli_epi32(vfitab,3);
1475
1476             /* REACTION-FIELD ELECTROSTATICS */
1477             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1478
1479             /* CUBIC SPLINE TABLE DISPERSION */
1480             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1481             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1482             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1483             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1484             _MM_TRANSPOSE4_PS(Y,F,G,H);
1485             Heps             = _mm_mul_ps(vfeps,H);
1486             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1487             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1488             fvdw6            = _mm_mul_ps(c6_00,FF);
1489
1490             /* CUBIC SPLINE TABLE REPULSION */
1491             vfitab           = _mm_add_epi32(vfitab,ifour);
1492             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1493             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1494             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1495             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1496             _MM_TRANSPOSE4_PS(Y,F,G,H);
1497             Heps             = _mm_mul_ps(vfeps,H);
1498             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1499             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1500             fvdw12           = _mm_mul_ps(c12_00,FF);
1501             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1502
1503             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1504
1505             fscal            = _mm_add_ps(felec,fvdw);
1506
1507             fscal            = _mm_and_ps(fscal,cutoff_mask);
1508
1509             /* Calculate temporary vectorial force */
1510             tx               = _mm_mul_ps(fscal,dx00);
1511             ty               = _mm_mul_ps(fscal,dy00);
1512             tz               = _mm_mul_ps(fscal,dz00);
1513
1514             /* Update vectorial force */
1515             fix0             = _mm_add_ps(fix0,tx);
1516             fiy0             = _mm_add_ps(fiy0,ty);
1517             fiz0             = _mm_add_ps(fiz0,tz);
1518
1519             fjx0             = _mm_add_ps(fjx0,tx);
1520             fjy0             = _mm_add_ps(fjy0,ty);
1521             fjz0             = _mm_add_ps(fjz0,tz);
1522
1523             }
1524
1525             /**************************
1526              * CALCULATE INTERACTIONS *
1527              **************************/
1528
1529             if (gmx_mm_any_lt(rsq01,rcutoff2))
1530             {
1531
1532             /* REACTION-FIELD ELECTROSTATICS */
1533             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1534
1535             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1536
1537             fscal            = felec;
1538
1539             fscal            = _mm_and_ps(fscal,cutoff_mask);
1540
1541             /* Calculate temporary vectorial force */
1542             tx               = _mm_mul_ps(fscal,dx01);
1543             ty               = _mm_mul_ps(fscal,dy01);
1544             tz               = _mm_mul_ps(fscal,dz01);
1545
1546             /* Update vectorial force */
1547             fix0             = _mm_add_ps(fix0,tx);
1548             fiy0             = _mm_add_ps(fiy0,ty);
1549             fiz0             = _mm_add_ps(fiz0,tz);
1550
1551             fjx1             = _mm_add_ps(fjx1,tx);
1552             fjy1             = _mm_add_ps(fjy1,ty);
1553             fjz1             = _mm_add_ps(fjz1,tz);
1554
1555             }
1556
1557             /**************************
1558              * CALCULATE INTERACTIONS *
1559              **************************/
1560
1561             if (gmx_mm_any_lt(rsq02,rcutoff2))
1562             {
1563
1564             /* REACTION-FIELD ELECTROSTATICS */
1565             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1566
1567             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1568
1569             fscal            = felec;
1570
1571             fscal            = _mm_and_ps(fscal,cutoff_mask);
1572
1573             /* Calculate temporary vectorial force */
1574             tx               = _mm_mul_ps(fscal,dx02);
1575             ty               = _mm_mul_ps(fscal,dy02);
1576             tz               = _mm_mul_ps(fscal,dz02);
1577
1578             /* Update vectorial force */
1579             fix0             = _mm_add_ps(fix0,tx);
1580             fiy0             = _mm_add_ps(fiy0,ty);
1581             fiz0             = _mm_add_ps(fiz0,tz);
1582
1583             fjx2             = _mm_add_ps(fjx2,tx);
1584             fjy2             = _mm_add_ps(fjy2,ty);
1585             fjz2             = _mm_add_ps(fjz2,tz);
1586
1587             }
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             if (gmx_mm_any_lt(rsq10,rcutoff2))
1594             {
1595
1596             /* REACTION-FIELD ELECTROSTATICS */
1597             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1598
1599             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1600
1601             fscal            = felec;
1602
1603             fscal            = _mm_and_ps(fscal,cutoff_mask);
1604
1605             /* Calculate temporary vectorial force */
1606             tx               = _mm_mul_ps(fscal,dx10);
1607             ty               = _mm_mul_ps(fscal,dy10);
1608             tz               = _mm_mul_ps(fscal,dz10);
1609
1610             /* Update vectorial force */
1611             fix1             = _mm_add_ps(fix1,tx);
1612             fiy1             = _mm_add_ps(fiy1,ty);
1613             fiz1             = _mm_add_ps(fiz1,tz);
1614
1615             fjx0             = _mm_add_ps(fjx0,tx);
1616             fjy0             = _mm_add_ps(fjy0,ty);
1617             fjz0             = _mm_add_ps(fjz0,tz);
1618
1619             }
1620
1621             /**************************
1622              * CALCULATE INTERACTIONS *
1623              **************************/
1624
1625             if (gmx_mm_any_lt(rsq11,rcutoff2))
1626             {
1627
1628             /* REACTION-FIELD ELECTROSTATICS */
1629             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1630
1631             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1632
1633             fscal            = felec;
1634
1635             fscal            = _mm_and_ps(fscal,cutoff_mask);
1636
1637             /* Calculate temporary vectorial force */
1638             tx               = _mm_mul_ps(fscal,dx11);
1639             ty               = _mm_mul_ps(fscal,dy11);
1640             tz               = _mm_mul_ps(fscal,dz11);
1641
1642             /* Update vectorial force */
1643             fix1             = _mm_add_ps(fix1,tx);
1644             fiy1             = _mm_add_ps(fiy1,ty);
1645             fiz1             = _mm_add_ps(fiz1,tz);
1646
1647             fjx1             = _mm_add_ps(fjx1,tx);
1648             fjy1             = _mm_add_ps(fjy1,ty);
1649             fjz1             = _mm_add_ps(fjz1,tz);
1650
1651             }
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             if (gmx_mm_any_lt(rsq12,rcutoff2))
1658             {
1659
1660             /* REACTION-FIELD ELECTROSTATICS */
1661             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1662
1663             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1664
1665             fscal            = felec;
1666
1667             fscal            = _mm_and_ps(fscal,cutoff_mask);
1668
1669             /* Calculate temporary vectorial force */
1670             tx               = _mm_mul_ps(fscal,dx12);
1671             ty               = _mm_mul_ps(fscal,dy12);
1672             tz               = _mm_mul_ps(fscal,dz12);
1673
1674             /* Update vectorial force */
1675             fix1             = _mm_add_ps(fix1,tx);
1676             fiy1             = _mm_add_ps(fiy1,ty);
1677             fiz1             = _mm_add_ps(fiz1,tz);
1678
1679             fjx2             = _mm_add_ps(fjx2,tx);
1680             fjy2             = _mm_add_ps(fjy2,ty);
1681             fjz2             = _mm_add_ps(fjz2,tz);
1682
1683             }
1684
1685             /**************************
1686              * CALCULATE INTERACTIONS *
1687              **************************/
1688
1689             if (gmx_mm_any_lt(rsq20,rcutoff2))
1690             {
1691
1692             /* REACTION-FIELD ELECTROSTATICS */
1693             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1694
1695             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1696
1697             fscal            = felec;
1698
1699             fscal            = _mm_and_ps(fscal,cutoff_mask);
1700
1701             /* Calculate temporary vectorial force */
1702             tx               = _mm_mul_ps(fscal,dx20);
1703             ty               = _mm_mul_ps(fscal,dy20);
1704             tz               = _mm_mul_ps(fscal,dz20);
1705
1706             /* Update vectorial force */
1707             fix2             = _mm_add_ps(fix2,tx);
1708             fiy2             = _mm_add_ps(fiy2,ty);
1709             fiz2             = _mm_add_ps(fiz2,tz);
1710
1711             fjx0             = _mm_add_ps(fjx0,tx);
1712             fjy0             = _mm_add_ps(fjy0,ty);
1713             fjz0             = _mm_add_ps(fjz0,tz);
1714
1715             }
1716
1717             /**************************
1718              * CALCULATE INTERACTIONS *
1719              **************************/
1720
1721             if (gmx_mm_any_lt(rsq21,rcutoff2))
1722             {
1723
1724             /* REACTION-FIELD ELECTROSTATICS */
1725             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1726
1727             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1728
1729             fscal            = felec;
1730
1731             fscal            = _mm_and_ps(fscal,cutoff_mask);
1732
1733             /* Calculate temporary vectorial force */
1734             tx               = _mm_mul_ps(fscal,dx21);
1735             ty               = _mm_mul_ps(fscal,dy21);
1736             tz               = _mm_mul_ps(fscal,dz21);
1737
1738             /* Update vectorial force */
1739             fix2             = _mm_add_ps(fix2,tx);
1740             fiy2             = _mm_add_ps(fiy2,ty);
1741             fiz2             = _mm_add_ps(fiz2,tz);
1742
1743             fjx1             = _mm_add_ps(fjx1,tx);
1744             fjy1             = _mm_add_ps(fjy1,ty);
1745             fjz1             = _mm_add_ps(fjz1,tz);
1746
1747             }
1748
1749             /**************************
1750              * CALCULATE INTERACTIONS *
1751              **************************/
1752
1753             if (gmx_mm_any_lt(rsq22,rcutoff2))
1754             {
1755
1756             /* REACTION-FIELD ELECTROSTATICS */
1757             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1758
1759             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1760
1761             fscal            = felec;
1762
1763             fscal            = _mm_and_ps(fscal,cutoff_mask);
1764
1765             /* Calculate temporary vectorial force */
1766             tx               = _mm_mul_ps(fscal,dx22);
1767             ty               = _mm_mul_ps(fscal,dy22);
1768             tz               = _mm_mul_ps(fscal,dz22);
1769
1770             /* Update vectorial force */
1771             fix2             = _mm_add_ps(fix2,tx);
1772             fiy2             = _mm_add_ps(fiy2,ty);
1773             fiz2             = _mm_add_ps(fiz2,tz);
1774
1775             fjx2             = _mm_add_ps(fjx2,tx);
1776             fjy2             = _mm_add_ps(fjy2,ty);
1777             fjz2             = _mm_add_ps(fjz2,tz);
1778
1779             }
1780
1781             fjptrA             = f+j_coord_offsetA;
1782             fjptrB             = f+j_coord_offsetB;
1783             fjptrC             = f+j_coord_offsetC;
1784             fjptrD             = f+j_coord_offsetD;
1785
1786             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1787                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1788
1789             /* Inner loop uses 297 flops */
1790         }
1791
1792         if(jidx<j_index_end)
1793         {
1794
1795             /* Get j neighbor index, and coordinate index */
1796             jnrlistA         = jjnr[jidx];
1797             jnrlistB         = jjnr[jidx+1];
1798             jnrlistC         = jjnr[jidx+2];
1799             jnrlistD         = jjnr[jidx+3];
1800             /* Sign of each element will be negative for non-real atoms.
1801              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1802              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1803              */
1804             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1805             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1806             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1807             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1808             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1809             j_coord_offsetA  = DIM*jnrA;
1810             j_coord_offsetB  = DIM*jnrB;
1811             j_coord_offsetC  = DIM*jnrC;
1812             j_coord_offsetD  = DIM*jnrD;
1813
1814             /* load j atom coordinates */
1815             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1816                                               x+j_coord_offsetC,x+j_coord_offsetD,
1817                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1818
1819             /* Calculate displacement vector */
1820             dx00             = _mm_sub_ps(ix0,jx0);
1821             dy00             = _mm_sub_ps(iy0,jy0);
1822             dz00             = _mm_sub_ps(iz0,jz0);
1823             dx01             = _mm_sub_ps(ix0,jx1);
1824             dy01             = _mm_sub_ps(iy0,jy1);
1825             dz01             = _mm_sub_ps(iz0,jz1);
1826             dx02             = _mm_sub_ps(ix0,jx2);
1827             dy02             = _mm_sub_ps(iy0,jy2);
1828             dz02             = _mm_sub_ps(iz0,jz2);
1829             dx10             = _mm_sub_ps(ix1,jx0);
1830             dy10             = _mm_sub_ps(iy1,jy0);
1831             dz10             = _mm_sub_ps(iz1,jz0);
1832             dx11             = _mm_sub_ps(ix1,jx1);
1833             dy11             = _mm_sub_ps(iy1,jy1);
1834             dz11             = _mm_sub_ps(iz1,jz1);
1835             dx12             = _mm_sub_ps(ix1,jx2);
1836             dy12             = _mm_sub_ps(iy1,jy2);
1837             dz12             = _mm_sub_ps(iz1,jz2);
1838             dx20             = _mm_sub_ps(ix2,jx0);
1839             dy20             = _mm_sub_ps(iy2,jy0);
1840             dz20             = _mm_sub_ps(iz2,jz0);
1841             dx21             = _mm_sub_ps(ix2,jx1);
1842             dy21             = _mm_sub_ps(iy2,jy1);
1843             dz21             = _mm_sub_ps(iz2,jz1);
1844             dx22             = _mm_sub_ps(ix2,jx2);
1845             dy22             = _mm_sub_ps(iy2,jy2);
1846             dz22             = _mm_sub_ps(iz2,jz2);
1847
1848             /* Calculate squared distance and things based on it */
1849             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1850             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1851             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1852             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1853             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1854             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1855             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1856             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1857             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1858
1859             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1860             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1861             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1862             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1863             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1864             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1865             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1866             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1867             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1868
1869             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1870             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1871             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1872             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1873             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1874             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1875             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1876             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1877             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1878
1879             fjx0             = _mm_setzero_ps();
1880             fjy0             = _mm_setzero_ps();
1881             fjz0             = _mm_setzero_ps();
1882             fjx1             = _mm_setzero_ps();
1883             fjy1             = _mm_setzero_ps();
1884             fjz1             = _mm_setzero_ps();
1885             fjx2             = _mm_setzero_ps();
1886             fjy2             = _mm_setzero_ps();
1887             fjz2             = _mm_setzero_ps();
1888
1889             /**************************
1890              * CALCULATE INTERACTIONS *
1891              **************************/
1892
1893             if (gmx_mm_any_lt(rsq00,rcutoff2))
1894             {
1895
1896             r00              = _mm_mul_ps(rsq00,rinv00);
1897             r00              = _mm_andnot_ps(dummy_mask,r00);
1898
1899             /* Calculate table index by multiplying r with table scale and truncate to integer */
1900             rt               = _mm_mul_ps(r00,vftabscale);
1901             vfitab           = _mm_cvttps_epi32(rt);
1902             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1903             vfitab           = _mm_slli_epi32(vfitab,3);
1904
1905             /* REACTION-FIELD ELECTROSTATICS */
1906             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1907
1908             /* CUBIC SPLINE TABLE DISPERSION */
1909             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1910             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1911             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1912             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1913             _MM_TRANSPOSE4_PS(Y,F,G,H);
1914             Heps             = _mm_mul_ps(vfeps,H);
1915             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1916             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1917             fvdw6            = _mm_mul_ps(c6_00,FF);
1918
1919             /* CUBIC SPLINE TABLE REPULSION */
1920             vfitab           = _mm_add_epi32(vfitab,ifour);
1921             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1922             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1923             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1924             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1925             _MM_TRANSPOSE4_PS(Y,F,G,H);
1926             Heps             = _mm_mul_ps(vfeps,H);
1927             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1928             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1929             fvdw12           = _mm_mul_ps(c12_00,FF);
1930             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1931
1932             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1933
1934             fscal            = _mm_add_ps(felec,fvdw);
1935
1936             fscal            = _mm_and_ps(fscal,cutoff_mask);
1937
1938             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1939
1940             /* Calculate temporary vectorial force */
1941             tx               = _mm_mul_ps(fscal,dx00);
1942             ty               = _mm_mul_ps(fscal,dy00);
1943             tz               = _mm_mul_ps(fscal,dz00);
1944
1945             /* Update vectorial force */
1946             fix0             = _mm_add_ps(fix0,tx);
1947             fiy0             = _mm_add_ps(fiy0,ty);
1948             fiz0             = _mm_add_ps(fiz0,tz);
1949
1950             fjx0             = _mm_add_ps(fjx0,tx);
1951             fjy0             = _mm_add_ps(fjy0,ty);
1952             fjz0             = _mm_add_ps(fjz0,tz);
1953
1954             }
1955
1956             /**************************
1957              * CALCULATE INTERACTIONS *
1958              **************************/
1959
1960             if (gmx_mm_any_lt(rsq01,rcutoff2))
1961             {
1962
1963             /* REACTION-FIELD ELECTROSTATICS */
1964             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1965
1966             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1967
1968             fscal            = felec;
1969
1970             fscal            = _mm_and_ps(fscal,cutoff_mask);
1971
1972             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1973
1974             /* Calculate temporary vectorial force */
1975             tx               = _mm_mul_ps(fscal,dx01);
1976             ty               = _mm_mul_ps(fscal,dy01);
1977             tz               = _mm_mul_ps(fscal,dz01);
1978
1979             /* Update vectorial force */
1980             fix0             = _mm_add_ps(fix0,tx);
1981             fiy0             = _mm_add_ps(fiy0,ty);
1982             fiz0             = _mm_add_ps(fiz0,tz);
1983
1984             fjx1             = _mm_add_ps(fjx1,tx);
1985             fjy1             = _mm_add_ps(fjy1,ty);
1986             fjz1             = _mm_add_ps(fjz1,tz);
1987
1988             }
1989
1990             /**************************
1991              * CALCULATE INTERACTIONS *
1992              **************************/
1993
1994             if (gmx_mm_any_lt(rsq02,rcutoff2))
1995             {
1996
1997             /* REACTION-FIELD ELECTROSTATICS */
1998             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1999
2000             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
2001
2002             fscal            = felec;
2003
2004             fscal            = _mm_and_ps(fscal,cutoff_mask);
2005
2006             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2007
2008             /* Calculate temporary vectorial force */
2009             tx               = _mm_mul_ps(fscal,dx02);
2010             ty               = _mm_mul_ps(fscal,dy02);
2011             tz               = _mm_mul_ps(fscal,dz02);
2012
2013             /* Update vectorial force */
2014             fix0             = _mm_add_ps(fix0,tx);
2015             fiy0             = _mm_add_ps(fiy0,ty);
2016             fiz0             = _mm_add_ps(fiz0,tz);
2017
2018             fjx2             = _mm_add_ps(fjx2,tx);
2019             fjy2             = _mm_add_ps(fjy2,ty);
2020             fjz2             = _mm_add_ps(fjz2,tz);
2021
2022             }
2023
2024             /**************************
2025              * CALCULATE INTERACTIONS *
2026              **************************/
2027
2028             if (gmx_mm_any_lt(rsq10,rcutoff2))
2029             {
2030
2031             /* REACTION-FIELD ELECTROSTATICS */
2032             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
2033
2034             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
2035
2036             fscal            = felec;
2037
2038             fscal            = _mm_and_ps(fscal,cutoff_mask);
2039
2040             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2041
2042             /* Calculate temporary vectorial force */
2043             tx               = _mm_mul_ps(fscal,dx10);
2044             ty               = _mm_mul_ps(fscal,dy10);
2045             tz               = _mm_mul_ps(fscal,dz10);
2046
2047             /* Update vectorial force */
2048             fix1             = _mm_add_ps(fix1,tx);
2049             fiy1             = _mm_add_ps(fiy1,ty);
2050             fiz1             = _mm_add_ps(fiz1,tz);
2051
2052             fjx0             = _mm_add_ps(fjx0,tx);
2053             fjy0             = _mm_add_ps(fjy0,ty);
2054             fjz0             = _mm_add_ps(fjz0,tz);
2055
2056             }
2057
2058             /**************************
2059              * CALCULATE INTERACTIONS *
2060              **************************/
2061
2062             if (gmx_mm_any_lt(rsq11,rcutoff2))
2063             {
2064
2065             /* REACTION-FIELD ELECTROSTATICS */
2066             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
2067
2068             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2069
2070             fscal            = felec;
2071
2072             fscal            = _mm_and_ps(fscal,cutoff_mask);
2073
2074             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2075
2076             /* Calculate temporary vectorial force */
2077             tx               = _mm_mul_ps(fscal,dx11);
2078             ty               = _mm_mul_ps(fscal,dy11);
2079             tz               = _mm_mul_ps(fscal,dz11);
2080
2081             /* Update vectorial force */
2082             fix1             = _mm_add_ps(fix1,tx);
2083             fiy1             = _mm_add_ps(fiy1,ty);
2084             fiz1             = _mm_add_ps(fiz1,tz);
2085
2086             fjx1             = _mm_add_ps(fjx1,tx);
2087             fjy1             = _mm_add_ps(fjy1,ty);
2088             fjz1             = _mm_add_ps(fjz1,tz);
2089
2090             }
2091
2092             /**************************
2093              * CALCULATE INTERACTIONS *
2094              **************************/
2095
2096             if (gmx_mm_any_lt(rsq12,rcutoff2))
2097             {
2098
2099             /* REACTION-FIELD ELECTROSTATICS */
2100             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
2101
2102             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2103
2104             fscal            = felec;
2105
2106             fscal            = _mm_and_ps(fscal,cutoff_mask);
2107
2108             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2109
2110             /* Calculate temporary vectorial force */
2111             tx               = _mm_mul_ps(fscal,dx12);
2112             ty               = _mm_mul_ps(fscal,dy12);
2113             tz               = _mm_mul_ps(fscal,dz12);
2114
2115             /* Update vectorial force */
2116             fix1             = _mm_add_ps(fix1,tx);
2117             fiy1             = _mm_add_ps(fiy1,ty);
2118             fiz1             = _mm_add_ps(fiz1,tz);
2119
2120             fjx2             = _mm_add_ps(fjx2,tx);
2121             fjy2             = _mm_add_ps(fjy2,ty);
2122             fjz2             = _mm_add_ps(fjz2,tz);
2123
2124             }
2125
2126             /**************************
2127              * CALCULATE INTERACTIONS *
2128              **************************/
2129
2130             if (gmx_mm_any_lt(rsq20,rcutoff2))
2131             {
2132
2133             /* REACTION-FIELD ELECTROSTATICS */
2134             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
2135
2136             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2137
2138             fscal            = felec;
2139
2140             fscal            = _mm_and_ps(fscal,cutoff_mask);
2141
2142             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2143
2144             /* Calculate temporary vectorial force */
2145             tx               = _mm_mul_ps(fscal,dx20);
2146             ty               = _mm_mul_ps(fscal,dy20);
2147             tz               = _mm_mul_ps(fscal,dz20);
2148
2149             /* Update vectorial force */
2150             fix2             = _mm_add_ps(fix2,tx);
2151             fiy2             = _mm_add_ps(fiy2,ty);
2152             fiz2             = _mm_add_ps(fiz2,tz);
2153
2154             fjx0             = _mm_add_ps(fjx0,tx);
2155             fjy0             = _mm_add_ps(fjy0,ty);
2156             fjz0             = _mm_add_ps(fjz0,tz);
2157
2158             }
2159
2160             /**************************
2161              * CALCULATE INTERACTIONS *
2162              **************************/
2163
2164             if (gmx_mm_any_lt(rsq21,rcutoff2))
2165             {
2166
2167             /* REACTION-FIELD ELECTROSTATICS */
2168             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
2169
2170             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2171
2172             fscal            = felec;
2173
2174             fscal            = _mm_and_ps(fscal,cutoff_mask);
2175
2176             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2177
2178             /* Calculate temporary vectorial force */
2179             tx               = _mm_mul_ps(fscal,dx21);
2180             ty               = _mm_mul_ps(fscal,dy21);
2181             tz               = _mm_mul_ps(fscal,dz21);
2182
2183             /* Update vectorial force */
2184             fix2             = _mm_add_ps(fix2,tx);
2185             fiy2             = _mm_add_ps(fiy2,ty);
2186             fiz2             = _mm_add_ps(fiz2,tz);
2187
2188             fjx1             = _mm_add_ps(fjx1,tx);
2189             fjy1             = _mm_add_ps(fjy1,ty);
2190             fjz1             = _mm_add_ps(fjz1,tz);
2191
2192             }
2193
2194             /**************************
2195              * CALCULATE INTERACTIONS *
2196              **************************/
2197
2198             if (gmx_mm_any_lt(rsq22,rcutoff2))
2199             {
2200
2201             /* REACTION-FIELD ELECTROSTATICS */
2202             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2203
2204             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2205
2206             fscal            = felec;
2207
2208             fscal            = _mm_and_ps(fscal,cutoff_mask);
2209
2210             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2211
2212             /* Calculate temporary vectorial force */
2213             tx               = _mm_mul_ps(fscal,dx22);
2214             ty               = _mm_mul_ps(fscal,dy22);
2215             tz               = _mm_mul_ps(fscal,dz22);
2216
2217             /* Update vectorial force */
2218             fix2             = _mm_add_ps(fix2,tx);
2219             fiy2             = _mm_add_ps(fiy2,ty);
2220             fiz2             = _mm_add_ps(fiz2,tz);
2221
2222             fjx2             = _mm_add_ps(fjx2,tx);
2223             fjy2             = _mm_add_ps(fjy2,ty);
2224             fjz2             = _mm_add_ps(fjz2,tz);
2225
2226             }
2227
2228             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2229             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2230             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2231             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2232
2233             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2234                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2235
2236             /* Inner loop uses 298 flops */
2237         }
2238
2239         /* End of innermost loop */
2240
2241         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2242                                               f+i_coord_offset,fshift+i_shift_offset);
2243
2244         /* Increment number of inner iterations */
2245         inneriter                  += j_index_end - j_index_start;
2246
2247         /* Outer loop uses 18 flops */
2248     }
2249
2250     /* Increment number of outer iterations */
2251     outeriter        += nri;
2252
2253     /* Update outer/inner flops */
2254
2255     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*298);
2256 }