src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwNone_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW3W3_VF_avx_128_fma_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            None
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSh_VdwNone_GeomW3W3_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128i          ewitab;
  93     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  94     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  95     real             *ewtab;
  96     __m128           dummy_mask,cutoff_mask;
  97     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  98     __m128           one     = _mm_set1_ps(1.0);
  99     __m128           two     = _mm_set1_ps(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = _mm_set1_ps(fr->epsfac);
 112     charge           = mdatoms->chargeA;
 113
 114     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 115     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 116     beta2            = _mm_mul_ps(beta,beta);
 117     beta3            = _mm_mul_ps(beta,beta2);
 118     ewtab            = fr->ic->tabq_coul_FDV0;
 119     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 120     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 121
 122     /* Setup water-specific parameters */
 123     inr              = nlist->iinr[0];
 124     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 125     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 126     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 127
 128     jq0              = _mm_set1_ps(charge[inr+0]);
 129     jq1              = _mm_set1_ps(charge[inr+1]);
 130     jq2              = _mm_set1_ps(charge[inr+2]);
 131     qq00             = _mm_mul_ps(iq0,jq0);
 132     qq01             = _mm_mul_ps(iq0,jq1);
 133     qq02             = _mm_mul_ps(iq0,jq2);
 134     qq10             = _mm_mul_ps(iq1,jq0);
 135     qq11             = _mm_mul_ps(iq1,jq1);
 136     qq12             = _mm_mul_ps(iq1,jq2);
 137     qq20             = _mm_mul_ps(iq2,jq0);
 138     qq21             = _mm_mul_ps(iq2,jq1);
 139     qq22             = _mm_mul_ps(iq2,jq2);
 140
 141     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 142     rcutoff_scalar   = fr->rcoulomb;
 143     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 144     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 145
 146     /* Avoid stupid compiler warnings */
 147     jnrA = jnrB = jnrC = jnrD = 0;
 148     j_coord_offsetA = 0;
 149     j_coord_offsetB = 0;
 150     j_coord_offsetC = 0;
 151     j_coord_offsetD = 0;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     for(iidx=0;iidx<4*DIM;iidx++)
 157     {
 158         scratch[iidx] = 0.0;
 159     }
 160
 161     /* Start outer loop over neighborlists */
 162     for(iidx=0; iidx<nri; iidx++)
 163     {
 164         /* Load shift vector for this list */
 165         i_shift_offset   = DIM*shiftidx[iidx];
 166
 167         /* Load limits for loop over neighbors */
 168         j_index_start    = jindex[iidx];
 169         j_index_end      = jindex[iidx+1];
 170
 171         /* Get outer coordinate index */
 172         inr              = iinr[iidx];
 173         i_coord_offset   = DIM*inr;
 174
 175         /* Load i particle coords and add shift vector */
 176         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 177                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 178
 179         fix0             = _mm_setzero_ps();
 180         fiy0             = _mm_setzero_ps();
 181         fiz0             = _mm_setzero_ps();
 182         fix1             = _mm_setzero_ps();
 183         fiy1             = _mm_setzero_ps();
 184         fiz1             = _mm_setzero_ps();
 185         fix2             = _mm_setzero_ps();
 186         fiy2             = _mm_setzero_ps();
 187         fiz2             = _mm_setzero_ps();
 188
 189         /* Reset potential sums */
 190         velecsum         = _mm_setzero_ps();
 191
 192         /* Start inner kernel loop */
 193         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 194         {
 195
 196             /* Get j neighbor index, and coordinate index */
 197             jnrA             = jjnr[jidx];
 198             jnrB             = jjnr[jidx+1];
 199             jnrC             = jjnr[jidx+2];
 200             jnrD             = jjnr[jidx+3];
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203             j_coord_offsetC  = DIM*jnrC;
 204             j_coord_offsetD  = DIM*jnrD;
 205
 206             /* load j atom coordinates */
 207             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 208                                               x+j_coord_offsetC,x+j_coord_offsetD,
 209                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 210
 211             /* Calculate displacement vector */
 212             dx00             = _mm_sub_ps(ix0,jx0);
 213             dy00             = _mm_sub_ps(iy0,jy0);
 214             dz00             = _mm_sub_ps(iz0,jz0);
 215             dx01             = _mm_sub_ps(ix0,jx1);
 216             dy01             = _mm_sub_ps(iy0,jy1);
 217             dz01             = _mm_sub_ps(iz0,jz1);
 218             dx02             = _mm_sub_ps(ix0,jx2);
 219             dy02             = _mm_sub_ps(iy0,jy2);
 220             dz02             = _mm_sub_ps(iz0,jz2);
 221             dx10             = _mm_sub_ps(ix1,jx0);
 222             dy10             = _mm_sub_ps(iy1,jy0);
 223             dz10             = _mm_sub_ps(iz1,jz0);
 224             dx11             = _mm_sub_ps(ix1,jx1);
 225             dy11             = _mm_sub_ps(iy1,jy1);
 226             dz11             = _mm_sub_ps(iz1,jz1);
 227             dx12             = _mm_sub_ps(ix1,jx2);
 228             dy12             = _mm_sub_ps(iy1,jy2);
 229             dz12             = _mm_sub_ps(iz1,jz2);
 230             dx20             = _mm_sub_ps(ix2,jx0);
 231             dy20             = _mm_sub_ps(iy2,jy0);
 232             dz20             = _mm_sub_ps(iz2,jz0);
 233             dx21             = _mm_sub_ps(ix2,jx1);
 234             dy21             = _mm_sub_ps(iy2,jy1);
 235             dz21             = _mm_sub_ps(iz2,jz1);
 236             dx22             = _mm_sub_ps(ix2,jx2);
 237             dy22             = _mm_sub_ps(iy2,jy2);
 238             dz22             = _mm_sub_ps(iz2,jz2);
 239
 240             /* Calculate squared distance and things based on it */
 241             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 242             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 243             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 244             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 245             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 246             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 247             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 248             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 249             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 250
 251             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 252             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 253             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 254             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 255             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 256             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 257             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 258             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 259             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 260
 261             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 262             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 263             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 264             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 265             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 266             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 267             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 268             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 269             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 270
 271             fjx0             = _mm_setzero_ps();
 272             fjy0             = _mm_setzero_ps();
 273             fjz0             = _mm_setzero_ps();
 274             fjx1             = _mm_setzero_ps();
 275             fjy1             = _mm_setzero_ps();
 276             fjz1             = _mm_setzero_ps();
 277             fjx2             = _mm_setzero_ps();
 278             fjy2             = _mm_setzero_ps();
 279             fjz2             = _mm_setzero_ps();
 280
 281             /**************************
 282              * CALCULATE INTERACTIONS *
 283              **************************/
 284
 285             if (gmx_mm_any_lt(rsq00,rcutoff2))
 286             {
 287
 288             r00              = _mm_mul_ps(rsq00,rinv00);
 289
 290             /* EWALD ELECTROSTATICS */
 291
 292             /* Analytical PME correction */
 293             zeta2            = _mm_mul_ps(beta2,rsq00);
 294             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 295             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 296             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 297             felec            = _mm_mul_ps(qq00,felec);
 298             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 299             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 300             velec            = _mm_mul_ps(qq00,velec);
 301
 302             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 303
 304             /* Update potential sum for this i atom from the interaction with this j atom. */
 305             velec            = _mm_and_ps(velec,cutoff_mask);
 306             velecsum         = _mm_add_ps(velecsum,velec);
 307
 308             fscal            = felec;
 309
 310             fscal            = _mm_and_ps(fscal,cutoff_mask);
 311
 312              /* Update vectorial force */
 313             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 314             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 315             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 316
 317             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 318             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 319             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 320
 321             }
 322
 323             /**************************
 324              * CALCULATE INTERACTIONS *
 325              **************************/
 326
 327             if (gmx_mm_any_lt(rsq01,rcutoff2))
 328             {
 329
 330             r01              = _mm_mul_ps(rsq01,rinv01);
 331
 332             /* EWALD ELECTROSTATICS */
 333
 334             /* Analytical PME correction */
 335             zeta2            = _mm_mul_ps(beta2,rsq01);
 336             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 337             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 338             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 339             felec            = _mm_mul_ps(qq01,felec);
 340             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 341             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 342             velec            = _mm_mul_ps(qq01,velec);
 343
 344             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velec            = _mm_and_ps(velec,cutoff_mask);
 348             velecsum         = _mm_add_ps(velecsum,velec);
 349
 350             fscal            = felec;
 351
 352             fscal            = _mm_and_ps(fscal,cutoff_mask);
 353
 354              /* Update vectorial force */
 355             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 356             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 357             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 358
 359             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 360             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 361             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 362
 363             }
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             if (gmx_mm_any_lt(rsq02,rcutoff2))
 370             {
 371
 372             r02              = _mm_mul_ps(rsq02,rinv02);
 373
 374             /* EWALD ELECTROSTATICS */
 375
 376             /* Analytical PME correction */
 377             zeta2            = _mm_mul_ps(beta2,rsq02);
 378             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 379             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 380             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 381             felec            = _mm_mul_ps(qq02,felec);
 382             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 383             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 384             velec            = _mm_mul_ps(qq02,velec);
 385
 386             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 387
 388             /* Update potential sum for this i atom from the interaction with this j atom. */
 389             velec            = _mm_and_ps(velec,cutoff_mask);
 390             velecsum         = _mm_add_ps(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             fscal            = _mm_and_ps(fscal,cutoff_mask);
 395
 396              /* Update vectorial force */
 397             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 398             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 399             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 400
 401             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 402             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 403             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 404
 405             }
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             if (gmx_mm_any_lt(rsq10,rcutoff2))
 412             {
 413
 414             r10              = _mm_mul_ps(rsq10,rinv10);
 415
 416             /* EWALD ELECTROSTATICS */
 417
 418             /* Analytical PME correction */
 419             zeta2            = _mm_mul_ps(beta2,rsq10);
 420             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 421             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 422             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 423             felec            = _mm_mul_ps(qq10,felec);
 424             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 425             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 426             velec            = _mm_mul_ps(qq10,velec);
 427
 428             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velec            = _mm_and_ps(velec,cutoff_mask);
 432             velecsum         = _mm_add_ps(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436             fscal            = _mm_and_ps(fscal,cutoff_mask);
 437
 438              /* Update vectorial force */
 439             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 440             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 441             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 442
 443             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 444             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 445             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 446
 447             }
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             if (gmx_mm_any_lt(rsq11,rcutoff2))
 454             {
 455
 456             r11              = _mm_mul_ps(rsq11,rinv11);
 457
 458             /* EWALD ELECTROSTATICS */
 459
 460             /* Analytical PME correction */
 461             zeta2            = _mm_mul_ps(beta2,rsq11);
 462             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 463             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 464             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 465             felec            = _mm_mul_ps(qq11,felec);
 466             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 467             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
 468             velec            = _mm_mul_ps(qq11,velec);
 469
 470             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velec            = _mm_and_ps(velec,cutoff_mask);
 474             velecsum         = _mm_add_ps(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478             fscal            = _mm_and_ps(fscal,cutoff_mask);
 479
 480              /* Update vectorial force */
 481             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 482             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 483             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 484
 485             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 486             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 487             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 488
 489             }
 490
 491             /**************************
 492              * CALCULATE INTERACTIONS *
 493              **************************/
 494
 495             if (gmx_mm_any_lt(rsq12,rcutoff2))
 496             {
 497
 498             r12              = _mm_mul_ps(rsq12,rinv12);
 499
 500             /* EWALD ELECTROSTATICS */
 501
 502             /* Analytical PME correction */
 503             zeta2            = _mm_mul_ps(beta2,rsq12);
 504             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 505             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 506             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 507             felec            = _mm_mul_ps(qq12,felec);
 508             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 509             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
 510             velec            = _mm_mul_ps(qq12,velec);
 511
 512             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 513
 514             /* Update potential sum for this i atom from the interaction with this j atom. */
 515             velec            = _mm_and_ps(velec,cutoff_mask);
 516             velecsum         = _mm_add_ps(velecsum,velec);
 517
 518             fscal            = felec;
 519
 520             fscal            = _mm_and_ps(fscal,cutoff_mask);
 521
 522              /* Update vectorial force */
 523             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 524             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 525             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 526
 527             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 528             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 529             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 530
 531             }
 532
 533             /**************************
 534              * CALCULATE INTERACTIONS *
 535              **************************/
 536
 537             if (gmx_mm_any_lt(rsq20,rcutoff2))
 538             {
 539
 540             r20              = _mm_mul_ps(rsq20,rinv20);
 541
 542             /* EWALD ELECTROSTATICS */
 543
 544             /* Analytical PME correction */
 545             zeta2            = _mm_mul_ps(beta2,rsq20);
 546             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 547             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 548             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 549             felec            = _mm_mul_ps(qq20,felec);
 550             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 551             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 552             velec            = _mm_mul_ps(qq20,velec);
 553
 554             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 555
 556             /* Update potential sum for this i atom from the interaction with this j atom. */
 557             velec            = _mm_and_ps(velec,cutoff_mask);
 558             velecsum         = _mm_add_ps(velecsum,velec);
 559
 560             fscal            = felec;
 561
 562             fscal            = _mm_and_ps(fscal,cutoff_mask);
 563
 564              /* Update vectorial force */
 565             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 566             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 567             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 568
 569             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 570             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 571             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 572
 573             }
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             if (gmx_mm_any_lt(rsq21,rcutoff2))
 580             {
 581
 582             r21              = _mm_mul_ps(rsq21,rinv21);
 583
 584             /* EWALD ELECTROSTATICS */
 585
 586             /* Analytical PME correction */
 587             zeta2            = _mm_mul_ps(beta2,rsq21);
 588             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 589             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 590             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 591             felec            = _mm_mul_ps(qq21,felec);
 592             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 593             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
 594             velec            = _mm_mul_ps(qq21,velec);
 595
 596             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 597
 598             /* Update potential sum for this i atom from the interaction with this j atom. */
 599             velec            = _mm_and_ps(velec,cutoff_mask);
 600             velecsum         = _mm_add_ps(velecsum,velec);
 601
 602             fscal            = felec;
 603
 604             fscal            = _mm_and_ps(fscal,cutoff_mask);
 605
 606              /* Update vectorial force */
 607             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 608             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 609             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 610
 611             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 612             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 613             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 614
 615             }
 616
 617             /**************************
 618              * CALCULATE INTERACTIONS *
 619              **************************/
 620
 621             if (gmx_mm_any_lt(rsq22,rcutoff2))
 622             {
 623
 624             r22              = _mm_mul_ps(rsq22,rinv22);
 625
 626             /* EWALD ELECTROSTATICS */
 627
 628             /* Analytical PME correction */
 629             zeta2            = _mm_mul_ps(beta2,rsq22);
 630             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 631             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 632             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 633             felec            = _mm_mul_ps(qq22,felec);
 634             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 635             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
 636             velec            = _mm_mul_ps(qq22,velec);
 637
 638             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 639
 640             /* Update potential sum for this i atom from the interaction with this j atom. */
 641             velec            = _mm_and_ps(velec,cutoff_mask);
 642             velecsum         = _mm_add_ps(velecsum,velec);
 643
 644             fscal            = felec;
 645
 646             fscal            = _mm_and_ps(fscal,cutoff_mask);
 647
 648              /* Update vectorial force */
 649             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 650             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 651             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 652
 653             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 654             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 655             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 656
 657             }
 658
 659             fjptrA             = f+j_coord_offsetA;
 660             fjptrB             = f+j_coord_offsetB;
 661             fjptrC             = f+j_coord_offsetC;
 662             fjptrD             = f+j_coord_offsetD;
 663
 664             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 665                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 666
 667             /* Inner loop uses 297 flops */
 668         }
 669
 670         if(jidx<j_index_end)
 671         {
 672
 673             /* Get j neighbor index, and coordinate index */
 674             jnrlistA         = jjnr[jidx];
 675             jnrlistB         = jjnr[jidx+1];
 676             jnrlistC         = jjnr[jidx+2];
 677             jnrlistD         = jjnr[jidx+3];
 678             /* Sign of each element will be negative for non-real atoms.
 679              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 680              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 681              */
 682             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 683             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 684             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 685             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 686             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 687             j_coord_offsetA  = DIM*jnrA;
 688             j_coord_offsetB  = DIM*jnrB;
 689             j_coord_offsetC  = DIM*jnrC;
 690             j_coord_offsetD  = DIM*jnrD;
 691
 692             /* load j atom coordinates */
 693             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 694                                               x+j_coord_offsetC,x+j_coord_offsetD,
 695                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 696
 697             /* Calculate displacement vector */
 698             dx00             = _mm_sub_ps(ix0,jx0);
 699             dy00             = _mm_sub_ps(iy0,jy0);
 700             dz00             = _mm_sub_ps(iz0,jz0);
 701             dx01             = _mm_sub_ps(ix0,jx1);
 702             dy01             = _mm_sub_ps(iy0,jy1);
 703             dz01             = _mm_sub_ps(iz0,jz1);
 704             dx02             = _mm_sub_ps(ix0,jx2);
 705             dy02             = _mm_sub_ps(iy0,jy2);
 706             dz02             = _mm_sub_ps(iz0,jz2);
 707             dx10             = _mm_sub_ps(ix1,jx0);
 708             dy10             = _mm_sub_ps(iy1,jy0);
 709             dz10             = _mm_sub_ps(iz1,jz0);
 710             dx11             = _mm_sub_ps(ix1,jx1);
 711             dy11             = _mm_sub_ps(iy1,jy1);
 712             dz11             = _mm_sub_ps(iz1,jz1);
 713             dx12             = _mm_sub_ps(ix1,jx2);
 714             dy12             = _mm_sub_ps(iy1,jy2);
 715             dz12             = _mm_sub_ps(iz1,jz2);
 716             dx20             = _mm_sub_ps(ix2,jx0);
 717             dy20             = _mm_sub_ps(iy2,jy0);
 718             dz20             = _mm_sub_ps(iz2,jz0);
 719             dx21             = _mm_sub_ps(ix2,jx1);
 720             dy21             = _mm_sub_ps(iy2,jy1);
 721             dz21             = _mm_sub_ps(iz2,jz1);
 722             dx22             = _mm_sub_ps(ix2,jx2);
 723             dy22             = _mm_sub_ps(iy2,jy2);
 724             dz22             = _mm_sub_ps(iz2,jz2);
 725
 726             /* Calculate squared distance and things based on it */
 727             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 728             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 729             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 730             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 731             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 732             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 733             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 734             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 735             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 736
 737             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 738             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 739             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 740             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 741             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 742             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 743             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 744             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 745             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 746
 747             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 748             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 749             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 750             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 751             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 752             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 753             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 754             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 755             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 756
 757             fjx0             = _mm_setzero_ps();
 758             fjy0             = _mm_setzero_ps();
 759             fjz0             = _mm_setzero_ps();
 760             fjx1             = _mm_setzero_ps();
 761             fjy1             = _mm_setzero_ps();
 762             fjz1             = _mm_setzero_ps();
 763             fjx2             = _mm_setzero_ps();
 764             fjy2             = _mm_setzero_ps();
 765             fjz2             = _mm_setzero_ps();
 766
 767             /**************************
 768              * CALCULATE INTERACTIONS *
 769              **************************/
 770
 771             if (gmx_mm_any_lt(rsq00,rcutoff2))
 772             {
 773
 774             r00              = _mm_mul_ps(rsq00,rinv00);
 775             r00              = _mm_andnot_ps(dummy_mask,r00);
 776
 777             /* EWALD ELECTROSTATICS */
 778
 779             /* Analytical PME correction */
 780             zeta2            = _mm_mul_ps(beta2,rsq00);
 781             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 782             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 783             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 784             felec            = _mm_mul_ps(qq00,felec);
 785             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 786             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 787             velec            = _mm_mul_ps(qq00,velec);
 788
 789             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 790
 791             /* Update potential sum for this i atom from the interaction with this j atom. */
 792             velec            = _mm_and_ps(velec,cutoff_mask);
 793             velec            = _mm_andnot_ps(dummy_mask,velec);
 794             velecsum         = _mm_add_ps(velecsum,velec);
 795
 796             fscal            = felec;
 797
 798             fscal            = _mm_and_ps(fscal,cutoff_mask);
 799
 800             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 801
 802              /* Update vectorial force */
 803             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 804             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 805             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 806
 807             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 808             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 809             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 810
 811             }
 812
 813             /**************************
 814              * CALCULATE INTERACTIONS *
 815              **************************/
 816
 817             if (gmx_mm_any_lt(rsq01,rcutoff2))
 818             {
 819
 820             r01              = _mm_mul_ps(rsq01,rinv01);
 821             r01              = _mm_andnot_ps(dummy_mask,r01);
 822
 823             /* EWALD ELECTROSTATICS */
 824
 825             /* Analytical PME correction */
 826             zeta2            = _mm_mul_ps(beta2,rsq01);
 827             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 828             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 829             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 830             felec            = _mm_mul_ps(qq01,felec);
 831             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 832             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 833             velec            = _mm_mul_ps(qq01,velec);
 834
 835             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_and_ps(velec,cutoff_mask);
 839             velec            = _mm_andnot_ps(dummy_mask,velec);
 840             velecsum         = _mm_add_ps(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_and_ps(fscal,cutoff_mask);
 845
 846             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 847
 848              /* Update vectorial force */
 849             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 850             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 851             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 852
 853             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 854             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 855             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 856
 857             }
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             if (gmx_mm_any_lt(rsq02,rcutoff2))
 864             {
 865
 866             r02              = _mm_mul_ps(rsq02,rinv02);
 867             r02              = _mm_andnot_ps(dummy_mask,r02);
 868
 869             /* EWALD ELECTROSTATICS */
 870
 871             /* Analytical PME correction */
 872             zeta2            = _mm_mul_ps(beta2,rsq02);
 873             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 874             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 875             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 876             felec            = _mm_mul_ps(qq02,felec);
 877             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 878             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 879             velec            = _mm_mul_ps(qq02,velec);
 880
 881             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 882
 883             /* Update potential sum for this i atom from the interaction with this j atom. */
 884             velec            = _mm_and_ps(velec,cutoff_mask);
 885             velec            = _mm_andnot_ps(dummy_mask,velec);
 886             velecsum         = _mm_add_ps(velecsum,velec);
 887
 888             fscal            = felec;
 889
 890             fscal            = _mm_and_ps(fscal,cutoff_mask);
 891
 892             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 893
 894              /* Update vectorial force */
 895             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 896             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 897             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 898
 899             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 900             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 901             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 902
 903             }
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             if (gmx_mm_any_lt(rsq10,rcutoff2))
 910             {
 911
 912             r10              = _mm_mul_ps(rsq10,rinv10);
 913             r10              = _mm_andnot_ps(dummy_mask,r10);
 914
 915             /* EWALD ELECTROSTATICS */
 916
 917             /* Analytical PME correction */
 918             zeta2            = _mm_mul_ps(beta2,rsq10);
 919             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 920             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 921             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 922             felec            = _mm_mul_ps(qq10,felec);
 923             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 924             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 925             velec            = _mm_mul_ps(qq10,velec);
 926
 927             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 928
 929             /* Update potential sum for this i atom from the interaction with this j atom. */
 930             velec            = _mm_and_ps(velec,cutoff_mask);
 931             velec            = _mm_andnot_ps(dummy_mask,velec);
 932             velecsum         = _mm_add_ps(velecsum,velec);
 933
 934             fscal            = felec;
 935
 936             fscal            = _mm_and_ps(fscal,cutoff_mask);
 937
 938             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 939
 940              /* Update vectorial force */
 941             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 942             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 943             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 944
 945             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 946             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 947             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 948
 949             }
 950
 951             /**************************
 952              * CALCULATE INTERACTIONS *
 953              **************************/
 954
 955             if (gmx_mm_any_lt(rsq11,rcutoff2))
 956             {
 957
 958             r11              = _mm_mul_ps(rsq11,rinv11);
 959             r11              = _mm_andnot_ps(dummy_mask,r11);
 960
 961             /* EWALD ELECTROSTATICS */
 962
 963             /* Analytical PME correction */
 964             zeta2            = _mm_mul_ps(beta2,rsq11);
 965             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 966             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 967             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 968             felec            = _mm_mul_ps(qq11,felec);
 969             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 970             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
 971             velec            = _mm_mul_ps(qq11,velec);
 972
 973             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 974
 975             /* Update potential sum for this i atom from the interaction with this j atom. */
 976             velec            = _mm_and_ps(velec,cutoff_mask);
 977             velec            = _mm_andnot_ps(dummy_mask,velec);
 978             velecsum         = _mm_add_ps(velecsum,velec);
 979
 980             fscal            = felec;
 981
 982             fscal            = _mm_and_ps(fscal,cutoff_mask);
 983
 984             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 985
 986              /* Update vectorial force */
 987             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 988             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 989             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 990
 991             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 992             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 993             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 994
 995             }
 996
 997             /**************************
 998              * CALCULATE INTERACTIONS *
 999              **************************/
1000
1001             if (gmx_mm_any_lt(rsq12,rcutoff2))
1002             {
1003
1004             r12              = _mm_mul_ps(rsq12,rinv12);
1005             r12              = _mm_andnot_ps(dummy_mask,r12);
1006
1007             /* EWALD ELECTROSTATICS */
1008
1009             /* Analytical PME correction */
1010             zeta2            = _mm_mul_ps(beta2,rsq12);
1011             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1012             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1013             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1014             felec            = _mm_mul_ps(qq12,felec);
1015             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1016             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
1017             velec            = _mm_mul_ps(qq12,velec);
1018
1019             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1020
1021             /* Update potential sum for this i atom from the interaction with this j atom. */
1022             velec            = _mm_and_ps(velec,cutoff_mask);
1023             velec            = _mm_andnot_ps(dummy_mask,velec);
1024             velecsum         = _mm_add_ps(velecsum,velec);
1025
1026             fscal            = felec;
1027
1028             fscal            = _mm_and_ps(fscal,cutoff_mask);
1029
1030             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1031
1032              /* Update vectorial force */
1033             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1034             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1035             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1036
1037             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1038             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1039             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1040
1041             }
1042
1043             /**************************
1044              * CALCULATE INTERACTIONS *
1045              **************************/
1046
1047             if (gmx_mm_any_lt(rsq20,rcutoff2))
1048             {
1049
1050             r20              = _mm_mul_ps(rsq20,rinv20);
1051             r20              = _mm_andnot_ps(dummy_mask,r20);
1052
1053             /* EWALD ELECTROSTATICS */
1054
1055             /* Analytical PME correction */
1056             zeta2            = _mm_mul_ps(beta2,rsq20);
1057             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1058             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1059             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1060             felec            = _mm_mul_ps(qq20,felec);
1061             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1062             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
1063             velec            = _mm_mul_ps(qq20,velec);
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              /* Update vectorial force */
1079             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1080             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1081             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1082
1083             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1084             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1085             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1086
1087             }
1088
1089             /**************************
1090              * CALCULATE INTERACTIONS *
1091              **************************/
1092
1093             if (gmx_mm_any_lt(rsq21,rcutoff2))
1094             {
1095
1096             r21              = _mm_mul_ps(rsq21,rinv21);
1097             r21              = _mm_andnot_ps(dummy_mask,r21);
1098
1099             /* EWALD ELECTROSTATICS */
1100
1101             /* Analytical PME correction */
1102             zeta2            = _mm_mul_ps(beta2,rsq21);
1103             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1104             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1105             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1106             felec            = _mm_mul_ps(qq21,felec);
1107             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1108             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
1109             velec            = _mm_mul_ps(qq21,velec);
1110
1111             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1112
1113             /* Update potential sum for this i atom from the interaction with this j atom. */
1114             velec            = _mm_and_ps(velec,cutoff_mask);
1115             velec            = _mm_andnot_ps(dummy_mask,velec);
1116             velecsum         = _mm_add_ps(velecsum,velec);
1117
1118             fscal            = felec;
1119
1120             fscal            = _mm_and_ps(fscal,cutoff_mask);
1121
1122             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1123
1124              /* Update vectorial force */
1125             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1126             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1127             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1128
1129             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1130             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1131             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1132
1133             }
1134
1135             /**************************
1136              * CALCULATE INTERACTIONS *
1137              **************************/
1138
1139             if (gmx_mm_any_lt(rsq22,rcutoff2))
1140             {
1141
1142             r22              = _mm_mul_ps(rsq22,rinv22);
1143             r22              = _mm_andnot_ps(dummy_mask,r22);
1144
1145             /* EWALD ELECTROSTATICS */
1146
1147             /* Analytical PME correction */
1148             zeta2            = _mm_mul_ps(beta2,rsq22);
1149             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1150             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1151             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1152             felec            = _mm_mul_ps(qq22,felec);
1153             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1154             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
1155             velec            = _mm_mul_ps(qq22,velec);
1156
1157             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1158
1159             /* Update potential sum for this i atom from the interaction with this j atom. */
1160             velec            = _mm_and_ps(velec,cutoff_mask);
1161             velec            = _mm_andnot_ps(dummy_mask,velec);
1162             velecsum         = _mm_add_ps(velecsum,velec);
1163
1164             fscal            = felec;
1165
1166             fscal            = _mm_and_ps(fscal,cutoff_mask);
1167
1168             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1169
1170              /* Update vectorial force */
1171             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1172             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1173             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1174
1175             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1176             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1177             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1178
1179             }
1180
1181             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1182             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1183             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1184             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1185
1186             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1187                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1188
1189             /* Inner loop uses 306 flops */
1190         }
1191
1192         /* End of innermost loop */
1193
1194         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1195                                               f+i_coord_offset,fshift+i_shift_offset);
1196
1197         ggid                        = gid[iidx];
1198         /* Update potential energies */
1199         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1200
1201         /* Increment number of inner iterations */
1202         inneriter                  += j_index_end - j_index_start;
1203
1204         /* Outer loop uses 19 flops */
1205     }
1206
1207     /* Increment number of outer iterations */
1208     outeriter        += nri;
1209
1210     /* Update outer/inner flops */
1211
1212     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*306);
1213 }
1214 /*
1215  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW3W3_F_avx_128_fma_single
1216  * Electrostatics interaction: Ewald
1217  * VdW interaction:            None
1218  * Geometry:                   Water3-Water3
1219  * Calculate force/pot:        Force
1220  */
1221 void
1222 nb_kernel_ElecEwSh_VdwNone_GeomW3W3_F_avx_128_fma_single
1223                     (t_nblist * gmx_restrict                nlist,
1224                      rvec * gmx_restrict                    xx,
1225                      rvec * gmx_restrict                    ff,
1226                      t_forcerec * gmx_restrict              fr,
1227                      t_mdatoms * gmx_restrict               mdatoms,
1228                      nb_kernel_data_t * gmx_restrict        kernel_data,
1229                      t_nrnb * gmx_restrict                  nrnb)
1230 {
1231     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1232      * just 0 for non-waters.
1233      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1234      * jnr indices corresponding to data put in the four positions in the SIMD register.
1235      */
1236     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1237     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1238     int              jnrA,jnrB,jnrC,jnrD;
1239     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1240     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1241     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1242     real             rcutoff_scalar;
1243     real             *shiftvec,*fshift,*x,*f;
1244     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1245     real             scratch[4*DIM];
1246     __m128           fscal,rcutoff,rcutoff2,jidxall;
1247     int              vdwioffset0;
1248     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1249     int              vdwioffset1;
1250     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1251     int              vdwioffset2;
1252     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1253     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1254     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1255     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1256     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1257     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1258     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1259     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1260     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1261     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1262     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1263     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1264     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1265     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1266     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1267     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1268     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1269     real             *charge;
1270     __m128i          ewitab;
1271     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1272     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1273     real             *ewtab;
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
1292     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1293     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
1294     beta2            = _mm_mul_ps(beta,beta);
1295     beta3            = _mm_mul_ps(beta,beta2);
1296     ewtab            = fr->ic->tabq_coul_F;
1297     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1298     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1299
1300     /* Setup water-specific parameters */
1301     inr              = nlist->iinr[0];
1302     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1303     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1304     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1305
1306     jq0              = _mm_set1_ps(charge[inr+0]);
1307     jq1              = _mm_set1_ps(charge[inr+1]);
1308     jq2              = _mm_set1_ps(charge[inr+2]);
1309     qq00             = _mm_mul_ps(iq0,jq0);
1310     qq01             = _mm_mul_ps(iq0,jq1);
1311     qq02             = _mm_mul_ps(iq0,jq2);
1312     qq10             = _mm_mul_ps(iq1,jq0);
1313     qq11             = _mm_mul_ps(iq1,jq1);
1314     qq12             = _mm_mul_ps(iq1,jq2);
1315     qq20             = _mm_mul_ps(iq2,jq0);
1316     qq21             = _mm_mul_ps(iq2,jq1);
1317     qq22             = _mm_mul_ps(iq2,jq2);
1318
1319     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1320     rcutoff_scalar   = fr->rcoulomb;
1321     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1322     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1323
1324     /* Avoid stupid compiler warnings */
1325     jnrA = jnrB = jnrC = jnrD = 0;
1326     j_coord_offsetA = 0;
1327     j_coord_offsetB = 0;
1328     j_coord_offsetC = 0;
1329     j_coord_offsetD = 0;
1330
1331     outeriter        = 0;
1332     inneriter        = 0;
1333
1334     for(iidx=0;iidx<4*DIM;iidx++)
1335     {
1336         scratch[iidx] = 0.0;
1337     }
1338
1339     /* Start outer loop over neighborlists */
1340     for(iidx=0; iidx<nri; iidx++)
1341     {
1342         /* Load shift vector for this list */
1343         i_shift_offset   = DIM*shiftidx[iidx];
1344
1345         /* Load limits for loop over neighbors */
1346         j_index_start    = jindex[iidx];
1347         j_index_end      = jindex[iidx+1];
1348
1349         /* Get outer coordinate index */
1350         inr              = iinr[iidx];
1351         i_coord_offset   = DIM*inr;
1352
1353         /* Load i particle coords and add shift vector */
1354         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1355                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1356
1357         fix0             = _mm_setzero_ps();
1358         fiy0             = _mm_setzero_ps();
1359         fiz0             = _mm_setzero_ps();
1360         fix1             = _mm_setzero_ps();
1361         fiy1             = _mm_setzero_ps();
1362         fiz1             = _mm_setzero_ps();
1363         fix2             = _mm_setzero_ps();
1364         fiy2             = _mm_setzero_ps();
1365         fiz2             = _mm_setzero_ps();
1366
1367         /* Start inner kernel loop */
1368         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1369         {
1370
1371             /* Get j neighbor index, and coordinate index */
1372             jnrA             = jjnr[jidx];
1373             jnrB             = jjnr[jidx+1];
1374             jnrC             = jjnr[jidx+2];
1375             jnrD             = jjnr[jidx+3];
1376             j_coord_offsetA  = DIM*jnrA;
1377             j_coord_offsetB  = DIM*jnrB;
1378             j_coord_offsetC  = DIM*jnrC;
1379             j_coord_offsetD  = DIM*jnrD;
1380
1381             /* load j atom coordinates */
1382             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1383                                               x+j_coord_offsetC,x+j_coord_offsetD,
1384                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1385
1386             /* Calculate displacement vector */
1387             dx00             = _mm_sub_ps(ix0,jx0);
1388             dy00             = _mm_sub_ps(iy0,jy0);
1389             dz00             = _mm_sub_ps(iz0,jz0);
1390             dx01             = _mm_sub_ps(ix0,jx1);
1391             dy01             = _mm_sub_ps(iy0,jy1);
1392             dz01             = _mm_sub_ps(iz0,jz1);
1393             dx02             = _mm_sub_ps(ix0,jx2);
1394             dy02             = _mm_sub_ps(iy0,jy2);
1395             dz02             = _mm_sub_ps(iz0,jz2);
1396             dx10             = _mm_sub_ps(ix1,jx0);
1397             dy10             = _mm_sub_ps(iy1,jy0);
1398             dz10             = _mm_sub_ps(iz1,jz0);
1399             dx11             = _mm_sub_ps(ix1,jx1);
1400             dy11             = _mm_sub_ps(iy1,jy1);
1401             dz11             = _mm_sub_ps(iz1,jz1);
1402             dx12             = _mm_sub_ps(ix1,jx2);
1403             dy12             = _mm_sub_ps(iy1,jy2);
1404             dz12             = _mm_sub_ps(iz1,jz2);
1405             dx20             = _mm_sub_ps(ix2,jx0);
1406             dy20             = _mm_sub_ps(iy2,jy0);
1407             dz20             = _mm_sub_ps(iz2,jz0);
1408             dx21             = _mm_sub_ps(ix2,jx1);
1409             dy21             = _mm_sub_ps(iy2,jy1);
1410             dz21             = _mm_sub_ps(iz2,jz1);
1411             dx22             = _mm_sub_ps(ix2,jx2);
1412             dy22             = _mm_sub_ps(iy2,jy2);
1413             dz22             = _mm_sub_ps(iz2,jz2);
1414
1415             /* Calculate squared distance and things based on it */
1416             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1417             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1418             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1419             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1420             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1421             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1422             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1423             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1424             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1425
1426             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1427             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1428             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1429             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1430             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1431             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1432             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1433             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1434             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1435
1436             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1437             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1438             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1439             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1440             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1441             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1442             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1443             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1444             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1445
1446             fjx0             = _mm_setzero_ps();
1447             fjy0             = _mm_setzero_ps();
1448             fjz0             = _mm_setzero_ps();
1449             fjx1             = _mm_setzero_ps();
1450             fjy1             = _mm_setzero_ps();
1451             fjz1             = _mm_setzero_ps();
1452             fjx2             = _mm_setzero_ps();
1453             fjy2             = _mm_setzero_ps();
1454             fjz2             = _mm_setzero_ps();
1455
1456             /**************************
1457              * CALCULATE INTERACTIONS *
1458              **************************/
1459
1460             if (gmx_mm_any_lt(rsq00,rcutoff2))
1461             {
1462
1463             r00              = _mm_mul_ps(rsq00,rinv00);
1464
1465             /* EWALD ELECTROSTATICS */
1466
1467             /* Analytical PME correction */
1468             zeta2            = _mm_mul_ps(beta2,rsq00);
1469             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1470             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1471             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1472             felec            = _mm_mul_ps(qq00,felec);
1473
1474             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1475
1476             fscal            = felec;
1477
1478             fscal            = _mm_and_ps(fscal,cutoff_mask);
1479
1480              /* Update vectorial force */
1481             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1482             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1483             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1484
1485             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1486             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1487             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1488
1489             }
1490
1491             /**************************
1492              * CALCULATE INTERACTIONS *
1493              **************************/
1494
1495             if (gmx_mm_any_lt(rsq01,rcutoff2))
1496             {
1497
1498             r01              = _mm_mul_ps(rsq01,rinv01);
1499
1500             /* EWALD ELECTROSTATICS */
1501
1502             /* Analytical PME correction */
1503             zeta2            = _mm_mul_ps(beta2,rsq01);
1504             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1505             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1506             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1507             felec            = _mm_mul_ps(qq01,felec);
1508
1509             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1510
1511             fscal            = felec;
1512
1513             fscal            = _mm_and_ps(fscal,cutoff_mask);
1514
1515              /* Update vectorial force */
1516             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1517             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1518             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1519
1520             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1521             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1522             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1523
1524             }
1525
1526             /**************************
1527              * CALCULATE INTERACTIONS *
1528              **************************/
1529
1530             if (gmx_mm_any_lt(rsq02,rcutoff2))
1531             {
1532
1533             r02              = _mm_mul_ps(rsq02,rinv02);
1534
1535             /* EWALD ELECTROSTATICS */
1536
1537             /* Analytical PME correction */
1538             zeta2            = _mm_mul_ps(beta2,rsq02);
1539             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1540             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1541             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1542             felec            = _mm_mul_ps(qq02,felec);
1543
1544             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_and_ps(fscal,cutoff_mask);
1549
1550              /* Update vectorial force */
1551             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1552             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1553             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1554
1555             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1556             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1557             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1558
1559             }
1560
1561             /**************************
1562              * CALCULATE INTERACTIONS *
1563              **************************/
1564
1565             if (gmx_mm_any_lt(rsq10,rcutoff2))
1566             {
1567
1568             r10              = _mm_mul_ps(rsq10,rinv10);
1569
1570             /* EWALD ELECTROSTATICS */
1571
1572             /* Analytical PME correction */
1573             zeta2            = _mm_mul_ps(beta2,rsq10);
1574             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1575             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1576             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1577             felec            = _mm_mul_ps(qq10,felec);
1578
1579             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1580
1581             fscal            = felec;
1582
1583             fscal            = _mm_and_ps(fscal,cutoff_mask);
1584
1585              /* Update vectorial force */
1586             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1587             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1588             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1589
1590             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1591             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1592             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1593
1594             }
1595
1596             /**************************
1597              * CALCULATE INTERACTIONS *
1598              **************************/
1599
1600             if (gmx_mm_any_lt(rsq11,rcutoff2))
1601             {
1602
1603             r11              = _mm_mul_ps(rsq11,rinv11);
1604
1605             /* EWALD ELECTROSTATICS */
1606
1607             /* Analytical PME correction */
1608             zeta2            = _mm_mul_ps(beta2,rsq11);
1609             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1610             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1611             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1612             felec            = _mm_mul_ps(qq11,felec);
1613
1614             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1615
1616             fscal            = felec;
1617
1618             fscal            = _mm_and_ps(fscal,cutoff_mask);
1619
1620              /* Update vectorial force */
1621             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1622             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1623             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1624
1625             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1626             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1627             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1628
1629             }
1630
1631             /**************************
1632              * CALCULATE INTERACTIONS *
1633              **************************/
1634
1635             if (gmx_mm_any_lt(rsq12,rcutoff2))
1636             {
1637
1638             r12              = _mm_mul_ps(rsq12,rinv12);
1639
1640             /* EWALD ELECTROSTATICS */
1641
1642             /* Analytical PME correction */
1643             zeta2            = _mm_mul_ps(beta2,rsq12);
1644             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1645             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1646             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1647             felec            = _mm_mul_ps(qq12,felec);
1648
1649             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1650
1651             fscal            = felec;
1652
1653             fscal            = _mm_and_ps(fscal,cutoff_mask);
1654
1655              /* Update vectorial force */
1656             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1657             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1658             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1659
1660             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1661             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1662             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1663
1664             }
1665
1666             /**************************
1667              * CALCULATE INTERACTIONS *
1668              **************************/
1669
1670             if (gmx_mm_any_lt(rsq20,rcutoff2))
1671             {
1672
1673             r20              = _mm_mul_ps(rsq20,rinv20);
1674
1675             /* EWALD ELECTROSTATICS */
1676
1677             /* Analytical PME correction */
1678             zeta2            = _mm_mul_ps(beta2,rsq20);
1679             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1680             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1681             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1682             felec            = _mm_mul_ps(qq20,felec);
1683
1684             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1685
1686             fscal            = felec;
1687
1688             fscal            = _mm_and_ps(fscal,cutoff_mask);
1689
1690              /* Update vectorial force */
1691             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1692             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1693             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1694
1695             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1696             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1697             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1698
1699             }
1700
1701             /**************************
1702              * CALCULATE INTERACTIONS *
1703              **************************/
1704
1705             if (gmx_mm_any_lt(rsq21,rcutoff2))
1706             {
1707
1708             r21              = _mm_mul_ps(rsq21,rinv21);
1709
1710             /* EWALD ELECTROSTATICS */
1711
1712             /* Analytical PME correction */
1713             zeta2            = _mm_mul_ps(beta2,rsq21);
1714             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1715             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1716             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1717             felec            = _mm_mul_ps(qq21,felec);
1718
1719             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1720
1721             fscal            = felec;
1722
1723             fscal            = _mm_and_ps(fscal,cutoff_mask);
1724
1725              /* Update vectorial force */
1726             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1727             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1728             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1729
1730             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1731             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1732             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1733
1734             }
1735
1736             /**************************
1737              * CALCULATE INTERACTIONS *
1738              **************************/
1739
1740             if (gmx_mm_any_lt(rsq22,rcutoff2))
1741             {
1742
1743             r22              = _mm_mul_ps(rsq22,rinv22);
1744
1745             /* EWALD ELECTROSTATICS */
1746
1747             /* Analytical PME correction */
1748             zeta2            = _mm_mul_ps(beta2,rsq22);
1749             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1750             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1751             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1752             felec            = _mm_mul_ps(qq22,felec);
1753
1754             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1755
1756             fscal            = felec;
1757
1758             fscal            = _mm_and_ps(fscal,cutoff_mask);
1759
1760              /* Update vectorial force */
1761             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1762             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1763             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1764
1765             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1766             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1767             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1768
1769             }
1770
1771             fjptrA             = f+j_coord_offsetA;
1772             fjptrB             = f+j_coord_offsetB;
1773             fjptrC             = f+j_coord_offsetC;
1774             fjptrD             = f+j_coord_offsetD;
1775
1776             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1777                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1778
1779             /* Inner loop uses 279 flops */
1780         }
1781
1782         if(jidx<j_index_end)
1783         {
1784
1785             /* Get j neighbor index, and coordinate index */
1786             jnrlistA         = jjnr[jidx];
1787             jnrlistB         = jjnr[jidx+1];
1788             jnrlistC         = jjnr[jidx+2];
1789             jnrlistD         = jjnr[jidx+3];
1790             /* Sign of each element will be negative for non-real atoms.
1791              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1792              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1793              */
1794             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1795             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1796             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1797             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1798             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1799             j_coord_offsetA  = DIM*jnrA;
1800             j_coord_offsetB  = DIM*jnrB;
1801             j_coord_offsetC  = DIM*jnrC;
1802             j_coord_offsetD  = DIM*jnrD;
1803
1804             /* load j atom coordinates */
1805             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1806                                               x+j_coord_offsetC,x+j_coord_offsetD,
1807                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1808
1809             /* Calculate displacement vector */
1810             dx00             = _mm_sub_ps(ix0,jx0);
1811             dy00             = _mm_sub_ps(iy0,jy0);
1812             dz00             = _mm_sub_ps(iz0,jz0);
1813             dx01             = _mm_sub_ps(ix0,jx1);
1814             dy01             = _mm_sub_ps(iy0,jy1);
1815             dz01             = _mm_sub_ps(iz0,jz1);
1816             dx02             = _mm_sub_ps(ix0,jx2);
1817             dy02             = _mm_sub_ps(iy0,jy2);
1818             dz02             = _mm_sub_ps(iz0,jz2);
1819             dx10             = _mm_sub_ps(ix1,jx0);
1820             dy10             = _mm_sub_ps(iy1,jy0);
1821             dz10             = _mm_sub_ps(iz1,jz0);
1822             dx11             = _mm_sub_ps(ix1,jx1);
1823             dy11             = _mm_sub_ps(iy1,jy1);
1824             dz11             = _mm_sub_ps(iz1,jz1);
1825             dx12             = _mm_sub_ps(ix1,jx2);
1826             dy12             = _mm_sub_ps(iy1,jy2);
1827             dz12             = _mm_sub_ps(iz1,jz2);
1828             dx20             = _mm_sub_ps(ix2,jx0);
1829             dy20             = _mm_sub_ps(iy2,jy0);
1830             dz20             = _mm_sub_ps(iz2,jz0);
1831             dx21             = _mm_sub_ps(ix2,jx1);
1832             dy21             = _mm_sub_ps(iy2,jy1);
1833             dz21             = _mm_sub_ps(iz2,jz1);
1834             dx22             = _mm_sub_ps(ix2,jx2);
1835             dy22             = _mm_sub_ps(iy2,jy2);
1836             dz22             = _mm_sub_ps(iz2,jz2);
1837
1838             /* Calculate squared distance and things based on it */
1839             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1840             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1841             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1842             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1843             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1844             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1845             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1846             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1847             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1848
1849             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1850             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1851             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1852             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1853             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1854             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1855             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1856             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1857             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1858
1859             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1860             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1861             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1862             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1863             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1864             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1865             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1866             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1867             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1868
1869             fjx0             = _mm_setzero_ps();
1870             fjy0             = _mm_setzero_ps();
1871             fjz0             = _mm_setzero_ps();
1872             fjx1             = _mm_setzero_ps();
1873             fjy1             = _mm_setzero_ps();
1874             fjz1             = _mm_setzero_ps();
1875             fjx2             = _mm_setzero_ps();
1876             fjy2             = _mm_setzero_ps();
1877             fjz2             = _mm_setzero_ps();
1878
1879             /**************************
1880              * CALCULATE INTERACTIONS *
1881              **************************/
1882
1883             if (gmx_mm_any_lt(rsq00,rcutoff2))
1884             {
1885
1886             r00              = _mm_mul_ps(rsq00,rinv00);
1887             r00              = _mm_andnot_ps(dummy_mask,r00);
1888
1889             /* EWALD ELECTROSTATICS */
1890
1891             /* Analytical PME correction */
1892             zeta2            = _mm_mul_ps(beta2,rsq00);
1893             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1894             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1895             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1896             felec            = _mm_mul_ps(qq00,felec);
1897
1898             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1899
1900             fscal            = felec;
1901
1902             fscal            = _mm_and_ps(fscal,cutoff_mask);
1903
1904             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1905
1906              /* Update vectorial force */
1907             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1908             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1909             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1910
1911             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1912             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1913             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1914
1915             }
1916
1917             /**************************
1918              * CALCULATE INTERACTIONS *
1919              **************************/
1920
1921             if (gmx_mm_any_lt(rsq01,rcutoff2))
1922             {
1923
1924             r01              = _mm_mul_ps(rsq01,rinv01);
1925             r01              = _mm_andnot_ps(dummy_mask,r01);
1926
1927             /* EWALD ELECTROSTATICS */
1928
1929             /* Analytical PME correction */
1930             zeta2            = _mm_mul_ps(beta2,rsq01);
1931             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1932             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1933             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1934             felec            = _mm_mul_ps(qq01,felec);
1935
1936             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1937
1938             fscal            = felec;
1939
1940             fscal            = _mm_and_ps(fscal,cutoff_mask);
1941
1942             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1943
1944              /* Update vectorial force */
1945             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1946             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1947             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1948
1949             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1950             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1951             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1952
1953             }
1954
1955             /**************************
1956              * CALCULATE INTERACTIONS *
1957              **************************/
1958
1959             if (gmx_mm_any_lt(rsq02,rcutoff2))
1960             {
1961
1962             r02              = _mm_mul_ps(rsq02,rinv02);
1963             r02              = _mm_andnot_ps(dummy_mask,r02);
1964
1965             /* EWALD ELECTROSTATICS */
1966
1967             /* Analytical PME correction */
1968             zeta2            = _mm_mul_ps(beta2,rsq02);
1969             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1970             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1971             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1972             felec            = _mm_mul_ps(qq02,felec);
1973
1974             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1975
1976             fscal            = felec;
1977
1978             fscal            = _mm_and_ps(fscal,cutoff_mask);
1979
1980             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1981
1982              /* Update vectorial force */
1983             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1984             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1985             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1986
1987             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1988             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1989             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1990
1991             }
1992
1993             /**************************
1994              * CALCULATE INTERACTIONS *
1995              **************************/
1996
1997             if (gmx_mm_any_lt(rsq10,rcutoff2))
1998             {
1999
2000             r10              = _mm_mul_ps(rsq10,rinv10);
2001             r10              = _mm_andnot_ps(dummy_mask,r10);
2002
2003             /* EWALD ELECTROSTATICS */
2004
2005             /* Analytical PME correction */
2006             zeta2            = _mm_mul_ps(beta2,rsq10);
2007             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
2008             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2009             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2010             felec            = _mm_mul_ps(qq10,felec);
2011
2012             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
2013
2014             fscal            = felec;
2015
2016             fscal            = _mm_and_ps(fscal,cutoff_mask);
2017
2018             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2019
2020              /* Update vectorial force */
2021             fix1             = _mm_macc_ps(dx10,fscal,fix1);
2022             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
2023             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
2024
2025             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
2026             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
2027             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
2028
2029             }
2030
2031             /**************************
2032              * CALCULATE INTERACTIONS *
2033              **************************/
2034
2035             if (gmx_mm_any_lt(rsq11,rcutoff2))
2036             {
2037
2038             r11              = _mm_mul_ps(rsq11,rinv11);
2039             r11              = _mm_andnot_ps(dummy_mask,r11);
2040
2041             /* EWALD ELECTROSTATICS */
2042
2043             /* Analytical PME correction */
2044             zeta2            = _mm_mul_ps(beta2,rsq11);
2045             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
2046             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2047             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2048             felec            = _mm_mul_ps(qq11,felec);
2049
2050             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2051
2052             fscal            = felec;
2053
2054             fscal            = _mm_and_ps(fscal,cutoff_mask);
2055
2056             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2057
2058              /* Update vectorial force */
2059             fix1             = _mm_macc_ps(dx11,fscal,fix1);
2060             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
2061             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
2062
2063             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
2064             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
2065             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
2066
2067             }
2068
2069             /**************************
2070              * CALCULATE INTERACTIONS *
2071              **************************/
2072
2073             if (gmx_mm_any_lt(rsq12,rcutoff2))
2074             {
2075
2076             r12              = _mm_mul_ps(rsq12,rinv12);
2077             r12              = _mm_andnot_ps(dummy_mask,r12);
2078
2079             /* EWALD ELECTROSTATICS */
2080
2081             /* Analytical PME correction */
2082             zeta2            = _mm_mul_ps(beta2,rsq12);
2083             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
2084             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2085             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2086             felec            = _mm_mul_ps(qq12,felec);
2087
2088             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2089
2090             fscal            = felec;
2091
2092             fscal            = _mm_and_ps(fscal,cutoff_mask);
2093
2094             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2095
2096              /* Update vectorial force */
2097             fix1             = _mm_macc_ps(dx12,fscal,fix1);
2098             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
2099             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
2100
2101             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
2102             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
2103             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
2104
2105             }
2106
2107             /**************************
2108              * CALCULATE INTERACTIONS *
2109              **************************/
2110
2111             if (gmx_mm_any_lt(rsq20,rcutoff2))
2112             {
2113
2114             r20              = _mm_mul_ps(rsq20,rinv20);
2115             r20              = _mm_andnot_ps(dummy_mask,r20);
2116
2117             /* EWALD ELECTROSTATICS */
2118
2119             /* Analytical PME correction */
2120             zeta2            = _mm_mul_ps(beta2,rsq20);
2121             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
2122             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2123             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2124             felec            = _mm_mul_ps(qq20,felec);
2125
2126             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2127
2128             fscal            = felec;
2129
2130             fscal            = _mm_and_ps(fscal,cutoff_mask);
2131
2132             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2133
2134              /* Update vectorial force */
2135             fix2             = _mm_macc_ps(dx20,fscal,fix2);
2136             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
2137             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
2138
2139             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
2140             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
2141             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
2142
2143             }
2144
2145             /**************************
2146              * CALCULATE INTERACTIONS *
2147              **************************/
2148
2149             if (gmx_mm_any_lt(rsq21,rcutoff2))
2150             {
2151
2152             r21              = _mm_mul_ps(rsq21,rinv21);
2153             r21              = _mm_andnot_ps(dummy_mask,r21);
2154
2155             /* EWALD ELECTROSTATICS */
2156
2157             /* Analytical PME correction */
2158             zeta2            = _mm_mul_ps(beta2,rsq21);
2159             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
2160             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2161             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2162             felec            = _mm_mul_ps(qq21,felec);
2163
2164             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2165
2166             fscal            = felec;
2167
2168             fscal            = _mm_and_ps(fscal,cutoff_mask);
2169
2170             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2171
2172              /* Update vectorial force */
2173             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2174             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2175             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2176
2177             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2178             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2179             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2180
2181             }
2182
2183             /**************************
2184              * CALCULATE INTERACTIONS *
2185              **************************/
2186
2187             if (gmx_mm_any_lt(rsq22,rcutoff2))
2188             {
2189
2190             r22              = _mm_mul_ps(rsq22,rinv22);
2191             r22              = _mm_andnot_ps(dummy_mask,r22);
2192
2193             /* EWALD ELECTROSTATICS */
2194
2195             /* Analytical PME correction */
2196             zeta2            = _mm_mul_ps(beta2,rsq22);
2197             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
2198             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2199             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2200             felec            = _mm_mul_ps(qq22,felec);
2201
2202             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2203
2204             fscal            = felec;
2205
2206             fscal            = _mm_and_ps(fscal,cutoff_mask);
2207
2208             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2209
2210              /* Update vectorial force */
2211             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2212             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2213             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2214
2215             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2216             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2217             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2218
2219             }
2220
2221             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2222             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2223             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2224             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2225
2226             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2227                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2228
2229             /* Inner loop uses 288 flops */
2230         }
2231
2232         /* End of innermost loop */
2233
2234         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2235                                               f+i_coord_offset,fshift+i_shift_offset);
2236
2237         /* Increment number of inner iterations */
2238         inneriter                  += j_index_end - j_index_start;
2239
2240         /* Outer loop uses 18 flops */
2241     }
2242
2243     /* Increment number of outer iterations */
2244     outeriter        += nri;
2245
2246     /* Update outer/inner flops */
2247
2248     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*288);
2249 }